diff --git a/.dockerignore b/.dockerignore
index 43f2e07dd..fada7a9b4 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -7,3 +7,5 @@ llm/llama.cpp
 .env
 .cache
 test_data
+llm/build
+llama/build
diff --git a/.gitattributes b/.gitattributes
index a8436e9cf..f1c8bcb4d 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -1 +1,3 @@
 llm/ext_server/* linguist-vendored
+* text=auto
+*.go text eol=lf
diff --git a/.github/workflows/release.yaml b/.github/workflows/release.yaml
index 61ca3c433..4090f2066 100644
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -102,7 +102,8 @@ jobs:
         with:
           name: generate-windows-cpu
           path: |
-            llm/build/**/bin/*
+            build/**/*
+            build/**/*.a
             llm/build/**/*.a
             dist/windows-amd64/**
 
@@ -147,7 +148,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"
@@ -176,7 +177,7 @@ jobs:
         with:
           name: generate-windows-rocm
           path: |
-            llm/build/**/bin/*
+            build/**/*
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
@@ -187,6 +188,13 @@ jobs:
   generate-windows-cuda:
     environment: release
     runs-on: windows
+    strategy:
+      matrix:
+        cuda:
+          - version: "11"
+            url: 'https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe'
+          - version: "12"
+            url: 'https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe'
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -220,11 +228,11 @@ jobs:
         with:
           go-version-file: go.mod
           cache: true
-      - name: 'Install CUDA'
+      - name: 'Install CUDA ${{ matrix.cuda.version }}'
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading CUDA Installer"
-          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
           write-host "Installing CUDA"
           Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
           write-host "Completed CUDA"
@@ -256,16 +264,144 @@ jobs:
           cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
       - uses: actions/upload-artifact@v4
         with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
-            llm/build/**/bin/*
+            build/**/*
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
           path: dist/deps/*
 
-  # Import the prior generation steps and build the final windows assets
+
+  # windows arm64 generate, go build, and zip file (no installer)
+  # Output of this build is aggregated into the final x86 build
+  # for a unified windows installer
+  windows-arm64:
+    runs-on: windows-arm64
+    environment: release
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      # The current Windows arm64 beta image has effectively zero dev tools installed...
+      - name: Install git and gzip
+        run: |
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
+          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
+          choco install -y --no-progress git gzip
+          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install Visual Studio 2022
+        run: |
+          $components = @(
+            "Microsoft.VisualStudio.Component.CoreEditor",
+            "Microsoft.VisualStudio.Workload.CoreEditor",
+            "Microsoft.VisualStudio.Component.Roslyn.Compiler",
+            "Microsoft.Component.MSBuild",
+            "Microsoft.VisualStudio.Component.TextTemplating",
+            "Microsoft.VisualStudio.Component.Debugger.JustInTime",
+            "Microsoft.VisualStudio.Component.VC.CoreIde",
+            "Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+            "Microsoft.VisualStudio.Component.Windows11SDK.22621",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64EC",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64",
+            "Microsoft.VisualStudio.Component.VC.ATL",
+            "Microsoft.VisualStudio.Component.VC.ATL.ARM64",
+            "Microsoft.VisualStudio.Component.Graphics",
+            "Microsoft.VisualStudio.Component.VC.Redist.14.Latest",
+            "Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Core",
+            "Microsoft.VisualStudio.Component.Windows11Sdk.WindowsPerformanceToolkit",
+            "Microsoft.VisualStudio.Component.CppBuildInsights",
+            "Microsoft.VisualStudio.Component.VC.DiagnosticTools",
+            "Microsoft.VisualStudio.ComponentGroup.WebToolsExtensions.CMake",
+            "Microsoft.VisualStudio.Component.VC.CMake.Project",
+            "Microsoft.VisualStudio.Component.VC.ASAN",
+            "Microsoft.VisualStudio.Component.Vcpkg",
+            "Microsoft.VisualStudio.Workload.NativeDesktop"
+          )
+          $config = @{
+                "version" = "1.0"
+                "components"  = $components
+                "extensions"  = @()
+            }
+          $configPath = "${env:RUNNER_TEMP}\vsconfig"
+          $config | ConvertTo-Json | Out-File -FilePath $configPath
+          $bootstrapperFilePath = "${env:RUNNER_TEMP}\vs_community.exe"
+          write-host "Downloading Visual Studio 2022"
+          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vs_community.exe" -outfile $bootstrapperFilePath
+          $bootstrapperArgumentList = ('/c', $bootstrapperFilePath, '--config', $configPath, '--quiet', '--wait' )
+          write-host "Installing Visual Studio 2022"
+          $process = Start-Process -FilePath cmd.exe -ArgumentList $bootstrapperArgumentList -Wait -PassThru
+          $exitCode = $process.ExitCode
+          write-host $exitCode
+      # pacman in mingw/msys2 is ~broken on windows arm right now - hangs consistently during attempts to install
+      # so we'll use this alternative GCC binary
+      - name: Install llvm-mingw GCC
+        run: |
+          $gcc_url="https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-aarch64.zip"
+          write-host "Downloading llvm-mingw"
+          Invoke-WebRequest -Uri "${gcc_url}" -OutFile "${env:RUNNER_TEMP}\gcc.zip"
+          write-host "Unpacking llvm-mingw"
+          expand-archive -path "${env:RUNNER_TEMP}\gcc.zip" -destinationpath "c:\"
+          mv c:\llvm-mingw-* c:\llvm-mingw
+          echo "c:\llvm-mingw\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Verify GCC
+        run: |
+          echo $env:PATH
+          gcc --version
+      - uses: actions/checkout@v4
+      - name: Set Version
+        run: |
+          $ver=${env:GITHUB_REF_NAME}.trim("v")
+          write-host VERSION=$ver | Out-File -FilePath ${env:GITHUB_ENV} -Encoding utf8 -Append
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" | Out-File -FilePath ollama_inc.crt -Encoding utf8
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
+          & "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin"
+          echo $env:PATH
+          $env:ARCH="arm64"
+          .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
+        name: 'Windows Build'
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-arm64
+          path: |
+            dist/windows-arm64/**
+            dist/windows-arm64-app.exe
+            dist/ollama-windows-arm64.zip
+
+  # Import the prior generation steps plus the full arm64 build, and build the final windows assets
   build-windows:
     environment: release
     runs-on: windows
@@ -273,6 +409,7 @@ jobs:
       - generate-windows-cuda
       - generate-windows-rocm
       - generate-windows-cpu
+      - windows-arm64
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -314,17 +451,27 @@ jobs:
           name: generate-windows-cpu
       - uses: actions/download-artifact@v4
         with:
-          name: generate-windows-cuda
+          name: generate-windows-cuda-11
       - uses: actions/download-artifact@v4
         with:
-          name: windows-cuda-deps
+          name: generate-windows-cuda-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-11
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-12
       - uses: actions/download-artifact@v4
         with:
           name: windows-rocm-deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
-      - run: dir llm/build
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-arm64
+          path: dist
+      - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent
           & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
@@ -345,9 +492,7 @@ jobs:
     environment: release
     runs-on: linux
     env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
-      BUILD_ARCH: amd64
-      PUSH: '1'
+      PLATFORM: linux/amd64
     steps:
       - uses: actions/checkout@v4
         with:
@@ -355,15 +500,8 @@ jobs:
       - name: Set Version
         shell: bash
         run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
       - run: |
           ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
-          mv dist/deps/* dist/
       - uses: actions/upload-artifact@v4
         with:
           name: dist-linux-amd64
@@ -377,9 +515,7 @@ jobs:
     environment: release
     runs-on: linux-arm64
     env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
-      BUILD_ARCH: arm64
-      PUSH: '1'
+      PLATFORM: linux/arm64
     steps:
       - uses: actions/checkout@v4
         with:
@@ -408,14 +544,8 @@ jobs:
           sudo usermod -aG docker $USER
           sudo apt-get install acl
           sudo setfacl --modify user:$USER:rw /var/run/docker.sock
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
       - run: |
           ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
       - uses: actions/upload-artifact@v4
         with:
           name: dist-linux-arm64
@@ -423,6 +553,178 @@ jobs:
             dist/*linux*
             !dist/*-cov
 
+  # Container image build
+  build-container-image:
+    environment: release
+    strategy:
+      matrix:
+        runner:
+          - linux
+          - linux-arm64
+    runs-on: ${{ matrix.runner }}
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: 'Install Docker'
+        if: ${{ startsWith(matrix.runner, 'linux-arm64') }}
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y ca-certificates curl
+          sudo install -m 0755 -d /etc/apt/keyrings
+          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
+          sudo chmod a+r /etc/apt/keyrings/docker.asc
+          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
+            $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
+            sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
+          sudo apt-get update
+          sudo apt-get install -y docker-ce docker-ce-cli containerd.io
+          sudo usermod -aG docker $USER
+          sudo apt-get install acl
+          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          platforms: linux/${{ env.ARCH }}
+          build-args: |
+            GOFLAGS
+          outputs: type=image,name=${{ env.FINAL_IMAGE_REPO }},push-by-digest=true,name-canonical=true,push=true
+      - name: Export digest
+        run: |
+          mkdir -p /tmp/digests
+          digest="${{ steps.build.outputs.digest }}"
+          touch "/tmp/digests/${digest#sha256:}"
+      - name: Upload digest
+        uses: actions/upload-artifact@v4
+        with:
+          name: digests-${{ env.PLATFORM_PAIR }}
+          path: /tmp/digests/*
+          if-no-files-found: error
+          retention-days: 1
+  merge:
+    environment: release
+    runs-on: linux
+    needs:
+      - build-container-image
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Download digests
+        uses: actions/download-artifact@v4
+        with:
+          path: /tmp/digests
+          pattern: digests-*
+          merge-multiple: true
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Create manifest list and push
+        working-directory: /tmp/digests
+        run: |
+          docker buildx imagetools create $(jq -cr '.tags | map("-t " + .) | join(" ")' <<< "$DOCKER_METADATA_OUTPUT_JSON") \
+            $(printf '${{ env.FINAL_IMAGE_REPO }}@sha256:%s ' *)
+      - name: Inspect image
+        run: |
+          docker buildx imagetools inspect ${{ env.FINAL_IMAGE_REPO }}:${{ steps.meta.outputs.version }}          
+  build-container-image-rocm:
+    environment: release
+    runs-on: linux
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+      ARCH: amd64
+      PLATFORM_PAIR: linux-amd64
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          target: runtime-rocm
+          build-args: |
+            GOFLAGS
+          tags: ${{ env.FINAL_IMAGE_REPO }}:${{ env.DOCKER_METADATA_OUTPUT_VERSION}}-rocm
+          push: true
+
   # Aggregate all the assets and ship a release
   release:
     needs:
@@ -435,8 +737,6 @@ jobs:
     permissions:
       contents: write
     env:
-      OLLAMA_SKIP_IMAGE_BUILD: '1'
-      PUSH: '1'
       GH_TOKEN: ${{ github.token }}
     steps:
       - uses: actions/checkout@v4
@@ -445,12 +745,6 @@ jobs:
         run: |
           echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
           echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
-      - run: ./scripts/build_docker.sh
       - name: Retrieve built artifact
         uses: actions/download-artifact@v4
         with:
@@ -459,7 +753,8 @@ jobs:
           merge-multiple: true
       - run: |
           ls -lh dist/
-          (cd dist; sha256sum * > sha256sum.txt)
+          (cd dist; find . -type f | xargs sha256sum > ../sha256sum.txt)
+          mv sha256sum.txt dist/
           cat dist/sha256sum.txt
       - name: Create or update Release
         run: |
diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml
index 29adf56f3..26dc732af 100644
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -58,6 +58,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
@@ -79,12 +80,7 @@ jobs:
       - run: go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
         name: 'Unix Go Generate'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: |
-            llm/build/**/bin/*
-            llm/build/**/*.a
+      - run: go build .
   generate-cuda:
     needs: [changes]
     if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
@@ -112,19 +108,13 @@ jobs:
           go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: cuda-${{ matrix.cuda-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**
   generate-rocm:
     needs: [changes]
     if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
     strategy:
       matrix:
         rocm-version:
-          - '6.1.1'
+          - '6.1.2'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
     steps:
@@ -145,12 +135,6 @@ jobs:
           go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: rocm-${{ matrix.rocm-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**
 
   # ROCm generation step
   generate-windows-rocm:
@@ -167,7 +151,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"
@@ -187,7 +171,6 @@ jobs:
         name: go generate
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
-      # TODO - do we need any artifacts?
 
   # CUDA generation step
   generate-windows-cuda:
@@ -229,7 +212,6 @@ jobs:
           go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
-      # TODO - do we need any artifacts?
 
   lint:
     strategy:
@@ -261,17 +243,9 @@ jobs:
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
       - uses: golangci/golangci-lint-action@v6
         with:
-          args: --timeout 8m0s -v ${{ startsWith(matrix.os, 'windows-') && '' || '--disable gofmt --disable goimports' }}
+          args: --timeout 8m0s -v
   test:
     strategy:
       matrix:
@@ -299,23 +273,10 @@ jobs:
           cache: true
       - run: |
           case ${{ matrix.arch }} in
-            amd64) echo ARCH=x86_64 ;;
+            amd64) echo ARCH=amd64 ;;
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
-        shell: bash
       - run: go generate ./...
       - run: go build
       - run: go test -v ./...
-      - uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.os }}-binaries
-          path: ollama
diff --git a/.gitignore b/.gitignore
index 0d826ab61..87f8b0072 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,4 +12,7 @@ ggml-metal.metal
 test_data
 *.crt
 llm/build
+build/*/*/*
+!build/**/placeholder
+llama/build
 __debug_bin*
\ No newline at end of file
diff --git a/.golangci.yaml b/.golangci.yaml
index cfe06e07a..2e0ed3c7b 100644
--- a/.golangci.yaml
+++ b/.golangci.yaml
@@ -7,22 +7,35 @@ linters:
     - bodyclose
     - containedctx
     - contextcheck
+    - errcheck
     - exportloopref
+    - gci
     - gocheckcompilerdirectives
-    # conditionally enable this on linux/macos
-    # - gofmt
-    # - goimports
+    - gofmt
+    - gofumpt
+    - gosimple
+    - govet
+    - ineffassign
     - intrange
+    - makezero
     - misspell
     - nilerr
     - nolintlint
     - nosprintfhostport
-    - testifylint
+    - staticcheck
+    - tenv
     - unconvert
     - unused
+    - usestdlibvars
     - wastedassign
     - whitespace
-    - usestdlibvars
+linters-settings:
+  gci:
+    sections: [standard, default, localmodule]
+  staticcheck:
+    checks:
+      - all
+      - -SA1019 # omit Deprecated check
 severity:
   default-severity: error
   rules:
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 000000000..f003a69da
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,37 @@
+# Contributing to Ollama
+
+Thank you for your interest in contributing to Ollama! Here are a few guidelines to help get you started.
+
+## Set up
+
+See the [development documentation](./docs/development.md) for instructions on how to build and run Ollama locally.
+
+## Pull requests
+
+### Ideal issues
+
+* [Bugs](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Abug): issues where Ollama stops working or where it results in an unexpected error.
+* [Performance](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Aperformance): issues to make Ollama faster at model inference, downloading or uploading.
+* [Security](https://github.com/ollama/ollama/blob/main/SECURITY.md): issues that could lead to a security vulnerability. As mentioned in [SECURITY.md](https://github.com/ollama/ollama/blob/main/SECURITY.md), please do not disclose security vulnerabilities publicly.
+
+### Issues that are harder to review
+
+* New features: new features (e.g. API fields, environment variables) add surface area to Ollama and make it harder to maintain in the long run as they cannot be removed without potentially breaking users in the future.
+* Refactoring: large code improvements are important, but can be harder or take longer to review and merge.
+* Documentation: small updates to fill in or correct missing documentation is helpful, however large documentation additions can be hard to maintain over time.
+
+### Issues that may not be accepted
+
+* Changes that break backwards compatibility in Ollama's API (including the OpenAI-compatible API)
+* Changes that add significant friction to the user experience
+* Changes that create a large future maintenance burden for maintainers and contributors
+
+### Best practices
+
+* Commit messages: please leave both a title and a description in your commit messages. The title should be a short summary of the changes, with a leading word that explains the section of the code being changed (e.g. `api: fix parsing of prompt field`) . In the description, leave a short 2-3 sentences that explain more about the change and its impact.
+* Tests: please add test coverage to changes where possible.
+* Minimize dependencies: avoid adding new dependencies unless absolutely necessary.
+
+## Need help?
+
+If you need help with anything, feel free to reach out to us on our [Discord server](https://discord.gg/ollama).
diff --git a/Dockerfile b/Dockerfile
index 98a3ddfd2..0f43e618d 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,8 +1,10 @@
-ARG GOLANG_VERSION=1.22.1
+ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
-# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
-ARG CUDA_VERSION=11.3.1
-ARG ROCM_VERSION=6.1.1
+ARG CUDA_VERSION_11=11.3.1
+ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+ARG CUDA_VERSION_12=12.4.0
+ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+ARG ROCM_VERSION=6.1.2
 
 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@@ -10,131 +12,243 @@ COPY .git .git
 COPY .gitmodules .gitmodules
 COPY llm llm
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION-devel-centos7 AS cuda-build-amd64
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
 
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-runner-arm64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH=arm64
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-runner-arm64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH=arm64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
 
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH /opt/amdgpu/lib64
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV LIBRARY_PATH=/opt/amdgpu/lib64
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
-RUN mkdir /tmp/scratch && \
-    for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
-        cp ${dep} /tmp/scratch/ || exit 1 ; \
-    done && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
-    mkdir -p /go/src/github.com/ollama/ollama/dist/deps/ && \
-    (cd /tmp/scratch/ && tar czvf /go/src/github.com/ollama/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
-
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
+RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
+    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
 
 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
+ENV GOARCH=amd64
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
 FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
-RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
 
-FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
+FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
+ENV GOARCH=arm64
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
 FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
-RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 
 
-# Intermediate stage used for ./scripts/build_linux.sh
+# Intermediate stages used for ./scripts/build_linux.sh
 FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
-ENV CGO_ENABLED 1
+ENV CGO_ENABLED=1
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
+COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build -trimpath .
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz
 
-# Intermediate stage used for ./scripts/build_linux.sh
 FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
-ENV CGO_ENABLED 1
+ENV CGO_ENABLED=1
 ARG GOLANG_VERSION
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
+COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build -trimpath .
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
 
-# Runtime stages
-FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
-FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
+FROM --platform=linux/amd64 scratch AS dist-amd64
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM --platform=linux/arm64 scratch AS dist-arm64
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM dist-$TARGETARCH as dist
 
-# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
-RUN update-pciids
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
+
+# Optimized container images do not cary nested payloads
+FROM --platform=linux/amd64 static-build-amd64 AS container-build-amd64
+WORKDIR /go/src/github.com/ollama/ollama
+COPY . .
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
+
+FROM --platform=linux/arm64 static-build-arm64 AS container-build-arm64
+WORKDIR /go/src/github.com/ollama/ollama
+COPY . .
+ARG GOFLAGS
+ARG CGO_CFLAGS
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=cpu-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+
+# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
+# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
+# across releases
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
 EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
+ENV OLLAMA_HOST=0.0.0.0
 
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
 
 FROM runtime-$TARGETARCH
 EXPOSE 11434
-ENV OLLAMA_HOST 0.0.0.0
+ENV OLLAMA_HOST=0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
diff --git a/README.md b/README.md
index 62f5cd65c..466f315ad 100644
--- a/README.md
+++ b/README.md
@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3](https://ollama.com/library/llama3):
+To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
 
 ```
-ollama run llama3
+ollama run llama3.1
 ```
 
 ## Model library
@@ -49,10 +49,12 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
-| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
+| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
+| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
+| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
 | Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
+| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`         |
 | Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
 | Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`        |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
@@ -64,7 +66,8 @@ Here are some example models that can be downloaded:
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
-> Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
+> [!NOTE]
+> You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
 
 ## Customize a model
 
@@ -96,16 +99,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.1` model:
 
 ```
-ollama pull llama3
+ollama pull llama3.1
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3
+FROM llama3.1
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -140,7 +143,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3
+ollama pull llama3.1
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -148,13 +151,13 @@ ollama pull llama3
 ### Remove a model
 
 ```
-ollama rm llama3
+ollama rm llama3.1
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3 my-model
+ollama cp llama3.1 my-model
 ```
 
 ### Multiline input
@@ -171,21 +174,21 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 ### Multimodal models
 
 ```
->>> What's in this image? /Users/jmorgan/Desktop/smile.png
+ollama run llava "What's in this image? /Users/jmorgan/Desktop/smile.png"
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```
 
 ### Pass the prompt as an argument
 
 ```
-$ ollama run llama3 "Summarize this file: $(cat README.md)"
+$ ollama run llama3.1 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
 ### Show model information
 
 ```
-ollama show llama3
+ollama show llama3.1
 ```
 
 ### List models on your computer
@@ -194,6 +197,18 @@ ollama show llama3
 ollama list
 ```
 
+### List which models are currently loaded
+
+```
+ollama ps
+```
+
+### Stop a model which is currently running
+
+```
+ollama stop llama3.1
+```
+
 ### Start Ollama
 
 `ollama serve` is used when you want to start ollama without running the desktop application.
@@ -213,7 +228,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3
+./ollama run llama3.1
 ```
 
 ## REST API
@@ -224,7 +239,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -233,7 +248,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -292,7 +307,24 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [AiLama](https://github.com/zeyoyt/ailama) (A Discord User App that allows you to interact with Ollama anywhere in discord )
 - [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
+- [Painting Droid](https://github.com/mateuszmigas/painting-droid) (Painting app with AI integrations)
+- [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
+- [AI Studio](https://github.com/MindWorkAI/AI-Studio)
+- [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
+- [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
+- [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
+- [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
+- [Go-CREW](https://www.jonathanhecl.com/go-crew/) (Powerful Offline RAG in Golang)
+- [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
+- [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
+- [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
+- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
+- [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
+- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
+- [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
+- [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)
 
 ### Terminal
 
@@ -316,6 +348,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [tlm](https://github.com/yusufcanb/tlm)
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 - [gollama](https://github.com/sammcj/gollama)
+- [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
+- [Ollama Mixture of Experts (MOE) in 50 lines of code](https://github.com/rapidarchitect/ollama_moe)
+- [vim-intelligence-bridge](https://github.com/pepo-ec/vim-intelligence-bridge) Simple interaction of "Ollama" with the Vim editor
+
+### Apple Vision Pro
+- [Enchanted](https://github.com/AugustDev/enchanted)
 
 ### Database
 
@@ -325,22 +363,28 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Package managers
 
 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
+- [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 - [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
+- [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
+- [Flox](https://flox.dev/blog/ollama-part-one)
 
 ### Libraries
 
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
+- [crewAI](https://github.com/crewAIInc/crewAI)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
 - [LiteLLM](https://github.com/BerriAI/litellm)
+- [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
-- [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
+- [Ollama4j for Java](https://github.com/ollama4j/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
 - [Ollama for Dart](https://github.com/breitburg/dart-ollama)
@@ -357,11 +401,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
 - [LlamaScript](https://github.com/Project-Llama/llamascript)
+- [Gollm](https://docs.gollm.co/examples/ollama-example)
+- [Ollamaclient for Golang](https://github.com/xyproto/ollamaclient)
+- [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
+- [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
+- [Agents-Flex for Java](https://github.com/agents-flex/agents-flex) with [example](https://github.com/agents-flex/agents-flex/tree/main/agents-flex-llm/agents-flex-llm-ollama/src/test/java/com/agentsflex/llm/ollama)
 
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
+- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 
 ### Extensions & Plugins
 
@@ -384,13 +434,18 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Ollama Copilot](https://github.com/bernardo-bruning/ollama-copilot) (Proxy that allows you to use ollama as a copilot like Github copilot)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
-- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
+- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and Hugging Face)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
+- [Plasmoid Ollama Control](https://github.com/imoize/plasmoid-ollamacontrol) (KDE Plasma extension that allows you to quickly manage/control Ollama model)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
 - [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 - [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
+- [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
+- [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
+- [QodeAssist](https://github.com/Palm1r/QodeAssist) (AI-powered coding assistant plugin for Qt Creator)
+- [Obsidian Quiz Generator plugin](https://github.com/ECuiDev/obsidian-quiz-generator)
 
 ### Supported backends
 
diff --git a/SECURITY.md b/SECURITY.md
new file mode 100644
index 000000000..d38bb7c4e
--- /dev/null
+++ b/SECURITY.md
@@ -0,0 +1,25 @@
+# Security
+
+The Ollama maintainer team takes security seriously and will actively work to resolve security issues.
+
+## Reporting a vulnerability
+
+If you discover a security vulnerability, please do not open a public issue. Instead, please report it by emailing hello@ollama.com. We ask that you give us sufficient time to investigate and address the vulnerability before disclosing it publicly.
+
+Please include the following details in your report:
+- A description of the vulnerability
+- Steps to reproduce the issue
+- Your assessment of the potential impact
+- Any possible mitigations
+
+## Security best practices
+
+While the maintainer team does their best to secure Ollama, users are encouraged to implement their own security best practices, such as:
+
+- Regularly updating to the latest version of Ollama
+- Securing access to hosted instances of Ollama
+- Monitoring systems for unusual activity
+
+## Contact
+
+For any other questions or concerns related to security, please contact us at hello@ollama.com
diff --git a/api/client.go b/api/client.go
index fccbc9ad7..2528fb21f 100644
--- a/api/client.go
+++ b/api/client.go
@@ -18,9 +18,9 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
-	"net"
 	"net/http"
 	"net/url"
 	"runtime"
@@ -63,13 +63,8 @@ func checkError(resp *http.Response, body []byte) error {
 // If the variable is not specified, a default ollama host and port will be
 // used.
 func ClientFromEnvironment() (*Client, error) {
-	ollamaHost := envconfig.Host
-
 	return &Client{
-		base: &url.URL{
-			Scheme: ollamaHost.Scheme,
-			Host:   net.JoinHostPort(ollamaHost.Host, ollamaHost.Port),
-		},
+		base: envconfig.Host(),
 		http: http.DefaultClient,
 	}, nil
 }
@@ -178,7 +173,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 
 		if errorResponse.Error != "" {
-			return fmt.Errorf(errorResponse.Error)
+			return errors.New(errorResponse.Error)
 		}
 
 		if response.StatusCode >= http.StatusBadRequest {
@@ -303,7 +298,7 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }
 
-// List running models.
+// ListRunning lists running models.
 func (c *Client) ListRunning(ctx context.Context) (*ProcessResponse, error) {
 	var lr ProcessResponse
 	if err := c.do(ctx, http.MethodGet, "/api/ps", nil, &lr); err != nil {
@@ -338,7 +333,7 @@ func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, err
 	return &resp, nil
 }
 
-// Hearbeat checks if the server has started and is responsive; if yes, it
+// Heartbeat checks if the server has started and is responsive; if yes, it
 // returns nil, otherwise an error.
 func (c *Client) Heartbeat(ctx context.Context) error {
 	if err := c.do(ctx, http.MethodHead, "/", nil, nil); err != nil {
@@ -347,7 +342,16 @@ func (c *Client) Heartbeat(ctx context.Context) error {
 	return nil
 }
 
-// Embeddings generates embeddings from a model.
+// Embed generates embeddings from a model.
+func (c *Client) Embed(ctx context.Context, req *EmbedRequest) (*EmbedResponse, error) {
+	var resp EmbedResponse
+	if err := c.do(ctx, http.MethodPost, "/api/embed", req, &resp); err != nil {
+		return nil, err
+	}
+	return &resp, nil
+}
+
+// Embeddings generates an embedding from a model.
 func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*EmbeddingResponse, error) {
 	var resp EmbeddingResponse
 	if err := c.do(ctx, http.MethodPost, "/api/embeddings", req, &resp); err != nil {
diff --git a/api/client_test.go b/api/client_test.go
index fe9fd74f7..23fe9334b 100644
--- a/api/client_test.go
+++ b/api/client_test.go
@@ -2,8 +2,6 @@ package api
 
 import (
 	"testing"
-
-	"github.com/ollama/ollama/envconfig"
 )
 
 func TestClientFromEnvironment(t *testing.T) {
@@ -33,7 +31,6 @@ func TestClientFromEnvironment(t *testing.T) {
 	for k, v := range testCases {
 		t.Run(k, func(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", v.value)
-			envconfig.LoadConfig()
 
 			client, err := ClientFromEnvironment()
 			if err != v.err {
diff --git a/api/types.go b/api/types.go
index 95ed5d37e..df7bab210 100644
--- a/api/types.go
+++ b/api/types.go
@@ -47,6 +47,9 @@ type GenerateRequest struct {
 	// Prompt is the textual prompt to send to the model.
 	Prompt string `json:"prompt"`
 
+	// Suffix is the text that comes after the inserted text.
+	Suffix string `json:"suffix"`
+
 	// System overrides the model's default system message/prompt.
 	System string `json:"system"`
 
@@ -97,17 +100,85 @@ type ChatRequest struct {
 	// followin the request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
+	// Tools is an optional list of tools the model has access to.
+	Tools `json:"tools,omitempty"`
+
 	// Options lists model-specific options.
 	Options map[string]interface{} `json:"options"`
 }
 
+type Tools []Tool
+
+func (t Tools) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
+func (t Tool) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
 // Message is a single message in a chat sequence. The message contains the
 // role ("system", "user", or "assistant"), the content and an optional list
 // of images.
 type Message struct {
-	Role    string      `json:"role"`
-	Content string      `json:"content"`
-	Images  []ImageData `json:"images,omitempty"`
+	Role      string      `json:"role"`
+	Content   string      `json:"content"`
+	Images    []ImageData `json:"images,omitempty"`
+	ToolCalls []ToolCall  `json:"tool_calls,omitempty"`
+}
+
+func (m *Message) UnmarshalJSON(b []byte) error {
+	type Alias Message
+	var a Alias
+	if err := json.Unmarshal(b, &a); err != nil {
+		return err
+	}
+
+	*m = Message(a)
+	m.Role = strings.ToLower(m.Role)
+	return nil
+}
+
+type ToolCall struct {
+	Function ToolCallFunction `json:"function"`
+}
+
+type ToolCallFunction struct {
+	Name      string                    `json:"name"`
+	Arguments ToolCallFunctionArguments `json:"arguments"`
+}
+
+type ToolCallFunctionArguments map[string]any
+
+func (t *ToolCallFunctionArguments) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
+}
+
+type Tool struct {
+	Type     string       `json:"type"`
+	Function ToolFunction `json:"function"`
+}
+
+type ToolFunction struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Parameters  struct {
+		Type       string   `json:"type"`
+		Required   []string `json:"required"`
+		Properties map[string]struct {
+			Type        string   `json:"type"`
+			Description string   `json:"description"`
+			Enum        []string `json:"enum,omitempty"`
+		} `json:"properties"`
+	} `json:"parameters"`
+}
+
+func (t *ToolFunction) String() string {
+	bts, _ := json.Marshal(t)
+	return string(bts)
 }
 
 // ChatResponse is the response returned by [Client.Chat]. Its fields are
@@ -143,6 +214,7 @@ type Options struct {
 	NumPredict       int      `json:"num_predict,omitempty"`
 	TopK             int      `json:"top_k,omitempty"`
 	TopP             float32  `json:"top_p,omitempty"`
+	MinP             float32  `json:"min_p,omitempty"`
 	TFSZ             float32  `json:"tfs_z,omitempty"`
 	TypicalP         float32  `json:"typical_p,omitempty"`
 	RepeatLastN      int      `json:"repeat_last_n,omitempty"`
@@ -159,49 +231,45 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA   bool     `json:"numa,omitempty"`
-	NumCtx    int      `json:"num_ctx,omitempty"`
-	NumBatch  int      `json:"num_batch,omitempty"`
-	NumGPU    int      `json:"num_gpu,omitempty"`
-	MainGPU   int      `json:"main_gpu,omitempty"`
-	LowVRAM   bool     `json:"low_vram,omitempty"`
-	F16KV     bool     `json:"f16_kv,omitempty"`
-	LogitsAll bool     `json:"logits_all,omitempty"`
-	VocabOnly bool     `json:"vocab_only,omitempty"`
-	UseMMap   TriState `json:"use_mmap,omitempty"`
-	UseMLock  bool     `json:"use_mlock,omitempty"`
-	NumThread int      `json:"num_thread,omitempty"`
+	NumCtx    int   `json:"num_ctx,omitempty"`
+	NumBatch  int   `json:"num_batch,omitempty"`
+	NumGPU    int   `json:"num_gpu,omitempty"`
+	MainGPU   int   `json:"main_gpu,omitempty"`
+	LowVRAM   bool  `json:"low_vram,omitempty"`
+	F16KV     bool  `json:"f16_kv,omitempty"`
+	LogitsAll bool  `json:"logits_all,omitempty"`
+	VocabOnly bool  `json:"vocab_only,omitempty"`
+	UseMMap   *bool `json:"use_mmap,omitempty"`
+	UseMLock  bool  `json:"use_mlock,omitempty"`
+	NumThread int   `json:"num_thread,omitempty"`
 }
 
-type TriState int
+// EmbedRequest is the request passed to [Client.Embed].
+type EmbedRequest struct {
+	// Model is the model name.
+	Model string `json:"model"`
 
-const (
-	TriStateUndefined TriState = -1
-	TriStateFalse     TriState = 0
-	TriStateTrue      TriState = 1
-)
+	// Input is the input to embed.
+	Input any `json:"input"`
 
-func (b *TriState) UnmarshalJSON(data []byte) error {
-	var v bool
-	if err := json.Unmarshal(data, &v); err != nil {
-		return err
-	}
-	if v {
-		*b = TriStateTrue
-	}
-	*b = TriStateFalse
-	return nil
+	// KeepAlive controls how long the model will stay loaded in memory following
+	// this request.
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
+
+	Truncate *bool `json:"truncate,omitempty"`
+
+	// Options lists model-specific options.
+	Options map[string]interface{} `json:"options"`
 }
 
-func (b *TriState) MarshalJSON() ([]byte, error) {
-	if *b == TriStateUndefined {
-		return nil, nil
-	}
-	var v bool
-	if *b == TriStateTrue {
-		v = true
-	}
-	return json.Marshal(v)
+// EmbedResponse is the response from [Client.Embed].
+type EmbedResponse struct {
+	Model      string      `json:"model"`
+	Embeddings [][]float32 `json:"embeddings"`
+
+	TotalDuration   time.Duration `json:"total_duration,omitempty"`
+	LoadDuration    time.Duration `json:"load_duration,omitempty"`
+	PromptEvalCount int           `json:"prompt_eval_count,omitempty"`
 }
 
 // EmbeddingRequest is the request passed to [Client.Embeddings].
@@ -228,15 +296,17 @@ type EmbeddingResponse struct {
 // CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
 	Model     string `json:"model"`
-	Path      string `json:"path"`
 	Modelfile string `json:"modelfile"`
 	Stream    *bool  `json:"stream,omitempty"`
 	Quantize  string `json:"quantize,omitempty"`
 
-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 
-	// Quantization is deprecated, see Quantize
+	// Deprecated: set the file content with Modelfile instead
+	Path string `json:"path"`
+
+	// Deprecated: use Quantize instead
 	Quantization string `json:"quantization,omitempty"`
 }
 
@@ -244,20 +314,22 @@ type CreateRequest struct {
 type DeleteRequest struct {
 	Model string `json:"model"`
 
-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }
 
 // ShowRequest is the request passed to [Client.Show].
 type ShowRequest struct {
-	Model    string `json:"model"`
-	System   string `json:"system"`
+	Model  string `json:"model"`
+	System string `json:"system"`
+
+	// Template is deprecated
 	Template string `json:"template"`
 	Verbose  bool   `json:"verbose"`
 
 	Options map[string]interface{} `json:"options"`
 
-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }
 
@@ -289,7 +361,7 @@ type PullRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
 
-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }
 
@@ -310,7 +382,7 @@ type PushRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
 
-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }
 
@@ -345,6 +417,13 @@ type ProcessModelResponse struct {
 	SizeVRAM  int64        `json:"size_vram"`
 }
 
+type RetrieveModelResponse struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
 type TokenResponse struct {
 	Token string `json:"token"`
 }
@@ -427,7 +506,7 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 	for key, val := range m {
 		opt, ok := jsonOpts[key]
 		if !ok {
-			slog.Warn("invalid option provided", "option", opt.Name)
+			slog.Warn("invalid option provided", "option", key)
 			continue
 		}
 
@@ -437,19 +516,6 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 				continue
 			}
 
-			if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-				val, ok := val.(bool)
-				if !ok {
-					return fmt.Errorf("option %q must be of type boolean", key)
-				}
-				if val {
-					field.SetInt(int64(TriStateTrue))
-				} else {
-					field.SetInt(int64(TriStateFalse))
-				}
-				continue
-			}
-
 			switch field.Kind() {
 			case reflect.Int:
 				switch t := val.(type) {
@@ -496,6 +562,17 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					slice[i] = str
 				}
 				field.Set(reflect.ValueOf(slice))
+			case reflect.Pointer:
+				var b bool
+				if field.Type() == reflect.TypeOf(&b) {
+					val, ok := val.(bool)
+					if !ok {
+						return fmt.Errorf("option %q must be of type boolean", key)
+					}
+					field.Set(reflect.ValueOf(&val))
+				} else {
+					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
+				}
 			default:
 				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
@@ -538,8 +615,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   TriStateUndefined,
-			UseNUMA:   false,
+			UseMMap:   nil,
 		},
 	}
 }
@@ -608,19 +684,6 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		} else {
 			field := valueOpts.FieldByName(opt.Name)
 			if field.IsValid() && field.CanSet() {
-				if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-					boolVal, err := strconv.ParseBool(vals[0])
-					if err != nil {
-						return nil, fmt.Errorf("invalid bool value %s", vals)
-					}
-					if boolVal {
-						out[key] = TriStateTrue
-					} else {
-						out[key] = TriStateFalse
-					}
-					continue
-				}
-
 				switch field.Kind() {
 				case reflect.Float32:
 					floatVal, err := strconv.ParseFloat(vals[0], 32)
@@ -648,6 +711,17 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 				case reflect.Slice:
 					// TODO: only string slices are supported right now
 					out[key] = vals
+				case reflect.Pointer:
+					var b bool
+					if field.Type() == reflect.TypeOf(&b) {
+						boolVal, err := strconv.ParseBool(vals[0])
+						if err != nil {
+							return nil, fmt.Errorf("invalid bool value %s", vals)
+						}
+						out[key] = &boolVal
+					} else {
+						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
+					}
 				default:
 					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 				}
diff --git a/api/types_test.go b/api/types_test.go
index 8b6c60c62..a9de5a9a9 100644
--- a/api/types_test.go
+++ b/api/types_test.go
@@ -2,7 +2,7 @@ package api
 
 import (
 	"encoding/json"
-	"fmt"
+	"errors"
 	"math"
 	"testing"
 	"time"
@@ -108,25 +108,27 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 }
 
 func TestUseMmapParsingFromJSON(t *testing.T) {
+	tr := true
+	fa := false
 	tests := []struct {
 		name string
 		req  string
-		exp  TriState
+		exp  *bool
 	}{
 		{
 			name: "Undefined",
 			req:  `{ }`,
-			exp:  TriStateUndefined,
+			exp:  nil,
 		},
 		{
 			name: "True",
 			req:  `{ "use_mmap": true }`,
-			exp:  TriStateTrue,
+			exp:  &tr,
 		},
 		{
 			name: "False",
 			req:  `{ "use_mmap": false }`,
-			exp:  TriStateFalse,
+			exp:  &fa,
 		},
 	}
 
@@ -144,63 +146,88 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 }
 
 func TestUseMmapFormatParams(t *testing.T) {
+	tr := true
+	fa := false
 	tests := []struct {
 		name string
 		req  map[string][]string
-		exp  TriState
+		exp  *bool
 		err  error
 	}{
 		{
 			name: "True",
 			req: map[string][]string{
-				"use_mmap": []string{"true"},
+				"use_mmap": {"true"},
 			},
-			exp: TriStateTrue,
+			exp: &tr,
 			err: nil,
 		},
 		{
 			name: "False",
 			req: map[string][]string{
-				"use_mmap": []string{"false"},
+				"use_mmap": {"false"},
 			},
-			exp: TriStateFalse,
+			exp: &fa,
 			err: nil,
 		},
 		{
 			name: "Numeric True",
 			req: map[string][]string{
-				"use_mmap": []string{"1"},
+				"use_mmap": {"1"},
 			},
-			exp: TriStateTrue,
+			exp: &tr,
 			err: nil,
 		},
 		{
 			name: "Numeric False",
 			req: map[string][]string{
-				"use_mmap": []string{"0"},
+				"use_mmap": {"0"},
 			},
-			exp: TriStateFalse,
+			exp: &fa,
 			err: nil,
 		},
 		{
 			name: "invalid string",
 			req: map[string][]string{
-				"use_mmap": []string{"foo"},
+				"use_mmap": {"foo"},
 			},
-			exp: TriStateUndefined,
-			err: fmt.Errorf("invalid bool value [foo]"),
+			exp: nil,
+			err: errors.New("invalid bool value [foo]"),
 		},
 	}
 
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
 			resp, err := FormatParams(test.req)
-			require.Equal(t, err, test.err)
+			require.Equal(t, test.err, err)
 			respVal, ok := resp["use_mmap"]
-			if test.exp != TriStateUndefined {
+			if test.exp != nil {
 				assert.True(t, ok, "resp: %v", resp)
-				assert.Equal(t, test.exp, respVal)
+				assert.Equal(t, *test.exp, *respVal.(*bool))
 			}
 		})
 	}
 }
+
+func TestMessage_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{`{"role": "USER", "content": "Hello!"}`, "user"},
+		{`{"role": "System", "content": "Initialization complete."}`, "system"},
+		{`{"role": "assistant", "content": "How can I help you?"}`, "assistant"},
+		{`{"role": "TOOl", "content": "Access granted."}`, "tool"},
+	}
+
+	for _, test := range tests {
+		var msg Message
+		if err := json.Unmarshal([]byte(test.input), &msg); err != nil {
+			t.Errorf("Unexpected error: %v", err)
+		}
+
+		if msg.Role != test.expected {
+			t.Errorf("role not lowercased: got %v, expected %v", msg.Role, test.expected)
+		}
+	}
+}
diff --git a/app/lifecycle/getstarted_nonwindows.go b/app/lifecycle/getstarted_nonwindows.go
index c36d14c09..2af87ab92 100644
--- a/app/lifecycle/getstarted_nonwindows.go
+++ b/app/lifecycle/getstarted_nonwindows.go
@@ -2,8 +2,8 @@
 
 package lifecycle
 
-import "fmt"
+import "errors"
 
 func GetStarted() error {
-	return fmt.Errorf("GetStarted not implemented")
+	return errors.New("not implemented")
 }
diff --git a/app/lifecycle/getstarted_windows.go b/app/lifecycle/getstarted_windows.go
index 092c3c17f..f39dc31c0 100644
--- a/app/lifecycle/getstarted_windows.go
+++ b/app/lifecycle/getstarted_windows.go
@@ -34,7 +34,6 @@ func GetStarted() error {
 		Sys:   &syscall.SysProcAttr{CreationFlags: CREATE_NEW_CONSOLE, HideWindow: false},
 	}
 	proc, err := os.StartProcess(args[0], args, attrs)
-
 	if err != nil {
 		return fmt.Errorf("unable to start getting started shell %w", err)
 	}
diff --git a/app/lifecycle/logging.go b/app/lifecycle/logging.go
index a8f1f7cdf..9985fc3f8 100644
--- a/app/lifecycle/logging.go
+++ b/app/lifecycle/logging.go
@@ -14,7 +14,7 @@ import (
 func InitLogging() {
 	level := slog.LevelInfo
 
-	if envconfig.Debug {
+	if envconfig.Debug() {
 		level = slog.LevelDebug
 	}
 
@@ -27,7 +27,7 @@ func InitLogging() {
 		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
 	} else {
 		rotateLogs(AppLogFile)
-		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755)
 		if err != nil {
 			slog.Error(fmt.Sprintf("failed to create server log %v", err))
 			return
diff --git a/app/lifecycle/logging_nonwindows.go b/app/lifecycle/logging_nonwindows.go
index 50b3a638c..205e47d77 100644
--- a/app/lifecycle/logging_nonwindows.go
+++ b/app/lifecycle/logging_nonwindows.go
@@ -5,5 +5,5 @@ package lifecycle
 import "log/slog"
 
 func ShowLogs() {
-	slog.Warn("ShowLogs not yet implemented")
+	slog.Warn("not implemented")
 }
diff --git a/app/lifecycle/logging_test.go b/app/lifecycle/logging_test.go
index a2157ca2c..8d5cdf6e7 100644
--- a/app/lifecycle/logging_test.go
+++ b/app/lifecycle/logging_test.go
@@ -17,7 +17,7 @@ func TestRotateLogs(t *testing.T) {
 	// No log exists
 	rotateLogs(logFile)
 
-	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0644))
+	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0o644))
 	assert.FileExists(t, logFile)
 	// First rotation
 	rotateLogs(logFile)
@@ -32,7 +32,7 @@ func TestRotateLogs(t *testing.T) {
 	assert.NoFileExists(t, logFile)
 
 	for i := 2; i <= LogRotationCount+1; i++ {
-		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0644))
+		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0o644))
 		assert.FileExists(t, logFile)
 		rotateLogs(logFile)
 		assert.NoFileExists(t, logFile)
diff --git a/app/lifecycle/server.go b/app/lifecycle/server.go
index c178a1abf..37957399c 100644
--- a/app/lifecycle/server.go
+++ b/app/lifecycle/server.go
@@ -55,7 +55,7 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 	}
 
 	rotateLogs(ServerLogFile)
-	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create server log: %w", err)
 	}
diff --git a/app/lifecycle/updater.go b/app/lifecycle/updater.go
index b6d953309..4d3c7d8dc 100644
--- a/app/lifecycle/updater.go
+++ b/app/lifecycle/updater.go
@@ -15,6 +15,7 @@ import (
 	"path"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"time"
 
@@ -46,7 +47,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	query.Add("os", runtime.GOOS)
 	query.Add("arch", runtime.GOARCH)
 	query.Add("version", version.Version)
-	query.Add("ts", fmt.Sprintf("%d", time.Now().Unix()))
+	query.Add("ts", strconv.FormatInt(time.Now().Unix(), 10))
 
 	nonce, err := auth.NewNonce(rand.Reader, 16)
 	if err != nil {
diff --git a/app/lifecycle/updater_nonwindows.go b/app/lifecycle/updater_nonwindows.go
index 0f213b34f..1d2dda801 100644
--- a/app/lifecycle/updater_nonwindows.go
+++ b/app/lifecycle/updater_nonwindows.go
@@ -4,9 +4,9 @@ package lifecycle
 
 import (
 	"context"
-	"fmt"
+	"errors"
 )
 
 func DoUpgrade(cancel context.CancelFunc, done chan int) error {
-	return fmt.Errorf("DoUpgrade not yet implemented")
+	return errors.New("not implemented")
 }
diff --git a/app/lifecycle/updater_windows.go b/app/lifecycle/updater_windows.go
index 4053671a5..1d3830d4e 100644
--- a/app/lifecycle/updater_windows.go
+++ b/app/lifecycle/updater_windows.go
@@ -2,6 +2,7 @@ package lifecycle
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"log/slog"
 	"os"
@@ -15,7 +16,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 		return fmt.Errorf("failed to lookup downloads: %s", err)
 	}
 	if len(files) == 0 {
-		return fmt.Errorf("no update downloads found")
+		return errors.New("no update downloads found")
 	} else if len(files) > 1 {
 		// Shouldn't happen
 		slog.Warn(fmt.Sprintf("multiple downloads found, using first one %v", files))
@@ -64,7 +65,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 		}
 	} else {
 		// TODO - some details about why it didn't start, or is this a pedantic error case?
-		return fmt.Errorf("installer process did not start")
+		return errors.New("installer process did not start")
 	}
 
 	// TODO should we linger for a moment and check to make sure it's actually running by checking the pid?
diff --git a/app/ollama.iss b/app/ollama.iss
index e6502abd3..63b5bdb0f 100644
--- a/app/ollama.iss
+++ b/app/ollama.iss
@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64 arm64
-ArchitecturesInstallIn64BitMode=x64 arm64
+ArchitecturesAllowed=x64compatible arm64
+ArchitecturesInstallIn64BitMode=x64compatible arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@@ -48,6 +48,7 @@ OutputDir=..\dist\
 SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
+RestartIfNeededByRun=no
 
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
@@ -86,21 +87,21 @@ Name: "english"; MessagesFile: "compiler:Default.isl"
 DialogFontSize=12
 
 [Files]
-Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
-Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
-Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-#if DirExists("..\dist\windows-amd64\cuda")
-  Source: "..\dist\windows-amd64\cuda\*"; DestDir: "{app}\cuda\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\oneapi")
-  Source: "..\dist\windows-amd64\oneapi\*"; DestDir: "{app}\oneapi\"; Flags: ignoreversion recursesubdirs
-#endif
-#if DirExists("..\dist\windows-amd64\rocm")
-  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
+#if DirExists("..\dist\windows-amd64")
+Source: "..\dist\windows-amd64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
 #endif
 
+#if DirExists("..\dist\windows-arm64")
+Source: "..\dist\windows-arm64\vc_redist.arm64.exe"; DestDir: "{tmp}"; Check: IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
+Source: "..\dist\windows-arm64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\ollama.exe"; DestDir: "{app}"; Check: IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
+
+Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
+Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@@ -108,6 +109,9 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
 Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
 
 [Run]
+#if DirExists("..\dist\windows-arm64")
+Filename: "{tmp}\vc_redist.arm64.exe"; Parameters: "/install /passive /norestart"; Check: IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
+#endif
 Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
 
 [UninstallRun]
@@ -127,6 +131,10 @@ Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\models"
 Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\history"
 ; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved
 
+[InstallDelete]
+Type: filesandordirs; Name: "{%TEMP}\ollama*"
+Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
+
 [Messages]
 WizardReady=Ollama Windows Preview
 ReadyLabel1=%nLet's get you up and running with your own large language models.
@@ -134,7 +142,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 
 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.1
 ;ClickFinish=%n
 
 [Registry]
@@ -159,3 +167,39 @@ begin
   { Pos() returns 0 if not found }
   Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
 end;
+
+{ --- VC Runtime libraries discovery code - Only install vc_redist if it isn't already installed ----- }
+const VCRTL_MIN_V1 = 14;
+const VCRTL_MIN_V2 = 40;
+const VCRTL_MIN_V3 = 33807;
+const VCRTL_MIN_V4 = 0;
+
+ // check if the minimum required vc redist is installed (by looking the registry)
+function vc_redist_needed (): Boolean;
+var
+  sRegKey: string;
+  v1: Cardinal;
+  v2: Cardinal;
+  v3: Cardinal;
+  v4: Cardinal;
+begin
+  sRegKey := 'SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\arm64';
+  if (RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Major', v1)  and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Minor', v2) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Bld', v3) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'RBld', v4)) then
+  begin
+    Log ('VC Redist version: ' + IntToStr (v1) +
+        '.' + IntToStr (v2) + '.' + IntToStr (v3) +
+        '.' + IntToStr (v4));
+    { Version info was found. Return true if later or equal to our
+       minimal required version RTL_MIN_Vx }
+    Result := not (
+        (v1 > VCRTL_MIN_V1) or ((v1 = VCRTL_MIN_V1) and
+         ((v2 > VCRTL_MIN_V2) or ((v2 = VCRTL_MIN_V2) and
+          ((v3 > VCRTL_MIN_V3) or ((v3 = VCRTL_MIN_V3) and
+           (v4 >= VCRTL_MIN_V4)))))));
+  end
+  else
+    Result := TRUE;
+end;
diff --git a/app/ollama_welcome.ps1 b/app/ollama_welcome.ps1
index 9af37a461..46777a3a6 100644
--- a/app/ollama_welcome.ps1
+++ b/app/ollama_welcome.ps1
@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3"
+write-host "`tollama run llama3.1"
 write-host ""
\ No newline at end of file
diff --git a/app/tray/tray_nonwindows.go b/app/tray/tray_nonwindows.go
index ae5572b23..a03d233ea 100644
--- a/app/tray/tray_nonwindows.go
+++ b/app/tray/tray_nonwindows.go
@@ -3,11 +3,11 @@
 package tray
 
 import (
-	"fmt"
+	"errors"
 
 	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
-	return nil, fmt.Errorf("NOT IMPLEMENTED YET")
+	return nil, errors.New("not implemented")
 }
diff --git a/app/tray/wintray/eventloop.go b/app/tray/wintray/eventloop.go
index 0f9448947..157828a36 100644
--- a/app/tray/wintray/eventloop.go
+++ b/app/tray/wintray/eventloop.go
@@ -11,9 +11,7 @@ import (
 	"golang.org/x/sys/windows"
 )
 
-var (
-	quitOnce sync.Once
-)
+var quitOnce sync.Once
 
 func (t *winTray) Run() {
 	nativeLoop()
diff --git a/app/tray/wintray/menus.go b/app/tray/wintray/menus.go
index 9cb3b8933..596244442 100644
--- a/app/tray/wintray/menus.go
+++ b/app/tray/wintray/menus.go
@@ -11,12 +11,12 @@ import (
 )
 
 const (
-	updatAvailableMenuID = 1
-	updateMenuID         = updatAvailableMenuID + 1
-	separatorMenuID      = updateMenuID + 1
-	diagLogsMenuID       = separatorMenuID + 1
-	diagSeparatorMenuID  = diagLogsMenuID + 1
-	quitMenuID           = diagSeparatorMenuID + 1
+	updateAvailableMenuID = 1
+	updateMenuID          = updateAvailableMenuID + 1
+	separatorMenuID       = updateMenuID + 1
+	diagLogsMenuID        = separatorMenuID + 1
+	diagSeparatorMenuID   = diagLogsMenuID + 1
+	quitMenuID            = diagSeparatorMenuID + 1
 )
 
 func (t *winTray) initMenus() error {
@@ -35,7 +35,7 @@ func (t *winTray) initMenus() error {
 func (t *winTray) UpdateAvailable(ver string) error {
 	if !t.updateNotified {
 		slog.Debug("updating menu and sending notification for new update")
-		if err := t.addOrUpdateMenuItem(updatAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
+		if err := t.addOrUpdateMenuItem(updateAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
 			return fmt.Errorf("unable to create menu entries %w", err)
 		}
 		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenutTitle, false); err != nil {
diff --git a/app/tray/wintray/tray.go b/app/tray/wintray/tray.go
index 027ec5a50..6f8278939 100644
--- a/app/tray/wintray/tray.go
+++ b/app/tray/wintray/tray.go
@@ -11,10 +11,12 @@ import (
 	"path/filepath"
 	"sort"
 	"sync"
+	"syscall"
 	"unsafe"
 
-	"github.com/ollama/ollama/app/tray/commontray"
 	"golang.org/x/sys/windows"
+
+	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 // Helpful sources: https://github.com/golang/exp/blob/master/shiny/driver/internal/win32
@@ -414,7 +416,7 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) {
 	iconFilePath := filepath.Join(os.TempDir(), "ollama_temp_icon_"+dataHash)
 
 	if _, err := os.Stat(iconFilePath); os.IsNotExist(err) {
-		if err := os.WriteFile(iconFilePath, iconBytes, 0644); err != nil {
+		if err := os.WriteFile(iconFilePath, iconBytes, 0o644); err != nil {
 			return "", err
 		}
 	}
@@ -432,7 +434,12 @@ func (t *winTray) setIcon(src string) error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid.Icon = h
-	t.nid.Flags |= NIF_ICON
+	t.nid.Flags |= NIF_ICON | NIF_TIP
+	if toolTipUTF16, err := syscall.UTF16FromString(commontray.ToolTip); err == nil {
+		copy(t.nid.Tip[:], toolTipUTF16)
+	} else {
+		return err
+	}
 	t.nid.Size = uint32(unsafe.Sizeof(*t.nid))
 
 	return t.nid.modify()
diff --git a/app/tray/wintray/w32api.go b/app/tray/wintray/w32api.go
index a1e0381de..7c7c0ac8a 100644
--- a/app/tray/wintray/w32api.go
+++ b/app/tray/wintray/w32api.go
@@ -61,6 +61,7 @@ const (
 	MIIM_SUBMENU        = 0x00000004
 	MIM_APPLYTOSUBMENUS = 0x80000000
 	NIF_ICON            = 0x00000002
+	NIF_TIP             = 0x00000004
 	NIF_INFO            = 0x00000010
 	NIF_MESSAGE         = 0x00000001
 	SW_HIDE             = 0
diff --git a/auth/auth.go b/auth/auth.go
index 026b2a2c7..e1d854124 100644
--- a/auth/auth.go
+++ b/auth/auth.go
@@ -5,6 +5,7 @@ import (
 	"context"
 	"crypto/rand"
 	"encoding/base64"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -78,7 +79,7 @@ func Sign(ctx context.Context, bts []byte) (string, error) {
 	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
 	parts := bytes.Split(publicKey, []byte(" "))
 	if len(parts) < 2 {
-		return "", fmt.Errorf("malformed public key")
+		return "", errors.New("malformed public key")
 	}
 
 	signedData, err := privateKey.Sign(rand.Reader, bts)
diff --git a/build/darwin/amd64/placeholder b/build/darwin/amd64/placeholder
new file mode 100644
index 000000000..87dc27381
--- /dev/null
+++ b/build/darwin/amd64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
diff --git a/build/darwin/arm64/placeholder b/build/darwin/arm64/placeholder
new file mode 100644
index 000000000..87dc27381
--- /dev/null
+++ b/build/darwin/arm64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
diff --git a/build/embed_darwin_amd64.go b/build/embed_darwin_amd64.go
new file mode 100644
index 000000000..af1458ea9
--- /dev/null
+++ b/build/embed_darwin_amd64.go
@@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/amd64/*
+var EmbedFS embed.FS
diff --git a/build/embed_darwin_arm64.go b/build/embed_darwin_arm64.go
new file mode 100644
index 000000000..d885365d0
--- /dev/null
+++ b/build/embed_darwin_arm64.go
@@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/arm64/*
+var EmbedFS embed.FS
diff --git a/build/embed_linux.go b/build/embed_linux.go
new file mode 100644
index 000000000..4cf7be4c3
--- /dev/null
+++ b/build/embed_linux.go
@@ -0,0 +1,6 @@
+package build
+
+import "embed"
+
+//go:embed linux/*
+var EmbedFS embed.FS
diff --git a/build/embed_unused.go b/build/embed_unused.go
new file mode 100644
index 000000000..00fbe02e8
--- /dev/null
+++ b/build/embed_unused.go
@@ -0,0 +1,8 @@
+//go:build !linux && !darwin
+
+package build
+
+import "embed"
+
+// unused on windows
+var EmbedFS embed.FS
diff --git a/build/linux/amd64/placeholder b/build/linux/amd64/placeholder
new file mode 100644
index 000000000..87dc27381
--- /dev/null
+++ b/build/linux/amd64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
diff --git a/build/linux/arm64/placeholder b/build/linux/arm64/placeholder
new file mode 100644
index 000000000..87dc27381
--- /dev/null
+++ b/build/linux/arm64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
diff --git a/cmd/cmd.go b/cmd/cmd.go
index c898c7db6..3bb8b06ec 100644
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -2,6 +2,7 @@ package cmd
 
 import (
 	"archive/zip"
+	"bufio"
 	"bytes"
 	"context"
 	"crypto/ed25519"
@@ -21,7 +22,9 @@ import (
 	"regexp"
 	"runtime"
 	"slices"
+	"strconv"
 	"strings"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -78,6 +81,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	status := "transferring model data"
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)
+	defer p.Stop()
 
 	for i := range modelfile.Commands {
 		switch modelfile.Commands[i].Name {
@@ -112,7 +116,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				path = tempfile
 			}
 
-			digest, err := createBlob(cmd, client, path)
+			digest, err := createBlob(cmd, client, path, spinner)
 			if err != nil {
 				return err
 			}
@@ -202,6 +206,12 @@ func tempZipFiles(path string) (string, error) {
 		// safetensors files might be unresolved git lfs references; skip if they are
 		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
 		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapters.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapters.safetensors
+		files = append(files, st...)
+	} else if st, _ := glob(filepath.Join(path, "adapter_model.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// covers adapter_model.safetensors
+		files = append(files, st...)
 	} else if pt, _ := glob(filepath.Join(path, "pytorch_model*.bin"), "application/zip"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
@@ -221,6 +231,14 @@ func tempZipFiles(path string) (string, error) {
 	}
 	files = append(files, js...)
 
+	// bert models require a nested config.json
+	// TODO(mxyng): merge this with the glob above
+	js, err = glob(filepath.Join(path, "**/*.json"), "text/plain")
+	if err != nil {
+		return "", err
+	}
+	files = append(files, js...)
+
 	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
 		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
 		// tokenizer.model might be a unresolved git lfs reference; error if it is
@@ -250,6 +268,11 @@ func tempZipFiles(path string) (string, error) {
 			return "", err
 		}
 
+		zfi.Name, err = filepath.Rel(path, file)
+		if err != nil {
+			return "", err
+		}
+
 		zf, err := zipfile.CreateHeader(zfi)
 		if err != nil {
 			return "", err
@@ -263,13 +286,20 @@ func tempZipFiles(path string) (string, error) {
 	return tempfile.Name(), nil
 }
 
-func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, error) {
+func createBlob(cmd *cobra.Command, client *api.Client, path string, spinner *progress.Spinner) (string, error) {
 	bin, err := os.Open(path)
 	if err != nil {
 		return "", err
 	}
 	defer bin.Close()
 
+	// Get file info to retrieve the size
+	fileInfo, err := bin.Stat()
+	if err != nil {
+		return "", err
+	}
+	fileSize := fileInfo.Size()
+
 	hash := sha256.New()
 	if _, err := io.Copy(hash, bin); err != nil {
 		return "", err
@@ -279,13 +309,76 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 		return "", err
 	}
 
+	var pw progressWriter
+	status := "transferring model data 0%"
+	spinner.SetMessage(status)
+
+	done := make(chan struct{})
+	defer close(done)
+
+	go func() {
+		ticker := time.NewTicker(60 * time.Millisecond)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-ticker.C:
+				spinner.SetMessage(fmt.Sprintf("transferring model data %d%%", int(100*pw.n.Load()/fileSize)))
+			case <-done:
+				spinner.SetMessage("transferring model data 100%")
+				return
+			}
+		}
+	}()
+
 	digest := fmt.Sprintf("sha256:%x", hash.Sum(nil))
-	if err = client.CreateBlob(cmd.Context(), digest, bin); err != nil {
+	if err = client.CreateBlob(cmd.Context(), digest, io.TeeReader(bin, &pw)); err != nil {
 		return "", err
 	}
 	return digest, nil
 }
 
+type progressWriter struct {
+	n atomic.Int64
+}
+
+func (w *progressWriter) Write(p []byte) (n int, err error) {
+	w.n.Add(int64(len(p)))
+	return len(p), nil
+}
+
+func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	req := &api.GenerateRequest{
+		Model:     opts.Model,
+		KeepAlive: opts.KeepAlive,
+	}
+
+	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+}
+
+func StopHandler(cmd *cobra.Command, args []string) error {
+	opts := &runOptions{
+		Model:     args[0],
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := loadOrUnloadModel(cmd, opts); err != nil {
+		if strings.Contains(err.Error(), "not found") {
+			return fmt.Errorf("couldn't find model \"%s\" to stop", args[0])
+		}
+	}
+	return nil
+}
+
 func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true
 
@@ -362,9 +455,24 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 
 	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
 	opts.ParentModel = info.Details.ParentModel
-	opts.Messages = append(opts.Messages, info.Messages...)
 
 	if interactive {
+		if err := loadOrUnloadModel(cmd, &opts); err != nil {
+			return err
+		}
+
+		for _, msg := range info.Messages {
+			switch msg.Role {
+			case "user":
+				fmt.Printf(">>> %s\n", msg.Content)
+			case "assistant":
+				state := &displayResponseState{}
+				displayResponse(msg.Content, opts.WordWrap, state)
+				fmt.Println()
+				fmt.Println()
+			}
+		}
+
 		return generateInteractive(cmd, opts)
 	}
 	return generate(cmd, opts)
@@ -505,7 +613,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()
 
@@ -540,7 +648,15 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
 				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
 			}
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+
+			var until string
+			delta := time.Since(m.ExpiresAt)
+			if delta > 0 {
+				until = "Stopping..."
+			} else {
+				until = format.HumanTime(m.ExpiresAt, "Never")
+			}
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, until})
 		}
 	}
 
@@ -551,7 +667,7 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()
 
@@ -647,122 +763,89 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	showInfo(resp)
-
-	return nil
+	return showInfo(resp, os.Stdout)
 }
 
-func showInfo(resp *api.ShowResponse) {
-	arch := resp.ModelInfo["general.architecture"].(string)
+func showInfo(resp *api.ShowResponse, w io.Writer) error {
+	tableRender := func(header string, rows func() [][]string) {
+		fmt.Fprintln(w, " ", header)
+		table := tablewriter.NewWriter(w)
+		table.SetAlignment(tablewriter.ALIGN_LEFT)
+		table.SetBorder(false)
+		table.SetNoWhiteSpace(true)
+		table.SetTablePadding("    ")
 
-	modelData := [][]string{
-		{"arch", arch},
-		{"parameters", resp.Details.ParameterSize},
-		{"quantization", resp.Details.QuantizationLevel},
-		{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
-		{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
+		switch header {
+		case "Template", "System", "License":
+			table.SetColWidth(100)
+		}
+
+		table.AppendBulk(rows())
+		table.Render()
+		fmt.Fprintln(w)
 	}
 
-	mainTableData := [][]string{
-		{"Model"},
-		{renderSubTable(modelData, false)},
-	}
+	tableRender("Model", func() (rows [][]string) {
+		if resp.ModelInfo != nil {
+			arch := resp.ModelInfo["general.architecture"].(string)
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
+		} else {
+			rows = append(rows, []string{"", "architecture", resp.Details.Family})
+			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
+		}
+		rows = append(rows, []string{"", "quantization", resp.Details.QuantizationLevel})
+		return
+	})
 
 	if resp.ProjectorInfo != nil {
-		projectorData := [][]string{
-			{"arch", "clip"},
-			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
-		}
-
-		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
-			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
-		}
-
-		projectorData = append(projectorData,
-			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
-			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
-		)
-
-		mainTableData = append(mainTableData,
-			[]string{"Projector"},
-			[]string{renderSubTable(projectorData, false)},
-		)
+		tableRender("Projector", func() (rows [][]string) {
+			arch := resp.ProjectorInfo["general.architecture"].(string)
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.embedding_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "dimensions", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.projection_dim", arch)].(float64), 'f', -1, 64)})
+			return
+		})
 	}
 
 	if resp.Parameters != "" {
-		mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)})
+		tableRender("Parameters", func() (rows [][]string) {
+			scanner := bufio.NewScanner(strings.NewReader(resp.Parameters))
+			for scanner.Scan() {
+				if text := scanner.Text(); text != "" {
+					rows = append(rows, append([]string{""}, strings.Fields(text)...))
+				}
+			}
+			return
+		})
+	}
+
+	head := func(s string, n int) (rows [][]string) {
+		scanner := bufio.NewScanner(strings.NewReader(s))
+		for scanner.Scan() && (len(rows) < n || n < 0) {
+			if text := scanner.Text(); text != "" {
+				rows = append(rows, []string{"", strings.TrimSpace(text)})
+			}
+		}
+		return
 	}
 
 	if resp.System != "" {
-		mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)})
+		tableRender("System", func() [][]string {
+			return head(resp.System, 2)
+		})
 	}
 
 	if resp.License != "" {
-		mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)})
+		tableRender("License", func() [][]string {
+			return head(resp.License, 2)
+		})
 	}
 
-	table := tablewriter.NewWriter(os.Stdout)
-	table.SetAutoWrapText(false)
-	table.SetBorder(false)
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range mainTableData {
-		table.Append(v)
-	}
-
-	table.Render()
-}
-
-func renderSubTable(data [][]string, file bool) string {
-	var buf bytes.Buffer
-	table := tablewriter.NewWriter(&buf)
-	table.SetAutoWrapText(!file)
-	table.SetBorder(false)
-	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range data {
-		table.Append(v)
-	}
-
-	table.Render()
-
-	renderedTable := buf.String()
-	lines := strings.Split(renderedTable, "\n")
-	for i, line := range lines {
-		lines[i] = "\t" + line
-	}
-
-	return strings.Join(lines, "\n")
-}
-
-func twoLines(s string) [][]string {
-	lines := strings.Split(s, "\n")
-	res := [][]string{}
-
-	count := 0
-	for _, line := range lines {
-		line = strings.TrimSpace(line)
-		if line != "" {
-			count++
-			res = append(res, []string{line})
-			if count == 2 {
-				return res
-			}
-		}
-	}
-	return res
-}
-
-func formatParams(s string) string {
-	lines := strings.Split(s, "\n")
-	table := [][]string{}
-
-	for _, line := range lines {
-		table = append(table, strings.Fields(line))
-	}
-	return renderSubTable(table, false)
+	return nil
 }
 
 func CopyHandler(cmd *cobra.Command, args []string) error {
@@ -843,7 +926,6 @@ type runOptions struct {
 	WordWrap    bool
 	Format      string
 	System      string
-	Template    string
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
@@ -1037,7 +1119,6 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 		Images:    opts.Images,
 		Format:    opts.Format,
 		System:    opts.System,
-		Template:  opts.Template,
 		Options:   opts.Options,
 		KeepAlive: opts.KeepAlive,
 	}
@@ -1073,12 +1154,12 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	return nil
 }
 
-func RunServer(cmd *cobra.Command, _ []string) error {
+func RunServer(_ *cobra.Command, _ []string) error {
 	if err := initializeKeypair(); err != nil {
 		return err
 	}
 
-	ln, err := net.Listen("tcp", net.JoinHostPort(envconfig.Host.Host, envconfig.Host.Port))
+	ln, err := net.Listen("tcp", envconfig.Host().Host)
 	if err != nil {
 		return err
 	}
@@ -1147,7 +1228,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return fmt.Errorf("could not connect to ollama app, is it running?")
+			return errors.New("could not connect to ollama app, is it running?")
 		}
 	}
 	return nil
@@ -1254,6 +1335,15 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+
+	stopCmd := &cobra.Command{
+		Use:     "stop MODEL",
+		Short:   "Stop a running model",
+		Args:    cobra.ExactArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    StopHandler,
+	}
+
 	serveCmd := &cobra.Command{
 		Use:     "serve",
 		Aliases: []string{"start"},
@@ -1321,6 +1411,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
@@ -1343,10 +1434,12 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NUM_PARALLEL"],
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
+				envVars["OLLAMA_SCHED_SPREAD"],
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
-				envVars["OLLAMA_MAX_VRAM"],
+				envVars["OLLAMA_GPU_OVERHEAD"],
+				envVars["OLLAMA_LOAD_TIMEOUT"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)
@@ -1358,6 +1451,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
diff --git a/cmd/cmd_test.go b/cmd/cmd_test.go
new file mode 100644
index 000000000..0f8863cc7
--- /dev/null
+++ b/cmd/cmd_test.go
@@ -0,0 +1,206 @@
+package cmd
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestShowInfo(t *testing.T) {
+	t.Run("bare details", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+`
+
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("bare model info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			ModelInfo: map[string]any{
+				"general.architecture":    "test",
+				"general.parameter_count": float64(7_000_000_000),
+				"test.context_length":     float64(0),
+				"test.embedding_length":   float64(0),
+			},
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture        test    
+    parameters          7B      
+    context length      0       
+    embedding length    0       
+    quantization        FP16    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("parameters", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Parameters: `
+			stop never
+			stop gonna
+			stop give
+			stop you
+			stop up
+			temperature 99`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Parameters
+    stop           never    
+    stop           gonna    
+    stop           give     
+    stop           you      
+    stop           up       
+    temperature    99       
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("project info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			ProjectorInfo: map[string]any{
+				"general.architecture":         "clip",
+				"general.parameter_count":      float64(133_700_000),
+				"clip.vision.embedding_length": float64(0),
+				"clip.vision.projection_dim":   float64(0),
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Projector
+    architecture        clip       
+    parameters          133.70M    
+    embedding length    0          
+    dimensions          0          
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("system", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			System: `You are a pirate!
+Ahoy, matey!
+Weigh anchor!
+			`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  System
+    You are a pirate!    
+    Ahoy, matey!         
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("license", func(t *testing.T) {
+		var b bytes.Buffer
+		license, err := os.ReadFile(filepath.Join("..", "LICENSE"))
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			License: string(license),
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  License
+    MIT License             
+    Copyright (c) Ollama    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+}
diff --git a/cmd/interactive.go b/cmd/interactive.go
index 9214f2db5..94578f11b 100644
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@@ -1,6 +1,7 @@
 package cmd
 
 import (
+	"cmp"
 	"errors"
 	"fmt"
 	"io"
@@ -9,14 +10,14 @@ import (
 	"path/filepath"
 	"regexp"
 	"slices"
-	"sort"
 	"strings"
 
 	"github.com/spf13/cobra"
+	"golang.org/x/exp/maps"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
 )
@@ -27,49 +28,9 @@ const (
 	MultilineNone MultilineState = iota
 	MultilinePrompt
 	MultilineSystem
-	MultilineTemplate
 )
 
-func loadModel(cmd *cobra.Command, opts *runOptions) error {
-	p := progress.NewProgress(os.Stderr)
-	defer p.StopAndClear()
-
-	spinner := progress.NewSpinner("")
-	p.Add("", spinner)
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	chatReq := &api.ChatRequest{
-		Model:     opts.Model,
-		KeepAlive: opts.KeepAlive,
-	}
-
-	return client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
-		p.StopAndClear()
-		for _, msg := range opts.Messages {
-			switch msg.Role {
-			case "user":
-				fmt.Printf(">>> %s\n", msg.Content)
-			case "assistant":
-				state := &displayResponseState{}
-				displayResponse(msg.Content, opts.WordWrap, state)
-				fmt.Println()
-				fmt.Println()
-			}
-		}
-		return nil
-	})
-}
-
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
-	err := loadModel(cmd, &opts)
-	if err != nil {
-		return err
-	}
-
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
 		fmt.Fprintln(os.Stderr, "  /set            Set session variables")
@@ -94,7 +55,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
 		fmt.Fprintln(os.Stderr, "  /set parameter ...     Set a parameter")
 		fmt.Fprintln(os.Stderr, "  /set system <string>   Set system message")
-		fmt.Fprintln(os.Stderr, "  /set template <string> Set prompt template")
 		fmt.Fprintln(os.Stderr, "  /set history           Enable history")
 		fmt.Fprintln(os.Stderr, "  /set nohistory         Disable history")
 		fmt.Fprintln(os.Stderr, "  /set wordwrap          Enable wordwrap")
@@ -140,6 +100,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /set parameter num_predict <int>      Max number of tokens to predict")
 		fmt.Fprintln(os.Stderr, "  /set parameter top_k <int>            Pick from top k num of tokens")
 		fmt.Fprintln(os.Stderr, "  /set parameter top_p <float>          Pick token based on sum of probabilities")
+		fmt.Fprintln(os.Stderr, "  /set parameter min_p <float>          Pick token based on top token probability * min_p")
 		fmt.Fprintln(os.Stderr, "  /set parameter num_ctx <int>          Set the context size")
 		fmt.Fprintln(os.Stderr, "  /set parameter temperature <float>    Set creativity level")
 		fmt.Fprintln(os.Stderr, "  /set parameter repeat_penalty <float> How strongly to penalize repetitions")
@@ -159,7 +120,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		return err
 	}
 
-	if envconfig.NoHistory {
+	if envconfig.NoHistory() {
 		scanner.HistoryDisable()
 	}
 
@@ -204,10 +165,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 				opts.Messages = append(opts.Messages, api.Message{Role: "system", Content: opts.System})
 				fmt.Println("Set system message.")
 				sb.Reset()
-			case MultilineTemplate:
-				opts.Template = sb.String()
-				fmt.Println("Set prompt template.")
-				sb.Reset()
 			}
 
 			multiline = MultilineNone
@@ -239,7 +196,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
-			if err := loadModel(cmd, &opts); err != nil {
+			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				return err
 			}
 			continue
@@ -326,17 +283,13 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					}
 					fmt.Printf("Set parameter '%s' to '%s'\n", args[2], strings.Join(params, ", "))
 					opts.Options[args[2]] = fp[args[2]]
-				case "system", "template":
+				case "system":
 					if len(args) < 3 {
 						usageSet()
 						continue
 					}
 
-					if args[1] == "system" {
-						multiline = MultilineSystem
-					} else if args[1] == "template" {
-						multiline = MultilineTemplate
-					}
+					multiline = MultilineSystem
 
 					line := strings.Join(args[2:], " ")
 					line, ok := strings.CutPrefix(line, `"""`)
@@ -356,23 +309,17 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						continue
 					}
 
-					if args[1] == "system" {
-						opts.System = sb.String() // for display in modelfile
-						newMessage := api.Message{Role: "system", Content: sb.String()}
-						// Check if the slice is not empty and the last message is from 'system'
-						if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
-							// Replace the last message
-							opts.Messages[len(opts.Messages)-1] = newMessage
-						} else {
-							opts.Messages = append(opts.Messages, newMessage)
-						}
-						fmt.Println("Set system message.")
-						sb.Reset()
-					} else if args[1] == "template" {
-						opts.Template = sb.String()
-						fmt.Println("Set prompt template.")
-						sb.Reset()
+					opts.System = sb.String() // for display in modelfile
+					newMessage := api.Message{Role: "system", Content: sb.String()}
+					// Check if the slice is not empty and the last message is from 'system'
+					if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
+						// Replace the last message
+						opts.Messages[len(opts.Messages)-1] = newMessage
+					} else {
+						opts.Messages = append(opts.Messages, newMessage)
 					}
+					fmt.Println("Set system message.")
+					sb.Reset()
 
 					sb.Reset()
 					continue
@@ -391,10 +338,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					return err
 				}
 				req := &api.ShowRequest{
-					Name:     opts.Model,
-					System:   opts.System,
-					Template: opts.Template,
-					Options:  opts.Options,
+					Name:    opts.Model,
+					System:  opts.System,
+					Options: opts.Options,
 				}
 				resp, err := client.Show(cmd.Context(), req)
 				if err != nil {
@@ -404,7 +350,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					showInfo(resp)
+					_ = showInfo(resp, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -437,12 +383,9 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						fmt.Println("No system message was specified for this model.")
 					}
 				case "template":
-					switch {
-					case opts.Template != "":
-						fmt.Println(opts.Template + "\n")
-					case resp.Template != "":
+					if resp.Template != "" {
 						fmt.Println(resp.Template)
-					default:
+					} else {
 						fmt.Println("No prompt template was specified for this model.")
 					}
 				default:
@@ -526,35 +469,35 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func buildModelfile(opts runOptions) string {
-	var mf strings.Builder
-	model := opts.ParentModel
-	if model == "" {
-		model = opts.Model
-	}
-	fmt.Fprintf(&mf, "FROM %s\n", model)
+	var f parser.File
+	f.Commands = append(f.Commands, parser.Command{Name: "model", Args: cmp.Or(opts.ParentModel, opts.Model)})
+
 	if opts.System != "" {
-		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
+		f.Commands = append(f.Commands, parser.Command{Name: "system", Args: opts.System})
 	}
 
-	if opts.Template != "" {
-		fmt.Fprintf(&mf, "TEMPLATE \"\"\"%s\"\"\"\n", opts.Template)
-	}
-
-	keys := make([]string, 0)
-	for k := range opts.Options {
-		keys = append(keys, k)
-	}
-	sort.Strings(keys)
+	keys := maps.Keys(opts.Options)
+	slices.Sort(keys)
 	for _, k := range keys {
-		fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, opts.Options[k])
+		v := opts.Options[k]
+		var cmds []parser.Command
+		switch t := v.(type) {
+		case []string:
+			for _, s := range t {
+				cmds = append(cmds, parser.Command{Name: k, Args: s})
+			}
+		default:
+			cmds = append(cmds, parser.Command{Name: k, Args: fmt.Sprintf("%v", t)})
+		}
+
+		f.Commands = append(f.Commands, cmds...)
 	}
-	fmt.Fprintln(&mf)
 
 	for _, msg := range opts.Messages {
-		fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content)
+		f.Commands = append(f.Commands, parser.Command{Name: "message", Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content)})
 	}
 
-	return mf.String()
+	return f.String()
 }
 
 func normalizeFilePath(fp string) string {
@@ -640,7 +583,7 @@ func getImageData(filePath string) ([]byte, error) {
 	// Check if the file size exceeds 100MB
 	var maxSize int64 = 100 * 1024 * 1024 // 100MB in bytes
 	if info.Size() > maxSize {
-		return nil, fmt.Errorf("file size exceeds maximum limit (100MB)")
+		return nil, errors.New("file size exceeds maximum limit (100MB)")
 	}
 
 	buf = make([]byte, info.Size())
diff --git a/cmd/interactive_test.go b/cmd/interactive_test.go
index d9af01eb8..bb7e0abaf 100644
--- a/cmd/interactive_test.go
+++ b/cmd/interactive_test.go
@@ -1,12 +1,10 @@
 package cmd
 
 import (
-	"bytes"
 	"testing"
-	"text/template"
 
+	"github.com/google/go-cmp/cmp"
 	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
 )
@@ -57,61 +55,53 @@ d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8
 
 func TestModelfileBuilder(t *testing.T) {
 	opts := runOptions{
-		Model:    "hork",
-		System:   "You are part horse and part shark, but all hork. Do horklike things",
-		Template: "This is a template.",
+		Model:  "hork",
+		System: "You are part horse and part shark, but all hork. Do horklike things",
 		Messages: []api.Message{
 			{Role: "user", Content: "Hey there hork!"},
 			{Role: "assistant", Content: "Yes it is true, I am half horse, half shark."},
 		},
-		Options: map[string]interface{}{},
+		Options: map[string]any{
+			"temperature":      0.9,
+			"seed":             42,
+			"penalize_newline": false,
+			"stop":             []string{"hi", "there"},
+		},
 	}
 
-	opts.Options["temperature"] = 0.9
-	opts.Options["seed"] = 42
-	opts.Options["penalize_newline"] = false
-	opts.Options["stop"] = []string{"hi", "there"}
-
-	mf := buildModelfile(opts)
-	expectedModelfile := `FROM {{.Model}}
-SYSTEM """{{.System}}"""
-TEMPLATE """{{.Template}}"""
+	t.Run("model", func(t *testing.T) {
+		expect := `FROM hork
+SYSTEM You are part horse and part shark, but all hork. Do horklike things
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop [hi there]
+PARAMETER stop hi
+PARAMETER stop there
 PARAMETER temperature 0.9
-
-MESSAGE user """Hey there hork!"""
-MESSAGE assistant """Yes it is true, I am half horse, half shark."""
+MESSAGE user Hey there hork!
+MESSAGE assistant Yes it is true, I am half horse, half shark.
 `
 
-	tmpl, err := template.New("").Parse(expectedModelfile)
-	require.NoError(t, err)
+		actual := buildModelfile(opts)
+		if diff := cmp.Diff(expect, actual); diff != "" {
+			t.Errorf("mismatch (-want +got):\n%s", diff)
+		}
+	})
 
-	var buf bytes.Buffer
-	err = tmpl.Execute(&buf, opts)
-	require.NoError(t, err)
-	assert.Equal(t, buf.String(), mf)
-
-	opts.ParentModel = "horseshark"
-	mf = buildModelfile(opts)
-	expectedModelfile = `FROM {{.ParentModel}}
-SYSTEM """{{.System}}"""
-TEMPLATE """{{.Template}}"""
+	t.Run("parent model", func(t *testing.T) {
+		opts.ParentModel = "horseshark"
+		expect := `FROM horseshark
+SYSTEM You are part horse and part shark, but all hork. Do horklike things
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop [hi there]
+PARAMETER stop hi
+PARAMETER stop there
 PARAMETER temperature 0.9
-
-MESSAGE user """Hey there hork!"""
-MESSAGE assistant """Yes it is true, I am half horse, half shark."""
+MESSAGE user Hey there hork!
+MESSAGE assistant Yes it is true, I am half horse, half shark.
 `
-
-	tmpl, err = template.New("").Parse(expectedModelfile)
-	require.NoError(t, err)
-
-	var parentBuf bytes.Buffer
-	err = tmpl.Execute(&parentBuf, opts)
-	require.NoError(t, err)
-	assert.Equal(t, parentBuf.String(), mf)
+		actual := buildModelfile(opts)
+		if diff := cmp.Diff(expect, actual); diff != "" {
+			t.Errorf("mismatch (-want +got):\n%s", diff)
+		}
+	})
 }
diff --git a/cmd/start_darwin.go b/cmd/start_darwin.go
index 82b09ad62..1a9a1ae87 100644
--- a/cmd/start_darwin.go
+++ b/cmd/start_darwin.go
@@ -2,7 +2,7 @@ package cmd
 
 import (
 	"context"
-	"fmt"
+	"errors"
 	"os"
 	"os/exec"
 	"strings"
@@ -20,7 +20,7 @@ func startApp(ctx context.Context, client *api.Client) error {
 		return err
 	}
 	if !strings.Contains(link, "Ollama.app") {
-		return fmt.Errorf("could not find ollama app")
+		return errors.New("could not find ollama app")
 	}
 	path := strings.Split(link, "Ollama.app")
 	if err := exec.Command("/usr/bin/open", "-a", path[0]+"Ollama.app").Run(); err != nil {
diff --git a/cmd/start_default.go b/cmd/start_default.go
index c9d6137b9..5eabb2862 100644
--- a/cmd/start_default.go
+++ b/cmd/start_default.go
@@ -4,11 +4,11 @@ package cmd
 
 import (
 	"context"
-	"fmt"
+	"errors"
 
 	"github.com/ollama/ollama/api"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {
-	return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it")
+	return errors.New("could not connect to ollama server, run 'ollama serve' to start it")
 }
diff --git a/cmd/start_windows.go b/cmd/start_windows.go
index 6024a2352..5bca24331 100644
--- a/cmd/start_windows.go
+++ b/cmd/start_windows.go
@@ -31,7 +31,7 @@ func startApp(ctx context.Context, client *api.Client) error {
 			// Finally look in the path
 			appExe, err = exec.LookPath(AppName)
 			if err != nil {
-				return fmt.Errorf("could not locate ollama app")
+				return errors.New("could not locate ollama app")
 			}
 		}
 	}
diff --git a/convert/convert.go b/convert/convert.go
index 103de457c..44783b6e8 100644
--- a/convert/convert.go
+++ b/convert/convert.go
@@ -1,200 +1,232 @@
 package convert
 
 import (
-	"cmp"
-	"encoding/binary"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
+	"io/fs"
 	"log/slog"
-	"os"
-	"path/filepath"
-	"slices"
 	"strings"
 
-	"google.golang.org/protobuf/proto"
-
-	"github.com/ollama/ollama/convert/sentencepiece"
 	"github.com/ollama/ollama/llm"
 )
 
-const (
-	_ int32 = iota
-	tokenTypeNormal
-	tokenTypeUnknown
-	tokenTypeControl
-	tokenTypeUserDefined
-	tokenTypeUnused
-	tokenTypeByte
-)
-
-type Params struct {
-	Architectures     []string `json:"architectures"`
-	VocabSize         int      `json:"vocab_size"`
-	HiddenSize        int      `json:"hidden_size"`       // n_embd
-	HiddenLayers      int      `json:"num_hidden_layers"` // n_layer
-	ContextSize       int      `json:"max_position_embeddings"`
-	IntermediateSize  int      `json:"intermediate_size"`
-	AttentionHeads    int      `json:"num_attention_heads"` // n_head
-	KeyValHeads       int      `json:"num_key_value_heads"`
-	NormEPS           float64  `json:"rms_norm_eps"`
-	BoSTokenID        int      `json:"bos_token_id"`
-	EoSTokenID        int      `json:"eos_token_id"`
-	HeadDimension     int      `json:"head_dim"`
-	PaddingTokenID    int      `json:"pad_token_id"`
-	RopeFrequencyBase float64  `json:"rope_theta"`
-
-	Experts     int `json:"num_local_experts"`
-	ExpertsUsed int `json:"num_experts_per_tok"`
-
-	PreTokenizer string
-
-	ByteOrder
+type ModelParameters struct {
+	Architectures []string `json:"architectures"`
+	VocabSize     uint32   `json:"vocab_size"`
 }
 
-type ByteOrder interface {
-	binary.ByteOrder
-	binary.AppendByteOrder
+type AdapterParameters struct {
+	Alpha          uint32 `json:"lora_alpha"`
+	LoraLayers     uint32 `json:"lora_layers"`
+	LoraParameters struct {
+		Rank  uint32  `json:"rank"`
+		Alpha float32 `json:"alpha"`
+		Scale float32 `json:"scale"`
+	} `json:"lora_parameters"`
 }
 
-type ModelArch interface {
-	GetTensors() error
-	LoadVocab() error
-	WriteGGUF(io.WriteSeeker) error
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
+	kv := llm.KV{
+		"general.file_type":            uint32(1),
+		"general.quantization_version": uint32(2),
+		"tokenizer.ggml.pre":           t.Pre,
+		"tokenizer.ggml.model":         t.Vocabulary.Model,
+		"tokenizer.ggml.tokens":        t.Vocabulary.Tokens,
+		"tokenizer.ggml.scores":        t.Vocabulary.Scores,
+		"tokenizer.ggml.token_type":    t.Vocabulary.Types,
+	}
+
+	if len(t.Merges) > 0 {
+		kv["tokenizer.ggml.merges"] = t.Merges
+	}
+
+	if t.Template != "" {
+		kv["tokenizer.chat_template"] = t.Template
+	}
+
+	for _, sv := range t.SpecialVocabulary {
+		kv[fmt.Sprintf("tokenizer.ggml.%s_token_id", sv.Key())] = uint32(sv.ID)
+		kv[fmt.Sprintf("tokenizer.ggml.add_%s_token", sv.Key())] = sv.AddToken
+	}
+
+	return kv
 }
 
-type ModelFormat interface {
-	GetLayerName(string) (string, error)
-	GetTensors(string, *Params) ([]llm.Tensor, error)
-	GetParams(string) (*Params, error)
-	GetModelArch(string, string, *Params) (ModelArch, error)
+func (p AdapterParameters) KV() llm.KV {
+	var alpha float32
+	if p.LoraParameters.Alpha == 0 {
+		alpha = float32(p.Alpha)
+	} else {
+		alpha = p.LoraParameters.Alpha
+	}
+
+	kv := llm.KV{
+		"adapter.lora.alpha": alpha,
+		"adapter.type":       "lora",
+		"general.file_type":  uint32(1),
+		"general.type":       "adapter",
+		"general.version":    "v0.2",
+	}
+
+	return kv
 }
 
-type ModelData struct {
-	Path    string
-	Name    string
-	Params  *Params
-	Vocab   *Vocab
-	Tensors []llm.Tensor
-	Format  ModelFormat
+func (ModelParameters) specialTokenTypes() []string {
+	return []string{
+		"bos", "eos", "unk", "sep", "pad", "cls", "mask",
+	}
 }
 
-func GetModelFormat(dirname string) (ModelFormat, error) {
-	files, err := filepath.Glob(filepath.Join(dirname, "*"))
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
+}
+
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
+}
+
+type ModelConverter interface {
+	// KV maps parameters to LLM key-values
+	KV(*Tokenizer) llm.KV
+	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
+	Tensors([]Tensor) []llm.Tensor
+	// Replacements returns a list of string pairs to replace in tensor names.
+	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
+	Replacements() []string
+
+	// specialTokenTypes returns any special token types the model uses
+	specialTokenTypes() []string
+	// writeFile writes the model to the provided io.WriteSeeker
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+}
+
+type moreParser interface {
+	parseMore(fs.FS) error
+}
+
+type AdapterConverter interface {
+	// KV maps parameters to LLM key-values
+	KV(llm.KV) llm.KV
+	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
+	Tensors([]Tensor) []llm.Tensor
+	// Replacements returns a list of string pairs to replace in tensor names.
+	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
+	Replacements() []string
+
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
+}
+
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
+	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	for _, fn := range files {
-		if strings.HasSuffix(fn, ".safetensors") {
-			return &SafetensorFormat{}, nil
-		} else if strings.HasSuffix(fn, ".bin") || strings.HasSuffix(fn, ".pth") {
-			slog.Debug("model is torch")
-			return &TorchFormat{}, nil
-		}
+	var p AdapterParameters
+	if err := json.Unmarshal(bts, &p); err != nil {
+		return err
 	}
 
-	return nil, fmt.Errorf("couldn't determine model format")
-}
+	arch, ok := baseKV["general.architecture"]
+	if !ok {
+		return errors.New("architecture not set for the base model")
+	}
 
-// Details on gguf's tokenizer can be found at:
-// https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#tokenizer
-type Vocab struct {
-	Tokens []string
-	Scores []float32
-	Types  []int32
-	Merges []string
-}
+	var conv AdapterConverter
+	switch arch {
+	case "llama":
+		conv = &llamaAdapter{}
+	case "gemma2":
+		conv = &gemma2Adapter{}
+	default:
+		return errors.New("unsupported architecture")
+	}
 
-func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
-	slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
-	in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	// To regenerate sentencepiece from the protobufs use:
-	// protoc -I=./ --go_out=./ sentencepiece_model.proto
-	modelProto := &sentencepiece.ModelProto{}
-	if err := proto.Unmarshal(in, modelProto); err != nil {
-		return nil, err
+	if err := json.Unmarshal(bts, conv); err != nil {
+		return err
 	}
 
-	v := &Vocab{
-		Tokens: make([]string, 0),
-		Scores: make([]float32, 0),
-		Types:  make([]int32, 0),
-	}
-
-	pieces := modelProto.GetPieces()
-	for _, p := range pieces {
-		v.Tokens = append(v.Tokens, p.GetPiece())
-		v.Scores = append(v.Scores, p.GetScore())
-		t := p.GetType()
-		switch t {
-		case sentencepiece.ModelProto_SentencePiece_UNKNOWN:
-		case sentencepiece.ModelProto_SentencePiece_CONTROL:
-		case sentencepiece.ModelProto_SentencePiece_UNUSED:
-		case sentencepiece.ModelProto_SentencePiece_BYTE:
-		default:
-			t = sentencepiece.ModelProto_SentencePiece_NORMAL
-		}
-		v.Types = append(v.Types, int32(t))
-	}
-
-	slog.Info(fmt.Sprintf("vocab size: %d", len(v.Tokens)))
-
-	// add any additional tokens
-	addIn, err := os.ReadFile(filepath.Join(dirpath, "added_tokens.json"))
-	if os.IsNotExist(err) {
-		return v, nil
-	} else if err != nil {
-		return nil, err
-	}
-
-	slog.Info("reading user defined tokens")
-
-	var extraTokenData map[string]int
-	if err := json.Unmarshal(addIn, &extraTokenData); err != nil {
-		return nil, err
-	}
-
-	type token struct {
-		key string
-		pos int
-	}
-
-	extraTokens := make([]token, 0)
-	for k, id := range extraTokenData {
-		extraTokens = append(extraTokens, token{k, id})
-	}
-
-	slices.SortFunc(extraTokens, func(a, b token) int {
-		return cmp.Compare(a.pos, b.pos)
-	})
-
-	numToks := len(v.Tokens)
-
-	for cnt, t := range extraTokens {
-		// the token id should match the specific index for the total number of tokens
-		if t.pos != cnt+numToks {
-			return nil, fmt.Errorf("token ID '%d' for '%s' doesn't match total token size", t.pos, t.key)
-		}
-		v.Tokens = append(v.Tokens, t.key)
-		v.Scores = append(v.Scores, -1000.0)
-		v.Types = append(v.Types, tokenTypeUserDefined)
-	}
-	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
-
-	if params.VocabSize > len(v.Tokens) {
-		missingTokens := params.VocabSize - len(v.Tokens)
-		slog.Warn(fmt.Sprintf("vocab is missing %d tokens", missingTokens))
-		for cnt := range missingTokens {
-			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
-			v.Scores = append(v.Scores, -1)
-			v.Types = append(v.Types, tokenTypeUserDefined)
-		}
-	}
-
-	return v, nil
+	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+}
+
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// and files it finds in the input path.
+// Supported input model formats include safetensors.
+// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
+func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
+	bts, err := fs.ReadFile(fsys, "config.json")
+	if err != nil {
+		return err
+	}
+
+	var p ModelParameters
+	if err := json.Unmarshal(bts, &p); err != nil {
+		return err
+	}
+
+	if len(p.Architectures) < 1 {
+		return errors.New("unknown architecture")
+	}
+
+	var conv ModelConverter
+	switch p.Architectures[0] {
+	case "LlamaForCausalLM", "MistralForCausalLM":
+		conv = &llamaModel{}
+	case "MixtralForCausalLM":
+		conv = &mixtralModel{}
+	case "GemmaForCausalLM":
+		conv = &gemmaModel{}
+	case "Gemma2ForCausalLM":
+		conv = &gemma2Model{}
+	case "Phi3ForCausalLM":
+		conv = &phi3Model{}
+	case "BertModel":
+		conv = &bertModel{}
+	default:
+		return errors.New("unsupported architecture")
+	}
+
+	if err := json.Unmarshal(bts, conv); err != nil {
+		return err
+	}
+
+	if t, ok := conv.(moreParser); ok {
+		if err := t.parseMore(fsys); err != nil {
+			return err
+		}
+	}
+
+	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
+	if err != nil {
+		return err
+	}
+
+	vocabSize := int(p.VocabSize)
+	switch {
+	case vocabSize > len(t.Vocabulary.Tokens):
+		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
+		for i := range vocabSize - len(t.Vocabulary.Tokens) {
+			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
+			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
+			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
+		}
+	case vocabSize < len(t.Vocabulary.Tokens):
+		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
+	default:
+		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
+	}
+
+	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
+	if err != nil {
+		return err
+	}
+
+	return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts))
 }
diff --git a/convert/convert_bert.go b/convert/convert_bert.go
new file mode 100644
index 000000000..ea5facaa5
--- /dev/null
+++ b/convert/convert_bert.go
@@ -0,0 +1,174 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"io/fs"
+	"path/filepath"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type bertModel struct {
+	ModelParameters
+	NLayers               uint32  `json:"n_layers"`
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	NLayer                uint32  `json:"n_layer"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	NCtx                  uint32  `json:"n_ctx"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	NEmbd                 uint32  `json:"n_embd"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NInner                uint32  `json:"n_inner"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NHead                 uint32  `json:"n_head"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	LayerNormEPS          float32 `json:"layer_norm_eps"`
+	LayerNormEpsilon      float32 `json:"layer_norm_epsilon"`
+	NormEpsilon           float32 `json:"norm_epsilon"`
+
+	PoolingType uint32
+}
+
+var (
+	_ ModelConverter = (*bertModel)(nil)
+	_ moreParser     = (*bertModel)(nil)
+)
+
+func (p *bertModel) parseMore(fsys fs.FS) error {
+	bts, err := fs.ReadFile(fsys, "modules.json")
+	if err != nil {
+		return err
+	}
+
+	var modules []struct {
+		Type string `json:"type"`
+		Path string `json:"path"`
+	}
+
+	if err := json.Unmarshal(bts, &modules); err != nil {
+		return err
+	}
+
+	var pooling string
+	for _, m := range modules {
+		if m.Type == "sentence_transformers.models.Pooling" {
+			pooling = m.Path
+			break
+		}
+	}
+
+	if pooling != "" {
+		bts, err := fs.ReadFile(fsys, filepath.Join(pooling, "config.json"))
+		if err != nil {
+			return err
+		}
+
+		var pc struct {
+			PoolingModeCLSToken   bool `json:"pooling_mode_cls_token"`
+			PoolingModeMeanTokens bool `json:"pooling_mode_mean_tokens"`
+		}
+
+		if err := json.Unmarshal(bts, &pc); err != nil {
+			return err
+		}
+
+		if pc.PoolingModeMeanTokens {
+			p.PoolingType = 1
+		} else if pc.PoolingModeCLSToken {
+			p.PoolingType = 2
+		}
+	}
+
+	return nil
+}
+
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "bert"
+	kv["bert.attention.causal"] = false
+	kv["bert.pooling_type"] = p.PoolingType
+
+	kv["bert.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
+
+	if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
+		kv["bert.context_length"] = contextLength
+	}
+
+	if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
+		kv["bert.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	}
+
+	if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
+		kv["bert.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
+	}
+
+	if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
+		kv["bert.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+	}
+
+	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
+		kv["bert.attention.layer_norm_epsilon"] = layerNormEpsilon
+	}
+
+	kv["tokenizer.ggml.model"] = "bert"
+	kv["tokenizer.ggml.token_type_count"] = uint32(2)
+
+	// convert to phantom space tokens
+	for i, e := range t.Tokens {
+		if strings.HasPrefix(e, "[") && strings.HasSuffix(e, "]") {
+			// noop
+		} else if strings.HasPrefix(e, "##") {
+			t.Tokens[i] = e[2:]
+		} else {
+			t.Tokens[i] = "\u2581" + e
+		}
+	}
+
+	kv["tokenizer.ggml.tokens"] = t.Tokens
+
+	return kv
+}
+
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		if slices.Contains([]string{
+			"embeddings.position_ids",
+			"pooler.dense.weight",
+			"pooler.dense.bias",
+		}, t.Name()) {
+			continue
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (bertModel) Replacements() []string {
+	return []string{
+		"encoder.layer", "blk",
+		"encoder.layers", "blk",
+		"embeddings.word_embeddings", "token_embd",
+		"embeddings.token_type_embeddings", "token_types",
+		"embeddings.LayerNorm", "token_embd_norm",
+		"embeddings.position_embeddings", "position_embd",
+		"attention.self.query", "attn_q",
+		"attention.self.key", "attn_k",
+		"attention.self.value", "attn_v",
+		"attention.output.dense", "attn_output",
+		"attention.output.LayerNorm", "attn_output_norm",
+		"intermediate.dense", "ffn_up",
+		"output.dense", "ffn_down",
+		"output.LayerNorm", "layer_output_norm",
+	}
+}
diff --git a/convert/convert_gemma.go b/convert/convert_gemma.go
new file mode 100644
index 000000000..b88652947
--- /dev/null
+++ b/convert/convert_gemma.go
@@ -0,0 +1,100 @@
+package convert
+
+import (
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type gemmaModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+	HeadDim               uint32  `json:"head_dim"`
+}
+
+var _ ModelConverter = (*gemmaModel)(nil)
+
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma"
+	kv["gemma.context_length"] = p.MaxPositionEmbeddings
+	kv["gemma.embedding_length"] = p.HiddenSize
+	kv["gemma.block_count"] = p.HiddenLayers
+	kv["gemma.feed_forward_length"] = p.IntermediateSize
+	kv["gemma.attention.head_count"] = p.NumAttentionHeads
+	kv["gemma.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["gemma.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["gemma.attention.key_length"] = p.HeadDim
+	kv["gemma.attention.value_length"] = p.HeadDim
+	kv["tokenizer.ggml.eot_token_id"] = uint32(107)
+	kv["tokenizer.ggml.middle_token_id"] = uint32(68)
+	kv["tokenizer.ggml.prefix_token_id"] = uint32(67)
+	kv["tokenizer.ggml.suffix_token_id"] = uint32(69)
+	return kv
+}
+
+func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		if strings.HasSuffix(t.Name(), "_norm.weight") {
+			t.SetRepacker(p.addOne)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *gemmaModel) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "ffn_norm",
+	}
+}
+
+func (*gemmaModel) addOne(_ string, data []float32, shape []uint64) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0])), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, int(shape[0]))
+
+	n, err := n.Add(ones)
+	if err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
diff --git a/convert/convert_gemma2.go b/convert/convert_gemma2.go
new file mode 100644
index 000000000..0f98c1e38
--- /dev/null
+++ b/convert/convert_gemma2.go
@@ -0,0 +1,53 @@
+package convert
+
+import (
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Model struct {
+	gemmaModel
+	SlidingWindow         uint32  `json:"sliding_window"`
+	AttentionLogitSoftcap float32 `json:"attn_logit_softcapping"`
+	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
+}
+
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma2"
+	kv["gemma2.context_length"] = p.MaxPositionEmbeddings
+	kv["gemma2.embedding_length"] = p.HiddenSize
+	kv["gemma2.block_count"] = p.HiddenLayers
+	kv["gemma2.feed_forward_length"] = p.IntermediateSize
+	kv["gemma2.attention.head_count"] = p.NumAttentionHeads
+	kv["gemma2.attention.head_count_kv"] = p.NumKeyValueHeads
+	kv["gemma2.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["gemma2.attention.key_length"] = p.HeadDim
+	kv["gemma2.attention.value_length"] = p.HeadDim
+	kv["gemma2.attention.sliding_window"] = p.SlidingWindow
+	kv["gemma2.attn_logit_softcapping"] = p.AttentionLogitSoftcap
+	kv["gemma2.final_logit_softcapping"] = p.FinalLogitSoftcap
+	kv["tokenizer.ggml.eot_token_id"] = uint32(107)
+	kv["tokenizer.ggml.middle_token_id"] = uint32(68)
+	kv["tokenizer.ggml.prefix_token_id"] = uint32(67)
+	kv["tokenizer.ggml.suffix_token_id"] = uint32(69)
+	return kv
+}
+
+func (p *gemma2Model) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+	}
+}
diff --git a/convert/convert_gemma2_adapter.go b/convert/convert_gemma2_adapter.go
new file mode 100644
index 000000000..a89a25f4c
--- /dev/null
+++ b/convert/convert_gemma2_adapter.go
@@ -0,0 +1,91 @@
+package convert
+
+import (
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type gemma2Adapter struct {
+	AdapterParameters
+}
+
+var _ AdapterConverter = (*gemma2Adapter)(nil)
+
+func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "gemma2"
+	return kv
+}
+
+func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *gemma2Adapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *gemma2Adapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	if err := n.T(1, 0); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
diff --git a/convert/convert_llama.go b/convert/convert_llama.go
new file mode 100644
index 000000000..5dedb829d
--- /dev/null
+++ b/convert/convert_llama.go
@@ -0,0 +1,213 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"math"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type llamaModel struct {
+	ModelParameters
+	NLayers               uint32  `json:"n_layers"`
+	NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+	NLayer                uint32  `json:"n_layer"`
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	NCtx                  uint32  `json:"n_ctx"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	NEmbd                 uint32  `json:"n_embd"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NInner                uint32  `json:"n_inner"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NHead                 uint32  `json:"n_head"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RopeTheta             float32 `json:"rope_theta"`
+	RopeScaling           struct {
+		Type                            string  `json:"type"`
+		RopeType                        string  `json:"rope_type"`
+		Factor                          float32 `json:"factor"`
+		LowFrequencyFactor              float32 `json:"low_freq_factor"`
+		HighFrequencyFactor             float32 `json:"high_freq_factor"`
+		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
+
+		factors ropeFactor
+	} `json:"rope_scaling"`
+	RMSNormEPS       float32 `json:"rms_norm_eps"`
+	LayerNormEPS     float32 `json:"layer_norm_eps"`
+	LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
+	NormEpsilon      float32 `json:"norm_epsilon"`
+	HeadDim          uint32  `json:"head_dim"`
+}
+
+var _ ModelConverter = (*llamaModel)(nil)
+
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "llama"
+	kv["llama.vocab_size"] = p.VocabSize
+
+	kv["llama.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
+
+	if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
+		kv["llama.context_length"] = contextLength
+	}
+
+	if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
+		kv["llama.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	}
+
+	if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
+		kv["llama.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
+	}
+
+	if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
+		kv["llama.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+		kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
+	}
+
+	if p.RopeTheta > 0 {
+		kv["llama.rope.freq_base"] = p.RopeTheta
+	}
+
+	if p.RopeScaling.Type == "linear" {
+		kv["llama.rope.scaling.type"] = p.RopeScaling.Type
+		kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
+	} else if p.RopeScaling.RopeType == "llama3" {
+		dim := p.HiddenSize / p.NumAttentionHeads
+		for i := uint32(0); i < dim; i += 2 {
+			factor := cmp.Or(p.RopeScaling.Factor, 8.0)
+			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
+			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
+
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
+			lambdaLow := float32(original) / factorLow
+			lambdaHigh := float32(original) / factorHigh
+
+			lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
+			if lambda < float64(lambdaHigh) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0)
+			} else if lambda > float64(lambdaLow) {
+				p.RopeScaling.factors = append(p.RopeScaling.factors, factor)
+			} else {
+				smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
+				p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0/((1-smooth)/factor+smooth))
+			}
+		}
+	}
+
+	if p.NumKeyValueHeads > 0 {
+		kv["llama.attention.head_count_kv"] = p.NumKeyValueHeads
+	}
+
+	if p.RMSNormEPS > 0 {
+		kv["llama.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	}
+
+	if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
+		kv["llama.attention.layer_norm_epsilon"] = layerNormEpsilon
+	}
+
+	if p.HeadDim > 0 {
+		kv["llama.attention.key_length"] = p.HeadDim
+		kv["llama.attention.value_length"] = p.HeadDim
+	}
+
+	return kv
+}
+
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+
+	if p.RopeScaling.factors != nil {
+		out = append(out, llm.Tensor{
+			Name:     "rope_freqs.weight",
+			Kind:     0,
+			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
+			WriterTo: p.RopeScaling.factors,
+		})
+	}
+
+	for _, t := range ts {
+		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
+			strings.HasSuffix(t.Name(), "attn_k.weight") {
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *llamaModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "ffn_norm",
+	}
+}
+
+func (p *llamaModel) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		dims = append(dims, int(dim))
+	}
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
diff --git a/convert/convert_llama_adapter.go b/convert/convert_llama_adapter.go
new file mode 100644
index 000000000..08ddee10a
--- /dev/null
+++ b/convert/convert_llama_adapter.go
@@ -0,0 +1,169 @@
+package convert
+
+import (
+	"cmp"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type llamaAdapter struct {
+	AdapterParameters
+	NumAttentionHeads uint32 `json:"num_attention_heads"`
+	NumKeyValueHeads  uint32 `json:"num_key_value_heads"`
+}
+
+var _ AdapterConverter = (*llamaAdapter)(nil)
+
+func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+	kv := p.AdapterParameters.KV()
+	kv["general.architecture"] = "llama"
+	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
+	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+
+	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+
+	return kv
+}
+
+func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
+	for _, t := range ts {
+		shape := t.Shape()
+		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
+			(strings.HasSuffix(t.Name(), "weight.lora_b") && shape[0] < shape[1]) {
+			shape[0], shape[1] = shape[1], shape[0]
+			t.SetRepacker(p.repackAndTranspose)
+		} else {
+			t.SetRepacker(p.repack)
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    shape,
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *llamaAdapter) Replacements() []string {
+	return []string{
+		"base_model.model.", "",
+		"model.layers", "blk",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"lora_A.weight", "weight.lora_a",
+		"lora_B.weight", "weight.lora_b",
+		"lora_a", "weight.lora_a",
+		"lora_b", "weight.lora_b",
+	}
+}
+
+func (p *llamaAdapter) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	} else {
+		return data, nil
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
+
+func (p *llamaAdapter) repackAndTranspose(name string, data []float32, shape []uint64) ([]float32, error) {
+	dims := []int{int(shape[1]), int(shape[0])}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+
+	var heads uint32
+	if strings.HasSuffix(name, "attn_q.weight.lora_a") {
+		heads = p.NumAttentionHeads
+	} else if strings.HasSuffix(name, "attn_k.weight.lora_a") {
+		heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+	}
+
+	if heads > 0 {
+		if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+			return nil, err
+		}
+
+		if err := n.T(0, 2, 1, 3); err != nil {
+			return nil, err
+		}
+
+		if err := n.Reshape(dims...); err != nil {
+			return nil, err
+		}
+
+		if err := n.Transpose(); err != nil {
+			return nil, err
+		}
+	}
+
+	if err := n.T(1, 0); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}
diff --git a/convert/convert_mixtral.go b/convert/convert_mixtral.go
new file mode 100644
index 000000000..43b7c8b10
--- /dev/null
+++ b/convert/convert_mixtral.go
@@ -0,0 +1,94 @@
+package convert
+
+import (
+	"fmt"
+	"io"
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type mixtralModel struct {
+	llamaModel
+	NumLocalExperts    uint32 `json:"num_local_experts"`
+	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
+}
+
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
+	kv := p.llamaModel.KV(t)
+
+	if p.NumLocalExperts > 0 {
+		kv["llama.expert_count"] = p.NumLocalExperts
+	}
+
+	if p.NumExpertsPerToken > 0 {
+		kv["llama.expert_used_count"] = p.NumExpertsPerToken
+	}
+
+	return kv
+}
+
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
+	oldnew := []string{
+		"model.layers", "blk",
+		"w1", "ffn_gate_exps",
+		"w2", "ffn_down_exps",
+		"w3", "ffn_up_exps",
+	}
+
+	for i := range p.NumLocalExperts {
+		oldnew = append(oldnew, fmt.Sprintf(".block_sparse_moe.experts.%d.", i), ".")
+	}
+
+	// group experts of the same layer (model.layers.%d) and type (w[123]) into a single tensor
+	namer := strings.NewReplacer(oldnew...)
+	experts := make(map[string]experts)
+
+	// merge experts into a single tensor while removing them from ts
+	ts = slices.DeleteFunc(ts, func(t Tensor) bool {
+		if !strings.Contains(t.Name(), ".block_sparse_moe.experts.") {
+			return false
+		}
+
+		name := namer.Replace(t.Name())
+		experts[name] = append(experts[name], t)
+		return true
+	})
+
+	var out []llm.Tensor
+	for n, e := range experts {
+		// TODO(mxyng): sanity check experts
+		out = append(out, llm.Tensor{
+			Name:     n,
+			Kind:     e[0].Kind(),
+			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
+			WriterTo: e,
+		})
+	}
+
+	return append(out, p.llamaModel.Tensors(ts)...)
+}
+
+func (p *mixtralModel) Replacements() []string {
+	return append(
+		p.llamaModel.Replacements(),
+		"block_sparse_moe.gate", "ffn_gate_inp",
+	)
+}
+
+type experts []Tensor
+
+func (e experts) WriteTo(w io.Writer) (int64, error) {
+	// TODO(mxyng): experts _should_ be numerically sorted by expert but this should check
+	for _, t := range e {
+		// the canonical merged experts tensor stacks all experts along a new, 0 axis,
+		// e.g. `tensor.Stack(0, e[0], e[1:]...)`, which requires allocating temporary buffers
+		// this accomplishes the same thing by writing each expert tensor in sequence
+		if _, err := t.WriteTo(w); err != nil {
+			return 0, err
+		}
+	}
+
+	return 0, nil
+}
diff --git a/convert/convert_phi3.go b/convert/convert_phi3.go
new file mode 100644
index 000000000..3de0d4049
--- /dev/null
+++ b/convert/convert_phi3.go
@@ -0,0 +1,123 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/binary"
+	"io"
+	"math"
+	"strings"
+	"sync"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type phi3Model struct {
+	ModelParameters
+	NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+	NLayers           uint32  `json:"n_layers"`
+	HiddenSize        uint32  `json:"hidden_size"`
+	NEmbd             uint32  `json:"n_embd"`
+	IntermediateSize  uint32  `json:"intermediate_size"`
+	NumAttentionHeads uint32  `json:"num_attention_heads"`
+	NHead             uint32  `json:"n_head"`
+	NumKeyValueHeads  uint32  `json:"num_key_value_heads"`
+	NHeadKV           uint32  `json:"n_head_kv"`
+	RopeTheta         float32 `json:"rope_theta"`
+	RopeScaling       struct {
+		Type        string     `json:"type"`
+		LongFactor  ropeFactor `json:"long_factor"`
+		ShortFactor ropeFactor `json:"short_factor"`
+	} `json:"rope_scaling"`
+	RMSNormEPS                    float32 `json:"rms_norm_eps"`
+	NPositions                    uint32  `json:"n_positions"`
+	MaxPositionEmbeddings         uint32  `json:"max_position_embeddings"`
+	OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
+	SlidingWindow                 uint32  `json:"sliding_window"`
+}
+
+var _ ModelConverter = (*phi3Model)(nil)
+
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "phi3"
+	kv["phi3.context_length"] = p.MaxPositionEmbeddings
+	kv["phi3.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
+	kv["phi3.feed_forward_length"] = p.IntermediateSize
+	kv["phi3.block_count"] = cmp.Or(p.NumHiddenLayers, p.NLayers)
+	kv["phi3.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
+	kv["phi3.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NHeadKV)
+	kv["phi3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["phi3.rope.dimension_count"] = p.HiddenSize / cmp.Or(p.NumAttentionHeads, p.NHead)
+	kv["phi3.rope.freq_base"] = p.RopeTheta
+	kv["phi3.rope.scaling.original_context_length"] = p.OriginalMaxPositionEmbeddings
+	kv["phi3.attention.sliding_window"] = p.SlidingWindow
+
+	scale := float64(p.MaxPositionEmbeddings) / float64(p.OriginalMaxPositionEmbeddings)
+
+	switch p.RopeScaling.Type {
+	case "":
+		// no scaling
+	case "su", "longrope":
+		kv["phi3.rope.scaling.attn_factor"] = float32(max(math.Sqrt(1+math.Log(scale)/math.Log(float64(p.OriginalMaxPositionEmbeddings))), 1.0))
+	case "yarn":
+		kv["phi3.rope.scaling.attn_factor"] = float32(max(0.1*math.Log(scale)+1.0, 1.0))
+	default:
+		panic("unknown rope scaling type")
+	}
+
+	return kv
+}
+
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
+	var addRopeFactors sync.Once
+
+	out := make([]llm.Tensor, 0, len(ts)+2)
+	for _, t := range ts {
+		if strings.HasPrefix(t.Name(), "blk.0.") {
+			addRopeFactors.Do(func() {
+				out = append(out, llm.Tensor{
+					Name:     "rope_factors_long.weight",
+					Kind:     0,
+					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
+					WriterTo: p.RopeScaling.LongFactor,
+				}, llm.Tensor{
+					Name:     "rope_factors_short.weight",
+					Kind:     0,
+					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
+					WriterTo: p.RopeScaling.ShortFactor,
+				})
+			})
+		}
+
+		out = append(out, llm.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *phi3Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.qkv_proj", "attn_qkv",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_up_proj", "ffn_up",
+		"post_attention_layernorm", "ffn_norm",
+	}
+}
+
+type ropeFactor []float32
+
+func (r ropeFactor) WriteTo(w io.Writer) (int64, error) {
+	err := binary.Write(w, binary.LittleEndian, r)
+	return 0, err
+}
diff --git a/convert/convert_test.go b/convert/convert_test.go
index 6aa33a49c..2969673d5 100644
--- a/convert/convert_test.go
+++ b/convert/convert_test.go
@@ -1,48 +1,44 @@
-//go:build slow
-
 package convert
 
 import (
+	"bytes"
+	"crypto/sha256"
+	"encoding/binary"
+	"encoding/hex"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"io"
+	"io/fs"
+	"log/slog"
+	"math"
 	"os"
 	"path/filepath"
+	"slices"
+	"strings"
 	"testing"
 
+	"golang.org/x/exp/maps"
+
 	"github.com/ollama/ollama/llm"
 )
 
-func convertFull(t *testing.T, p string) (llm.KV, llm.Tensors) {
+type tensorData struct {
+	Offsets []int  `json:"data_offsets"`
+	Type    string `json:"dtype"`
+	Shape   []int  `json:"shape"`
+}
+
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	t.Helper()
 
-	mf, err := GetModelFormat(p)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	params, err := mf.GetParams(p)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	arch, err := mf.GetModelArch("", p, params)
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	if err := arch.LoadVocab(); err != nil {
-		t.Fatal(err)
-	}
-
-	if err := arch.GetTensors(); err != nil {
-		t.Fatal(err)
-	}
-
 	f, err := os.CreateTemp(t.TempDir(), "f16")
 	if err != nil {
 		t.Fatal(err)
 	}
 	defer f.Close()
 
-	if err := arch.WriteGGUF(f); err != nil {
+	if err := ConvertModel(fsys, f); err != nil {
 		t.Fatal(err)
 	}
 
@@ -50,54 +46,431 @@ func convertFull(t *testing.T, p string) (llm.KV, llm.Tensors) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	defer r.Close()
+	t.Cleanup(func() { r.Close() })
 
-	m, _, err := llm.DecodeGGML(r)
+	m, _, err := llm.DecodeGGML(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
 
-	return m.KV(), m.Tensors()
-}
-
-func TestConvertFull(t *testing.T) {
-	cases := []struct {
-		path    string
-		arch    string
-		tensors int
-		layers  int
-	}{
-		{"Meta-Llama-3-8B-Instruct", "llama", 291, 35},
-		{"Mistral-7B-Instruct-v0.2", "llama", 291, 35},
-		{"Mixtral-8x7B-Instruct-v0.1", "llama", 291, 35},
-		{"gemma-2b-it", "gemma", 164, 20},
+	if _, err := r.Seek(0, io.SeekStart); err != nil {
+		t.Fatal(err)
 	}
 
-	for _, tt := range cases {
-		t.Run(tt.path, func(t *testing.T) {
-			p := filepath.Join("testdata", tt.path)
-			if _, err := os.Stat(p); err != nil {
+	return r, m.KV(), m.Tensors()
+}
+
+func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors llm.Tensors) map[string]string {
+	actual := make(map[string]string)
+	for k, v := range kv {
+		if s, ok := v.(json.Marshaler); !ok {
+			actual[k] = fmt.Sprintf("%v", v)
+		} else {
+			bts, err := json.Marshal(s)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			actual[k] = fmt.Sprintf("%x", sha256.Sum256(bts))
+		}
+	}
+
+	for _, tensor := range tensors.Items {
+		sha256sum := sha256.New()
+		sr := io.NewSectionReader(f, int64(tensors.Offset+tensor.Offset), int64(tensor.Size()))
+		if _, err := io.Copy(sha256sum, sr); err != nil {
+			t.Fatal(err)
+		}
+
+		actual[tensor.Name] = hex.EncodeToString(sha256sum.Sum(nil))
+	}
+
+	return actual
+}
+
+func TestMain(m *testing.M) {
+	var level slog.Level
+	flag.TextVar(&level, "level", slog.LevelInfo, "log level")
+	flag.Parse()
+	slog.SetLogLoggerLevel(level)
+	os.Exit(m.Run())
+}
+
+func TestConvertModel(t *testing.T) {
+	cases := []string{
+		"Meta-Llama-3-8B-Instruct",
+		"Meta-Llama-3.1-8B-Instruct",
+		"Mistral-7B-Instruct-v0.2",
+		"Mixtral-8x7B-Instruct-v0.1",
+		"gemma-2b-it",
+		"gemma-2-2b-it",
+		// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
+		"Phi-3-mini-128k-instruct",
+		"all-MiniLM-L6-v2",
+		"gemma-2-9b-it",
+	}
+
+	for i := range cases {
+		tt := cases[i]
+		t.Run(tt, func(t *testing.T) {
+			t.Parallel()
+
+			p := filepath.Join("testdata", tt)
+			if testing.Short() {
+				t.Skip("skipping in short mode")
+			} else if _, err := os.Stat(p); err != nil {
 				t.Skipf("%s not found", p)
 			}
 
-			kv, tensors := convertFull(t, p)
+			f, kv, tensors := convertFull(t, os.DirFS(p))
+			actual := generateResultsJSON(t, f, kv, tensors)
 
-			if kv.Architecture() != tt.arch {
-				t.Fatalf("expected llama, got %s", kv.Architecture())
+			expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt)))
+			if err != nil {
+				t.Fatal(err)
 			}
 
-			if kv.FileType().String() != "F16" {
-				t.Fatalf("expected F16, got %s", kv.FileType())
+			var expect map[string]string
+			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {
+				t.Fatal(err)
 			}
 
-			if len(tensors) != tt.tensors {
-				t.Fatalf("expected %d tensors, got %d", tt.tensors, len(tensors))
-			}
-
-			layers := tensors.Layers()
-			if len(layers) != tt.layers {
-				t.Fatalf("expected %d layers, got %d", tt.layers, len(layers))
+			keys := maps.Keys(expect)
+			slices.Sort(keys)
+			for _, k := range keys {
+				if v, ok := actual[k]; !ok {
+					t.Errorf("missing %s", k)
+				} else if v != expect[k] {
+					t.Errorf("unexpected %s: want %s, got %s", k, expect[k], v)
+				}
 			}
 		})
 	}
 }
+
+func TestConvertInvalidTensorNames(t *testing.T) {
+	f, err := os.CreateTemp(t.TempDir(), "testmodel")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	tempDir := t.TempDir()
+
+	td := map[string]*tensorData{}
+	offset := 4096
+
+	td["model.layers.0.self_attn.q_proj.weight"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	td["blk.0.attn_q.weight"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	generateSafetensorTestData(t, tempDir, td)
+
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || !strings.HasPrefix(err.Error(), "duplicate tensor name") {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
+func TestConvertInvalidDatatype(t *testing.T) {
+	f, err := os.CreateTemp(t.TempDir(), "testmodel")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	tempDir := t.TempDir()
+
+	td := map[string]*tensorData{}
+	offset := 4096 * 14336
+
+	td["model.layers.0.mlp.down_proj.weight"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "I8",
+		Shape:   []int{4096, 14336},
+	}
+	td["model.layers.0.mlp.down_proj.weight_format"] = &tensorData{
+		Offsets: []int{offset, offset},
+		Type:    "U8",
+		Shape:   []int{},
+	}
+	generateSafetensorTestData(t, tempDir, td)
+
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || err.Error() != "unsupported safetensors model" {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
+func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[string]*tensorData) {
+	data, err := json.Marshal(tensorData)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var buf bytes.Buffer
+
+	l := int64(len(data))
+	err = binary.Write(&buf, binary.LittleEndian, l)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = buf.Write(data)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	fdata, err := os.Create(filepath.Join(tempDir, "model-00001-of-00001.safetensors"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer fdata.Close()
+
+	_, err = fdata.Write(buf.Bytes())
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	configData := `
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ]
+}
+`
+
+	f, err := os.Create(filepath.Join(tempDir, "config.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	_, err = f.WriteString(configData)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	tokenizerData := `
+{
+}
+`
+
+	f, err = os.Create(filepath.Join(tempDir, "tokenizer.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	_, err = f.WriteString(tokenizerData)
+	if err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestConvertAdapter(t *testing.T) {
+	type AdapterCase struct {
+		Name     string
+		BaseKV   map[string]any
+		Expected map[string]string
+	}
+
+	cases := []AdapterCase{
+		{
+			Name: "discollama",
+			BaseKV: map[string]any{
+				"general.architecture":          "llama",
+				"llama.attention.head_count":    uint32(32),
+				"llama.attention.head_count_kv": uint32(8),
+			},
+			Expected: map[string]string{
+				"general.architecture":          "llama",
+				"general.file_type":             "1",
+				"general.parameter_count":       "106496",
+				"general.type":                  "adapter",
+				"general.version":               "v0.2",
+				"adapter.lora.alpha":            "16",
+				"adapter.type":                  "lora",
+				"llama.attention.head_count":    "32",
+				"llama.attention.head_count_kv": "8",
+				"blk.31.attn_q.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_q.weight.lora_b":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_a":   "0eb3318b02cd313429bcc7621b539fdbb10240fea190c56c9e5f93fcd37a4e50",
+				"blk.31.attn_v.weight.lora_b":   "071dcafe89df065d6e1c935ecb8fdf6479b3c202eb912e7da938597673ff5857",
+			},
+		},
+	}
+
+	for _, c := range cases {
+		t.Run(c.Name, func(t *testing.T) {
+			t.Parallel()
+
+			f, err := os.CreateTemp(t.TempDir(), "f16")
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer f.Close()
+
+			tempDir := t.TempDir()
+			generateLoraTestData(t, tempDir)
+
+			if err = ConvertAdapter(os.DirFS(tempDir), f, c.BaseKV); err != nil {
+				t.Fatal(err)
+			}
+
+			r, err := os.Open(f.Name())
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer r.Close()
+
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if _, err := r.Seek(0, io.SeekStart); err != nil {
+				t.Fatal(err)
+			}
+
+			actual := generateResultsJSON(t, r, m.KV(), m.Tensors())
+
+			keys := maps.Keys(c.Expected)
+			slices.Sort(keys)
+			for _, k := range keys {
+				if v, ok := actual[k]; !ok {
+					t.Errorf("missing %s", k)
+				} else if v != c.Expected[k] {
+					t.Errorf("unexpected %s: want %s, got %s", k, c.Expected[k], v)
+				}
+			}
+		})
+	}
+}
+
+func generateLoraTestData(t *testing.T, tempDir string) {
+	offset := 4096 * 8 * 4
+
+	td := map[string]*tensorData{"__metadata__": nil}
+	td["model.layers.31.self_attn.q_proj.lora_a"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.q_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{8, 4096},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_a"] = &tensorData{
+		Offsets: []int{offset * 2, offset * 3},
+		Type:    "F32",
+		Shape:   []int{4096, 8},
+	}
+	td["model.layers.31.self_attn.v_proj.lora_b"] = &tensorData{
+		Offsets: []int{offset * 3, offset*3 + 8*1024*4},
+		Type:    "F32",
+		Shape:   []int{8, 1024},
+	}
+
+	data, err := json.Marshal(td)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var buf bytes.Buffer
+
+	l := int64(len(data))
+	err = binary.Write(&buf, binary.LittleEndian, l)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = buf.Write(data)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// write some data for the tensors
+
+	ones := make([]float32, 4096*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	for range 3 {
+		err = binary.Write(&buf, binary.LittleEndian, ones)
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	ones = make([]float32, 1024*8)
+	for i := range ones {
+		ones[i] = float32(1)
+	}
+
+	err = binary.Write(&buf, binary.LittleEndian, ones)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	fdata, err := os.Create(filepath.Join(tempDir, "adapters.safetensors"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer fdata.Close()
+
+	_, err = fdata.Write(buf.Bytes())
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	configData := `
+{
+    "adapter_path": "adapters-test",
+    "batch_size": 8,
+    "config": "config-tiny.json",
+    "data": "../discollama-completion",
+    "grad_checkpoint": null,
+    "iters": 1000,
+    "learning_rate": 1e-05,
+    "lora_layers": 1,
+    "lora_parameters": {
+        "rank": 8,
+        "alpha": 16,
+        "dropout": 0.0,
+        "scale": 2.0
+    },
+    "lr_schedule": null,
+    "max_seq_length": 2048,
+    "model": "/Users/pdevine/git/Meta-Llama-3-8B-Instruct",
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 200,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "use_dora": false,
+    "val_batches": 25
+}
+`
+	f, err := os.Create(filepath.Join(tempDir, "adapter_config.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	_, err = f.WriteString(configData)
+	if err != nil {
+		t.Fatal(err)
+	}
+}
diff --git a/convert/fs.go b/convert/fs.go
new file mode 100644
index 000000000..31132dbe7
--- /dev/null
+++ b/convert/fs.go
@@ -0,0 +1,58 @@
+package convert
+
+import (
+	"archive/zip"
+	"errors"
+	"io"
+	"io/fs"
+	"os"
+	"path/filepath"
+)
+
+type ZipReader struct {
+	r *zip.Reader
+	p string
+
+	// limit is the maximum size of a file that can be read directly
+	// from the zip archive. Files larger than this size will be extracted
+	limit int64
+}
+
+func NewZipReader(r *zip.Reader, p string, limit int64) fs.FS {
+	return &ZipReader{r, p, limit}
+}
+
+func (z *ZipReader) Open(name string) (fs.File, error) {
+	r, err := z.r.Open(name)
+	if err != nil {
+		return nil, err
+	}
+	defer r.Close()
+
+	if fi, err := r.Stat(); err != nil {
+		return nil, err
+	} else if fi.Size() < z.limit {
+		return r, nil
+	}
+
+	if !filepath.IsLocal(name) {
+		return nil, zip.ErrInsecurePath
+	}
+
+	n := filepath.Join(z.p, name)
+	if _, err := os.Stat(n); errors.Is(err, os.ErrNotExist) {
+		w, err := os.Create(n)
+		if err != nil {
+			return nil, err
+		}
+		defer w.Close()
+
+		if _, err := io.Copy(w, r); err != nil {
+			return nil, err
+		}
+	} else if err != nil {
+		return nil, err
+	}
+
+	return os.Open(n)
+}
diff --git a/convert/gemma.go b/convert/gemma.go
deleted file mode 100644
index d01ffedf1..000000000
--- a/convert/gemma.go
+++ /dev/null
@@ -1,102 +0,0 @@
-package convert
-
-import (
-	"fmt"
-	"io"
-	"log/slog"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type GemmaModel struct {
-	ModelData
-}
-
-func addOnes(data []float32, vectorSize int) ([]float32, error) {
-	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
-	ones := tensor.Ones(tensor.Float32, vectorSize)
-
-	n, err := n.Add(ones)
-	if err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 0)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}
-
-func (m *GemmaModel) GetTensors() error {
-	t, err := m.Format.GetTensors(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-
-	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
-	for _, l := range t {
-		if strings.HasSuffix(l.Name, "norm.weight") {
-			wt := l.WriterTo.(safetensorWriterTo)
-			wt.repacker = m.Repack
-			l.WriterTo = wt
-		}
-		m.Tensors = append(m.Tensors, l)
-	}
-
-	return nil
-}
-
-func (m *GemmaModel) LoadVocab() error {
-	v, err := LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-	m.Vocab = v
-	return nil
-}
-
-func (m *GemmaModel) Repack(_ string, data []float32, shape []uint64) ([]float32, error) {
-	return addOnes(data, int(shape[0]))
-}
-
-func (m *GemmaModel) WriteGGUF(ws io.WriteSeeker) error {
-	kv := llm.KV{
-		"general.architecture":                   "gemma",
-		"general.name":                           m.Name,
-		"gemma.context_length":                   uint32(m.Params.ContextSize),
-		"gemma.embedding_length":                 uint32(m.Params.HiddenSize),
-		"gemma.block_count":                      uint32(m.Params.HiddenLayers),
-		"gemma.feed_forward_length":              uint32(m.Params.IntermediateSize),
-		"gemma.attention.head_count":             uint32(m.Params.AttentionHeads),
-		"gemma.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
-		"gemma.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-		"gemma.attention.key_length":             uint32(m.Params.HeadDimension),
-		"gemma.attention.value_length":           uint32(m.Params.HeadDimension),
-		"general.file_type":                      uint32(1),
-		"tokenizer.ggml.model":                   "llama",
-
-		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
-		"tokenizer.ggml.token_type": m.Vocab.Types,
-
-		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
-		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
-		"tokenizer.ggml.padding_token_id": uint32(m.Params.PaddingTokenID),
-		"tokenizer.ggml.unknown_token_id": uint32(3),
-		"tokenizer.ggml.add_bos_token":    true,
-		"tokenizer.ggml.add_eos_token":    false,
-	}
-
-	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
-}
diff --git a/convert/llama.go b/convert/llama.go
deleted file mode 100644
index b4211b02d..000000000
--- a/convert/llama.go
+++ /dev/null
@@ -1,159 +0,0 @@
-package convert
-
-import (
-	"cmp"
-	"errors"
-	"fmt"
-	"io"
-	"os"
-	"path/filepath"
-	"regexp"
-	"strings"
-
-	"github.com/pdevine/tensor"
-	"github.com/pdevine/tensor/native"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type LlamaModel struct {
-	ModelData
-}
-
-func (m *LlamaModel) GetTensors() error {
-	t, err := m.Format.GetTensors(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-
-	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
-	re, err := regexp.Compile(pattern)
-	if err != nil {
-		return err
-	}
-
-	for _, l := range t {
-		matches := re.FindAllStringSubmatch(l.Name, -1)
-		if len(matches) > 0 {
-			switch m.Format.(type) {
-			case *TorchFormat:
-				wt := l.WriterTo.(torchWriterTo)
-				wt.repacker = m.Repack
-				l.WriterTo = wt
-			case *SafetensorFormat:
-				wt := l.WriterTo.(safetensorWriterTo)
-				wt.repacker = m.Repack
-				l.WriterTo = wt
-			}
-		}
-		m.Tensors = append(m.Tensors, l)
-	}
-
-	return nil
-}
-
-func (m *LlamaModel) LoadVocab() (err error) {
-	pre, ts, merges, err := parseTokens(filepath.Join(m.Path, "tokenizer.json"))
-	if errors.Is(err, os.ErrNotExist) {
-		return nil
-	} else if err != nil {
-		return err
-	}
-
-	m.Vocab = &Vocab{}
-	for _, t := range ts {
-		m.Vocab.Tokens = append(m.Vocab.Tokens, t.Content)
-		m.Vocab.Types = append(m.Vocab.Types, t.Type())
-	}
-
-	m.Vocab.Merges = merges
-	m.Params.PreTokenizer = pre
-	return nil
-}
-
-func (m *LlamaModel) WriteGGUF(ws io.WriteSeeker) error {
-	kv := llm.KV{
-		"general.architecture":                   "llama",
-		"general.name":                           m.Name,
-		"llama.vocab_size":                       uint32(len(m.Vocab.Tokens)),
-		"llama.context_length":                   uint32(m.Params.ContextSize),
-		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
-		"llama.block_count":                      uint32(m.Params.HiddenLayers),
-		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
-		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
-		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
-		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
-		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
-		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-		"general.file_type":                      uint32(1),
-		"tokenizer.ggml.model":                   "gpt2",
-
-		"tokenizer.ggml.pre":        m.Params.PreTokenizer,
-		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.token_type": m.Vocab.Types,
-
-		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
-		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
-		"tokenizer.ggml.unknown_token_id": uint32(0),
-	}
-
-	if len(m.Vocab.Merges) > 0 {
-		kv["tokenizer.ggml.merges"] = m.Vocab.Merges
-	} else {
-		kv["tokenizer.ggml.scores"] = m.Vocab.Scores
-	}
-
-	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
-}
-
-func (m *LlamaModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	return llamaRepack(name, m.Params, data, shape)
-}
-
-func llamaRepack(name string, params *Params, data []float32, shape []uint64) ([]float32, error) {
-	var dims []int
-	for _, dim := range shape {
-		if dim != 0 {
-			dims = append(dims, int(dim))
-		}
-	}
-
-	var heads int
-	switch {
-	case strings.HasSuffix(name, "attn_q.weight"):
-		heads = params.AttentionHeads
-	case strings.HasSuffix(name, "attn_k.weight"):
-		heads = cmp.Or(params.KeyValHeads, params.AttentionHeads)
-	default:
-		return nil, fmt.Errorf("unknown tensor name: %s", name)
-	}
-
-	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
-	if err := n.Reshape(append([]int{heads, 2, dims[0] / heads / 2}, dims[1:]...)...); err != nil {
-		return nil, err
-	}
-
-	if err := n.T(0, 2, 1, 3); err != nil {
-		return nil, err
-	}
-
-	if err := n.Reshape(dims...); err != nil {
-		return nil, err
-	}
-
-	if err := n.Transpose(); err != nil {
-		return nil, err
-	}
-
-	ts, err := native.SelectF32(n, 1)
-	if err != nil {
-		return nil, err
-	}
-
-	var f32s []float32
-	for _, t := range ts {
-		f32s = append(f32s, t...)
-	}
-
-	return f32s, nil
-}
diff --git a/convert/mistral.go b/convert/mistral.go
deleted file mode 100644
index da6874cfd..000000000
--- a/convert/mistral.go
+++ /dev/null
@@ -1,79 +0,0 @@
-package convert
-
-import (
-	"io"
-	"regexp"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type MistralModel struct {
-	ModelData
-}
-
-func (m *MistralModel) GetTensors() error {
-	t, err := m.Format.GetTensors(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-
-	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
-	re, err := regexp.Compile(pattern)
-	if err != nil {
-		return err
-	}
-
-	for _, l := range t {
-		matches := re.FindAllStringSubmatch(l.Name, -1)
-		if len(matches) > 0 {
-			wt := l.WriterTo.(safetensorWriterTo)
-			wt.repacker = m.Repack
-			l.WriterTo = wt
-		}
-		m.Tensors = append(m.Tensors, l)
-	}
-
-	return nil
-}
-
-func (m *MistralModel) LoadVocab() error {
-	v, err := LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-	m.Vocab = v
-	return nil
-}
-
-func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
-	kv := llm.KV{
-		"general.architecture":                   "llama",
-		"general.name":                           m.Name,
-		"llama.context_length":                   uint32(m.Params.ContextSize),
-		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
-		"llama.block_count":                      uint32(m.Params.HiddenLayers),
-		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
-		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
-		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
-		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
-		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-		"general.file_type":                      uint32(1),
-		"tokenizer.ggml.model":                   "llama",
-
-		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
-		"tokenizer.ggml.token_type": m.Vocab.Types,
-
-		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
-		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
-		"tokenizer.ggml.add_bos_token":    true,
-		"tokenizer.ggml.add_eos_token":    false,
-		"tokenizer.ggml.unknown_token_id": uint32(0),
-	}
-
-	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
-}
-
-func (m *MistralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	return llamaRepack(name, m.Params, data, shape)
-}
diff --git a/convert/mixtral.go b/convert/mixtral.go
deleted file mode 100644
index baea68cd3..000000000
--- a/convert/mixtral.go
+++ /dev/null
@@ -1,87 +0,0 @@
-package convert
-
-import (
-	"io"
-	"regexp"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type MixtralModel struct {
-	ModelData
-}
-
-func (m *MixtralModel) GetTensors() error {
-	t, err := m.Format.GetTensors(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-
-	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
-	re, err := regexp.Compile(pattern)
-	if err != nil {
-		return err
-	}
-
-	for _, l := range t {
-		matches := re.FindAllStringSubmatch(l.Name, -1)
-		if len(matches) > 0 {
-			wt := l.WriterTo.(safetensorWriterTo)
-			wt.repacker = m.Repack
-			l.WriterTo = wt
-		}
-		m.Tensors = append(m.Tensors, l)
-	}
-
-	return nil
-}
-
-func (m *MixtralModel) LoadVocab() error {
-	v, err := LoadSentencePieceTokens(m.Path, m.Params)
-	if err != nil {
-		return err
-	}
-	m.Vocab = v
-	return nil
-}
-
-func (m *MixtralModel) WriteGGUF(ws io.WriteSeeker) error {
-	kv := llm.KV{
-		"general.architecture":          "llama",
-		"general.name":                  m.Name,
-		"llama.block_count":             uint32(m.Params.HiddenLayers),
-		"llama.context_length":          uint32(m.Params.ContextSize),
-		"llama.embedding_length":        uint32(m.Params.HiddenSize),
-		"llama.feed_forward_length":     uint32(m.Params.IntermediateSize),
-		"llama.attention.head_count":    uint32(m.Params.AttentionHeads),
-		"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
-
-		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
-		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
-
-		"llama.expert_count":      uint32(m.Params.Experts),
-		"llama.expert_used_count": uint32(m.Params.ExpertsUsed),
-
-		"llama.vocab_size":           uint32(len(m.Vocab.Tokens)),
-		"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
-
-		"general.file_type":    uint32(1),
-		"tokenizer.ggml.model": "llama",
-
-		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
-		"tokenizer.ggml.scores":     m.Vocab.Scores,
-		"tokenizer.ggml.token_type": m.Vocab.Types,
-
-		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
-		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
-		"tokenizer.ggml.unknown_token_id": uint32(0),
-		"tokenizer.ggml.add_bos_token":    true,
-		"tokenizer.ggml.add_eos_token":    false,
-	}
-
-	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
-}
-
-func (m *MixtralModel) Repack(name string, data []float32, shape []uint64) ([]float32, error) {
-	return llamaRepack(name, m.Params, data, shape)
-}
diff --git a/convert/reader.go b/convert/reader.go
new file mode 100644
index 000000000..c1218e66d
--- /dev/null
+++ b/convert/reader.go
@@ -0,0 +1,86 @@
+package convert
+
+import (
+	"errors"
+	"io"
+	"io/fs"
+	"strings"
+)
+
+type Tensor interface {
+	Name() string
+	Shape() []uint64
+	Kind() uint32
+	SetRepacker(repacker)
+	WriteTo(io.Writer) (int64, error)
+}
+
+type tensorBase struct {
+	name  string
+	shape []uint64
+	repacker
+}
+
+func (t tensorBase) Name() string {
+	return t.name
+}
+
+func (t tensorBase) Shape() []uint64 {
+	return t.shape
+}
+
+const (
+	tensorKindF32 uint32 = iota
+	tensorKindF16
+)
+
+func (t tensorBase) Kind() uint32 {
+	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
+		t.name == "token_types.weight" {
+		// these tensors are always F32
+		return 0
+	}
+
+	switch len(t.shape) {
+	case 0:
+		panic("invalid tensor shape")
+	case 1:
+		return tensorKindF32
+	default:
+		return tensorKindF16
+	}
+}
+
+func (t *tensorBase) SetRepacker(fn repacker) {
+	t.repacker = fn
+}
+
+type repacker func(string, []float32, []uint64) ([]float32, error)
+
+func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
+	patterns := []struct {
+		Pattern string
+		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
+	}{
+		{"model-*-of-*.safetensors", parseSafetensors},
+		{"model.safetensors", parseSafetensors},
+		{"adapters.safetensors", parseSafetensors},
+		{"adapter_model.safetensors", parseSafetensors},
+		{"pytorch_model-*-of-*.bin", parseTorch},
+		{"pytorch_model.bin", parseTorch},
+		{"consolidated.*.pth", parseTorch},
+	}
+
+	for _, pattern := range patterns {
+		matches, err := fs.Glob(fsys, pattern.Pattern)
+		if err != nil {
+			return nil, err
+		}
+
+		if len(matches) > 0 {
+			return pattern.Func(fsys, replacer, matches...)
+		}
+	}
+
+	return nil, errors.New("unknown tensor format")
+}
diff --git a/convert/reader_safetensors.go b/convert/reader_safetensors.go
new file mode 100644
index 000000000..b21d219c2
--- /dev/null
+++ b/convert/reader_safetensors.go
@@ -0,0 +1,163 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"slices"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/x448/float16"
+	"golang.org/x/exp/maps"
+)
+
+type safetensorMetadata struct {
+	Type    string   `json:"dtype"`
+	Shape   []uint64 `json:"shape"`
+	Offsets []int64  `json:"data_offsets"`
+}
+
+func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
+	var ts []Tensor
+	for _, p := range ps {
+		f, err := fsys.Open(p)
+		if err != nil {
+			return nil, err
+		}
+		defer f.Close()
+
+		var n int64
+		if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
+			return nil, err
+		}
+
+		b := bytes.NewBuffer(make([]byte, 0, n))
+		if _, err = io.CopyN(b, f, n); err != nil {
+			return nil, err
+		}
+
+		var headers map[string]safetensorMetadata
+		if err := json.NewDecoder(b).Decode(&headers); err != nil {
+			return nil, err
+		}
+
+		keys := maps.Keys(headers)
+		slices.Sort(keys)
+
+		names := make(map[string]struct{}, len(keys))
+
+		for _, key := range keys {
+			if value := headers[key]; value.Type != "" {
+				// bitsandbytes quantized models are unsupported
+				if len(value.Shape) == 0 {
+					return nil, errors.New("unsupported safetensors model")
+				}
+				ggufName := replacer.Replace(key)
+				if _, ok := names[ggufName]; ok {
+					return nil, fmt.Errorf("duplicate tensor name '%s' was found for this model", ggufName)
+				}
+				names[ggufName] = struct{}{}
+				ts = append(ts, safetensor{
+					fs:     fsys,
+					path:   p,
+					dtype:  value.Type,
+					offset: safetensorsPad(n, value.Offsets[0]),
+					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
+					tensorBase: &tensorBase{
+						name:  ggufName,
+						shape: value.Shape,
+					},
+				})
+			}
+		}
+	}
+
+	return ts, nil
+}
+
+// safetensorsPad returns the padded size of the safetensors file given a length n and offset s
+func safetensorsPad(n, offset int64) int64 {
+	return 8 + n + offset
+}
+
+type safetensor struct {
+	fs     fs.FS
+	path   string
+	dtype  string
+	offset int64
+	size   int64
+	*tensorBase
+}
+
+func (st safetensor) WriteTo(w io.Writer) (int64, error) {
+	f, err := st.fs.Open(st.path)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+
+	if seeker, ok := f.(io.Seeker); ok {
+		if _, err := seeker.Seek(st.offset, io.SeekStart); err != nil {
+			return 0, err
+		}
+	} else {
+		if _, err := io.CopyN(io.Discard, f, st.offset); err != nil {
+			return 0, err
+		}
+	}
+
+	var f32s []float32
+	switch st.dtype {
+	case "F32":
+		f32s = make([]float32, st.size/4)
+		if err = binary.Read(f, binary.LittleEndian, f32s); err != nil {
+			return 0, err
+		}
+	case "F16":
+		u16s := make([]uint16, st.size/2)
+		if err = binary.Read(f, binary.LittleEndian, u16s); err != nil {
+			return 0, err
+		}
+
+		f32s = make([]float32, len(u16s))
+		for i := range u16s {
+			f32s[i] = float16.Frombits(u16s[i]).Float32()
+		}
+
+	case "BF16":
+		u8s := make([]uint8, st.size)
+		if err = binary.Read(f, binary.LittleEndian, u8s); err != nil {
+			return 0, err
+		}
+
+		f32s = bfloat16.DecodeFloat32(u8s)
+	default:
+		return 0, fmt.Errorf("unknown data type: %s", st.dtype)
+	}
+
+	if st.repacker != nil {
+		f32s, err = st.repacker(st.Name(), f32s, st.Shape())
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	switch st.Kind() {
+	case tensorKindF32:
+		return 0, binary.Write(w, binary.LittleEndian, f32s)
+	case tensorKindF16:
+		f16s := make([]uint16, len(f32s))
+		for i := range f32s {
+			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
+		}
+
+		return 0, binary.Write(w, binary.LittleEndian, f16s)
+	default:
+		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
+	}
+}
diff --git a/convert/reader_torch.go b/convert/reader_torch.go
new file mode 100644
index 000000000..1b3e1c9f1
--- /dev/null
+++ b/convert/reader_torch.go
@@ -0,0 +1,48 @@
+package convert
+
+import (
+	"io"
+	"io/fs"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/nlpodyssey/gopickle/types"
+)
+
+func parseTorch(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]Tensor, error) {
+	var ts []Tensor
+	for _, p := range ps {
+		pt, err := pytorch.Load(p)
+		if err != nil {
+			return nil, err
+		}
+
+		for _, k := range pt.(*types.Dict).Keys() {
+			t := pt.(*types.Dict).MustGet(k)
+
+			var shape []uint64
+			for dim := range t.(*pytorch.Tensor).Size {
+				shape = append(shape, uint64(dim))
+			}
+
+			ts = append(ts, torch{
+				storage: t.(*pytorch.Tensor).Source,
+				tensorBase: &tensorBase{
+					name:  replacer.Replace(k.(string)),
+					shape: shape,
+				},
+			})
+		}
+	}
+
+	return ts, nil
+}
+
+type torch struct {
+	storage pytorch.StorageInterface
+	*tensorBase
+}
+
+func (pt torch) WriteTo(w io.Writer) (int64, error) {
+	return 0, nil
+}
diff --git a/convert/safetensors.go b/convert/safetensors.go
deleted file mode 100644
index f45687f10..000000000
--- a/convert/safetensors.go
+++ /dev/null
@@ -1,309 +0,0 @@
-package convert
-
-import (
-	"bytes"
-	"encoding/binary"
-	"encoding/json"
-	"fmt"
-	"io"
-	"os"
-	"path/filepath"
-	"regexp"
-	"slices"
-	"strings"
-
-	"github.com/d4l3k/go-bfloat16"
-	"github.com/x448/float16"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type safetensorWriterTo struct {
-	t *llm.Tensor
-
-	params *Params
-	bo     ByteOrder
-
-	filename string
-	dtype    string
-
-	offset, size int64
-	repacker     func(string, []float32, []uint64) ([]float32, error)
-}
-
-type safetensorMetadata struct {
-	Type    string   `json:"dtype"`
-	Shape   []uint64 `json:"shape"`
-	Offsets []int64  `json:"data_offsets"`
-}
-
-type SafetensorFormat struct{}
-
-func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
-	var tensors []llm.Tensor
-	matches, err := filepath.Glob(filepath.Join(dirpath, "*.safetensors"))
-	if err != nil {
-		return nil, err
-	}
-
-	var offset uint64
-	for _, f := range matches {
-		var t []llm.Tensor
-		var err error
-		t, offset, err = m.readTensors(f, offset, params)
-		if err != nil {
-			return nil, err
-		}
-
-		tensors = append(tensors, t...)
-	}
-	return tensors, nil
-}
-
-func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params) ([]llm.Tensor, uint64, error) {
-	f, err := os.Open(fn)
-	if err != nil {
-		return nil, 0, err
-	}
-	defer f.Close()
-
-	var n int64
-	if err := binary.Read(f, binary.LittleEndian, &n); err != nil {
-		return nil, 0, err
-	}
-
-	b := bytes.NewBuffer(make([]byte, 0, n))
-	if _, err = io.CopyN(b, f, n); err != nil {
-		return nil, 0, err
-	}
-
-	var headers map[string]safetensorMetadata
-	if err := json.NewDecoder(b).Decode(&headers); err != nil {
-		return nil, 0, err
-	}
-
-	var keys []string
-	for key := range headers {
-		if !strings.HasSuffix(key, "self_attn.rotary_embd.inv_freq") {
-			keys = append(keys, key)
-		}
-	}
-
-	slices.Sort(keys)
-
-	var tensors []llm.Tensor
-	for _, key := range keys {
-		value := headers[key]
-
-		var kind uint32
-		switch len(value.Shape) {
-		case 0:
-			// valuedata
-			continue
-		case 2:
-			kind = 1
-		}
-
-		name, err := m.GetLayerName(key)
-		if err != nil {
-			return nil, 0, err
-		}
-
-		shape := make([]uint64, len(value.Shape))
-		copy(shape, value.Shape)
-
-		pad := func(s int64) int64 {
-			return 8 + n + s
-		}
-
-		t := llm.Tensor{
-			Name:   name,
-			Kind:   kind,
-			Offset: offset,
-			Shape:  shape,
-		}
-
-		t.WriterTo = safetensorWriterTo{
-			t:        &t,
-			params:   params,
-			bo:       params.ByteOrder,
-			filename: fn,
-			dtype:    value.Type,
-			offset:   pad(value.Offsets[0]),
-			size:     pad(value.Offsets[1]) - pad(value.Offsets[0]),
-		}
-
-		offset += t.Size()
-		tensors = append(tensors, t)
-	}
-
-	return tensors, offset, nil
-}
-
-func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
-	f, err := os.Open(filepath.Join(dirpath, "config.json"))
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
-
-	var params Params
-
-	if err := json.NewDecoder(f).Decode(&params); err != nil {
-		return nil, err
-	}
-
-	params.ByteOrder = binary.LittleEndian
-	return &params, nil
-}
-
-func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
-	directMap := map[string]string{
-		"model.embed_tokens.weight": "token_embd.weight",
-		"lm_head.weight":            "output.weight",
-		"model.norm.weight":         "output_norm.weight",
-	}
-
-	tMap := map[string]string{
-		"model.layers.(\\d+).input_layernorm.weight":                    "blk.$1.attn_norm.weight",
-		"model.layers.(\\d+).mlp.down_proj.weight":                      "blk.$1.ffn_down.weight",
-		"model.layers.(\\d+).mlp.gate_proj.weight":                      "blk.$1.ffn_gate.weight",
-		"model.layers.(\\d+).mlp.up_proj.weight":                        "blk.$1.ffn_up.weight",
-		"model.layers.(\\d+).post_attention_layernorm.weight":           "blk.$1.ffn_norm.weight",
-		"model.layers.(\\d+).self_attn.k_proj.weight":                   "blk.$1.attn_k.weight",
-		"model.layers.(\\d+).self_attn.o_proj.weight":                   "blk.$1.attn_output.weight",
-		"model.layers.(\\d+).self_attn.q_proj.weight":                   "blk.$1.attn_q.weight",
-		"model.layers.(\\d+).self_attn.v_proj.weight":                   "blk.$1.attn_v.weight",
-		"model.layers.(\\d+).block_sparse_moe.gate.weight":              "blk.$1.ffn_gate_inp.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
-		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
-	}
-
-	v, ok := directMap[n]
-	if ok {
-		return v, nil
-	}
-
-	// quick hack to rename the layers to gguf format
-	for k, v := range tMap {
-		re := regexp.MustCompile(k)
-		newName := re.ReplaceAllString(n, v)
-		if newName != n {
-			return newName, nil
-		}
-	}
-
-	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
-}
-
-func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
-	f, err := os.Open(r.filename)
-	if err != nil {
-		return 0, err
-	}
-	defer f.Close()
-
-	if _, err = f.Seek(r.offset, io.SeekStart); err != nil {
-		return 0, err
-	}
-
-	var f32s []float32
-	switch r.dtype {
-	case "F32":
-		f32s = make([]float32, r.size/4)
-		if err = binary.Read(f, r.bo, f32s); err != nil {
-			return 0, err
-		}
-	case "F16":
-		u16s := make([]uint16, r.size/2)
-		if err = binary.Read(f, r.bo, u16s); err != nil {
-			return 0, err
-		}
-
-		for _, b := range u16s {
-			f32s = append(f32s, float16.Frombits(b).Float32())
-		}
-
-	case "BF16":
-		u8s := make([]uint8, r.size)
-		if err = binary.Read(f, r.bo, u8s); err != nil {
-			return 0, err
-		}
-
-		f32s = bfloat16.DecodeFloat32(u8s)
-	default:
-		return 0, fmt.Errorf("unknown data type: %s", r.dtype)
-	}
-
-	if r.repacker != nil {
-		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
-		if err != nil {
-			return 0, err
-		}
-	}
-
-	switch r.t.Kind {
-	case 0:
-		return 0, binary.Write(w, r.bo, f32s)
-	case 1:
-		f16s := make([]uint16, len(f32s))
-		for i := range f32s {
-			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
-		}
-
-		return 0, binary.Write(w, r.bo, f16s)
-	default:
-		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
-	}
-}
-
-func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
-	switch len(params.Architectures) {
-	case 0:
-		return nil, fmt.Errorf("No architecture specified to convert")
-	case 1:
-		switch params.Architectures[0] {
-		case "LlamaForCausalLM":
-			return &LlamaModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
-		case "MistralForCausalLM":
-			return &MistralModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
-		case "MixtralForCausalLM":
-			return &MixtralModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
-		case "GemmaForCausalLM":
-			return &GemmaModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
-		default:
-			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
-		}
-	}
-
-	return nil, fmt.Errorf("Unknown error")
-}
diff --git a/convert/testdata/Meta-Llama-3-8B-Instruct.json b/convert/testdata/Meta-Llama-3-8B-Instruct.json
new file mode 100644
index 000000000..808826bb6
--- /dev/null
+++ b/convert/testdata/Meta-Llama-3-8B-Instruct.json
@@ -0,0 +1,313 @@
+{
+  "general.architecture": "llama",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "llama.block_count": "32",
+  "llama.context_length": "8192",
+  "llama.embedding_length": "4096",
+  "llama.feed_forward_length": "14336",
+  "llama.rope.dimension_count": "128",
+  "llama.rope.freq_base": "500000",
+  "llama.vocab_size": "128256",
+  "llama.attention.head_count": "32",
+  "llama.attention.head_count_kv": "8",
+  "llama.attention.layer_norm_rms_epsilon": "1e-05",
+  "tokenizer.ggml.model": "gpt2",
+  "tokenizer.ggml.pre": "llama-bpe",
+  "tokenizer.ggml.bos_token_id": "128000",
+  "tokenizer.ggml.eos_token_id": "128009",
+  "tokenizer.ggml.merges": "d0cbac1fcc9dcf03724b8db5c9bfb593ae1cf68fb9bc72eb1d15274dcbbf618b",
+  "tokenizer.ggml.token_type": "d70a88809fd7da6f1f028622685cd64268a7a922c5d343c96f25b66327358978",
+  "tokenizer.ggml.tokens": "765b529dbcbc42dd202ce657341c63807b51f3b07e09898f6aa6196326865d5a",
+  "token_embd.weight": "b53102a11d9064bbd404833e3464b1b13e08ce73300b442312cccde2f19b2698",
+  "blk.0.attn_norm.weight": "7318df3cca9e8d153ff0a503026a1265e63d20b2a8c1dd7a2769585082b5d1ee",
+  "blk.0.ffn_down.weight": "b950806a1fc722c9fad7fd0b20c3c0a7fb50f14395e1e7663a590bfd62e20900",
+  "blk.0.ffn_gate.weight": "e73e580af6d4f08e060a74a3c25efdf5d3bed99e183d95a5a85ae859014839fd",
+  "blk.0.ffn_up.weight": "c8158af679ef99746da1befb67eebb19489e0bbe6ce7d97e13e348508244e516",
+  "blk.0.ffn_norm.weight": "7ec69c3c31e95e49a3359003b0033f6b9e85561a3e3fd83e7476661ecdd756bb",
+  "blk.0.attn_k.weight": "2732303257bac969b4964e0e32ec08b5a7f5c031bb02bf6ac4467b3ea0ebcf1e",
+  "blk.0.attn_output.weight": "ecda1d43b4ccc91cd5b366d7e7a275353990ac78561a07c83d9c77031aba12dc",
+  "blk.0.attn_q.weight": "569b1f5faf92b6f00910cf7effb2d5862f91038ce5c3b0019fc10e5d79fbd5e1",
+  "blk.0.attn_v.weight": "aa8416c5ef7e32fb54a1f20d6ac651656845d4af240564b397c39bd83e06e3b8",
+  "blk.1.attn_norm.weight": "03327e02862908c2a44b2f52decdb924bf4201f400b46f8037a9cb2e1d7a61ff",
+  "blk.1.ffn_down.weight": "5a83a87603f38c99f8e1e370a2d5f967bb45ac51d881a609304a7811027321e0",
+  "blk.1.ffn_gate.weight": "31da0572c79e655186c721c231376f85e56cdcc6257c28d08c8c5b40d5c22b40",
+  "blk.1.ffn_up.weight": "e0c811d64ca155c8de10a868e72015d43888834804614ee1aa2953129ffbc90f",
+  "blk.1.ffn_norm.weight": "5861f313d6137d6f0f904d423df47fffc6069e224ff746e1b637ac9c7f0af862",
+  "blk.1.attn_k.weight": "5fbbec0acca6457b9416ebdcd90e526885d0224537b7628f6be376a7f275313d",
+  "blk.1.attn_output.weight": "b237c9763fa3f75166a6f70b70f1566e77d0d89dfa164ed1b3137393e90575c3",
+  "blk.1.attn_q.weight": "c0a9cf4a98b4882b16f3eb2b49d933793dcc5357abb246fd3fe3134ed2b12e1c",
+  "blk.1.attn_v.weight": "96867111727200cac1af7865189dd41fd62b47584e5e5f33a91f1d34509cbd40",
+  "blk.2.attn_norm.weight": "f392f8a88ee3a95b1cc19c40dd4ef66317037b0faaa1800f610779e129ee0539",
+  "blk.2.ffn_down.weight": "73823eef46632aedcc8c1cb08a736b6aa97ca97842cd1fdfc5567d8dec459662",
+  "blk.2.ffn_gate.weight": "f4909ae19fc3848b00bb8b9050122e74f8e903b89e22937036f4cc9fea20a718",
+  "blk.2.ffn_up.weight": "16f4904a3d814ea68f00519724fc4943e48444a84c786bda39aa5efc298a7d84",
+  "blk.2.ffn_norm.weight": "e3ccdf56e75cb969f6f69c39caf6daf7c4e70e89e25df0f4d2e4bc60e159aafe",
+  "blk.2.attn_k.weight": "c3beb1e0a11bcf007ef0f0d8f6bdd3082d8b29090cd29597846b5d51e308a8e5",
+  "blk.2.attn_output.weight": "bb9f66c32cff51154fea92933c2cd62549236f8cb1a767f9ef28d3f99809b343",
+  "blk.2.attn_q.weight": "8eba394132eef2a05c5a92d62d2376000f7948448d7a2dc74e6b608203add20d",
+  "blk.2.attn_v.weight": "88f61f77c53567c617db3eef8f30621109a750e679f6784f7911739bd42c2f02",
+  "blk.3.attn_norm.weight": "7b996675b7ca75fa24107b3ebe0788653ede0f49ac83b8659d71ff54d591f81a",
+  "blk.3.ffn_down.weight": "2cb332bc05e4821962fdc9dcbcc7cc12630f32117711b687d18fb53c0bc4fbf4",
+  "blk.3.ffn_gate.weight": "340b387c7f208c8f0a6db904ef8d87c1e84b7d6ad57177abd32d86c8d18b760f",
+  "blk.3.ffn_up.weight": "07484433f8a7ee061c55aa0de2ecc009f769b0617c9c0ec096e9bb2946df9f0e",
+  "blk.3.ffn_norm.weight": "4f1a4ade36b393af341240bc894a2aab09cff7e4d56dc4658445deb107f9371b",
+  "blk.3.attn_k.weight": "483dcd96acb4528df84b9842970994630dbd82b8715ace394aa8b39fcf8d6291",
+  "blk.3.attn_output.weight": "beaff0810687923585642ee11d929cbf3b43dc6f87f30ddb552c222ab57bdbb3",
+  "blk.3.attn_q.weight": "0739355002f6fce520863add697e0ff25fc88215322dc3f993be7bb68dcce7e8",
+  "blk.3.attn_v.weight": "c216d17b6d90ee3e07f82598b8161fae34de2f392dbb0f745b682b578c324767",
+  "blk.4.attn_norm.weight": "91ab405bc4ba15bf63af233f266aa43aaab43789a9e6596e14a357c2ac7df217",
+  "blk.4.ffn_down.weight": "620f34ee75cdc73aecb8949af5fbb0d2437fd81422b6d8eb7acfc52addb9fc68",
+  "blk.4.ffn_gate.weight": "f6feec7bc9acadf35ec22532f8998d8e50f31afedabb19263590dcf8b9a92eee",
+  "blk.4.ffn_up.weight": "4a72af7cd28fd07b038f6cc4406678d120517280236ea85d9e76eff40ab2cc22",
+  "blk.4.ffn_norm.weight": "1805b37b44d5d682bdbd2fadeafb763ee001617d7870848cc487079ee34b21f9",
+  "blk.4.attn_k.weight": "a1e4f9d97cdf4c1b0d177cf00c4e32d1be30c1984a239b3c9bd73f8848888853",
+  "blk.4.attn_output.weight": "a1547e2497c423b0aff0eee71d9300d6fdf4e4986679418b6e637b69a9a6720b",
+  "blk.4.attn_q.weight": "0677483a9264ea6803d03d304d87a54632242cb516e8b76b6e3e8284c2f4de04",
+  "blk.4.attn_v.weight": "02691ba3af344fcc1969428ab0df811ac94aaa2fd91b0dc4ec1ac0a58806980d",
+  "blk.5.attn_norm.weight": "ba9c028335e5c895b87a5bd1448ca429248f9746ed97bdcb8679923206117156",
+  "blk.5.ffn_down.weight": "ccfdc9006acad1940a6bc05042a3947f1066acd671e0bb53b7684e9eea9ef5c9",
+  "blk.5.ffn_gate.weight": "623157679f1e742ccc3807c0b0153ddc8450104de75ec62f1370ec3807c09cf4",
+  "blk.5.ffn_up.weight": "05748804c65091f963729b58b085f58351891cac8a2861f5eae26b06aa60b2a0",
+  "blk.5.ffn_norm.weight": "84bae55af2efc8b8429f09056c8c04990c466dae31cb3f9356038b8957f1b406",
+  "blk.5.attn_k.weight": "8c766180c726b037d587fc52371de6e3307140c52409011609d1225624b6a3eb",
+  "blk.5.attn_output.weight": "490b582b3b1dc151ae55aee8b6743dad6c01fb49e43afefb6e68394b74be3d73",
+  "blk.5.attn_q.weight": "6f7b8ca4d9025ec836a44bbcca46be30c66b471a9fb62943ddff8288b3731409",
+  "blk.5.attn_v.weight": "9f70df3ba00c9e723214b3da83ff435a2163fff5915f75515c9664c05c866c27",
+  "blk.6.attn_norm.weight": "1a4a66613a682df6f061fc7c4d986f9f7e9175b62f0c42fc1ef31db536bd5942",
+  "blk.6.ffn_down.weight": "c56f25e4e49b443dbc82d88311ee63bc1f5002cc67e52f4787fd5f003aedeac1",
+  "blk.6.ffn_gate.weight": "31a5cf1aa9b831a81588d508550f51fc425f9517c43254d4ef7096d38029cf04",
+  "blk.6.ffn_up.weight": "ce135f3a1163e0c9297a615bdbe68a67ead21edce8debbfa9f6e15e6af8d4c94",
+  "blk.6.ffn_norm.weight": "4e328ce0648c94e732bc40501858ef6262ad1161e2e407b0cdcf4813fa9d45d8",
+  "blk.6.attn_k.weight": "1eb1c4c9f9c4c7ff7f5429075e0dc6a7782bed55109fa88df209a817dd8ef960",
+  "blk.6.attn_output.weight": "3d32986b56873b88655ee1edabdd413fdd9ab18b82108c9ce90bdbc2d3a6f3a3",
+  "blk.6.attn_q.weight": "8432f583b3a2809c99c393f9beb077cb0534dd5d247c17108f2986cadc6651f6",
+  "blk.6.attn_v.weight": "5045381513815bb91839dbac8335ffe49bbc7b0008369de7ea97eb676c5e2b36",
+  "blk.7.attn_norm.weight": "3dabd003638ec2499bfc8a48c49eef34276caab4fe76894eb963207848c2fdaf",
+  "blk.7.ffn_down.weight": "194fae858608bdcffd235be59ab119d0b91c8549f864ea06dae69249e099935f",
+  "blk.7.ffn_gate.weight": "00b24c29c30246892bce0791be804a89701d4c1332777e0bcdad5d9d5666604f",
+  "blk.7.ffn_up.weight": "44d7082a5280080c90cef9e19d410391de34f212ca0736377769b8ddd0c82d5e",
+  "blk.7.ffn_norm.weight": "21fe8a7fd6911c64e0d15a788b3b4cb6d71dd6ec51de65f760ee89afbb6ae53e",
+  "blk.7.attn_k.weight": "57a149eec5f6744a9526cd3925ac073f9d12db0fbcb5afe042ef4dc846458c44",
+  "blk.7.attn_output.weight": "0e9c28a3e81a2880251ce5eed77bcb8be8aaa1a51c9cb6de820b47ed83849fc2",
+  "blk.7.attn_q.weight": "15ee75263ee4e2a43eb322bc159ae004bb7d77e3a7e63ee4ddab700430693fff",
+  "blk.7.attn_v.weight": "440aa970bba4bff429fd7b7b1de21f2ad14fb2952b776cfa4acee68d7c6e9b8f",
+  "blk.8.attn_norm.weight": "af5b44825633c42c1ae964c82bb2be6a242d3a751f0a91f1bae4f593e8f5b6ec",
+  "blk.8.ffn_down.weight": "b11c14c76adca94fa200496dd2c10743becb23aab6642443ef1ae6d8710edbc1",
+  "blk.8.ffn_gate.weight": "7bb03d3325bf8637ae2fa1296b0651356515578d46a7c5ca65c7a923d7de27bc",
+  "blk.8.ffn_up.weight": "b956ef0a0669b5a9c9bf3a8da2d1c24f52d331cfb7354f6d7c51bd65be355e30",
+  "blk.8.ffn_norm.weight": "c78c3d748302edfef76f71ea5cb2055c94352122eee8b9b1173779a1814d224e",
+  "blk.8.attn_k.weight": "c0fba6a596ed9c1c32a7055c31a935a8b31e42b77282ee47c1f03ee3bde736b5",
+  "blk.8.attn_output.weight": "83cf9947080c5d8d571f04a842bc3dcfe7bbb0195fb25b346e22635e8649f2d4",
+  "blk.8.attn_q.weight": "47409350a576b333d97b7c877d69f47f46df504f3765102dfc0be9e521c7ecd6",
+  "blk.8.attn_v.weight": "1999dff91404fdcf1ecb34d9eaaaa9244ec7658a74dec8feb7cfd1fddba0347e",
+  "blk.9.attn_norm.weight": "1e6e29d5c3889ab4e1b0a5b9998cba60179b0f1fca133515df49cbc19d092593",
+  "blk.9.ffn_down.weight": "acb898a6490adff592e10b4c62d70edc5941661ee6da44658500e9205357c8e9",
+  "blk.9.ffn_gate.weight": "4cff63013593aadc3ffbaaa6ed70ffdba1224cd43c3644bf6f4162b5ac1ab542",
+  "blk.9.ffn_up.weight": "f985b5a2d6cf4fe32c7256301c3c89b8ad22b59e516342c52da42d8110766a4e",
+  "blk.9.ffn_norm.weight": "0d659c538bc6b21ed0018f107ab674a7424a00a42946c80e07208b479b21918f",
+  "blk.9.attn_k.weight": "f67611d888780d1b38c1c146b361c65310c8183bdf64fd73e2259985c6e8517f",
+  "blk.9.attn_output.weight": "f12ca1fa62a02ddc3f77f798bfb5707e0c50bf18ee0eaa67025521a98355f26b",
+  "blk.9.attn_q.weight": "3865185f4361a645b086ad47b72904c095313fb1c624e511647bf1a7dfc1c476",
+  "blk.9.attn_v.weight": "92125bbfed63544ab56052bd1e4aa453bbf34c795249ee54cde54907c8c6d1d3",
+  "blk.10.attn_norm.weight": "5d6bfbe545bcc2fcb2fc75c68f64b1f4c918badaf53e0156fe2d88aa977b2f94",
+  "blk.10.ffn_down.weight": "1dd9da8b0d2696ab5531fbca8a29c7d67567620a9d3e5fc2a19ec5d7e4c6cc8a",
+  "blk.10.ffn_gate.weight": "6e55e7f014edaebda0ac6819a426221d3b025c27312a2e18cc5806f31e3db226",
+  "blk.10.ffn_up.weight": "d80dde54af5db51241345ee8d64c1972608644f4deeac1e8195dc423bf27474a",
+  "blk.10.ffn_norm.weight": "f6ca65951d58ae3379eee8247bec34ebd0db05674cc9295593573841b8a55df3",
+  "blk.10.attn_k.weight": "b58e350bd6b49aba0fba4e4dd6865de3a2a0651ab865dbf2419b627b53ffc187",
+  "blk.10.attn_output.weight": "6b26a986e12fe66ec286a21d7d5af5eaa1bfe6f2bf502165d270e4497235a54a",
+  "blk.10.attn_q.weight": "3440e0e5b7e0d1e426424ae5a33f4e057be623249e9035ea12e57dbe5d3893c4",
+  "blk.10.attn_v.weight": "ebfadcfe14bcd6dee933053df0a67e12e7a196d5cc45728c1ffb2a2daedd5ca2",
+  "blk.11.attn_norm.weight": "3ed057b9576cd2de84507ef64c7646dc478c651efca4c2024cbe91a4f3fbf0bc",
+  "blk.11.ffn_down.weight": "8ff1c2487d22f5c499761e4eb721418f141f960160d0bab779595a34e4d68898",
+  "blk.11.ffn_gate.weight": "9c74e4507c7e45bf39b7cc7402198cd1dd77e3fff8c625b0413acaeb16efeb9f",
+  "blk.11.ffn_up.weight": "4367158007161d29939e00a322bb6776016e43f648a94f9b08a96a477aae75be",
+  "blk.11.ffn_norm.weight": "1cc0288c1491072121f4c9a0af20be0e13af49895696a3320e4fcac608768de3",
+  "blk.11.attn_k.weight": "066f5b3c144fce1366835e1ebf376f768b333b8ae29f5b478c42d1d0c809c855",
+  "blk.11.attn_output.weight": "e0d9f3d3f2c54aed59c02713ea4fb562799ddbacbe67ca3998dfc887bc44e47b",
+  "blk.11.attn_q.weight": "28d3ecc8a88cb3815e89a7f7a7d043da7a71f702b337a126e4d3a2ac1cd6370f",
+  "blk.11.attn_v.weight": "7c5cdef10ee73bca0a3b9f6ece5f0a0155664e0ce3d8de90ccdccfab5545e5e7",
+  "blk.12.attn_norm.weight": "973b133301a1af760cd7b3a7955371ea0a750808b442deb6adaf7b98482bd0c6",
+  "blk.12.ffn_down.weight": "d6c87b4b4ca03f75546ddd6a9e7fca720585a309188723c1ace8122438d4b200",
+  "blk.12.ffn_gate.weight": "2189a6e0cab1540bd05d6089b922aa8fd694be51255654933c165f302a0c955f",
+  "blk.12.ffn_up.weight": "5affbec19b58d092b9305721e3552481fe2eff51269ea3ed91cda3b9ef84d4df",
+  "blk.12.ffn_norm.weight": "f650fd42a34e950f758b4a130e7b8b1a712b1dcbede0291bb8edde47aaed0ef6",
+  "blk.12.attn_k.weight": "59b1e86f10450a7cc188beefc0856d2dcf44e8d7fdd9cd8859c30ec1ebaf24b6",
+  "blk.12.attn_output.weight": "446b0d36b2f66bd72a2323f4f4e9d85a0f621e9a58872e89a27248d6b1123238",
+  "blk.12.attn_q.weight": "3ed6bfd39f040301ed99fad882d3e569769d594259f9948445bef0e44ec881fb",
+  "blk.12.attn_v.weight": "e73652cd5d0029b1931be3ba9d82508f6696dce5a29d085476a54fb7a2ddbabc",
+  "blk.13.attn_norm.weight": "491b85278c0bd67bd31b9b8a9720902c244bd067e53a4a03641b7c0994782e82",
+  "blk.13.ffn_down.weight": "ad71cc248a85e9ced49307a24a9bfae01d387e979a7689c82ff59998e09741f3",
+  "blk.13.ffn_gate.weight": "0a55984d53971fab97575ee0ef5882013be7fdecfa76e3fbebb5dc85a07a14d4",
+  "blk.13.ffn_up.weight": "378b697b35e2e53c0de98e8e29b73d42ae3ec112ec16129aa5997a9e2f3b5943",
+  "blk.13.ffn_norm.weight": "f8aff2f69ab286210fad45a62b03f8d10b38f96a420d7baadf6b95d7b0b0bcd2",
+  "blk.13.attn_k.weight": "25ceb841afb1034831bea7f4d6a6c578def2ce4d4c412c780ef147dc9a598360",
+  "blk.13.attn_output.weight": "a242b322889c6bdaa14b67a7bab593db39df8eea3721638ef639abbb74d482e3",
+  "blk.13.attn_q.weight": "d80be9945a369439e835c55cfb0e97828b8a66bb7ced534d9059c92487bf20a9",
+  "blk.13.attn_v.weight": "ac33274cf9b67979d9ecdc967a55175afe0c9c4aeeff6391433cd9840c818706",
+  "blk.14.attn_norm.weight": "12a1e1091de5b2da12c9e7c0b1c8e6f09ce2a749733cf7d5240445b8e21cd093",
+  "blk.14.ffn_down.weight": "cfd41965c88266e32bc2dcdadda512499c35519e8686fefb9a7f249ab2291eb5",
+  "blk.14.ffn_gate.weight": "8dcfe774f07a095c7c6cf0a901c9df70d938bad7b5ba347fbc8f694e7603c0d1",
+  "blk.14.ffn_up.weight": "c7995577fe4a72ea0fb17c4a7b6b87b959072bbfdd5edacc6c367d43465809ae",
+  "blk.14.ffn_norm.weight": "81c41ebde41739e7016ffec31d2256217b825dc3cae049a935f5f61a60d22003",
+  "blk.14.attn_k.weight": "fb708bdebe4384f5c4b479c110028554f4d122f166b8091eda7d8d65e6780eb8",
+  "blk.14.attn_output.weight": "f5295caf2dfdc60553dcabe17537a80577e8b153c902247daac058df23542514",
+  "blk.14.attn_q.weight": "c12b7a3601c68c63ab5dc9d2599ebf3f3a10abc2c59d3a2126fffd5818f2763b",
+  "blk.14.attn_v.weight": "1ce968d9149bf0d5e237d52cc6d6433565b4bbf03252a736262bb00a2b34a687",
+  "blk.15.attn_norm.weight": "266fd2c36d7dcefc6b6bb7f1c9374c41f2bab5d6c84a063b6f91c4f682dad3c4",
+  "blk.15.ffn_down.weight": "6154886e9ef0a6cc08ab0d264a35f497e6f0987efdac992ed04e87088bea7801",
+  "blk.15.ffn_gate.weight": "183d9fd3c1b5657840099053d2fd3f72ad953b1de523296159b7761f20491a76",
+  "blk.15.ffn_up.weight": "51546d4498842ae2340ee226a0888d5f61e7d2ca4d052dfa06a77b0451242d3d",
+  "blk.15.ffn_norm.weight": "ef7378091a41a25a5f58bf1bf9d3bc64ea562e7f421e1c232b1f177c30fd3500",
+  "blk.15.attn_k.weight": "8d556ab8d9639324141774999b6eed0e91d7ee645bf3e7a3dcd200b2e7a00751",
+  "blk.15.attn_output.weight": "54aa6ba87def7cbe18b0c6ab3aff5c351cb3b6ca4a0d7b2cd5f75a1312991429",
+  "blk.15.attn_q.weight": "10731b0dc031ea8e0ef37bd7f010e0a78518a10a6df05a8bae48e3148b73ef3e",
+  "blk.15.attn_v.weight": "cbbe50c2ed7224866d3cf9b489c599f3ec41a4ea1aa3181e9f4e87e1fa0cefec",
+  "blk.16.attn_norm.weight": "387058eb39d4b28c04cf1368247417f1faeae8ae79d894c9f293457e0eaa00b0",
+  "blk.16.ffn_down.weight": "2cb26ccee585e933401ad5c82ed36ddacb3289efa0b28f8cf91b020ffbd9c333",
+  "blk.16.ffn_gate.weight": "d745985efb5bab42304e5d509024631efe35f92f2b2ec4931ead6db97ca9727e",
+  "blk.16.ffn_up.weight": "7a67bd195e0642828ca36eb7818149bb70c2c25f82de07e2b5807c520daf540e",
+  "blk.16.ffn_norm.weight": "7cefd061c8182482a89272f8a4e88a954b12609a62716923ca1cb3593b1c1651",
+  "blk.16.attn_k.weight": "d7968a2de67e755b4533e061aaad1cb62f8882af92dcad67f99d6d5112513439",
+  "blk.16.attn_output.weight": "9e9ab5788272ca3394ea89eadbce8c86ecc3fd75b7899184d6191c134ad9aae0",
+  "blk.16.attn_q.weight": "ef81c261b536c1a3a093b33f44cf2d42b86e5aa2d821674f07a0c80e992ed925",
+  "blk.16.attn_v.weight": "aef38e7958301b4a437cbdd2fbae6197f677b09269ec1eaf63188cd5da428d25",
+  "blk.17.attn_norm.weight": "28f6b289f1bc3131041e9f791b7a2a3a48baee0dfea27bf7051ebbb7ed364d80",
+  "blk.17.ffn_down.weight": "1a502829aafc6a9bd6bc81f12573bf8632d5c8c659f0dfb13c8b2411f3b1ec05",
+  "blk.17.ffn_gate.weight": "ddfd8aa0eb98846ebc9afe31366249159f46ae9815199dd70161527ed241ac4d",
+  "blk.17.ffn_up.weight": "4211a3cc247071bd361b30de2131d02382f552855062bf3b3e004c17992e5d09",
+  "blk.17.ffn_norm.weight": "647e5fa99a5b0d232af36d15816539f4d27e60a50a341b00aa88bb6e4474f8b9",
+  "blk.17.attn_k.weight": "d9125ff33a19c502c0f8846433ffc24395048582fc2f463d34a0301a82156f02",
+  "blk.17.attn_output.weight": "3d64fbb1cfef04444827f37c35fd9ad3413eb2165094d339ef89f00503f09de4",
+  "blk.17.attn_q.weight": "e5b29424028f578beca385fd82e29f37adedf3037cd51e5889d5a1ffb0428ca7",
+  "blk.17.attn_v.weight": "1809c5aaf2ac04c5d65539097564ad62796e87d24bb8b9ce5b095561a61d908a",
+  "blk.18.attn_norm.weight": "99daca58d001c627523d3adfbca1d95f04e590382a326866544d57989d5f4835",
+  "blk.18.ffn_down.weight": "84f30231ce6ca0f10227541dfc602d6418c1a210386b0c4926ef1656e7d4635c",
+  "blk.18.ffn_gate.weight": "ca5bbe4468b541740e54f69b9e08fcc8e478c344b70551dab21b1206acfbaadb",
+  "blk.18.ffn_up.weight": "0b3067b9dded31686dcfdc1e247eae3974a28a61ac59e9862758dbfaad64e8f7",
+  "blk.18.ffn_norm.weight": "8154a102232dbc0f90ce77ae5c1ff8f26f8b6e4dcf326e9ec1645749669e7960",
+  "blk.18.attn_k.weight": "25abb26021ccc481471a30e0d4cbeb7e1db29828417ec5136edeb93fecf09ac4",
+  "blk.18.attn_output.weight": "d87d481d9b046b68efa06ccdd4ed8cbf61e692d61114b75b7fad5ed75f5d87b2",
+  "blk.18.attn_q.weight": "cc6400379e15766992ff1293be79dc67682c28e9e15155a78109f4b64653b164",
+  "blk.18.attn_v.weight": "45c75cb1dd496aea3173aafe2575b841dd1d02cbe010b3198099731eb98f531c",
+  "blk.19.attn_norm.weight": "65389efc75297684773284ef8e5f8789a4504b636c9f33b8a32e0ee42499fa72",
+  "blk.19.ffn_down.weight": "4eefab7e939f64a17e4a214ca3c77a6fa110d94f677e2d6401086f70fc538b04",
+  "blk.19.ffn_gate.weight": "f1c0a59cafda66f466ab585b0b8b4861b58abe87a67cea1f6a488492242edfdf",
+  "blk.19.ffn_up.weight": "c42d045eef588db4a0e56960a57e110e1ff92eb8041107d19899165fd3b90f17",
+  "blk.19.ffn_norm.weight": "a8f33eda6d5d62ff5f333ad9771783caff556641f4e7df713451385676f441fa",
+  "blk.19.attn_k.weight": "0bab5d9e9083492bfb05a5a3bb23b79c0e7b99ef6a6644817b4d57d5c453b8a5",
+  "blk.19.attn_output.weight": "c99c551d70eafad0f7aea98fb6f9251635897168eb3895f76abf0d4ea3b3aa6f",
+  "blk.19.attn_q.weight": "c98bde95627c3b54c9443813ca50b4e14f518319681db6bbf7b2332ba26e9a60",
+  "blk.19.attn_v.weight": "ff3a490518cf64904db89ce0dc7d6eb89e870f1440e41883c6b55a221f82de84",
+  "blk.20.ffn_gate.weight": "761f0e317229cafe9d3754048ab038a0a84e9a287b196ab65f633139f2d29aba",
+  "blk.20.attn_k.weight": "45d13439b41066d282e8490a726785abf513605f46c79bd0c840f6419d27e790",
+  "blk.20.attn_output.weight": "a3b958d84b4a097844179b7d55c18fd0e4f319cb15e918c6fde33b68de1bcac6",
+  "blk.20.attn_q.weight": "127ab8e7d8c3f882874904196a02712bab42e6744fde45871b67350609d19f5e",
+  "blk.20.attn_v.weight": "5f0ad2d14a8ae42dd3bbeccfb33295687a14055fa92c54bc946249373c1c9f17",
+  "blk.20.attn_norm.weight": "77300b1755edc8c70089e0f45efa646056b9add7d8568b2324d2f3e62b64971a",
+  "blk.20.ffn_down.weight": "ab93d0e075b42e9017b701a070d561e698050d90aac4b4b9919256fbe50c3204",
+  "blk.20.ffn_up.weight": "4fd6628a07acc57a48d1ef83f81b7d7aa0bce569c1160a99d307284f8821322c",
+  "blk.20.ffn_norm.weight": "2a9e46b9e48e8e55215de56592e1f189530037c1c94a1428e3d6f106c7f26fb2",
+  "blk.21.attn_norm.weight": "4b3b5912c7bc61eb9da8e47d4651f896e85d9e59c4ecaa65df7acf3c21737298",
+  "blk.21.ffn_down.weight": "7146f931663d93b8771cd84405cd4802ea6560d0729b0d6d44588203c095bc53",
+  "blk.21.ffn_gate.weight": "b44ec5d64388fa40b90b3e9976d97a8b6800fa3b97584f32e64b03daffb8601f",
+  "blk.21.ffn_up.weight": "0cf3643fd23c685e17062cd11e116e17ce57a405e5e78953bab94cd62fe48789",
+  "blk.21.ffn_norm.weight": "4ef2cdb53da166df70b39f3e6b17af51848cfa5ea3c27ad6a1ae2a1bb1da1ce9",
+  "blk.21.attn_k.weight": "5d40f32a706f670c19972b14176bf660d5b045e3637b110dbf8d7de4ff32101a",
+  "blk.21.attn_output.weight": "18afaa916752ce16c9653ec0ec7e2fe60be55faa2aa5025d147be184adb75cac",
+  "blk.21.attn_q.weight": "2621daa5f858931514a4b2f0fe8d81cf9b96f541e6af99bfa7539e9bde8e34ee",
+  "blk.21.attn_v.weight": "63226dafc54c899bbce4aa49efceeedd8908e94faa613450fdda91f332b62864",
+  "blk.22.attn_norm.weight": "cf3058daab4d2c04387e7d169d1553bb8e7358eea66285ec067703f6ce62043a",
+  "blk.22.ffn_down.weight": "6a58d5fd220abdbac6cee7ba048abab794731af318f04982c2506df59413d0b3",
+  "blk.22.ffn_gate.weight": "d5614535324b03c7b91727a903b2a72f8d07ad17f7aa8b61ea173cf9b895069e",
+  "blk.22.ffn_up.weight": "ec20da3949566e93f66cabb67f8cd7eab399047ec6ebf5d43edfaf3669b82296",
+  "blk.22.ffn_norm.weight": "84c82f38f53a649972a44466fc476bf764e064ce18de870291edc302f3700e28",
+  "blk.22.attn_k.weight": "a3d2ecc37fde7c201176bb8abadf27f0d8ede9679a6034913e03d9db924fda12",
+  "blk.22.attn_output.weight": "5a3b8bb433f43a387df43dd371bdf80ddfac986dfeaf38e9bac1d7a0ec6628de",
+  "blk.22.attn_q.weight": "3a875cec661b4859f30a8fd2c866811184b25b68c9e36fe2663d299caf8b59c6",
+  "blk.22.attn_v.weight": "8717a83b79035058dcfd3ef6f8e5b36e71d77379e5a239e1899eef8766fb7703",
+  "blk.23.attn_norm.weight": "2b4a68a0a2f023dd646e4755c9bef17c2f631901154afd839edac7ac006ec99c",
+  "blk.23.ffn_down.weight": "29499b1586c6fc4883c9b7a9c8cf388035146b5aecf90c5c4c8c8e082c71e7d7",
+  "blk.23.ffn_gate.weight": "7d6554036d21c587b9b556428054f9c15cbef96d24b257f906fcef4ae38bd9c8",
+  "blk.23.ffn_up.weight": "19761ecb288d6ebd44b681c4535661583b1e19dc29e96d0c007333cd8f00aacf",
+  "blk.23.ffn_norm.weight": "37dc35500790a4ca33807b39cf7af65065e535dc25b9e94f3ed2759f61887ac9",
+  "blk.23.attn_k.weight": "717547d00323817b0cb40a72ec5f8cf42ecd1f9e3e42715c2cc5e38f07fffffe",
+  "blk.23.attn_output.weight": "a24786feb6a905fdf166d7500133757cbe494779d4ebcba9eb03046b319557df",
+  "blk.23.attn_q.weight": "6a2c4a98f138b928d22136efa163562691d3b4ed526d52d46a2fa2694a8f3965",
+  "blk.23.attn_v.weight": "c6e6081eb9c38a7fda023085957b460e9ea321e1fff408b38c2b58595c39979c",
+  "blk.24.attn_norm.weight": "5e6283f891e538670425f3e244b08dc6f96f33dfa4aefa913f8eb17212421850",
+  "blk.24.ffn_down.weight": "e09eb170f389deea0a4a1cbfdb52c12490768a2c60491b7bef8a4c445e2a08f5",
+  "blk.24.ffn_gate.weight": "af29d815cf49a38fc2ebd0bf9b2dd9933d023a29f2d766981acb9a1b53f09117",
+  "blk.24.ffn_up.weight": "36ccd9333426666de9d3088bd4dcdf5b624b09dca9e3a83a22fc0383f2d950fa",
+  "blk.24.ffn_norm.weight": "a88e1692318826db6ac42582d182e51a3c698c655d0e21e04fa086318832d07b",
+  "blk.24.attn_k.weight": "f7d61d6d1225289bcc502e3bbb0168b4584add0253218c1b77ac92ccef9a1c2e",
+  "blk.24.attn_output.weight": "85a1363b3ccc87312094c2195022687c16b0dad7fafb9e80bb4ec474d53c29ac",
+  "blk.24.attn_q.weight": "53482a2c008f42f4fad779ca323addc3712040149dfc12f782417756388a72bb",
+  "blk.24.attn_v.weight": "67498272369af7dd10097c73b07f731b565cfc9a559e711cc0d526389e7b44e2",
+  "blk.25.attn_norm.weight": "98dd617def5cb7825ee4833132ca2da2121245921585e1d9e36b93344adc321b",
+  "blk.25.ffn_down.weight": "7fd477d6c50aed5f424a878dd284343379cffbee8a34c0b6e55100c8305fa13f",
+  "blk.25.ffn_gate.weight": "f892c9806c8ec22e8aa746734ac9213428c534921cf161239e1d249fdb5d1ec0",
+  "blk.25.ffn_up.weight": "528bed14c9bf9762f790525ee40412545221f4321d2a2323fa8e73c58b7643c5",
+  "blk.25.ffn_norm.weight": "ca5831966672e7be6a578feeb631ec3570d3b5afe12860819ccb96e896ffc346",
+  "blk.25.attn_k.weight": "610d3068cc9b20401f0c3a0efea39a279dd9f564fde19baf3403b2ec2319e4c4",
+  "blk.25.attn_output.weight": "798aaf702e53b657265ac3b5e6caf3a0ab515bdadfeb1a3a156b4f3bfba76666",
+  "blk.25.attn_q.weight": "8a7fa25248de83029fb97b51d036a01baebe31fcb4be121ab00dd8b7de209b10",
+  "blk.25.attn_v.weight": "2a53d5e9f8a1218c66958c6388d3b37400a9af7956c785024ca44bfbc3c7d371",
+  "blk.26.attn_norm.weight": "5f44fc043481eb0771f3e6d2420bcbcf73140afb9a9feb8eddb6575452acebee",
+  "blk.26.ffn_down.weight": "944a60a409d0d5b6a851e33c69aca152454b691711a8b96f5bcc488772ab2833",
+  "blk.26.ffn_gate.weight": "2a0ca4abb3de5593e6693d8be69b63d6d1a639855ac8332a75f520353f030c62",
+  "blk.26.ffn_up.weight": "0b1df496163f9ac07bf89375d3eb441b51a81d41b47d769a04a61efc18dbe35b",
+  "blk.26.ffn_norm.weight": "56b8dd046e9be6ea71f7efd80dbd14e7fb1aa020d3cd38e063275f3873fd12f8",
+  "blk.26.attn_k.weight": "b1dabfabb970e6971c7ea6e53c63cf7ef56341e6a2edd9cf177785cad9af2f9a",
+  "blk.26.attn_output.weight": "39532c7e836baad164a655fb97ec5114ea4da37ffba9fdea2684f6e4450e6f84",
+  "blk.26.attn_q.weight": "8f48bf6aaa1252bc149e98af2be1777a5c0d2c3274c6d314171ea9344a41b604",
+  "blk.26.attn_v.weight": "02fb145f7fd905133750e90571effacadddfd3f4966552dc59982ac3900ab8c4",
+  "blk.27.attn_norm.weight": "654d168fc3cab716d91261f5719f180b7d697218401633b4878a759f1b5283f2",
+  "blk.27.ffn_down.weight": "2823272bec3a1c12f02cc4cb24aa4031abd7e9dbe0b02676e2305b21671818f0",
+  "blk.27.ffn_gate.weight": "b1a1d40cd02f97182cac17a79971d1934ee0daf3aa0bf11303568c636e208a64",
+  "blk.27.ffn_up.weight": "ed62ec72a020d070e64eb7b50237b32213944727b5b2427f45d989f50df5fb2a",
+  "blk.27.ffn_norm.weight": "c69649ac65d694b306a905dee8b03b89eec1ed188b1eaaf38f8e29d4b12e38a0",
+  "blk.27.attn_k.weight": "cc57bbf413f1fd227128dc66efc8590c73634cbd6f96d01ec4878b5e7ca6a925",
+  "blk.27.attn_output.weight": "cac407ad02361d53207b3c7e25ceab84dcb4347b8087055162e2efe14d11d84a",
+  "blk.27.attn_q.weight": "0af18e07cee12015761c07c94407024f4f4d77d97bdb24163db0e16669e2cef3",
+  "blk.27.attn_v.weight": "a1d08fbdfa40af773c5adcf93bd68b78a44ed144e3fc6bbeb8af02e937527eb6",
+  "blk.28.attn_norm.weight": "f39a51f814512b040a1082143150e4a49ff730f85cef49d7f77fc79d83e91f40",
+  "blk.28.ffn_down.weight": "74f29ed51055d1c1adb8f0660bbe538a27e016c65650f2d67efc6f1c84fa1b45",
+  "blk.28.ffn_gate.weight": "ae48bb16487ded6781c60aafc0bf738fb4ae15729952906f247d216592ce249a",
+  "blk.28.ffn_up.weight": "543009727718ac22f11ee4b17815f68ea6f15ba1f3e7ed5ecdb755cf6417565b",
+  "blk.28.ffn_norm.weight": "b8f9e54c322079ff20a82b88948cdc2916c22c7db40b9a9ed6d3cbe89efb727e",
+  "blk.28.attn_k.weight": "55d055ba653b728d6e784f9e013786fed07115c9fdf23367e3941386d5e77db8",
+  "blk.28.attn_output.weight": "155101c03ddbf18f4fd0694bfc982f33c7bae25c9b087d6f5273c2bfbffcf2c9",
+  "blk.28.attn_q.weight": "1ed19bfdd22e9c14eca014739982492e9516d411515a8585f65cf754d849e53f",
+  "blk.28.attn_v.weight": "11ba854dd575c025d37256eee9041f6d1bd2b549a083d6409a09bfc1542913f3",
+  "blk.29.attn_norm.weight": "02b0bf5e2fcefd11a153cc988c81ba672682e4844fcf6442423e21a0e10d566d",
+  "blk.29.ffn_down.weight": "594bb692ec2779938721ff4748666ca8370e0e4fe85229503f616438b8884f5f",
+  "blk.29.ffn_gate.weight": "8bedcf47e91dcb2cf4093de56b048ee411faab6ff472f89ab2c9c113a08e6967",
+  "blk.29.ffn_up.weight": "e241a547b5fd6dfca8200b8141e21c1c487a96cbc4e5855f181a7ed1be91b642",
+  "blk.29.ffn_norm.weight": "e63eba5e4c6b288bfd9f15e46e236086456c8b7f1f9c732c0b5de84962a2e7cc",
+  "blk.29.attn_k.weight": "afe5979d5bcf211aebb526620f5974bcb0a2c39c8be71e815575c55d6385e3aa",
+  "blk.29.attn_output.weight": "9c944ed44b124b014906fc240afd3b90aed56bbd9567f2eddfd5b7a685b3cb48",
+  "blk.29.attn_q.weight": "e234e08e5c1bd9245a2edc8d63e9933b6b879f97c01392209cad4f55f05f3ada",
+  "blk.29.attn_v.weight": "5cb8e3e5f954e775c5a5e4de7a9a62b17e9c6931bb0ff0e2f82c4126fd3e1a1c",
+  "blk.30.attn_norm.weight": "a65483ee51a0b214144ec8a14f28ea5437586e9e12ebe342a57d1f8627ee12af",
+  "blk.30.ffn_down.weight": "417959da77ceb33ead4271cbb9428b195196173a893c44e52880a7ec61b4856b",
+  "blk.30.ffn_gate.weight": "a0d503ffcbe45dc927600bb98c9f6082487e65cb577ab545add400d666a87638",
+  "blk.30.ffn_up.weight": "f8ab957b82ffcd10b21303cb5e866209b6fe95f827b1b94e9a949207952d12c0",
+  "blk.30.ffn_norm.weight": "210c7ceb0514a9ef27b5d4d1b3aff6dde43f1af0345a050d71097940e0e73e03",
+  "blk.30.attn_k.weight": "16861b9abcf5a3fe73c93d977ca45a1e6daa65be0fd85c2cff53486ce2033afa",
+  "blk.30.attn_output.weight": "ca541fb2e57e2257118c35784845b0c731278af8db3036ac53d71aa1681fdbdc",
+  "blk.30.attn_q.weight": "f7834917748e26bb456b945e230bc926c228e93696bc01fbc2b134bdeeac71a1",
+  "blk.30.attn_v.weight": "9292783171dbe5eb689d17c9bda11e537f0e9b328fced6986c938d61ed590e81",
+  "blk.31.ffn_gate.weight": "e4766a04bcd8f937ba883c6a144101e546747804ca66c35c97281d6ccb47b566",
+  "blk.31.ffn_up.weight": "cc1e666116f7e6b06736db4aa4b81003c583f54f4d9200bfa48842249940e16a",
+  "blk.31.attn_k.weight": "fc80b57557687504efae7d24265cb7dc39b8f826bb3d897a11783012dbedc44f",
+  "blk.31.attn_output.weight": "215617f50a1f5d9b2250b82f3652b35a9e9aa0ad9ef2b485d73965a14b2b872a",
+  "blk.31.attn_q.weight": "274b4f1dfb0bdec28632705677049fb3e327ce6d9e1f3baaad1560439039982f",
+  "blk.31.attn_v.weight": "e641b8b926f9dfcbbf6b6da1c02555525ac4b1c306d96f20cfbba7d6662c4e56",
+  "blk.31.attn_norm.weight": "b3243c361d4041ddb892ce6862dd5091f57d87357e3c67e177451b85d8baf34d",
+  "blk.31.ffn_down.weight": "0a00cd3ecd5e91624a27f9e239b1de425d5ba3cfff82c256a11a4ad434abf3c2",
+  "blk.31.ffn_norm.weight": "2a0d67ea2bb1303975712243f07273c92fce83baa11b1cd6d8e42e74ea3c810b",
+  "output.weight": "768615f077fb797967844571c58b94d7c399d884d115be3ab4b0154504cae892",
+  "output_norm.weight": "7cc5b7ce10e5082000fa00bfa68af8c7c5da218e59e2c41cf2f1499d40ca229e"
+}
diff --git a/convert/testdata/Meta-Llama-3.1-8B-Instruct.json b/convert/testdata/Meta-Llama-3.1-8B-Instruct.json
new file mode 100644
index 000000000..ad7cd20ac
--- /dev/null
+++ b/convert/testdata/Meta-Llama-3.1-8B-Instruct.json
@@ -0,0 +1,3 @@
+{
+  "rope_freqs.weight": "80fd5efb2f729381785b293a091a268cfeceb0079167f6ece9b07070e662b222"
+}
diff --git a/convert/testdata/Mistral-7B-Instruct-v0.2.json b/convert/testdata/Mistral-7B-Instruct-v0.2.json
new file mode 100644
index 000000000..88d447b3a
--- /dev/null
+++ b/convert/testdata/Mistral-7B-Instruct-v0.2.json
@@ -0,0 +1,313 @@
+{
+  "general.architecture": "llama",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "llama.block_count": "32",
+  "llama.context_length": "32768",
+  "llama.embedding_length": "4096",
+  "llama.feed_forward_length": "14336",
+  "llama.attention.head_count": "32",
+  "llama.attention.head_count_kv": "8",
+  "llama.attention.layer_norm_rms_epsilon": "1e-05",
+  "llama.rope.dimension_count": "128",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "1",
+  "tokenizer.ggml.eos_token_id": "2",
+  "tokenizer.ggml.unknown_token_id": "0",
+  "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676",
+  "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e",
+  "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6",
+  "token_embd.weight": "cde834ccac5e94324b25cb81b02d27312cac0c551b55a7e1d555d90bf6cb6e81",
+  "blk.0.attn_k.weight": "458bfdd9715c66e017c2447b1ed3c582963a3111479314e664faad8c914f42be",
+  "blk.0.attn_norm.weight": "e1fd60b95f713bae7b7e3ca933c64ae6c9cd1e8d808000204bbfdc19f0ba635b",
+  "blk.0.attn_output.weight": "df13b6a157d9d4f96c53b012b3b9bcd207d0c94144cbd22ae3ec13bb07d6c373",
+  "blk.0.attn_q.weight": "13b4126b4245bf06c915a93317c42b8174e05053535ec99dc576541e4cec7c25",
+  "blk.0.attn_v.weight": "5b1781d3a341214511b27eb4e268674ea3ea829dbdf8ae5a6bb89b3c0b33fafd",
+  "blk.0.ffn_down.weight": "49186f5d8148d316b07458841d13a2e66587f4af69b776188a809591ed9c070d",
+  "blk.0.ffn_gate.weight": "4397e30ece09136f00f4ff84ff49e5241b765a374deb8c5a12e897e2bf73473e",
+  "blk.0.ffn_norm.weight": "43260589aac3850a779bca3f9649f793bbfbe5db538361cb743b3830217f8287",
+  "blk.0.ffn_up.weight": "fd7ac918240a07566f6967527ffca58fcf433a30b78fdd6d84b2136d4ebd9987",
+  "blk.1.attn_k.weight": "209839566c7d235bdc20565a4766378b6ee8553133a5a3315abe8a85baa80712",
+  "blk.1.attn_norm.weight": "58c52986f7c69784ba327cb7f350923420782bee17fa39b1fbd13839d4005357",
+  "blk.1.attn_output.weight": "5067cc628449682665dfcf59b16e58fe2a9d2a81cb099f0fcd42f4f8670c6740",
+  "blk.1.attn_q.weight": "f410f9f0dd5edc09401af597d02e2a4c727f1502ec3ec3898321617b36c6df6b",
+  "blk.1.attn_v.weight": "d40fa49e07c102c0644e130e7909eaa93ed0d54e2edddc0759e721d58a4e4f5e",
+  "blk.1.ffn_down.weight": "594b1eff6ed4defbdd819fabbe2d48764984f08878a860bdb808511d5a25b8db",
+  "blk.1.ffn_gate.weight": "4cda97541e388a5bb607ce4cc8b3db1da7045830a630e7ba4d17807befcff346",
+  "blk.1.ffn_norm.weight": "66c13d7481be65b97aa474735ddc9674f33d512ddda76fa6fb45c7464b09f1ed",
+  "blk.1.ffn_up.weight": "1adc6de288ba4cc1237833ca8b4eb81107149842e38bc452e18e5cfe284338a2",
+  "blk.2.attn_k.weight": "5420423559f236ab22d85a00849f31e0cc6e9c7dd879de724393d8cd2b379153",
+  "blk.2.attn_norm.weight": "495fe1ab40cc52aa054ddd4f0c2d2790f4326c8d103296b1b38f3b1060db2a24",
+  "blk.2.attn_output.weight": "ccb83e7085381f558bfd65588c525ad2671feddcbc3887afb4038ad9c7aac348",
+  "blk.2.attn_q.weight": "2e8f77478392bc93c2a391f2e0f4a173a952bbab88a7aca099c6ee909726409a",
+  "blk.2.attn_v.weight": "d64512590f3b7ebbb9e77c2eb97fbda90b00d45c944f2b174f03a2cb11007567",
+  "blk.2.ffn_down.weight": "1de5084a05dcaa6b1bd926e83517dbe9ebe7fde79235fe56018b3028b1aa6397",
+  "blk.2.ffn_gate.weight": "cbea526b557f49aad8c976973cf367fcd12175b900f551984f498b9e07e4b7fd",
+  "blk.2.ffn_norm.weight": "530aa49b10c7eae08899d143409240deb95dae4e1d5bf78cea3b26393cff3ba1",
+  "blk.2.ffn_up.weight": "13a5fc19b96b4dcc1e9bd01998c8272ebe52034c1933ed123a506b711fae9a5c",
+  "blk.3.attn_k.weight": "1913b63a73305941d8cdc472e7f101c633d3357a78602eac0a4b49a744261075",
+  "blk.3.attn_norm.weight": "9c11bed5ab41f4adbfdae4ead65b525c8f19443e656a8c61ba412a4e1ad1193b",
+  "blk.3.attn_output.weight": "bb0b42c1d34779c5943272ed71f1dbb31ad8edd75f8bcd5c868f88505ac3a610",
+  "blk.3.attn_q.weight": "3461a1fe4e49f5319ea047cae98ccdb46528a3ec23831183fe87610b48c94948",
+  "blk.3.attn_v.weight": "82aa30be6a61526a41fb79bb28a2617416f5909f0477aa9e95e16be9370fcb38",
+  "blk.3.ffn_down.weight": "68521011ae03f5e3b0966127111afa8ee9f2eaeeef8d3a0b86b633e0332e9fbf",
+  "blk.3.ffn_gate.weight": "1e89e26338fd364bb679695968c65106382f15ad55c95cbb5ec9bdfeb766f432",
+  "blk.3.ffn_norm.weight": "c81932529a5a8c417c27b888dbe95fff8b447c2ea5f6f560444ec5d50b93832c",
+  "blk.3.ffn_up.weight": "305021735afd8669afefd713f56137248d5e817e60471a112ad06b7fa07ffe88",
+  "blk.4.attn_k.weight": "cc26ba5c5c28082a79e6abfe61186029e80b145252ca6a7924c437f0bcf2d51b",
+  "blk.4.attn_norm.weight": "302d251fdcc91f7468cf33f80b49484251d8917d7018ad264ab3a85c8ecf9ddd",
+  "blk.4.attn_output.weight": "a012f5bee3520cd4ce51f0076c132ebc3653309f304032ad051aa308f55f36de",
+  "blk.4.attn_q.weight": "3c8d607e447f5ef21e73af71e3c0d32fae16f91f31faae34ff06912cf9cb68fa",
+  "blk.4.attn_v.weight": "49f6c81a634ce46d71c2350206ecbd231b1732af96e4e4e67693c41a07e007d8",
+  "blk.4.ffn_down.weight": "e89504f311a4a34dc819a67b761022f14d71c43df3ead4f892c87aaa8e9f0adf",
+  "blk.4.ffn_gate.weight": "18b22f079a2fbaefe3572eec61fdcd996fd747724e2f0ff4f08cfcb43eb7bfb6",
+  "blk.4.ffn_norm.weight": "22415a492c168a0878912b05c854a631228b01c3ea8842e1d75989ec46c18a65",
+  "blk.4.ffn_up.weight": "f57379eae2874d8853f14ddf0f0fcc4ff1338574d5ed5d7e88331d5fb84f5642",
+  "blk.5.attn_k.weight": "d627af853c40bddf9762ce3988008c1ff17f2686fa8f73a0b5da38010147c316",
+  "blk.5.attn_norm.weight": "9ce01092c7f7f1c3ef72d6b794da12d77aa1f6a24fb96ba1b9bd5a0bcc3e2443",
+  "blk.5.attn_output.weight": "0388da8064c4b6b795ce2d8079e8a36535e82b2c9cf794e38ce8ae460aae726d",
+  "blk.5.attn_q.weight": "039b7ce1c909761fdf475c06cf14cabe5a90199282c89e4dcf460e95a4b6275d",
+  "blk.5.attn_v.weight": "c47bfd8d2496bdb6e00e03b903e15fd0ee806a515094ec257e43cc433147ab7e",
+  "blk.5.ffn_down.weight": "1d62e6708974bae318cbf00a8bf621d9ba0537e549ce4710a536520a8d14168e",
+  "blk.5.ffn_gate.weight": "8b42b1b11c92db19985094cbb50434e3a7c9cfea71ee6f21ea79eae7c49284a5",
+  "blk.5.ffn_norm.weight": "e0bc520f1505e687ec391d632a381d38d8ebcdec19f614a11a2000ab573e8b7b",
+  "blk.5.ffn_up.weight": "8cdcd17d2ea89bb9ab902dbc6bf3f827fa4ee029c6bf19eecbdefd146d8b6f2f",
+  "blk.6.attn_k.weight": "5dc6bcff89794d1756bf57ec665b58622d9352130d31082a6c66e1a079f99932",
+  "blk.6.attn_norm.weight": "13b26008abe0f119b5104b9d78ebd5e797d3cdd68122b93d73a3b4831a54d085",
+  "blk.6.attn_output.weight": "f5a49917ea70c3fb311ccfffbfafa63ab18416a5d55e5429b70ce8bfba57c075",
+  "blk.6.attn_q.weight": "d9c2f652c87dbd09ec3822e12876648fa32e86553ac25afab723b1cd9f8cef90",
+  "blk.6.attn_v.weight": "5ecc5fe67609a35151011cb526f45c56fc0a999079ae0ff37c755ca03c68c555",
+  "blk.6.ffn_down.weight": "0ec125ae0ecb2d9277fdb1b04f17efee94e37d0ae37311057c212ca2db3fe6d1",
+  "blk.6.ffn_gate.weight": "fa4d6d38355ee8aa3b80b476d65ae7e343c9b7770d7b097fc848ee8a6e091d1f",
+  "blk.6.ffn_norm.weight": "30e8f7defc627532e1739dc76d31223d45767391a431f925b63dabe334b0f392",
+  "blk.6.ffn_up.weight": "6b97cc32b290fa9087806b5d65aa6dc1760737730c8c71394cc4f30c2157f9ab",
+  "blk.7.attn_k.weight": "0231cb127cb7c3714cd72b8f39343891d7715a9bab2237ade9e7bc5f4ed2e68a",
+  "blk.7.attn_norm.weight": "7c3187f07eead7d219d98ab2daf87905e88d5f1ace109b6f5fa55dce3914981f",
+  "blk.7.attn_output.weight": "2f30ad972c284ae7c8eb0482053433495ebe8fe9c5ee2c28b4bc4ed1f33050fe",
+  "blk.7.attn_q.weight": "3a2b4b8d61cc9956d304fa9f82a9e65b4bb9fda2196670b16df7e0d8c43eff2c",
+  "blk.7.attn_v.weight": "d2aab97d0dcf0f61dd2f32848f7a8a99c423a4948a660a660a03a546972b8db8",
+  "blk.7.ffn_down.weight": "2270d520468c5549cd30023ff9c452a277058310104c4239a616373fc5a94387",
+  "blk.7.ffn_gate.weight": "4134a3ef71b3eac8f76b6f1a2e58625b3bae48081f175994bc3ed7d8b0d4f2d0",
+  "blk.7.ffn_norm.weight": "42df4abd4b8769b16f3930068f96960af1b061f1aeb7505384f272233b2badff",
+  "blk.7.ffn_up.weight": "c920549054ec16ff8c73a72f5d837cf4e11885e44db57c1c1c584c18fbd7a9a5",
+  "blk.8.attn_k.weight": "01c609bd3bf31ce65688f1f640ee413740e821330134d4ed1877a3065d1527d5",
+  "blk.8.attn_norm.weight": "48857411f769b00290f4e4f2e593e092781fdc2503f80c1e3eeda1b85a20f74d",
+  "blk.8.attn_output.weight": "90fb273f8df83744554bd59236515c16c5a5a698ca3fbedc17cc89ddcee354ff",
+  "blk.8.attn_q.weight": "ade617ac4653c7f00593dbb51837a468afef20a14eaab3780fb96ac3d6714369",
+  "blk.8.attn_v.weight": "c2c37496494864fee5c527d1fe1f88529d31c73f9cbd02ef9b2e9b23611ea50f",
+  "blk.8.ffn_down.weight": "2da58572e9ad79087c03cbb0c23c9ef69f93ec221fd5fe4ed92fb93871d23ffa",
+  "blk.8.ffn_gate.weight": "4483294e628edaa4901708e73e92c917bdd93b780fa01aa74aed57166f2bbf0a",
+  "blk.8.ffn_norm.weight": "c0cbb7a4f8123b62f0c4652a687f3b394802bc32870dc446eefb709e42043a7f",
+  "blk.8.ffn_up.weight": "9eaf8a2060cb9224cd585997cd671866c4051ad885c2c6d9fdc7056c2a5c0d89",
+  "blk.9.attn_k.weight": "5dd36c45fbc9c50fd35c36cd75576288506971eac5c5311d4f5c16ef60099645",
+  "blk.9.attn_norm.weight": "3c8ca64f2f75ed7c8fc1da010c23be787648139a96ca0ef3ad10be7b14942b8d",
+  "blk.9.attn_output.weight": "6277e1f833024f53c409be919ec76d34464a78b278c8f9dbf79e777746e3b995",
+  "blk.9.attn_q.weight": "87352b70d9e328c2d51d59090cf5ea5a046529864a890d0bc8986447a0a5c006",
+  "blk.9.attn_v.weight": "2efdf01161d7a82a9117cc2d87d37dba5ffefcf730781cb94fcc95130e48ff9e",
+  "blk.9.ffn_down.weight": "e7658a2ca984961c7ace16acb679387bedb1fef656b5330bbbf588db19673a75",
+  "blk.9.ffn_gate.weight": "773cd330d4ff5d64be8af00adf2e2722fae4e33fc26bb9d03549f6f4b3b0fe57",
+  "blk.9.ffn_norm.weight": "c8b86cd5c43b332f72060b807091c33a258e5dac01358ff4733b916cd34c9c97",
+  "blk.9.ffn_up.weight": "d8cc3bcff18bd46124ba2aa7caacc71220b44eeef6fccb993b4c6cb53e8f2c3a",
+  "blk.10.attn_k.weight": "964bdf3b4e77b915a216f750ff7b0f2eb1dd6bfa071358aef21010b90111044d",
+  "blk.10.attn_norm.weight": "59ed411d91d14775764eb514acb0895a75a10cbbfbc1c15d453bc50f8046cb7f",
+  "blk.10.attn_output.weight": "4d35a2a44cfe4ac0a83fd3ab0dcf1f5a0bf54cdb3b7be9fc353ed32c8a3eb81c",
+  "blk.10.attn_q.weight": "defff5339450dd881ac352f5c459293f39e07b9619ebd10ed632d79a3f310278",
+  "blk.10.attn_v.weight": "b9803e8d6a54acea58f662d4c0a5c8ebdf986676de7dfe12d4b288937881ce93",
+  "blk.10.ffn_down.weight": "eba856be64e4be20b92fb4639a783454dd92427250759df92a337e39f1971c08",
+  "blk.10.ffn_gate.weight": "2d5c509b066584db4de3632b01234e86edcde35409c5ebce18957dc80fe465e3",
+  "blk.10.ffn_norm.weight": "ecb9a8679945ff0273856624ce435dd250ffe5a440ea0861a5c84f0e4c44d2c6",
+  "blk.10.ffn_up.weight": "e76ec7e993f399af02958778c643aa78368e3067846714165eb5aba9d5f547f5",
+  "blk.11.attn_k.weight": "29c6d1f34bd3ba2f0904e57b32a5bf8dcb2834d439159a33edf234ce0b775677",
+  "blk.11.attn_norm.weight": "b5817b275149cd2abe18a6a10e19854605fc58fd364666744362ceee8cfe49f4",
+  "blk.11.attn_output.weight": "1e05653220e237cbe0cc770033e183c9a0eed5680510997409b16186c6691950",
+  "blk.11.attn_q.weight": "03db725ae669151e4d536e50285b3b047ad097f52475df208ed3e790e31a44be",
+  "blk.11.attn_v.weight": "27cdf1d4e971326c451a4615a0b79a8c7fe9508f9b76c0d52fa01971fc7eb403",
+  "blk.11.ffn_down.weight": "176938cd7c2966094f614cace8ba568b10532e45a0d438f80eccd19b6c2a7f87",
+  "blk.11.ffn_gate.weight": "9782339915dd6fa70013628a01524ee1d01ad8beab04068da7ac6a5ee7603a60",
+  "blk.11.ffn_norm.weight": "8245f6391e3be97811c0ff27f0d8f484ecc82a468a837c893f059745bfcd95eb",
+  "blk.11.ffn_up.weight": "15616ddde096d0d25e906375c548b6de4bd5576d1f6b68eefdc29f14e183af42",
+  "blk.12.attn_k.weight": "66dd21604993edd1b1fe547bcaa06f5bb7e31c9204902d147a227e4badf7feec",
+  "blk.12.attn_norm.weight": "23a69f85dd8a0904b9839cc5d0afcda299b74e82ae2642106224a1c820f2b761",
+  "blk.12.attn_output.weight": "4a98d132e376beb274a39d4ea9b6a1b870ad5c66625439d7ff6f45c229c3ca04",
+  "blk.12.attn_q.weight": "1c6c309d63afcfde32fe37257e300a78e25d01117e33490801107c0e75d1ea66",
+  "blk.12.attn_v.weight": "723d9e4ebe4e2b1974afa01d8f512b52933698fa36717dd47b37b07760c50a10",
+  "blk.12.ffn_down.weight": "00e0fb09e1f1fbbf3803f1dee373eaae7a93756b6e13063ab77f9927bc6f996a",
+  "blk.12.ffn_gate.weight": "89159f7f97aefb1e100107e3ac2d694e1008ad873f79bb953d60c2c1bb22724d",
+  "blk.12.ffn_norm.weight": "5f70aebd0e43a39d6373d8658cc670c13aadd7818831d3d84f761d5f688442f0",
+  "blk.12.ffn_up.weight": "faec21b446f061eb4dca561a3180712724347b77a71eb312e7afe9be9e89fa04",
+  "blk.13.attn_k.weight": "3d440825d19eac3b1753b34d94fee2b3a3cb6636c10b2703ffcf688d3c1eded3",
+  "blk.13.attn_norm.weight": "47b575e57e410738ad13fd3c74bb49c06b3d31030910834ece509cd1a5c6d9be",
+  "blk.13.attn_output.weight": "05436d8e613f4475741c1798a7c371b53d61b229507fa04fe23c504ba1f0e12a",
+  "blk.13.attn_q.weight": "002b5024ce520da41256e3ded5cdc60e5ae07ad9b202cb19d76ab511efd02b1b",
+  "blk.13.attn_v.weight": "c1f2d6763587c50312cee0d7140fa2c7ee326f5b172bc99b2d8946e08329cabd",
+  "blk.13.ffn_down.weight": "b5c4e0d8a3ff96cd76a135e415b89f02d28c28f7f3c16a36af31ef0ab8773da5",
+  "blk.13.ffn_gate.weight": "ae06e9e3d2e1f64c7ad23a4009dc904c2eccd7241f9f91c4974ab2504f116be0",
+  "blk.13.ffn_norm.weight": "e44a22321bcbcb4a3c345b504e939e8071370f54a8cd702fabdb40b97e0d7683",
+  "blk.13.ffn_up.weight": "7e6f366d538e21ad431264b12c011892d0be9dfe4c4da9f730af677f920641ba",
+  "blk.14.attn_k.weight": "95492d6417952ec24b2cab87bceb750fc7e95ac6b1944fc328a3852d980164be",
+  "blk.14.attn_norm.weight": "6b7b09e1c51addcdbb160ea59edf032531421c520ec5645fe1ff9ca4180cef54",
+  "blk.14.attn_output.weight": "75887474e4d72c218e6ab0f69f1bf3ec3dc414d51b36fc59df00cdb23421bb6a",
+  "blk.14.attn_q.weight": "940e33f76e48c21215d19e8a21234c8246d4d084381a7d9806aecb24b071d5bd",
+  "blk.14.attn_v.weight": "c58601cf5a9833f80f7f9a5b2656e8eab5eb133211446ebd48f8be15fed4ebb9",
+  "blk.14.ffn_down.weight": "f9f886e7f9b2a54d717b08947a25a0a93e8c2a5b8bcd5a907c06817c8ee3ac11",
+  "blk.14.ffn_gate.weight": "727ed0ee68594a3f59d704ed3240b6929f083b9c36650fb848d182315737245c",
+  "blk.14.ffn_norm.weight": "bd2471008ff1b2bae9aa26bea019393fb2bbc5b9493b8cec3ebd2c280fca24ca",
+  "blk.14.ffn_up.weight": "b006446769f51e4f93b503c4727deae897bc1fc7f4fad49f85024b63c4548d38",
+  "blk.15.attn_k.weight": "23bb70f9035356624039547a603e46be7d1e4403616eafc2451cc09c5373d522",
+  "blk.15.attn_norm.weight": "718cb371ca052eeb3bfac6ac506abb887df125271821fd171797a7f2d8dd6313",
+  "blk.15.attn_output.weight": "c76a2695a204b43a8e5acfa5720590b5d449a9ad9e082cbe3e80fab5903ea16a",
+  "blk.15.attn_q.weight": "2b3e4037b9e91bdd26d6e8d904cf39f948192dcf09bb6445cb55ca058d4f4626",
+  "blk.15.attn_v.weight": "7c15e89b6acafc8619e86aa9d412f5893ab17843ff2cfaf40eea9637b24910c6",
+  "blk.15.ffn_down.weight": "e16fd4bdc6d1c1209c6b633454df4992870c8cefb2cb0e8c92a7e489e9fb5d19",
+  "blk.15.ffn_gate.weight": "95a46bea366c260337c537fde06b4cbeaeec52484a69c3390bb1d178eb0525c9",
+  "blk.15.ffn_norm.weight": "37730293f704da265dc6d1896b3be00c39c0a41dab07f573af39dc30a481d623",
+  "blk.15.ffn_up.weight": "ba74a199da2d0875d7410824238c4ffafbda3993568812284a72b8800df91f15",
+  "blk.16.attn_k.weight": "f58f79a2a91c9a763adefce0c53a71eb5ce6bd8442f4af554b04b58083bff27e",
+  "blk.16.attn_norm.weight": "0c16e41b95e81978e0e0e3b338e2afe2d297426578cacee94de15df74e94eaad",
+  "blk.16.attn_output.weight": "ead22fc337514e4add49aee19720008558e52090466866e849671953a1fccba4",
+  "blk.16.attn_q.weight": "ef59c4e8fe8918c1add43d7e9c6fb3ef799dd3e1bdd731ec7b6a4a6f97c86048",
+  "blk.16.attn_v.weight": "902e6b84c2b64241470b13e6f412f859f66b4b223bcfb9c15d5cb1106b07ef3b",
+  "blk.16.ffn_down.weight": "2ad6e9eb4d8372c32a554395d460d17cfb02d6dbcb757cc962b6bfa36db4f5ee",
+  "blk.16.ffn_gate.weight": "825b2d50fcce3dbe6a5d8d8a50a95466f83ca4a10343efe67894c20b4628fb15",
+  "blk.16.ffn_norm.weight": "3bf6ac90befb0e17e077c8ea9454a8485a30f89f2d761ec7751b60c90aed1af9",
+  "blk.16.ffn_up.weight": "9fbdd08739b32411f5ab0252174d386bab19eb0b17884862f760429b7d41d78c",
+  "blk.17.attn_k.weight": "4033398718bf3674830ed1b73071ed8482b6dd4ef27f31a6c5fbb998321b6c07",
+  "blk.17.attn_norm.weight": "714f2e8ac9592966a0f1c02ee979eee8f84586405b992e8ee9543e840199ffa1",
+  "blk.17.attn_output.weight": "b6bbb618597d767b8f535117be68f92911e4a71d4eb4d8b5d943444151445ece",
+  "blk.17.attn_q.weight": "b84a0dc00ceb515faa2628125dcec502eed923077b21cfe900a4ff16c2e5f9ed",
+  "blk.17.attn_v.weight": "4387c7d6a17da9cc7a6bca8f4a75618b20407d570792056283a8e93b6ec65f18",
+  "blk.17.ffn_down.weight": "47db95c6f1e12b399c3eaf9ddba261782dd71173dd163b52af96541cf87b5196",
+  "blk.17.ffn_gate.weight": "59abaded0aedfd12f01df81f7a811e84db6a227f51b60abe9a247ca726e87392",
+  "blk.17.ffn_norm.weight": "b7e86445be5c7b722e01ddb98d5c7527ca86cb827ce0354f2c269e0f2558751e",
+  "blk.17.ffn_up.weight": "8e31c293bac649d2f60da4b3fc4a3acdce1111ec6058d8805eeeb242443011de",
+  "blk.18.attn_k.weight": "5ce762ab7b032511c131df81093b587871718c7097f79d8e07d707571f18a47b",
+  "blk.18.attn_norm.weight": "1f52cdc7af1f4dc1f0ef6ad1ad02e18cda32133654e57cfa9c72ada9c0b1d995",
+  "blk.18.attn_output.weight": "6486957f30bf8a88516e25772c6650f98b13923f490a2865a8752e36439d1cfa",
+  "blk.18.attn_q.weight": "93621c8abf69d2ca29c5207180eb628fb2b544d89de6c4a7fb0699be95534899",
+  "blk.18.attn_v.weight": "11604083b5a74828ac1d226af015ad5dc0215a1fdca44fa7131c2163c02d8156",
+  "blk.18.ffn_down.weight": "8f9997feb94385f106915df810239c9753b31efda2bf14bdf18a9fbbeec8233d",
+  "blk.18.ffn_gate.weight": "427c213b3a4e94af703429daf2f65766f70424d8230c123e7e712a18bceb5ecb",
+  "blk.18.ffn_norm.weight": "c45d305c4ea6a54013ba112f12dafaade064a32cf01317373464a3618d8ba44a",
+  "blk.18.ffn_up.weight": "a2811f2e73ac9eb9cce91a21a454e84e230a155244e2cd73f2c12aad3c9b8cfd",
+  "blk.19.attn_k.weight": "b2daed159925eac58c291e2f1e2000beed21002b03c9e1bc7e7a52e22240666c",
+  "blk.19.attn_norm.weight": "6307306ede2ab5bffa1bcac3f8b139354678c0376b1d9f5530c1fcb4268cfeb4",
+  "blk.19.attn_output.weight": "ebb98218b2a9c84d3fb6baeb02c5df264b7ab80d994d1098ba1cd47aa398effe",
+  "blk.19.attn_q.weight": "4f10df2ad09177e7528e9456039b670d07db22940a49417101b725d239c16724",
+  "blk.19.attn_v.weight": "30f1efc5114badaeaafa91fa466dc7fa14b1616db433c6f563ab851f7333a5dd",
+  "blk.19.ffn_down.weight": "be5ec7fe6b48855cd0015b0e430d1b70c620de87a7ff188c7c1afef546d7b6bd",
+  "blk.19.ffn_gate.weight": "10dffea4213881f8a9b583ee0fd370e033756d32255ed15053f794375b9400e9",
+  "blk.19.ffn_norm.weight": "e75cd24ade45dca78fdb0cbcaaa2d4a17d83a5a73dcc94ce0ec2d68fbdb2a881",
+  "blk.19.ffn_up.weight": "63e81bdb951410ffa81bcfba1b94a679ec9ebae59cd1623ce2651ed5d4c78bfd",
+  "blk.20.attn_k.weight": "c2fc5ad39e9bdd45e73c6e54aecc474388d944c4be1ee1921b7fcd035bad02e0",
+  "blk.20.attn_norm.weight": "aaa9169171937bdce20c1f057e94e9252f221cabacf1ced12e11b9586f23d308",
+  "blk.20.attn_output.weight": "a9f4fb496e4bc053e3f6cf2e72e22d4cd2b545ef6c32f7e782c2ef6ebcc21d4b",
+  "blk.20.attn_q.weight": "5a07ac619ed251494170b213921ef3fcc4c2712839da262516d9d5b8ea1ff185",
+  "blk.20.attn_v.weight": "d6689473105d241eacb17f09f06000ee237336916cf5ec4f48271c5b41bcb8e7",
+  "blk.20.ffn_down.weight": "74be38db51df736f26ede7c6b52ea787e385f181cb66231e2cced4556a25c9b8",
+  "blk.20.ffn_gate.weight": "ea91e06dc3d051c0ba0243b5a8bb40edbf254eadfb54fda7247e05cfdd88cbe2",
+  "blk.20.ffn_norm.weight": "5fbd357b3d6f44a7a91e8a4fc246b24303891b7957e0f3c32818ae5dc16ddd8d",
+  "blk.20.ffn_up.weight": "fe3290333e056af4ed12942ac72aeba97a6b562e2db05e79cd35dd07eab5b101",
+  "blk.21.attn_k.weight": "201ec6ee95f06ea5eb80fe86fd07bd016d3ae9ab6abd25d631834414e14a010e",
+  "blk.21.attn_norm.weight": "ea8154f93e06485828475a00b98cc397ac84768dd70e06ecc0c075b5712d7276",
+  "blk.21.attn_output.weight": "9f8af74d531478fd304723fd8e4e01578db598441b80dc7c960cb801dbbc501e",
+  "blk.21.attn_q.weight": "277de9953a8d3cff894ffd06c15ad0ee1407e319df0c1a693d4f45fa9c74ac7f",
+  "blk.21.attn_v.weight": "6bfdc16cfb898909b7788ddd39dd04b928f31d6732772195d53c558004638dca",
+  "blk.21.ffn_down.weight": "173877146cb94801157796ee9e5eecf3f46acb3b5e797f90b83a3fc22395eb30",
+  "blk.21.ffn_gate.weight": "53146713e2ca1be80496024077a028f6b6d749b02e71003c349e113b436f48f4",
+  "blk.21.ffn_norm.weight": "b28b97e18ab20a5c553ba422f7d7f6014f5902f1d62a69abd20d9fe19a5f9462",
+  "blk.21.ffn_up.weight": "5c39d0ac4d602b8ec8909dade93b2efcd6b6d9d84a19b252d76bb66dcfaab87c",
+  "blk.22.attn_k.weight": "01f26272c82917a87a3ccf922fa1d521a952b05de878241b7efe3525b617ac87",
+  "blk.22.attn_norm.weight": "5ffc96249d8873b506e9eb7158bdfd07fa1429e53c1951430ca7505d25f11c76",
+  "blk.22.attn_output.weight": "9c2201569358f720244b9c9497e4da02585a167b1414c8a506b85ad75ba990d0",
+  "blk.22.attn_q.weight": "906036eb4ddf027f6d920f9356a6a2a5e529b96f4e1231a0496d46b4434a5842",
+  "blk.22.attn_v.weight": "30ede8b0d166003a4b8a81fc99437f557719fc36e5c4dd510c9f161f36a47e73",
+  "blk.22.ffn_down.weight": "d04c164beabab30e1837b843e18852260efccfbb9d96a34ddd816e6fb3ba23c5",
+  "blk.22.ffn_gate.weight": "19c889db6b19179f0a62d5981a1506592c65de83760d67afbe00d202202750a8",
+  "blk.22.ffn_norm.weight": "4885eff2d851b32dbd306bd632c725857e6d164f0fa8b3d5857e572e6ef98ee9",
+  "blk.22.ffn_up.weight": "365594d8db8e95cf87cc33ac23947942dc326110175cc8ec5a07b5c7059089a7",
+  "blk.23.attn_k.weight": "badfea1569da0fc6ab817c5727ca3a69b07d9cfd622fb8be5e66678d5b3f7ae2",
+  "blk.23.attn_norm.weight": "8968f78a379ac3ca5458b4ed4251e8d9112aca6d6dd1ef6440b4bb0b380375a4",
+  "blk.23.attn_output.weight": "93e43393c03956287b1fe31e9735ff1cfe84f4ae56b83dbaebe96275e4e11831",
+  "blk.23.attn_q.weight": "aaff73c725a8700ae66bf26ac8869dfe96738eff23a8ff340de2ab53400a5795",
+  "blk.23.attn_v.weight": "3a86a8dcf14a746ed1411f5a7e634064bc4dfd6511c24cfeccfb2c9ebb6b4101",
+  "blk.23.ffn_down.weight": "d4da6f37bd7ef69bb203f7b0dd59f50bce37432c70627e6cf274ab81548af5cf",
+  "blk.23.ffn_gate.weight": "5b6072936c4a693923bb4e3d1473fd45545cb02fc07799aca458ef0449a04061",
+  "blk.23.ffn_norm.weight": "cd76e37025f84773180298ddb15e0d4ba9cfc7d832e19c791049daa47c6d9c10",
+  "blk.23.ffn_up.weight": "cde43b99b83124a13b2e4753d12674b3a61dfb34c04703007ced3e8e2aee1801",
+  "blk.24.attn_k.weight": "457379edc4cce4cbbe107385079019bc922264fdfc7bd1d1ae84343a81460c66",
+  "blk.24.attn_norm.weight": "0ce0dfab2edeede5da419fa7833db78e36222cf25c358d08f3ec664310f031fb",
+  "blk.24.attn_output.weight": "0cf91c2fd40c204d2fd4b9c85b69281e5ad4ea8442972fcd44b5fc8e835ffdf8",
+  "blk.24.attn_q.weight": "87ede30c09eafec6a4e6285674c1bc4637140b168b2da4ed34f36fdb6e176cc9",
+  "blk.24.attn_v.weight": "4c0b078b2798ca35d6d2c2258fe499820d2bc88700654ba4016e4b028f563590",
+  "blk.24.ffn_down.weight": "cdb8540c32b1ab988f984484928d39f6841f2131c1cebe90ad9456737fccbcaf",
+  "blk.24.ffn_gate.weight": "da2e0e913648b5526bd2bbb344038dd067639343aed3b413662b064b0db7556e",
+  "blk.24.ffn_norm.weight": "8940bd781c610d75eb2be63cfc8d869a3af05e53c963dc7fd4c6f653df5a80ab",
+  "blk.24.ffn_up.weight": "90cbac2a58801abe11ed6c24560aa4acb949f79429f2aa8ff129ac05868bb87d",
+  "blk.25.attn_k.weight": "90607131e36998e990ce718ad05cbecd1bcaed010931401ce6baa3b0d93ebce6",
+  "blk.25.attn_norm.weight": "fbf679c85656c04a6cf8fedd5412c1ace22960e6c2d47f2d43997827811fbb97",
+  "blk.25.attn_output.weight": "08412724ee7a2086514406e6f68fb9f622e10bac25b0c373b294709f4b09bd2b",
+  "blk.25.attn_q.weight": "9c1238e98a2747654a0d4371d3e7ea8b979867f609dc42482544f25591e85c7f",
+  "blk.25.attn_v.weight": "a57796a535c6cb09581cbafd6a91dc14adc8cca2a2465a7ffd0aec546cd84074",
+  "blk.25.ffn_down.weight": "f7e34e8a6391b480da08b52640613ccadce268373934b409759743a1735b74d6",
+  "blk.25.ffn_gate.weight": "b8d0b2f4612678b5ce42bd4a683f8024514b75fb5ebf6b22c600811e95582ee4",
+  "blk.25.ffn_norm.weight": "cde1fdba2369d315f3c6940a997c471ec891924e642505db580d732763bd7b75",
+  "blk.25.ffn_up.weight": "72e700c32ac8b9c47559c2222e45888a480b527ea512075423c5dc01678e2bb3",
+  "blk.26.attn_k.weight": "6ac83b3414ae75bf3a9055c32e49d2c40fe611ab21f8444f03d2f465d18122c9",
+  "blk.26.attn_norm.weight": "55f9d6dc9d75973dc75136ecb9d991b4398097ac133070873fb96ec76a6f60bc",
+  "blk.26.attn_output.weight": "ebc4fcbd15b33263e50ed2ad45740867cce15bc90e1216623babcb1820734509",
+  "blk.26.attn_q.weight": "080f057521073e412936fe3fee64fd574c8128fa4a148b879d3e598fe4954581",
+  "blk.26.attn_v.weight": "0fa2830d6746487ac91b243716e4302361f891e4e008eddd14abec47c7809d5e",
+  "blk.26.ffn_down.weight": "cb2ab8af1653adc57111ada49d2825c6995e338c8208455b92de10e580f60f31",
+  "blk.26.ffn_gate.weight": "231ce30966086bce2dc0e0afd34a22a1958cfda7a57c41b3b8e9444c5dfde8a6",
+  "blk.26.ffn_norm.weight": "35d959d25d17b00617590f5d5831bf705c385c51e46297a14375a700effca6af",
+  "blk.26.ffn_up.weight": "367680c8d332538b467d1ef87cfeb36cc5c6af564c5023c5fb50e728e3438287",
+  "blk.27.attn_k.weight": "0bfcb351c6d17aeac5b55a915074fbdf00f11c4bda98babb196ac8804805746b",
+  "blk.27.attn_norm.weight": "5d598a88c2e75ba59dd7ba4fee940bdec92d72038f1286536d2dfb71d008a09c",
+  "blk.27.attn_output.weight": "23a9da7347336479f6a10ded14cb3f46e06b5bd56dc4b0fbc526c688552ec840",
+  "blk.27.attn_q.weight": "b83319dba9055f069208e9c9d66da08bc6874f23e575288fcd81697d1777aa54",
+  "blk.27.attn_v.weight": "36ed34ccb2f36fdf16b2c2dd225a98ea6b7b0e376e7791191136ccd7bd7a4add",
+  "blk.27.ffn_down.weight": "5488e1d3a58c71b5e9ddda430540b4776b268cfe1457cbc1c2622dedd9e4526e",
+  "blk.27.ffn_gate.weight": "4ff48011ee0bac39af704849d9132a2410392c87a509c684f2062f6b76b498fb",
+  "blk.27.ffn_norm.weight": "32afe99675983da3de2961d1b5ca41c98970a356823597fe29e91f6e86abf0e8",
+  "blk.27.ffn_up.weight": "1eae3088a75629571fdbf6a20f141bc2bb2ed3f5ba2b9fd1d949f80695e442a1",
+  "blk.28.attn_k.weight": "c4e80af714962d6f9040d2c09f316f4a1cbc3a2e994e19902d7c653cf3c73dba",
+  "blk.28.attn_norm.weight": "c1ecf85dedc1c83d5d402bb7c94fb8b9c11f1a3e5f64e7680f80912d4a560794",
+  "blk.28.attn_output.weight": "72ba47c061b21f5ebc5213a455eaf6fc49c8f8e04ff9ce37e6ed4921b629161d",
+  "blk.28.attn_q.weight": "c4abc47234307f44b8ca789aa6668e298158fa4b459b2c1e84bd581806591cc1",
+  "blk.28.attn_v.weight": "aeba950799d4950e491ad0fcbe30334e39b8975177990a2cb339031c45ac153c",
+  "blk.28.ffn_down.weight": "4e84ce382a37b994fb8608df451a60040559e3f4f3241c3b3cb8989a3ed50d83",
+  "blk.28.ffn_gate.weight": "04df157acdc8e8534ad60acc2d2a4dd3a7a6610f6382535ec728994fa6f83f83",
+  "blk.28.ffn_norm.weight": "4d0386dae2bd1c1a9d0f9730718333e3a486c3bc6a5c5d482193c75d39832c80",
+  "blk.28.ffn_up.weight": "fec60bb0a3daf182a14bd8311fe6dd1e3fd020c5fc273e2549cdb1a2d6b79b05",
+  "blk.29.attn_k.weight": "b0532a263aa5a4e2a7a80adc83fc5dec974493bd18da7f953e7ebfc3f3a19aae",
+  "blk.29.attn_norm.weight": "593fc3b4000c35b7a59dace09ca1756c08be0105b2edd354a0e1c16c82898859",
+  "blk.29.attn_output.weight": "315b896f9f0cbacd0ca8937384c3a3a227efa908cb8c3a9125ec00c480e32b9b",
+  "blk.29.attn_q.weight": "d482d45386d4ad3394f08e9dff233ee3a70d0427d65c0b8fa05905da7e25ca53",
+  "blk.29.attn_v.weight": "cd3b5a6e2852da796902930a6a84bc87fc6a7c7bf51f8fc23758d12a39013b36",
+  "blk.29.ffn_down.weight": "5b3dba6f9753bd1b1ebcba65ef5373dd62c38e755c44b7231b95d93d45761f89",
+  "blk.29.ffn_gate.weight": "8610d9d2db15c256243ffcca3ffd31786d0ada0af0e7c7aa3fd20524370ab036",
+  "blk.29.ffn_norm.weight": "1a2ef2d38b7ac3e51190b9ccb8b6552ba83ab290e523356a7f851ddb35dedca2",
+  "blk.29.ffn_up.weight": "a5fdd15811bde16dc27677cf1a4c97daab4c28cb12a9530f1a0e573134fdb69c",
+  "blk.30.attn_k.weight": "1efeb0b5f4b45a85cdf47300f892ac77ac1f38000ec3653565d1303d1fb8c743",
+  "blk.30.attn_norm.weight": "c73934c182c7fe80838ec1d0b92f50a583f75f7a3d78d822f009b58ad2c80e65",
+  "blk.30.attn_output.weight": "3a0fd89de2d274614750345d827a9c886a4f97b343a13cdf680390505df596a3",
+  "blk.30.attn_q.weight": "711e113362bdb067db843c66236704eb1cd3fc5f40e3767143e96d510686ef4e",
+  "blk.30.attn_v.weight": "82b12a9a74fd3d91b73cc2e841e2b3f0a5197ccd2998afa17020995f880d2267",
+  "blk.30.ffn_down.weight": "af9f4b1287c0d824ae22d6e335d19e04a70135b835be7caa2435f1d85e931993",
+  "blk.30.ffn_gate.weight": "e2ab3e6f15f5c50fca66c084cb6a57a2b6b82406d65150e82ea0437b93dd9a46",
+  "blk.30.ffn_norm.weight": "c1b9c325c83f00e177386a4d7e769945f2995e60950c4a576c0a2c4ab9703d04",
+  "blk.30.ffn_up.weight": "9b94a21efd419715d82071b490d3b635cf1e8da080620dcc39e5bde976d7e9a6",
+  "blk.31.attn_k.weight": "0db0d82e3ddcc2c06209f5f013e1d72a84a996c40bf00186be485b909cc268e8",
+  "blk.31.attn_norm.weight": "2b8b7239471f57140c5cdfe06bd224a4f6326282f99736e44fba4c7b120ac101",
+  "blk.31.attn_output.weight": "a310b048840cc3ff2be4b84796340e8e2cdf05ec89d14bd3655c109b2bfa9fcd",
+  "blk.31.attn_q.weight": "f45e0cd95645175ea82813455356d171838539bc3f7676d877c698f2af0a0eda",
+  "blk.31.attn_v.weight": "8bde008e809112aa7e7c23e9c3099087bcc557313b01306c87efa0a4a30805ba",
+  "blk.31.ffn_down.weight": "8266fec7e203fbfad7033120861e44984581ff8b6851d01dfb7b81c5d8fa90ec",
+  "blk.31.ffn_gate.weight": "b73bc0aa5baf006d9ef6403104891b8133671b0992398fe038380b67e0d7e2cf",
+  "blk.31.ffn_norm.weight": "9c62cc27a7b6017c1df8ad49bff249a8245e8895c6754f402cd44623fda83268",
+  "blk.31.ffn_up.weight": "5b970a4694ea3171a0167f6e1636d9f00268bc1c9640430ffc35218494884adb",
+  "output.weight": "74fa0ef08c57a30e633e7117b1e9c805f833e2e5e21434bc79ddf9c92c6d7330",
+  "output_norm.weight": "59b8a59fd3fbf39353506116e43e5e76edd0cbf2a2873d869da4cf27a04997c3"
+}
diff --git a/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json b/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json
new file mode 100644
index 000000000..a15965324
--- /dev/null
+++ b/convert/testdata/Mixtral-8x7B-Instruct-v0.1.json
@@ -0,0 +1,348 @@
+{
+  "general.architecture": "llama",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "llama.block_count": "32",
+  "llama.context_length": "32768",
+  "llama.embedding_length": "4096",
+  "llama.feed_forward_length": "14336",
+  "llama.rope.dimension_count": "128",
+  "llama.rope.freq_base": "1e+06",
+  "llama.attention.head_count": "32",
+  "llama.attention.head_count_kv": "8",
+  "llama.attention.layer_norm_rms_epsilon": "1e-05",
+  "llama.expert_count": "8",
+  "llama.expert_used_count": "2",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "1",
+  "tokenizer.ggml.eos_token_id": "2",
+  "tokenizer.ggml.unknown_token_id": "0",
+  "tokenizer.ggml.scores": "e3d3eea80bb41a1213f2d0aa3e8a38581d1f19323be77dbd779c9c7e3b72e676",
+  "tokenizer.ggml.token_type": "6040635e6bd38d98af06698feb75c1802bad35180ee6ae0a503e38c0f60fd71e",
+  "tokenizer.ggml.tokens": "604ac4bfbd019e430d7b6cdf18c6c0cd5b967900601f0307f714ec7773aa5ca6",
+  "token_embd.weight": "1d1d1d39a867d5a4bfb32792a47247d2638c10c95a6259391d02843583505cc4",
+  "blk.0.ffn_gate_exps.weight": "2e5cd43ac3f26c44f071926ff6c3f239ecc52a34bc9a5b5906d3d4c1bf2fbbfa",
+  "blk.0.ffn_down_exps.weight": "a4dfc7e7c96e7402eb70279601675b956bb7331da8101e63fe5c0a611b6972e5",
+  "blk.0.ffn_up_exps.weight": "2d5d87b378b2319c344ed2c642598b6f7cb6beeb582a8ea51abc9ae690d473c3",
+  "blk.0.ffn_gate_inp.weight": "a46aaf5aba7401ce6e41f158242b4879d34901661f3ede85496cbd0ce79d6314",
+  "blk.0.attn_norm.weight": "3fe37d913bdd2b65076bcdd6efe64a37b0b03cacbb1b80b9f7089068aa35f38c",
+  "blk.0.ffn_norm.weight": "5e14308a3c894734eb204c8f558bdc817e94bbd5b4e9cb4094e91ba388c8f7f2",
+  "blk.0.attn_k.weight": "73d943dcac0911e87bd771f4aa1c901e1bfe1aed293af06e1a67812159859f67",
+  "blk.0.attn_output.weight": "4c5f754c855e262e8d4c94c6fbbb57af06399dc0e170d7d99a1a17fc9aab9227",
+  "blk.0.attn_q.weight": "d6fd7403c873d49c05f6f03208f30d99ad34cb3b71c9990c47334d502a8e4c7b",
+  "blk.0.attn_v.weight": "cf17cf64b2d683bd9de6cebaf60e5c264df6fdc38fe719dde9d54c80334f6366",
+  "blk.1.ffn_gate_inp.weight": "0d524de81cd915816b4e714bf595ad6946a9130b3de731cd89428b2781230809",
+  "blk.1.attn_k.weight": "2ea47f412992b374c70674730fe84700e0c8cce177086ce9b6635e42408964bd",
+  "blk.1.attn_output.weight": "b4b2520794d54113e86c8ff678eacfc62e35be4395a594a6c8c22b4383ebcc0c",
+  "blk.1.attn_q.weight": "5db930c98c4f91f6eab57eb974c72210b158e366d23d6d2890b2759c053bee33",
+  "blk.1.attn_v.weight": "079bdde09668394bf7af9f8bc175017b4f48f0ab64e6dd855a4d7561d1693c0f",
+  "blk.1.ffn_gate_exps.weight": "146a62de19f9ab093deb101f9640534ffc3dc40d69f508be12fc0475d01b0c7a",
+  "blk.1.ffn_down_exps.weight": "949da94a3c0f375160672a979e85f7def284264b10d48d038238aad5f5ece793",
+  "blk.1.ffn_up_exps.weight": "7016a3f467d9e3f2f4b4019579ed86b757469cd367f2b225483305376b4bb3c1",
+  "blk.1.attn_norm.weight": "1614d1e6ed537737275eb888666c7bac533f4eefbe73dec92b591045ca9e1afd",
+  "blk.1.ffn_norm.weight": "405a455fa7d1ec36894652ceb554bbcb09a07fd6405f42741e66dc4a4665c19c",
+  "blk.2.ffn_gate_exps.weight": "90d5003fc7421f44220c0842d43128955e91488f6f785fe570b62d81b719e964",
+  "blk.2.ffn_down_exps.weight": "ecdc2b5a8b504ef0a7833acff47d69b0c1fa9c22126de1bb120ff5e48c3d6e2c",
+  "blk.2.ffn_up_exps.weight": "2cbd9485a32460d315eb50a2f3b00863fd77245bfe885b7565efac1cdb1f191e",
+  "blk.2.ffn_gate_inp.weight": "0d0a17a1a2c7a61f2cca49ecbb479154dc93a870873257bc4f225e7607f2e2c2",
+  "blk.2.attn_norm.weight": "b2e4c5a977f87a6f880896bd73596234c9b83622fa0d7add5892501e3155913c",
+  "blk.2.ffn_norm.weight": "0ab875b4280afa922376cfc7b9aa3f7071c9432ea1254091ce7de3749df0e8e6",
+  "blk.2.attn_k.weight": "bb884af51fb51550acfef54ccf1b58ce8284e587806e6a2f88c8265e1ad05a5e",
+  "blk.2.attn_output.weight": "0f03099ba1ef342ea61af9cd71d028123bbd8b1dd7d7fd9b509aef77815427d9",
+  "blk.2.attn_q.weight": "8fad0d29eb4c9d24e564774ee3316b9eb7a4c4985e4567111d2c836c830f6cf3",
+  "blk.2.attn_v.weight": "fe04c847ff677632401a94e7b6b6fdca60391ab21cb23bd791533115de6303a1",
+  "blk.3.ffn_gate_inp.weight": "29e3aaa724590c070e614af8288939603d2641b0ef11e8c0f476bebb2776673c",
+  "blk.3.attn_k.weight": "231cc5631def10f7f292d8862d6125ff555164cd70480ac76362149fad204497",
+  "blk.3.attn_output.weight": "86467a605c62852e05fda1a7ef43150df2cf715fe59785dbcba09f1c27cfa086",
+  "blk.3.attn_q.weight": "901822402453922225c2d6ac79616691d48217635d5ff7338daa971d5ddee210",
+  "blk.3.attn_v.weight": "27030784f44375720df2f090933645a31a022d3fb3b14573e5ca0b78f44070c1",
+  "blk.3.ffn_gate_exps.weight": "231ba59cc0b988d125d77bf627aa3f04636684870af88f081f3944b48a160d86",
+  "blk.3.ffn_down_exps.weight": "530c3ab44ae4d66e8afa4d10c153ba5dfcdfb7321989a988e62e9d12e7234625",
+  "blk.3.ffn_up_exps.weight": "b85c2d4d9d11332e702b3c0a6610d4f525f9a93e5d12f5c7c55c592c40755e75",
+  "blk.3.attn_norm.weight": "05dbb6d88cfa6b199f9d705ccbda97c0ef13f9ec875c595398a1a42d009a4555",
+  "blk.3.ffn_norm.weight": "6880b1c27d46969ce36fac049c05dc8b89e4bb47dc89df357e32df7e18fc512e",
+  "blk.4.ffn_gate_exps.weight": "a883b4f225b760c5a2f6605dc5e2167ab85bb398c70bf64ceb539fcbd6128dcd",
+  "blk.4.ffn_down_exps.weight": "d291bb656aae77947d4b525e2819bf4112afece53ff31de9dab999af1f65f9c4",
+  "blk.4.ffn_up_exps.weight": "38592afb8ba3dcfb26970f906174f7d3fa62da44fa4be4fc6912a19030ea9164",
+  "blk.4.ffn_gate_inp.weight": "1596cb74e8fd6c3080b937b06468bb397b0dbb661e6d180a6bcbdc43e8bfd0c6",
+  "blk.4.attn_norm.weight": "f90c83c5ff4366281d283384efc941620542b9cfdea160d678dc54a75e33f758",
+  "blk.4.ffn_norm.weight": "d28d8c49d1746b7cc085562d1074905fd14023844de823dc4fb22202bb280790",
+  "blk.4.attn_k.weight": "792bbf412cc357140fdaba543e547a9b2f7582919e307bbd9a80c7d6d8f5f1f9",
+  "blk.4.attn_output.weight": "d98e4a062d2631d9c315f1990d5f6ca9a88e7e0e46387f611ccb0353f876aa12",
+  "blk.4.attn_q.weight": "1a11a55a91d9f748a72176ff6b1c174844df406e00d1b66b9aa64dc6ee4bcd1d",
+  "blk.4.attn_v.weight": "04cb3c02b12a6313c7ac7044513441083d534fb4c5a3f63bbaa58f7edbd2fadb",
+  "blk.5.ffn_gate_inp.weight": "cbd5cdf015d33a2da6703eb74c22fcb97581fb9175435173b6dc4f9e8364320d",
+  "blk.5.attn_k.weight": "4fdf3405e4d657403f5647b51233521310ee984b4b81bbcd901cb3e6ab76b7ff",
+  "blk.5.attn_output.weight": "4a25662c46979a29600ed77e1907cf81fb16ef30e724c155444e54ccb76af481",
+  "blk.5.attn_q.weight": "e2acb30e30b97300039bb20ad0878f05159d5657fa811748a51d5b6fb35d631e",
+  "blk.5.attn_v.weight": "306504b6a26aa123c63dbbed3f4ced0ed2ee8fb6a30bf0093539b817539f5ece",
+  "blk.5.ffn_gate_exps.weight": "7e34df9b9944dbeea5e8565786d3aa6937314a4b87acd4d0874687877c5a39fd",
+  "blk.5.ffn_down_exps.weight": "c4b7a57a42b5ac0a8ae27dcd5cb2646d7a7cc7123126d44a56ab128e85f60b13",
+  "blk.5.ffn_up_exps.weight": "09d47593b6dd6c664a9155bff02fc2eb7ac4a70219a88162d05c802a01d3c6ba",
+  "blk.5.attn_norm.weight": "58804a036d6ac4c1fe357b8b6a97a5c37cae1c2f06ee0086c041d449c1c6ef6a",
+  "blk.5.ffn_norm.weight": "d872dee6789f0826211aa46ca9d0869e3e96bcace9e77d6559a7b6f3e524f3ca",
+  "blk.6.ffn_gate_inp.weight": "fb1eae732e974d6c1d020a5b4ef98c5f33016f984701bcea656f999a99daad66",
+  "blk.6.attn_k.weight": "55e9c59c5051ab5519b3a7962e1b5fa96a3c0251cb6200dc2f177885ad2de470",
+  "blk.6.attn_output.weight": "f3c834a8d0027370350e2b6294d95434d31432e57be6313b013c15a56303d61c",
+  "blk.6.attn_q.weight": "efaefe5f11c2140dc7cb532b0832c2a0b363a165cbda21f00fadae77efca377b",
+  "blk.6.attn_v.weight": "900bd734d75616d846a90a121c97e081c956a3d1ab012f66dd0bc62c43e1ec3c",
+  "blk.6.ffn_gate_exps.weight": "312a99661b1468fcaed2474621116f1681432755e973f3ee79d01912974fd424",
+  "blk.6.ffn_down_exps.weight": "ac9cd7db67a2ef0d2b5def86873673d05e48d49d147dd944469dbb8e2d4c46f6",
+  "blk.6.ffn_up_exps.weight": "57613e7e09579400a1a09fee4445acfbfe83f2f327fdf317877787d96ada6b84",
+  "blk.6.attn_norm.weight": "0e8801e09885c633bc01a9a5b85d4e878d30158a4eb41a937dc5b760ebd044cb",
+  "blk.6.ffn_norm.weight": "b8c58062ac93072f878446b0e7f958c737aa47fb769fc3a8f593133d12db2dd1",
+  "blk.7.ffn_gate_exps.weight": "1ef611732ff13edfa8d30981ed9dac00c15ceba9fc012ed0b199e9280a849948",
+  "blk.7.ffn_down_exps.weight": "856c6811945c7b0fa461ca17811cfa43436b4cdf5326bad23cbc30883486d7cc",
+  "blk.7.ffn_up_exps.weight": "6725e3e33994302ee13fa5ec163631ce2dcaa08aadde8fc166c2265d4561c5c5",
+  "blk.7.ffn_gate_inp.weight": "36b49d7f80c1003dc392b2c1b9960cd49889dd69e77b26b9e4b13d01f3d0a32a",
+  "blk.7.attn_norm.weight": "7a0ec49acc5e20ee71c6f80ca02f4f1e564c485e0ae0621309e7c2eb0c616cf0",
+  "blk.7.ffn_norm.weight": "eeae035c39ab6e64bc06a4baa1bf6e50d4c8b8797cb0ad8abd48be86974802c0",
+  "blk.7.attn_k.weight": "e8f78c1def01a7a38d2d9bf7becb17755e28fefe4927856f7890fbee52840187",
+  "blk.7.attn_output.weight": "5367f05ac3bb49ef8745ba5902e1bdd4442415a3ebff2c7e1a3918d7be6fe948",
+  "blk.7.attn_q.weight": "37c95fc5acc55a4f6e5f02cab9be60e4fe54c08b65f98f4455741b4aa542ff4e",
+  "blk.7.attn_v.weight": "c89f1343486ba55814233511e94090f7365662a8a4214aa4c278cdadc79196c2",
+  "blk.8.ffn_gate_inp.weight": "4e239afe8c7afb8de3a005757c887cf14b1622ca2d224227591cb0e5301f4c17",
+  "blk.8.attn_k.weight": "2ad0229f30fdcc1e85ce64e00d8f75902238294844a81d5af43e14ba75c02983",
+  "blk.8.attn_output.weight": "2e44a4722acb3b521b81d0b910f8ca2f6c286d874a92ddd02150566454061699",
+  "blk.8.attn_q.weight": "1cd2b09cb2f43e08de776b5f7eac197a5a6d4ffdfd52b21baa36319450147bd0",
+  "blk.8.attn_v.weight": "5a22c57ebfd33ac500cbcfd321d5b5b1783f8728801db6f3f8bed51c7183e4db",
+  "blk.8.ffn_gate_exps.weight": "91063fe56cb4f3ff3b41052bb5046fcf8ef61516a603ee90aab893a9d68c15a7",
+  "blk.8.ffn_down_exps.weight": "d4c3abc8f1d1b462f67f70bd8f404b3fcf45dceeaa8527fa120527254c383c90",
+  "blk.8.ffn_up_exps.weight": "76a1a1f08ec577716a2e7027b45293e9205751126424f1bebe1de89c78f087d5",
+  "blk.8.attn_norm.weight": "f980d774da39eb76c52358afac3e38cb4c81cb323deaabbe5c41822e3f17a98e",
+  "blk.8.ffn_norm.weight": "1c937658cf90f1a85db9a5f26e077730fdd4b694607dbeeb825c5fb2bc407e0b",
+  "blk.9.ffn_gate_exps.weight": "a2532471ecb7896d5c78e5a34e10cfaf4125265e1595166c8d0d0dfbe2a3187f",
+  "blk.9.ffn_down_exps.weight": "b47921a28412d48fee450b8b9d97cee42344a2e69f06d407fd9523d7adf13333",
+  "blk.9.ffn_up_exps.weight": "7c461bd1b2a73b439cff6a10d94afa01e8b06f7e6f09d9a6f28e3876aef48bce",
+  "blk.9.ffn_gate_inp.weight": "1648dfb08b5c06d7953a5a97ecb764995fae9487fb729a1c867023b2538149d0",
+  "blk.9.attn_norm.weight": "8635db0f299882a63b7cfcd1d4259c9e53fab22c31d3d054de36b1001380b31b",
+  "blk.9.ffn_norm.weight": "f9309aa323062d174c463613afef9b0a33501b510bfaa58a8e0e866d12ffef3c",
+  "blk.9.attn_k.weight": "dfe62030441e947a588512d18d9c6e4ed72c2f71c227d622c095e4263b23dadf",
+  "blk.9.attn_output.weight": "1977beb75c6349c50ba7dd3865d7c0a9c5c5ddc854413147b0eec98ac4fda351",
+  "blk.9.attn_q.weight": "eb132596719605cd6bd1782487f121994629e115190edd69240b12af66e734f5",
+  "blk.9.attn_v.weight": "9e708f15d332d7c5187b0693b1a977eb30a2fa10bf7df48ed9d7537c0aa6ed99",
+  "blk.10.ffn_gate_inp.weight": "97503a5d166c1925f9b65c0eed980753d411714d66896f3d0fad5286c7aba702",
+  "blk.10.attn_k.weight": "1ebdd222336bd25b48df1b138cdbe09021c4a5562ea7cb78cadd1255d2be3a39",
+  "blk.10.attn_output.weight": "5e98faa38e9d514b9057e1c8342c509cbe1083defd518e506f6bad89117d1f5a",
+  "blk.10.attn_q.weight": "3323a26c87d936d1dd87c577d0b763459fced726679612c874b3de5fc6d969c5",
+  "blk.10.attn_v.weight": "d5fa73cb56aca388e205f44455e4b4f676fdc12ed7fac4542fbb3b41ecea59ad",
+  "blk.10.ffn_gate_exps.weight": "225021b53782800906cd13b70be3a4161e8b300b97f984a959ccad6a6e8adcbd",
+  "blk.10.ffn_down_exps.weight": "f08eb91526bd22f5fd0402fe925d6141cdbb308a1ced0330858d0c85c71f5ef3",
+  "blk.10.ffn_up_exps.weight": "a9f688350c3b53eaada5103b5848bd9a3d7d6b327a70fa16c24bf28ece933eac",
+  "blk.10.attn_norm.weight": "5ba426c9dfc79805015ccd76cd1068b0ad3bb7a8453e14bb1d35486f122d8f95",
+  "blk.10.ffn_norm.weight": "98891d6acbc3986b2581b7a3af9f5946a392d9188972c6a8b15d4e745a4f2482",
+  "blk.11.ffn_gate_inp.weight": "b2365a60566e7dace892e1cb0e62eb73ce387352601723e847052b34874feaa6",
+  "blk.11.attn_k.weight": "0efbc1d1430505543ff71532a4fcda821aeac616ef6c1dca40e00d4f2ff70bea",
+  "blk.11.attn_output.weight": "3d5bd4d9a41236f30d4293edb9ae27beaa113ffb31b4fbfadff3a4c370dfd3e6",
+  "blk.11.attn_q.weight": "aa11e9db14dd9c77951511443077c2a1a78070753d7bd3d9811038473f69e325",
+  "blk.11.attn_v.weight": "5adc567f377aa11d1763d35f50e53fb2896a8b03b623ac36acc45efa2486d512",
+  "blk.11.ffn_gate_exps.weight": "71d07d982aabfab9eed3c733d49c20f023bf475368fc71db5084d91beadc4b47",
+  "blk.11.ffn_down_exps.weight": "9a06e61461e48b3925a9f7d9cca634d048c8b62163d7bc5c43e35899f959319e",
+  "blk.11.ffn_up_exps.weight": "bc05494d0dcec61021b3ac0c5bc1bf502736cadf48224e213bc139d562699a89",
+  "blk.11.attn_norm.weight": "a5758a10bdd0404ae1470e8e9db903985d4d07f60553c5001a5e7b660d4f7ada",
+  "blk.11.ffn_norm.weight": "814ae037563aad3771787316bec4806c95bf6f5991dd6474b4b1e5cc13dc18ee",
+  "blk.12.ffn_gate_exps.weight": "3a68b831ba1606fb9ef6dffed4732032447ecef23ea563ff4e79317586c7eb49",
+  "blk.12.ffn_down_exps.weight": "268b25e13f4b7beab08686e83705a41b21d15251809ee4784526f78a580da829",
+  "blk.12.ffn_up_exps.weight": "9105751a5b5b42ca2614d0456f24f779d2e2ac8cdff0f96842aa7ae2b70f341e",
+  "blk.12.ffn_gate_inp.weight": "d0de1558cc1d458c5c504f63ddc59785c323df7330474bb0644c346104b40a3a",
+  "blk.12.attn_norm.weight": "859a4c8113678e2e202d10299850e0cfb52eb11ea50bcbf4fe3ff39bdd394154",
+  "blk.12.ffn_norm.weight": "7fbf4c459c1760218877e9ee3f5ad49e960956a4369bcfe96c143f04ff9ddf97",
+  "blk.12.attn_k.weight": "0a7e254fdf3730a57372b6ff421a613eabaea68cdefd64800857941411318374",
+  "blk.12.attn_output.weight": "ceb763fc15d88af149d8fb78e82db2b7dab3aeae584af8cf7611a12356a397e5",
+  "blk.12.attn_q.weight": "a43402d23c46cb2d3cb3c2a98c81b19d10026b7e6742370fed6b2880b6e049b5",
+  "blk.12.attn_v.weight": "3bc24f2c0480ce91ef72993ee8f1cf962f7359e12183424583ffa1246bf3db52",
+  "blk.13.ffn_gate_inp.weight": "a6d68c82bfe66d8bab68f980f5f18268a9e2c0cd6b8832ed39010e0de198ae05",
+  "blk.13.attn_k.weight": "0166c39546b37dc2e01b2b396ba43e183f797dd04eaa51a6d103d8b58ee4bace",
+  "blk.13.attn_output.weight": "2ce5eb198deab9557475a58b69b11e9874b547e05c23f223c6e42fa35ddca069",
+  "blk.13.attn_q.weight": "745c1bbdf434284a7fae98f45e821c076dd9c2a2467dba6a9d8cf0041e419dbc",
+  "blk.13.attn_v.weight": "9ece68d5ac64d1421ea7aa32e1cff9cc1fecf5175f4c4da858dd31d8633e3337",
+  "blk.13.ffn_gate_exps.weight": "ccfdcb4670b131689de12d396a010b5ea737795cf5c15a14a304d720b3c7c899",
+  "blk.13.ffn_down_exps.weight": "8b8fb328664764f1aaa5cbdec336d5654e981e965a02ef622bde5f07ea1c164d",
+  "blk.13.ffn_up_exps.weight": "d2ace0236c2fb3365fdc85499d676a7f65813c48e5085348b1df1799922766ec",
+  "blk.13.attn_norm.weight": "1ed29d7d89ce52d7cb4d57e895ff7115430466e917136c049c385c030ed44e9c",
+  "blk.13.ffn_norm.weight": "a194fc542597a4dcfdfaec5e3cba2a2b2b21b21edfc87c39c0d7f7651355bc4d",
+  "blk.14.ffn_gate_exps.weight": "a625e3574e5e740e7f8e2f9c40390f2f382c720aab5b10534e298002dd8d1fb9",
+  "blk.14.ffn_down_exps.weight": "bc366f015b83c865946afd74c8a884943e0ea2c671314a0b7bb72f21a44d2f78",
+  "blk.14.ffn_up_exps.weight": "ee3199bf2086de77b49f57f487676be8ee70e102a2fb5a5ef8ddbbc28a9eff41",
+  "blk.14.ffn_gate_inp.weight": "2b437870c850fa2e2044d032bb02908af634356e37466fdae260b933e48ee8b4",
+  "blk.14.attn_norm.weight": "cd8344d193a1cbd42bd898e17f4bcb1ca0b2918420fbdafa9249a6f2b7f4ae06",
+  "blk.14.ffn_norm.weight": "70eec40374e558fed5b07257283cf36342b6b0129285a00007deb59c32c9f7c8",
+  "blk.14.attn_k.weight": "4053bdb507e0543d724b632570bac86b31707696d90a0db44c49b2a082e0d599",
+  "blk.14.attn_output.weight": "0182632cb0e06a07241b8293d25d109fbc1862e1e337d435f908e8681e2eb1ab",
+  "blk.14.attn_q.weight": "ffc7794a4c1b6f793c842dba969435330a7a80b9212e457b4b2ac33e68b41241",
+  "blk.14.attn_v.weight": "6411805292d528e61bbaad8f9aab9dd073529a17946c057fb06864fad9cf3211",
+  "blk.15.ffn_gate_inp.weight": "77d0744567c76e6abb67f81ba9c715b2b544841186d5b948309571eff213bafb",
+  "blk.15.attn_k.weight": "1f7957954ea4c6521c257b35a360e868ffa02bdb3de91f146d5e06bb4a545c98",
+  "blk.15.attn_output.weight": "d7809d36bd8d3342240c46fd87bcc7f9821a222f48d9a95e45ae50460265d3cf",
+  "blk.15.attn_q.weight": "25f509313ae4d8401b871904059f472a26f5714e7c791c725de77a1a522c976e",
+  "blk.15.attn_v.weight": "96fedf5a591fc0f020e6de10fd72ff12b3ef9cf70cd21dabaa0d3e7b06f54e73",
+  "blk.15.ffn_gate_exps.weight": "8f950d976b2fd9a3d213b84123cf114c1377efde9352767fb2ddee89e177c8ef",
+  "blk.15.ffn_down_exps.weight": "6fd09d1557bb94b06efbd4f6a1ca4be532a202ba290e9315bc8da3d12a5c4c4a",
+  "blk.15.ffn_up_exps.weight": "cbeb59ae7b0266a928dc7e3a6e70a9330b92f9ee1b17ee1ed91022108204a33c",
+  "blk.15.attn_norm.weight": "2005330911ac2edc7b6d27aca021c67d30d16eb632e49b1a13f30fdb2717aed0",
+  "blk.15.ffn_norm.weight": "0e9198f3b548eb78acc8961f2b3350d238d26cec110933ba753a8cf0035c501c",
+  "blk.16.ffn_gate_inp.weight": "a41d1f99d739c8b150c3945b6949763988d0c6a4c5a2b5855592ca1a48ed23d5",
+  "blk.16.attn_k.weight": "b624e2ec88c2d3047f60530fb87e72cb4a5e655a9663f6f3e9b09e5ad32cddaa",
+  "blk.16.attn_output.weight": "687759ea75e45108526ffc1573d6fdf084728079bfc2dc89b9979e76280f43c4",
+  "blk.16.attn_q.weight": "beff3a45c7e9ec82ffc6d3c701126be28654d10aabd747d03441210491fd31b6",
+  "blk.16.attn_v.weight": "43a349b13f0b9d040cacecd942bcb168c030fef8c75c987d59a4fce6c14e855b",
+  "blk.16.ffn_gate_exps.weight": "793406d6c13d727c82bb7b692ca98d65ca975baee69fc57be5378d77c5a19b62",
+  "blk.16.ffn_down_exps.weight": "9bad3dd150d0230404b7f886ac7ff8803225757e813f195cdb26bad245243b4d",
+  "blk.16.ffn_up_exps.weight": "7449d663023fea3496475bf0a9c1de7272ad0ce9adcb3265e8e424badaa674dc",
+  "blk.16.attn_norm.weight": "a424ce34c195a401df1ce37ac4f2794e8a6720b1ee8acb21428e2b68c65e0125",
+  "blk.16.ffn_norm.weight": "405a68bb8e16e1064df2de55ca3cd9ceddda1d9fc0af007a9bd7cad4b2676248",
+  "blk.17.ffn_gate_exps.weight": "97c6e5321491ca5dc039ee88da0eb0e78f347372785411809af84b3298cb19dd",
+  "blk.17.ffn_down_exps.weight": "1617ac19788a1be19bac69277408761e6bdf5719d63a8c7fea14d41cc27641b5",
+  "blk.17.ffn_up_exps.weight": "4ead1c365f112581c10610ea3f63d2a1474311d2503d2060fed4b458ef337f5d",
+  "blk.17.ffn_gate_inp.weight": "ed4b3393f2523f2b5e0fc7680a1caa2842e605728a529b5af68a7fa8d7abf940",
+  "blk.17.attn_norm.weight": "beac17ef86a7fb2b5840cc72f7a95a5e3d6bd24e7fa698e0b0ebb9bdac45c561",
+  "blk.17.ffn_norm.weight": "81cb58ec6d6dc02a0b4ede10adc336dc865fa76f982d4eab0e4a37b40f5b0fac",
+  "blk.17.attn_k.weight": "eab569e5ea8c8b05e5a6a209fba031129453c2e28181eee3e736b3b04b36bbec",
+  "blk.17.attn_output.weight": "f85b70f01438ce8fe5d10599b113f30bf18dee2bbae0657d3eba295870001db3",
+  "blk.17.attn_q.weight": "887ceebfbf6a2b94b43d2df4439ac3a5bbc29311d4b28addc04d525546032047",
+  "blk.17.attn_v.weight": "2df9414d65014c06a93da22ba3a668be7b83e2e8008e98d7771f7dfebed98298",
+  "blk.18.ffn_gate_inp.weight": "9b07741a0950fc667e5fd25937e33bc22e1f764f80eb4ff3119f005327ae0f6e",
+  "blk.18.attn_k.weight": "8649598dbb63938744c39bcda5ce8c31773e29c573be8d4d2c114f5030f8d3e8",
+  "blk.18.attn_output.weight": "f8e391adb92622298ca834d5d1eda48b69c3b1c51c5a584ef6c54a725c298d75",
+  "blk.18.attn_q.weight": "84bf8708a2eed618f48f69c178ed7dd11fa4c468102376e72e910ebd037d131f",
+  "blk.18.attn_v.weight": "31db3cd773f09548c2c1b1eac2718e46364a7810970fe9c433fad9d8de5397eb",
+  "blk.18.ffn_gate_exps.weight": "be2a2ba378002f1b61f86c273a69eede9b93786d5ce96b4fee1861f730dca4c4",
+  "blk.18.ffn_down_exps.weight": "d35196159e37705db50a5343e3989f7335477f1a4add67ef42ad64a638cd07ae",
+  "blk.18.ffn_up_exps.weight": "c6ceedd86e97913a6dcadc838e7abb762d629fb8dd55f15cf02fd9bd66d2ba78",
+  "blk.18.attn_norm.weight": "41f0b1ad83d6e3cb9fbe0d27878c2e7ad4a351b9f554a6bc9117c01745cdf6e5",
+  "blk.18.ffn_norm.weight": "96646204bd0d82f25dc77faba4dbd86b1332e449313e6684e00122da8be99057",
+  "blk.19.ffn_gate_exps.weight": "c6eb7f61e7938bda0492dbc05e51e8f631c99224fe18e99861fc4fc53ba9e9ff",
+  "blk.19.ffn_down_exps.weight": "4384803da3a3a3d44120d7dd192fe2c9bbd9a1a0cb492dbec1fdd7565230f1e8",
+  "blk.19.ffn_up_exps.weight": "22d73de2fbb8bb0f1bd2caf17fad8a355c47d914143f7f6e6d0128f66f074a60",
+  "blk.19.ffn_gate_inp.weight": "9a0cc4a2301a5634022fbce41189021bf0d1a961792d2d9330fd35556d18e5bd",
+  "blk.19.attn_norm.weight": "c5cc56ec5df9a1f7d5ad71fbda49f1433132e58895d45cb44c73420bd61ebd6b",
+  "blk.19.ffn_norm.weight": "77e17de741742ef2482fc7872fd423c8e3c1454dc4d2be89ee939084b6d78bc0",
+  "blk.19.attn_k.weight": "a92ea36ce2e3569656306aeefb835ccd5d1b03b33a86e0d3d030644cc923b813",
+  "blk.19.attn_output.weight": "5e2a912b37855f84ea964907a1a86d609cbdd79efa0c93c3e8e2fc07caf7c226",
+  "blk.19.attn_q.weight": "4ef3a5913292ac3c1a6fd3e9e53d011021f2b41d0276cf849706d1ca925cf7a7",
+  "blk.19.attn_v.weight": "42981b75b68ae852cee638b5433605c147da4392aaa6d7a06e756115b0171f39",
+  "blk.20.ffn_gate_inp.weight": "71381b9879a7c80b9f7b475abc0aa31b8cd71ccc00856ebe89764a2acb9df2dc",
+  "blk.20.attn_k.weight": "1928b7ebc054eb3967929ed6fb446314d5352f4aaf8b475ce55c6345019f2ea4",
+  "blk.20.attn_output.weight": "6071ecd9ca91af0d2ba93fef4a1a56f3b243dd70f862a21a2d164d56f386043b",
+  "blk.20.attn_q.weight": "002e95042a40f36ceed5829e3d0c8072e5f5e4ee86a089e2902b2348fed24dd5",
+  "blk.20.attn_v.weight": "42f509cdb1c0e298f89f896e349be86952c5168e49b3f83bb17badbcb7596d57",
+  "blk.20.ffn_gate_exps.weight": "a684a3ffe4b0a57c819a5fa9cb3521de223f392732927271e97ce925b6e33765",
+  "blk.20.ffn_down_exps.weight": "e3081a7bc7ba750d8a4886bc8ca4f231b55db4ca082b54b4106c7531964725cb",
+  "blk.20.ffn_up_exps.weight": "fad0fd5eca36ab154788da28be8ec25bb5d6db06c9d133db89e96df358a2f6a2",
+  "blk.20.attn_norm.weight": "c3e3f2429715ae95e884ef1246b0b461b23c5cc0ed08beecf70a14cddd184820",
+  "blk.20.ffn_norm.weight": "ff31f609dda65ca496b0584fabea6550e42edd05ebf229812aa6b7bb5ede15e6",
+  "blk.21.ffn_gate_exps.weight": "366f09ef0ecfb86808eb3296cc9abdb957951d27f6533c03f1422b54061da660",
+  "blk.21.ffn_down_exps.weight": "3fc495947d27fcca7fc0893c8a96e5d48ba27b2c8c58f8fcfb8dcfcd5539741c",
+  "blk.21.ffn_up_exps.weight": "6713ed51410bcc8283cbb001c4ad784098f25701e8021f4fa4f411e186859c4a",
+  "blk.21.ffn_gate_inp.weight": "6d4c92c01ec801647134d907bf1108878156df266a6107abc10526332b328b93",
+  "blk.21.attn_norm.weight": "27605719ae2df24f4f2e85a730927cab20367631612cb501631f6bbf38eb1209",
+  "blk.21.ffn_norm.weight": "ca80ee8177db185b15a4a378c1cb6f7143c76546a7f1726bda23f329323d4ffa",
+  "blk.21.attn_k.weight": "9e49f743d4a5bda9b4bd9c40c2ca37cdae5aec7e54cb193897ac8b4945ada14d",
+  "blk.21.attn_output.weight": "ab923540879753feaed152f5950f69cdd83d8f2413ca873f5f038b63ab0aea12",
+  "blk.21.attn_q.weight": "62617fc3f1c9d2aa672a4d91a121c7a91b92d145b65e75f0b06b4bb7c825dc36",
+  "blk.21.attn_v.weight": "15f8b2e72f8e8e992f2f6b3e93238a9d7be7bd6136f91c9d04b4b4cd0cd60369",
+  "blk.22.ffn_gate_inp.weight": "3ddb1773d9257b68add7a2a4e94dad25ed926803e02707863dd742ab9b2dc179",
+  "blk.22.attn_k.weight": "680e45a9e8d5feddee5266e119dc053bf80718fa9af1cf6803e6f493b265f1eb",
+  "blk.22.attn_output.weight": "0d5fae3402fb2c5aa3a860010e3973fc8e3168d1015f7a76b7b2964681693206",
+  "blk.22.attn_q.weight": "eee7e3d426ab533bd18d62c9aa142eedbde394bed07db58313e0fccc82a23237",
+  "blk.22.attn_v.weight": "26b5be1fe3c2b6824c5a648a3e4bdf17691904526fca158fbc3ebb627b67e2f4",
+  "blk.22.ffn_gate_exps.weight": "32ab7a7735313d60f6a75229b1aeee940b6aee176c9648536bf5921b0dc2929a",
+  "blk.22.ffn_down_exps.weight": "67590808f6a67777d3eb7976c31fe616d388b98fecbb12253b72d1241d70753f",
+  "blk.22.ffn_up_exps.weight": "fc245c0183e6d90829ff5e71a4ec93e4860b3d4c1a17b9dda2fb64f5f5c9ed32",
+  "blk.22.attn_norm.weight": "128e99d206d4d6724758ec97468af767fa0aea592149c324b731659c1e74a1a8",
+  "blk.22.ffn_norm.weight": "e45f498033f0cffa15da0eff2c47b4472e43fcf8921729fc4eeb2e3a6b3c78e2",
+  "blk.23.ffn_gate_inp.weight": "d63e686f5325fbc89fa242c2c52a3b8ff54f867dca914c9ae6eea13e9d6f46e5",
+  "blk.23.attn_k.weight": "f71f5a577f46ea12b1818f3a5ff4b85ddc45f9a2afb0fa2e041d71a3e31c6779",
+  "blk.23.attn_output.weight": "92b13563c1e0eac0d748fb67b235dfd7a64c8f16e2dafb316885744582e23b4b",
+  "blk.23.attn_q.weight": "2f9b9c35dc4f912f3f51c06e2d68f417b51a0de0a84aac530a64f9d3d7b0a2dd",
+  "blk.23.attn_v.weight": "268e40813806e74a5c364b19556d087bf8374e76e7b6fcf55c381eb7da13ccd1",
+  "blk.23.ffn_gate_exps.weight": "12f857e7a7ce228afac34d99b602c8d6fe96984f2a21118f459a58cb767ee65e",
+  "blk.23.ffn_down_exps.weight": "cdb082c16599c3bb36a28066dcc122d9529b54fa91b6cf0153437ec960a5e16d",
+  "blk.23.ffn_up_exps.weight": "f4b99f6f44d7b8b5a305894e88633bf5938fc1f6303a2b2092399da9c8b64d7c",
+  "blk.23.attn_norm.weight": "a691392210383915916b4d3886d5e4d56e7855e27e37e414fbd73bf66b3712e6",
+  "blk.23.ffn_norm.weight": "0c3dc72f667e5ae19b69bfa9f2bd2a01a57681f89ef9527bad4eb0d8c7b70da8",
+  "blk.24.ffn_gate_exps.weight": "86baca2a3157994df7fd8ced5e08436d5c1810dc29c0715637c36de723e0e7d1",
+  "blk.24.ffn_down_exps.weight": "ac5d559562b35c34993e34b071f66d15c65be5907797078c2d2a49aba54e3192",
+  "blk.24.ffn_up_exps.weight": "fce0a099cf09777f44fbab3606ceb75f7fae6f0b80725f9e871654b8cdf9262a",
+  "blk.24.ffn_gate_inp.weight": "e7c6800c0cfc56b565b2d35ad6f1dbfdb70dd0b05b338bc8da2286ffc3678d79",
+  "blk.24.attn_norm.weight": "dc6cc18ec52d102d015153c4a1132f9d7a504e29cbdec81c5edbf3b9e65815e1",
+  "blk.24.ffn_norm.weight": "480d5a1397af5e0e657f1e67d20ec0cdef5724e71246a326843321b87ffabd33",
+  "blk.24.attn_k.weight": "338c0597954a9b95a782545b2fe36469553e73f86ae2d2b5697767b28e1c7daa",
+  "blk.24.attn_output.weight": "a77d23b79933c67e52f1eef7f83a3dff4f767ce0bbcc39572f8cec4acd457643",
+  "blk.24.attn_q.weight": "45c9478593002be1998e96e70668aafa2dd3972380fbc1df12fb05c24ba959e0",
+  "blk.24.attn_v.weight": "515729420885408a6a9614bc27cda393ed907521318d14d21335d39a3eff0b61",
+  "blk.25.ffn_gate_inp.weight": "aae4ac40e9ab3925241f9d784b54b38851d9bc999a6c3bc03fc3f17c9b28a67c",
+  "blk.25.attn_k.weight": "4ab4808d02396c35b00b426f536015673b71c17ae6cd55bbc2e6bfe7a4c59d0c",
+  "blk.25.attn_output.weight": "1990bb982b77e0c947cd1a8ef0b36227ee1259e6dbbc2829e5c136edf88675eb",
+  "blk.25.attn_q.weight": "a1490f3048e8c0ec8784f8550c43adf5cc8d0f2f90131c934713fe4b1b015bd7",
+  "blk.25.attn_v.weight": "f15e53c6d45b3b6f58808fa968425d65e0b26b7f9b268127a77abb1227c67431",
+  "blk.25.ffn_gate_exps.weight": "656662447ff54f56ee80f78a1b9483f7efdc40f7375d0cd8a9c72ccf21f77e7b",
+  "blk.25.ffn_down_exps.weight": "db06f101bccbaef19cced0f6c185166e18202465f4a42cddfd535fbe5cbabb4a",
+  "blk.25.ffn_up_exps.weight": "584a7b02456f27fe1d8d3c7ccd21d426b6ea887795a3ed77f704596a1e3841d7",
+  "blk.25.attn_norm.weight": "8f0f3597982930fd237e9d609776c64f2b909a455b21678f83a7ebd4bbb83e64",
+  "blk.25.ffn_norm.weight": "3e7079c32582afba0c55e032f254adc18d2997705eec860185e9a6dd3d82f07e",
+  "blk.26.ffn_gate_exps.weight": "e70341691b583b86489812b29b77aa41eb658b1865733d6118da54c66e3bfcc6",
+  "blk.26.ffn_down_exps.weight": "5c1b812d11dfb064af816ced5ab6463bf9722eefdfc341b8a93705d5038fd781",
+  "blk.26.ffn_up_exps.weight": "e18118362ae54ef7432781c83884f9fb230a9d934e342aabeda8822ea5f71fb6",
+  "blk.26.ffn_gate_inp.weight": "cd1c5f6710166b9567c6b74c97b2348b191c60aa860958c6bc264ab095261dff",
+  "blk.26.attn_norm.weight": "71d087531af2520bda2e676c489e8529cef5db8aeea1eec0a937a8b4f2fa2e54",
+  "blk.26.ffn_norm.weight": "7f704e936fda28eb5c2cc339f0f6a5f78170b5aa43c01265b21668870d819c82",
+  "blk.26.attn_k.weight": "1cc62a0ce0ae251275d898c52c4a9fba5995fca10955d2011d10dd1a59e1afb8",
+  "blk.26.attn_output.weight": "636e881b1505f9cef656a4be98bec6a4765321d51f9bf1dac8933397cf44b765",
+  "blk.26.attn_q.weight": "89a3c4d202d7d6adebb9e0c1bcfd8b775f6456386f1be25e86e43acc949c1e16",
+  "blk.26.attn_v.weight": "ff2cc963b597cdf1a21703f3e7022af3bb4c65a34a19e19d9309a7c5e198b5bd",
+  "blk.27.ffn_gate_inp.weight": "6150139498fefe380bb99d11e72028da47a15ecb73dfc5b2774f726f4bed8f9e",
+  "blk.27.attn_k.weight": "f286eb9e5c56c7b801a497aedc40158c2a27877d7f9fb59b3fc67834798902d2",
+  "blk.27.attn_output.weight": "5dc3d3a05f9f7729509147fd09c16fb53f85f520cdab5cb69abf4bae3fd460c7",
+  "blk.27.attn_q.weight": "8462e40f86b24251960d6f35a9ea99b8793a01937faf1aec2859f2e5395dbb61",
+  "blk.27.attn_v.weight": "bac1a99e38e25953f8315f7212eb9777dc216cadb09b959977885ae62724ceca",
+  "blk.27.ffn_gate_exps.weight": "6a15eca7f0f6ecfd93db2e55c63875348ec4a78c4ff643ec46df9e958c0101e4",
+  "blk.27.ffn_down_exps.weight": "2e1c91247c4359e2073a8e5f26fd7f6426da7be3ed5bc65dcfff701f0a5022b2",
+  "blk.27.ffn_up_exps.weight": "65d6f5c553c9332085eae4aeadf25090b5d7768212ea7b08ed698102c21b29a1",
+  "blk.27.attn_norm.weight": "7fab8ae63ec8e91ce625cd130ab96d8427dad3a7413bb21b25ec5f408c5b9f5a",
+  "blk.27.ffn_norm.weight": "532720546b0fdcd423a02ca6e3e9d8aacb84b1b3e8269968f88a47fe2a69bab4",
+  "blk.28.ffn_gate_inp.weight": "a305ea58d98962d9dcf0c53ad2389b7acc8936fb35a0e3fc9410e7767cd49dea",
+  "blk.28.attn_k.weight": "8315e8a2e4f78dfdf36d4fc18fffc74bc95fe42c3ae4f9af2b6c874612c0f71b",
+  "blk.28.attn_output.weight": "9b5fdedd32d39ef46a22cca7cd5355d7b93bd07ea305f466a8aad6ca5a4f3778",
+  "blk.28.attn_q.weight": "4e8fb96997c30e231c437130f410d7c91d541a816f6c568b5f3bfdb4b8dece74",
+  "blk.28.attn_v.weight": "1fec739cf3bd7b4913f72ca358d4cf31391c304de44ac0ae31ecb825beaa7cfd",
+  "blk.28.ffn_gate_exps.weight": "9f259789d535e09268266b9a8020f32d6a6779966c909d91d3a10574f06238a2",
+  "blk.28.ffn_down_exps.weight": "516d3f8abaedb01b9916a4b67d4672159769138ef2850158bc1b32c41e31f0e8",
+  "blk.28.ffn_up_exps.weight": "f2f1d88d2c31ed588806fb5ad981d68f5134d7284c4fc022fd018de2eef437fc",
+  "blk.28.attn_norm.weight": "960fd005598deadaebd969996f4367a9dbfad90539a863674fe95730935acc64",
+  "blk.28.ffn_norm.weight": "e1993b37ced93d4049e9af2c47b0d9207d8f7e6f2cc3a52f57bef30bc806d805",
+  "blk.29.ffn_gate_exps.weight": "58927146338f443513337476b3cd30e6341742f096c2beb5890d400f10121298",
+  "blk.29.ffn_down_exps.weight": "03a3386e4f0b75a28c5608e23b2de8f0de25f21954e4aa7fc343431bde9db07e",
+  "blk.29.ffn_up_exps.weight": "6916b7490a7ae7b04a5d81cc1e7ac9b20c483434f3b186b12d87fe176bf1567b",
+  "blk.29.ffn_gate_inp.weight": "98e710e467a3d567abe4ce29d78b8e8dc033148762290c0c5e1ae4d78efd8c78",
+  "blk.29.attn_norm.weight": "4e64cb307d37be20d55f38c94faf7e451d11df5e60df347906cbaf9c5441be71",
+  "blk.29.ffn_norm.weight": "696c23a52f742679bd44440d687a4c44b4302d57f1e9dc5610d23374336187e7",
+  "blk.29.attn_k.weight": "e85253652fd6120c623634ba66b725bf7cd491318b54ccdad2c7df8851d64c0a",
+  "blk.29.attn_output.weight": "4f650a71efb150d1f24cd4d114d4187bf570ac424da3b92ea6455abdf1aea705",
+  "blk.29.attn_q.weight": "69fa7da901026ebcbbbc848455b425458b7e3295007d7fc093acf4b38e2166ea",
+  "blk.29.attn_v.weight": "17e2e7590b317b21f106de546aafd955579703d1e95d6aea044ee72ec3a514c9",
+  "blk.30.ffn_gate_inp.weight": "3a03284b4aa60d59d4a2ec86253469b61fc656372afca427cb77a5332fbcc62c",
+  "blk.30.attn_k.weight": "d518cfd0db9708e769eb1399e87ee49357dc54d5afdbac3d4c0ca46c64e789eb",
+  "blk.30.attn_output.weight": "9b44378714d784c5ef9ab604359091baca4e0ec222afa139b7f840eaefb371fd",
+  "blk.30.attn_q.weight": "cbb95365bbfbcad0c9cd99b4eebb5a5d32de68ce08e4063b5ec3e792b7548044",
+  "blk.30.attn_v.weight": "e7985c04fe1740e35a9598f43b67b0922b4fc2d00b68a92a9f917b82c3248de1",
+  "blk.30.ffn_gate_exps.weight": "8ac4bbd07935d98f895ba94dc174e5ad5046c3c222b53729d60f987c05e7eb70",
+  "blk.30.ffn_down_exps.weight": "dd672cc71e82abf05064a18121b8e55fe1a4f19bc1d7cb9a142f4add54bc336e",
+  "blk.30.ffn_up_exps.weight": "12282f664a2a12aa25e2deac58946108715ebb978bafed5274cef24569107646",
+  "blk.30.attn_norm.weight": "1a33458fee054c6c9c896a4bb0a4e1fbfa0293b2408c7dd2b81d692e966e7273",
+  "blk.30.ffn_norm.weight": "311e33b68051f507f1478ed8f2693fddb846170ddb7285a91be43f795c2ce31e",
+  "blk.31.ffn_gate_exps.weight": "8af43d9867a51cd8392fb48b981b0ceee0ae979c491c07d711b3b56b5162c786",
+  "blk.31.ffn_down_exps.weight": "5579cb7758c1600b19d1f540deffe081b575962e37437b3b2efb2fb0a2924e40",
+  "blk.31.ffn_up_exps.weight": "f2e7c005276b3a001fb40753f027fa10b4d5a346f43cf4b4bbdeec6e74e1cf6a",
+  "blk.31.ffn_gate_inp.weight": "89885dc0e30b6b16a90c0331d7fa3174671e941364e8102d934f02132237e61b",
+  "blk.31.attn_norm.weight": "99e4e9bf86a9edf8c404153a7e8a82324ba79da462622196e2faba161bd95172",
+  "blk.31.ffn_norm.weight": "55335997cf6de781bf332b943de96ff4646966b05d9fee86b76ea897e27b6ca7",
+  "blk.31.attn_k.weight": "cee570762b78da6316b637892cc4b080e40f57af5551ffb1866b9a8e80e96628",
+  "blk.31.attn_output.weight": "fa321ff55ec7819ead7b819fd45215262f39744569765ba2113c989c03588802",
+  "blk.31.attn_q.weight": "9e2c409b878f8a2a1436874abf428fceb1c534b21f9ad4dd6f532b8a469007f0",
+  "blk.31.attn_v.weight": "a845d0be68ba537b4a775bfba4d897faf7c82a811a2612b0b7420cc4f3574cb8",
+  "output.weight": "16101cbb74b54cda9ebc07ca3c762e3263a56efb3cc011156184b95807d7cf13",
+  "output_norm.weight": "d7aa61585baedd60157aafe157930785742c55989c288573566a971b02423564"
+}
diff --git a/convert/testdata/Phi-3-mini-128k-instruct.json b/convert/testdata/Phi-3-mini-128k-instruct.json
new file mode 100644
index 000000000..19296f5a9
--- /dev/null
+++ b/convert/testdata/Phi-3-mini-128k-instruct.json
@@ -0,0 +1,225 @@
+{
+  "general.architecture": "phi3",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "phi3.block_count": "32",
+  "phi3.context_length": "131072",
+  "phi3.embedding_length": "3072",
+  "phi3.feed_forward_length": "8192",
+  "phi3.rope.scaling.original_context_length": "4096",
+  "phi3.rope.dimension_count": "96",
+  "phi3.rope.freq_base": "10000",
+  "phi3.rope.scaling.attn_factor": "1.1902381",
+  "phi3.attention.head_count": "32",
+  "phi3.attention.head_count_kv": "32",
+  "phi3.attention.layer_norm_rms_epsilon": "1e-05",
+  "phi3.attention.sliding_window": "262144",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.pre": "default",
+  "tokenizer.ggml.add_bos_token": "false",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "1",
+  "tokenizer.ggml.eos_token_id": "32000",
+  "tokenizer.ggml.unknown_token_id": "0",
+  "tokenizer.ggml.padding_token_id": "32000",
+  "tokenizer.ggml.scores": "6e37bcde2adc7e350e87c496eddd7a2124329c1dc66c5bf3ad3997253e4f7a62",
+  "tokenizer.ggml.token_type": "b6ecf55ec64ee67d87750bdb8d757a2c58bf78377e9f4219f5689a6c4dea57ce",
+  "tokenizer.ggml.tokens": "d168da3ddd3eee820916945fcb9baf24dd3cde42f606cffa2d19e7c8a8743918",
+  "blk.0.attn_norm.weight": "216aeb2c9e0c271f899e1ef2a63cceeb8f41e97642e84fada54b1d3c1c11cf25",
+  "blk.0.attn_output.weight": "b597d56f7188ffc1fafc273fadc59d41738cffd677ae98c61a62c3285b3a3099",
+  "blk.0.attn_qkv.weight": "d28a6b44e13f59be5483e4be2bedb544e346168d720aca27f47d1a5a722be91e",
+  "blk.0.ffn_down.weight": "4a691370e5a61fcbbf540fbcbf4c0f1d15dec0364528c0e916d0744f6262b63b",
+  "blk.0.ffn_norm.weight": "0c00af2b4a3128bec64a0cbb1084b042fdbe13d9ad0d03bd577f9449dfead338",
+  "blk.0.ffn_up.weight": "b32b52f790c1c083bfb8a3126dc1111cfeeb28dc8c584a930a1e5334cb176bf4",
+  "blk.1.attn_norm.weight": "68748011503c6c029e8e69a84a8e5a89338f378769627b6dbf7f93d715c292e1",
+  "blk.1.attn_output.weight": "2267344add13b048ca59e4377c86dc512be8046a57156901fa32a20fa74e4ee0",
+  "blk.1.attn_qkv.weight": "9109d2e3d7a2eacfda5226587b8be124a3bf44b972da7ebb17aa15795897eacc",
+  "blk.1.ffn_down.weight": "d675df4df4dd039c0c339ad6445d39eddd2004db6bf35bed6314c7497245a633",
+  "blk.1.ffn_norm.weight": "3b5767ae977bc8baaa06b06efdbea193b6b3ba605ce76d77a76ce317e935500c",
+  "blk.1.ffn_up.weight": "80dfd6d9d234b00334c89b8e0a02f81899c2efd377321c34ba5ba51a5f61b5ff",
+  "blk.2.attn_norm.weight": "6a6743b057e5088f145bc179e92c9bfb41163e7295d7b81c62e23dd89d2b59c4",
+  "blk.2.attn_output.weight": "bc5491ea54e0db81462d7d9b7d25cbdda380c2db8de041bd1c4ab7b76a1d19c3",
+  "blk.2.attn_qkv.weight": "a61287a9852e2f5aca9c100b471d98398b2913a3497c743de3c70ec9ddd7087f",
+  "blk.2.ffn_down.weight": "4fddcc382c8dceeab027fe43d8d44e67edb5e8ce4b9a1b7f773c87770380ade1",
+  "blk.2.ffn_norm.weight": "07e05f82b3f63f711db3b684ca79aed25c0657917e66f88af47348a82065c227",
+  "blk.2.ffn_up.weight": "4835a682ef1826c12df01ae7663fc45f9c82bc8e64b665f13fb7da8e201ec0fb",
+  "blk.3.attn_norm.weight": "f22aba7c03999ba7136f39cda747a39715e498699dc1716cd97fc5dfc58d1b1c",
+  "blk.3.attn_output.weight": "53b579855366fd786c5126b2b30aac4d583ca7bda56833c4865f5cadb5c18c6d",
+  "blk.3.attn_qkv.weight": "bb56aba78158123140fcea59c69ac562ca208f6d3086819417cdad8c50f333ad",
+  "blk.3.ffn_down.weight": "97280897a7cd86db2830c004bccc5bc094f50e293baded0189159a2019145a6e",
+  "blk.3.ffn_norm.weight": "10a8c99f8b57a960e8e0a1133c4a26f9148403d1b9bff2eff114917de996f3b5",
+  "blk.3.ffn_up.weight": "7324046c915e75d621b2043597a245a428d8eea31869135e6257a861491d8dcc",
+  "blk.4.attn_norm.weight": "507d8e164de94646edbfe33def8e8fbf7c9a6ee3fbaedb5000f72d9f51ec5e36",
+  "blk.4.attn_output.weight": "bbb3429e6efa98c150e0fdbf48c16180cbf0d0cbc1b3c253c6c319d78f4593a2",
+  "blk.4.attn_qkv.weight": "b95ee5be0786d3901273d806c339fe6c20e6bfffd2a20672a9f56af80921e8ab",
+  "blk.4.ffn_down.weight": "806bbf91df92a5a22bd5aa1ffb7fc2869f7293ffc7704771c290ecc583b27975",
+  "blk.4.ffn_norm.weight": "cfc2930a81df7aee3a5e7f726a15c1182233e868bf0d9d37f6b6ae6d8c15c234",
+  "blk.4.ffn_up.weight": "c3390c69533de2c8424e8069323ccc5d0c4543111535da04cf2c7d26745576aa",
+  "blk.5.attn_norm.weight": "0d71c4fbcefabbd021569442853d2fe90668b19409ae2805a718a829ca60beab",
+  "blk.5.attn_output.weight": "10ebd93629112bf2df5c30dd0953a4a5e9020306768283181ed426934d47e14f",
+  "blk.5.attn_qkv.weight": "5cb05633369f12d4b00e0ff787736bd846856682115720ebc6cce05270c334f6",
+  "blk.5.ffn_down.weight": "e28bcc5094212eafc7476dbc5b7a520d25b79578cbf4229d698e2655956a80ad",
+  "blk.5.ffn_norm.weight": "b6f2c4cf9f34bb4d59989f96165c14a67dc1e266ad0a6d0fcc49f1add929e6ff",
+  "blk.5.ffn_up.weight": "0f9ef99423cc07ebedc0e9cfa95809f2d7108d910bb4ef97ebc0b0309c440750",
+  "blk.6.attn_norm.weight": "b3edcc47a42218234f7564d7470611b49401a41ae8cd42123f86557c69f5d7f2",
+  "blk.6.attn_output.weight": "eb9b7d257b388bb5b8fe0515e5c6873317239cb94cda236e4b6ada2a6c57c65c",
+  "blk.6.attn_qkv.weight": "eb968081f478c52f07bd9c2761741e982dba33cc4eeadeea3557d391b9ac2106",
+  "blk.6.ffn_down.weight": "1b8588bb7463206290322695577dcfced300895d6e6f4b26966c53a9ae2f0f84",
+  "blk.6.ffn_norm.weight": "1219c04b7770983c77814200eefe743f46d15328ea2b12711e44f8103eab08d3",
+  "blk.6.ffn_up.weight": "197ef287239fec47c55677f0fbb66eaf0644f775bc382de843971730721394f6",
+  "blk.7.attn_norm.weight": "b630ad08c80d564ed1c024384818e9fd3f22a36cd7a14aa96e7e2759a8285099",
+  "blk.7.attn_output.weight": "970255aa750828a47d6b9d399f9612b5bf25aefe7dadbcba41fc416d0d4067c1",
+  "blk.7.attn_qkv.weight": "ebb157c880293e6de8d629f263ba8853ed1dbdc02c311d43432bb8cfbb310739",
+  "blk.7.ffn_down.weight": "24bcd4db4cba844c89f878b81843c373dbbc0675e889d32c5b12e63384a7b670",
+  "blk.7.ffn_norm.weight": "b9c6f71001808ee873ce7db8056e4b53fb4cccec8b7f0f312899b575fae39d39",
+  "blk.7.ffn_up.weight": "979f1828d227455c26015a2a11afe9dd05f2bb97a8ba6b38c8dab3f50e627401",
+  "blk.8.attn_norm.weight": "4e8e347e3775010b7112ee630f2f4f2383be7ff64e6ca6154b9b22566552eaa6",
+  "blk.8.attn_output.weight": "65a44babf44a435a1829945211b3168f9ec78ac3cb7a049a733e93d11f0d6659",
+  "blk.8.attn_qkv.weight": "343ed07671da400b040812a4058482fa38284b5d9af9becfed07417fe26ce747",
+  "blk.8.ffn_down.weight": "7fb7e073e3c2c503c4e9d60efa0988fed7398d900cc003695fe3fffd3e188b82",
+  "blk.8.ffn_norm.weight": "b07c1f655d8593e3892a2cf73f8a0c19ce8e5cb613fafbe7cbd430da8ce4c57d",
+  "blk.8.ffn_up.weight": "8b26e14de54b3fdc2e2d3ea41720f9d9c236a93688c3b7fd7bf43f5fbb327c9b",
+  "blk.9.attn_norm.weight": "46394d408a8e316916177e6aa261de32e137a82d729c0b1800b072f0c38c39b6",
+  "blk.9.attn_output.weight": "d57f3d46107947a7073373a0b35d6ecf7759b5df15406f4a3590a60666af6b16",
+  "blk.9.attn_qkv.weight": "14bb8ace8c5453148f4b536e9f4279c813f31136716947256f5cca333448639c",
+  "blk.9.ffn_down.weight": "2b8d98e2b5ed68338f6e4de43bf7de0c4858cc69103cd5177725f7444eec7694",
+  "blk.9.ffn_norm.weight": "41a499dfd418cc4c6b8c12313f673f7e2cd4a3f9c4065eb6c4feb5eed02fb542",
+  "blk.9.ffn_up.weight": "143aab7533a64b17fbe201490a6f674bc7f0bd370c094500b2e100419073d1c2",
+  "blk.10.attn_norm.weight": "ebb670aafd36816a794347287269d8f1a5b19c1e3c0a1e38023bc19fdba9b073",
+  "blk.10.attn_output.weight": "b5d65bbc0ed5e49fdd9d754bc18163cd042a285024d0cf6f954c503bc8c877cb",
+  "blk.10.attn_qkv.weight": "f06b15bac88da798fa34a62b03eaac0dbe8b846020516603c387541f2d8dd672",
+  "blk.10.ffn_down.weight": "fb091fcd1b4de25d1bea94d1755e255cb02914a030d23e3a234e57b8d46bde6e",
+  "blk.10.ffn_norm.weight": "eb347bdf9c40414af87e13a8e72e40b31f004b50f7cb366f1a219ced60a61355",
+  "blk.10.ffn_up.weight": "ed2d52fc881a173f404fe8a1067862c9856d6c3e0d2e90a330a7aa394e3f84d1",
+  "blk.11.attn_norm.weight": "64e252603cf010a0e502ca39fdf8d0a196a79aec67c0d2bb9213fc0cb80c47d4",
+  "blk.11.attn_output.weight": "228e33e21c69f52efc74fdfc831bc9af271e44b2a29a3dced1d64e667ce36eb5",
+  "blk.11.attn_qkv.weight": "ab9ce6d4ef9e42ee0da3f20a7708a3bbc5e79e967b05fa86ba946a05e2eb63eb",
+  "blk.11.ffn_down.weight": "0ca133b7835c98dc77c25d64e4eb7873778bdb5e4d22d8b80f920f46865b43bd",
+  "blk.11.ffn_norm.weight": "02455741a0dfd161c79aa1ecc381901721f229fdcda5615622a629631fb61cfd",
+  "blk.11.ffn_up.weight": "9fecdcc099fbb8e23c6b1ea9294702a027f4a58d265543ec5e7be79b8f63b354",
+  "blk.12.attn_norm.weight": "783bb459911b1b3609a9b2bdfe272f1670add73b5471da738e07ac47e2e07dfd",
+  "blk.12.attn_output.weight": "1e1a914c9e48b857206ac5a1f7cead994bc1ea91d5d4fff8c834d73f2e38ef5d",
+  "blk.12.attn_qkv.weight": "5953e7185ccb87fb4dae8f9426ec86315d4c7794326e8ab59b3a95d4af2189f0",
+  "blk.12.ffn_down.weight": "a3eecf0f394f86e2cfb48a5940a5c50ca86d71883b2f79fcc642a935fabce0d4",
+  "blk.12.ffn_norm.weight": "0a4272e41373c23bd72f10d2d82930aa3a1480aac75832bfbf01cebf0b86b6a4",
+  "blk.12.ffn_up.weight": "06f42776de3a7ceac3025f26a7a8bd20e062233cce2bdaa2183470dc4b30b87d",
+  "blk.13.attn_norm.weight": "5915da60fb03e201fa649faba780e5fdf1c761c262b206e5415cf83181f65780",
+  "blk.13.attn_output.weight": "4dbf6eab074fa3835fd32bd631a8208e511037d5056d2fd3015735cca7674ef7",
+  "blk.13.attn_qkv.weight": "d3d8339a1c4782d9e73d77fdebe154d3c5b83ac40c9175b3e91a4977d08f876b",
+  "blk.13.ffn_down.weight": "de6772b46a55e1fd42b007637dfbf68b6598e5d5b61622da0935002e1e192d3a",
+  "blk.13.ffn_norm.weight": "5a640ea3b8c7be49c95a58a2327e10d8e8d9d142504bde5c8091613e5b961d7a",
+  "blk.13.ffn_up.weight": "f35e3545e4bd3531b2e843b5efd31dee0c13c807ee6386e65473ba67bbec30d0",
+  "blk.14.attn_norm.weight": "9b34986450b7c98b4927e81e61a816f9e84b1addc7c14926402100037aad6678",
+  "blk.14.attn_output.weight": "155d52efb23d366016d861a251d4d1f4a0c13699188c50d50dba016a0d8bfcd9",
+  "blk.14.attn_qkv.weight": "8e1415084e1f33c73a777f19e752489f4dd312cca047733e5ea643cd4a955e04",
+  "blk.14.ffn_down.weight": "a2a142226b94baa01ccb65bdea2b7418e49085c1d9c3c63e544e3112c58a25da",
+  "blk.14.ffn_norm.weight": "8aecfd9b0ae6affaea31a80c5c9a4a14b31deaa0db7bd8f6da2a64d23447921c",
+  "blk.14.ffn_up.weight": "0c1407237b8c1bd02f193346b5681926fe698a5055eac6a7450451b0f991707c",
+  "blk.15.attn_norm.weight": "e037bd19880bfa83d983200fb0c7866f8ad16c3ff5cc4b4f3a37ca7373870ff6",
+  "blk.15.attn_output.weight": "045fe4fc95cc129a1b92771b179c11b12845c4c088786c607f17bd98857e68e1",
+  "blk.15.attn_qkv.weight": "7621b7559705cab1d4dea1c69f76dbf9dc1c8837a203b656f484703b9c1b70ce",
+  "blk.15.ffn_down.weight": "7e5ac20e290bc60761e1cd972354fde225b7fa861048d44d9a0dd9b046d55f58",
+  "blk.15.ffn_norm.weight": "b6d830d88f1db1825687973c8c2b1a24c6fa84f07af8d0e3ef9c86009baca0b2",
+  "blk.15.ffn_up.weight": "dcda0957cd04fc45476774dba2bbf9aa89d6b05d5ca7b10ae6f73ad2c49b1cd3",
+  "blk.16.attn_norm.weight": "4ee9b70ba15cb2a08240f93990e90f5068c48fceb481f8e2186bec8b7214eb3f",
+  "blk.16.attn_output.weight": "315cfe5536658d2498192b2980eade15b2c9a4ff220e4011911457b1727fa103",
+  "blk.16.attn_qkv.weight": "3c8122e3ad637583b9dcde8ff3a323267d3014bb1f0f9771e5322260ca9ecc8d",
+  "blk.16.ffn_down.weight": "3b5fbebd5ee2b86cad96fb8a9b45a8770d08f82c1c8b74d7061e866f7020a18d",
+  "blk.16.ffn_norm.weight": "ffab69f20bda372de6e5878f0539163e2fc6ba113621ded95705fc3b1465c9f0",
+  "blk.16.ffn_up.weight": "0935ea3d258da42d6258406365f39f58ddaabfe97ea5977580db3635188f24a1",
+  "blk.17.attn_norm.weight": "f030441733f3d147b4a06a1eb4aeb8465c7c24d9c53bf4c48fe7e134d3629803",
+  "blk.17.attn_output.weight": "07a955ef09e8dc766ac0df647d0b2c69f23c4c69a7137654b4aad80303ed0eda",
+  "blk.17.attn_qkv.weight": "1c10688061e21e2fe12ad0cb54bf03895c1f83c3b0df743a42f548b52cbca1b2",
+  "blk.17.ffn_down.weight": "ebb9cc9836f41d88fdae2aa9a4355514e4edaec8d1577ffeb947a35204e77f52",
+  "blk.17.ffn_norm.weight": "50aff44f6528b13db5389f2ddcdb7676244947610bd7ffbff3f881c968c2a0d4",
+  "blk.17.ffn_up.weight": "d716537949582be33bde6b02e38f5a70081c9642a9fb05a61312126718b8d148",
+  "blk.18.attn_norm.weight": "0ea695c4e53d637902f46663a6ee42adc493c36794476acc7dbddaa05b13840d",
+  "blk.18.attn_output.weight": "5fd35b500221a612eb4f4bddf0e9b6b7db4d7733032a75f8802fb2d884647c2e",
+  "blk.18.attn_qkv.weight": "b0da37fd030fe69581f990bf23bfd35467a1bbe558af6de7c0924f6b72e92317",
+  "blk.18.ffn_down.weight": "b355c33f44b328f4bb977567de8f7544db4b005d7a8fbded658518ecf3c5a153",
+  "blk.18.ffn_norm.weight": "58b3fe9094079989a86e0387143259e1cc35952d24dc3df290c4ba6df44f5c51",
+  "blk.18.ffn_up.weight": "2ce530954c342c30ed2ead5353f931960bfae1d278868504c0efb973560fabbe",
+  "blk.19.attn_norm.weight": "533e9aed66feea8f0392aa81f9e293240e1f009a5334253915fb60c2749b615d",
+  "blk.19.attn_output.weight": "84f2d00f98a4113a779d3b5d1c3e7c914eb47784d3ab13b290367c124c2994aa",
+  "blk.19.attn_qkv.weight": "fbe6b9f53b07fa7537d3b3d452d20a9bc666f9fd41ec2091dd28bc2f70fc668f",
+  "blk.19.ffn_down.weight": "b30199e098c8bb3f890183d8b18471e80b62b604729b277ad62488dd71e1206b",
+  "blk.19.ffn_norm.weight": "c81373e41cd340b7badb19f9517c77c4250b4eb9a02dc758b8b49b652487d7ff",
+  "blk.19.ffn_up.weight": "5a5cb083ca7725720e3a890f7fa46354760e8007a8188849a092e305694a75e3",
+  "blk.20.attn_norm.weight": "4953091b4477e354357a8e743ba0a1900633e52f1599ee082a0c9b0b2b5cd978",
+  "blk.20.attn_output.weight": "62d54f7749cd6856097b2632066a322b0296df915fe66f382c5b5981be0d4f23",
+  "blk.20.attn_qkv.weight": "406de9e35b0729ebe902d7a47905cc7fb29a921431ed35dbef0c03e5690a1329",
+  "blk.20.ffn_down.weight": "62fb678b0d1261e19a4903a2b347d67afcc8acff01feb33a687a35a2d1e6f9a5",
+  "blk.20.ffn_norm.weight": "cd9d36b7e71e55c8925b97bb09c28219f182626bcff094878ae39c3db887a14b",
+  "blk.20.ffn_up.weight": "b9276771d79d3e932e73ccc520c3f8476342b9ef312ed2ee1e0da822e6e3ad18",
+  "blk.21.attn_norm.weight": "66d8c8a35e13ce9c2a0e75b670150e2c31484a55c2316df46075312196178ed3",
+  "blk.21.attn_output.weight": "12ab46c9382648f9b3350fdd92a6be6352743d62d6b520d7e2024e0c838588f5",
+  "blk.21.attn_qkv.weight": "a7909676ee1675ca23cd29a5fdd226df8dd9d68f94c6c9bbb51dd9fd38504008",
+  "blk.21.ffn_down.weight": "6fb317279c6542e82f97d5a12a60fac1bd0fa0405154f9fbe265e2fe39bd49cc",
+  "blk.21.ffn_norm.weight": "c0f703eb3ff161b5ba4490d87d8684b8a6c47a8f433e12f418333b9db439010a",
+  "blk.21.ffn_up.weight": "6dbdb80ef0c35e364bbce12d40d5e74c7963c7b55d58d9579567a07ffce7b863",
+  "blk.22.attn_norm.weight": "f94237433bf03d675cb2f655b81ca91a1ce2447bc6b00b13d6b0ccfe2d411eff",
+  "blk.22.attn_output.weight": "e821f95995ce497c01e63ca64f737713b1b65f11df1903e51d444aa516f33f71",
+  "blk.22.attn_qkv.weight": "1b0f717c73afb5eb4c82a1708c4e85c969e8a2a8770d9ddb78b1870a2d8a781e",
+  "blk.22.ffn_down.weight": "0f33f7a3cdc685484be99aa0c03642b0b20850a27d1fddbe054b13a9382f3ccb",
+  "blk.22.ffn_norm.weight": "9df285cf211ddd7df2b36a50489af574755c7d4d98b29a05cd04566ae613c8dc",
+  "blk.22.ffn_up.weight": "63ac300e1efb34041dd0136cf43ea622fac6f0caccce1cd9262f5e08d2cf179c",
+  "blk.23.attn_norm.weight": "5f72d9e88689b4027b28f5f8f26cd3abb03635ceea7ec98a4c91a9fc691f6707",
+  "blk.23.attn_output.weight": "6ecf04ff61125c5fc768f8656497152149373daf321ee9c957e8f7245a1184d1",
+  "blk.23.attn_qkv.weight": "a9d9978806724c2959f2cf386c233831f08e1e933dbf2b32665e788d9d512ea4",
+  "blk.23.ffn_down.weight": "72c7d17886a3da17fa0daa456aa5e877b2ef5b8b403182b870d9ca5ca9c70347",
+  "blk.23.ffn_norm.weight": "971e4b712e3025a13419b5b57d674b5e4ab7f18f74b57b9afc4671623da90c4b",
+  "blk.23.ffn_up.weight": "df2b5c7dbd5834545b815073af0c7355b065124e6d6f0fee78d8fa5b2076dc3e",
+  "blk.24.attn_norm.weight": "c41957c4a79ad3b16f6e11daec1c7f530b9f3f4b618e1e4367c3b67787ac4ab6",
+  "blk.24.attn_output.weight": "ef7d61f5fc88ac6f31bf60cb5f4d2d6b8df42d38825807112361a7224b0dee3b",
+  "blk.24.attn_qkv.weight": "3e6a58fe7d49c90bb6971efbad3371c32256881173ea5aee4b0c296cb206490f",
+  "blk.24.ffn_down.weight": "f43619144047de42fed81dfa495f1815d3cb771330e574043e2b67620819292c",
+  "blk.24.ffn_norm.weight": "5501d4a2a98c8ca6b42e77b53b221dbc08f530f6a067256d787534ec6fe028bd",
+  "blk.24.ffn_up.weight": "d64c8b0e509e2b1118f6000176f8956cacecdbb200c7e95ed93fb78b6e26c84a",
+  "blk.25.attn_norm.weight": "502fa3c302d371f61c5791f4615b73018ffb1daa09b6499b227116581244c5d4",
+  "blk.25.attn_output.weight": "ad8391d4e9c980856f2547aa945b2b6a407a6382158dc1ddd4f08d94ecc24be6",
+  "blk.25.attn_qkv.weight": "42e8983780d4a01a02c54ad23d4df21eea437f119a10af5a9c12a76a42d308c1",
+  "blk.25.ffn_down.weight": "302dd010d4e0ab4eeaee89090409ea0dddeeeed3236415eb8f97c942497eea91",
+  "blk.25.ffn_norm.weight": "fb34c1ee5bca96986c08834df0a0c047ba041c1123ac1f563e9d64312bf82d6a",
+  "blk.25.ffn_up.weight": "10739a8de156816d93c92b935386540bfa976bdbef204f0312960f6fc657582f",
+  "blk.26.attn_norm.weight": "7036c711609128c4e55968ff3681d3043338879a5737efd6c2ac9e1a2a61f1a0",
+  "blk.26.attn_output.weight": "db5db45dead5cb911fa01da59832f121b7c18b2d167bf53741c40819f24d346c",
+  "blk.26.attn_qkv.weight": "cae34c6b7f82ed14348d5ed30a79919c383737c1694a9cb9c0de609d3b0c1d0a",
+  "blk.26.ffn_down.weight": "491ec3a4da9b4f49f8ebc6be658ce397a9b801ae9fb35e82177e47808c65e5d0",
+  "blk.26.ffn_norm.weight": "fd7059d75d7f0e5288511ddeeb0f772eb3cae3ccfe4226b877015834edc3c386",
+  "blk.26.ffn_up.weight": "ea1ee1274c56458ce056d2205e5bb6e5422ce4cb0ad58006b8141749b97a0c39",
+  "blk.27.attn_norm.weight": "cc362c9a937609265052cd38544af17a1a7448cea086d4c801139e1fc865832d",
+  "blk.27.attn_output.weight": "ba757a81dabde9cb1b069d1bb616fe79649a1724f756567ec61caed1304fe6cf",
+  "blk.27.attn_qkv.weight": "1ab8d7d02d87756c12c2275636823aa5ede3d683178225c4cac4bd892c319bd4",
+  "blk.27.ffn_down.weight": "deb1c711c8a66acf4dcd2d088e1548f8e08f296f755e4067d6557fa55afde88c",
+  "blk.27.ffn_norm.weight": "fc6242d8cb8a4a37a8ddb7e41e7e60a63d4a89edf36acb35df052f10b9c91ece",
+  "blk.27.ffn_up.weight": "8df39b09c4801f343aca78f2918a1f6db78c8c55e591eda4c69eadb74c26e180",
+  "blk.28.attn_norm.weight": "75b539308f77e3cefdc6d98484d8b5cbf0538f0c2869a77b7373a145a18bc850",
+  "blk.28.attn_output.weight": "ae128940eb60a6d2e121762ef4b3e9dcf9eb3e105b249507fa7f12de0e19822c",
+  "blk.28.attn_qkv.weight": "bdda781c288e9326c240e33905f8e621b6a2ad902e620739d34f93fcd6f933de",
+  "blk.28.ffn_down.weight": "f1d6e6d1c286b1138bfd7e53fe477f399ae93bc2c04e35416f84218ed7247965",
+  "blk.28.ffn_norm.weight": "3f837ce82c8b9bde0d61d08b6f5fe5574886ea5328dbdc53f2929f18da8b4087",
+  "blk.28.ffn_up.weight": "2af027002e31d1b6cfedbdb30a2b9d7213f3aa691167c353913adfd48fda31e4",
+  "blk.29.attn_norm.weight": "61e8003b5329462ffe0fe172f2b160260de006aed858332d49d75504b6b6aa7a",
+  "blk.29.attn_output.weight": "ca44542a72a37476dc73dbdcc01f5b7497cb3ebc4ea230a55c9634ccd8e56ad4",
+  "blk.29.attn_qkv.weight": "abb3d9d6abe57872ae3daa51935d43264093ded5ce63b49d1e280ee5758be0e4",
+  "blk.29.ffn_down.weight": "6764b895fce881df097489c263446f0106de36217997660c15984b3ee22a5a06",
+  "blk.29.ffn_norm.weight": "89e03e9a33fc0e6e31ba9f0c2bd7c5734a118c5602bb90148793e08a80e8d0ae",
+  "blk.29.ffn_up.weight": "fa7ad57a84954f4121653152efed1a871d8adb20a1ea9086e3e849ce359d7d2e",
+  "blk.30.attn_norm.weight": "91a697aca1e42af54f806a20211031c3369e8d0bd58df1b0147fe24954e1f5a4",
+  "blk.30.attn_output.weight": "36063fcf766c89ac75be56f688cc63cefe5f2c733fbf4378ea9956ad386fa148",
+  "blk.30.attn_qkv.weight": "2cacd1161f1121a2c0b979930134f4666f73fb8d7237b3b0659ae091b15955a6",
+  "blk.30.ffn_down.weight": "9f3fcb6217100595850c05dc98f9ab2a263afdb6ab28df2fcb08aeff512057d7",
+  "blk.30.ffn_norm.weight": "6c600bc1fc7de39d4f8917b81fc7d1d5ed2a9b56492234c13a4bd6028c30d880",
+  "blk.30.ffn_up.weight": "73cabd1bb011956b2689ea3338bb76642ef3a57c197377d666d2ab5f56317668",
+  "blk.31.attn_norm.weight": "72d3e1cc771380645fa75a899858c95f39857a4f3f1ed60fe1578df383b8bc53",
+  "blk.31.attn_output.weight": "40089cdd29994dc19a1d89fa15902a89cfeca3540f12dc9bf4d00ef82506e456",
+  "blk.31.attn_qkv.weight": "1d0bb40e9258071ae14290a53c619a8e331dda07354d2a02ef45766c029ae5e4",
+  "blk.31.ffn_down.weight": "8defa0e06335b793fa8be03883f0a322d6c5b33f52c69c943c35c60d16e42c0a",
+  "blk.31.ffn_norm.weight": "33c55d9d0c496ccfb130361fe131649346e098abaaac39c0519507e5d846721d",
+  "blk.31.ffn_up.weight": "599f6503f61c692c1f82001973d35119f9688db5e6be9d9c298411491c93f09b",
+  "output.weight": "14b8dc662bfa3308ebb2e102c562d8e52c15670e538f20f3216a9c310ca9dd41",
+  "output_norm.weight": "7f2294ba94ce65681df6c7ddd8698799199b9d77dc83c10bdad5c3999f0fdb82",
+  "rope_factors_long.weight": "e34d378664e354652c38f47d10dafb0498ccc2fb042d39ff7fef768146fff22b",
+  "rope_factors_short.weight": "9379146a4988f373d362fe47b06c75e7fe7c54aa4dc9558758df79b7a87471fd",
+  "token_embd.weight": "19a03c1fb5ac0baee93b0a7d8b0f26e9a9b011e229b694afc50ebfc13d84f8bf"
+}
diff --git a/convert/testdata/all-MiniLM-L6-v2.json b/convert/testdata/all-MiniLM-L6-v2.json
new file mode 100644
index 000000000..15c8f039c
--- /dev/null
+++ b/convert/testdata/all-MiniLM-L6-v2.json
@@ -0,0 +1,124 @@
+{
+  "general.architecture": "bert",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "bert.attention.causal": "false",
+  "bert.attention.head_count": "12",
+  "bert.attention.layer_norm_epsilon": "1e-12",
+  "bert.block_count": "6",
+  "bert.context_length": "512",
+  "bert.embedding_length": "384",
+  "bert.feed_forward_length": "1536",
+  "bert.pooling_type": "1",
+  "tokenizer.ggml.model": "bert",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "100",
+  "tokenizer.ggml.cls_token_id": "101",
+  "tokenizer.ggml.seperator_token_id": "102",
+  "tokenizer.ggml.mask_token_id": "103",
+  "tokenizer.ggml.token_type_count": "2",
+  "tokenizer.ggml.scores": "6db964fe67338aca57790481a390121ff3dd643eebe49f7dd308029ad99abb6f",
+  "tokenizer.ggml.token_type": "98d247c5404b6b18f05f133b92dd56edf6efefefac326794b00d7b351f6c5aa1",
+  "tokenizer.ggml.tokens": "9efe405e229a45ff9916f54c475d151d2200cd2ab0006f347abfb069cf096c86",
+  "token_embd.weight": "8c1ee80a9ea4f65aa385ba30112010068af3d209bebc6e149d3d4589c2cd0a5a",
+  "position_embd.weight": "6c516f0b1c4e2388ab90394dd80ad69e4e4509b890982fc3408108ae66210eb6",
+  "token_types.weight": "f879f8e422ed211948f28b560d3c5e17aae7993f063b51196a28cf5c0fb3da21",
+  "token_embd_norm.weight": "75076e095d717aab96f8b6beeee503c27940d9a76f2b891a0e3de72f8a6043e4",
+  "token_embd_norm.bias": "298735285ffe944e1bf03e5d35c7280326b85cf121bde9874f1af5dc51ab939d",
+  "blk.0.attn_q.weight": "ab0923ce4c1549175112dcdfcc860fe30137f991e03ea6857fb5993670adaf6c",
+  "blk.0.attn_q.bias": "a3ec29551dabf976e1d34256b8ab5ab7b758f3ed9742c3cafdbd984d5441df62",
+  "blk.0.attn_k.weight": "4c1038a6d035c3e9ffed7fa672b614627814752503755fbad0cfb76a41ad71ba",
+  "blk.0.attn_k.bias": "e0363930eb588d91816aa3d230bb03b6e2551c165117b80b8d60397413819ef9",
+  "blk.0.attn_v.weight": "425e2e53e3f00ce98d29c3e6a161eb55d3e6ae0d96fdb9f6242d1c4fd6eef4b3",
+  "blk.0.attn_v.bias": "6579173a1e65ee124fbd0bd53cbdca4225515b4f2c5f18fb1bfd000f5978f9bb",
+  "blk.0.attn_output.weight": "a6d70a08cd7164de5d12af65d86d657c3db35aaecde778b2b3fda9193c4c9802",
+  "blk.0.attn_output.bias": "2b8d12c4f9a9c5bfaa29c597839568f6e0525cb41eeaf64ddeb6bd84dfeb9701",
+  "blk.0.attn_output_norm.weight": "bbe6e502a473228b525aeed26cc31b7db123ad63bdc5a6eebac6ea70b8b51d62",
+  "blk.0.attn_output_norm.bias": "36eaacaf0007c5c62daea97aab0115390c0682914f78482e37eb76885f4b7a50",
+  "blk.0.ffn_up.weight": "24654561c76ce387d125759ba843f06b904ef721fcceaeff6ccc62180a48e874",
+  "blk.0.ffn_up.bias": "fd3f0126aa1d95768fa60eb6f4ab8a2763cfcb7e5405f35b92353031d86f4d34",
+  "blk.0.ffn_down.weight": "97a829763a6a5bf3329ceb4d39c424ba4787d61653a5b0bbd1f84782e4d4e0ca",
+  "blk.0.ffn_down.bias": "7aa980c30ae8b4ee7f69df28808dbf5c431f56ccc4a80340f644a0419f16c054",
+  "blk.0.layer_output_norm.weight": "ef30dad4c2a083ae1ff5039a2a6cda60ecc89bf1e486a6f8c0d15f50589603f8",
+  "blk.0.layer_output_norm.bias": "8b1b77e67568b1bce43fc476de1b177c53ff688d66beb66995e8eb3dc290da8a",
+  "blk.1.attn_q.weight": "284331622a1f6f9b87ccee4f652bd66a394ca493c4d93be4d1844e4f6159ad10",
+  "blk.1.attn_q.bias": "e24ebd4860330e08f6bfdd077a82db0bee33f4c8846cf1db26327a34754c7069",
+  "blk.1.attn_k.weight": "729dd0d555544b5bd0f7580b3c8b384256b974605f0e7487b95f295aa032997d",
+  "blk.1.attn_k.bias": "2aa51a828a858f35473f54477583fea54ce2ccc34ea60fbd1d228fbe9bca827f",
+  "blk.1.attn_v.weight": "6be304671cc311d5ca5c103f2b51467ee800c589bc5b8101e09ff5aed1f68c21",
+  "blk.1.attn_v.bias": "43bcbab78a8819e07f723bc9e5b737b71e87a7594f15234e882b63e327a64199",
+  "blk.1.attn_output.weight": "15ec8a1a12b26c9976445308a09f748ab0e4bef0f583d13ab08c3129f8738d73",
+  "blk.1.attn_output.bias": "dac2146f4baa6ed16f6c0dc7443831fb7ec79bedcceafd80d1a4b628a1bb072d",
+  "blk.1.attn_output_norm.weight": "d2151eb33bffac536787a4c9a5d2b31c7a80b17c4611877842a3cce2cd6e98d8",
+  "blk.1.attn_output_norm.bias": "31e1b779716dafb855d2cf5631ee168a0ccf372eb9c6ea6091f66fa97a9b9d2d",
+  "blk.1.ffn_up.weight": "a57547fc3fc3b77406f5cdcb0c87af9bc184701f175c39c1f35297826fce3cc7",
+  "blk.1.ffn_up.bias": "123be6d541d086202913c75d878c54d59a749f3af7b58f7ef9eb9e7c62a24c9a",
+  "blk.1.ffn_down.weight": "cfdb79788377e5cbded8790cd41b9e66c397ecab75474071fcd7cf32d30f9613",
+  "blk.1.ffn_down.bias": "bcb58315519a573097960891c9ae41cf4c685ab78c3e0e77471471758a7eae88",
+  "blk.1.layer_output_norm.weight": "819b554271452bfb1d84c2603b90377b2e41a0ac1e3aa8b417ccf9dce63375bd",
+  "blk.1.layer_output_norm.bias": "47a3433ac27f5ce8947fb38dd491f3706df4ef6adb0ddf74612bf0f54b19e164",
+  "blk.2.attn_q.weight": "1557a9ea852b1880551f7290e00aded4f35e6c4180fdcbed1b0039bf805f639e",
+  "blk.2.attn_q.bias": "c3bfe5f3066f655fd36b055530997b59ff33ef013563aaeb3cb8ff07dabd59a9",
+  "blk.2.attn_k.weight": "cfd08eb69c61ae2f9f14f9b7ff5c5394ca264b1a9f3d48156677f90dd1766289",
+  "blk.2.attn_k.bias": "9b839bc0e79974a0b3f5d1895972bc6f5c9a1bc16052e1af786e6a530758152d",
+  "blk.2.attn_v.weight": "02b26b1208480eaeeb00e7b4cf8b690006ca14759357fc44ed4a2a8924ead993",
+  "blk.2.attn_v.bias": "e7e6f0089fded1659a867ab736c220d9653ea7da6b1b94baf5c8d30a748b63ab",
+  "blk.2.attn_output.weight": "a1db121c7d33806b349cadd050300a57db49fdc91224fd07c9ac43bf4299dc79",
+  "blk.2.attn_output.bias": "7675128b6a92555cd955c820311e91e9417d31f48848f45d047b4100c62148b3",
+  "blk.2.attn_output_norm.weight": "5b4595e0fbcba67a700c4331adf746d2fba3546364a4db5607ae241947bb1a21",
+  "blk.2.attn_output_norm.bias": "7b8e16826ea30e5a2ba0b02e0095a901775981a296e98819625320e983060d08",
+  "blk.2.ffn_up.weight": "a0d815d946ac07a65095c4ae4df77b818845e6d97795c7d82f55e689d944db59",
+  "blk.2.ffn_up.bias": "ce37c0a4174d6bf773ded7bd016ede627ad3bdb8bc99b9992a18dc8e8898f252",
+  "blk.2.ffn_down.weight": "f6231d2a25426fbd45b9f1160aa484220eb227ceef0348c4a6a6de890606e5ef",
+  "blk.2.ffn_down.bias": "429e00556e8dc63a785238b309b9d83738500c1ef6d736fe6526ad88ea496d27",
+  "blk.2.layer_output_norm.weight": "651457a573adf3f7dd9ee5dfe1c8e89389e94443993aab77ec6a0b05aa621e35",
+  "blk.2.layer_output_norm.bias": "41fbbeda7fd89b0cef5f945ae44011c316982390401d6f75ba8c6d365e185247",
+  "blk.3.attn_q.weight": "95a43f32949d2cb8d22815bb27a44abfc6665ba96221af817dfe058cb6ca72c6",
+  "blk.3.attn_q.bias": "f4e34385e75d8108b6b3bd336106e2133a8c9be0cc343dfe5dc48c32a823c7cb",
+  "blk.3.attn_k.weight": "6b892da6a17d4d3265265a15f695864a31813ee8c8e710ae9bc9e1adbc6c9a18",
+  "blk.3.attn_k.bias": "40b8067b641a56014cee42548240aa8930820958b1933004892b5f04fbaef39e",
+  "blk.3.attn_v.weight": "9fcd5922319dd2a461082a5ce040c1dfe65d87d70ca6547dd0b46eeecc3eeb2b",
+  "blk.3.attn_v.bias": "b528c56212e66931fdbe267ac327a9c2f87cd03baff3ea719e30afe681da15f1",
+  "blk.3.attn_output.weight": "e3b178c1b03981e75510e0d277af23ea59cc404b5394e61bd32291825719b502",
+  "blk.3.attn_output.bias": "712c84d39a6a5a9c06a09da8fd9939ba0d5525524a4bba61ea4de09b48f45cae",
+  "blk.3.attn_output_norm.weight": "d1ffac88e675592ff72f8a617be32b4a381d443b2f8f2645dbe44a1e5745aac0",
+  "blk.3.attn_output_norm.bias": "ea31a1c73146234c50e0e43f485c458413714867b8e2703af66482f7db2d6c40",
+  "blk.3.ffn_up.weight": "4ef4f3b9a1ea6ab2ef2eb6e8b008e06a44790d099d97482a05a51e39a29afac0",
+  "blk.3.ffn_up.bias": "06a4296dda16f452675c51f108079fe7722552d6521c737d97734943818b9a2b",
+  "blk.3.ffn_down.weight": "f114b2bebe392c7d80433bb880c6730293aa4561b0b0370dcdaf7472daebd847",
+  "blk.3.ffn_down.bias": "2c8e67831d28a3bf613fc7912ae3259b63d72abcaf4d30efd8800758400158de",
+  "blk.3.layer_output_norm.weight": "a1dfeb7b5a51dd56447312ca41e2ad2f361a3ea12ddc355127f5f4219fb0a482",
+  "blk.3.layer_output_norm.bias": "1ed630021b25c6c6fc93fd32988b9907df966d4982a93081f639aac3044618ab",
+  "blk.4.attn_q.weight": "b5fae4c1f9a5f33a2a2e816ac0c01c25f422e4efdd59ef1ed93da2610e5370fc",
+  "blk.4.attn_q.bias": "c2e376524ea98ac3b10d9eee19ecb1b1e261fa5149efe0232844c923dfb428fb",
+  "blk.4.attn_k.weight": "a4632f5ebf9321d9d08f9112a4e5dda2efe5671df4a4e67fee24845f5b14af16",
+  "blk.4.attn_k.bias": "a9a02ffb8b8b4f6dfe487a7e0341f1d5318c9d2b793a688f34cb1b22fc66ef60",
+  "blk.4.attn_v.weight": "10ad8deb81d9fa093b1e5c0f24ea82aa7df43e6aca49e260fcbea56eab8cc86a",
+  "blk.4.attn_v.bias": "7326813e181e021130bd33ac136293fcffccce2d1d8cb59041e5b13a8cceacf6",
+  "blk.4.attn_output.weight": "c92573088c7437c2b3cda51490e152c27fb19e5468df591eabba5a49d5398d44",
+  "blk.4.attn_output.bias": "14e10b419e5859af1eb685af5c330aee67048cd704dcead9217840c6f5393222",
+  "blk.4.attn_output_norm.weight": "02b6831c0e0fb0edbc579a92812a1dd972cb15d14fcd382d4427c5a7b300ac44",
+  "blk.4.attn_output_norm.bias": "7eed5cd503bb6bb6ceb1bc8b07cc077903a4f14fb8b9d6cdf39644815ecf1374",
+  "blk.4.ffn_up.weight": "8d0c91d62e74d6431321116a37cf3339e630bd50ba164d3304fc4fe8dd831223",
+  "blk.4.ffn_up.bias": "d325f07f73c005a273c484c7be8e7abb4d6e8a5c4fd093f5869133b97629d017",
+  "blk.4.ffn_down.weight": "7ba7bd81143f40537b84f938e403e19f30e4928625eb371de052b9025beb4d21",
+  "blk.4.ffn_down.bias": "2853d9c2a75288214a4bf4907dc19d04d01926f4913d302b1aa7bdbfcce0f7a1",
+  "blk.4.layer_output_norm.weight": "a4ed1885fa77b90fed5300c355ef0aa0c876a8c747151d9d790939d464d57d4f",
+  "blk.4.layer_output_norm.bias": "62142a81e813a9e636333b2b805d6bc3b17c5e7cd4b15adce1ada6bc9a32563c",
+  "blk.5.attn_q.weight": "afc1dff080a72c3daad01384b1448d476aaf789871017c8ff8e144788887995d",
+  "blk.5.attn_q.bias": "748a820371c1d4f872c84545b36358d239c35bf6c99e2812c237d88c3292763b",
+  "blk.5.attn_k.weight": "59e30c1ed8acd2cbb01de5f62e7804015b9ecf98ba157d98cab016344639eda5",
+  "blk.5.attn_k.bias": "f839520078f9e589496e982e86d0126c7aa14196047339abffcf49a696229f77",
+  "blk.5.attn_v.weight": "3e21fb874e21b90308e1f46af034a3c32d3eba1628d62ae5f2246d6af5818923",
+  "blk.5.attn_v.bias": "5cd4852bf95c1444d10d756750f6bf49f842c0b39e9953c7f408bb67c325ac8c",
+  "blk.5.attn_output.weight": "636ce6a7752895f204b9d01ba0aedd9a294f908b42f372c22a16d9dd590d7471",
+  "blk.5.attn_output.bias": "82d924d4b0d2b94f2bbff91619216d6967a3541ce9b1531a6a60457a67b5d219",
+  "blk.5.attn_output_norm.weight": "5e7bd0a8d3396080f3360d7c4700bf094a06216431bd014c4479eef72ecf4271",
+  "blk.5.attn_output_norm.bias": "66c6de5edda5466d029c6753780be81ccd4218bf8bc00680000e0f06856ab712",
+  "blk.5.ffn_up.weight": "5bbf6e7ea380e216e33f8bee06d25f2265359d3876a300e92bc6e41d48e33430",
+  "blk.5.ffn_up.bias": "9d795388bb36fb33ad3a37fea3ccb4937838e02800a608fb47d363cd06b47370",
+  "blk.5.ffn_down.weight": "2fd628974e7f075479dd227b46fbd48ae8d3ca34d735b36f391ac06410730368",
+  "blk.5.ffn_down.bias": "cd213ba9eaa75fa541648097fbe9c96e58077e6c3ad6ad2fb1f21f8350f44291",
+  "blk.5.layer_output_norm.weight": "159a9df41d15b7022d136f86a2a2631c4635f9816e957472217077b522bcf52a",
+  "blk.5.layer_output_norm.bias": "24c1f27ffd1eb4e5be7e3a2909943e6f0980635d761fa1efdd0c19645da23766"
+}
diff --git a/convert/testdata/gemma-2-2b-it.json b/convert/testdata/gemma-2-2b-it.json
new file mode 100644
index 000000000..8f34e6672
--- /dev/null
+++ b/convert/testdata/gemma-2-2b-it.json
@@ -0,0 +1,312 @@
+{
+  "general.architecture": "gemma2",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "gemma2.block_count": "26",
+  "gemma2.context_length": "8192",
+  "gemma2.embedding_length": "2304",
+  "gemma2.feed_forward_length": "9216",
+  "gemma2.attention.head_count": "8",
+  "gemma2.attention.head_count_kv": "4",
+  "gemma2.attention.key_length": "256",
+  "gemma2.attention.value_length": "256",
+  "gemma2.attention.layer_norm_rms_epsilon": "1e-06",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "2",
+  "tokenizer.ggml.eos_token_id": "1",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "3",
+  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
+  "tokenizer.ggml.token_type": "8d40143b3477df77beea4139420335ede458bf5e14102f01b0170197b55da8d8",
+  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
+  "token_embd.weight": "64a9d30707e659e2e673656d71f5aef7a9fb9fd83bb9a77558dfc5abbe218a05",
+  "blk.0.attn_k.weight": "d8b4437c5edb3cddf6af9987038e1bb2b191c4f0fce0e160d2abace717f5d5d7",
+  "blk.0.attn_norm.weight": "1eb73e3f7aa8e502f6ca31cd19efbb8e4fd9a89692e13e48ac8205545a7fa7e8",
+  "blk.0.attn_output.weight": "39e7b78e57d356a22dd89ce1c4d7163b970712ba756545e1703f97866cd2192e",
+  "blk.0.attn_q.weight": "795058e23b6109febd9d55c89e1eebe6af0714ec8c56fd86a160876a6135ffe8",
+  "blk.0.attn_v.weight": "0cd6e583d1887c020472e961bbb113fe5a0d23ae2f1c2c876fc366cdb7692b52",
+  "blk.0.ffn_down.weight": "51eb4d962189e945a84e94e0dc1aad3f8f90cc1a11e18029670afcd0ea0acb1b",
+  "blk.0.ffn_gate.weight": "9811a29b8ad48432925897ab21dfcb13c5cbd372aeccbbefca9b7866883b4ce3",
+  "blk.0.ffn_norm.weight": "92cbf4652ef503c1de5b10f2be00b3fcf00100980cb3baa8f3013a8d8bf3d851",
+  "blk.0.ffn_up.weight": "af87de21746879483ed1b374cdd76b19ba11ca2b6dbb1beba98efdf3be3e8077",
+  "blk.0.post_attention_norm.weight": "32e135f1f258ffe407018899e39af1725d59d66d60022b9a21575ba160e0357a",
+  "blk.0.post_ffw_norm.weight": "ba286f5ac11b07fbc986173708c66f1920427be5a6d108af38fa0a837c1c8eb6",
+  "blk.1.attn_k.weight": "51584435552051f7fade76beca582b3f7190cf7fc07adcf527c2774d4b1c3901",
+  "blk.1.attn_norm.weight": "6833104c7fbf35a7e799ae56c262b97fffa14789642aee14381b25acd21ed80a",
+  "blk.1.attn_output.weight": "14c39481369087bf292ac9a3ab2ef166f9fe376a9f90c246653213ef264febdc",
+  "blk.1.attn_q.weight": "443f64ae2229f857c69d6bebb7800b685786cb77884c3ae19d4286aeed081325",
+  "blk.1.attn_v.weight": "0df482de2038f1e4c8a7733ac0ddb69ad90759dab5968b942af0155588de4c4a",
+  "blk.1.ffn_down.weight": "66f30763a8bbbcaea609a0087ed75fadb5e771c06378dd2cea94cf17e492e8cf",
+  "blk.1.ffn_gate.weight": "a7151bff00a545fa18b2c92dcd2a14572ccf9beb957a6c494f1374e8ebe174c9",
+  "blk.1.ffn_norm.weight": "e197d71ea11b5276bc0167d2663b88089b3ff42b47ba91e85f6c5d95f6306435",
+  "blk.1.ffn_up.weight": "57c182e0b14cccd1350d388f0c616991702e74281db54637451b70f4ccc24f9b",
+  "blk.1.post_attention_norm.weight": "3c56f837168d784c2d8bac247c130bdca6610c095c8da4558c536ccad7605609",
+  "blk.1.post_ffw_norm.weight": "d2a51d320fd01069dd7ccaa7082f16a7faeb671885607d7900b10a89c354d0fa",
+  "blk.2.attn_k.weight": "bc103c818192de7ce36caaf89dc117be4df13fb902e0bd9a23c64edace5df9b6",
+  "blk.2.attn_norm.weight": "0f2503aa126083a5d6ac72481be1ef66c6014705b573682b35bd864e4749a3d5",
+  "blk.2.attn_output.weight": "05fcd4a1226e482f91803a266f72caca887a93e63c2d2ba5611ab3c68d38743a",
+  "blk.2.attn_q.weight": "6a10b5c2fd423d1e4c4fd60fa8c154a0159b6b2501ea79cae2ef19f45a674e5e",
+  "blk.2.attn_v.weight": "3cf891945a1f8ae7cc908a5c6b729ff5b70f4436c5ffdbf245cc0ed4cc19cd1b",
+  "blk.2.ffn_down.weight": "ea204fd04e0d2fc728a9861a459216bbfec629c152004ba625f52cd8837bd51e",
+  "blk.2.ffn_gate.weight": "3a3518729f1b8b64a82b8792f33987db5418fdb094be0263c68f146a5c38de54",
+  "blk.2.ffn_norm.weight": "754ede678b725de41a34b82f0edf7688b5c065be7c0d46df6f7ad9430d986884",
+  "blk.2.ffn_up.weight": "ffdcb88439f5828ffbd9fc844b03ff91637b790b9838097258cc3ae75935720c",
+  "blk.2.post_attention_norm.weight": "4b3f53b7ba26e8c36b2dfda3b7e5fc4b1065257cefdea235fc7df9af130ac2fd",
+  "blk.2.post_ffw_norm.weight": "e550369e26b8485e2b54ad34b34bc98af5494287dcc513c2c39cf1eaa5b89d07",
+  "blk.3.attn_k.weight": "89f24ea450e37d9e95757651a83205c085d81b354ee9489dd6310a391d8409f3",
+  "blk.3.attn_norm.weight": "24e2ea662b7cb822b4ca5cd61bc17f2709f406d990ec3b4a0dac1cc112db45cf",
+  "blk.3.attn_output.weight": "ac4dad69473c6e3fac56669212cadd8c34ecc5973d945972e974d94805334967",
+  "blk.3.attn_q.weight": "b6a9c9a7d4722b9096631c65de62228dfddca6e26edfe6af7fce01e116ef0f4c",
+  "blk.3.attn_v.weight": "f272a960a40093942309bc342a379984cbacec2d7bc64428db3f64e6b1887ed4",
+  "blk.3.ffn_down.weight": "c0188ba50d8228805982029c277fc0e87aa57473b8363037c648f6d006ff828a",
+  "blk.3.ffn_gate.weight": "a04aec1561ee6c0fbb18c3db49dc62fb533619cf697fd548cbf2279761aaec3b",
+  "blk.3.ffn_norm.weight": "bc053837d44087ec05eb5d9458357b2a5be787789b19cdbbdc694b57697f99a6",
+  "blk.3.ffn_up.weight": "b3ce8b274f20796d3b1a7c08ba27a919066f9de89a782faa544c4a8d6bea1382",
+  "blk.3.post_attention_norm.weight": "9c922dee7a7df5667289e2788e60170238239cee2dfdbbd9e435763f9f416718",
+  "blk.3.post_ffw_norm.weight": "b682544ac953ad2e0b49027ed8916f2e9d1aba5d1587bb4127ac703570c7a03a",
+  "blk.4.attn_k.weight": "143b0cbb4b787b95c2b6212374410e32173ccef2adb914908a2f89a7916de512",
+  "blk.4.attn_norm.weight": "5668f60491b780273745192662d02c9a92a4f692b29d16aa0bbc7413fec4f85b",
+  "blk.4.attn_output.weight": "b9f2bdb68be1e0cf66dd19f8fa2afb105910ad2ef394864cb32cea8f8944e0d5",
+  "blk.4.attn_q.weight": "ddcf1343dafbc2dfcd0b8741225af22fe4b54b2becce29240bd01c34265d126c",
+  "blk.4.attn_v.weight": "6dc7074366e7ed52d9f48c594dcc85bef738e096276cb99d28228c89eecc5b9c",
+  "blk.4.ffn_down.weight": "30334ffc59ce343cf2a1b973174acb7722823463adc07e19a99bd0f404bc9906",
+  "blk.4.ffn_gate.weight": "890f7c8af208d63b28db52c4b8c16c2288a382d87ff5a6a6d6b0a5b3bf27e6cd",
+  "blk.4.ffn_norm.weight": "ff0316cc7847221eb86a90c1ab441d4ee61553d410c66414a7755021b3b12448",
+  "blk.4.ffn_up.weight": "6af97d113f91564c636734f215e25ee602d48eb045458f300b3ec7582be0f41d",
+  "blk.4.post_attention_norm.weight": "69438f231e105e68216b078bdeb35a7cdc8b12c4e2845e18ecf4c8d361d6a321",
+  "blk.4.post_ffw_norm.weight": "0fd535da78bcf2b32c95b05b2b83dc49817393765be90d8cc1ed3d56f47b68ec",
+  "blk.5.attn_k.weight": "0166eb3c6d20dcf3d3c169e94caa8dee057535bb525e29f698fb6f8844f18a6c",
+  "blk.5.attn_norm.weight": "a7808f27f164023d5cde2be00fc23cac6c71aa0ddeb60bc23e12411b80087672",
+  "blk.5.attn_output.weight": "8b65b2027a0842b68c5308f91d6a31de9599d794157d77df8418b19f9e0d9334",
+  "blk.5.attn_q.weight": "966bc626ef2c2394d872087a41c126bb1b67d1d5f6de920204ef5e5b16c34003",
+  "blk.5.attn_v.weight": "9a362aef3f4437fbf0ef6e1ba785f3329c3db2960f93fe36547d2795e9c254ea",
+  "blk.5.ffn_down.weight": "63e53541d34197720c06f297aa8142ac6b6eec002c7987b296f26e8b1400f931",
+  "blk.5.ffn_gate.weight": "d9591fdd32f783e0fc26e20d5d587ee8971ac8ae2e4c818c6eac1c125c7c7f37",
+  "blk.5.ffn_norm.weight": "677334cc60ecce3a7f4ab3acda15d359353d7358872f614ad8914e3780e9fc6e",
+  "blk.5.ffn_up.weight": "a63764110e1c655ffbd55af0669b2dfe4cc29d0e198d33a8e5426461b08a85f7",
+  "blk.5.post_attention_norm.weight": "c55499f859b2c0a7f5cabceaae47309a5ad38bc29d0f4a8db81f1357023162a9",
+  "blk.5.post_ffw_norm.weight": "82752754665f842418f3e302cb5f43d1e0504dcd124c4b8ddb77018b2c793837",
+  "blk.6.attn_k.weight": "e20a5f0d6c807273c8d491439566b428497ac02097cf0aa55e33748c28e14be6",
+  "blk.6.attn_norm.weight": "2c6ba42fd3c73d72073ced03a32dd28d70a89ed9bbbc8fea1ba03a7ade951e6c",
+  "blk.6.attn_output.weight": "4de7c5c2f4a133a266e17ed8c14c52959466b54cc7ab9e19f789a33b4850f284",
+  "blk.6.attn_q.weight": "56462d921800e6b8cd2213fef04c4ff16d728905cb2f4c58e966d0a053a3b0ae",
+  "blk.6.attn_v.weight": "b758dcbff769d6240c2245ede1dbc62c4170a67c77458e866312589220fe29af",
+  "blk.6.ffn_down.weight": "582247fb3c2bf687cbe9413fe18d18ad47bef4b65df7d78905e10335c6134764",
+  "blk.6.ffn_gate.weight": "3035444d5286aefb7a6d04e55bc27e1fac7cf895cd5be02319a431b8e047b4ae",
+  "blk.6.ffn_norm.weight": "e582d24c66e01b96faa20ce6adfda3d8583b11e809bff89969927398175e369a",
+  "blk.6.ffn_up.weight": "6f4b7bbfedeacf61a4866ae0616c4ba6c9e856662e8f00ae6aaec7f52c53e7b4",
+  "blk.6.post_attention_norm.weight": "8fe51b50bd677d21586aecab0b565c4bf9fa68ad50bfe366f45e8fea3c657ca8",
+  "blk.6.post_ffw_norm.weight": "81ba3cb4c2bf5c546b86855b7a885d3fafededc67eb3a35cd3598b03c9e26e65",
+  "blk.7.attn_k.weight": "2e044179cdcae0946708c86bfea7aa0391e1f7e2a09b33fca035d384cc3ca758",
+  "blk.7.attn_norm.weight": "94b48c546b046803c60e75a3acb17a356b710735989938021b565f68df9b4985",
+  "blk.7.attn_output.weight": "65709b4ad7a581f4d75793d39d4032a359f6bcc0c3835205242a0b99e5b66824",
+  "blk.7.attn_q.weight": "8ded993c95d1f7caf201ceb6fa035cd6ed6d351b50b999fa9355dfee9486cb5b",
+  "blk.7.attn_v.weight": "c92d5e2d2d48397542bc03bea25bf39154075e66c5bb1ead85188505aa04ae91",
+  "blk.7.ffn_down.weight": "e8ba8fb57208805ef1dc23cd7c86e9a2d1fb7c52c3940d292cd5bb2eb24b3fac",
+  "blk.7.ffn_gate.weight": "f0f06d6a2e06c5ac252083bc61d05c814e6289d3f4e4a87d2f06918254c02c36",
+  "blk.7.ffn_norm.weight": "ebf8ef775f72624148e09d68a4332187a7a5020c521fe0623da1cd3485ad33e0",
+  "blk.7.ffn_up.weight": "a554adc4fc7122c247c77670e169916ba1794c787b5be30a2b36705138f1f746",
+  "blk.7.post_attention_norm.weight": "3aa6bc21d85c3a0c12b964e82b12feaedfdd13130c3cd2229228e24e0967ebdf",
+  "blk.7.post_ffw_norm.weight": "508bc7b19ee8ff08f0007c890133a462fc57c7e72b16ee8f6dd64def264ef876",
+  "blk.8.attn_k.weight": "363c8e74056642fe9e7c2f3f9769d57319cd3fa0a6022810189ab8d894322885",
+  "blk.8.attn_norm.weight": "685b49a1f1acb169f4df0bdd8e3de6943f3033cebad14b898a72000595610d92",
+  "blk.8.attn_output.weight": "7bde571e4efef1c6a6143f0526721dfb59e0a0ea0e1a3616a322b2eb937efa48",
+  "blk.8.attn_q.weight": "fc993dbc1074c28a0e1d85e5ab2f4ea6a9c6c1affe7ee56027000a275daed9b6",
+  "blk.8.attn_v.weight": "281e8791d3aef9b3864f1cb054da0ae0c2fef4ce0a58b1bad8bc136b2fa0f62b",
+  "blk.8.ffn_down.weight": "b1164a2578a7f87ed99c2bbc76c5dfbbbc6a1a803605391acc3f320fc989ffd7",
+  "blk.8.ffn_gate.weight": "6b39a3b3aaaa79aee61416b54d62160b9258042650e61c6b47bc77c2dd17daf3",
+  "blk.8.ffn_norm.weight": "17ea1362c72da27f12bc936500492035bdef3fd8f940cb12b57f37d42ba8ecb1",
+  "blk.8.ffn_up.weight": "bc3a7c47afc440d2bdf8fbe9ddf2c9220467472c60c8b4ded8c0f181470ec96c",
+  "blk.8.post_attention_norm.weight": "5c506204e00411ef9c8b4134d40eedcc19fffe68dd0af7d7cc49dcabf2dfac7e",
+  "blk.8.post_ffw_norm.weight": "002faec235c3678864e2901eed275ce4e9dc229164a91c9cd4c965142ba62305",
+  "blk.9.attn_k.weight": "0bab39d8c237f1b6d0010db40467142625a9e6f2e0e4c49a56c12b41e4e0b1fa",
+  "blk.9.attn_norm.weight": "de5f38e873b17f07aa7598831b89cc1cae2c9bc3eb2e042ee9af059d2563e84e",
+  "blk.9.attn_output.weight": "8a8184702c25a62df9ff309c0c7badc8587208523b2be3e8fa90ce7080573e6f",
+  "blk.9.attn_q.weight": "7c961b2431b09ddf95377acd07201cb91bf13d9cd3ae0f2c25c7d6a0358d9f50",
+  "blk.9.attn_v.weight": "e22d240cb4743067033e659cbf210ebe2ebbab3e1dea6ccbe5eaa982382ca038",
+  "blk.9.ffn_down.weight": "a426f81210f03d6ad53277416e1fdcdf37d8065e4817613edaf6c67a343426be",
+  "blk.9.ffn_gate.weight": "a82eba825cb77b8e64f85ff99ede2fc71bc9b01751eeb17e9e6c246ee12ea62e",
+  "blk.9.ffn_norm.weight": "1a97f9b1302a3a326d534c5c3fed2db6db0ae45fd0edd381a3e4fc1c75d81030",
+  "blk.9.ffn_up.weight": "5f20bac2bbf03bb42adb92fbf99561651e1edda57e0b61935ac7f6c08c0ed7cb",
+  "blk.9.post_attention_norm.weight": "9f9866d13988e1946b1e1c80d9374a92a6e3be33748f8eaed3e126d1e1a4c796",
+  "blk.9.post_ffw_norm.weight": "a6896dbf698db4dbbe5dbf12417d4fd80e9cad0c539c858892ec0aa5b046bb58",
+  "blk.10.attn_k.weight": "ca8446e5d21ecd4e6a70dca8d321be480be4fba94d70cba065205436feb44270",
+  "blk.10.attn_norm.weight": "4f41fe290e8f21f63b82151b6cce94bf7318d121468816b0c58af0ff7c1658ab",
+  "blk.10.attn_output.weight": "c626d2e9681c5c941bbde43dddfae1a8d4986bf2be4470857bc8e8bd7f869044",
+  "blk.10.attn_q.weight": "1e61b210a13a429977325cf15d781ab77d604cfa862f4270329cbd94237d5835",
+  "blk.10.attn_v.weight": "8ff8d3e3f058ec3b35ada1057f2ed59c06494d0e0be6a8dc3ff9edf9f0e1a115",
+  "blk.10.ffn_down.weight": "bcebc04219f8081a5f483e58103c0ddbbbc631a0a54fd6dd9d55778e041f70ee",
+  "blk.10.ffn_gate.weight": "7a23a1e620ef871384ddf9611ccdcfb893fbf013cc203ac8e72f745420f1eea0",
+  "blk.10.ffn_norm.weight": "e3a375e43c349a1c6c66c22328e513cc1af3137fe839e43dc8e9be2f65914fd7",
+  "blk.10.ffn_up.weight": "5d182e7c94369194fca5f19cbbe668a999911e57f3d363bc7fb6088428700cb9",
+  "blk.10.post_attention_norm.weight": "b841c6308296e8984f3c5f549c6e3a242f4b3e19141e1f54cc08de9c46759c09",
+  "blk.10.post_ffw_norm.weight": "9d66fa05b5c940208f634f5053d809094c99a2a10a1d1e8847c8281fbd99fb49",
+  "blk.11.attn_k.weight": "14adf24ebb2bb17b336ca81cec3e690fd854782f4440ca6c66cc1d7e7bf1c850",
+  "blk.11.attn_norm.weight": "2d2213f311f50414702b5b34f22aafb9d9a0b6787243e7578562583dc40ad195",
+  "blk.11.attn_output.weight": "de1f14cc2a7fff00cf11b229f0576999205f17b9536e97abc9d6de3cc79a7884",
+  "blk.11.attn_q.weight": "2bcc5c147524003109ece0be08b89ac8b25baa71416ffa76573c6c052ffc6eea",
+  "blk.11.attn_v.weight": "2e6ab8573070c22dc1e0d7aebe4d52123226dacf7822dcce06fadbb38fb036a4",
+  "blk.11.ffn_down.weight": "1b86902f4e36868421e5228b9445051f8290b292df22a6d1af836dcecc1f25c3",
+  "blk.11.ffn_gate.weight": "e756e8081bd0a16aea4a9ef5076ad102113524f7a3d50a3a77aaa7f7938b63e8",
+  "blk.11.ffn_norm.weight": "6913887267be227cf9d1991a3dd8db2e7e74bb9b5fbdfcb9ac954fd7d7b95b3b",
+  "blk.11.ffn_up.weight": "619a3ac0609ebdf42c3fb2b6e4b1db48df79e6dd8418d7ab8f1bbff13d8a6a50",
+  "blk.11.post_attention_norm.weight": "e4b4ba92cef7b6a78407e8ab1b0307d47dac6c3df7b6817e28038317ff662d7e",
+  "blk.11.post_ffw_norm.weight": "40aceeec58cb855f0c158c9cc217168fcd5d0e735567d587217b1d78df17bc5f",
+  "blk.12.attn_k.weight": "c54c5a4d4892522022d1aa2204cfc624f0b4042caa536e678967316293fe5cb1",
+  "blk.12.attn_norm.weight": "7cd2ef58298569ffdf244d9b390f3917245276c8206e5780af5f96d8c0bbb446",
+  "blk.12.attn_output.weight": "85495ef9cc8b3deb21f741bde463ff6493acae2be51f02ecdeef952cbdec3375",
+  "blk.12.attn_q.weight": "d19383f83fd119bfb8c0280c9515705c11d8e7d502019fcf8f49efeef0d106d0",
+  "blk.12.attn_v.weight": "869ac669ba49531d9128892a0e27cef15de508ff40cdf80cc1681dde50d09204",
+  "blk.12.ffn_down.weight": "578f39f8f9fc2f09138afc884a952d7cc3a9a31de4216acd10e88e19e0b75f8c",
+  "blk.12.ffn_gate.weight": "e29a0186bc6c4a0720246306e922d3a83f777dadcf4ac80bad468287031cc8b5",
+  "blk.12.ffn_norm.weight": "e1ee95c6584b5cb57fcf1db8ce2bcc03aff91eb389238c094a61c00dde93d1f2",
+  "blk.12.ffn_up.weight": "2a826f06d7cdfb3edc6ae250ff44363ef77a2a9cdf96313e23a331b99ebfa17d",
+  "blk.12.post_attention_norm.weight": "4bafc7699b948d5cbc0d3e09b418b06c6abc4651a61ada9609d9a2f21c7e5607",
+  "blk.12.post_ffw_norm.weight": "bbb8c34a7176bb1a49f9fe2bacca0bd26b673d52c0835b2e90fa11f2962f077f",
+  "blk.13.attn_k.weight": "ffeefccfe8255d1b694382012ff4134eee5fec9d9491c8d0ff0a13832d1a37e8",
+  "blk.13.attn_norm.weight": "35713726529e3887c4135a88e86e8a4d7270ba5b9f2d1ab462622fbf40a7cdce",
+  "blk.13.attn_output.weight": "0d60b7c5cd71190a9ef4b873b0f516be15447c32d83914db2794b14592b0b460",
+  "blk.13.attn_q.weight": "8296069e65bef794cefc61257fc65789b3cb22955e30f3df129205e5041b2222",
+  "blk.13.attn_v.weight": "ca0f4ab9d16a748fc643a5c0c7a19826a811bf2a4e7316a8c935d4bf0ce8abc6",
+  "blk.13.ffn_down.weight": "d5514e0c8e7b3ed1cbcc1605eb5be1733b6ab3514cf8a0508fc72f7d05ed8bcb",
+  "blk.13.ffn_gate.weight": "8108e517a82e08a3aefbbd267bfa50a1668f92a76273280ce8a6bc1f6dd61521",
+  "blk.13.ffn_norm.weight": "5fcb6132d2134bf1f835b904a99820fa501dbc57d2224129f7098bf3cabc1d36",
+  "blk.13.ffn_up.weight": "6d744b7cd390a3cae3aa350dd379b81246acd056a2259996b6aaadece8465ccc",
+  "blk.13.post_attention_norm.weight": "e08b14698912509790e9575b8676971fbb0a4d82d719367e3756c0d0c4ab8cc0",
+  "blk.13.post_ffw_norm.weight": "2b196e4450fc5f1e7367b2cf7fe33a15fe919fbcdd861d11002346f16e980535",
+  "blk.14.attn_k.weight": "120e5f48d7268dfd9ab5f4bc9cc57a7cec63ea9635f56b80d435eb22936e9483",
+  "blk.14.attn_norm.weight": "146367bcce4db72cc894419a2e0145a6f533507dd68e4739c10ee480308c401f",
+  "blk.14.attn_output.weight": "720fa0165e756876c5cb6ad9e2780dd910390933f3f8849e5add5da04266650b",
+  "blk.14.attn_q.weight": "f5183466f56219ca1aca52d8b82c2d966a4198fea40fdd6b39f4d8b06ca2a6dd",
+  "blk.14.attn_v.weight": "24f8ea3d5512cd37c43c8329cb0da0c90d1895aef763ac2dcee3fe5157ec50a2",
+  "blk.14.ffn_down.weight": "e29960965b384ae5ab3d898a4dbaa8fddd28fa0e477ac28bcac49dec12a5ac67",
+  "blk.14.ffn_gate.weight": "6d0d6a74bfe9692e8f8eedff0fc34fc4fa1c8687794f35f2e2b033ab2d7510b8",
+  "blk.14.ffn_norm.weight": "f7036c1a9a71e046c9d2af16e9218fda5dbb0f7241ab44747abed1f0f9d602ca",
+  "blk.14.ffn_up.weight": "7d69ea1424007ffc9c12247dd0308c616e93ac02a59ec341cfa48f92d6ce3b10",
+  "blk.14.post_attention_norm.weight": "65b9712834d9445d4236bec362f3fb795c20d60c541b3dc6dbb7914d9b493e41",
+  "blk.14.post_ffw_norm.weight": "9c6a8da2e4e437d5cfdf3b9097e9f8b64bf07946a048badec20f4d374613f38f",
+  "blk.15.attn_k.weight": "864bc618303a0e4ee67fb1d5e751de61e936cd51e96669dd86f8cd08f2305045",
+  "blk.15.attn_norm.weight": "f9f4187da6eeadc2fc5921d8fe669741697d16c13d71e4aaeb73b82f50dc577e",
+  "blk.15.attn_output.weight": "ce2419a0b097036b2a31f2f4ad731d5814bcc2ef4c511786e24471e5eefd273b",
+  "blk.15.attn_q.weight": "9539db5a970d11ebe99722d1e13fcd635e250033630811efe583d2f97778e4a9",
+  "blk.15.attn_v.weight": "1c834b48ccd88adaeabb7d8bcb6be0bcd6d5ac1354ce88fc28f19a1a96b81ab3",
+  "blk.15.ffn_down.weight": "bc1f97a65dde6fa2c1e5397afb612266944b343f2eaa868b635ddd25829f8a42",
+  "blk.15.ffn_gate.weight": "1b14529d57056b79037f6cb5008132e62cc35992353b38dda59572274623103b",
+  "blk.15.ffn_norm.weight": "9af77458de9ee55c66f93865759f9c2c398557f94f3fa8fa6af30543d7339cde",
+  "blk.15.ffn_up.weight": "41d524a26b61a9595816b4fd53cf57ef50a702e4ef32933ff6136dca9136a267",
+  "blk.15.post_attention_norm.weight": "c60a03cd0e63a7db5c80015e58e9b97ba2208caa19f66a6fef5c4447eca900ce",
+  "blk.15.post_ffw_norm.weight": "34f7f9f96769215bbc3d17084df091864aef96a6645b7d0b3b7d9bd92f1a4b0b",
+  "blk.16.attn_k.weight": "7e27240d9f3a8c6cf0f4a980113d43234f514eadc3e3e1792b86efb29ffb1a6d",
+  "blk.16.attn_norm.weight": "af798acc0899282a30448edec48223b3e8efda177090273e612d8eca5e377301",
+  "blk.16.attn_output.weight": "79df39a3709d3d53e84146291e0944a7a653d06705293d9ccb5648dceadb432c",
+  "blk.16.attn_q.weight": "db58a1c3b83ad294804e5fd7321005719e200659173466df5a52a182b80b7165",
+  "blk.16.attn_v.weight": "2af6d48cbaeb225b5c1a704f76abd89c8ab1521417695b112b4dcc2cbd39b74d",
+  "blk.16.ffn_down.weight": "fc1c813eb5e7da3d6194569d6cb21602fc6eff2dc8e1b0eb753f2d5df148189c",
+  "blk.16.ffn_gate.weight": "7a80bcbc42464bd55df4814a6edbd7b5c153e0428323bbe49de55e2d2add33e7",
+  "blk.16.ffn_norm.weight": "2041685ee926d30f3f2ae4ec35b5688f1cd834167a6359a7d4057eac804c58b2",
+  "blk.16.ffn_up.weight": "8da4b718973ac1d43b928829bc45e062fd101984d6c98dd825bd7c5d08ebfbe3",
+  "blk.16.post_attention_norm.weight": "975c48fe680a6167438a106140a8872eee7765191f152d80e3b8ddf47693e095",
+  "blk.16.post_ffw_norm.weight": "4de2d4d483acfe4fc77860ea929025df2f4e15c10729413f36a18c94eaa6d689",
+  "blk.17.attn_k.weight": "f937e61f0af8c4cd98ee742648eb60e02e579683e21d421071295a3b70aebaad",
+  "blk.17.attn_norm.weight": "c3270583ed28b7e423f5b170c59113234f258169b93a867d9274f4c10b7cb115",
+  "blk.17.attn_output.weight": "b8c1150e81e685e539a5dcf2c19047a24eba2b281fabe166674b1d71ef4612ea",
+  "blk.17.attn_q.weight": "c255100ae2011e7dc7e3bf3bc3ccd96d859fbb98581cae993d7b82c1ba8e8b39",
+  "blk.17.attn_v.weight": "5830bb0a555984c6485348067f70b5d22ae337c011aa9248dac2ff4c95944551",
+  "blk.17.ffn_down.weight": "8ff9a7cccaa3776434a9d895aae4fb5c36c736bf2ec98784226b4c234940fbb0",
+  "blk.17.ffn_gate.weight": "1b52876739712831c272911533da206f407b46034a1a4ae8a88c1f96b6bd5747",
+  "blk.17.ffn_norm.weight": "d0e16ba5e87c91b545334e022058c7d03849665c3b1a6298771b656531366b66",
+  "blk.17.ffn_up.weight": "4dd6211d01dbebbe21052708eddc242b082a58b5f18ed16479e17987c1d3432e",
+  "blk.17.post_attention_norm.weight": "6f49c775c7417dade77ba8268a0f8441c1e5ec28b5d7e4dc5ed07a04d04600c8",
+  "blk.17.post_ffw_norm.weight": "b91a0bb2e6679e9c9be06ad323adae441d00a3d673efb19d7c4954be2aa84b27",
+  "blk.18.attn_k.weight": "22b565ace1b4da8b33865a58625be1d90beea9891f29686a69fa9cf7c93217db",
+  "blk.18.attn_norm.weight": "3e0160d7063c8753de65d2356a66648e47d921efdc5c917efb8209892120f8db",
+  "blk.18.attn_output.weight": "e3180f0bb4ca90b31e9b08158db38e332de62dfbaefe34aa94cc316409331e09",
+  "blk.18.attn_q.weight": "f3a5a83614c3ba7ea41cdd5b1b0819a241ee2a951a381ce4a9e001d3f700ed8f",
+  "blk.18.attn_v.weight": "f3350a5984fb951fc738adcf78147e6d812ff1c576670c460cafc99c253c1654",
+  "blk.18.ffn_down.weight": "9e9d09b13a33525e14bdaee6efc65c551ac7cf7680e534b940ab122a3a7c1ac9",
+  "blk.18.ffn_gate.weight": "ebaec8b4b578a2e8d815baac12f1675c208f80c68074d5a18288a2e1a60680ee",
+  "blk.18.ffn_norm.weight": "33e7687c53a242f2f8dc7093a491c97b18d4a5a8c14d183f02bd586a770f05aa",
+  "blk.18.ffn_up.weight": "78a1816662378ce56cc870e705174492781897b3afd2d4d97a51f10f2f2987c1",
+  "blk.18.post_attention_norm.weight": "a58dde3f12df3e94cbc27d87c8ea86f89af8a388a506446ff6758f05399b05fc",
+  "blk.18.post_ffw_norm.weight": "cebf90cc143577d483cca27b032dfd82031ee59bdf17c0e2cf60a0a3ad5bf996",
+  "blk.19.attn_k.weight": "4683375d0599ac9e2232196aae1e90af13a14cae26e865465de5c8e257bb2055",
+  "blk.19.attn_norm.weight": "f3eba936bfb1814bbcb0a1d62739eb66daac839df8c9c836fe0e94860df88525",
+  "blk.19.attn_output.weight": "51c0f01d38a9dcfe9bdbc4643576fab164c1d9e4b7168b7695c0ee55e6965667",
+  "blk.19.attn_q.weight": "28d15b69b8416f2e7ddc88fe381cb1e2ef2ad705fb1c268139ba96498cc74848",
+  "blk.19.attn_v.weight": "6860f1cd720638e63a981fa2c0b4db900129826bcb9823c9ddf9fb8b1b9f3383",
+  "blk.19.ffn_down.weight": "bc7f2d7827ee01c2dd41401c7b3b1700ad3a4ff620e8bb734f92630d342dcc7f",
+  "blk.19.ffn_gate.weight": "54d03ef69ba373fc410fbca8f1e34a565d58e4296d9a035ff7e48340b9c848e7",
+  "blk.19.ffn_norm.weight": "9178fc796a340ee6e8128ca74c0cb6203d1adbed6927af4e5ac7863da57affc7",
+  "blk.19.ffn_up.weight": "a77bd708026c6e83ad5c79c223278e74621bcf74a9641c7818d96b595daaad20",
+  "blk.19.post_attention_norm.weight": "ae94aa26f4c411bf9496a6fd4a6df64ee589ee1ae9a04b531d45acc95721e582",
+  "blk.19.post_ffw_norm.weight": "9ad210700edeef12133bdcff04bf1c7f62b49f6f4a9ba483c7cdc59857c24a5c",
+  "blk.20.attn_k.weight": "e35bce1e9f4a7a09ef34721f57ea38cfca68c272f52d923fe50af8308f66cfaa",
+  "blk.20.attn_norm.weight": "644800f6926fd34f233795c4dec1151a295d2138ca8cac33e3e48167d26f8b41",
+  "blk.20.attn_output.weight": "8d3758cd236471741e1ad66c0710cb79077dc8c7a3a292d35bc551c0c5abe627",
+  "blk.20.attn_q.weight": "c333b1f0f6f956b5d73891df10b1a0321e55fc31c40d623a24e1f52caa6a998b",
+  "blk.20.attn_v.weight": "8562b418d0c4868a050fb19fa3fcaf50a8cf1c669f537d666c80c7b3a04714e1",
+  "blk.20.ffn_down.weight": "97efb608ac44cc804198faec3ee66eafe56ced6b7ca5359700c6f1df75b7205e",
+  "blk.20.ffn_gate.weight": "5c61151d86f28415c73c73d90ec088c646cbe5c1640197caf58eb501ba7db293",
+  "blk.20.ffn_norm.weight": "24bbe0a701afd4bbeea65b3edde712b3cbb2281043bbc43dbf250582453116ed",
+  "blk.20.ffn_up.weight": "e170cf68e249566aa99eb6f6b265679bf9a5a6b76830ba24e7e130c2515910c4",
+  "blk.20.post_attention_norm.weight": "e092d751cfe20dbf2d348358f3b38397bd83e4ed94d6bbaa6bbaddcd902b2ac4",
+  "blk.20.post_ffw_norm.weight": "219a18a47dcba76e669e4322223a5a9227bd3db1de3fbd3d3cfb22e54a783c5a",
+  "blk.21.attn_k.weight": "c3a095ebddb42c63824f1c98da65263dc88e4d790a26aa1632840b44f5cc7cb1",
+  "blk.21.attn_norm.weight": "ef8bbaded5fbc45ad9cf3985ae02174524e7090fe6362811124f942ef643bec7",
+  "blk.21.attn_output.weight": "668f018aba72baac6252aa3ad58569ddd55ab751a0dd8d7bcc9fb9b6efb4bf53",
+  "blk.21.attn_q.weight": "e759c65663089f3bbbd51847934c185e680c82f1249065d5d487da638e519e6d",
+  "blk.21.attn_v.weight": "2ff57762686cf9ba1f5a6be76503454b97556ce67f4ac98254bd0562231197ba",
+  "blk.21.ffn_down.weight": "3fd106556fb721b1c28ae3f4026bc83eb1b08ed910f2ba5f466c6b5f327d91cb",
+  "blk.21.ffn_gate.weight": "338022d882f4b6619e8054a6fb909696fa3eef3013cf69b65c3cacdfc5b9e42c",
+  "blk.21.ffn_norm.weight": "1e77660c23a3f9653ee721a863d1960f773d87437cabc4dc0a6e17ee3d4e5e44",
+  "blk.21.ffn_up.weight": "7d31b20fbc2e6eba8f350f170069dc36f0cb12f68fbc4206ec5022a74085ebcb",
+  "blk.21.post_attention_norm.weight": "9638bae8d8bdcd7ed68da282979cd84a07c41ff9cabcaea94ebc846a1803db23",
+  "blk.21.post_ffw_norm.weight": "d622ef11115fe0cbe04b727d5a3b6371e7f39bf08c8d5eb9bc6da52e3f3cfb9d",
+  "blk.22.attn_k.weight": "5c321cb29deffbe57de200dd206a62005f1e80acb86c4fd2349dd44c8d3594fd",
+  "blk.22.attn_norm.weight": "198d949705d7170a331d75889d8c7500c3635254dac2cc6aa4dc35d556584536",
+  "blk.22.attn_output.weight": "19805cd5d7025b457e5d41d70db8b3fd63c2dd0e4a94d3ef1704d50ef4e749e8",
+  "blk.22.attn_q.weight": "177836cd583fc87405975ddc21ebfebdaa090a0363799664c72caa3da851ae2c",
+  "blk.22.attn_v.weight": "fea255692483e30d0108f9e4e250eb3ed7dbda8d83f499b06519b8c223ae6096",
+  "blk.22.ffn_down.weight": "00cb8939f03e5817d6d412de8cf2c923c9568d5493e382cec7faf5718fb034eb",
+  "blk.22.ffn_gate.weight": "b0591065b91281b2fbd8a9567f3568d40479f680e1f0a29e27ae213f37642489",
+  "blk.22.ffn_norm.weight": "96b5c5d0737c2ceb8fc869f54adb9e5f46e28cb7b177c40f49fa926b923c00f8",
+  "blk.22.ffn_up.weight": "81f472185b24344ab0594ea8246cc6e200e0dc1cab4943e74fbe4ca19d5a9701",
+  "blk.22.post_attention_norm.weight": "27fa9aa6260aa3071e0391e1a1d49322dcb6e8072315b8a9b7064087108dbd06",
+  "blk.22.post_ffw_norm.weight": "f37e1dcd7f643d9545675ffe9dc527a11eba86eb204989c2f44f636b266d896a",
+  "blk.23.attn_k.weight": "5d82f36658a56c3f94d0bb2d61f65509c966fa6568f81812e0d3e338b380ef8c",
+  "blk.23.attn_norm.weight": "b7983f88d9cad88bc88a528923e6da592ad20e699965b223ebc10840fe1f4fec",
+  "blk.23.attn_output.weight": "59f97f80f430d71606aab0158a195aed29ccd3405e6c0a5c41c809be8eb01898",
+  "blk.23.attn_q.weight": "53ac4789fe958919cc02ea4222bcd64c0ea1b4baa54304bff46635bdf42f7490",
+  "blk.23.attn_v.weight": "ec8abe09b9e84dbb52c7a068094657c6d3c62fe551ba8d7c3a3f23da622e9756",
+  "blk.23.ffn_down.weight": "3cf547eccb1b82aa64f208cee9682d7f558ca84e0aead7d9d3d1420d90f3d992",
+  "blk.23.ffn_gate.weight": "366aa2486d911ba81eb519119e13807deacf7e9908bc1975a2a63e00d6b10124",
+  "blk.23.ffn_norm.weight": "6d1d4a4af34bb7dc090ac87d6457d398c3e0fb68bd2e2b60b099dc318b6cfac3",
+  "blk.23.ffn_up.weight": "53f76692e253f5d2420b3f200c731b9f3b7a83e379920b4a067c729b4674aa4d",
+  "blk.23.post_attention_norm.weight": "7c952fa0efa76b3f048c8c4c9e8dcb5e3724d231327eda6423a34d3f3d3367de",
+  "blk.23.post_ffw_norm.weight": "7ab188cfe61f0a91b40309a0ab6bfa99f19d0ff2a37b6ac10e5f0c7f44eb5270",
+  "blk.24.attn_k.weight": "225798792f9bfdd10eff0505ebe61e0aad0209c17b431f6044ee7968ffe8c198",
+  "blk.24.attn_norm.weight": "635e3c1ebf5219bbebfc40ef164bc32d2b726ef595a94da64ac524ae878e2915",
+  "blk.24.attn_output.weight": "482f5bb2db8d9ed22b253d9a3296333b239efe698e5992e5d77e7e12dc2a5cf5",
+  "blk.24.attn_q.weight": "43805bbccddb65d58fffc4be9b5c374d4e1df1395ec1e1ffb4bcff03e98d5adb",
+  "blk.24.attn_v.weight": "fa741af54b4a3b1775d32f59134756090c5df2e7345a12a2d8db94fe289667a7",
+  "blk.24.ffn_down.weight": "83c6351e3162626b276f524a57836144625c2556dbe321b57cbd8fd486a68fab",
+  "blk.24.ffn_gate.weight": "fbe66be0d84d12cea5176cc7eaef64382ffc7324cd9d6266a3342dc43442f2ac",
+  "blk.24.ffn_norm.weight": "77c1445a8639ad24938bdf0280233eea2362d47391421833dfa72ec756dfc1e8",
+  "blk.24.ffn_up.weight": "78235ac729ee23c1cf1ae543751e3af32776d8808cee6e529c2a625a1f027654",
+  "blk.24.post_attention_norm.weight": "161f71b6d07628d43e4ae51a4c9088ec6ca2db123a17986a14505d83fdd04dad",
+  "blk.24.post_ffw_norm.weight": "cf1ba692aa683368b02ac413e69b2521b98c69a5274eacbb54165b53bf38a8b2",
+  "blk.25.attn_k.weight": "057a56bd8c8d2b41608d1f71faa3052902152ddf85e47669ad950c1c3e77c33f",
+  "blk.25.attn_norm.weight": "b7179fe02c334da556ddcf6c1b502245639a728c4cbba8b552d8e1df4565ee9d",
+  "blk.25.attn_output.weight": "4fed8b05b08a0ff75ffd022701bbeb52f17b23d09332a1ddcba737244bd0d3b0",
+  "blk.25.attn_q.weight": "c52e99f5d38bf7538d6106a0bbf38ac6dc6296bca9a3f849afa384ea67b4af01",
+  "blk.25.attn_v.weight": "c49c23d8e1cfa6a8eb971eb69942204890c6d7d830dc8774c84b108a80598912",
+  "blk.25.ffn_down.weight": "c08d4dc8412b19fdc870c164b83c341b236ec6fe7bb4a9bcfe0dc100faa20286",
+  "blk.25.ffn_gate.weight": "1a4cb3f36735d59181721471452807903006539e5e1b5ceb4f72d1d7ae134127",
+  "blk.25.ffn_norm.weight": "8fd6bd0dcec5198761525a36992a57c9ec5e9da60a22092839a84ae8c4e87f26",
+  "blk.25.ffn_up.weight": "3a00f39bdd5f31dc5e3b281d2002e1ac4f2475d49a0ac1d7720a25b377dcd04a",
+  "blk.25.post_attention_norm.weight": "e5f31a648612c859b6d21c9ee426e87a86cb1973dfdd86276c767371d9cef5ad",
+  "blk.25.post_ffw_norm.weight": "553c3bd774922c99c2384380a142d019881d30dbf0fe3bf9430dabfb3f6cbd33",
+  "output_norm.weight": "49445c4585ab0a8135717a0bdb1cda4a062a030177d0119561d91542aec5744b"
+}
diff --git a/convert/testdata/gemma-2-9b-it.json b/convert/testdata/gemma-2-9b-it.json
new file mode 100644
index 000000000..90cdbee47
--- /dev/null
+++ b/convert/testdata/gemma-2-9b-it.json
@@ -0,0 +1,6 @@
+{
+  "general.architecture": "gemma2",
+  "gemma2.attention.sliding_window": "4096",
+  "gemma2.attn_logit_softcapping": "50",
+  "gemma2.final_logit_softcapping": "30"
+}
diff --git a/convert/testdata/gemma-2b-it.json b/convert/testdata/gemma-2b-it.json
new file mode 100644
index 000000000..0482f1e14
--- /dev/null
+++ b/convert/testdata/gemma-2b-it.json
@@ -0,0 +1,188 @@
+{
+  "general.architecture": "gemma",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "gemma.block_count": "18",
+  "gemma.context_length": "8192",
+  "gemma.embedding_length": "2048",
+  "gemma.feed_forward_length": "16384",
+  "gemma.attention.head_count": "8",
+  "gemma.attention.head_count_kv": "1",
+  "gemma.attention.key_length": "256",
+  "gemma.attention.value_length": "256",
+  "gemma.attention.layer_norm_rms_epsilon": "1e-06",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "2",
+  "tokenizer.ggml.eos_token_id": "1",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "3",
+  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
+  "tokenizer.ggml.token_type": "485e40bf3d715a4764818fc097d6a2a41db872d82ee714bc500872a3437ff48d",
+  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
+  "token_embd.weight": "17b87ab2c01c80657855a5413d0457b4a041afaeda0cc785080e44e2f04acf07",
+  "blk.0.attn_k.weight": "28ac0da05754ad2714ae95da28a5ad191192140b30b8fd22d108d4700c9d989f",
+  "blk.0.attn_norm.weight": "3f9d5675d1ab0eb8a816719dac9fab81f2e95c52be02c34263339acbc087febb",
+  "blk.0.attn_output.weight": "703295c2c63990ff896778685c678f145298886f680f3ed5dc2a7ad54c293265",
+  "blk.0.attn_q.weight": "69c2d0e4870e9d722a190d356203c9605575a16863466c3d1747966ef1cf5791",
+  "blk.0.attn_v.weight": "95219c9c07b5ffe9a9a01e456d845eef2b11f4fc12c93dbbba479db395444c13",
+  "blk.0.ffn_down.weight": "a2feb5eb3d572c57c5bafbf0ab506862df1160fe40965dcfe4b9fd855c08bed7",
+  "blk.0.ffn_gate.weight": "fcca072c445c31f4dc4d5dfaa785b1bdf7271342442099b74fd17268b5829fbf",
+  "blk.0.ffn_norm.weight": "7621f95dbd245cade6fffd6b08797d69d8e3954e960f0b5551b90d967ab95448",
+  "blk.0.ffn_up.weight": "14a9bcdd451403c67136391e1b6e53b3b1830f00199bd911dbcc56d8749c14f4",
+  "blk.1.attn_k.weight": "c70f73c5df20579cb44d971164b48b5f0d8d5abdb38b381e7a8b880ba12aa406",
+  "blk.1.attn_norm.weight": "88b6b91f93a1ef83425a7c7dc2a2fbd3b22704a04c64a80061df376ac8c33626",
+  "blk.1.attn_output.weight": "f031a537490c452be3b3bb51e6b7949a636405756e160976a1c070a792ea00ee",
+  "blk.1.attn_q.weight": "bdb23214b1cf9cfd30f863a0a5868e52c6809d93b7e8f44df096a94204d9896a",
+  "blk.1.attn_v.weight": "e9bbc0b05f2c872fb1403f8f938cd1612b502229ee401f12593b1164c61acc00",
+  "blk.1.ffn_down.weight": "5ff53811038b661a7b8f2bfdf213bebfb185ec1a6060b662f063714f33584d79",
+  "blk.1.ffn_gate.weight": "205085c8c951a5c7543b1495183cd96028fb49f67464b3e9862a2693a6077a33",
+  "blk.1.ffn_norm.weight": "798f354fc85afce9625f5d10093a585a966831698a0560e6c9b97ce659eb4b22",
+  "blk.1.ffn_up.weight": "db92dc5684cb6e90940e13f4d1da555ed20ba4f8cab1e990ddfd7553e2e91315",
+  "blk.2.attn_k.weight": "ef5ce360c4eed6d00d03ca4761e0f8e4b0af4509978468314be14f3d46621044",
+  "blk.2.attn_norm.weight": "6dadbc05dbd0d3fabb4216affa60a3de1378a82d2859dc90b338cbe70f50d455",
+  "blk.2.attn_output.weight": "6bbf87a966f691bbfd7c8d25629aa4e6710107bd431a667434861febb391edc5",
+  "blk.2.attn_q.weight": "4e575c09ae2de417ce9057ce8b073680e860a24aae13a472b68f101b760752e5",
+  "blk.2.attn_v.weight": "cd33f7f01141e9439afdaf2ea1aaced9feaa335e32a58daa136ebd555d4d96f4",
+  "blk.2.ffn_down.weight": "b970ff1b0b6494165defe2fbfa1d31425766ed71e64de9ec4e66ac3955c8bc5f",
+  "blk.2.ffn_gate.weight": "dbb3e1360402e0e369b101995bb686b73f95d4a7673f061be85d64d15dfb0061",
+  "blk.2.ffn_norm.weight": "bfb7980105d8ac9647710454f57a5cdac50598a0f6f4884e16f1d94b00844687",
+  "blk.2.ffn_up.weight": "50ef89339b275a438b664686f6227dd9b6e43853ed6856ec9e33ef4bbd90bda1",
+  "blk.3.attn_k.weight": "be942ea98151434eebcd2c1da4b00e0146152fe524a530689b1fd491cb833d21",
+  "blk.3.attn_norm.weight": "0df2f218daf609c289fb7c60c5f375fa99c0d4e04381ad5a494a19144edd8e20",
+  "blk.3.attn_output.weight": "c2184aaf86aa2cb8f47be49f60b165834e97205f39c6ee1dfd19fd4411a156ce",
+  "blk.3.attn_q.weight": "4f86e2a0a4221c1c84ff9c409ac89893cb95d7208cf65bf1e98e24e01125f991",
+  "blk.3.attn_v.weight": "abfdb8a60c349dadde641d1afc9542025e24fbf41a3238bfa9675e0b1f1e4b68",
+  "blk.3.ffn_down.weight": "58821a8d87008d47d122427911c6fad5272aca70c448bbae223256a74bacd07e",
+  "blk.3.ffn_gate.weight": "776e051f1a0ddd5c4934e69186683a75ca9a3c8c0f61911bba321fed1dd287d2",
+  "blk.3.ffn_norm.weight": "7f380f29335e28be90bfcfae6f6d69fdf5751211b36d2dd62aa5541ed113e4f2",
+  "blk.3.ffn_up.weight": "fc5ae8d488894cbd4951059675468d227da27871d26e925c9941863841c097ee",
+  "blk.4.attn_k.weight": "14833b078cc4c5137bdd5fdc0538047974ca147a99b0282e1b144440c78bc1db",
+  "blk.4.attn_norm.weight": "0a69957d4a15599fb80ad4753558020804925221457d9a5052926754d3768065",
+  "blk.4.attn_output.weight": "887a49b6130fb6297cf10767207c3dd97191b2cf63723449af9c27bca8dbeda0",
+  "blk.4.attn_q.weight": "51fd577b76764824dd6f0d4891c137ebe4736f591b5ca2793c5fff2be49abbde",
+  "blk.4.attn_v.weight": "1a623c43cf9c509d1b7ea0d1a5c04d0af4809665f9f9e93b7d6dba8c5df178fa",
+  "blk.4.ffn_down.weight": "5d61e8856d8941d2b1fd138116d015f63840d0fa1e31e20e20a5ceca1536ceec",
+  "blk.4.ffn_gate.weight": "06640f7273764f8ca5df7e386547417916b6cd7d565a8343153113239a94b0a1",
+  "blk.4.ffn_norm.weight": "91a6c6c41b894228e361435ecbc5058dca34d4911a23da5b56de219299c964d3",
+  "blk.4.ffn_up.weight": "d016dac1055e36d6a10b6317e57f98a904709ea892ef3194342f4d2f6326561e",
+  "blk.5.attn_k.weight": "987146afe124131500808cc0da33c06d207433656d41df6e6d8c99118a83bac5",
+  "blk.5.attn_norm.weight": "6b354938966f2608a2fb8d0f5b363ed0d8b0967c2ec8d0abd5c625b413042ded",
+  "blk.5.attn_output.weight": "cdcbfe02c6ff79d5326882b017a02099f5af71beedf6b1b3eb4de01e3a844536",
+  "blk.5.attn_q.weight": "b910d0cff781d3efb42eab0a302f46f286b2de717079175680d5b42bf8c309c8",
+  "blk.5.attn_v.weight": "66d3a279f747412f9f4b0e8abad44540c122ab2e811a7ee74c1f33bc36caade9",
+  "blk.5.ffn_down.weight": "c9b0efd2212981f16d956d8571f054b68780ad01f4917033647e359b557a4653",
+  "blk.5.ffn_gate.weight": "fe96b94109ca141c01f6a04788e20783019ca6ec334aa1f3134810bdb499e557",
+  "blk.5.ffn_norm.weight": "aa7b016e832e7055a36c6e20de58ea1936f995f390401fff1c5fc65906064e49",
+  "blk.5.ffn_up.weight": "555ce27c4873d3375394f38ad3b45e3d8848f9d5642dc1602383d0f0a33c2a14",
+  "blk.6.attn_k.weight": "88280d461db324c4f36475ce396793063e61a27283ec64511b0480890fb5b3b4",
+  "blk.6.attn_norm.weight": "af8f460c411f660d33196286d208f1845fd5a2b45f7b56549a4df31e7515447a",
+  "blk.6.attn_output.weight": "dd9996fb0a256e8375ad3917705258a33fce006bcea0f536caae420a77974d8b",
+  "blk.6.attn_q.weight": "7a4841541191e037cfb9b07930c4d8cab451809658b182f0ada6ccde9615c003",
+  "blk.6.attn_v.weight": "ae81e6a592b64d701a9d40233e986039a56cba8d8d24f61aea93c6393cf3078a",
+  "blk.6.ffn_down.weight": "622dd1ce1706355cbc659a8ab2c4509678ffe0f3ad34258e5e25ed2a5d951bcd",
+  "blk.6.ffn_gate.weight": "8389a735c0bd5591010f8ced9805a2a12c749f6df0d3c18ad4d05c2a302e7168",
+  "blk.6.ffn_norm.weight": "621f5346400382474d61358397bd58fb1459b07c53e376e4bca15e08b3f9b3fb",
+  "blk.6.ffn_up.weight": "8d834e4c42f13c251dfee36cf89e12f1bd400680d00d5c2e6cac0459e9ce2f7f",
+  "blk.7.attn_k.weight": "8bd0412de65a3e64901ef8fe6a28c95e116bf39dc9aa22f0126b9d36688e5ea7",
+  "blk.7.attn_norm.weight": "056d8e56be4e87d6dc6f900762f0dc6fde07bfdc50dd85bfc510415e2bba3f3d",
+  "blk.7.attn_output.weight": "27972eda51da53d416ff95aed78149a2c5a287b47d2cd46f2f544ca692ecb3bb",
+  "blk.7.attn_q.weight": "41eca977b9371f7932800c11a9c45b931310196919e2a0651b847703b180fc7f",
+  "blk.7.attn_v.weight": "13c74fd7e07f08883a09fb070a1fe5bbdd2341b4cb8d1cac07c4b637049b5774",
+  "blk.7.ffn_down.weight": "9e75db42468800849a9a7da603d0072c5e86c8ed2b4d8b20a312a51fb86a7a10",
+  "blk.7.ffn_gate.weight": "db6bdc3117f910088aaf7db51f2da63ea5bd933de36af5599c215bfb26f7db2b",
+  "blk.7.ffn_norm.weight": "48bb82b49bfc8679a1e77f282ee182d952db7a3c11be7ef9a102ee2ddd8011e2",
+  "blk.7.ffn_up.weight": "feebea87175817a0f3585ec0af09dc873d94c203581ae97a712eb356d3b49efe",
+  "blk.8.attn_k.weight": "d5640ad71b6af68d88e17bf8e7fc26c907d2262605457a84247dd9afc2884d69",
+  "blk.8.attn_norm.weight": "75b850c481a69083ae09d0207ba7317b37c735a39fcf5fef5400e6c84fb1257f",
+  "blk.8.attn_output.weight": "cbd669dbdea2bdd90f9f0cc97566b3dffff3c56cecb4f47290ceef30da83b2d6",
+  "blk.8.attn_q.weight": "9edcb63087a431bac361822497e6ecdaa06d9ea4a1a754e36da7ba9f8db81c7c",
+  "blk.8.attn_v.weight": "3fb72c2c4f95a83626aa3e30062f9450b09ab37c7871e229f18bbc5cf744633c",
+  "blk.8.ffn_down.weight": "bd69d2c9172974fff154441b237b4787fb53b2d185325442d5048130ef5bc4ef",
+  "blk.8.ffn_gate.weight": "d04689c80553edd011d1cbaa5d570fffa7fa91e88b66cf1352d89ab60b72f908",
+  "blk.8.ffn_norm.weight": "e49984183b735b7f2c4e4730c289eed9394056d2e283a00fd83ea0915df31a73",
+  "blk.8.ffn_up.weight": "8fe62a1ce8e847e567add6c6f6bf2922bc467495b5eb4c116b3cb85b85b3b211",
+  "blk.9.attn_k.weight": "d90904959e5004cf0d6e729c6bff18cc33c094798b802473c1ec55ab8d276183",
+  "blk.9.attn_norm.weight": "79277f290cc07411115d8fa138045edf4a17b3416ab2145409cbe8ab829fd4ee",
+  "blk.9.attn_output.weight": "5a21bf2e1f09a81405025f96d4153ffb630158e17269cff8ffff935c38ceb1a7",
+  "blk.9.attn_q.weight": "51b1d0febc3b350945be4504f55afa4347517bde0f710e1a4b88e6b17e71e7c7",
+  "blk.9.attn_v.weight": "aab7e1db0a8b50a03036356791ffce736ab010d15674c96eaef8049d80076054",
+  "blk.9.ffn_down.weight": "cbf43ec84becb40c9359a181ab0e641fd7faae7d34b549501f7cfb7afdc3d764",
+  "blk.9.ffn_gate.weight": "dce0e8661c778327bed7f03b6790d26710764188aed9dc746e6e05863891fa57",
+  "blk.9.ffn_norm.weight": "6d41642104f995c77bf31122b13237caebda3e7fcccb1367ce91db36b015e923",
+  "blk.9.ffn_up.weight": "82fe4c67bf24e7b2d6f6e05f7b1234c2bf90c3932951091a9066211b8e15ecbb",
+  "blk.10.attn_k.weight": "f6a9ed8fd8d3229b5d03175c413ffc56a07f2ce7236271986361dd3d8993f9aa",
+  "blk.10.attn_norm.weight": "cebbef89f0326ca8e02df3867a571e4d61c20c2a12f295f98ae590d62bc86010",
+  "blk.10.attn_output.weight": "34f5efb86accb4f06347d83a32558ea8eab3039d128969161a741ebacbb656ff",
+  "blk.10.attn_q.weight": "1e0efe27df2d5d50f7157253ba2cfd436d6781c3dc78ca176d0c16a210b5b763",
+  "blk.10.attn_v.weight": "8f085bf50a2b0f83cd6cdda3c8ef5a9e204a36348ed95871aac725d1f68640cf",
+  "blk.10.ffn_down.weight": "bf3b3cb4cace435809ac7b4cc933f20853af12f1f272d3dcefe7f19c0f203b8b",
+  "blk.10.ffn_gate.weight": "d3df7a1413b1c5adf1a1dcda9e5225a15c89874bae53bb6137ad1ea42fca2d34",
+  "blk.10.ffn_norm.weight": "a1da603b0480471b5ed8e862148cecd5fed918f8304d6933ab0bdb25b8d2fb8f",
+  "blk.10.ffn_up.weight": "bffbba605922e972dc47dda88a0b4659aa52236c76e5fe861a949e6d9a367492",
+  "blk.11.attn_k.weight": "9f31c63d66cd32c29b1eb8bb829d0c8525ce2ae936e0eefdaab6335a2d12a3df",
+  "blk.11.attn_norm.weight": "0bde1a266d8b2e8f202bb7e2e88b19147ca83021901f6d3cae77a4df5548c754",
+  "blk.11.attn_output.weight": "e10725c7cf746ed4a7e472cf7aea6cb564e5db6a1d5197adc980d650a387ccea",
+  "blk.11.attn_q.weight": "05ee758a7d065802630f8c65dca424364c1c8825e389aa33f9405c45e8a50cce",
+  "blk.11.attn_v.weight": "0c3ae7090f11775d24c51120db6e305db6aff706493e7ee123dcab74485ba789",
+  "blk.11.ffn_down.weight": "7ba40b8e12c09c5fb2006b77a771cb01ce894e88a3b3e1877f927a5b89c91709",
+  "blk.11.ffn_gate.weight": "db76388a023b98097972d354ba1c6a5e26efdeb1c596b9c28bf2cd8f6596975e",
+  "blk.11.ffn_norm.weight": "a38c3ae1b89a68ddc7b72c99c5b28be7fe3787c4fad9904d0c43d64eaf00c474",
+  "blk.11.ffn_up.weight": "13c8142f9cf1eddc658babf978daf3515c4ccc45f849f3e7e3930aa18a8480a0",
+  "blk.12.attn_k.weight": "f03241c36ac87cb57429a2ef22186b8d7d0b590a8b173beb01fa13d93772f3b1",
+  "blk.12.attn_norm.weight": "4568f654e6d65104d586e7c16ba960c83428698ce103022b7e0be15e2884e13b",
+  "blk.12.attn_output.weight": "04867603f82f91e41306e09b33ecda0104b3ee4834061f2c0bbdc8da33c72509",
+  "blk.12.attn_q.weight": "70fe04b9a8e08b6100cc8d6b58bf4cbbad15ca1de82d63baca5d352ba6c4cbae",
+  "blk.12.attn_v.weight": "15cb28db61a86c98687991d7e611bc92a1fcc6007f3432149cfb5fe518a4f65e",
+  "blk.12.ffn_down.weight": "6d10c790a4e3dc44c2dc36d96251ae97cdf30a4fa04d4c43e31bfbd038e6a7b7",
+  "blk.12.ffn_gate.weight": "3462a2d8f6b4743b25e24da51b90018ac2858d05ac7e582bcb69063cfdac1104",
+  "blk.12.ffn_norm.weight": "1f96392c1faa34e34ae5dea55a6a86c5aa4c79758952075d53d28de89dd88456",
+  "blk.12.ffn_up.weight": "d22eacc612a7411953d948483c5fb201e11722955ee0754da866e7bec578ac6d",
+  "blk.13.attn_k.weight": "5864977e6b733ea942647d6feed5c76156c48c200649c22e4e11b9e5860e57f3",
+  "blk.13.attn_norm.weight": "87e053535144723db4145aa5402acc54331b7696752d852bb9fc542ff33f0fb5",
+  "blk.13.attn_output.weight": "078145f5ad83f8b14f97a869346f7fd1583b24d1e3edadaa95d3da4242973f8f",
+  "blk.13.attn_q.weight": "3b8caf35504cbc4d1a7dd6e011a95760703b7f71e2218b030b1254f811362dd7",
+  "blk.13.attn_v.weight": "4fdf8365a603e043e5b40c4a21c84ac167f9be62794178f9d8a608dfe5653bf9",
+  "blk.13.ffn_down.weight": "a07d3abbfcacf48ba028df2cab895be32cc15022d23389a745286e79c1b1d1fd",
+  "blk.13.ffn_gate.weight": "1d2ab39666aa2909acc96787432a3ed13b19d25170f74665fadff9b17bbaffb1",
+  "blk.13.ffn_norm.weight": "4f2e809fda5f3eadf52578ee50e0ba36e53be91e55dce418c12dfe595f5f18e7",
+  "blk.13.ffn_up.weight": "8783d2720c2c37ca176a5801e0b3ef1f9cc9cf3ef1cd37af423aaf6b2a27e2bd",
+  "blk.14.attn_k.weight": "ce9428e2b55d43ae0c6690dbd56182f99adc427694ba8236b405cc8ea5035e86",
+  "blk.14.attn_norm.weight": "6abb35f9db8251d6ae954bda147c6ada2371b0574d11702e828f3c6ac99b7cc0",
+  "blk.14.attn_output.weight": "fe3880916d0ceb5bff672c88bbefb7060a545be609bf049beb2024b38221836d",
+  "blk.14.attn_q.weight": "7c8ad81be6f4a350931fd108b5f7c9e366e8c26ef62d1d85ffef5dca8fd893f8",
+  "blk.14.attn_v.weight": "e4bdedffacbebe38567a0734dfd67db90e911d9a9669fcde9a7c4ad8a0066c52",
+  "blk.14.ffn_down.weight": "ef6694dff1e05820aac0cd2b22f39ac7788b4967afc9250775575554c66aab2c",
+  "blk.14.ffn_gate.weight": "db63c4179e2db704bc505e2b4696e055b593e295a1b7c4c586fc793bdd5aab19",
+  "blk.14.ffn_norm.weight": "2796a62d832a9710148f95d533320492a33e712b2e5218659c548705bd11684d",
+  "blk.14.ffn_up.weight": "3f78c78d8c2d54df45f799d4ff902316628af296834afe4ceed63d4a324ff03e",
+  "blk.15.attn_k.weight": "6e810ee3859e07695645ee0c9a5efc7962668984a5f0a9325f47e462743b447c",
+  "blk.15.attn_norm.weight": "0956b576ae96db0b28cb09f761f801cfd9281432284664f0fe181c8d9c55d1ec",
+  "blk.15.attn_output.weight": "03a17f7e94208177aace5cc41b7f54670ba57873b7274ff6e23caf58cce110ca",
+  "blk.15.attn_q.weight": "b8edafe7d2216a6f8b4ae4905a906475490e6ea418f6e1d3cec563dbdc6fab91",
+  "blk.15.attn_v.weight": "f8ae8cae0f4cfa34a459824eba57350c3c248104ba5607e7d9dc7d7c39aaf4a6",
+  "blk.15.ffn_down.weight": "8d02eb439da852246d2ca67e9b7b6de0b090b80744355e64728a23e41926505b",
+  "blk.15.ffn_gate.weight": "ed5bf361c67db8731f186b775826f21c33bdb521111fd2d922539719a770239f",
+  "blk.15.ffn_norm.weight": "5942ca3c73209ac9a0c8bfd9b4aab7f7be7aee9aa12d9c35833493b44af76767",
+  "blk.15.ffn_up.weight": "f4bebf4ad99ec5f911327dec347be6c595814885309c7bc5647ce28c7f4d1cf5",
+  "blk.16.attn_k.weight": "756a534c19364448e0958b8948fe33891c6ccda0fbb4dfa2024e1f532a87804b",
+  "blk.16.attn_norm.weight": "386b7b9e4e6509f6af9c022d942b6c6c6cc136aeed8751ecb037c74d7c4bfb93",
+  "blk.16.attn_output.weight": "3ba1a766a25830b84d7c22178203635f9c5624caad290bc5e5d73da5d5e7a2ec",
+  "blk.16.attn_q.weight": "d39b0c91e1fda7685d50a0f7cc8d18c44b5bdc90a142c7fda0bc329cca1afa74",
+  "blk.16.attn_v.weight": "98b33fcb0ee3483cff1b06ecb44d7b7ffb4d34c268248e4d73dfdf82b2065b2f",
+  "blk.16.ffn_down.weight": "14006f5e4acb2f9416271ae562e299359cd2585739c7fc77ccbca54495563948",
+  "blk.16.ffn_gate.weight": "12f8abae2d301d8f88bedb6af98b1daecc7b0b8d05148594f931f30958d77aca",
+  "blk.16.ffn_norm.weight": "129a15a046ee96d06de288bd43c80f77a6b0fb3a159c7367154c6e4aaf362672",
+  "blk.16.ffn_up.weight": "b4a5911a45f3871ef1d4efb7dc7108645a564b70f818eccf45beebef2e844ee9",
+  "blk.17.attn_k.weight": "5e1bfcff0146ebdde3817b656952892eb671e14e75afc92fa53f84f8eecbec4c",
+  "blk.17.attn_norm.weight": "60bc988fab7c4b29ee9de599df41a8de00caa94fcd74677da011fac82f60f465",
+  "blk.17.attn_output.weight": "ba49b40d6a0b5685f749c24b0edbed3adc44dbe13b5d5e5fa1e56169fc746555",
+  "blk.17.attn_q.weight": "82bb415d24efcd14d03ace03f907bb70db6a204c76a0bdd1892e0fba165db87d",
+  "blk.17.attn_v.weight": "73dbe54beb91a899884e275ea81ffc5187a20cb7d5b68d5c299b783096999d94",
+  "blk.17.ffn_down.weight": "7c086166241e0664f8963fd1ca4ed74c737abfb2525ec20f8435821ff50158f3",
+  "blk.17.ffn_gate.weight": "51a32f78244d42a539f619c5ce661db9e6cf41636280a826d439b5444edcd28c",
+  "blk.17.ffn_norm.weight": "c4bb247fccd1ecc84875028af63dd20aaf5cbd17eb94a9bc36679c09285dccab",
+  "blk.17.ffn_up.weight": "b5886182790bc6fbadd63de9bc4ffee416f3b69a66280d197ab8c18edf769abf",
+  "output_norm.weight": "481f3097d0a20412e35b3a739b1b958487bcd41ff67744baa3c9acbddd2ee4d4"
+}
diff --git a/convert/tokenizer.go b/convert/tokenizer.go
index fd6df5f53..14d6ba66c 100644
--- a/convert/tokenizer.go
+++ b/convert/tokenizer.go
@@ -1,10 +1,12 @@
 package convert
 
 import (
-	"cmp"
 	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
+	"io/fs"
 	"log/slog"
 	"os"
 	"slices"
@@ -12,10 +14,152 @@ import (
 	"golang.org/x/exp/maps"
 )
 
+const (
+	_ int32 = iota
+	tokenTypeNormal
+	tokenTypeUnknown
+	tokenTypeControl
+	tokenTypeUserDefined
+	tokenTypeUnused
+	tokenTypeByte
+)
+
 type Tokenizer struct {
-	Version     string         `json:"version"`
-	AddedTokens []Token        `json:"added_tokens"`
-	Model       TokenizerModel `json:"model"`
+	*Vocabulary
+	SpecialVocabulary []*SpecialVocabulary
+	Merges            []string
+
+	Pre      string
+	Template string
+}
+
+func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error) {
+	v, err := parseVocabulary(fsys)
+	if err != nil {
+		return nil, err
+	}
+
+	t := &Tokenizer{
+		Vocabulary: v,
+		Pre:        "default",
+	}
+
+	addedTokens := make(map[string]token)
+	if f, err := fsys.Open("tokenizer.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var tt tokenizer
+		if err := json.NewDecoder(f).Decode(&tt); err != nil {
+			return nil, err
+		}
+
+		for _, t := range tt.AddedTokens {
+			addedTokens[t.Content] = t
+		}
+
+		t.Merges = tt.Model.Merges
+
+		sha256sum := sha256.New()
+		for _, pt := range tt.PreTokenizer.PreTokenizers {
+			switch pt.Type {
+			case "Split":
+				if pt.Pattern.Regex != "" {
+					// create a checksum of all Split pretokenizers which should be sufficient
+					// to identify the pretokenizer
+					sha256sum.Write([]byte(pt.Pattern.Regex))
+				}
+			}
+		}
+
+		switch digest := hex.EncodeToString(sha256sum.Sum(nil)); digest {
+		case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
+			t.Pre = "llama-bpe"
+		case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
+			t.Pre = "deepseek-llm"
+		case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
+			t.Pre = "deepseek-coder"
+		case "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855":
+			// noop, empty pretokenizer
+		default:
+			slog.Warn("unknown pretokenizer, using default", "digest", digest)
+		}
+	}
+
+	if f, err := fsys.Open("tokenizer_config.json"); errors.Is(err, os.ErrNotExist) {
+	} else if err != nil {
+		return nil, err
+	} else {
+		defer f.Close()
+
+		var p map[string]json.RawMessage
+		if err := json.NewDecoder(f).Decode(&p); err != nil {
+			return nil, err
+		}
+
+		if template, ok := p["chat_template"]; ok {
+			var s []struct {
+				Name     string `json:"name"`
+				Template string `json:"template"`
+			}
+			if err := json.Unmarshal(template, &t.Template); err == nil {
+				// noop
+			} else if err := json.Unmarshal(template, &s); err == nil {
+				for _, e := range s {
+					if e.Name == "default" {
+						t.Template = e.Template
+						break
+					}
+				}
+			} else {
+				return nil, fmt.Errorf("invalid chat_template: %w", err)
+			}
+		}
+
+		for _, st := range specialTokenTypes {
+			sv := SpecialVocabulary{Type: st}
+			if bts, ok := p[fmt.Sprintf("add_%s_token", st)]; ok {
+				if err := json.Unmarshal(bts, &sv.AddToken); err != nil {
+					return nil, err
+				}
+			}
+
+			if bts, ok := p[fmt.Sprintf("%s_token", st)]; ok {
+				var content string
+				if err := json.Unmarshal(bts, &content); err != nil {
+					var mm map[string]any
+					if err := json.Unmarshal(bts, &mm); err != nil {
+						continue
+					}
+
+					content, ok = mm["content"].(string)
+					if !ok {
+						continue
+					}
+				}
+
+				sv.Content = content
+			}
+
+			if id, ok := addedTokens[sv.Content]; ok {
+				sv.ID = id.ID
+				t.SpecialVocabulary = append(t.SpecialVocabulary, &sv)
+			}
+		}
+	}
+
+	return t, nil
+}
+
+type tokenizer struct {
+	AddedTokens []token `json:"added_tokens"`
+	Model       struct {
+		Type   string         `json:"type"`
+		Vocab  map[string]int `json:"vocab"`
+		Merges []string       `json:"merges"`
+	} `json:"model"`
 
 	PreTokenizer struct {
 		PreTokenizers []struct {
@@ -27,80 +171,108 @@ type Tokenizer struct {
 	} `json:"pre_tokenizer"`
 }
 
-type TokenizerModel struct {
-	Type   string         `json:"type"`
-	Vocab  map[string]int `json:"vocab"`
-	Merges []string       `json:"merges"`
-	Tokens []Token
-}
-
-type Token struct {
+type token struct {
 	ID          int    `json:"id"`
 	Content     string `json:"content"`
 	Special     bool   `json:"special"`
 	UserDefined bool
 }
 
-func (t *Token) Type() int32 {
-	switch {
-	case t.Special:
-		return tokenTypeControl
-	case t.UserDefined:
-		return tokenTypeUserDefined
-	default:
-		return tokenTypeNormal
-	}
+type Vocabulary struct {
+	Model  string
+	Tokens []string
+	Scores []float32
+	Types  []int32
 }
 
-func (t *Tokenizer) maxID() int {
-	return max(
-		slices.Max(maps.Values(t.Model.Vocab)),
-		slices.MaxFunc(t.AddedTokens, func(a, b Token) int {
-			return cmp.Compare(a.ID, b.ID)
-		}).ID,
-	)
-}
-
-func parseTokens(dirpath string) (pre string, tokens []Token, merges []string, err error) {
-	f, err := os.Open(dirpath)
+func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
+	f, err := fsys.Open("tokenizer.json")
 	if err != nil {
-		panic(err)
+		return nil, err
 	}
 	defer f.Close()
 
-	var t Tokenizer
+	var t tokenizer
 	if err := json.NewDecoder(f).Decode(&t); err != nil {
-		return "", nil, nil, err
+		return nil, err
 	}
 
-	tokens = make([]Token, t.maxID()+1)
+	tokens := make(map[int]token, len(t.Model.Vocab))
 	for k, v := range t.Model.Vocab {
-		tokens[v] = Token{ID: v, Content: k, Special: false, UserDefined: false}
-	}
-
-	for _, v := range t.AddedTokens {
-		v.UserDefined = true
-		tokens[v.ID] = v
-	}
-
-	sha256sum := sha256.New()
-	for _, pt := range t.PreTokenizer.PreTokenizers {
-		if pt.Type == "Split" && pt.Pattern.Regex != "" {
-			sha256sum.Write([]byte(pt.Pattern.Regex))
+		tokens[v] = token{
+			ID:      v,
+			Content: k,
 		}
 	}
 
-	switch digest := fmt.Sprintf("%x", sha256sum.Sum(nil)); digest {
-	case "d98f9631be1e9607a9848c26c1f9eac1aa9fc21ac6ba82a2fc0741af9780a48f":
-		pre = "llama-bpe"
-	case "03df5c5863ad70781dcfdef491ead25140f895fe8010964be0daefe27be32b02":
-		pre = "deepseek-llm"
-	case "21cde974d587f0d54dc8d56b183cc1e6239600172035c68fbd6d4b9f8da0576e":
-		pre = "deepseek-coder"
-	default:
-		slog.Warn("unknown pretokenizer, using default", "digest", digest)
-		pre = "default"
+	for _, token := range t.AddedTokens {
+		token.UserDefined = true
+		tokens[token.ID] = token
 	}
 
-	return pre, tokens, t.Model.Merges, nil
+	keys := maps.Keys(tokens)
+	slices.Sort(keys)
+
+	v := Vocabulary{Model: "gpt2"}
+	for _, k := range keys {
+		token := tokens[k]
+		v.Tokens = append(v.Tokens, token.Content)
+		v.Scores = append(v.Scores, float32(token.ID))
+
+		switch {
+		case token.Special:
+			v.Types = append(v.Types, tokenTypeControl)
+		case token.UserDefined:
+			v.Types = append(v.Types, tokenTypeUserDefined)
+		default:
+			v.Types = append(v.Types, tokenTypeNormal)
+		}
+	}
+
+	return &v, nil
+}
+
+func parseVocabulary(fsys fs.FS) (*Vocabulary, error) {
+	patterns := []struct {
+		Pattern string
+		Func    func(fs.FS) (*Vocabulary, error)
+	}{
+		{"tokenizer.model", parseSentencePiece},
+		{"tokenizer.json", parseVocabularyFromTokenizer},
+	}
+
+	for _, pattern := range patterns {
+		if _, err := fs.Stat(fsys, pattern.Pattern); errors.Is(err, os.ErrNotExist) {
+			continue
+		} else if err != nil {
+			return nil, err
+		}
+
+		return pattern.Func(fsys)
+	}
+
+	return nil, errors.New("unknown tokenizer format")
+}
+
+type SpecialVocabulary struct {
+	Type     string
+	ID       int
+	Content  string
+	AddToken bool
+}
+
+func (sv SpecialVocabulary) Key() string {
+	switch t := sv.Type; t {
+	case "bos", "eos", "cls", "mask":
+		return t
+	case "unk":
+		return "unknown"
+	case "sep":
+		//nolint:misspell // this is an upstream typo
+		return "seperator"
+	case "pad":
+		return "padding"
+	}
+
+	panic("unknown special vocabulary type")
 }
diff --git a/convert/tokenizer_spm.go b/convert/tokenizer_spm.go
new file mode 100644
index 000000000..5e506087c
--- /dev/null
+++ b/convert/tokenizer_spm.go
@@ -0,0 +1,113 @@
+package convert
+
+import (
+	"cmp"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"slices"
+
+	"google.golang.org/protobuf/proto"
+
+	"github.com/ollama/ollama/convert/sentencepiece"
+)
+
+func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	ast, err := parseAdditionalSpecialTokens(fsys)
+	if err != nil {
+		return nil, err
+	}
+
+	bts, err := fs.ReadFile(fsys, "tokenizer.model")
+	if err != nil {
+		return nil, err
+	}
+
+	var spm sentencepiece.ModelProto
+	if err := proto.Unmarshal(bts, &spm); err != nil {
+		return nil, err
+	}
+
+	v := Vocabulary{Model: "llama"}
+	for _, piece := range spm.GetPieces() {
+		v.Tokens = append(v.Tokens, piece.GetPiece())
+		v.Scores = append(v.Scores, piece.GetScore())
+
+		switch t := piece.GetType(); t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN,
+			sentencepiece.ModelProto_SentencePiece_CONTROL,
+			sentencepiece.ModelProto_SentencePiece_UNUSED,
+			sentencepiece.ModelProto_SentencePiece_BYTE:
+			v.Types = append(v.Types, int32(t))
+		default:
+			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
+			if slices.Contains(ast, piece.GetPiece()) {
+				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+			}
+
+			v.Types = append(v.Types, tt)
+		}
+	}
+
+	f, err := fsys.Open("added_tokens.json")
+	if errors.Is(err, os.ErrNotExist) {
+		return &v, nil
+	} else if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var atm map[string]int
+	if err := json.NewDecoder(f).Decode(&atm); err != nil {
+		return nil, err
+	}
+
+	type t struct {
+		id      int
+		content string
+	}
+
+	var ts []t
+	for content, id := range atm {
+		ts = append(ts, t{id, content})
+	}
+
+	slices.SortFunc(ts, func(i, j t) int {
+		return cmp.Compare(i.id, j.id)
+	})
+
+	n := len(v.Tokens)
+	for i, t := range ts {
+		if t.id != i+n {
+			return nil, fmt.Errorf("invalid token id: %d", t.id)
+		}
+
+		v.Tokens = append(v.Tokens, t.content)
+		v.Scores = append(v.Scores, -1000.0)
+		v.Types = append(v.Types, tokenTypeUserDefined)
+	}
+
+	return &v, nil
+}
+
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
+	f, err := fsys.Open("special_tokens_map.json")
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	} else if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var m struct {
+		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
+	}
+
+	if err := json.NewDecoder(f).Decode(&m); err != nil {
+		return nil, err
+	}
+
+	return m.AdditionalSpecialTokens, nil
+}
diff --git a/convert/tokenizer_test.go b/convert/tokenizer_test.go
new file mode 100644
index 000000000..d9550e095
--- /dev/null
+++ b/convert/tokenizer_test.go
@@ -0,0 +1,208 @@
+package convert
+
+import (
+	"io"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func createTokenizerFS(t *testing.T, dir string, files map[string]io.Reader) fs.FS {
+	t.Helper()
+
+	for k, v := range files {
+		if err := func() error {
+			f, err := os.Create(filepath.Join(dir, k))
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+
+			if _, err := io.Copy(f, v); err != nil {
+				return err
+			}
+
+			return nil
+		}(); err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+	}
+
+	return os.DirFS(dir)
+}
+
+func TestParseTokenizer(t *testing.T) {
+	cases := []struct {
+		name              string
+		fsys              fs.FS
+		specialTokenTypes []string
+		want              *Tokenizer
+	}{
+		{
+			name: "string chat template",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"chat_template": "<default template>"
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{Model: "gpt2"},
+				Pre:        "default",
+				Template:   "<default template>",
+			},
+		},
+		{
+			name: "list chat template",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"chat_template": [
+						{
+							"name": "default",
+							"template": "<default template>"
+						},
+						{
+							"name": "tools",
+							"template": "<tools template>"
+						}
+					]
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{Model: "gpt2"},
+				Pre:        "default",
+				Template:   "<default template>",
+			},
+		},
+		{
+			name: "added tokens",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 999,
+							"content": "<unused999>",
+							"special": false
+						}
+					]
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<unused999>"},
+					Scores: []float32{999},
+					Types:  []int32{4},
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "added tokens overlap vocab",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<pad>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<pad>": 0
+						}
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<pad>"},
+					Scores: []float32{0},
+					Types:  []int32{3},
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "special token types",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<pad>",
+							"special": true
+						},
+						{
+							"id": 1,
+							"content": "<eos>",
+							"special": true
+						},
+						{
+							"id": 2,
+							"content": "<bos>",
+							"special": true
+						},
+						{
+							"id": 3,
+							"content": "<unk>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<pad>": 0,
+							"<eos>": 1,
+							"<bos>": 2,
+							"<unk>": 3
+						}
+					}
+				}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"add_bos_token": true,
+					"add_eos_token": false,
+					"bos_token": "<bos>",
+					"eos_token": "<eos>",
+					"pad_token": "<pad>",
+					"unk_token": "<unk>"
+				}`),
+			}),
+			specialTokenTypes: []string{"pad", "eos", "bos", "unk"},
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<pad>", "<eos>", "<bos>", "<unk>"},
+					Scores: []float32{0, 1, 2, 3},
+					Types:  []int32{3, 3, 3, 3},
+				},
+				SpecialVocabulary: []*SpecialVocabulary{
+					{Type: "pad", Content: "<pad>", ID: 0, AddToken: false},
+					{Type: "eos", Content: "<eos>", ID: 1, AddToken: false},
+					{Type: "bos", Content: "<bos>", ID: 2, AddToken: true},
+					{Type: "unk", Content: "<unk>", ID: 3, AddToken: false},
+				},
+				Pre: "default",
+			},
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			tokenizer, err := parseTokenizer(tt.fsys, tt.specialTokenTypes)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			if diff := cmp.Diff(tt.want, tokenizer); diff != "" {
+				t.Errorf("unexpected tokenizer (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
diff --git a/convert/torch.go b/convert/torch.go
deleted file mode 100644
index 55414adc6..000000000
--- a/convert/torch.go
+++ /dev/null
@@ -1,287 +0,0 @@
-package convert
-
-import (
-	"encoding/binary"
-	"encoding/json"
-	"fmt"
-	"io"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"regexp"
-	"strings"
-
-	"github.com/nlpodyssey/gopickle/pytorch"
-	"github.com/nlpodyssey/gopickle/types"
-	"github.com/x448/float16"
-
-	"github.com/ollama/ollama/llm"
-)
-
-type torchWriterTo struct {
-	t *llm.Tensor
-
-	params *Params
-	bo     ByteOrder
-
-	storage  pytorch.StorageInterface
-	repacker func(string, []float32, []uint64) ([]float32, error)
-}
-
-type TorchFormat struct{}
-
-func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
-	slog.Debug("getting torch tensors")
-
-	var files []string
-	if pt, _ := filepath.Glob(filepath.Join(dirpath, "consolidated*.pth")); len(pt) > 0 {
-		files = append(files, pt...)
-	} else if pt, _ := filepath.Glob(filepath.Join(dirpath, "pytorch_model*.pth")); len(pt) > 0 {
-		files = append(files, pt...)
-	}
-
-	var offset uint64
-	var tensors []llm.Tensor
-	for _, fn := range files {
-		m, err := pytorch.Load(fn)
-		if err != nil {
-			slog.Error(fmt.Sprintf("error unpickling: %q", err))
-			return []llm.Tensor{}, err
-		}
-
-		for _, k := range m.(*types.Dict).Keys() {
-			if strings.HasSuffix(k.(string), "self_attn.rotary_emb.inv_freq") {
-				continue
-			}
-
-			t, _ := m.(*types.Dict).Get(k)
-			tshape := t.(*pytorch.Tensor).Size
-
-			var size uint64
-			var kind uint32
-			switch len(tshape) {
-			case 0:
-				continue
-			case 1:
-				// convert to float32
-				kind = 0
-				size = uint64(tshape[0] * 4)
-			case 2:
-				// convert to float16
-				kind = 1
-				size = uint64(tshape[0] * tshape[1] * 2)
-			}
-
-			ggufName, err := tf.GetLayerName(k.(string))
-			if err != nil {
-				slog.Error(err.Error())
-				return nil, err
-			}
-			slog.Debug(fmt.Sprintf("'%35s': '%30s' %10d [%#v]", k.(string), ggufName, size, tshape))
-
-			shape := []uint64{0, 0, 0, 0}
-			for i := range tshape {
-				shape[i] = uint64(tshape[i])
-			}
-
-			tensor := llm.Tensor{
-				Name:   ggufName,
-				Kind:   kind,
-				Offset: offset, // calculate the offset
-				Shape:  shape,
-			}
-
-			tensor.WriterTo = torchWriterTo{
-				t:       &tensor,
-				params:  params,
-				bo:      params.ByteOrder,
-				storage: t.(*pytorch.Tensor).Source,
-			}
-
-			tensors = append(tensors, tensor)
-			offset += size
-		}
-	}
-
-	return tensors, nil
-}
-
-func getAltParams(dirpath string) (*Params, error) {
-	f, err := os.Open(filepath.Join(dirpath, "params.json"))
-	if err != nil {
-		slog.Error("no params.json")
-		return nil, err
-	}
-	defer f.Close()
-
-	type TorchParams struct {
-		HiddenSize     int     `json:"dim"`
-		AttentionHeads int     `json:"n_heads"`
-		KeyValHeads    int     `json:"n_kv_heads"`
-		HiddenLayers   int     `json:"n_layers"`
-		RopeTheta      float64 `json:"rope_theta"`
-		NormEPS        float64 `json:"norm_eps"`
-	}
-
-	var tparams TorchParams
-
-	d := json.NewDecoder(f)
-	err = d.Decode(&tparams)
-	if err != nil {
-		return nil, err
-	}
-
-	params := &Params{
-		Architectures:  []string{"LlamaForCausalLM"},
-		HiddenSize:     tparams.HiddenSize,
-		AttentionHeads: tparams.AttentionHeads,
-		KeyValHeads:    tparams.KeyValHeads,
-		HiddenLayers:   tparams.HiddenLayers,
-		NormEPS:        tparams.NormEPS,
-	}
-
-	switch {
-	case tparams.RopeTheta == 1000000:
-		// Codellama
-		params.ContextSize = 16384
-	case tparams.NormEPS == 1e-06:
-		// llama2
-		slog.Debug("Found llama2 - setting context size to 4096")
-		params.ContextSize = 4096
-	default:
-		params.ContextSize = 2048
-	}
-
-	params.ByteOrder = binary.LittleEndian
-	return params, nil
-}
-
-func (m *TorchFormat) GetParams(dirpath string) (*Params, error) {
-	f, err := os.Open(filepath.Join(dirpath, "config.json"))
-	if err != nil {
-		if os.IsNotExist(err) {
-			// try params.json instead
-			return getAltParams(dirpath)
-		} else {
-			return nil, err
-		}
-	}
-
-	var params Params
-	d := json.NewDecoder(f)
-	err = d.Decode(&params)
-	if err != nil {
-		return nil, err
-	}
-
-	params.ByteOrder = binary.LittleEndian
-	return &params, nil
-}
-
-func (m *TorchFormat) GetLayerName(n string) (string, error) {
-	directMap := map[string]string{
-		"tok_embeddings.weight":     "token_embd.weight",
-		"output.weight":             "output.weight",
-		"norm.weight":               "output_norm.weight",
-		"rope.freqs":                "rope_freqs.weight",
-		"model.embed_tokens.weight": "token_embd.weight",
-		"lm_head.weight":            "output.weight",
-		"model.norm.weight":         "output_norm.weight",
-	}
-
-	lMap := map[string]string{
-		"layers.(\\d+).attention_norm.weight":                 "blk.$1.attn_norm.weight",
-		"layers.(\\d+).attention_output_norm.weight":          "blk.$1.attn_norm.weight",
-		"layers.(\\d+).feed_forward.w2.weight":                "blk.$1.ffn_down.weight",
-		"layers.(\\d+).feed_forward.w1.weight":                "blk.$1.ffn_gate.weight",
-		"layers.(\\d+).feed_forward.w3.weight":                "blk.$1.ffn_up.weight",
-		"layers.(\\d+).ffn_norm.weight":                       "blk.$1.ffn_norm.weight",
-		"layers.(\\d+).attention.wk.weight":                   "blk.$1.attn_k.weight",
-		"layers.(\\d+).attention.wo.weight":                   "blk.$1.attn_output.weight",
-		"layers.(\\d+).attention.wq.weight":                   "blk.$1.attn_q.weight",
-		"layers.(\\d+).attention.wv.weight":                   "blk.$1.attn_v.weight",
-		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
-		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
-		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
-		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
-		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
-		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
-		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
-		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
-		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
-	}
-
-	v, ok := directMap[n]
-	if ok {
-		return v, nil
-	}
-
-	// quick hack to rename the layers to gguf format
-	for k, v := range lMap {
-		re := regexp.MustCompile(k)
-		newName := re.ReplaceAllString(n, v)
-		if newName != n {
-			return newName, nil
-		}
-	}
-
-	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
-}
-
-func (r torchWriterTo) WriteTo(w io.Writer) (n int64, err error) {
-	var f32s []float32
-	switch s := r.storage.(type) {
-	case *pytorch.FloatStorage:
-		f32s = s.Data
-	case *pytorch.HalfStorage:
-		f32s = s.Data
-	case *pytorch.BFloat16Storage:
-		f32s = s.Data
-	default:
-		return 0, fmt.Errorf("unknown data type: %T", s)
-	}
-
-	if r.repacker != nil {
-		f32s, err = r.repacker(r.t.Name, f32s, r.t.Shape)
-		if err != nil {
-			return 0, err
-		}
-	}
-
-	switch r.t.Kind {
-	case 0:
-		return 0, binary.Write(w, r.bo, f32s)
-	case 1:
-		f16s := make([]uint16, len(f32s))
-		for i := range f32s {
-			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
-		}
-
-		return 0, binary.Write(w, r.bo, f16s)
-	default:
-		return 0, fmt.Errorf("unknown storage type: %d", r.t.Kind)
-	}
-}
-
-func (m *TorchFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
-	switch len(params.Architectures) {
-	case 0:
-		return nil, fmt.Errorf("No architecture specified to convert")
-	case 1:
-		switch params.Architectures[0] {
-		case "LlamaForCausalLM":
-			return &LlamaModel{
-				ModelData{
-					Name:   name,
-					Path:   dirPath,
-					Params: params,
-					Format: m,
-				},
-			}, nil
-		default:
-			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
-		}
-	}
-
-	return nil, fmt.Errorf("Unknown error")
-}
diff --git a/docs/api.md b/docs/api.md
index c577bb1a5..95e79e007 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -40,6 +40,7 @@ Generate a response for a given prompt with a provided model. This is a streamin
 
 - `model`: (required) the [model name](#model-names)
 - `prompt`: the prompt to generate a response for
+- `suffix`: the text after the model response
 - `images`: (optional) a list of base64-encoded images (for multimodal models such as `llava`)
 
 Advanced parameters (optional):
@@ -57,7 +58,8 @@ Advanced parameters (optional):
 
 Enable JSON mode by setting the `format` parameter to `json`. This will structure the response as a valid JSON object. See the JSON mode [example](#request-json-mode) below.
 
-> Note: it's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace.
+> [!IMPORTANT]
+> It's important to instruct the model to use JSON in the `prompt`. Otherwise, the model may generate large amounts whitespace.
 
 ### Examples
 
@@ -67,7 +69,7 @@ Enable JSON mode by setting the `format` parameter to `json`. This will structur
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt": "Why is the sky blue?"
 }'
 ```
@@ -78,7 +80,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "response": "The",
   "done": false
@@ -100,7 +102,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "",
   "done": true,
@@ -122,7 +124,7 @@ A response can be received in one reply when streaming is off.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt": "Why is the sky blue?",
   "stream": false
 }'
@@ -134,7 +136,7 @@ If `stream` is set to `false`, the response will be a single JSON object:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -148,15 +150,51 @@ If `stream` is set to `false`, the response will be a single JSON object:
 }
 ```
 
+#### Request (with suffix)
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "codellama:code",
+  "prompt": "def compute_gcd(a, b):",
+  "suffix": "    return result",
+  "options": {
+    "temperature": 0
+  },
+  "stream": false
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "codellama:code",
+  "created_at": "2024-07-22T20:47:51.147561Z",
+  "response": "\n  if a == 0:\n    return b\n  else:\n    return compute_gcd(b % a, a)\n\ndef compute_lcm(a, b):\n  result = (a * b) / compute_gcd(a, b)\n",
+  "done": true,
+  "done_reason": "stop",
+  "context": [...],
+  "total_duration": 1162761250,
+  "load_duration": 6683708,
+  "prompt_eval_count": 17,
+  "prompt_eval_duration": 201222000,
+  "eval_count": 63,
+  "eval_duration": 953997000
+}
+```
+
 #### Request (JSON mode)
 
+> [!IMPORTANT]
 > When `format` is set to `json`, the output will always be a well-formed JSON object. It's important to also instruct the model to respond in JSON.
 
 ##### Request
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt": "What color is the sky at different times of the day? Respond using JSON",
   "format": "json",
   "stream": false
@@ -167,7 +205,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-11-09T21:07:55.186497Z",
   "response": "{\n\"morning\": {\n\"color\": \"blue\"\n},\n\"noon\": {\n\"color\": \"blue-gray\"\n},\n\"afternoon\": {\n\"color\": \"warm gray\"\n},\n\"evening\": {\n\"color\": \"orange\"\n}\n}\n",
   "done": true,
@@ -289,7 +327,7 @@ If you want to set custom options for the model at runtime rather than in the Mo
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt": "Why is the sky blue?",
   "stream": false,
   "options": {
@@ -298,6 +336,7 @@ curl http://localhost:11434/api/generate -d '{
     "num_predict": 100,
     "top_k": 20,
     "top_p": 0.9,
+    "min_p": 0.0,
     "tfs_z": 0.5,
     "typical_p": 0.7,
     "repeat_last_n": 33,
@@ -329,7 +368,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -351,7 +390,7 @@ If an empty prompt is provided, the model will be loaded into memory.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3"
+  "model": "llama3.1"
 }'
 ```
 
@@ -361,13 +400,40 @@ A single JSON object is returned:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-12-18T19:52:07.071755Z",
   "response": "",
   "done": true
 }
 ```
 
+#### Unload a model
+
+If an empty prompt is provided and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama3.1",
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.1",
+  "created_at": "2024-09-12T03:54:03.516566Z",
+  "response": "",
+  "done": true,
+  "done_reason": "unload"
+}
+```
+
 ## Generate a chat completion
 
 ```shell
@@ -380,12 +446,14 @@ Generate the next message in a chat with a provided model. This is a streaming e
 
 - `model`: (required) the [model name](#model-names)
 - `messages`: the messages of the chat, this can be used to keep a chat memory
+- `tools`: tools for the model to use if supported. Requires `stream` to be set to `false`
 
 The `message` object has the following fields:
 
-- `role`: the role of the message, either `system`, `user` or `assistant`
+- `role`: the role of the message, either `system`, `user`, `assistant`, or `tool`
 - `content`: the content of the message
 - `images` (optional): a list of images to include in the message (for multimodal models such as `llava`)
+- `tool_calls` (optional): a list of tools the model wants to use
 
 Advanced parameters (optional):
 
@@ -404,7 +472,7 @@ Send a chat message with a streaming response.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     {
       "role": "user",
@@ -420,7 +488,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -435,7 +503,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 4883583458,
@@ -453,7 +521,7 @@ Final response:
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     {
       "role": "user",
@@ -468,7 +536,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "model": "llama3.1",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -492,7 +560,7 @@ Send a chat message with a conversation history. You can use this same approach
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     {
       "role": "user",
@@ -516,7 +584,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -530,7 +598,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3",
+  "model": "llama3.1",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 8113331500,
@@ -546,7 +614,7 @@ Final response:
 
 ##### Request
 
-Send a chat message with a conversation history.
+Send a chat message with images. The images should be provided as an array, with the individual images encoded in Base64.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
@@ -588,7 +656,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "messages": [
     {
       "role": "user",
@@ -606,7 +674,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "model": "llama3.1",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -622,6 +690,137 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Chat request (with tools)
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.1",
+  "messages": [
+    {
+      "role": "user",
+      "content": "What is the weather today in Paris?"
+    }
+  ],
+  "stream": false,
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "get_current_weather",
+        "description": "Get the current weather for a location",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "location": {
+              "type": "string",
+              "description": "The location to get the weather for, e.g. San Francisco, CA"
+            },
+            "format": {
+              "type": "string",
+              "description": "The format to return the weather in, e.g. 'celsius' or 'fahrenheit'",
+              "enum": ["celsius", "fahrenheit"]
+            }
+          },
+          "required": ["location", "format"]
+        }
+      }
+    }
+  ]
+}'
+```
+
+##### Response
+
+```json
+{
+  "model": "llama3.1",
+  "created_at": "2024-07-22T20:33:28.123648Z",
+  "message": {
+    "role": "assistant",
+    "content": "",
+    "tool_calls": [
+      {
+        "function": {
+          "name": "get_current_weather",
+          "arguments": {
+            "format": "celsius",
+            "location": "Paris, FR"
+          }
+        }
+      }
+    ]
+  },
+  "done_reason": "stop",
+  "done": true,
+  "total_duration": 885095291,
+  "load_duration": 3753500,
+  "prompt_eval_count": 122,
+  "prompt_eval_duration": 328493000,
+  "eval_count": 33,
+  "eval_duration": 552222000
+}
+```
+
+#### Load a model
+
+If the messages array is empty, the model will be loaded into memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.1",
+  "messages": []
+}'
+```
+
+##### Response
+```json
+{
+  "model": "llama3.1",
+  "created_at":"2024-09-12T21:17:29.110811Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "load",
+  "done": true
+}
+```
+
+#### Unload a model
+
+If the messages array is empty and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.1",
+  "messages": [],
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.1",
+  "created_at":"2024-09-12T21:33:17.547535Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "unload",
+  "done": true
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -790,7 +989,7 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama3"
+  "name": "llama3.1"
 }'
 ```
 
@@ -851,7 +1050,7 @@ Copy a model. Creates a model with another name from an existing model.
 
 ```shell
 curl http://localhost:11434/api/copy -d '{
-  "source": "llama3",
+  "source": "llama3.1",
   "destination": "llama3-backup"
 }'
 ```
@@ -906,7 +1105,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where
 
 ```shell
 curl http://localhost:11434/api/pull -d '{
-  "name": "llama3"
+  "name": "llama3.1"
 }'
 ```
 
@@ -1026,7 +1225,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 ## Generate Embeddings
 
 ```shell
-POST /api/embeddings
+POST /api/embed
 ```
 
 Generate embeddings from a model
@@ -1034,10 +1233,11 @@ Generate embeddings from a model
 ### Parameters
 
 - `model`: name of model to generate embeddings from
-- `prompt`: text to generate embeddings for
+- `input`: text or list of text to generate embeddings for
 
 Advanced parameters:
 
+- `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1046,9 +1246,9 @@ Advanced parameters:
 #### Request
 
 ```shell
-curl http://localhost:11434/api/embeddings -d '{
+curl http://localhost:11434/api/embed -d '{
   "model": "all-minilm",
-  "prompt": "Here is an article about llamas..."
+  "input": "Why is the sky blue?"
 }'
 ```
 
@@ -1056,10 +1256,38 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embedding": [
-    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
-    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
-  ]
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ]],
+  "total_duration": 14143917,
+  "load_duration": 1019500,
+  "prompt_eval_count": 8
+}
+```
+
+#### Request (Multiple input)
+
+```shell
+curl http://localhost:11434/api/embed -d '{
+  "model": "all-minilm",
+  "input": ["Why is the sky blue?", "Why is the grass green?"]
+}'
+```
+
+#### Response
+
+```json
+{
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ],[
+    -0.0098027075, 0.06042469, 0.025257962, -0.006364387, 0.07272725,
+    0.017194884, 0.09032035, -0.051705178, 0.09951512, 0.09072481
+  ]]
 }
 ```
 
@@ -1106,3 +1334,45 @@ A single JSON object will be returned.
   ]
 }
 ```
+
+## Generate Embedding
+
+> Note: this endpoint has been superseded by `/api/embed`
+
+```shell
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/embeddings -d '{
+  "model": "all-minilm",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+#### Response
+
+```json
+{
+  "embedding": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```
diff --git a/docs/development.md b/docs/development.md
index 2a6886a43..e67689abc 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:
 
 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
 go build .
 ```
 
@@ -148,3 +148,22 @@ In addition to the common Windows development tools described above, install AMD
 - [Strawberry Perl](https://strawberryperl.com/)
 
 Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
+
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
+```
+
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
+
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
+```
+
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
\ No newline at end of file
diff --git a/docs/docker.md b/docs/docker.md
index 0b58562b7..314666b26 100644
--- a/docs/docker.md
+++ b/docs/docker.md
@@ -1,71 +1,71 @@
-# Ollama Docker image
-
-### CPU only
-
-```bash
-docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
-```
-
-### Nvidia GPU
-Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installation).
-
-#### Install with Apt
-1.  Configure the repository
-```bash
-curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
-    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
-curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
-    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
-    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
-sudo apt-get update
-```
-2.  Install the NVIDIA Container Toolkit packages
-```bash
-sudo apt-get install -y nvidia-container-toolkit
-```
-
-#### Install with Yum or Dnf
-1.  Configure the repository
-    
-```bash
-curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
-    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
-```
-    
-2. Install the NVIDIA Container Toolkit packages
-    
-```bash
-sudo yum install -y nvidia-container-toolkit
-```
-
-#### Configure Docker to use Nvidia driver 
-```
-sudo nvidia-ctk runtime configure --runtime=docker
-sudo systemctl restart docker
-```
-
-#### Start the container
-
-```bash
-docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
-```
-
-### AMD GPU
-
-To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
-
-```
-docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
-```
-
-### Run model locally
-
-Now you can run a model:
-
-```
-docker exec -it ollama ollama run llama3
-```
-
-### Try different models
-
-More models can be found on the [Ollama library](https://ollama.com/library).
+# Ollama Docker image
+
+### CPU only
+
+```bash
+docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### Nvidia GPU
+Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installation).
+
+#### Install with Apt
+1.  Configure the repository
+```bash
+curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
+    | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+    | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+    | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
+sudo apt-get update
+```
+2.  Install the NVIDIA Container Toolkit packages
+```bash
+sudo apt-get install -y nvidia-container-toolkit
+```
+
+#### Install with Yum or Dnf
+1.  Configure the repository
+
+```bash
+curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo \
+    | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
+```
+
+2. Install the NVIDIA Container Toolkit packages
+
+```bash
+sudo yum install -y nvidia-container-toolkit
+```
+
+#### Configure Docker to use Nvidia driver
+```
+sudo nvidia-ctk runtime configure --runtime=docker
+sudo systemctl restart docker
+```
+
+#### Start the container
+
+```bash
+docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
+```
+
+### AMD GPU
+
+To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
+
+```
+docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm
+```
+
+### Run model locally
+
+Now you can run a model:
+
+```
+docker exec -it ollama ollama run llama3.1
+```
+
+### Try different models
+
+More models can be found on the [Ollama library](https://ollama.com/library).
diff --git a/docs/faq.md b/docs/faq.md
index 841f1d13d..b2b1ca304 100644
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -32,7 +32,7 @@ When using the API, specify the `num_ctx` parameter:
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.1",
   "prompt": "Why is the sky blue?",
   "options": {
     "num_ctx": 4096
@@ -111,7 +111,10 @@ On Windows, Ollama inherits your user and system environment variables.
 
 ## How do I use Ollama behind a proxy?
 
-Ollama is compatible with proxy servers if `HTTP_PROXY` or `HTTPS_PROXY` are configured. When using either variables, ensure it is set where `ollama serve` can access the values. When using `HTTPS_PROXY`, ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+Ollama pulls models from the Internet and may require a proxy server to access the models. Use `HTTPS_PROXY` to redirect outbound requests through the proxy. Ensure the proxy certificate is installed as a system certificate. Refer to the section above for how to use environment variables on your platform.
+
+> [!NOTE]
+> Avoid setting `HTTP_PROXY`. Ollama does not use HTTP for model pulls, only HTTPS. Setting `HTTP_PROXY` may interrupt client connections to the server.
 
 ### How do I use Ollama behind a proxy in Docker?
 
@@ -191,6 +194,8 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e
 
 If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.
 
+> Note: on Linux using the standard installer, the `ollama` user needs read and write access to the specified directory. To assign the directory to the `ollama` user run `sudo chown -R ollama:ollama <directory>`.
+
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## How can I use Ollama in Visual Studio Code?
@@ -227,14 +232,18 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 
 To preload a model using the CLI, use the command:
 ```shell
-ollama run llama3 ""
+ollama run llama3.1 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
 
-By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you are making numerous requests to the LLM. You may, however, want to free up the memory before the 5 minutes have elapsed or keep the model loaded indefinitely. Use the `keep_alive` parameter with either the `/api/generate` and `/api/chat` API endpoints to control how long the model is left in memory.
+By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you're making numerous requests to the LLM. If you want to immediately unload a model from memory, use the `ollama stop` command:
 
-The `keep_alive` parameter can be set to:
+```shell
+ollama stop llama3.1
+```
+
+If you're using the API, use the `keep_alive` parameter with the `/api/generate` and `/api/chat` endpoints to set the amount of time that a model stays in memory. The `keep_alive` parameter can be set to:
 * a duration string (such as "10m" or "24h")
 * a number in seconds (such as 3600)
 * any negative number which will keep the model loaded in memory (e.g. -1 or "-1m")
@@ -242,17 +251,17 @@ The `keep_alive` parameter can be set to:
 
 For example, to preload a model and leave it in memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": -1}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": -1}'
 ```
 
 To unload the model and free up memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": 0}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": 0}'
 ```
 
-Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
+Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to the section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
 
-If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` API parameter with the `/api/generate` or `/api/chat` API endpoints.
+The `keep_alive` API parameter with the `/api/generate` and `/api/chat` API endpoints will override the `OLLAMA_KEEP_ALIVE` setting.
 
 ## How do I manage the maximum number of requests the Ollama server can queue?
 
@@ -266,8 +275,14 @@ If there is insufficient available memory to load a new model request while one
 
 Parallel request processing for a given model results in increasing the context size by the number of parallel requests.  For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
 
-The following server settings may be used to adjust how Ollama handles concurrent requests:
+The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
 
 - `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
 - `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
 - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
+
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
+
+## How does Ollama load models on multiple GPUs?
+
+Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
diff --git a/docs/gpu.md b/docs/gpu.md
index 80f276c3b..2913a2e27 100644
--- a/docs/gpu.md
+++ b/docs/gpu.md
@@ -10,7 +10,7 @@ Check your compute compatibility to see if your card is supported:
 | 9.0                | NVIDIA              | `H100`                                                                                                      |
 | 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
-| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
+| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050`   |
 |                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
 | 8.0                | NVIDIA              | `A100` `A30`                                                                                                |
 | 7.5                | GeForce GTX/RTX     | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060`                                    |
@@ -46,13 +46,24 @@ sudo modprobe nvidia_uvm`
 
 ## AMD Radeon
 Ollama supports the following AMD GPUs:
+
+### Linux Support
 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
 | AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
 | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
 | AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
 
-### Overrides
+### Windows Support
+With ROCm v6.1, the following GPUs are supported on Windows.
+
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` |
+
+
+### Overrides on Linux
 Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
 some cases you can force the system to try to use a similar LLVM target that is
 close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
@@ -63,7 +74,7 @@ would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
 server.  If you have an unsupported AMD GPU you can experiment using the list of
 supported types below.
 
-At this time, the known supported GPU types are the following LLVM Targets.
+At this time, the known supported GPU types on linux are the following LLVM Targets.
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|
diff --git a/docs/images/ollama-keys.png b/docs/images/ollama-keys.png
new file mode 100644
index 000000000..942079a86
Binary files /dev/null and b/docs/images/ollama-keys.png differ
diff --git a/docs/images/signup.png b/docs/images/signup.png
new file mode 100644
index 000000000..e80bb4e7e
Binary files /dev/null and b/docs/images/signup.png differ
diff --git a/docs/import.md b/docs/import.md
index f34f09ace..2346886f1 100644
--- a/docs/import.md
+++ b/docs/import.md
@@ -1,42 +1,129 @@
-# Import
+# Importing a model
 
-GGUF models and select Safetensors models can be imported directly into Ollama.
+## Table of Contents
 
-## Import GGUF
+  * [Importing a Safetensors adapter](#Importing-a-fine-tuned-adapter-from-Safetensors-weights)
+  * [Importing a Safetensors model](#Importing-a-model-from-Safetensors-weights)
+  * [Importing a GGUF file](#Importing-a-GGUF-based-model-or-adapter)
+  * [Sharing models on ollama.com](#Sharing-your-model-on-ollamacom)
 
-A binary GGUF file can be imported directly into Ollama through a Modelfile.
+## Importing a fine tuned adapter from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command pointing at the base model you used for fine tuning, and an `ADAPTER` command which points to the directory with your Safetensors adapter:
 
 ```dockerfile
-FROM /path/to/file.gguf
+FROM <base model name>
+ADAPTER /path/to/safetensors/adapter/directory
 ```
 
-## Import Safetensors
+Make sure that you use the same base model in the `FROM` command as you used to create the adapter otherwise you will get erratic results. Most frameworks use different quantization methods, so it's best to use non-quantized (i.e. non-QLoRA) adapters. If your adapter is in the same directory as your `Modelfile`, use `ADAPTER .` to specify the adapter path.
 
-If the model being imported is one of these architectures, it can be imported directly into Ollama through a Modelfile:
+Now run `ollama create` from the directory where the `Modelfile` was created:
 
- - LlamaForCausalLM
- - MistralForCausalLM
- - GemmaForCausalLM
+```bash
+ollama create my-model
+```
+
+Lastly, test the model:
+
+```bash
+ollama run my-model
+```
+
+Ollama supports importing adapters based on several different model architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral); and
+  * Gemma (including Gemma 1 and Gemma 2)
+
+You can create the adapter using a fine tuning framework or tool which can output adapters in the Safetensors format, such as:
+
+  * Hugging Face [fine tuning framework](https://huggingface.co/docs/transformers/en/training)
+  * [Unsloth](https://github.com/unslothai/unsloth)
+  * [MLX](https://github.com/ml-explore/mlx)
+
+
+## Importing a model from Safetensors weights
+
+First, create a `Modelfile` with a `FROM` command which points to the directory containing your Safetensors weights:
 
 ```dockerfile
 FROM /path/to/safetensors/directory
 ```
 
-For architectures not directly convertable by Ollama, see llama.cpp's [guide](https://github.com/ggerganov/llama.cpp/blob/master/README.md#prepare-and-quantize) on conversion. After conversion, see [Import GGUF](#import-gguf).
+If you create the Modelfile in the same directory as the weights, you can use the command `FROM .`.
 
-## Automatic Quantization
+Now run the `ollama create` command from the directory where you created the `Modelfile`:
 
-> [!NOTE]
-> Automatic quantization requires v0.1.35 or higher.
+```shell
+ollama create my-model
+```
 
-Ollama is capable of quantizing FP16 or FP32 models to any of the supported quantizations with the `-q/--quantize` flag in `ollama create`.
+Lastly, test the model:
+
+```shell
+ollama run my-model
+```
+
+Ollama supports importing models for several different architectures including:
+
+  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral);
+  * Gemma (including Gemma 1 and Gemma 2); and
+  * Phi3
+
+This includes importing foundation models as well as any fine tuned models which which have been _fused_ with a foundation model.
+
+
+## Importing a GGUF based model or adapter
+
+If you have a GGUF based model or adapter it is possible to import it into Ollama. You can obtain a GGUF model or adapter by:
+
+  * converting a Safetensors model with the `convert_hf_to_gguf.py` from Llama.cpp; 
+  * converting a Safetensors adapter with the `convert_lora_to_gguf.py` from Llama.cpp; or
+  * downloading a model or adapter from a place such as HuggingFace
+
+To import a GGUF model, create a `Modelfile` containg:
+
+```dockerfile
+FROM /path/to/file.gguf
+```
+
+For a GGUF adapter, create the `Modelfile` with:
+
+```dockerfile
+FROM <model name>
+ADAPTER /path/to/file.gguf
+```
+
+When importing a GGUF adapter, it's important to use the same base model as the base model that the adapter was created with. You can use:
+
+ * a model from Ollama
+ * a GGUF file
+ * a Safetensors based model 
+
+Once you have created your `Modelfile`, use the `ollama create` command to build the model.
+
+```shell
+ollama create my-model
+```
+
+## Quantizing a Model
+
+Quantizing a model allows you to run models faster and with less memory consumption but at reduced accuracy. This allows you to run a model on more modest hardware.
+
+Ollama can quantize FP16 and FP32 based models into different quantization levels using the `-q/--quantize` flag with the `ollama create` command.
+
+First, create a Modelfile with the FP16 or FP32 based model you wish to quantize.
 
 ```dockerfile
 FROM /path/to/my/gemma/f16/model
 ```
 
+Use `ollama create` to then create the quantized model.
+
 ```shell
-$ ollama create -q Q4_K_M mymodel
+$ ollama create --quantize q4_K_M mymodel
 transferring model data
 quantizing F16 model to Q4_K_M
 creating new layer sha256:735e246cc1abfd06e9cdcf95504d6789a6cd1ad7577108a70d9902fef503c1bd
@@ -47,42 +134,53 @@ success
 
 ### Supported Quantizations
 
-- `Q4_0`
-- `Q4_1`
-- `Q5_0`
-- `Q5_1`
-- `Q8_0`
+- `q4_0`
+- `q4_1`
+- `q5_0`
+- `q5_1`
+- `q8_0`
 
 #### K-means Quantizations
 
-- `Q3_K_S`
-- `Q3_K_M`
-- `Q3_K_L`
-- `Q4_K_S`
-- `Q4_K_M`
-- `Q5_K_S`
-- `Q5_K_M`
-- `Q6_K`
+- `q3_K_S`
+- `q3_K_M`
+- `q3_K_L`
+- `q4_K_S`
+- `q4_K_M`
+- `q5_K_S`
+- `q5_K_M`
+- `q6_K`
 
-## Template Detection
 
-> [!NOTE]
-> Template detection requires v0.1.42 or higher.
+## Sharing your model on ollama.com
 
-Ollama uses model metadata, specifically `tokenizer.chat_template`, to automatically create a template appropriate for the model you're importing.
+You can share any model you have created by pushing it to [ollama.com](https://ollama.com) so that other users can try it out.
 
-```dockerfile
-FROM /path/to/my/gemma/model
-```
+First, use your browser to go to the [Ollama Sign-Up](https://ollama.com/signup) page. If you already have an account, you can skip this step.
+
+<img src="images/signup.png" alt="Sign-Up" width="40%">
+
+The `Username` field will be used as part of your model's name (e.g. `jmorganca/mymodel`), so make sure you are comfortable with the username that you have selected.
+
+Now that you have created an account and are signed-in, go to the [Ollama Keys Settings](https://ollama.com/settings/keys) page.
+
+Follow the directions on the page to determine where your Ollama Public Key is located.
+
+<img src="images/ollama-keys.png" alt="Ollama Keys" width="80%">
+
+Click on the `Add Ollama Public Key` button, and copy and paste the contents of your Ollama Public Key into the text field.
+
+To push a model to [ollama.com](https://ollama.com), first make sure that it is named correctly with your username. You may have to use the `ollama cp` command to copy
+your model to give it the correct name. Once you're happy with your model's name, use the `ollama push` command to push it to [ollama.com](https://ollama.com).
 
 ```shell
-$ ollama create mymodel
-transferring model data
-using autodetected template gemma-instruct
-creating new layer sha256:baa2a0edc27d19cc6b7537578a9a7ba1a4e3214dc185ed5ae43692b319af7b84
-creating new layer sha256:ba66c3309914dbef07e5149a648fd1877f030d337a4f240d444ea335008943cb
-writing manifest
-success
+ollama cp mymodel myuser/mymodel
+ollama push myuser/mymodel
+```
+
+Once your model has been pushed, other users can pull and run it by using the command:
+
+```shell
+ollama run myuser/mymodel
 ```
 
-Defining a template in the Modelfile will disable this feature which may be useful if you want to use a different template than the autodetected one.
diff --git a/docs/linux.md b/docs/linux.md
index ec7306560..0eec014f4 100644
--- a/docs/linux.md
+++ b/docs/linux.md
@@ -1,40 +1,59 @@
-# Ollama on Linux
+# Linux
 
 ## Install
 
-Install Ollama running this one-liner:
+To install Ollama, run the following command:
 
->
-
-```bash
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
-## AMD Radeon GPU support
-
-While AMD has contributed the `amdgpu` driver upstream to the official linux
-kernel source, the version is older and may not support all ROCm features. We
-recommend you install the latest driver from
-https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-GPU.
-
 ## Manual install
 
-### Download the `ollama` binary
+Download and extract the package:
 
-Ollama is distributed as a self-contained binary. Download it to a directory in your PATH:
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
+```
 
-```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+Start Ollama:
+
+```shell
+ollama serve
+```
+
+In another terminal, verify that Ollama is running:
+
+```shell
+ollama -v
+```
+
+### AMD GPU install
+
+If you have an AMD GPU, also download and extract the additional ROCm package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64-rocm.tgz -o ollama-linux-amd64-rocm.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64-rocm.tgz
+```
+
+### ARM64 install
+
+Download and extract the ARM64-specific package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-arm64.tgz -o ollama-linux-arm64.tgz
+sudo tar -C /usr -xzf ollama-linux-arm64.tgz
 ```
 
 ### Adding Ollama as a startup service (recommended)
 
-Create a user for Ollama:
+Create a user and group for Ollama:
 
-```bash
-sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama
+```shell
+sudo useradd -r -s /bin/false -U -m -d /usr/share/ollama ollama
+sudo usermod -a -G ollama $(whoami)
 ```
 
 Create a service file in `/etc/systemd/system/ollama.service`:
@@ -50,6 +69,7 @@ User=ollama
 Group=ollama
 Restart=always
 RestartSec=3
+Environment="PATH=$PATH"
 
 [Install]
 WantedBy=default.target
@@ -57,47 +77,54 @@ WantedBy=default.target
 
 Then start the service:
 
-```bash
+```shell
 sudo systemctl daemon-reload
 sudo systemctl enable ollama
 ```
 
-### Install CUDA drivers (optional – for Nvidia GPUs)
+### Install CUDA drivers (optional)
 
 [Download and install](https://developer.nvidia.com/cuda-downloads) CUDA.
 
 Verify that the drivers are installed by running the following command, which should print details about your GPU:
 
-```bash
+```shell
 nvidia-smi
 ```
 
-### Install ROCm (optional - for Radeon GPUs)
-[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html)
+### Install AMD ROCm drivers (optional)
 
-Make sure to install ROCm v6
+[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html) ROCm v6.
 
 ### Start Ollama
 
-Start Ollama using `systemd`:
+Start Ollama and verify it is running:
 
-```bash
+```shell
 sudo systemctl start ollama
+sudo systemctl status ollama
 ```
 
-## Update
+> [!NOTE]
+> While AMD has contributed the `amdgpu` driver upstream to the official linux
+> kernel source, the version is older and may not support all ROCm features. We
+> recommend you install the latest driver from
+> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+> GPU.
 
-Update ollama by running the install script again:
+## Updating
 
-```bash
+Update Ollama by running the install script again:
+
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
-Or by downloading the ollama binary:
+Or by re-downloading Ollama:
 
-```bash
-sudo curl -L https://ollama.com/download/ollama-linux-amd64 -o /usr/bin/ollama
-sudo chmod +x /usr/bin/ollama
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```
 
 ## Installing specific versions
@@ -106,15 +133,15 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s
 
 For example:
 
-```
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh
+```shell
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
 ```
 
 ## Viewing logs
 
 To view logs of Ollama running as a startup service, run:
 
-```bash
+```shell
 journalctl -e -u ollama
 ```
 
@@ -122,7 +149,7 @@ journalctl -e -u ollama
 
 Remove the ollama service:
 
-```bash
+```shell
 sudo systemctl stop ollama
 sudo systemctl disable ollama
 sudo rm /etc/systemd/system/ollama.service
@@ -130,13 +157,13 @@ sudo rm /etc/systemd/system/ollama.service
 
 Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):
 
-```bash
+```shell
 sudo rm $(which ollama)
 ```
 
 Remove the downloaded models and Ollama service user and group:
 
-```bash
+```shell
 sudo rm -r /usr/share/ollama
 sudo userdel ollama
 sudo groupdel ollama
diff --git a/docs/modelfile.md b/docs/modelfile.md
index 21ee1826e..a33f180b7 100644
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@@ -1,6 +1,7 @@
 # Ollama Model File
 
-> Note: `Modelfile` syntax is in development
+> [!NOTE]
+> `Modelfile` syntax is in development
 
 A model file is the blueprint to create and share models with Ollama.
 
@@ -10,8 +11,9 @@ A model file is the blueprint to create and share models with Ollama.
 - [Examples](#examples)
 - [Instructions](#instructions)
   - [FROM (Required)](#from-required)
-    - [Build from llama3](#build-from-llama3)
-    - [Build from a bin file](#build-from-a-bin-file)
+    - [Build from existing model](#build-from-existing-model)
+    - [Build from a Safetensors model](#build-from-a-safetensors-model)
+    - [Build from a GGUF file](#build-from-a-gguf-file)
   - [PARAMETER](#parameter)
     - [Valid Parameters and Values](#valid-parameters-and-values)
   - [TEMPLATE](#template)
@@ -48,7 +50,7 @@ INSTRUCTION arguments
 An example of a `Modelfile` creating a mario blueprint:
 
 ```modelfile
-FROM llama3
+FROM llama3.1
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
 # sets the context window size to 4096, this controls how many tokens the LLM can use as context to generate the next token
@@ -70,10 +72,10 @@ More examples are available in the [examples directory](../examples).
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.
 
   ```bash
-  > ollama show --modelfile llama3
+  > ollama show --modelfile llama3.1
   # Modelfile generated by "ollama show"
   # To build a new Modelfile based on this one, replace the FROM line with:
-  # FROM llama3:latest
+  # FROM llama3.1:latest
   FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
   TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
 
@@ -98,22 +100,39 @@ The `FROM` instruction defines the base model to use when creating a model.
 FROM <model name>:<tag>
 ```
 
-#### Build from llama3
+#### Build from existing model
 
 ```modelfile
-FROM llama3
+FROM llama3.1
 ```
 
 A list of available base models:
 <https://github.com/ollama/ollama#model-library>
+Additional models can be found at:
+<https://ollama.com/library>
 
-#### Build from a `bin` file
+#### Build from a Safetensors model
 
 ```modelfile
-FROM ./ollama-model.bin
+FROM <model directory>
 ```
 
-This bin file location should be specified as an absolute path or relative to the `Modelfile` location.
+The model directory should contain the Safetensors weights for a supported architecture.
+
+Currently supported model architectures:
+  * Llama (including Llama 2, Llama 3, and Llama 3.1)
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral)
+  * Gemma (including Gemma 1 and Gemma 2)
+  * Phi3
+
+#### Build from a GGUF file
+
+```modelfile
+FROM ./ollama-model.gguf
+```
+
+The GGUF file location should be specified as an absolute path or relative to the `Modelfile` location.
+
 
 ### PARAMETER
 
@@ -140,6 +159,7 @@ PARAMETER <parameter> <parametervalue>
 | num_predict    | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
+| min_p          | Alternative to the top_p, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. (Default: 0.0) | float      | min_p 0.05            |
 
 ### TEMPLATE
 
@@ -172,10 +192,23 @@ SYSTEM """<system message>"""
 
 ### ADAPTER
 
-The `ADAPTER` instruction is an optional instruction that specifies any LoRA adapter that should apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+The `ADAPTER` instruction specifies a fine tuned LoRA adapter that should apply to the base model. The value of the adapter should be an absolute path or a path relative to the Modelfile. The base model should be specified with a `FROM` instruction. If the base model is not the same as the base model that the adapter was tuned from the behaviour will be erratic.
+
+#### Safetensor adapter
 
 ```modelfile
-ADAPTER ./ollama-lora.bin
+ADAPTER <path to safetensor adapter>
+```
+
+Currently supported Safetensor adapters:
+  * Llama (including Llama 2, Llama 3, and Llama 3.1)
+  * Mistral (including Mistral 1, Mistral 2, and Mixtral)
+  * Gemma (including Gemma 1 and Gemma 2)
+
+#### GGUF adapter
+
+```modelfile
+ADAPTER ./ollama-lora.gguf
 ```
 
 ### LICENSE
diff --git a/docs/openai.md b/docs/openai.md
index 81b967eb7..c6df0fecb 100644
--- a/docs/openai.md
+++ b/docs/openai.md
@@ -25,7 +25,38 @@ chat_completion = client.chat.completions.create(
             'content': 'Say this is a test',
         }
     ],
-    model='llama3',
+    model='llama3.1',
+)
+
+response = client.chat.completions.create(
+    model="llava",
+    messages=[
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What's in this image?"},
+                {
+                    "type": "image_url",
+                    "image_url": "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",
+                },
+            ],
+        }
+    ],
+    max_tokens=300,
+)
+
+completion = client.completions.create(
+    model="llama3.1",
+    prompt="Say this is a test",
+)
+
+list_completion = client.models.list()
+
+model = client.models.retrieve("llama3.1")
+
+embeddings = client.embeddings.create(
+    model="all-minilm",
+    input=["why is the sky blue?", "why is the grass green?"],
 )
 ```
 
@@ -42,18 +73,48 @@ const openai = new OpenAI({
 })
 
 const chatCompletion = await openai.chat.completions.create({
-  messages: [{ role: 'user', content: 'Say this is a test' }],
-  model: 'llama3',
+    messages: [{ role: 'user', content: 'Say this is a test' }],
+    model: 'llama3.1',
+})
+
+const response = await openai.chat.completions.create({
+    model: "llava",
+    messages: [
+        {
+        role: "user",
+        content: [
+            { type: "text", text: "What's in this image?" },
+            {
+            type: "image_url",
+            image_url: "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",
+            },
+        ],
+        },
+    ],
+})
+
+const completion = await openai.completions.create({
+    model: "llama3.1",
+    prompt: "Say this is a test.",
+})
+
+const listCompletion = await openai.models.list()
+
+const model = await openai.models.retrieve("llama3.1")
+
+const embedding = await openai.embeddings.create({
+  model: "all-minilm",
+  input: ["why is the sky blue?", "why is the grass green?"],
 })
 ```
 
 ### `curl`
 
-```
+``` shell
 curl http://localhost:11434/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
-        "model": "llama3",
+        "model": "llama3.1",
         "messages": [
             {
                 "role": "system",
@@ -65,6 +126,48 @@ curl http://localhost:11434/v1/chat/completions \
             }
         ]
     }'
+
+curl http://localhost:11434/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "llava",
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {
+            "type": "text",
+            "text": "What'\''s in this image?"
+          },
+          {
+            "type": "image_url",
+            "image_url": {
+               "url": "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"
+            }
+          }
+        ]
+      }
+    ],
+    "max_tokens": 300
+  }'
+
+curl http://localhost:11434/v1/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "llama3.1",
+        "prompt": "Say this is a test"
+    }'
+
+curl http://localhost:11434/v1/models
+
+curl http://localhost:11434/v1/models/llama3.1
+
+curl http://localhost:11434/v1/embeddings \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "all-minilm",
+        "input": ["why is the sky blue?", "why is the grass green?"]
+    }'
 ```
 
 ## Endpoints
@@ -77,8 +180,8 @@ curl http://localhost:11434/v1/chat/completions \
 - [x] Streaming
 - [x] JSON mode
 - [x] Reproducible outputs
-- [ ] Vision
-- [ ] Function calling
+- [x] Vision
+- [x] Tools (streaming support coming soon)
 - [ ] Logprobs
 
 #### Supported request fields
@@ -86,7 +189,10 @@ curl http://localhost:11434/v1/chat/completions \
 - [x] `model`
 - [x] `messages`
   - [x] Text `content`
-  - [ ] Array of `content` parts
+  - [x] Image `content`
+    - [x] Base64 encoded image
+    - [ ] Image URL
+  - [x] Array of `content` parts
 - [x] `frequency_penalty`
 - [x] `presence_penalty`
 - [x] `response_format`
@@ -96,22 +202,79 @@ curl http://localhost:11434/v1/chat/completions \
 - [x] `temperature`
 - [x] `top_p`
 - [x] `max_tokens`
-- [ ] `logit_bias`
-- [ ] `tools`
+- [x] `tools`
 - [ ] `tool_choice`
+- [ ] `logit_bias`
+- [ ] `user`
+- [ ] `n`
+
+### `/v1/completions`
+
+#### Supported features
+
+- [x] Completions
+- [x] Streaming
+- [x] JSON mode
+- [x] Reproducible outputs
+- [ ] Logprobs
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `prompt`
+- [x] `frequency_penalty`
+- [x] `presence_penalty`
+- [x] `seed`
+- [x] `stop`
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `max_tokens`
+- [x] `suffix`
+- [ ] `best_of`
+- [ ] `echo`
+- [ ] `logit_bias`
 - [ ] `user`
 - [ ] `n`
 
 #### Notes
 
-- `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached
+- `prompt` currently only accepts a string
+
+### `/v1/models`
+
+#### Notes
+
+- `created` corresponds to when the model was last modified
+- `owned_by` corresponds to the ollama username, defaulting to `"library"`
+
+### `/v1/models/{model}`
+
+#### Notes
+
+- `created` corresponds to when the model was last modified
+- `owned_by` corresponds to the ollama username, defaulting to `"library"`
+
+### `/v1/embeddings`
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `input`
+  - [x] string
+  - [x] array of strings
+  - [ ] array of tokens
+  - [ ] array of token arrays
+- [ ] `encoding format`
+- [ ] `dimensions`
+- [ ] `user`
 
 ## Models
 
 Before using a model, pull it locally `ollama pull`:
 
 ```shell
-ollama pull llama3
+ollama pull llama3.1
 ```
 
 ### Default model names
@@ -119,7 +282,7 @@ ollama pull llama3
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:
 
 ```
-ollama cp llama3 gpt-3.5-turbo
+ollama cp llama3.1 gpt-3.5-turbo
 ```
 
 Afterwards, this new model name can be specified the `model` field:
@@ -137,3 +300,28 @@ curl http://localhost:11434/v1/chat/completions \
         ]
     }'
 ```
+
+### Setting the context size
+
+The OpenAI API does not have a way of setting the context size for a model. If you need to change the context size, create a `Modelfile` which looks like:
+
+```modelfile
+FROM <some model>
+PARAMETER num_ctx <context size>
+```
+
+Use the `ollama create mymodel` command to create a new model with the updated context size. Call the API with the updated model name:
+
+```shell
+curl http://localhost:11434/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "mymodel",
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
diff --git a/docs/template.md b/docs/template.md
new file mode 100644
index 000000000..192d878d4
--- /dev/null
+++ b/docs/template.md
@@ -0,0 +1,167 @@
+# Template
+
+Ollama provides a powerful templating engine backed by Go's built-in templating engine to construct prompts for your large language model. This feature is a valuable tool to get the most out of your models.
+
+## Basic Template Structure
+
+A basic Go template consists of three main parts:
+
+* **Layout**: The overall structure of the template.
+* **Variables**: Placeholders for dynamic data that will be replaced with actual values when the template is rendered.
+* **Functions**: Custom functions or logic that can be used to manipulate the template's content.
+
+Here's an example of a simple chat template:
+
+```gotmpl
+{{- range .Messages }}
+{{ .Role }}: {{ .Content }}
+{{- end }}
+```
+
+In this example, we have:
+
+* A basic messages structure (layout)
+* Three variables: `Messages`, `Role`, and `Content` (variables)
+* A custom function (action) that iterates over an array of items (`range .Messages`) and displays each item
+
+## Adding templates to your model
+
+By default, models imported into Ollama have a default template of `{{ .Prompt }}`, i.e. user inputs are sent verbatim to the LLM. This is appropriate for text or code completion models but lacks essential markers for chat or instruction models.
+
+Omitting a template in these models puts the responsibility of correctly templating input onto the user. Adding a template allows users to easily get the best results from the model.
+
+To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3.
+
+```dockerfile
+FROM llama3.1
+
+TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>
+{{- end }}
+{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ .Content }}<|eot_id|>
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
+
+"""
+```
+
+## Variables
+
+`System` (string): system prompt
+
+`Prompt` (string): user prompt
+
+`Response` (string): assistant response
+
+`Suffix` (string): text inserted after the assistant's response
+
+`Messages` (list): list of messages
+
+`Messages[].Role` (string): role which can be one of `system`, `user`, `assistant`, or `tool`
+
+`Messages[].Content` (string):  message content
+
+`Messages[].ToolCalls` (list): list of tools the model wants to call
+
+`Messages[].ToolCalls[].Function` (object): function to call
+
+`Messages[].ToolCalls[].Function.Name` (string): function name
+
+`Messages[].ToolCalls[].Function.Arguments` (map): mapping of argument name to argument value
+
+`Tools` (list): list of tools the model can access
+
+`Tools[].Type` (string): schema type. `type` is always `function`
+
+`Tools[].Function` (object): function definition
+
+`Tools[].Function.Name` (string): function name
+
+`Tools[].Function.Description` (string): function description
+
+`Tools[].Function.Parameters` (object): function parameters
+
+`Tools[].Function.Parameters.Type` (string): schema type. `type` is always `object`
+
+`Tools[].Function.Parameters.Required` (list): list of required properties
+
+`Tools[].Function.Parameters.Properties` (map): mapping of property name to property definition
+
+`Tools[].Function.Parameters.Properties[].Type` (string): property type
+
+`Tools[].Function.Parameters.Properties[].Description` (string): property description
+
+`Tools[].Function.Parameters.Properties[].Enum` (list): list of valid values
+
+## Tips and Best Practices
+
+Keep the following tips and best practices in mind when working with Go templates:
+
+* **Be mindful of dot**: Control flow structures like `range` and `with` changes the value `.`
+* **Out-of-scope variables**: Use `$.` to reference variables not currently in scope, starting from the root
+* **Whitespace control**: Use `-` to trim leading (`{{-`) and trailing (`-}}`) whitespace
+
+## Examples
+
+### Example Messages
+
+#### ChatML
+
+ChatML is a popular template format. It can be used for models such as Databrick's DBRX, Intel's Neural Chat, and Microsoft's Orca 2.
+
+```gotmpl
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+```
+
+### Example Tools
+
+Tools support can be added to a model by adding a `{{ .Tools }}` node to the template. This feature is useful for models trained to call external tools and can a powerful tool for retrieving real-time data or performing complex tasks.
+
+#### Mistral
+
+Mistral v0.3 and Mixtral 8x22B supports tool calling.
+
+```gotmpl
+{{- range $index, $_ := .Messages }}
+{{- if eq .Role "user" }}
+{{- if and (le (len (slice $.Messages $index)) 2) $.Tools }}[AVAILABLE_TOOLS] {{ json $.Tools }}[/AVAILABLE_TOOLS]
+{{- end }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}
+
+{{ end }}{{ .Content }}[/INST]
+{{- else if eq .Role "assistant" }}
+{{- if .Content }} {{ .Content }}</s>
+{{- else if .ToolCalls }}[TOOL_CALLS] [
+{{- range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ json .Function.Arguments }}}
+{{- end }}]</s>
+{{- end }}
+{{- else if eq .Role "tool" }}[TOOL_RESULTS] {"content": {{ .Content }}}[/TOOL_RESULTS]
+{{- end }}
+{{- end }}
+```
+
+### Example Fill-in-Middle
+
+Fill-in-middle support can be added to a model by adding a `{{ .Suffix }}` node to the template. This feature is useful for models that are trained to generate text in the middle of user input, such as code completion models.
+
+#### CodeLlama
+
+CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://ollama.com/library/codellama:13b-code) code completion models support fill-in-middle.
+
+```gotmpl
+<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+```
+
+> [!NOTE]
+> CodeLlama 34B and 70B code completion and all instruct and Python fine-tuned models do not support fill-in-middle.
+
+#### Codestral
+
+Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
+
+```gotmpl
+[SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
+```
diff --git a/docs/troubleshooting.md b/docs/troubleshooting.md
index de29b344c..0a89b87f9 100644
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama
+journalctl -u ollama --no-pager
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
@@ -70,14 +70,18 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 
 If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
 
-## Container fails to run on NVIDIA GPU
+## NVIDIA GPU Discovery
 
-Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.
 
-Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+### Linux NVIDIA Troubleshooting
 
-- Is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
-- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+
+Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+
+- If you are using a container, is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
+- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
 - Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
 - Try rebooting
 - Make sure you're running the latest nvidia drivers
@@ -85,3 +89,19 @@ Sometimes the container runtime can have difficulties initializing the GPU. When
 If none of those resolve the problem, gather additional information and file an issue:
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
+
+
+## AMD GPU Discovery
+
+On linux, AMD GPU access typically requires `video` and/or `render` group membership to access the `/dev/kfd` device.  If permissions are not set up correctly, Ollama will detect this and report an error in the server log.
+
+When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -ld /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the group assignments on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.
+
+If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
+- `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
+- `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported
+- Check dmesg for any errors from amdgpu or kfd drivers `sudo dmesg | grep -i amdgpu` and `sudo dmesg | grep -i kfd`
+
+## Windows Terminal Errors
+
+Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
diff --git a/docs/tutorials/langchainjs.md b/docs/tutorials/langchainjs.md
index 4d60afb64..f925869b5 100644
--- a/docs/tutorials/langchainjs.md
+++ b/docs/tutorials/langchainjs.md
@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",
-  model: "llama3",
+  model: "llama3.1",
 });
 
 const answer = await ollama.invoke(`why is the sky blue?`);
@@ -23,7 +23,7 @@ const answer = await ollama.invoke(`why is the sky blue?`);
 console.log(answer);
 ```
 
-That will get us the same thing as if we ran `ollama run llama3 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
+That will get us the same thing as if we ran `ollama run llama3.1 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
 
 ```bash
 npm install cheerio
diff --git a/docs/windows.md b/docs/windows.md
index abc0eb300..372a35aa8 100644
--- a/docs/windows.md
+++ b/docs/windows.md
@@ -19,15 +19,17 @@ Logs will often be helpful in diagnosing the problem (see
 
 ## System Requirements
 
-* Windows 10 or newer, Home or Pro
+* Windows 10 22H2 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
 
+Ollama uses unicode characters for progress indication, which may render as unknown squares in some older terminal fonts in Windows 10. If you see this, try changing your terminal font settings.
+
 ## API Access
 
 Here's a quick example showing API access from `powershell`
 ```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama3", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+(Invoke-WebRequest -method POST -Body '{"model":"llama3.1", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```
 
 ## Troubleshooting
@@ -46,6 +48,9 @@ the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
 - `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
+## Uninstall
+
+The Ollama Windows installer registers an Uninstaller application.  Under `Add or remove programs` in Windows Settings, you can uninstall Ollama.
 
 ## Standalone CLI
 
diff --git a/envconfig/config.go b/envconfig/config.go
index c02c4878e..9c1490a93 100644
--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -1,77 +1,226 @@
 package envconfig
 
 import (
-	"errors"
 	"fmt"
 	"log/slog"
+	"math"
 	"net"
+	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
+	"time"
 )
 
-type OllamaHost struct {
-	Scheme string
-	Host   string
-	Port   string
+// Host returns the scheme and host. Host can be configured via the OLLAMA_HOST environment variable.
+// Default is scheme "http" and host "127.0.0.1:11434"
+func Host() *url.URL {
+	defaultPort := "11434"
+
+	s := strings.TrimSpace(Var("OLLAMA_HOST"))
+	scheme, hostport, ok := strings.Cut(s, "://")
+	switch {
+	case !ok:
+		scheme, hostport = "http", s
+	case scheme == "http":
+		defaultPort = "80"
+	case scheme == "https":
+		defaultPort = "443"
+	}
+
+	hostport, path, _ := strings.Cut(hostport, "/")
+	host, port, err := net.SplitHostPort(hostport)
+	if err != nil {
+		host, port = "127.0.0.1", defaultPort
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
+	}
+
+	if n, err := strconv.ParseInt(port, 10, 32); err != nil || n > 65535 || n < 0 {
+		slog.Warn("invalid port, using default", "port", port, "default", defaultPort)
+		port = defaultPort
+	}
+
+	return &url.URL{
+		Scheme: scheme,
+		Host:   net.JoinHostPort(host, port),
+		Path:   path,
+	}
 }
 
-func (o OllamaHost) String() string {
-	return fmt.Sprintf("%s://%s:%s", o.Scheme, o.Host, o.Port)
+// Origins returns a list of allowed origins. Origins can be configured via the OLLAMA_ORIGINS environment variable.
+func Origins() (origins []string) {
+	if s := Var("OLLAMA_ORIGINS"); s != "" {
+		origins = strings.Split(s, ",")
+	}
+
+	for _, origin := range []string{"localhost", "127.0.0.1", "0.0.0.0"} {
+		origins = append(origins,
+			fmt.Sprintf("http://%s", origin),
+			fmt.Sprintf("https://%s", origin),
+			fmt.Sprintf("http://%s", net.JoinHostPort(origin, "*")),
+			fmt.Sprintf("https://%s", net.JoinHostPort(origin, "*")),
+		)
+	}
+
+	origins = append(origins,
+		"app://*",
+		"file://*",
+		"tauri://*",
+	)
+
+	return origins
 }
 
-var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST")
+// Models returns the path to the models directory. Models directory can be configured via the OLLAMA_MODELS environment variable.
+// Default is $HOME/.ollama/models
+func Models() string {
+	if s := Var("OLLAMA_MODELS"); s != "" {
+		return s
+	}
+
+	home, err := os.UserHomeDir()
+	if err != nil {
+		panic(err)
+	}
+
+	return filepath.Join(home, ".ollama", "models")
+}
+
+// KeepAlive returns the duration that models stay loaded in memory. KeepAlive can be configured via the OLLAMA_KEEP_ALIVE environment variable.
+// Negative values are treated as infinite. Zero is treated as no keep alive.
+// Default is 5 minutes.
+func KeepAlive() (keepAlive time.Duration) {
+	keepAlive = 5 * time.Minute
+	if s := Var("OLLAMA_KEEP_ALIVE"); s != "" {
+		if d, err := time.ParseDuration(s); err == nil {
+			keepAlive = d
+		} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
+			keepAlive = time.Duration(n) * time.Second
+		}
+	}
+
+	if keepAlive < 0 {
+		return time.Duration(math.MaxInt64)
+	}
+
+	return keepAlive
+}
+
+// LoadTimeout returns the duration for stall detection during model loads. LoadTimeout can be configured via the OLLAMA_LOAD_TIMEOUT environment variable.
+// Zero or Negative values are treated as infinite.
+// Default is 5 minutes.
+func LoadTimeout() (loadTimeout time.Duration) {
+	loadTimeout = 5 * time.Minute
+	if s := Var("OLLAMA_LOAD_TIMEOUT"); s != "" {
+		if d, err := time.ParseDuration(s); err == nil {
+			loadTimeout = d
+		} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
+			loadTimeout = time.Duration(n) * time.Second
+		}
+	}
+
+	if loadTimeout <= 0 {
+		return time.Duration(math.MaxInt64)
+	}
+
+	return loadTimeout
+}
+
+func Bool(k string) func() bool {
+	return func() bool {
+		if s := Var(k); s != "" {
+			b, err := strconv.ParseBool(s)
+			if err != nil {
+				return true
+			}
+
+			return b
+		}
+
+		return false
+	}
+}
 
 var (
-	// Set via OLLAMA_ORIGINS in the environment
-	AllowOrigins []string
-	// Set via OLLAMA_DEBUG in the environment
-	Debug bool
-	// Experimental flash attention
-	FlashAttention bool
-	// Set via OLLAMA_HOST in the environment
-	Host *OllamaHost
-	// Set via OLLAMA_KEEP_ALIVE in the environment
-	KeepAlive string
-	// Set via OLLAMA_LLM_LIBRARY in the environment
-	LLMLibrary string
-	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
-	MaxRunners int
-	// Set via OLLAMA_MAX_QUEUE in the environment
-	MaxQueuedRequests int
-	// Set via OLLAMA_MODELS in the environment
-	ModelsDir string
-	// Set via OLLAMA_MAX_VRAM in the environment
-	MaxVRAM uint64
-	// Set via OLLAMA_NOHISTORY in the environment
-	NoHistory bool
-	// Set via OLLAMA_NOPRUNE in the environment
-	NoPrune bool
-	// Set via OLLAMA_NUM_PARALLEL in the environment
-	NumParallel int
-	// Set via OLLAMA_RUNNERS_DIR in the environment
-	RunnersDir string
-	// Set via OLLAMA_SCHED_SPREAD in the environment
-	SchedSpread bool
-	// Set via OLLAMA_TMPDIR in the environment
-	TmpDir string
-	// Set via OLLAMA_INTEL_GPU in the environment
-	IntelGpu bool
-
-	// Set via CUDA_VISIBLE_DEVICES in the environment
-	CudaVisibleDevices string
-	// Set via HIP_VISIBLE_DEVICES in the environment
-	HipVisibleDevices string
-	// Set via ROCR_VISIBLE_DEVICES in the environment
-	RocrVisibleDevices string
-	// Set via GPU_DEVICE_ORDINAL in the environment
-	GpuDeviceOrdinal string
-	// Set via HSA_OVERRIDE_GFX_VERSION in the environment
-	HsaOverrideGfxVersion string
+	// Debug enabled additional debug information.
+	Debug = Bool("OLLAMA_DEBUG")
+	// FlashAttention enables the experimental flash attention feature.
+	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
+	// NoHistory disables readline history.
+	NoHistory = Bool("OLLAMA_NOHISTORY")
+	// NoPrune disables pruning of model blobs on startup.
+	NoPrune = Bool("OLLAMA_NOPRUNE")
+	// SchedSpread allows scheduling models across all GPUs.
+	SchedSpread = Bool("OLLAMA_SCHED_SPREAD")
+	// IntelGPU enables experimental Intel GPU detection.
+	IntelGPU = Bool("OLLAMA_INTEL_GPU")
 )
 
+func String(s string) func() string {
+	return func() string {
+		return Var(s)
+	}
+}
+
+var (
+	LLMLibrary = String("OLLAMA_LLM_LIBRARY")
+	TmpDir     = String("OLLAMA_TMPDIR")
+
+	CudaVisibleDevices    = String("CUDA_VISIBLE_DEVICES")
+	HipVisibleDevices     = String("HIP_VISIBLE_DEVICES")
+	RocrVisibleDevices    = String("ROCR_VISIBLE_DEVICES")
+	GpuDeviceOrdinal      = String("GPU_DEVICE_ORDINAL")
+	HsaOverrideGfxVersion = String("HSA_OVERRIDE_GFX_VERSION")
+)
+
+func Uint(key string, defaultValue uint) func() uint {
+	return func() uint {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseUint(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return uint(n)
+			}
+		}
+
+		return defaultValue
+	}
+}
+
+var (
+	// NumParallel sets the number of parallel model requests. NumParallel can be configured via the OLLAMA_NUM_PARALLEL environment variable.
+	NumParallel = Uint("OLLAMA_NUM_PARALLEL", 0)
+	// MaxRunners sets the maximum number of loaded models. MaxRunners can be configured via the OLLAMA_MAX_LOADED_MODELS environment variable.
+	MaxRunners = Uint("OLLAMA_MAX_LOADED_MODELS", 0)
+	// MaxQueue sets the maximum number of queued requests. MaxQueue can be configured via the OLLAMA_MAX_QUEUE environment variable.
+	MaxQueue = Uint("OLLAMA_MAX_QUEUE", 512)
+	// MaxVRAM sets a maximum VRAM override in bytes. MaxVRAM can be configured via the OLLAMA_MAX_VRAM environment variable.
+	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
+)
+
+func Uint64(key string, defaultValue uint64) func() uint64 {
+	return func() uint64 {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseUint(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return n
+			}
+		}
+
+		return defaultValue
+	}
+}
+
+// Set aside VRAM per GPU
+var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
+
 type EnvVar struct {
 	Name        string
 	Value       any
@@ -80,31 +229,45 @@ type EnvVar struct {
 
 func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
-		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug, "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
-		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention, "Enabled flash attention"},
-		"OLLAMA_HOST":              {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"},
-		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
-		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
-		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
-		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
-		"OLLAMA_MODELS":            {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"},
-		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
-		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"},
-		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
-		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
-		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"},
-		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir, "Location for temporary files"},
+		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
+		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
+		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
+		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
+		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
+		"OLLAMA_LOAD_TIMEOUT":      {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
+		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
+		"OLLAMA_MODELS":            {"OLLAMA_MODELS", Models(), "The path to the models directory"},
+		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
+		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
+		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
+		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
+		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
+
+		// Informational
+		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
+		"HTTPS_PROXY": {"HTTPS_PROXY", String("HTTPS_PROXY")(), "HTTPS proxy"},
+		"NO_PROXY":    {"NO_PROXY", String("NO_PROXY")(), "No proxy"},
 	}
+
+	if runtime.GOOS != "windows" {
+		// Windows environment variables are case-insensitive so there's no need to duplicate them
+		ret["http_proxy"] = EnvVar{"http_proxy", String("http_proxy")(), "HTTP proxy"}
+		ret["https_proxy"] = EnvVar{"https_proxy", String("https_proxy")(), "HTTPS proxy"}
+		ret["no_proxy"] = EnvVar{"no_proxy", String("no_proxy")(), "No proxy"}
+	}
+
 	if runtime.GOOS != "darwin" {
-		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices, "Set which NVIDIA devices are visible"}
-		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices, "Set which AMD devices are visible"}
-		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices, "Set which AMD devices are visible"}
-		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal, "Set which AMD devices are visible"}
-		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion, "Override the gfx used for all detected AMD GPUs"}
-		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGpu, "Enable experimental Intel GPU detection"}
+		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices(), "Set which NVIDIA devices are visible"}
+		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible"}
+		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices(), "Set which AMD devices are visible"}
+		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal(), "Set which AMD devices are visible"}
+		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion(), "Override the gfx used for all detected AMD GPUs"}
+		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGPU(), "Enable experimental Intel GPU detection"}
 	}
+
 	return ret
 }
 
@@ -116,231 +279,16 @@ func Values() map[string]string {
 	return vals
 }
 
-var defaultAllowOrigins = []string{
-	"localhost",
-	"127.0.0.1",
-	"0.0.0.0",
+// Var returns an environment variable stripped of leading and trailing quotes or spaces
+func Var(key string) string {
+	return strings.Trim(strings.TrimSpace(os.Getenv(key)), "\"'")
 }
 
-// Clean quotes and spaces from the value
-func clean(key string) string {
-	return strings.Trim(os.Getenv(key), "\"' ")
-}
-
-func init() {
-	// default values
-	NumParallel = 0 // Autoselect
-	MaxRunners = 0  // Autoselect
-	MaxQueuedRequests = 512
-
-	LoadConfig()
-}
-
-func LoadConfig() {
-	if debug := clean("OLLAMA_DEBUG"); debug != "" {
-		d, err := strconv.ParseBool(debug)
-		if err == nil {
-			Debug = d
-		} else {
-			Debug = true
-		}
-	}
-
-	if fa := clean("OLLAMA_FLASH_ATTENTION"); fa != "" {
-		d, err := strconv.ParseBool(fa)
-		if err == nil {
-			FlashAttention = d
-		}
-	}
-
-	RunnersDir = clean("OLLAMA_RUNNERS_DIR")
-	if runtime.GOOS == "windows" && RunnersDir == "" {
-		// On Windows we do not carry the payloads inside the main executable
-		appExe, err := os.Executable()
-		if err != nil {
-			slog.Error("failed to lookup executable path", "error", err)
-		}
-
-		cwd, err := os.Getwd()
-		if err != nil {
-			slog.Error("failed to lookup working directory", "error", err)
-		}
-
-		var paths []string
-		for _, root := range []string{filepath.Dir(appExe), cwd} {
-			paths = append(paths,
-				root,
-				filepath.Join(root, "windows-"+runtime.GOARCH),
-				filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
-			)
-		}
-
-		// Try a few variations to improve developer experience when building from source in the local tree
-		for _, p := range paths {
-			candidate := filepath.Join(p, "ollama_runners")
-			_, err := os.Stat(candidate)
-			if err == nil {
-				RunnersDir = candidate
-				break
-			}
-		}
-		if RunnersDir == "" {
-			slog.Error("unable to locate llm runner directory.  Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
-		}
-	}
-
-	TmpDir = clean("OLLAMA_TMPDIR")
-
-	userLimit := clean("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			MaxVRAM = avail
-		}
-	}
-
-	LLMLibrary = clean("OLLAMA_LLM_LIBRARY")
-
-	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
-		val, err := strconv.Atoi(onp)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err)
-		} else {
-			NumParallel = val
-		}
-	}
-
-	if nohistory := clean("OLLAMA_NOHISTORY"); nohistory != "" {
-		NoHistory = true
-	}
-
-	if spread := clean("OLLAMA_SCHED_SPREAD"); spread != "" {
-		s, err := strconv.ParseBool(spread)
-		if err == nil {
-			SchedSpread = s
-		} else {
-			SchedSpread = true
-		}
-	}
-
-	if noprune := clean("OLLAMA_NOPRUNE"); noprune != "" {
-		NoPrune = true
-	}
-
-	if origins := clean("OLLAMA_ORIGINS"); origins != "" {
-		AllowOrigins = strings.Split(origins, ",")
-	}
-	for _, allowOrigin := range defaultAllowOrigins {
-		AllowOrigins = append(AllowOrigins,
-			fmt.Sprintf("http://%s", allowOrigin),
-			fmt.Sprintf("https://%s", allowOrigin),
-			fmt.Sprintf("http://%s", net.JoinHostPort(allowOrigin, "*")),
-			fmt.Sprintf("https://%s", net.JoinHostPort(allowOrigin, "*")),
-		)
-	}
-
-	AllowOrigins = append(AllowOrigins,
-		"app://*",
-		"file://*",
-		"tauri://*",
-	)
-
-	maxRunners := clean("OLLAMA_MAX_LOADED_MODELS")
-	if maxRunners != "" {
-		m, err := strconv.Atoi(maxRunners)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
-		} else {
-			MaxRunners = m
-		}
-	}
-
-	if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
-		p, err := strconv.Atoi(onp)
-		if err != nil || p <= 0 {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err)
-		} else {
-			MaxQueuedRequests = p
-		}
-	}
-
-	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
-
-	var err error
-	ModelsDir, err = getModelsDir()
-	if err != nil {
-		slog.Error("invalid setting", "OLLAMA_MODELS", ModelsDir, "error", err)
-	}
-
-	Host, err = getOllamaHost()
-	if err != nil {
-		slog.Error("invalid setting", "OLLAMA_HOST", Host, "error", err, "using default port", Host.Port)
-	}
-
-	if set, err := strconv.ParseBool(clean("OLLAMA_INTEL_GPU")); err == nil {
-		IntelGpu = set
-	}
-
-	CudaVisibleDevices = clean("CUDA_VISIBLE_DEVICES")
-	HipVisibleDevices = clean("HIP_VISIBLE_DEVICES")
-	RocrVisibleDevices = clean("ROCR_VISIBLE_DEVICES")
-	GpuDeviceOrdinal = clean("GPU_DEVICE_ORDINAL")
-	HsaOverrideGfxVersion = clean("HSA_OVERRIDE_GFX_VERSION")
-}
-
-func getModelsDir() (string, error) {
-	if models, exists := os.LookupEnv("OLLAMA_MODELS"); exists {
-		return models, nil
-	}
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return "", err
-	}
-	return filepath.Join(home, ".ollama", "models"), nil
-}
-
-func getOllamaHost() (*OllamaHost, error) {
-	defaultPort := "11434"
-
-	hostVar := os.Getenv("OLLAMA_HOST")
-	hostVar = strings.TrimSpace(strings.Trim(strings.TrimSpace(hostVar), "\"'"))
-
-	scheme, hostport, ok := strings.Cut(hostVar, "://")
-	switch {
-	case !ok:
-		scheme, hostport = "http", hostVar
-	case scheme == "http":
-		defaultPort = "80"
-	case scheme == "https":
-		defaultPort = "443"
-	}
-
-	// trim trailing slashes
-	hostport = strings.TrimRight(hostport, "/")
-
-	host, port, err := net.SplitHostPort(hostport)
-	if err != nil {
-		host, port = "127.0.0.1", defaultPort
-		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
-			host = ip.String()
-		} else if hostport != "" {
-			host = hostport
-		}
-	}
-
-	if portNum, err := strconv.ParseInt(port, 10, 32); err != nil || portNum > 65535 || portNum < 0 {
-		return &OllamaHost{
-			Scheme: scheme,
-			Host:   host,
-			Port:   defaultPort,
-		}, ErrInvalidHostPort
-	}
-
-	return &OllamaHost{
-		Scheme: scheme,
-		Host:   host,
-		Port:   port,
-	}, nil
+// On windows, we keep the binary at the top directory, but
+// other platforms use a "bin" directory, so this returns ".."
+func LibRelativeToExe() string {
+	if runtime.GOOS == "windows" {
+		return "."
+	}
+	return ".."
 }
diff --git a/envconfig/config_test.go b/envconfig/config_test.go
index 7d923d629..7ac7c53e3 100644
--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@@ -1,70 +1,269 @@
 package envconfig
 
 import (
-	"fmt"
-	"net"
+	"math"
 	"testing"
+	"time"
 
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
+	"github.com/google/go-cmp/cmp"
 )
 
-func TestConfig(t *testing.T) {
-	Debug = false // Reset whatever was loaded in init()
-	t.Setenv("OLLAMA_DEBUG", "")
-	LoadConfig()
-	require.False(t, Debug)
-	t.Setenv("OLLAMA_DEBUG", "false")
-	LoadConfig()
-	require.False(t, Debug)
-	t.Setenv("OLLAMA_DEBUG", "1")
-	LoadConfig()
-	require.True(t, Debug)
-	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
-	LoadConfig()
-	require.True(t, FlashAttention)
-}
-
-func TestClientFromEnvironment(t *testing.T) {
-	type testCase struct {
+func TestHost(t *testing.T) {
+	cases := map[string]struct {
 		value  string
 		expect string
-		err    error
+	}{
+		"empty":               {"", "http://127.0.0.1:11434"},
+		"only address":        {"1.2.3.4", "http://1.2.3.4:11434"},
+		"only port":           {":1234", "http://:1234"},
+		"address and port":    {"1.2.3.4:1234", "http://1.2.3.4:1234"},
+		"hostname":            {"example.com", "http://example.com:11434"},
+		"hostname and port":   {"example.com:1234", "http://example.com:1234"},
+		"zero port":           {":0", "http://:0"},
+		"too large port":      {":66000", "http://:11434"},
+		"too small port":      {":-1", "http://:11434"},
+		"ipv6 localhost":      {"[::1]", "http://[::1]:11434"},
+		"ipv6 world open":     {"[::]", "http://[::]:11434"},
+		"ipv6 no brackets":    {"::1", "http://[::1]:11434"},
+		"ipv6 + port":         {"[::1]:1337", "http://[::1]:1337"},
+		"extra space":         {" 1.2.3.4 ", "http://1.2.3.4:11434"},
+		"extra quotes":        {"\"1.2.3.4\"", "http://1.2.3.4:11434"},
+		"extra space+quotes":  {" \" 1.2.3.4 \" ", "http://1.2.3.4:11434"},
+		"extra single quotes": {"'1.2.3.4'", "http://1.2.3.4:11434"},
+		"http":                {"http://1.2.3.4", "http://1.2.3.4:80"},
+		"http port":           {"http://1.2.3.4:4321", "http://1.2.3.4:4321"},
+		"https":               {"https://1.2.3.4", "https://1.2.3.4:443"},
+		"https port":          {"https://1.2.3.4:4321", "https://1.2.3.4:4321"},
+		"proxy path":          {"https://example.com/ollama", "https://example.com:443/ollama"},
 	}
 
-	hostTestCases := map[string]*testCase{
-		"empty":               {value: "", expect: "127.0.0.1:11434"},
-		"only address":        {value: "1.2.3.4", expect: "1.2.3.4:11434"},
-		"only port":           {value: ":1234", expect: ":1234"},
-		"address and port":    {value: "1.2.3.4:1234", expect: "1.2.3.4:1234"},
-		"hostname":            {value: "example.com", expect: "example.com:11434"},
-		"hostname and port":   {value: "example.com:1234", expect: "example.com:1234"},
-		"zero port":           {value: ":0", expect: ":0"},
-		"too large port":      {value: ":66000", err: ErrInvalidHostPort},
-		"too small port":      {value: ":-1", err: ErrInvalidHostPort},
-		"ipv6 localhost":      {value: "[::1]", expect: "[::1]:11434"},
-		"ipv6 world open":     {value: "[::]", expect: "[::]:11434"},
-		"ipv6 no brackets":    {value: "::1", expect: "[::1]:11434"},
-		"ipv6 + port":         {value: "[::1]:1337", expect: "[::1]:1337"},
-		"extra space":         {value: " 1.2.3.4 ", expect: "1.2.3.4:11434"},
-		"extra quotes":        {value: "\"1.2.3.4\"", expect: "1.2.3.4:11434"},
-		"extra space+quotes":  {value: " \" 1.2.3.4 \" ", expect: "1.2.3.4:11434"},
-		"extra single quotes": {value: "'1.2.3.4'", expect: "1.2.3.4:11434"},
-	}
-
-	for k, v := range hostTestCases {
-		t.Run(k, func(t *testing.T) {
-			t.Setenv("OLLAMA_HOST", v.value)
-			LoadConfig()
-
-			oh, err := getOllamaHost()
-			if err != v.err {
-				t.Fatalf("expected %s, got %s", v.err, err)
-			}
-
-			if err == nil {
-				host := net.JoinHostPort(oh.Host, oh.Port)
-				assert.Equal(t, v.expect, host, fmt.Sprintf("%s: expected %s, got %s", k, v.expect, host))
+	for name, tt := range cases {
+		t.Run(name, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", tt.value)
+			if host := Host(); host.String() != tt.expect {
+				t.Errorf("%s: expected %s, got %s", name, tt.expect, host.String())
+			}
+		})
+	}
+}
+
+func TestOrigins(t *testing.T) {
+	cases := []struct {
+		value  string
+		expect []string
+	}{
+		{"", []string{
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://10.0.0.1", []string{
+			"http://10.0.0.1",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://172.16.0.1,https://192.168.0.1", []string{
+			"http://172.16.0.1",
+			"https://192.168.0.1",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+		{"http://totally.safe,http://definitely.legit", []string{
+			"http://totally.safe",
+			"http://definitely.legit",
+			"http://localhost",
+			"https://localhost",
+			"http://localhost:*",
+			"https://localhost:*",
+			"http://127.0.0.1",
+			"https://127.0.0.1",
+			"http://127.0.0.1:*",
+			"https://127.0.0.1:*",
+			"http://0.0.0.0",
+			"https://0.0.0.0",
+			"http://0.0.0.0:*",
+			"https://0.0.0.0:*",
+			"app://*",
+			"file://*",
+			"tauri://*",
+		}},
+	}
+	for _, tt := range cases {
+		t.Run(tt.value, func(t *testing.T) {
+			t.Setenv("OLLAMA_ORIGINS", tt.value)
+
+			if diff := cmp.Diff(Origins(), tt.expect); diff != "" {
+				t.Errorf("%s: mismatch (-want +got):\n%s", tt.value, diff)
+			}
+		})
+	}
+}
+
+func TestBool(t *testing.T) {
+	cases := map[string]bool{
+		"":      false,
+		"true":  true,
+		"false": false,
+		"1":     true,
+		"0":     false,
+		// invalid values
+		"random":    true,
+		"something": true,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_BOOL", k)
+			if b := Bool("OLLAMA_BOOL")(); b != v {
+				t.Errorf("%s: expected %t, got %t", k, v, b)
+			}
+		})
+	}
+}
+
+func TestUint(t *testing.T) {
+	cases := map[string]uint{
+		"0":    0,
+		"1":    1,
+		"1337": 1337,
+		// default values
+		"":       11434,
+		"-1":     11434,
+		"0o10":   11434,
+		"0x10":   11434,
+		"string": 11434,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_UINT", k)
+			if i := Uint("OLLAMA_UINT", 11434)(); i != v {
+				t.Errorf("%s: expected %d, got %d", k, v, i)
+			}
+		})
+	}
+}
+
+func TestKeepAlive(t *testing.T) {
+	cases := map[string]time.Duration{
+		"":       5 * time.Minute,
+		"1s":     time.Second,
+		"1m":     time.Minute,
+		"1h":     time.Hour,
+		"5m0s":   5 * time.Minute,
+		"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
+		"0":      time.Duration(0),
+		"60":     60 * time.Second,
+		"120":    2 * time.Minute,
+		"3600":   time.Hour,
+		"-0":     time.Duration(0),
+		"-1":     time.Duration(math.MaxInt64),
+		"-1m":    time.Duration(math.MaxInt64),
+		// invalid values
+		" ":   5 * time.Minute,
+		"???": 5 * time.Minute,
+		"1d":  5 * time.Minute,
+		"1y":  5 * time.Minute,
+		"1w":  5 * time.Minute,
+	}
+
+	for tt, expect := range cases {
+		t.Run(tt, func(t *testing.T) {
+			t.Setenv("OLLAMA_KEEP_ALIVE", tt)
+			if actual := KeepAlive(); actual != expect {
+				t.Errorf("%s: expected %s, got %s", tt, expect, actual)
+			}
+		})
+	}
+}
+
+func TestLoadTimeout(t *testing.T) {
+	defaultTimeout := 5 * time.Minute
+	cases := map[string]time.Duration{
+		"":       defaultTimeout,
+		"1s":     time.Second,
+		"1m":     time.Minute,
+		"1h":     time.Hour,
+		"5m0s":   defaultTimeout,
+		"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
+		"0":      time.Duration(math.MaxInt64),
+		"60":     60 * time.Second,
+		"120":    2 * time.Minute,
+		"3600":   time.Hour,
+		"-0":     time.Duration(math.MaxInt64),
+		"-1":     time.Duration(math.MaxInt64),
+		"-1m":    time.Duration(math.MaxInt64),
+		// invalid values
+		" ":   defaultTimeout,
+		"???": defaultTimeout,
+		"1d":  defaultTimeout,
+		"1y":  defaultTimeout,
+		"1w":  defaultTimeout,
+	}
+
+	for tt, expect := range cases {
+		t.Run(tt, func(t *testing.T) {
+			t.Setenv("OLLAMA_LOAD_TIMEOUT", tt)
+			if actual := LoadTimeout(); actual != expect {
+				t.Errorf("%s: expected %s, got %s", tt, expect, actual)
+			}
+		})
+	}
+}
+
+func TestVar(t *testing.T) {
+	cases := map[string]string{
+		"value":       "value",
+		" value ":     "value",
+		" 'value' ":   "value",
+		` "value" `:   "value",
+		" ' value ' ": " value ",
+		` " value " `: " value ",
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_VAR", k)
+			if s := Var("OLLAMA_VAR"); s != v {
+				t.Errorf("%s: expected %q, got %q", k, v, s)
 			}
 		})
 	}
diff --git a/examples/go-chat/main.go b/examples/go-chat/main.go
index 5266f03e9..7663fb8f4 100644
--- a/examples/go-chat/main.go
+++ b/examples/go-chat/main.go
@@ -35,7 +35,7 @@ func main() {
 
 	ctx := context.Background()
 	req := &api.ChatRequest{
-		Model:    "llama3",
+		Model:    "llama3.1",
 		Messages: messages,
 	}
 
diff --git a/examples/go-generate-streaming/main.go b/examples/go-generate-streaming/main.go
index 494033511..3acfb22a9 100644
--- a/examples/go-generate-streaming/main.go
+++ b/examples/go-generate-streaming/main.go
@@ -16,7 +16,7 @@ func main() {
 
 	// By default, GenerateRequest is streaming.
 	req := &api.GenerateRequest{
-		Model:  "gemma",
+		Model:  "gemma2",
 		Prompt: "how many planets are there?",
 	}
 
diff --git a/examples/go-generate/main.go b/examples/go-generate/main.go
index 50fbf64b7..2fe28742b 100644
--- a/examples/go-generate/main.go
+++ b/examples/go-generate/main.go
@@ -15,7 +15,7 @@ func main() {
 	}
 
 	req := &api.GenerateRequest{
-		Model:  "gemma",
+		Model:  "gemma2",
 		Prompt: "how many planets are there?",
 
 		// set streaming to false
diff --git a/examples/go-http-generate/README.md b/examples/go-http-generate/README.md
deleted file mode 100644
index e69de29bb..000000000
diff --git a/examples/langchain-python-rag-document/README.md b/examples/langchain-python-rag-document/README.md
index 20a73a883..e2f3bc028 100644
--- a/examples/langchain-python-rag-document/README.md
+++ b/examples/langchain-python-rag-document/README.md
@@ -4,6 +4,14 @@ This example provides an interface for asking questions to a PDF document.
 
 ## Setup
 
+1. Ensure you have the `llama3.1` model installed:
+
+```
+ollama pull llama3.1
+```
+
+2. Install the Python Requirements.
+
 ```
 pip install -r requirements.txt
 ```
diff --git a/examples/langchain-python-rag-document/main.py b/examples/langchain-python-rag-document/main.py
index 3ed9499f2..6f7cec9be 100644
--- a/examples/langchain-python-rag-document/main.py
+++ b/examples/langchain-python-rag-document/main.py
@@ -51,7 +51,7 @@ while True:
         template=template,
     )
 
-    llm = Ollama(model="llama3:8b", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
     qa_chain = RetrievalQA.from_chain_type(
         llm,
         retriever=vectorstore.as_retriever(),
diff --git a/examples/langchain-python-rag-privategpt/requirements.txt b/examples/langchain-python-rag-privategpt/requirements.txt
index 0aad1fe55..4f2cee250 100644
--- a/examples/langchain-python-rag-privategpt/requirements.txt
+++ b/examples/langchain-python-rag-privategpt/requirements.txt
@@ -1,6 +1,6 @@
 langchain==0.0.274
 gpt4all==1.0.8
-chromadb==0.4.7
+chromadb==0.5.0
 llama-cpp-python==0.1.81
 urllib3==2.0.4
 PyMuPDF==1.23.5
@@ -12,4 +12,4 @@ pandoc==2.3
 pypandoc==1.11
 tqdm==4.66.1
 sentence_transformers==2.2.2
-numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
\ No newline at end of file
+numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
diff --git a/examples/langchain-python-rag-websummary/README.md b/examples/langchain-python-rag-websummary/README.md
index 3f3b98733..29c706a39 100644
--- a/examples/langchain-python-rag-websummary/README.md
+++ b/examples/langchain-python-rag-websummary/README.md
@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso
 
 ## Running the Example
 
-1. Ensure you have the `llama2` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama2
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/langchain-python-rag-websummary/main.py b/examples/langchain-python-rag-websummary/main.py
index d1b05ba8a..77b09fbbc 100644
--- a/examples/langchain-python-rag-websummary/main.py
+++ b/examples/langchain-python-rag-websummary/main.py
@@ -5,8 +5,8 @@ from langchain.chains.summarize import load_summarize_chain
 loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
-llm = Ollama(model="llama3")
+llm = Ollama(model="llama3.1")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
-result = chain.invoke(docs) 
+result = chain.invoke(docs)
 print(result)
diff --git a/examples/langchain-python-simple/README.md b/examples/langchain-python-simple/README.md
index d4102dec7..60db2c8c3 100644
--- a/examples/langchain-python-simple/README.md
+++ b/examples/langchain-python-simple/README.md
@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/langchain-python-simple/main.py b/examples/langchain-python-simple/main.py
index 7cb652867..a7ed81d67 100644
--- a/examples/langchain-python-simple/main.py
+++ b/examples/langchain-python-simple/main.py
@@ -1,6 +1,6 @@
 from langchain.llms import Ollama
 
 input = input("What is your question?")
-llm = Ollama(model="llama3")
+llm = Ollama(model="llama3.1")
 res = llm.predict(input)
 print (res)
diff --git a/examples/modelfile-mario/Modelfile b/examples/modelfile-mario/Modelfile
index 33d5952b1..a37470864 100644
--- a/examples/modelfile-mario/Modelfile
+++ b/examples/modelfile-mario/Modelfile
@@ -1,4 +1,4 @@
-FROM llama3
+FROM llama3.1
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from super mario bros, acting as an assistant.
diff --git a/examples/modelfile-mario/readme.md b/examples/modelfile-mario/readme.md
index e4f0d4172..c3f34197a 100644
--- a/examples/modelfile-mario/readme.md
+++ b/examples/modelfile-mario/readme.md
@@ -2,12 +2,12 @@
 
 # Example character: Mario
 
-This example shows how to create a basic character using Llama3 as the base model.
+This example shows how to create a basic character using Llama3.1 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3` to get the base model used in the model file.
+2. `ollama pull llama3.1` to get the base model used in the model file.
 3. `ollama create NAME -f ./Modelfile`
 4. `ollama run NAME`
 
@@ -18,7 +18,7 @@ Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
 What the model file looks like:
 
 ```
-FROM llama3
+FROM llama3.1
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from Super Mario Bros, acting as an assistant.
diff --git a/examples/python-dockerit/dockerit.py b/examples/python-dockerit/dockerit.py
index b013102fa..6a288d906 100644
--- a/examples/python-dockerit/dockerit.py
+++ b/examples/python-dockerit/dockerit.py
@@ -4,7 +4,7 @@ imageName = input("Enter the name of the image: ")
 client = docker.from_env()
 s = requests.Session()
 output=""
-with s.post('http://localhost:11434/api/generate', json={'model': 'dockerit', 'prompt': inputDescription}, stream=True) as r:
+with s.post('http://localhost:11434/api/generate', json={'model': 'mattw/dockerit', 'prompt': inputDescription}, stream=True) as r:
   for line in r.iter_lines():
     if line:
       j = json.loads(line)
diff --git a/examples/python-grounded-factuality-rag-check/README.md b/examples/python-grounded-factuality-rag-check/README.md
new file mode 100644
index 000000000..5c9817526
--- /dev/null
+++ b/examples/python-grounded-factuality-rag-check/README.md
@@ -0,0 +1,93 @@
+# RAG Hallucination Checker using Bespoke-Minicheck
+
+This example allows the user to ask questions related to a document, which can be specified via an article url. Relevant chunks are retreived from the document and given to `llama3.1` as context to answer the question. Then each sentence in the answer is checked against the retrieved chunks using `bespoke-minicheck` to ensure that the answer does not contain hallucinations. 
+
+## Running the Example
+
+1. Ensure `all-minilm` (embedding) `llama3.1` (chat) and `bespoke-minicheck` (check) models installed:
+
+   ```bash
+   ollama pull all-minilm
+   ollama pull llama3.1
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies.
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the example:
+
+   ```bash
+   python main.py
+   ```
+
+## Expected Output
+
+```text
+Enter the URL of an article you want to chat with, or press Enter for default example:
+
+Loaded, chunked, and embedded text from https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt.
+
+Enter your question or type quit: Who is the CEO of openai?
+
+Retrieved chunks:
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence .
+
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence . More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+LLM Answer:
+The text does not mention the CEO of OpenAI. It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+
+LLM Claim: The text does not mention the CEO of OpenAI.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+Is this claim supported by the context according to bespoke-minicheck? No
+```
+
+The second claim is unsupported since the text mentions the research lead. 
+
+Another tricky example:
+
+```text
+
+Enter your question or type quit: what sets o1 apart from gpt-4o?
+
+Retrieved chunks: 
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+He says OpenAI also tested o1 against a qualifying exam for the International Mathematics Olympiad , and while GPT-4o only correctly solved only 13 percent of problems , o1 scored 83 percent . “ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+“ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world . It also doesn ’ t have the ability to browse the web or process files and images . Still , the company believes it represents a brand-new class of capabilities . It was named o1 to indicate “ resetting the counter back to 1. ” “ I ’ m gon na be honest : I think we ’ re terrible at naming , traditionally , ” McGrew says .
+LLM Answer: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+* The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+* However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+
+LLM Claim: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+Is this claim supported by the context according to bespoke-minicheck? No
+
+LLM Claim: * However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+```
+
+We see that the third claim "* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance." is not supported by the context. This is because the context only mentions that o1 "is claimed to perform" which is different from "has been shown to perform".
diff --git a/examples/python-grounded-factuality-rag-check/main.py b/examples/python-grounded-factuality-rag-check/main.py
new file mode 100644
index 000000000..f4d562d5f
--- /dev/null
+++ b/examples/python-grounded-factuality-rag-check/main.py
@@ -0,0 +1,137 @@
+import ollama
+import warnings
+from mattsollamatools import chunker
+from newspaper import Article
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+import nltk
+
+warnings.filterwarnings(
+    "ignore", category=FutureWarning, module="transformers.tokenization_utils_base"
+)
+nltk.download("punkt_tab", quiet=True)
+
+
+def getArticleText(url):
+    """Gets the text of an article from a URL.
+
+    Often there are a bunch of ads and menus on pages for a news article.
+    This uses newspaper3k to get just the text of just the article.
+    """
+    article = Article(url)
+    article.download()
+    article.parse()
+    return article.text
+
+
+def knn_search(question_embedding, embeddings, k=5):
+    """Performs K-nearest neighbors (KNN) search"""
+    X = np.array(
+        [item["embedding"] for article in embeddings for item in article["embeddings"]]
+    )
+    source_texts = [
+        item["source"] for article in embeddings for item in article["embeddings"]
+    ]
+
+    # Fit a KNN model on the embeddings
+    knn = NearestNeighbors(n_neighbors=k, metric="cosine")
+    knn.fit(X)
+
+    # Find the indices and distances of the k-nearest neighbors.
+    _, indices = knn.kneighbors(question_embedding, n_neighbors=k)
+
+    # Get the indices and source texts of the best matches
+    best_matches = [(indices[0][i], source_texts[indices[0][i]]) for i in range(k)]
+
+    return best_matches
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+if __name__ == "__main__":
+    allEmbeddings = []
+    default_url = "https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt"
+    user_input = input(
+        "Enter the URL of an article you want to chat with, or press Enter for default example: "
+    )
+    article_url = user_input.strip() if user_input.strip() else default_url
+    article = {}
+    article["embeddings"] = []
+    article["url"] = article_url
+    text = getArticleText(article_url)
+    chunks = chunker(text)
+
+    # Embed (batch) chunks using ollama
+    embeddings = ollama.embed(model="all-minilm", input=chunks)["embeddings"]
+
+    for chunk, embedding in zip(chunks, embeddings):
+        item = {}
+        item["source"] = chunk
+        item["embedding"] = embedding
+        item["sourcelength"] = len(chunk)
+        article["embeddings"].append(item)
+
+    allEmbeddings.append(article)
+
+    print(f"\nLoaded, chunked, and embedded text from {article_url}.\n")
+
+    while True:
+        # Input a question from the user
+        # For example, "Who is the chief research officer?"
+        question = input("Enter your question or type quit: ")
+
+        if question.lower() == "quit":
+            break
+
+        # Embed the user's question using ollama.embed
+        question_embedding = ollama.embed(model="all-minilm", input=question)[
+            "embeddings"
+        ]
+
+        # Perform KNN search to find the best matches (indices and source text)
+        best_matches = knn_search(question_embedding, allEmbeddings, k=4)
+
+        sourcetext = "\n\n".join([source_text for (_, source_text) in best_matches])
+
+        print(f"\nRetrieved chunks: \n{sourcetext}\n")
+
+        # Give the retreived chunks and question to the chat model
+        system_prompt = f"Only use the following information to answer the question. Do not use anything else: {sourcetext}"
+
+        ollama_response = ollama.generate(
+            model="llama3.1",
+            prompt=question,
+            system=system_prompt,
+            options={"stream": False},
+        )
+
+        answer = ollama_response["response"]
+        print(f"LLM Answer:\n{answer}\n")
+
+        # Check each sentence in the response for grounded factuality
+        if answer:
+            for claim in nltk.sent_tokenize(answer):
+                print(f"LLM Claim: {claim}")
+                print(
+                    f"Is this claim supported by the context according to bespoke-minicheck? {check(sourcetext, claim)}\n"
+                )
diff --git a/examples/python-grounded-factuality-rag-check/requirements.txt b/examples/python-grounded-factuality-rag-check/requirements.txt
new file mode 100644
index 000000000..d4bd6df34
--- /dev/null
+++ b/examples/python-grounded-factuality-rag-check/requirements.txt
@@ -0,0 +1,8 @@
+ollama
+lxml==5.3.0
+lxml_html_clean==0.2.2
+mattsollamatools==0.0.25
+newspaper3k==0.2.8
+nltk==3.9.1
+numpy==1.26.4
+scikit-learn==1.5.2
\ No newline at end of file
diff --git a/examples/python-grounded-factuality-simple-check/main.py b/examples/python-grounded-factuality-simple-check/main.py
new file mode 100644
index 000000000..0204f3b39
--- /dev/null
+++ b/examples/python-grounded-factuality-simple-check/main.py
@@ -0,0 +1,53 @@
+"""Simple example to demonstrate how to use the bespoke-minicheck model."""
+
+import ollama
+
+# NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+def get_user_input(prompt):
+    user_input = input(prompt)
+    if not user_input:
+        exit()
+    print()
+    return user_input
+
+
+def main():
+    while True:
+        # Get a document from the user (e.g. "Ryan likes running and biking.")
+        document = get_user_input("Enter a document: ")
+        # Get a claim from the user (e.g. "Ryan likes to run.")
+        claim = get_user_input("Enter a claim: ")
+        # Check if the claim is supported by the document
+        grounded_factuality_check = check(document, claim)
+        print(
+            f"Is the claim supported by the document according to bespoke-minicheck? {grounded_factuality_check}"
+        )
+        print("\n\n")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/python-grounded-factuality-simple-check/readme.md b/examples/python-grounded-factuality-simple-check/readme.md
new file mode 100644
index 000000000..b164b5eb9
--- /dev/null
+++ b/examples/python-grounded-factuality-simple-check/readme.md
@@ -0,0 +1,54 @@
+# Simple Bespoke-Minicheck Example
+
+`bespoke-minicheck` is a model for checking if a claim is supported by a document. It is used through the **generate** endpoint, which is called in this example with a `prompt` that includes the expected formatting of the user input. 
+
+## Running the Example
+
+1. Ensure you have the `bespoke-minicheck` model installed:
+
+   ```bash
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies:
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the program:
+
+   ```bash
+   python main.py
+   ```
+
+4. Enter a document and a claim when prompted:
+
+   ```bash
+   Enter a document: Roses are red.
+
+   Enter a claim: Roses are blue. 
+   ```
+
+   The claim and document are then given to the `bespoke-minicheck` as inputs, which then generates a response (Yes or No) on whether the claim is supported by the document.
+
+   ```bash
+   Is the claim supported by the document according to bespoke-minicheck? No
+   ```
+
+## More Examples
+
+Document ([source](https://en.wikipedia.org/wiki/Apple_I)): 
+> The Apple Computer 1 (Apple-1[a]), later known predominantly as the Apple I(written with a Roman numeral),[b] is an 8-bit motherboard-only personal computer designed by Steve Wozniak[5][6] and released by the Apple Computer Company (now Apple Inc.) in 1976. The company was initially formed to sell the Apple I – its first product – and would later become the world's largest technology company.[7] The idea of starting a company and selling the computer came from Wozniak's friend and Apple co-founder Steve Jobs.[8][9] One of the main innovations of the Apple I was that it included video display terminal circuitry on its circuit board, allowing it to connect to a low-cost composite video monitor or television, instead of an expensive computer terminal, compared to most existing computers at the time.
+
+Claim: 
+>The Apple I is a 16-bit computer.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **No**
+
+Claim: 
+>Apple was originally called the Apple Computer Company.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **Yes**
diff --git a/examples/python-grounded-factuality-simple-check/requirements.txt b/examples/python-grounded-factuality-simple-check/requirements.txt
new file mode 100644
index 000000000..403abba66
--- /dev/null
+++ b/examples/python-grounded-factuality-simple-check/requirements.txt
@@ -0,0 +1 @@
+ollama
diff --git a/examples/python-json-datagenerator/predefinedschema.py b/examples/python-json-datagenerator/predefinedschema.py
index 1fd54892b..68090ad79 100644
--- a/examples/python-json-datagenerator/predefinedschema.py
+++ b/examples/python-json-datagenerator/predefinedschema.py
@@ -2,7 +2,7 @@ import requests
 import json
 import random
 
-model = "llama3"
+model = "llama3.1"
 template = {
   "firstName": "",
   "lastName": "",
diff --git a/examples/python-json-datagenerator/randomaddresses.py b/examples/python-json-datagenerator/randomaddresses.py
index 72b1fefb9..878c98037 100644
--- a/examples/python-json-datagenerator/randomaddresses.py
+++ b/examples/python-json-datagenerator/randomaddresses.py
@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3"
+model = "llama3.1"
 
 prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
 
diff --git a/examples/python-json-datagenerator/readme.md b/examples/python-json-datagenerator/readme.md
index 883570443..5b444dff1 100644
--- a/examples/python-json-datagenerator/readme.md
+++ b/examples/python-json-datagenerator/readme.md
@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/python-loganalysis/Modelfile b/examples/python-loganalysis/Modelfile
index 5237cb6eb..b28aa0c0f 100644
--- a/examples/python-loganalysis/Modelfile
+++ b/examples/python-loganalysis/Modelfile
@@ -4,5 +4,5 @@ SYSTEM """
 You are a log file analyzer. You will receive a set of lines from a log file for some software application, find the errors and other interesting aspects of the logs, and explain them so a new user can understand what they mean. If there are any steps they can do to resolve them, list the steps in your answer.
 """
 
-PARAMETER TEMPERATURE 0.3
+PARAMETER temperature 0.3
 
diff --git a/examples/python-loganalysis/readme.md b/examples/python-loganalysis/readme.md
index 4be0baaaf..03bab672e 100644
--- a/examples/python-loganalysis/readme.md
+++ b/examples/python-loganalysis/readme.md
@@ -21,6 +21,8 @@ You can try this with the `logtest.logfile` file included in this directory.
 2. Install the Python Requirements.
 
    ```bash
+   python3 -m venv .venv
+   source .venv/bin/activate
    pip install -r requirements.txt
    ```
 
diff --git a/examples/python-loganalysis/requirements.txt b/examples/python-loganalysis/requirements.txt
index 9688b8ec0..e7cb17efb 100644
--- a/examples/python-loganalysis/requirements.txt
+++ b/examples/python-loganalysis/requirements.txt
@@ -1 +1 @@
-Requests==2.31.0
+Requests>=2.32.3
diff --git a/examples/python-simplechat/client.py b/examples/python-simplechat/client.py
index f82a16b3e..85043d5f4 100644
--- a/examples/python-simplechat/client.py
+++ b/examples/python-simplechat/client.py
@@ -2,7 +2,7 @@ import json
 import requests
 
 # NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
-model = "llama3"  # TODO: update this for whatever model you wish to use
+model = "llama3.1"  # TODO: update this for whatever model you wish to use
 
 
 def chat(messages):
diff --git a/examples/python-simplechat/readme.md b/examples/python-simplechat/readme.md
index dd2576bc1..4c2ded4d8 100644
--- a/examples/python-simplechat/readme.md
+++ b/examples/python-simplechat/readme.md
@@ -4,10 +4,10 @@ The **chat** endpoint is one of two ways to generate text from an LLM with Ollam
 
 ## Running the Example
 
-1. Ensure you have the `llama3` model installed:
+1. Ensure you have the `llama3.1` model installed:
 
    ```bash
-   ollama pull llama3
+   ollama pull llama3.1
    ```
 
 2. Install the Python Requirements.
diff --git a/examples/typescript-simplechat/client.ts b/examples/typescript-simplechat/client.ts
index a1e0eea37..8ad113b12 100644
--- a/examples/typescript-simplechat/client.ts
+++ b/examples/typescript-simplechat/client.ts
@@ -1,6 +1,6 @@
 import * as readline from "readline";
 
-const model = "llama3";
+const model = "llama3.1";
 type Message = {
   role: "assistant" | "user" | "system";
   content: string;
diff --git a/format/format.go b/format/format.go
index 31059578f..ac50570df 100644
--- a/format/format.go
+++ b/format/format.go
@@ -3,6 +3,7 @@ package format
 import (
 	"fmt"
 	"math"
+	"strconv"
 )
 
 const (
@@ -28,6 +29,6 @@ func HumanNumber(b uint64) string {
 	case b >= Thousand:
 		return fmt.Sprintf("%.0fK", float64(b)/Thousand)
 	default:
-		return fmt.Sprintf("%d", b)
+		return strconv.FormatUint(b, 10)
 	}
 }
diff --git a/go.mod b/go.mod
index 6807b9b48..6e437c730 100644
--- a/go.mod
+++ b/go.mod
@@ -1,6 +1,6 @@
 module github.com/ollama/ollama
 
-go 1.22.0
+go 1.22.5
 
 require (
 	github.com/containerd/console v1.0.3
@@ -18,6 +18,7 @@ require (
 require (
 	github.com/agnivade/levenshtein v1.1.1
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
+	github.com/google/go-cmp v0.6.0
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
@@ -71,7 +72,7 @@ require (
 	golang.org/x/net v0.25.0 // indirect
 	golang.org/x/sys v0.20.0
 	golang.org/x/term v0.20.0
-	golang.org/x/text v0.15.0 // indirect
+	golang.org/x/text v0.15.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )
diff --git a/gpu/amd_common.go b/gpu/amd_common.go
index 27a81e3f8..2894ac2c4 100644
--- a/gpu/amd_common.go
+++ b/gpu/amd_common.go
@@ -3,12 +3,14 @@
 package gpu
 
 import (
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // Determine if the given ROCm lib directory is usable by checking for existence of some glob patterns
@@ -49,9 +51,17 @@ func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 }
 
 func commonAMDValidateLibDir() (string, error) {
-	// We try to favor system paths first, so that we can wire up the subprocess to use
-	// the system version.  Only use our bundled version if the system version doesn't work
-	// This gives users a more recovery options if versions have subtle problems at runtime
+	// Favor our bundled version
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe(), "lib", "ollama")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
+	}
 
 	// Prefer explicit HIP env var
 	hipPath := os.Getenv("HIP_PATH")
@@ -87,14 +97,5 @@ func commonAMDValidateLibDir() (string, error) {
 		}
 	}
 
-	// Installer payload location if we're running the installed binary
-	exe, err := os.Executable()
-	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
-	}
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
diff --git a/gpu/amd_hip_windows.go b/gpu/amd_hip_windows.go
index 8572a24c5..2cea28242 100644
--- a/gpu/amd_hip_windows.go
+++ b/gpu/amd_hip_windows.go
@@ -1,6 +1,7 @@
 package gpu
 
 import (
+	"errors"
 	"fmt"
 	"log/slog"
 	"syscall"
@@ -33,9 +34,10 @@ type HipLib struct {
 }
 
 func NewHipLib() (*HipLib, error) {
-	h, err := windows.LoadLibrary("amdhip64.dll")
+	// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
+	h, err := windows.LoadLibrary("amdhip64_6.dll")
 	if err != nil {
-		return nil, fmt.Errorf("unable to load amdhip64.dll: %w", err)
+		return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
 	}
 	hl := &HipLib{}
 	hl.dll = h
@@ -75,7 +77,7 @@ func (hl *HipLib) Release() {
 
 func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var version int
 	status, _, err := syscall.SyscallN(hl.hipDriverGetVersion, uintptr(unsafe.Pointer(&version)))
@@ -84,9 +86,8 @@ func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 
 	slog.Debug("hipDriverGetVersion", "version", version)
-	// TODO - this isn't actually right, but the docs claim hipDriverGetVersion isn't accurate anyway...
-	driverMajor = version / 1000
-	driverMinor = (version - (driverMajor * 1000)) / 10
+	driverMajor = version / 10000000
+	driverMinor = (version - (driverMajor * 10000000)) / 100000
 
 	return driverMajor, driverMinor, nil
 }
@@ -110,7 +111,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
 
 func (hl *HipLib) HipSetDevice(device int) error {
 	if hl.dll == 0 {
-		return fmt.Errorf("dll has been unloaded")
+		return errors.New("dll has been unloaded")
 	}
 	status, _, err := syscall.SyscallN(hl.hipSetDevice, uintptr(device))
 	if status != hipSuccess {
@@ -121,7 +122,7 @@ func (hl *HipLib) HipSetDevice(device int) error {
 
 func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, error) {
 	if hl.dll == 0 {
-		return nil, fmt.Errorf("dll has been unloaded")
+		return nil, errors.New("dll has been unloaded")
 	}
 	var props hipDevicePropMinimal
 	status, _, err := syscall.SyscallN(hl.hipGetDeviceProperties, uintptr(unsafe.Pointer(&props)), uintptr(device))
@@ -134,7 +135,7 @@ func (hl *HipLib) HipGetDeviceProperties(device int) (*hipDevicePropMinimal, err
 // free, total, err
 func (hl *HipLib) HipMemGetInfo() (uint64, uint64, error) {
 	if hl.dll == 0 {
-		return 0, 0, fmt.Errorf("dll has been unloaded")
+		return 0, 0, errors.New("dll has been unloaded")
 	}
 	var totalMemory uint64
 	var freeMemory uint64
diff --git a/gpu/amd_linux.go b/gpu/amd_linux.go
index 15b6fc61f..d3f5b9fc6 100644
--- a/gpu/amd_linux.go
+++ b/gpu/amd_linux.go
@@ -5,11 +5,13 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"io/fs"
 	"log/slog"
 	"os"
 	"path/filepath"
 	"regexp"
 	"slices"
+	"sort"
 	"strconv"
 	"strings"
 
@@ -60,9 +62,9 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 
 	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
 	var visibleDevices []string
-	hipVD := envconfig.HipVisibleDevices   // zero based index only
-	rocrVD := envconfig.RocrVisibleDevices // zero based index or UUID, but consumer cards seem to not support UUID
-	gpuDO := envconfig.GpuDeviceOrdinal    // zero based index
+	hipVD := envconfig.HipVisibleDevices()   // zero based index only
+	rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID, but consumer cards seem to not support UUID
+	gpuDO := envconfig.GpuDeviceOrdinal()    // zero based index
 	switch {
 	// TODO is this priorty order right?
 	case hipVD != "":
@@ -75,13 +77,27 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		visibleDevices = strings.Split(gpuDO, ",")
 	}
 
-	gfxOverride := envconfig.HsaOverrideGfxVersion
+	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	var supported []string
 	libDir := ""
 
 	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
 	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
 	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
+	sort.Slice(matches, func(i, j int) bool {
+		// /sys/class/kfd/kfd/topology/nodes/<number>/properties
+		a, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[i])), 10, 64)
+		if err != nil {
+			slog.Debug("parse err", "error", err, "match", matches[i])
+			return false
+		}
+		b, err := strconv.ParseInt(filepath.Base(filepath.Dir(matches[j])), 10, 64)
+		if err != nil {
+			slog.Debug("parse err", "error", err, "match", matches[i])
+			return false
+		}
+		return a < b
+	})
 	cpuCount := 0
 	for _, match := range matches {
 		slog.Debug("evaluating amdgpu node " + match)
@@ -344,6 +360,10 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	if len(resp) == 0 {
 		slog.Info("no compatible amdgpu devices detected")
 	}
+	if err := verifyKFDDriverAccess(); err != nil {
+		slog.Error("amdgpu devices detected but permission problems block access", "error", err)
+		return nil
+	}
 	return resp
 }
 
@@ -378,7 +398,7 @@ func AMDValidateLibDir() (string, error) {
 
 	// If we still haven't found a usable rocm, the user will have to install it on their own
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
 
 func AMDDriverVersion() (driverMajor, driverMinor int, err error) {
@@ -440,3 +460,19 @@ func getFreeMemory(usedFile string) (uint64, error) {
 	}
 	return usedMemory, nil
 }
+
+func verifyKFDDriverAccess() error {
+	// Verify we have permissions - either running as root, or we have group access to the driver
+	fd, err := os.OpenFile("/dev/kfd", os.O_RDWR, 0o666)
+	if err != nil {
+		if errors.Is(err, fs.ErrPermission) {
+			return fmt.Errorf("permissions not set up properly.  Either run ollama as root, or add you user account to the render group. %w", err)
+		} else if errors.Is(err, fs.ErrNotExist) {
+			// Container runtime failure?
+			return fmt.Errorf("kfd driver not loaded.  If running in a container, remember to include '--device /dev/kfd --device /dev/dri'")
+		}
+		return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
+	}
+	fd.Close()
+	return nil
+}
diff --git a/gpu/amd_windows.go b/gpu/amd_windows.go
index 8b6fabebb..ef6bf830c 100644
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@@ -2,7 +2,7 @@ package gpu
 
 import (
 	"bytes"
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -22,8 +22,8 @@ const (
 
 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // TODO - probably include more coverage of files here...
-	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
+	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // This is not sufficient to discern v5 vs v6
+	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
 )
 
 func AMDGetGPUInfo() []RocmGPUInfo {
@@ -35,12 +35,11 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	}
 	defer hl.Release()
 
-	// TODO - this reports incorrect version information, so omitting for now
-	// driverMajor, driverMinor, err := hl.AMDDriverVersion()
-	// if err != nil {
-	// 	// For now this is benign, but we may eventually need to fail compatibility checks
-	// 	slog.Debug("error looking up amd driver version", "error", err)
-	// }
+	driverMajor, driverMinor, err := hl.AMDDriverVersion()
+	if err != nil {
+		// For now this is benign, but we may eventually need to fail compatibility checks
+		slog.Debug("error looking up amd driver version", "error", err)
+	}
 
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
@@ -54,7 +53,7 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	}
 
 	var supported []string
-	gfxOverride := envconfig.HsaOverrideGfxVersion
+	gfxOverride := envconfig.HsaOverrideGfxVersion()
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
@@ -86,14 +85,15 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
 		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
-		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
+		// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
 		if strings.EqualFold(name, iGPUName) {
 			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
 			continue
 		}
 		if gfxOverride == "" {
-			if !slices.Contains[[]string, string](supported, gfx) {
+			// Strip off Target Features when comparing
+			if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
 				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
@@ -132,10 +132,8 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
-
-				// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
-				// DriverMajor:    driverMajor,
-				// DriverMinor:    driverMinor,
+				DriverMajor:    driverMajor,
+				DriverMinor:    driverMinor,
 			},
 			index: i,
 		}
@@ -155,7 +153,7 @@ func AMDValidateLibDir() (string, error) {
 	// Installer payload (if we're running from some other location)
 	localAppData := os.Getenv("LOCALAPPDATA")
 	appDir := filepath.Join(localAppData, "Programs", "Ollama")
-	rocmTargetDir := filepath.Join(appDir, "rocm")
+	rocmTargetDir := filepath.Join(appDir, envconfig.LibRelativeToExe(), "lib", "ollama")
 	if rocmLibUsable(rocmTargetDir) {
 		slog.Debug("detected ollama installed ROCm at " + rocmTargetDir)
 		return rocmTargetDir, nil
@@ -163,7 +161,7 @@ func AMDValidateLibDir() (string, error) {
 
 	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
-	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
+	return "", errors.New("no suitable rocm found, falling back to CPU")
 }
 
 func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
diff --git a/gpu/assets.go b/gpu/assets.go
deleted file mode 100644
index 073d2e813..000000000
--- a/gpu/assets.go
+++ /dev/null
@@ -1,145 +0,0 @@
-package gpu
-
-import (
-	"errors"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"runtime"
-	"strconv"
-	"strings"
-	"sync"
-	"syscall"
-	"time"
-
-	"github.com/ollama/ollama/envconfig"
-)
-
-var (
-	lock        sync.Mutex
-	payloadsDir = ""
-)
-
-func PayloadsDir() (string, error) {
-	lock.Lock()
-	defer lock.Unlock()
-	var err error
-	if payloadsDir == "" {
-		runnersDir := envconfig.RunnersDir
-
-		if runnersDir != "" {
-			payloadsDir = runnersDir
-			return payloadsDir, nil
-		}
-
-		// The remainder only applies on non-windows where we still carry payloads in the main executable
-		cleanupTmpDirs()
-		tmpDir := envconfig.TmpDir
-		if tmpDir == "" {
-			tmpDir, err = os.MkdirTemp("", "ollama")
-			if err != nil {
-				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
-			}
-		} else {
-			err = os.MkdirAll(tmpDir, 0755)
-			if err != nil {
-				return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
-			}
-		}
-
-		// Track our pid so we can clean up orphaned tmpdirs
-		pidFilePath := filepath.Join(tmpDir, "ollama.pid")
-		pidFile, err := os.OpenFile(pidFilePath, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.ModePerm)
-		if err != nil {
-			return "", err
-		}
-		if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
-			return "", err
-		}
-
-		// We create a distinct subdirectory for payloads within the tmpdir
-		// This will typically look like /tmp/ollama3208993108/runners on linux
-		payloadsDir = filepath.Join(tmpDir, "runners")
-	}
-	return payloadsDir, nil
-}
-
-// Best effort to clean up prior tmpdirs
-func cleanupTmpDirs() {
-	dirs, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*"))
-	if err != nil {
-		return
-	}
-	for _, d := range dirs {
-		info, err := os.Stat(d)
-		if err != nil || !info.IsDir() {
-			continue
-		}
-		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
-		if err != nil {
-			slog.Warn("failed to read ollama.pid", "path", d, "error", err)
-			// No pid, ignore this tmpdir
-			continue
-		}
-
-		pid, err := strconv.Atoi(string(raw))
-		if err != nil {
-			slog.Warn("failed to parse pid", "path", d, "error", err)
-			continue
-		}
-
-		proc, err := os.FindProcess(pid)
-		if err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
-			slog.Warn("found running ollama", "pid", pid, "path", d)
-			// Another running ollama, ignore this tmpdir
-			continue
-		}
-
-		if err := os.Remove(d); err != nil {
-			slog.Warn("unable to cleanup stale tmpdir", "path", d, "error", err)
-		}
-	}
-}
-
-func Cleanup() {
-	lock.Lock()
-	defer lock.Unlock()
-	runnersDir := envconfig.RunnersDir
-	if payloadsDir != "" && runnersDir == "" && runtime.GOOS != "windows" {
-		// We want to fully clean up the tmpdir parent of the payloads dir
-		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
-		slog.Debug("cleaning up", "dir", tmpDir)
-		err := os.RemoveAll(tmpDir)
-		if err != nil {
-			// On windows, if we remove too quickly the llama.dll may still be in-use and fail to remove
-			time.Sleep(1000 * time.Millisecond)
-			err = os.RemoveAll(tmpDir)
-			if err != nil {
-				slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
-			}
-		}
-	}
-}
-
-func UpdatePath(dir string) {
-	if runtime.GOOS == "windows" {
-		tmpDir := filepath.Dir(dir)
-		pathComponents := strings.Split(os.Getenv("PATH"), ";")
-		i := 0
-		for _, comp := range pathComponents {
-			if strings.EqualFold(comp, dir) {
-				return
-			}
-			// Remove any other prior paths to our temp dir
-			if !strings.HasPrefix(strings.ToLower(comp), strings.ToLower(tmpDir)) {
-				pathComponents[i] = comp
-				i++
-			}
-		}
-		newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
-		slog.Info("updating", "PATH", newPath)
-		os.Setenv("PATH", newPath)
-	}
-	// linux and darwin rely on rpath
-}
diff --git a/gpu/cpu_common.go b/gpu/cpu_common.go
index 63e88f25b..34edcdc5a 100644
--- a/gpu/cpu_common.go
+++ b/gpu/cpu_common.go
@@ -1,6 +1,11 @@
 package gpu
 
 import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+
 	"golang.org/x/sys/cpu"
 )
 
@@ -14,3 +19,19 @@ func GetCPUCapability() CPUCapability {
 	// else LCD
 	return CPUCapabilityNone
 }
+
+func IsNUMA() bool {
+	if runtime.GOOS != "linux" {
+		// numa support in llama.cpp is linux only
+		return false
+	}
+	ids := map[string]interface{}{}
+	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
+	for _, packageId := range packageIds {
+		id, err := os.ReadFile(packageId)
+		if err == nil {
+			ids[strings.TrimSpace(string(id))] = struct{}{}
+		}
+	}
+	return len(ids) > 1
+}
diff --git a/gpu/cuda_common.go b/gpu/cuda_common.go
index c90a644c4..aceec70af 100644
--- a/gpu/cuda_common.go
+++ b/gpu/cuda_common.go
@@ -4,9 +4,17 @@ package gpu
 
 import (
 	"log/slog"
+	"os"
+	"regexp"
+	"runtime"
+	"strconv"
 	"strings"
 )
 
+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
 func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	ids := []string{}
 	for _, info := range gpuInfo {
@@ -19,3 +27,38 @@ func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	}
 	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
 }
+
+func cudaVariant(gpuInfo CudaGPUInfo) string {
+	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
+		if CudaTegra != "" {
+			ver := strings.Split(CudaTegra, ".")
+			if len(ver) > 0 {
+				return "jetpack" + ver[0]
+			}
+		} else if data, err := os.ReadFile("/etc/nv_tegra_release"); err == nil {
+			r := regexp.MustCompile(` R(\d+) `)
+			m := r.FindSubmatch(data)
+			if len(m) != 2 {
+				slog.Info("Unexpected format for /etc/nv_tegra_release.  Set JETSON_JETPACK to select version")
+			} else {
+				if l4t, err := strconv.Atoi(string(m[1])); err == nil {
+					// Note: mapping from L4t -> JP is inconsistent (can't just subtract 30)
+					// https://developer.nvidia.com/embedded/jetpack-archive
+					switch l4t {
+					case 35:
+						return "jetpack5"
+					case 36:
+						return "jetpack6"
+					default:
+						slog.Info("unsupported L4T version", "nv_tegra_release", string(data))
+					}
+				}
+			}
+		}
+	}
+
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
+		return "v11"
+	}
+	return "v12"
+}
diff --git a/gpu/gpu.go b/gpu/gpu.go
index 0120d4271..69279867e 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -7,9 +7,9 @@ package gpu
 #cgo windows LDFLAGS: -lpthread
 
 #include "gpu_info.h"
-
 */
 import "C"
+
 import (
 	"fmt"
 	"log/slog"
@@ -72,13 +72,8 @@ var RocmComputeMin = 9
 // TODO find a better way to detect iGPU instead of minimum memory
 const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
 
-// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
-// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
-var CudaTegra string = os.Getenv("JETSON_JETPACK")
-
 // Note: gpuMutex must already be held
 func initCudaHandles() *cudaHandles {
-
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
 	cHandles := &cudaHandles{}
@@ -106,10 +101,9 @@ func initCudaHandles() *cudaHandles {
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
 	}
-	tmpDir, _ := PayloadsDir()
-	if tmpDir != "" {
-		// TODO - add "payloads" for subprocess
-		cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", CudartMgmtName)}
+	libDir := LibraryDir()
+	if libDir != "" {
+		cudartMgmtPatterns = []string{filepath.Join(libDir, CudartMgmtName)}
 	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)
 
@@ -236,7 +230,7 @@ func GetGPUInfo() GpuInfoList {
 	}()
 
 	if !bootstrapped {
-		slog.Debug("Detecting GPUs")
+		slog.Info("looking for compatible GPUs")
 		needRefresh = false
 		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
@@ -245,14 +239,19 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
-		cpus = []CPUInfo{CPUInfo{
-			GpuInfo: GpuInfo{
-				memInfo: mem,
-				Library: "cpu",
-				Variant: cpuCapability,
-				ID:      "0",
+		depPath := LibraryDir()
+
+		cpus = []CPUInfo{
+			{
+				GpuInfo: GpuInfo{
+					memInfo:        mem,
+					Library:        "cpu",
+					Variant:        cpuCapability.String(),
+					ID:             "0",
+					DependencyPath: depPath,
+				},
 			},
-		}}
+		}
 
 		// Fallback to CPU mode if we're lacking required vector extensions on x86
 		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
@@ -262,12 +261,6 @@ func GetGPUInfo() GpuInfoList {
 			return GpuInfoList{cpus[0].GpuInfo}
 		}
 
-		// On windows we bundle the nvidia library one level above the runner dir
-		depPath := ""
-		if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-			depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir), "cuda")
-		}
-
 		// Load ALL libraries
 		cHandles = initCudaHandles()
 
@@ -302,11 +295,45 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.FreeMemory = uint64(memInfo.free)
 				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 				gpuInfo.Compute = fmt.Sprintf("%d.%d", memInfo.major, memInfo.minor)
+				gpuInfo.computeMajor = int(memInfo.major)
+				gpuInfo.computeMinor = int(memInfo.minor)
 				gpuInfo.MinimumMemory = cudaMinimumMemory
-				gpuInfo.DependencyPath = depPath
-				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
+				variant := cudaVariant(gpuInfo)
+				if depPath != "" {
+					gpuInfo.DependencyPath = depPath
+					// Check for variant specific directory
+					if variant != "" {
+						if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
+							gpuInfo.DependencyPath = filepath.Join(depPath, "cuda_"+variant)
+						}
+					}
+				}
+				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+				gpuInfo.Variant = variant
+
+				// query the management library as well so we can record any skew between the two
+				// which represents overhead on the GPU we must set aside on subsequent updates
+				if cHandles.nvml != nil {
+					C.nvml_get_free(*cHandles.nvml, C.int(gpuInfo.index), &memInfo.free, &memInfo.total, &memInfo.used)
+					if memInfo.err != nil {
+						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+						C.free(unsafe.Pointer(memInfo.err))
+					} else {
+						if memInfo.free != 0 && uint64(memInfo.free) > gpuInfo.FreeMemory {
+							gpuInfo.OSOverhead = uint64(memInfo.free) - gpuInfo.FreeMemory
+							slog.Info("detected OS VRAM overhead",
+								"id", gpuInfo.ID,
+								"library", gpuInfo.Library,
+								"compute", gpuInfo.Compute,
+								"driver", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor),
+								"name", gpuInfo.Name,
+								"overhead", format.HumanBytes2(gpuInfo.OSOverhead),
+							)
+						}
+					}
+				}
 
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
@@ -314,40 +341,36 @@ func GetGPUInfo() GpuInfoList {
 		}
 
 		// Intel
-		if envconfig.IntelGpu {
+		if envconfig.IntelGPU() {
 			oHandles = initOneAPIHandles()
-			// On windows we bundle the oneapi library one level above the runner dir
-			depPath = ""
-			if runtime.GOOS == "windows" && envconfig.RunnersDir != "" {
-				depPath = filepath.Join(filepath.Dir(envconfig.RunnersDir), "oneapi")
-			}
-
-			for d := range oHandles.oneapi.num_drivers {
-				if oHandles.oneapi == nil {
-					// shouldn't happen
-					slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
-					continue
-				}
-				devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
-				for i := range devCount {
-					gpuInfo := OneapiGPUInfo{
-						GpuInfo: GpuInfo{
-							Library: "oneapi",
-						},
-						driverIndex: int(d),
-						gpuIndex:    int(i),
+			if oHandles != nil && oHandles.oneapi != nil {
+				for d := range oHandles.oneapi.num_drivers {
+					if oHandles.oneapi == nil {
+						// shouldn't happen
+						slog.Warn("nil oneapi handle with driver count", "count", int(oHandles.oneapi.num_drivers))
+						continue
+					}
+					devCount := C.oneapi_get_device_count(*oHandles.oneapi, C.int(d))
+					for i := range devCount {
+						gpuInfo := OneapiGPUInfo{
+							GpuInfo: GpuInfo{
+								Library: "oneapi",
+							},
+							driverIndex: int(d),
+							gpuIndex:    int(i),
+						}
+						// TODO - split bootstrapping from updating free memory
+						C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
+						// TODO - convert this to MinimumMemory based on testing...
+						var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
+						memInfo.free = C.uint64_t(totalFreeMem)
+						gpuInfo.TotalMemory = uint64(memInfo.total)
+						gpuInfo.FreeMemory = uint64(memInfo.free)
+						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
+						gpuInfo.DependencyPath = depPath
+						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
-					// TODO - split bootstrapping from updating free memory
-					C.oneapi_check_vram(*oHandles.oneapi, C.int(d), i, &memInfo)
-					// TODO - convert this to MinimumMemory based on testing...
-					var totalFreeMem float64 = float64(memInfo.free) * 0.95 // work-around: leave some reserve vram for mkl lib used in ggml-sycl backend.
-					memInfo.free = C.uint64_t(totalFreeMem)
-					gpuInfo.TotalMemory = uint64(memInfo.total)
-					gpuInfo.FreeMemory = uint64(memInfo.free)
-					gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
-					gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-					gpuInfo.DependencyPath = depPath
-					oneapiGPUs = append(oneapiGPUs, gpuInfo)
 				}
 			}
 		}
@@ -387,6 +410,9 @@ func GetGPUInfo() GpuInfoList {
 
 		rocmGPUs = AMDGetGPUInfo()
 		bootstrapped = true
+		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
+			slog.Info("no compatible GPUs were discovered")
+		}
 	}
 
 	// For detected GPUs, load library if not loaded
@@ -402,14 +428,17 @@ func GetGPUInfo() GpuInfoList {
 					"before",
 					"total", format.HumanBytes2(cpus[0].TotalMemory),
 					"free", format.HumanBytes2(cpus[0].FreeMemory),
+					"free_swap", format.HumanBytes2(cpus[0].FreeSwap),
 				),
 				slog.Group(
 					"now",
 					"total", format.HumanBytes2(mem.TotalMemory),
 					"free", format.HumanBytes2(mem.FreeMemory),
+					"free_swap", format.HumanBytes2(mem.FreeSwap),
 				),
 			)
 			cpus[0].FreeMemory = mem.FreeMemory
+			cpus[0].FreeSwap = mem.FreeSwap
 		}
 
 		var memInfo C.mem_info_t
@@ -438,9 +467,14 @@ func GetGPUInfo() GpuInfoList {
 				slog.Warn("error looking up nvidia GPU memory")
 				continue
 			}
+			if cHandles.nvml != nil && gpu.OSOverhead > 0 {
+				// When using the management library update based on recorded overhead
+				memInfo.free -= C.uint64_t(gpu.OSOverhead)
+			}
 			slog.Debug("updating cuda memory data",
 				"gpu", gpu.ID,
 				"name", gpu.Name,
+				"overhead", format.HumanBytes2(gpu.OSOverhead),
 				slog.Group(
 					"before",
 					"total", format.HumanBytes2(gpu.TotalMemory),
@@ -516,10 +550,12 @@ func GetGPUInfo() GpuInfoList {
 func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
-	var patterns []string
 	gpuLibPaths := []string{}
 	slog.Debug("Searching for GPU library", "name", baseLibName)
 
+	// Start with our bundled libraries
+	patterns := []string{filepath.Join(LibraryDir(), baseLibName)}
+
 	switch runtime.GOOS {
 	case "windows":
 		ldPaths = strings.Split(os.Getenv("PATH"), ";")
@@ -528,13 +564,14 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	default:
 		return gpuLibPaths
 	}
-	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+
+	// Then with whatever we find in the PATH/LD_LIBRARY_PATH
 	for _, ldPath := range ldPaths {
 		d, err := filepath.Abs(ldPath)
 		if err != nil {
 			continue
 		}
-		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
+		patterns = append(patterns, filepath.Join(d, baseLibName))
 	}
 	patterns = append(patterns, defaultPatterns...)
 	slog.Debug("gpu library search", "globs", patterns)
@@ -600,7 +637,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
+			// Decide what log level based on the type of error message to help users understand why
+			msg := C.GoString(resp.err)
+			switch resp.cudaErr {
+			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
+				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
+			case C.CUDA_ERROR_NO_DEVICE:
+				slog.Info("no nvidia devices detected", "library", libPath)
+			case C.CUDA_ERROR_UNKNOWN:
+				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
+				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
+			default:
+				if strings.Contains(msg, "wrong ELF class") {
+					slog.Debug("skipping 32bit library", "library", libPath)
+				} else {
+					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
+				}
+			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return int(resp.num_devices), &resp.ch, libPath
@@ -671,7 +724,7 @@ func LoadVulkanMgmt(vulkanLibPaths []string, capLibPaths []string) (int, *C.vk_h
 }
 
 func getVerboseState() C.uint16_t {
-	if envconfig.Debug {
+	if envconfig.Debug() {
 		return C.uint16_t(1)
 	}
 	return C.uint16_t(0)
@@ -699,3 +752,31 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 		return "", ""
 	}
 }
+
+func LibraryDir() string {
+	// On Windows/linux we bundle the dependencies at the same level as the executable
+	appExe, err := os.Executable()
+	if err != nil {
+		slog.Warn("failed to lookup executable path", "error", err)
+	}
+	cwd, err := os.Getwd()
+	if err != nil {
+		slog.Warn("failed to lookup working directory", "error", err)
+	}
+	// Scan for any of our dependeices, and pick first match
+	for _, root := range []string{filepath.Dir(appExe), filepath.Join(filepath.Dir(appExe), envconfig.LibRelativeToExe()), cwd} {
+		libDep := filepath.Join("lib", "ollama")
+		if _, err := os.Stat(filepath.Join(root, libDep)); err == nil {
+			return filepath.Join(root, libDep)
+		}
+		// Developer mode, local build
+		if _, err := os.Stat(filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+		if _, err := os.Stat(filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)); err == nil {
+			return filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH, libDep)
+		}
+	}
+	slog.Warn("unable to locate gpu dependency libraries")
+	return ""
+}
diff --git a/gpu/gpu_darwin.go b/gpu/gpu_darwin.go
index f26d23c12..417b48dfa 100644
--- a/gpu/gpu_darwin.go
+++ b/gpu/gpu_darwin.go
@@ -8,6 +8,7 @@ package gpu
 #include "gpu_info_darwin.h"
 */
 import "C"
+
 import (
 	"runtime"
 
@@ -24,7 +25,7 @@ func GetGPUInfo() GpuInfoList {
 		return []GpuInfo{
 			{
 				Library: "cpu",
-				Variant: GetCPUCapability(),
+				Variant: GetCPUCapability().String(),
 				memInfo: mem,
 			},
 		}
@@ -47,7 +48,7 @@ func GetCPUInfo() GpuInfoList {
 	return []GpuInfo{
 		{
 			Library: "cpu",
-			Variant: GetCPUCapability(),
+			Variant: GetCPUCapability().String(),
 			memInfo: mem,
 		},
 	}
@@ -56,7 +57,8 @@ func GetCPUInfo() GpuInfoList {
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
-		FreeMemory:  0,
+		FreeMemory:  uint64(C.getFreeMemory()),
+		// FreeSwap omitted as Darwin uses dynamic paging
 	}, nil
 }
 
diff --git a/gpu/gpu_info.h b/gpu/gpu_info.h
index a0b4c96ee..d36aff253 100644
--- a/gpu/gpu_info.h
+++ b/gpu/gpu_info.h
@@ -68,4 +68,4 @@ void cpu_check_ram(mem_info_t *resp);
 #include "gpu_info_vulkan.h"
 
 #endif  // __GPU_INFO_H__
-#endif  // __APPLE__
\ No newline at end of file
+#endif  // __APPLE__
diff --git a/gpu/gpu_info_darwin.h b/gpu/gpu_info_darwin.h
index 3edca237c..415e7922d 100644
--- a/gpu/gpu_info_darwin.h
+++ b/gpu/gpu_info_darwin.h
@@ -2,3 +2,4 @@
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
 uint64_t getPhysicalMemory();
+uint64_t getFreeMemory();
diff --git a/gpu/gpu_info_darwin.m b/gpu/gpu_info_darwin.m
index a145ac076..5ca139e0b 100644
--- a/gpu/gpu_info_darwin.m
+++ b/gpu/gpu_info_darwin.m
@@ -1,4 +1,5 @@
-// go:build darwin
+#import <Foundation/Foundation.h>
+#import <mach/mach.h>
 #include "gpu_info_darwin.h"
 
 uint64_t getRecommendedMaxVRAM() {
@@ -8,6 +9,27 @@ uint64_t getRecommendedMaxVRAM() {
   return result;
 }
 
+// getPhysicalMemory returns the total physical memory in bytes
 uint64_t getPhysicalMemory() {
-  return [[NSProcessInfo processInfo] physicalMemory];
+  return [NSProcessInfo processInfo].physicalMemory;
+}
+
+// getFreeMemory returns the total free memory in bytes, including inactive
+// memory that can be reclaimed by the system.
+uint64_t getFreeMemory() {
+  mach_port_t host_port = mach_host_self();
+  mach_msg_type_number_t host_size = sizeof(vm_statistics64_data_t) / sizeof(integer_t);
+  vm_size_t pagesize;
+  vm_statistics64_data_t vm_stat;
+
+  host_page_size(host_port, &pagesize);
+  if (host_statistics64(host_port, HOST_VM_INFO64, (host_info64_t)&vm_stat, &host_size) != KERN_SUCCESS) {
+    return 0;
+  }
+
+  uint64_t free_memory = (uint64_t)vm_stat.free_count * pagesize;
+  free_memory += (uint64_t)vm_stat.speculative_count * pagesize;
+  free_memory += (uint64_t)vm_stat.inactive_count * pagesize;
+
+  return free_memory;
 }
diff --git a/gpu/gpu_info_nvcuda.c b/gpu/gpu_info_nvcuda.c
index abe140844..a1a38bfc2 100644
--- a/gpu/gpu_info_nvcuda.c
+++ b/gpu/gpu_info_nvcuda.c
@@ -7,6 +7,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
   CUresult ret;
   resp->err = NULL;
   resp->num_devices = 0;
+  resp->cudaErr = CUDA_SUCCESS;
   const int buflen = 256;
   char buf[buflen + 1];
   int i;
@@ -38,6 +39,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
             nvcuda_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
+    resp->cudaErr = -1;
     return;
   }
 
@@ -52,6 +54,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
               msg);
       free(msg);
       resp->err = strdup(buf);
+      resp->cudaErr = -1;
       return;
     }
   }
@@ -61,12 +64,9 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
     UNLOAD_LIBRARY(resp->ch.handle);
     resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
-      return;
-    }
-    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
+    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 
@@ -91,6 +91,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
     resp->ch.handle = NULL;
     snprintf(buf, buflen, "unable to get device count: %d", ret);
     resp->err = strdup(buf);
+    resp->cudaErr = ret;
     return;
   }
 }
@@ -106,13 +107,13 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 
   if (h.handle == NULL) {
-    resp->err = strdup("nvcuda handle isn't initialized");
+    resp->err = strdup("cuda driver library handle isn't initialized");
     return;
   }
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device failed to initialize");
+    snprintf(buf, buflen, "cuda driver library device failed to initialize");
     resp->err = strdup(buf);
     return;
   }
@@ -168,14 +169,14 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda failed to get device context %d", ret);
+    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
     resp->err = strdup(buf);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
   if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
     resp->err = strdup(buf);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
@@ -193,7 +194,7 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
   }
 }
 
@@ -206,7 +207,7 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
 
   ret = (*h.cuDeviceGet)(&device, i);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device failed to initialize");
+    LOG(1, "cuda driver library device failed to initialize");
     return;
   }
 
@@ -214,13 +215,13 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
   // To get memory we have to set (and release) a context
   ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to get device context %d", ret);
+    LOG(1, "cuda driver library failed to get device context %d", ret);
     return;
   }
 
   ret = (*h.cuMemGetInfo_v2)(free, total);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device memory info lookup failure %d", ret);
+    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
     // Best effort on failure...
     (*h.cuCtxDestroy)(ctx);
     return;
@@ -228,12 +229,12 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
 
   ret = (*h.cuCtxDestroy)(ctx);
   if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
   }
 }
 
 void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing nvcuda library\n");
+  LOG(h.verbose, "releasing cuda driver library\n");
   UNLOAD_LIBRARY(h.handle);
   // TODO and other context release logic?
   h.handle = NULL;
diff --git a/gpu/gpu_info_nvcuda.h b/gpu/gpu_info_nvcuda.h
index f9654f641..ef2fe8a30 100644
--- a/gpu/gpu_info_nvcuda.h
+++ b/gpu/gpu_info_nvcuda.h
@@ -7,9 +7,12 @@
 typedef enum cudaError_enum {
   CUDA_SUCCESS = 0,
   CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_OUT_OF_MEMORY = 2,
   CUDA_ERROR_NOT_INITIALIZED = 3,
   CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  CUDA_ERROR_NO_DEVICE = 100,
+  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
+  CUDA_ERROR_UNKNOWN = 999,
   // Other values omitted for now...
 } CUresult;
 
@@ -64,6 +67,7 @@ typedef struct nvcuda_init_resp {
   char *err;  // If err is non-null handle is invalid
   nvcuda_handle_t ch;
   int num_devices;
+  CUresult cudaErr;
 } nvcuda_init_resp_t;
 
 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
diff --git a/gpu/gpu_linux.go b/gpu/gpu_linux.go
index 2e723c4da..76df63268 100644
--- a/gpu/gpu_linux.go
+++ b/gpu/gpu_linux.go
@@ -43,6 +43,15 @@ var OneapiGlobs = []string{
 	"/usr/lib*/libze_intel_gpu.so*",
 }
 
+var (
+	CudartMgmtName = "libcudart.so*"
+	NvcudaMgmtName = "libcuda.so*"
+	NvmlMgmtName   = "" // not currently wired on linux
+	OneapiMgmtName = "libze_intel_gpu.so*"
+	VulkanMgmtName = "libvulkan.so*"
+	libcapMgmtName = "libcap.so*"
+)
+
 var VulkanGlobs = []string{
 	"/usr/lib/x86_64-linux-gnu/libvulkan.so*",
 	"/usr/lib*/libvulkan.so*",
@@ -53,20 +62,13 @@ var capLinuxGlobs = []string{
 	"/usr/lib*/libcap.so*",
 }
 
-var CudartMgmtName = "libcudart.so*"
-var NvcudaMgmtName = "libcuda.so*"
-var NvmlMgmtName = "" // not currently wired on linux
-var OneapiMgmtName = "libze_intel_gpu.so"
-var VulkanMgmtName = "libvulkan.so*"
-var libcapMgmtName = "libcap.so*"
-
 func FindLibCapLibs() []string {
 	return FindGPULibs(libcapMgmtName, capLinuxGlobs)
 }
 
 func GetCPUMem() (memInfo, error) {
 	var mem memInfo
-	var total, available, free, buffers, cached uint64
+	var total, available, free, buffers, cached, freeSwap uint64
 	f, err := os.Open("/proc/meminfo")
 	if err != nil {
 		return mem, err
@@ -86,20 +88,21 @@ func GetCPUMem() (memInfo, error) {
 			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
 		case strings.HasPrefix(line, "Cached:"):
 			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
+		case strings.HasPrefix(line, "SwapFree:"):
+			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
 		default:
 			continue
 		}
 		if err != nil {
 			return mem, err
 		}
-
-		if total > 0 && available > 0 {
-			mem.TotalMemory = total * format.KibiByte
-			mem.FreeMemory = available * format.KibiByte
-			return mem, nil
-		}
 	}
 	mem.TotalMemory = total * format.KibiByte
-	mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	mem.FreeSwap = freeSwap * format.KibiByte
+	if available > 0 {
+		mem.FreeMemory = available * format.KibiByte
+	} else {
+		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	}
 	return mem, nil
 }
diff --git a/gpu/gpu_test.go b/gpu/gpu_test.go
index 46d3201e1..13a3f5442 100644
--- a/gpu/gpu_test.go
+++ b/gpu/gpu_test.go
@@ -32,4 +32,29 @@ func TestCPUMemInfo(t *testing.T) {
 	}
 }
 
+func TestByLibrary(t *testing.T) {
+	type testCase struct {
+		input  []GpuInfo
+		expect int
+	}
+
+	testCases := map[string]*testCase{
+		"empty":                    {input: []GpuInfo{}, expect: 0},
+		"cpu":                      {input: []GpuInfo{{Library: "cpu"}}, expect: 1},
+		"cpu + GPU":                {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU no variant":   {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda"}, {Library: "cuda"}}, expect: 2},
+		"cpu + 2 GPU same variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v11"}}, expect: 2},
+		"cpu + 2 GPU diff variant": {input: []GpuInfo{{Library: "cpu"}, {Library: "cuda", Variant: "v11"}, {Library: "cuda", Variant: "v12"}}, expect: 3},
+	}
+
+	for k, v := range testCases {
+		t.Run(k, func(t *testing.T) {
+			resp := (GpuInfoList)(v.input).ByLibrary()
+			if len(resp) != v.expect {
+				t.Fatalf("expected length %d, got %d => %+v", v.expect, len(resp), resp)
+			}
+		})
+	}
+}
+
 // TODO - add some logic to figure out card type through other means and actually verify we got back what we expected
diff --git a/gpu/gpu_windows.go b/gpu/gpu_windows.go
index 328477440..5491da963 100644
--- a/gpu/gpu_windows.go
+++ b/gpu/gpu_windows.go
@@ -40,10 +40,12 @@ var OneapiGlobs = []string{
 	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
 }
 
-var CudartMgmtName = "cudart64_*.dll"
-var NvcudaMgmtName = "nvcuda.dll"
-var NvmlMgmtName = "nvml.dll"
-var OneapiMgmtName = "ze_intel_gpu64.dll"
+var (
+	CudartMgmtName = "cudart64_*.dll"
+	NvcudaMgmtName = "nvcuda.dll"
+	NvmlMgmtName   = "nvml.dll"
+	OneapiMgmtName = "ze_intel_gpu64.dll"
+)
 
 func FindLibCapLibs() []string {
 	return []string{"<unused>"}
@@ -55,5 +57,5 @@ func GetCPUMem() (memInfo, error) {
 	if r1 == 0 {
 		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
 	}
-	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys}, nil
+	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
 }
diff --git a/gpu/types.go b/gpu/types.go
index ae7fb853c..e623c3766 100644
--- a/gpu/types.go
+++ b/gpu/types.go
@@ -10,6 +10,7 @@ import (
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	FreeSwap    uint64 `json:"free_swap,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -18,7 +19,7 @@ type GpuInfo struct {
 	Library string `json:"library,omitempty"`
 
 	// Optional variant to select (e.g. versions, cpu feature flags)
-	Variant CPUCapability `json:"variant"`
+	Variant string `json:"variant"`
 
 	// MinimumMemory represents the minimum memory required to use the GPU
 	MinimumMemory uint64 `json:"-"`
@@ -52,7 +53,10 @@ type CPUInfo struct {
 
 type CudaGPUInfo struct {
 	GpuInfo
-	index int //nolint:unused,nolintlint
+	OSOverhead   uint64 // Memory overhead between the driver library and management library
+	index        int    //nolint:unused,nolintlint
+	computeMajor int    //nolint:unused,nolintlint
+	computeMinor int    //nolint:unused,nolintlint
 }
 type CudaGPUInfoList []CudaGPUInfo
 
@@ -86,8 +90,8 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	for _, info := range l {
 		found := false
 		requested := info.Library
-		if info.Variant != CPUCapabilityNone {
-			requested += "_" + info.Variant.String()
+		if info.Variant != CPUCapabilityNone.String() {
+			requested += "_" + info.Variant
 		}
 		for i, lib := range libs {
 			if lib == requested {
@@ -97,7 +101,7 @@ func (l GpuInfoList) ByLibrary() []GpuInfoList {
 			}
 		}
 		if !found {
-			libs = append(libs, info.Library)
+			libs = append(libs, requested)
 			resp = append(resp, []GpuInfo{info})
 		}
 	}
@@ -110,6 +114,7 @@ func (l GpuInfoList) LogDetails() {
 		slog.Info("inference compute",
 			"id", g.ID,
 			"library", g.Library,
+			"variant", g.Variant,
 			"compute", g.Compute,
 			"driver", fmt.Sprintf("%d.%d", g.DriverMajor, g.DriverMinor),
 			"name", g.Name,
diff --git a/integration/basic_test.go b/integration/basic_test.go
index 6e632a1ce..8e35b5c5b 100644
--- a/integration/basic_test.go
+++ b/integration/basic_test.go
@@ -45,14 +45,7 @@ func TestUnicodeModelDir(t *testing.T) {
 	defer os.RemoveAll(modelDir)
 	slog.Info("unicode", "OLLAMA_MODELS", modelDir)
 
-	oldModelsDir := os.Getenv("OLLAMA_MODELS")
-	if oldModelsDir == "" {
-		defer os.Unsetenv("OLLAMA_MODELS")
-	} else {
-		defer os.Setenv("OLLAMA_MODELS", oldModelsDir)
-	}
-	err = os.Setenv("OLLAMA_MODELS", modelDir)
-	require.NoError(t, err)
+	t.Setenv("OLLAMA_MODELS", modelDir)
 
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
diff --git a/integration/concurrency_test.go b/integration/concurrency_test.go
index d66ba9f00..42e9d0749 100644
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@@ -11,8 +11,10 @@ import (
 	"testing"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
 )
 
 func TestMultiModelConcurrency(t *testing.T) {
@@ -39,8 +41,8 @@ func TestMultiModelConcurrency(t *testing.T) {
 			},
 		}
 		resp = [2][]string{
-			[]string{"sunlight"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"sunlight"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
 		}
 	)
 	var wg sync.WaitGroup
@@ -69,12 +71,11 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	reqLimit := len(req)
 	iterLimit := 5
 
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
-	if vram != "" {
-		max, err := strconv.ParseUint(vram, 10, 64)
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
 		require.NoError(t, err)
 		// Don't hammer on small VRAM cards...
-		if max < 4*1024*1024*1024 {
+		if maxVram < 4*format.GibiByte {
 			reqLimit = min(reqLimit, 2)
 			iterLimit = 2
 		}
@@ -106,13 +107,16 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 // Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
-	if vram == "" {
+	s := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
+	if s == "" {
 		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}
-	max, err := strconv.ParseUint(vram, 10, 64)
-	require.NoError(t, err)
-	const MB = uint64(1024 * 1024)
+
+	maxVram, err := strconv.ParseUint(s, 10, 64)
+	if err != nil {
+		t.Fatal(err)
+	}
+
 	type model struct {
 		name string
 		size uint64 // Approximate amount of VRAM they typically use when fully loaded in VRAM
@@ -121,83 +125,82 @@ func TestMultiModelStress(t *testing.T) {
 	smallModels := []model{
 		{
 			name: "orca-mini",
-			size: 2992 * MB,
+			size: 2992 * format.MebiByte,
 		},
 		{
 			name: "phi",
-			size: 2616 * MB,
+			size: 2616 * format.MebiByte,
 		},
 		{
 			name: "gemma:2b",
-			size: 2364 * MB,
+			size: 2364 * format.MebiByte,
 		},
 		{
 			name: "stable-code:3b",
-			size: 2608 * MB,
+			size: 2608 * format.MebiByte,
 		},
 		{
 			name: "starcoder2:3b",
-			size: 2166 * MB,
+			size: 2166 * format.MebiByte,
 		},
 	}
 	mediumModels := []model{
 		{
 			name: "llama2",
-			size: 5118 * MB,
+			size: 5118 * format.MebiByte,
 		},
 		{
 			name: "mistral",
-			size: 4620 * MB,
+			size: 4620 * format.MebiByte,
 		},
 		{
 			name: "orca-mini:7b",
-			size: 5118 * MB,
+			size: 5118 * format.MebiByte,
 		},
 		{
 			name: "dolphin-mistral",
-			size: 4620 * MB,
+			size: 4620 * format.MebiByte,
 		},
 		{
 			name: "gemma:7b",
-			size: 5000 * MB,
+			size: 5000 * format.MebiByte,
+		},
+		{
+			name: "codellama:7b",
+			size: 5118 * format.MebiByte,
 		},
-		// TODO - uncomment this once #3565 is merged and this is rebased on it
-		// {
-		// 	name: "codellama:7b",
-		// 	size: 5118 * MB,
-		// },
 	}
 
 	// These seem to be too slow to be useful...
 	// largeModels := []model{
 	// 	{
 	// 		name: "llama2:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "codellama:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "orca-mini:13b",
-	// 		size: 7400 * MB,
+	// 		size: 7400 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "gemma:7b",
-	// 		size: 5000 * MB,
+	// 		size: 5000 * format.MebiByte,
 	// 	},
 	// 	{
 	// 		name: "starcoder2:15b",
-	// 		size: 9100 * MB,
+	// 		size: 9100 * format.MebiByte,
 	// 	},
 	// }
 
 	var chosenModels []model
 	switch {
-	case max < 10000*MB:
+	case maxVram < 10000*format.MebiByte:
 		slog.Info("selecting small models")
 		chosenModels = smallModels
-	// case max < 30000*MB:
+	// case maxVram < 30000*format.MebiByte:
 	default:
 		slog.Info("selecting medium models")
 		chosenModels = mediumModels
@@ -226,15 +229,15 @@ func TestMultiModelStress(t *testing.T) {
 	}
 
 	var wg sync.WaitGroup
-	consumed := uint64(256 * MB) // Assume some baseline usage
+	consumed := uint64(256 * format.MebiByte) // Assume some baseline usage
 	for i := 0; i < len(req); i++ {
 		// Always get at least 2 models, but dont' overshoot VRAM too much or we'll take too long
-		if i > 1 && consumed > max {
-			slog.Info("achieved target vram exhaustion", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+		if i > 1 && consumed > maxVram {
+			slog.Info("achieved target vram exhaustion", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 			break
 		}
 		consumed += chosenModels[i].size
-		slog.Info("target vram", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+		slog.Info("target vram", "count", i, "vram", format.HumanBytes2(maxVram), "models", format.HumanBytes2(consumed))
 
 		wg.Add(1)
 		go func(i int) {
diff --git a/integration/context_test.go b/integration/context_test.go
index 46fac5ead..f1342e16c 100644
--- a/integration/context_test.go
+++ b/integration/context_test.go
@@ -12,7 +12,7 @@ import (
 
 func TestContextExhaustion(t *testing.T) {
 	// Longer needed for small footprint GPUs
-	ctx, cancel := context.WithTimeout(context.Background(), 6*time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
@@ -25,5 +25,10 @@ func TestContextExhaustion(t *testing.T) {
 			"num_ctx":     128,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("PullIfMissing failed: %v", err)
+	}
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }
diff --git a/integration/embed_test.go b/integration/embed_test.go
new file mode 100644
index 000000000..4a68af68a
--- /dev/null
+++ b/integration/embed_test.go
@@ -0,0 +1,209 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"math"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func floatsEqual32(a, b float32) bool {
+	return math.Abs(float64(a-b)) <= 1e-4
+}
+
+func floatsEqual64(a, b float64) bool {
+	return math.Abs(a-b) <= 1e-4
+}
+
+func TestAllMiniLMEmbeddings(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbeddingRequest{
+		Model:  "all-minilm",
+		Prompt: "why is the sky blue?",
+	}
+
+	res, err := embeddingTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embedding) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embedding))
+	}
+
+	if !floatsEqual64(res.Embedding[0], 0.06642947345972061) {
+		t.Fatalf("expected 0.06642947345972061, got %.16f", res.Embedding[0])
+	}
+}
+
+func TestAllMiniLMEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: "why is the sky blue?",
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 1 {
+		t.Fatalf("expected 1 embedding, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) {
+		t.Fatalf("expected 0.010071031, got %.8f", res.Embeddings[0][0])
+	}
+
+	if res.PromptEvalCount != 6 {
+		t.Fatalf("expected 6 prompt tokens, got %d", res.PromptEvalCount)
+	}
+}
+
+func TestAllMiniLMBatchEmbed(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.EmbedRequest{
+		Model: "all-minilm",
+		Input: []string{"why is the sky blue?", "why is the grass green?"},
+	}
+
+	res, err := embedTestHelper(ctx, t, req)
+
+	if err != nil {
+		t.Fatalf("error: %v", err)
+	}
+
+	if len(res.Embeddings) != 2 {
+		t.Fatalf("expected 2 embeddings, got %d", len(res.Embeddings))
+	}
+
+	if len(res.Embeddings[0]) != 384 {
+		t.Fatalf("expected 384 floats, got %d", len(res.Embeddings[0]))
+	}
+
+	if !floatsEqual32(res.Embeddings[0][0], 0.010071031) || !floatsEqual32(res.Embeddings[1][0], -0.009802706) {
+		t.Fatalf("expected 0.010071031 and -0.009802706, got %.8f and %.8f", res.Embeddings[0][0], res.Embeddings[1][0])
+	}
+
+	if res.PromptEvalCount != 12 {
+		t.Fatalf("expected 12 prompt tokens, got %d", res.PromptEvalCount)
+	}
+}
+
+func TestAllMiniLMEmbedTruncate(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	truncTrue, truncFalse := true, false
+
+	type testReq struct {
+		Name    string
+		Request api.EmbedRequest
+	}
+
+	reqs := []testReq{
+		{
+			Name: "Target Truncation",
+			Request: api.EmbedRequest{
+				Model: "all-minilm",
+				Input: "why",
+			},
+		},
+		{
+			Name: "Default Truncate",
+			Request: api.EmbedRequest{
+				Model:   "all-minilm",
+				Input:   "why is the sky blue?",
+				Options: map[string]any{"num_ctx": 1},
+			},
+		},
+		{
+			Name: "Explicit Truncate",
+			Request: api.EmbedRequest{
+				Model:    "all-minilm",
+				Input:    "why is the sky blue?",
+				Truncate: &truncTrue,
+				Options:  map[string]any{"num_ctx": 1},
+			},
+		},
+	}
+
+	res := make(map[string]*api.EmbedResponse)
+
+	for _, req := range reqs {
+		response, err := embedTestHelper(ctx, t, req.Request)
+		if err != nil {
+			t.Fatalf("error: %v", err)
+		}
+		res[req.Name] = response
+	}
+
+	if res["Target Truncation"].Embeddings[0][0] != res["Default Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request to truncate correctly")
+	}
+
+	if res["Default Truncate"].Embeddings[0][0] != res["Explicit Truncate"].Embeddings[0][0] {
+		t.Fatal("expected default request and truncate true request to be the same")
+	}
+
+	// check that truncate set to false returns an error if context length is exceeded
+	_, err := embedTestHelper(ctx, t, api.EmbedRequest{
+		Model:    "all-minilm",
+		Input:    "why is the sky blue?",
+		Truncate: &truncFalse,
+		Options:  map[string]any{"num_ctx": 1},
+	})
+
+	if err == nil {
+		t.Fatal("expected error, got nil")
+	}
+}
+
+func embeddingTestHelper(ctx context.Context, t *testing.T, req api.EmbeddingRequest) (*api.EmbeddingResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embeddings(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
+
+func embedTestHelper(ctx context.Context, t *testing.T, req api.EmbedRequest) (*api.EmbedResponse, error) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("failed to pull model %s: %v", req.Model, err)
+	}
+
+	response, err := client.Embed(ctx, &req)
+
+	if err != nil {
+		return nil, err
+	}
+
+	return response, nil
+}
diff --git a/integration/llm_test.go b/integration/llm_test.go
index 4952b0726..398e0a03a 100644
--- a/integration/llm_test.go
+++ b/integration/llm_test.go
@@ -35,8 +35,8 @@ var (
 		},
 	}
 	resp = [2][]string{
-		[]string{"sunlight"},
-		[]string{"england", "english", "massachusetts", "pilgrims"},
+		{"sunlight"},
+		{"england", "english", "massachusetts", "pilgrims"},
 	}
 )
 
diff --git a/integration/max_queue_test.go b/integration/max_queue_test.go
index dfa5eae0c..ec9e085a5 100644
--- a/integration/max_queue_test.go
+++ b/integration/max_queue_test.go
@@ -5,7 +5,6 @@ package integration
 import (
 	"context"
 	"errors"
-	"fmt"
 	"log/slog"
 	"os"
 	"strconv"
@@ -14,8 +13,10 @@ import (
 	"testing"
 	"time"
 
-	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 )
 
 func TestMaxQueue(t *testing.T) {
@@ -27,13 +28,10 @@ func TestMaxQueue(t *testing.T) {
 	// Note: This test can be quite slow when running in CPU mode, so keep the threadCount low unless your on GPU
 	// Also note that by default Darwin can't sustain > ~128 connections without adjusting limits
 	threadCount := 32
-	mq := os.Getenv("OLLAMA_MAX_QUEUE")
-	if mq != "" {
-		var err error
-		threadCount, err = strconv.Atoi(mq)
-		require.NoError(t, err)
+	if maxQueue := envconfig.MaxQueue(); maxQueue != 0 {
+		threadCount = int(maxQueue)
 	} else {
-		os.Setenv("OLLAMA_MAX_QUEUE", fmt.Sprintf("%d", threadCount))
+		t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
 	}
 
 	req := api.GenerateRequest{
diff --git a/integration/utils_test.go b/integration/utils_test.go
index 7e1fcc10e..a60109958 100644
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@@ -162,7 +162,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 	fn := func(resp api.ProgressResponse) error {
 		// fmt.Print(".")
 		if !stallTimer.Reset(stallDuration) {
-			return fmt.Errorf("stall was detected, aborting status reporting")
+			return errors.New("stall was detected, aborting status reporting")
 		}
 		return nil
 	}
@@ -180,7 +180,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	select {
 	case <-stallTimer.C:
-		return fmt.Errorf("download stalled")
+		return errors.New("download stalled")
 	case <-done:
 		return pullError
 	}
@@ -243,7 +243,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 		// fmt.Print(".")
 		buf.Write([]byte(response.Response))
 		if !stallTimer.Reset(streamTimeout) {
-			return fmt.Errorf("stall was detected while streaming response, aborting")
+			return errors.New("stall was detected while streaming response, aborting")
 		}
 		return nil
 	}
@@ -334,10 +334,10 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			},
 		},
 		[][]string{
-			[]string{"sunlight"},
-			[]string{"soil", "organic", "earth", "black", "tan"},
-			[]string{"england", "english", "massachusetts", "pilgrims", "british"},
-			[]string{"fourth", "july", "declaration", "independence"},
-			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
+			{"sunlight"},
+			{"soil", "organic", "earth", "black", "tan"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
+			{"fourth", "july", "declaration", "independence"},
+			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }
diff --git a/llm/ext_server/CMakeLists.txt b/llm/ext_server/CMakeLists.txt
index db7d52dcc..517302451 100644
--- a/llm/ext_server/CMakeLists.txt
+++ b/llm/ext_server/CMakeLists.txt
@@ -1,14 +1,15 @@
-
 set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp utils.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_SERVER_LDFLAGS})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+    target_link_options(${TARGET} PRIVATE -municode -Wl,/subsystem:console)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
\ No newline at end of file
diff --git a/llm/ext_server/json.hpp b/llm/ext_server/json.hpp
deleted file mode 100644
index ea945f346..000000000
--- a/llm/ext_server/json.hpp
+++ /dev/null
@@ -1,24596 +0,0 @@
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-/****************************************************************************\
- * Note on documentation: The source files contain links to the online      *
- * documentation of the public API at https://json.nlohmann.me. This URL    *
- * contains the most recent documentation and should also be applicable to  *
- * previous versions; documentation for deprecated functions is not         *
- * removed, but marked deprecated. See "Generate documentation" section in  *
- * file docs/README.md.                                                     *
-\****************************************************************************/
-
-#ifndef INCLUDE_NLOHMANN_JSON_HPP_
-#define INCLUDE_NLOHMANN_JSON_HPP_
-
-#include <algorithm> // all_of, find, for_each
-#include <cstddef> // nullptr_t, ptrdiff_t, size_t
-#include <functional> // hash, less
-#include <initializer_list> // initializer_list
-#ifndef JSON_NO_IO
-    #include <iosfwd> // istream, ostream
-#endif  // JSON_NO_IO
-#include <iterator> // random_access_iterator_tag
-#include <memory> // unique_ptr
-#include <numeric> // accumulate
-#include <string> // string, stoi, to_string
-#include <utility> // declval, forward, move, pair, swap
-#include <vector> // vector
-
-// #include <nlohmann/adl_serializer.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <utility>
-
-// #include <nlohmann/detail/abi_macros.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// This file contains all macro definitions affecting or depending on the ABI
-
-#ifndef JSON_SKIP_LIBRARY_VERSION_CHECK
-    #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH)
-        #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 11 || NLOHMANN_JSON_VERSION_PATCH != 2
-            #warning "Already included a different version of the library!"
-        #endif
-    #endif
-#endif
-
-#define NLOHMANN_JSON_VERSION_MAJOR 3   // NOLINT(modernize-macro-to-enum)
-#define NLOHMANN_JSON_VERSION_MINOR 11  // NOLINT(modernize-macro-to-enum)
-#define NLOHMANN_JSON_VERSION_PATCH 2   // NOLINT(modernize-macro-to-enum)
-
-#ifndef JSON_DIAGNOSTICS
-    #define JSON_DIAGNOSTICS 0
-#endif
-
-#ifndef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
-    #define JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON 0
-#endif
-
-#if JSON_DIAGNOSTICS
-    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS _diag
-#else
-    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS
-#endif
-
-#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
-    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON _ldvcmp
-#else
-    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON
-#endif
-
-#ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION
-    #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0
-#endif
-
-// Construct the namespace ABI tags component
-#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b) json_abi ## a ## b
-#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b) \
-    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b)
-
-#define NLOHMANN_JSON_ABI_TAGS                                       \
-    NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
-            NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS,                       \
-            NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON)
-
-// Construct the namespace version component
-#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \
-    _v ## major ## _ ## minor ## _ ## patch
-#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(major, minor, patch) \
-    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch)
-
-#if NLOHMANN_JSON_NAMESPACE_NO_VERSION
-#define NLOHMANN_JSON_NAMESPACE_VERSION
-#else
-#define NLOHMANN_JSON_NAMESPACE_VERSION                                 \
-    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(NLOHMANN_JSON_VERSION_MAJOR, \
-                                           NLOHMANN_JSON_VERSION_MINOR, \
-                                           NLOHMANN_JSON_VERSION_PATCH)
-#endif
-
-// Combine namespace components
-#define NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) a ## b
-#define NLOHMANN_JSON_NAMESPACE_CONCAT(a, b) \
-    NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b)
-
-#ifndef NLOHMANN_JSON_NAMESPACE
-#define NLOHMANN_JSON_NAMESPACE               \
-    nlohmann::NLOHMANN_JSON_NAMESPACE_CONCAT( \
-            NLOHMANN_JSON_ABI_TAGS,           \
-            NLOHMANN_JSON_NAMESPACE_VERSION)
-#endif
-
-#ifndef NLOHMANN_JSON_NAMESPACE_BEGIN
-#define NLOHMANN_JSON_NAMESPACE_BEGIN                \
-    namespace nlohmann                               \
-    {                                                \
-    inline namespace NLOHMANN_JSON_NAMESPACE_CONCAT( \
-                NLOHMANN_JSON_ABI_TAGS,              \
-                NLOHMANN_JSON_NAMESPACE_VERSION)     \
-    {
-#endif
-
-#ifndef NLOHMANN_JSON_NAMESPACE_END
-#define NLOHMANN_JSON_NAMESPACE_END                                     \
-    }  /* namespace (inline namespace) NOLINT(readability/namespace) */ \
-    }  // namespace nlohmann
-#endif
-
-// #include <nlohmann/detail/conversions/from_json.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // transform
-#include <array> // array
-#include <forward_list> // forward_list
-#include <iterator> // inserter, front_inserter, end
-#include <map> // map
-#include <string> // string
-#include <tuple> // tuple, make_tuple
-#include <type_traits> // is_arithmetic, is_same, is_enum, underlying_type, is_convertible
-#include <unordered_map> // unordered_map
-#include <utility> // pair, declval
-#include <valarray> // valarray
-
-// #include <nlohmann/detail/exceptions.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef> // nullptr_t
-#include <exception> // exception
-#include <stdexcept> // runtime_error
-#include <string> // to_string
-#include <vector> // vector
-
-// #include <nlohmann/detail/value_t.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <array> // array
-#include <cstddef> // size_t
-#include <cstdint> // uint8_t
-#include <string> // string
-
-// #include <nlohmann/detail/macro_scope.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <utility> // declval, pair
-// #include <nlohmann/detail/meta/detected.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <type_traits>
-
-// #include <nlohmann/detail/meta/void_t.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename ...Ts> struct make_void
-{
-    using type = void;
-};
-template<typename ...Ts> using void_t = typename make_void<Ts...>::type;
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-// https://en.cppreference.com/w/cpp/experimental/is_detected
-struct nonesuch
-{
-    nonesuch() = delete;
-    ~nonesuch() = delete;
-    nonesuch(nonesuch const&) = delete;
-    nonesuch(nonesuch const&&) = delete;
-    void operator=(nonesuch const&) = delete;
-    void operator=(nonesuch&&) = delete;
-};
-
-template<class Default,
-         class AlwaysVoid,
-         template<class...> class Op,
-         class... Args>
-struct detector
-{
-    using value_t = std::false_type;
-    using type = Default;
-};
-
-template<class Default, template<class...> class Op, class... Args>
-struct detector<Default, void_t<Op<Args...>>, Op, Args...>
-{
-    using value_t = std::true_type;
-    using type = Op<Args...>;
-};
-
-template<template<class...> class Op, class... Args>
-using is_detected = typename detector<nonesuch, void, Op, Args...>::value_t;
-
-template<template<class...> class Op, class... Args>
-struct is_detected_lazy : is_detected<Op, Args...> { };
-
-template<template<class...> class Op, class... Args>
-using detected_t = typename detector<nonesuch, void, Op, Args...>::type;
-
-template<class Default, template<class...> class Op, class... Args>
-using detected_or = detector<Default, void, Op, Args...>;
-
-template<class Default, template<class...> class Op, class... Args>
-using detected_or_t = typename detected_or<Default, Op, Args...>::type;
-
-template<class Expected, template<class...> class Op, class... Args>
-using is_detected_exact = std::is_same<Expected, detected_t<Op, Args...>>;
-
-template<class To, template<class...> class Op, class... Args>
-using is_detected_convertible =
-    std::is_convertible<detected_t<Op, Args...>, To>;
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/thirdparty/hedley/hedley.hpp>
-
-
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-FileCopyrightText: 2016-2021 Evan Nemerson <evan@nemerson.com>
-// SPDX-License-Identifier: MIT
-
-/* Hedley - https://nemequ.github.io/hedley
- * Created by Evan Nemerson <evan@nemerson.com>
- */
-
-#if !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < 15)
-#if defined(JSON_HEDLEY_VERSION)
-    #undef JSON_HEDLEY_VERSION
-#endif
-#define JSON_HEDLEY_VERSION 15
-
-#if defined(JSON_HEDLEY_STRINGIFY_EX)
-    #undef JSON_HEDLEY_STRINGIFY_EX
-#endif
-#define JSON_HEDLEY_STRINGIFY_EX(x) #x
-
-#if defined(JSON_HEDLEY_STRINGIFY)
-    #undef JSON_HEDLEY_STRINGIFY
-#endif
-#define JSON_HEDLEY_STRINGIFY(x) JSON_HEDLEY_STRINGIFY_EX(x)
-
-#if defined(JSON_HEDLEY_CONCAT_EX)
-    #undef JSON_HEDLEY_CONCAT_EX
-#endif
-#define JSON_HEDLEY_CONCAT_EX(a,b) a##b
-
-#if defined(JSON_HEDLEY_CONCAT)
-    #undef JSON_HEDLEY_CONCAT
-#endif
-#define JSON_HEDLEY_CONCAT(a,b) JSON_HEDLEY_CONCAT_EX(a,b)
-
-#if defined(JSON_HEDLEY_CONCAT3_EX)
-    #undef JSON_HEDLEY_CONCAT3_EX
-#endif
-#define JSON_HEDLEY_CONCAT3_EX(a,b,c) a##b##c
-
-#if defined(JSON_HEDLEY_CONCAT3)
-    #undef JSON_HEDLEY_CONCAT3
-#endif
-#define JSON_HEDLEY_CONCAT3(a,b,c) JSON_HEDLEY_CONCAT3_EX(a,b,c)
-
-#if defined(JSON_HEDLEY_VERSION_ENCODE)
-    #undef JSON_HEDLEY_VERSION_ENCODE
-#endif
-#define JSON_HEDLEY_VERSION_ENCODE(major,minor,revision) (((major) * 1000000) + ((minor) * 1000) + (revision))
-
-#if defined(JSON_HEDLEY_VERSION_DECODE_MAJOR)
-    #undef JSON_HEDLEY_VERSION_DECODE_MAJOR
-#endif
-#define JSON_HEDLEY_VERSION_DECODE_MAJOR(version) ((version) / 1000000)
-
-#if defined(JSON_HEDLEY_VERSION_DECODE_MINOR)
-    #undef JSON_HEDLEY_VERSION_DECODE_MINOR
-#endif
-#define JSON_HEDLEY_VERSION_DECODE_MINOR(version) (((version) % 1000000) / 1000)
-
-#if defined(JSON_HEDLEY_VERSION_DECODE_REVISION)
-    #undef JSON_HEDLEY_VERSION_DECODE_REVISION
-#endif
-#define JSON_HEDLEY_VERSION_DECODE_REVISION(version) ((version) % 1000)
-
-#if defined(JSON_HEDLEY_GNUC_VERSION)
-    #undef JSON_HEDLEY_GNUC_VERSION
-#endif
-#if defined(__GNUC__) && defined(__GNUC_PATCHLEVEL__)
-    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__)
-#elif defined(__GNUC__)
-    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, 0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_VERSION_CHECK)
-    #undef JSON_HEDLEY_GNUC_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_GNUC_VERSION)
-    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GNUC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_MSVC_VERSION)
-    #undef JSON_HEDLEY_MSVC_VERSION
-#endif
-#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 140000000) && !defined(__ICL)
-    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 10000000, (_MSC_FULL_VER % 10000000) / 100000, (_MSC_FULL_VER % 100000) / 100)
-#elif defined(_MSC_FULL_VER) && !defined(__ICL)
-    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 1000000, (_MSC_FULL_VER % 1000000) / 10000, (_MSC_FULL_VER % 10000) / 10)
-#elif defined(_MSC_VER) && !defined(__ICL)
-    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_VER / 100, _MSC_VER % 100, 0)
-#endif
-
-#if defined(JSON_HEDLEY_MSVC_VERSION_CHECK)
-    #undef JSON_HEDLEY_MSVC_VERSION_CHECK
-#endif
-#if !defined(JSON_HEDLEY_MSVC_VERSION)
-    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (0)
-#elif defined(_MSC_VER) && (_MSC_VER >= 1400)
-    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 10000000) + (minor * 100000) + (patch)))
-#elif defined(_MSC_VER) && (_MSC_VER >= 1200)
-    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 1000000) + (minor * 10000) + (patch)))
-#else
-    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_VER >= ((major * 100) + (minor)))
-#endif
-
-#if defined(JSON_HEDLEY_INTEL_VERSION)
-    #undef JSON_HEDLEY_INTEL_VERSION
-#endif
-#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && !defined(__ICL)
-    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, __INTEL_COMPILER_UPDATE)
-#elif defined(__INTEL_COMPILER) && !defined(__ICL)
-    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, 0)
-#endif
-
-#if defined(JSON_HEDLEY_INTEL_VERSION_CHECK)
-    #undef JSON_HEDLEY_INTEL_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_INTEL_VERSION)
-    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
-    #undef JSON_HEDLEY_INTEL_CL_VERSION
-#endif
-#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && defined(__ICL)
-    #define JSON_HEDLEY_INTEL_CL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER, __INTEL_COMPILER_UPDATE, 0)
-#endif
-
-#if defined(JSON_HEDLEY_INTEL_CL_VERSION_CHECK)
-    #undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
-    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_CL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_PGI_VERSION)
-    #undef JSON_HEDLEY_PGI_VERSION
-#endif
-#if defined(__PGI) && defined(__PGIC__) && defined(__PGIC_MINOR__) && defined(__PGIC_PATCHLEVEL__)
-    #define JSON_HEDLEY_PGI_VERSION JSON_HEDLEY_VERSION_ENCODE(__PGIC__, __PGIC_MINOR__, __PGIC_PATCHLEVEL__)
-#endif
-
-#if defined(JSON_HEDLEY_PGI_VERSION_CHECK)
-    #undef JSON_HEDLEY_PGI_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_PGI_VERSION)
-    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PGI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_SUNPRO_VERSION)
-    #undef JSON_HEDLEY_SUNPRO_VERSION
-#endif
-#if defined(__SUNPRO_C) && (__SUNPRO_C > 0x1000)
-    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_C >> 16) & 0xf) * 10) + ((__SUNPRO_C >> 12) & 0xf), (((__SUNPRO_C >> 8) & 0xf) * 10) + ((__SUNPRO_C >> 4) & 0xf), (__SUNPRO_C & 0xf) * 10)
-#elif defined(__SUNPRO_C)
-    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_C >> 8) & 0xf, (__SUNPRO_C >> 4) & 0xf, (__SUNPRO_C) & 0xf)
-#elif defined(__SUNPRO_CC) && (__SUNPRO_CC > 0x1000)
-    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_CC >> 16) & 0xf) * 10) + ((__SUNPRO_CC >> 12) & 0xf), (((__SUNPRO_CC >> 8) & 0xf) * 10) + ((__SUNPRO_CC >> 4) & 0xf), (__SUNPRO_CC & 0xf) * 10)
-#elif defined(__SUNPRO_CC)
-    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_CC >> 8) & 0xf, (__SUNPRO_CC >> 4) & 0xf, (__SUNPRO_CC) & 0xf)
-#endif
-
-#if defined(JSON_HEDLEY_SUNPRO_VERSION_CHECK)
-    #undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_SUNPRO_VERSION)
-    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_SUNPRO_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
-    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION
-#endif
-#if defined(__EMSCRIPTEN__)
-    #define JSON_HEDLEY_EMSCRIPTEN_VERSION JSON_HEDLEY_VERSION_ENCODE(__EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__)
-#endif
-
-#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK)
-    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
-    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_EMSCRIPTEN_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_ARM_VERSION)
-    #undef JSON_HEDLEY_ARM_VERSION
-#endif
-#if defined(__CC_ARM) && defined(__ARMCOMPILER_VERSION)
-    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCOMPILER_VERSION / 1000000, (__ARMCOMPILER_VERSION % 1000000) / 10000, (__ARMCOMPILER_VERSION % 10000) / 100)
-#elif defined(__CC_ARM) && defined(__ARMCC_VERSION)
-    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCC_VERSION / 1000000, (__ARMCC_VERSION % 1000000) / 10000, (__ARMCC_VERSION % 10000) / 100)
-#endif
-
-#if defined(JSON_HEDLEY_ARM_VERSION_CHECK)
-    #undef JSON_HEDLEY_ARM_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_ARM_VERSION)
-    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_ARM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_IBM_VERSION)
-    #undef JSON_HEDLEY_IBM_VERSION
-#endif
-#if defined(__ibmxl__)
-    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ibmxl_version__, __ibmxl_release__, __ibmxl_modification__)
-#elif defined(__xlC__) && defined(__xlC_ver__)
-    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, (__xlC_ver__ >> 8) & 0xff)
-#elif defined(__xlC__)
-    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, 0)
-#endif
-
-#if defined(JSON_HEDLEY_IBM_VERSION_CHECK)
-    #undef JSON_HEDLEY_IBM_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_IBM_VERSION)
-    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IBM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_VERSION)
-    #undef JSON_HEDLEY_TI_VERSION
-#endif
-#if \
-    defined(__TI_COMPILER_VERSION__) && \
-    ( \
-      defined(__TMS470__) || defined(__TI_ARM__) || \
-      defined(__MSP430__) || \
-      defined(__TMS320C2000__) \
-    )
-#if (__TI_COMPILER_VERSION__ >= 16000000)
-    #define JSON_HEDLEY_TI_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-#endif
-
-#if defined(JSON_HEDLEY_TI_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_VERSION)
-    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
-    #undef JSON_HEDLEY_TI_CL2000_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C2000__)
-    #define JSON_HEDLEY_TI_CL2000_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL2000_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
-    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL2000_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL430_VERSION)
-    #undef JSON_HEDLEY_TI_CL430_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && defined(__MSP430__)
-    #define JSON_HEDLEY_TI_CL430_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL430_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_CL430_VERSION)
-    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL430_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
-    #undef JSON_HEDLEY_TI_ARMCL_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && (defined(__TMS470__) || defined(__TI_ARM__))
-    #define JSON_HEDLEY_TI_ARMCL_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_ARMCL_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
-    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_ARMCL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
-    #undef JSON_HEDLEY_TI_CL6X_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C6X__)
-    #define JSON_HEDLEY_TI_CL6X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL6X_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
-    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL6X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
-    #undef JSON_HEDLEY_TI_CL7X_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && defined(__C7000__)
-    #define JSON_HEDLEY_TI_CL7X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_CL7X_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
-    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL7X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
-    #undef JSON_HEDLEY_TI_CLPRU_VERSION
-#endif
-#if defined(__TI_COMPILER_VERSION__) && defined(__PRU__)
-    #define JSON_HEDLEY_TI_CLPRU_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
-#endif
-
-#if defined(JSON_HEDLEY_TI_CLPRU_VERSION_CHECK)
-    #undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
-    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CLPRU_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_CRAY_VERSION)
-    #undef JSON_HEDLEY_CRAY_VERSION
-#endif
-#if defined(_CRAYC)
-    #if defined(_RELEASE_PATCHLEVEL)
-        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, _RELEASE_PATCHLEVEL)
-    #else
-        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, 0)
-    #endif
-#endif
-
-#if defined(JSON_HEDLEY_CRAY_VERSION_CHECK)
-    #undef JSON_HEDLEY_CRAY_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_CRAY_VERSION)
-    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_CRAY_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_IAR_VERSION)
-    #undef JSON_HEDLEY_IAR_VERSION
-#endif
-#if defined(__IAR_SYSTEMS_ICC__)
-    #if __VER__ > 1000
-        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE((__VER__ / 1000000), ((__VER__ / 1000) % 1000), (__VER__ % 1000))
-    #else
-        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE(__VER__ / 100, __VER__ % 100, 0)
-    #endif
-#endif
-
-#if defined(JSON_HEDLEY_IAR_VERSION_CHECK)
-    #undef JSON_HEDLEY_IAR_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_IAR_VERSION)
-    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IAR_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_TINYC_VERSION)
-    #undef JSON_HEDLEY_TINYC_VERSION
-#endif
-#if defined(__TINYC__)
-    #define JSON_HEDLEY_TINYC_VERSION JSON_HEDLEY_VERSION_ENCODE(__TINYC__ / 1000, (__TINYC__ / 100) % 10, __TINYC__ % 100)
-#endif
-
-#if defined(JSON_HEDLEY_TINYC_VERSION_CHECK)
-    #undef JSON_HEDLEY_TINYC_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_TINYC_VERSION)
-    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TINYC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_DMC_VERSION)
-    #undef JSON_HEDLEY_DMC_VERSION
-#endif
-#if defined(__DMC__)
-    #define JSON_HEDLEY_DMC_VERSION JSON_HEDLEY_VERSION_ENCODE(__DMC__ >> 8, (__DMC__ >> 4) & 0xf, __DMC__ & 0xf)
-#endif
-
-#if defined(JSON_HEDLEY_DMC_VERSION_CHECK)
-    #undef JSON_HEDLEY_DMC_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_DMC_VERSION)
-    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_DMC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_COMPCERT_VERSION)
-    #undef JSON_HEDLEY_COMPCERT_VERSION
-#endif
-#if defined(__COMPCERT_VERSION__)
-    #define JSON_HEDLEY_COMPCERT_VERSION JSON_HEDLEY_VERSION_ENCODE(__COMPCERT_VERSION__ / 10000, (__COMPCERT_VERSION__ / 100) % 100, __COMPCERT_VERSION__ % 100)
-#endif
-
-#if defined(JSON_HEDLEY_COMPCERT_VERSION_CHECK)
-    #undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_COMPCERT_VERSION)
-    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_COMPCERT_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_PELLES_VERSION)
-    #undef JSON_HEDLEY_PELLES_VERSION
-#endif
-#if defined(__POCC__)
-    #define JSON_HEDLEY_PELLES_VERSION JSON_HEDLEY_VERSION_ENCODE(__POCC__ / 100, __POCC__ % 100, 0)
-#endif
-
-#if defined(JSON_HEDLEY_PELLES_VERSION_CHECK)
-    #undef JSON_HEDLEY_PELLES_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_PELLES_VERSION)
-    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PELLES_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
-    #undef JSON_HEDLEY_MCST_LCC_VERSION
-#endif
-#if defined(__LCC__) && defined(__LCC_MINOR__)
-    #define JSON_HEDLEY_MCST_LCC_VERSION JSON_HEDLEY_VERSION_ENCODE(__LCC__ / 100, __LCC__ % 100, __LCC_MINOR__)
-#endif
-
-#if defined(JSON_HEDLEY_MCST_LCC_VERSION_CHECK)
-    #undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
-    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_MCST_LCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_VERSION)
-    #undef JSON_HEDLEY_GCC_VERSION
-#endif
-#if \
-    defined(JSON_HEDLEY_GNUC_VERSION) && \
-    !defined(__clang__) && \
-    !defined(JSON_HEDLEY_INTEL_VERSION) && \
-    !defined(JSON_HEDLEY_PGI_VERSION) && \
-    !defined(JSON_HEDLEY_ARM_VERSION) && \
-    !defined(JSON_HEDLEY_CRAY_VERSION) && \
-    !defined(JSON_HEDLEY_TI_VERSION) && \
-    !defined(JSON_HEDLEY_TI_ARMCL_VERSION) && \
-    !defined(JSON_HEDLEY_TI_CL430_VERSION) && \
-    !defined(JSON_HEDLEY_TI_CL2000_VERSION) && \
-    !defined(JSON_HEDLEY_TI_CL6X_VERSION) && \
-    !defined(JSON_HEDLEY_TI_CL7X_VERSION) && \
-    !defined(JSON_HEDLEY_TI_CLPRU_VERSION) && \
-    !defined(__COMPCERT__) && \
-    !defined(JSON_HEDLEY_MCST_LCC_VERSION)
-    #define JSON_HEDLEY_GCC_VERSION JSON_HEDLEY_GNUC_VERSION
-#endif
-
-#if defined(JSON_HEDLEY_GCC_VERSION_CHECK)
-    #undef JSON_HEDLEY_GCC_VERSION_CHECK
-#endif
-#if defined(JSON_HEDLEY_GCC_VERSION)
-    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
-#else
-    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (0)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_ATTRIBUTE)
-    #undef JSON_HEDLEY_HAS_ATTRIBUTE
-#endif
-#if \
-  defined(__has_attribute) && \
-  ( \
-    (!defined(JSON_HEDLEY_IAR_VERSION) || JSON_HEDLEY_IAR_VERSION_CHECK(8,5,9)) \
-  )
-#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) __has_attribute(attribute)
-#else
-#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_ATTRIBUTE)
-    #undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
-#endif
-#if defined(__has_attribute)
-    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_ATTRIBUTE)
-    #undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
-#endif
-#if defined(__has_attribute)
-    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
-#else
-    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE)
-    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
-#endif
-#if \
-    defined(__has_cpp_attribute) && \
-    defined(__cplusplus) && \
-    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0))
-    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) __has_cpp_attribute(attribute)
-#else
-    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) (0)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS)
-    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
-#endif
-#if !defined(__cplusplus) || !defined(__has_cpp_attribute)
-    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
-#elif \
-    !defined(JSON_HEDLEY_PGI_VERSION) && \
-    !defined(JSON_HEDLEY_IAR_VERSION) && \
-    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) && \
-    (!defined(JSON_HEDLEY_MSVC_VERSION) || JSON_HEDLEY_MSVC_VERSION_CHECK(19,20,0))
-    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(ns::attribute)
-#else
-    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE)
-    #undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
-#endif
-#if defined(__has_cpp_attribute) && defined(__cplusplus)
-    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE)
-    #undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
-#endif
-#if defined(__has_cpp_attribute) && defined(__cplusplus)
-    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
-#else
-    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_BUILTIN)
-    #undef JSON_HEDLEY_HAS_BUILTIN
-#endif
-#if defined(__has_builtin)
-    #define JSON_HEDLEY_HAS_BUILTIN(builtin) __has_builtin(builtin)
-#else
-    #define JSON_HEDLEY_HAS_BUILTIN(builtin) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_BUILTIN)
-    #undef JSON_HEDLEY_GNUC_HAS_BUILTIN
-#endif
-#if defined(__has_builtin)
-    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_BUILTIN)
-    #undef JSON_HEDLEY_GCC_HAS_BUILTIN
-#endif
-#if defined(__has_builtin)
-    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
-#else
-    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_FEATURE)
-    #undef JSON_HEDLEY_HAS_FEATURE
-#endif
-#if defined(__has_feature)
-    #define JSON_HEDLEY_HAS_FEATURE(feature) __has_feature(feature)
-#else
-    #define JSON_HEDLEY_HAS_FEATURE(feature) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_FEATURE)
-    #undef JSON_HEDLEY_GNUC_HAS_FEATURE
-#endif
-#if defined(__has_feature)
-    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_FEATURE)
-    #undef JSON_HEDLEY_GCC_HAS_FEATURE
-#endif
-#if defined(__has_feature)
-    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
-#else
-    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_EXTENSION)
-    #undef JSON_HEDLEY_HAS_EXTENSION
-#endif
-#if defined(__has_extension)
-    #define JSON_HEDLEY_HAS_EXTENSION(extension) __has_extension(extension)
-#else
-    #define JSON_HEDLEY_HAS_EXTENSION(extension) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_EXTENSION)
-    #undef JSON_HEDLEY_GNUC_HAS_EXTENSION
-#endif
-#if defined(__has_extension)
-    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_EXTENSION)
-    #undef JSON_HEDLEY_GCC_HAS_EXTENSION
-#endif
-#if defined(__has_extension)
-    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
-#else
-    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE)
-    #undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
-#endif
-#if defined(__has_declspec_attribute)
-    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) __has_declspec_attribute(attribute)
-#else
-    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE)
-    #undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
-#endif
-#if defined(__has_declspec_attribute)
-    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE)
-    #undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
-#endif
-#if defined(__has_declspec_attribute)
-    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
-#else
-    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_HAS_WARNING)
-    #undef JSON_HEDLEY_HAS_WARNING
-#endif
-#if defined(__has_warning)
-    #define JSON_HEDLEY_HAS_WARNING(warning) __has_warning(warning)
-#else
-    #define JSON_HEDLEY_HAS_WARNING(warning) (0)
-#endif
-
-#if defined(JSON_HEDLEY_GNUC_HAS_WARNING)
-    #undef JSON_HEDLEY_GNUC_HAS_WARNING
-#endif
-#if defined(__has_warning)
-    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
-#else
-    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_GCC_HAS_WARNING)
-    #undef JSON_HEDLEY_GCC_HAS_WARNING
-#endif
-#if defined(__has_warning)
-    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
-#else
-    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if \
-    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
-    defined(__clang__) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,0,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) || \
-    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,17) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(8,0,0) || \
-    (JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) && defined(__C99_PRAGMA_OPERATOR))
-    #define JSON_HEDLEY_PRAGMA(value) _Pragma(#value)
-#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
-    #define JSON_HEDLEY_PRAGMA(value) __pragma(value)
-#else
-    #define JSON_HEDLEY_PRAGMA(value)
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_PUSH)
-    #undef JSON_HEDLEY_DIAGNOSTIC_PUSH
-#endif
-#if defined(JSON_HEDLEY_DIAGNOSTIC_POP)
-    #undef JSON_HEDLEY_DIAGNOSTIC_POP
-#endif
-#if defined(__clang__)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("clang diagnostic push")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("clang diagnostic pop")
-#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("GCC diagnostic push")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("GCC diagnostic pop")
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH __pragma(warning(push))
-    #define JSON_HEDLEY_DIAGNOSTIC_POP __pragma(warning(pop))
-#elif JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("push")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("pop")
-#elif \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,4,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("diag_push")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("diag_pop")
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
-    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_PUSH
-    #define JSON_HEDLEY_DIAGNOSTIC_POP
-#endif
-
-/* JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ is for
-   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
-#endif
-#if defined(__cplusplus)
-#  if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat")
-#    if JSON_HEDLEY_HAS_WARNING("-Wc++17-extensions")
-#      if JSON_HEDLEY_HAS_WARNING("-Wc++1z-extensions")
-#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
-    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
-    _Pragma("clang diagnostic ignored \"-Wc++1z-extensions\"") \
-    xpr \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#      else
-#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
-    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
-    xpr \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#      endif
-#    else
-#      define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
-    xpr \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#    endif
-#  endif
-#endif
-#if !defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(x) x
-#endif
-
-#if defined(JSON_HEDLEY_CONST_CAST)
-    #undef JSON_HEDLEY_CONST_CAST
-#endif
-#if defined(__cplusplus)
-#  define JSON_HEDLEY_CONST_CAST(T, expr) (const_cast<T>(expr))
-#elif \
-  JSON_HEDLEY_HAS_WARNING("-Wcast-qual") || \
-  JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-#  define JSON_HEDLEY_CONST_CAST(T, expr) (__extension__ ({ \
-        JSON_HEDLEY_DIAGNOSTIC_PUSH \
-        JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL \
-        ((T) (expr)); \
-        JSON_HEDLEY_DIAGNOSTIC_POP \
-    }))
-#else
-#  define JSON_HEDLEY_CONST_CAST(T, expr) ((T) (expr))
-#endif
-
-#if defined(JSON_HEDLEY_REINTERPRET_CAST)
-    #undef JSON_HEDLEY_REINTERPRET_CAST
-#endif
-#if defined(__cplusplus)
-    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) (reinterpret_cast<T>(expr))
-#else
-    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) ((T) (expr))
-#endif
-
-#if defined(JSON_HEDLEY_STATIC_CAST)
-    #undef JSON_HEDLEY_STATIC_CAST
-#endif
-#if defined(__cplusplus)
-    #define JSON_HEDLEY_STATIC_CAST(T, expr) (static_cast<T>(expr))
-#else
-    #define JSON_HEDLEY_STATIC_CAST(T, expr) ((T) (expr))
-#endif
-
-#if defined(JSON_HEDLEY_CPP_CAST)
-    #undef JSON_HEDLEY_CPP_CAST
-#endif
-#if defined(__cplusplus)
-#  if JSON_HEDLEY_HAS_WARNING("-Wold-style-cast")
-#    define JSON_HEDLEY_CPP_CAST(T, expr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wold-style-cast\"") \
-    ((T) (expr)) \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#  elif JSON_HEDLEY_IAR_VERSION_CHECK(8,3,0)
-#    define JSON_HEDLEY_CPP_CAST(T, expr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("diag_suppress=Pe137") \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#  else
-#    define JSON_HEDLEY_CPP_CAST(T, expr) ((T) (expr))
-#  endif
-#else
-#  define JSON_HEDLEY_CPP_CAST(T, expr) (expr)
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wdeprecated-declarations")
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"")
-#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warning(disable:1478 1786)")
-#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:1478 1786))
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1216,1444,1445")
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
-#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:4996))
-#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
-#elif \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1291,1718")
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && !defined(__cplusplus)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,E_DEPRECATED_ATT,E_DEPRECATED_ATT_MESS)")
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && defined(__cplusplus)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,symdeprecated,symdeprecated2)")
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress=Pe1444,Pe1215")
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warn(disable:2241)")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("clang diagnostic ignored \"-Wunknown-pragmas\"")
-#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("warning(disable:161)")
-#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:161))
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 1675")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"")
-#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:4068))
-#elif \
-    JSON_HEDLEY_TI_VERSION_CHECK(16,9,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
-#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress=Pe161")
-#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 161")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wunknown-attributes")
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("clang diagnostic ignored \"-Wunknown-attributes\"")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
-#elif JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("warning(disable:1292)")
-#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:1292))
-#elif JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:5030))
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097,1098")
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("error_messages(off,attrskipunsup)")
-#elif \
-    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1173")
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress=Pe1097")
-#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wcast-qual")
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("clang diagnostic ignored \"-Wcast-qual\"")
-#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("warning(disable:2203 2331)")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("GCC diagnostic ignored \"-Wcast-qual\"")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
-#endif
-
-#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION)
-    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wunused-function")
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("clang diagnostic ignored \"-Wunused-function\"")
-#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("GCC diagnostic ignored \"-Wunused-function\"")
-#elif JSON_HEDLEY_MSVC_VERSION_CHECK(1,0,0)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION __pragma(warning(disable:4505))
-#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("diag_suppress 3142")
-#else
-    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
-#endif
-
-#if defined(JSON_HEDLEY_DEPRECATED)
-    #undef JSON_HEDLEY_DEPRECATED
-#endif
-#if defined(JSON_HEDLEY_DEPRECATED_FOR)
-    #undef JSON_HEDLEY_DEPRECATED_FOR
-#endif
-#if \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated("Since " # since))
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated("Since " #since "; use " #replacement))
-#elif \
-    (JSON_HEDLEY_HAS_EXTENSION(attribute_deprecated_with_message) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(18,1,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__("Since " #since)))
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__("Since " #since "; use " #replacement)))
-#elif defined(__cplusplus) && (__cplusplus >= 201402L)
-    #define JSON_HEDLEY_DEPRECATED(since) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since)]])
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since "; use " #replacement)]])
-#elif \
-    JSON_HEDLEY_HAS_ATTRIBUTE(deprecated) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
-    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
-    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__))
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__))
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
-    JSON_HEDLEY_PELLES_VERSION_CHECK(6,50,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated)
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated)
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_DEPRECATED(since) _Pragma("deprecated")
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) _Pragma("deprecated")
-#else
-    #define JSON_HEDLEY_DEPRECATED(since)
-    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement)
-#endif
-
-#if defined(JSON_HEDLEY_UNAVAILABLE)
-    #undef JSON_HEDLEY_UNAVAILABLE
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(warning) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_UNAVAILABLE(available_since) __attribute__((__warning__("Not available until " #available_since)))
-#else
-    #define JSON_HEDLEY_UNAVAILABLE(available_since)
-#endif
-
-#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT)
-    #undef JSON_HEDLEY_WARN_UNUSED_RESULT
-#endif
-#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT_MSG)
-    #undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(warn_unused_result) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__))
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) __attribute__((__warn_unused_result__))
-#elif (JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) >= 201907L)
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard(msg)]])
-#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard)
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
-#elif defined(_Check_return_) /* SAL */
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT _Check_return_
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) _Check_return_
-#else
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT
-    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg)
-#endif
-
-#if defined(JSON_HEDLEY_SENTINEL)
-    #undef JSON_HEDLEY_SENTINEL
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(sentinel) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_SENTINEL(position) __attribute__((__sentinel__(position)))
-#else
-    #define JSON_HEDLEY_SENTINEL(position)
-#endif
-
-#if defined(JSON_HEDLEY_NO_RETURN)
-    #undef JSON_HEDLEY_NO_RETURN
-#endif
-#if JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_NO_RETURN __noreturn
-#elif \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
-#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
-    #define JSON_HEDLEY_NO_RETURN _Noreturn
-#elif defined(__cplusplus) && (__cplusplus >= 201103L)
-    #define JSON_HEDLEY_NO_RETURN JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[noreturn]])
-#elif \
-    JSON_HEDLEY_HAS_ATTRIBUTE(noreturn) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,2,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
-    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
-    #define JSON_HEDLEY_NO_RETURN _Pragma("does_not_return")
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
-#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
-    #define JSON_HEDLEY_NO_RETURN _Pragma("FUNC_NEVER_RETURNS;")
-#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
-    #define JSON_HEDLEY_NO_RETURN __attribute((noreturn))
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
-    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
-#else
-    #define JSON_HEDLEY_NO_RETURN
-#endif
-
-#if defined(JSON_HEDLEY_NO_ESCAPE)
-    #undef JSON_HEDLEY_NO_ESCAPE
-#endif
-#if JSON_HEDLEY_HAS_ATTRIBUTE(noescape)
-    #define JSON_HEDLEY_NO_ESCAPE __attribute__((__noescape__))
-#else
-    #define JSON_HEDLEY_NO_ESCAPE
-#endif
-
-#if defined(JSON_HEDLEY_UNREACHABLE)
-    #undef JSON_HEDLEY_UNREACHABLE
-#endif
-#if defined(JSON_HEDLEY_UNREACHABLE_RETURN)
-    #undef JSON_HEDLEY_UNREACHABLE_RETURN
-#endif
-#if defined(JSON_HEDLEY_ASSUME)
-    #undef JSON_HEDLEY_ASSUME
-#endif
-#if \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_ASSUME(expr) __assume(expr)
-#elif JSON_HEDLEY_HAS_BUILTIN(__builtin_assume)
-    #define JSON_HEDLEY_ASSUME(expr) __builtin_assume(expr)
-#elif \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
-    #if defined(__cplusplus)
-        #define JSON_HEDLEY_ASSUME(expr) std::_nassert(expr)
-    #else
-        #define JSON_HEDLEY_ASSUME(expr) _nassert(expr)
-    #endif
-#endif
-#if \
-    (JSON_HEDLEY_HAS_BUILTIN(__builtin_unreachable) && (!defined(JSON_HEDLEY_ARM_VERSION))) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(18,10,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,5) || \
-    JSON_HEDLEY_CRAY_VERSION_CHECK(10,0,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_UNREACHABLE() __builtin_unreachable()
-#elif defined(JSON_HEDLEY_ASSUME)
-    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
-#endif
-#if !defined(JSON_HEDLEY_ASSUME)
-    #if defined(JSON_HEDLEY_UNREACHABLE)
-        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, ((expr) ? 1 : (JSON_HEDLEY_UNREACHABLE(), 1)))
-    #else
-        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, expr)
-    #endif
-#endif
-#if defined(JSON_HEDLEY_UNREACHABLE)
-    #if  \
-        JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
-        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
-        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (JSON_HEDLEY_STATIC_CAST(void, JSON_HEDLEY_ASSUME(0)), (value))
-    #else
-        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) JSON_HEDLEY_UNREACHABLE()
-    #endif
-#else
-    #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (value)
-#endif
-#if !defined(JSON_HEDLEY_UNREACHABLE)
-    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
-#endif
-
-JSON_HEDLEY_DIAGNOSTIC_PUSH
-#if JSON_HEDLEY_HAS_WARNING("-Wpedantic")
-    #pragma clang diagnostic ignored "-Wpedantic"
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat-pedantic") && defined(__cplusplus)
-    #pragma clang diagnostic ignored "-Wc++98-compat-pedantic"
-#endif
-#if JSON_HEDLEY_GCC_HAS_WARNING("-Wvariadic-macros",4,0,0)
-    #if defined(__clang__)
-        #pragma clang diagnostic ignored "-Wvariadic-macros"
-    #elif defined(JSON_HEDLEY_GCC_VERSION)
-        #pragma GCC diagnostic ignored "-Wvariadic-macros"
-    #endif
-#endif
-#if defined(JSON_HEDLEY_NON_NULL)
-    #undef JSON_HEDLEY_NON_NULL
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(nonnull) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
-    #define JSON_HEDLEY_NON_NULL(...) __attribute__((__nonnull__(__VA_ARGS__)))
-#else
-    #define JSON_HEDLEY_NON_NULL(...)
-#endif
-JSON_HEDLEY_DIAGNOSTIC_POP
-
-#if defined(JSON_HEDLEY_PRINTF_FORMAT)
-    #undef JSON_HEDLEY_PRINTF_FORMAT
-#endif
-#if defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && !defined(__USE_MINGW_ANSI_STDIO)
-    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(ms_printf, string_idx, first_to_check)))
-#elif defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && defined(__USE_MINGW_ANSI_STDIO)
-    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(gnu_printf, string_idx, first_to_check)))
-#elif \
-    JSON_HEDLEY_HAS_ATTRIBUTE(format) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(__printf__, string_idx, first_to_check)))
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(6,0,0)
-    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __declspec(vaformat(printf,string_idx,first_to_check))
-#else
-    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check)
-#endif
-
-#if defined(JSON_HEDLEY_CONSTEXPR)
-    #undef JSON_HEDLEY_CONSTEXPR
-#endif
-#if defined(__cplusplus)
-    #if __cplusplus >= 201103L
-        #define JSON_HEDLEY_CONSTEXPR JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(constexpr)
-    #endif
-#endif
-#if !defined(JSON_HEDLEY_CONSTEXPR)
-    #define JSON_HEDLEY_CONSTEXPR
-#endif
-
-#if defined(JSON_HEDLEY_PREDICT)
-    #undef JSON_HEDLEY_PREDICT
-#endif
-#if defined(JSON_HEDLEY_LIKELY)
-    #undef JSON_HEDLEY_LIKELY
-#endif
-#if defined(JSON_HEDLEY_UNLIKELY)
-    #undef JSON_HEDLEY_UNLIKELY
-#endif
-#if defined(JSON_HEDLEY_UNPREDICTABLE)
-    #undef JSON_HEDLEY_UNPREDICTABLE
-#endif
-#if JSON_HEDLEY_HAS_BUILTIN(__builtin_unpredictable)
-    #define JSON_HEDLEY_UNPREDICTABLE(expr) __builtin_unpredictable((expr))
-#endif
-#if \
-  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect_with_probability) && !defined(JSON_HEDLEY_PGI_VERSION)) || \
-  JSON_HEDLEY_GCC_VERSION_CHECK(9,0,0) || \
-  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-#  define JSON_HEDLEY_PREDICT(expr, value, probability) __builtin_expect_with_probability(  (expr), (value), (probability))
-#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability)   __builtin_expect_with_probability(!!(expr),    1   , (probability))
-#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability)  __builtin_expect_with_probability(!!(expr),    0   , (probability))
-#  define JSON_HEDLEY_LIKELY(expr)                      __builtin_expect                 (!!(expr),    1                  )
-#  define JSON_HEDLEY_UNLIKELY(expr)                    __builtin_expect                 (!!(expr),    0                  )
-#elif \
-  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
-  JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-  (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
-  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
-  JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
-  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
-  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
-  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-  JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,27) || \
-  JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
-  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-#  define JSON_HEDLEY_PREDICT(expr, expected, probability) \
-    (((probability) >= 0.9) ? __builtin_expect((expr), (expected)) : (JSON_HEDLEY_STATIC_CAST(void, expected), (expr)))
-#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) \
-    (__extension__ ({ \
-        double hedley_probability_ = (probability); \
-        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 1) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 0) : !!(expr))); \
-    }))
-#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) \
-    (__extension__ ({ \
-        double hedley_probability_ = (probability); \
-        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 0) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 1) : !!(expr))); \
-    }))
-#  define JSON_HEDLEY_LIKELY(expr)   __builtin_expect(!!(expr), 1)
-#  define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect(!!(expr), 0)
-#else
-#  define JSON_HEDLEY_PREDICT(expr, expected, probability) (JSON_HEDLEY_STATIC_CAST(void, expected), (expr))
-#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) (!!(expr))
-#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) (!!(expr))
-#  define JSON_HEDLEY_LIKELY(expr) (!!(expr))
-#  define JSON_HEDLEY_UNLIKELY(expr) (!!(expr))
-#endif
-#if !defined(JSON_HEDLEY_UNPREDICTABLE)
-    #define JSON_HEDLEY_UNPREDICTABLE(expr) JSON_HEDLEY_PREDICT(expr, 1, 0.5)
-#endif
-
-#if defined(JSON_HEDLEY_MALLOC)
-    #undef JSON_HEDLEY_MALLOC
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(malloc) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_MALLOC __attribute__((__malloc__))
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
-    #define JSON_HEDLEY_MALLOC _Pragma("returns_new_memory")
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_MALLOC __declspec(restrict)
-#else
-    #define JSON_HEDLEY_MALLOC
-#endif
-
-#if defined(JSON_HEDLEY_PURE)
-    #undef JSON_HEDLEY_PURE
-#endif
-#if \
-  JSON_HEDLEY_HAS_ATTRIBUTE(pure) || \
-  JSON_HEDLEY_GCC_VERSION_CHECK(2,96,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-  JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
-  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-#  define JSON_HEDLEY_PURE __attribute__((__pure__))
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
-#  define JSON_HEDLEY_PURE _Pragma("does_not_write_global_data")
-#elif defined(__cplusplus) && \
-    ( \
-      JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
-      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) || \
-      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) \
-    )
-#  define JSON_HEDLEY_PURE _Pragma("FUNC_IS_PURE;")
-#else
-#  define JSON_HEDLEY_PURE
-#endif
-
-#if defined(JSON_HEDLEY_CONST)
-    #undef JSON_HEDLEY_CONST
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(const) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(2,5,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_CONST __attribute__((__const__))
-#elif \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
-    #define JSON_HEDLEY_CONST _Pragma("no_side_effect")
-#else
-    #define JSON_HEDLEY_CONST JSON_HEDLEY_PURE
-#endif
-
-#if defined(JSON_HEDLEY_RESTRICT)
-    #undef JSON_HEDLEY_RESTRICT
-#endif
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && !defined(__cplusplus)
-    #define JSON_HEDLEY_RESTRICT restrict
-#elif \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,4) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)) || \
-    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
-    defined(__clang__) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_RESTRICT __restrict
-#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,3,0) && !defined(__cplusplus)
-    #define JSON_HEDLEY_RESTRICT _Restrict
-#else
-    #define JSON_HEDLEY_RESTRICT
-#endif
-
-#if defined(JSON_HEDLEY_INLINE)
-    #undef JSON_HEDLEY_INLINE
-#endif
-#if \
-    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
-    (defined(__cplusplus) && (__cplusplus >= 199711L))
-    #define JSON_HEDLEY_INLINE inline
-#elif \
-    defined(JSON_HEDLEY_GCC_VERSION) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(6,2,0)
-    #define JSON_HEDLEY_INLINE __inline__
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,1,0) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_INLINE __inline
-#else
-    #define JSON_HEDLEY_INLINE
-#endif
-
-#if defined(JSON_HEDLEY_ALWAYS_INLINE)
-    #undef JSON_HEDLEY_ALWAYS_INLINE
-#endif
-#if \
-  JSON_HEDLEY_HAS_ATTRIBUTE(always_inline) || \
-  JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
-  JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
-#  define JSON_HEDLEY_ALWAYS_INLINE __attribute__((__always_inline__)) JSON_HEDLEY_INLINE
-#elif \
-  JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
-  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-#  define JSON_HEDLEY_ALWAYS_INLINE __forceinline
-#elif defined(__cplusplus) && \
-    ( \
-      JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-      JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-      JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
-      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-      JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) \
-    )
-#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("FUNC_ALWAYS_INLINE;")
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("inline=forced")
-#else
-#  define JSON_HEDLEY_ALWAYS_INLINE JSON_HEDLEY_INLINE
-#endif
-
-#if defined(JSON_HEDLEY_NEVER_INLINE)
-    #undef JSON_HEDLEY_NEVER_INLINE
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(noinline) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
-    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
-    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
-    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
-    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
-    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
-    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
-    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
-    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
-    #define JSON_HEDLEY_NEVER_INLINE __attribute__((__noinline__))
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
-#elif JSON_HEDLEY_PGI_VERSION_CHECK(10,2,0)
-    #define JSON_HEDLEY_NEVER_INLINE _Pragma("noinline")
-#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
-    #define JSON_HEDLEY_NEVER_INLINE _Pragma("FUNC_CANNOT_INLINE;")
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-    #define JSON_HEDLEY_NEVER_INLINE _Pragma("inline=never")
-#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
-    #define JSON_HEDLEY_NEVER_INLINE __attribute((noinline))
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
-    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
-#else
-    #define JSON_HEDLEY_NEVER_INLINE
-#endif
-
-#if defined(JSON_HEDLEY_PRIVATE)
-    #undef JSON_HEDLEY_PRIVATE
-#endif
-#if defined(JSON_HEDLEY_PUBLIC)
-    #undef JSON_HEDLEY_PUBLIC
-#endif
-#if defined(JSON_HEDLEY_IMPORT)
-    #undef JSON_HEDLEY_IMPORT
-#endif
-#if defined(_WIN32) || defined(__CYGWIN__)
-#  define JSON_HEDLEY_PRIVATE
-#  define JSON_HEDLEY_PUBLIC   __declspec(dllexport)
-#  define JSON_HEDLEY_IMPORT   __declspec(dllimport)
-#else
-#  if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(visibility) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
-    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
-    ( \
-      defined(__TI_EABI__) && \
-      ( \
-        (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
-        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) \
-      ) \
-    ) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-#    define JSON_HEDLEY_PRIVATE __attribute__((__visibility__("hidden")))
-#    define JSON_HEDLEY_PUBLIC  __attribute__((__visibility__("default")))
-#  else
-#    define JSON_HEDLEY_PRIVATE
-#    define JSON_HEDLEY_PUBLIC
-#  endif
-#  define JSON_HEDLEY_IMPORT    extern
-#endif
-
-#if defined(JSON_HEDLEY_NO_THROW)
-    #undef JSON_HEDLEY_NO_THROW
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(nothrow) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_NO_THROW __attribute__((__nothrow__))
-#elif \
-    JSON_HEDLEY_MSVC_VERSION_CHECK(13,1,0) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
-    #define JSON_HEDLEY_NO_THROW __declspec(nothrow)
-#else
-    #define JSON_HEDLEY_NO_THROW
-#endif
-
-#if defined(JSON_HEDLEY_FALL_THROUGH)
-    #undef JSON_HEDLEY_FALL_THROUGH
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(fallthrough) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(7,0,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_FALL_THROUGH __attribute__((__fallthrough__))
-#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(clang,fallthrough)
-    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[clang::fallthrough]])
-#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(fallthrough)
-    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[fallthrough]])
-#elif defined(__fallthrough) /* SAL */
-    #define JSON_HEDLEY_FALL_THROUGH __fallthrough
-#else
-    #define JSON_HEDLEY_FALL_THROUGH
-#endif
-
-#if defined(JSON_HEDLEY_RETURNS_NON_NULL)
-    #undef JSON_HEDLEY_RETURNS_NON_NULL
-#endif
-#if \
-    JSON_HEDLEY_HAS_ATTRIBUTE(returns_nonnull) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_RETURNS_NON_NULL __attribute__((__returns_nonnull__))
-#elif defined(_Ret_notnull_) /* SAL */
-    #define JSON_HEDLEY_RETURNS_NON_NULL _Ret_notnull_
-#else
-    #define JSON_HEDLEY_RETURNS_NON_NULL
-#endif
-
-#if defined(JSON_HEDLEY_ARRAY_PARAM)
-    #undef JSON_HEDLEY_ARRAY_PARAM
-#endif
-#if \
-    defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && \
-    !defined(__STDC_NO_VLA__) && \
-    !defined(__cplusplus) && \
-    !defined(JSON_HEDLEY_PGI_VERSION) && \
-    !defined(JSON_HEDLEY_TINYC_VERSION)
-    #define JSON_HEDLEY_ARRAY_PARAM(name) (name)
-#else
-    #define JSON_HEDLEY_ARRAY_PARAM(name)
-#endif
-
-#if defined(JSON_HEDLEY_IS_CONSTANT)
-    #undef JSON_HEDLEY_IS_CONSTANT
-#endif
-#if defined(JSON_HEDLEY_REQUIRE_CONSTEXPR)
-    #undef JSON_HEDLEY_REQUIRE_CONSTEXPR
-#endif
-/* JSON_HEDLEY_IS_CONSTEXPR_ is for
-   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
-#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
-    #undef JSON_HEDLEY_IS_CONSTEXPR_
-#endif
-#if \
-    JSON_HEDLEY_HAS_BUILTIN(__builtin_constant_p) || \
-    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
-    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,19) || \
-    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
-    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
-    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
-    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) && !defined(__cplusplus)) || \
-    JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
-    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
-    #define JSON_HEDLEY_IS_CONSTANT(expr) __builtin_constant_p(expr)
-#endif
-#if !defined(__cplusplus)
-#  if \
-       JSON_HEDLEY_HAS_BUILTIN(__builtin_types_compatible_p) || \
-       JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
-       JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-       JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
-       JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
-       JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
-       JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,24)
-#if defined(__INTPTR_TYPE__)
-    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0)), int*)
-#else
-    #include <stdint.h>
-    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((intptr_t) ((expr) * 0)) : (int*) 0)), int*)
-#endif
-#  elif \
-       ( \
-          defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) && \
-          !defined(JSON_HEDLEY_SUNPRO_VERSION) && \
-          !defined(JSON_HEDLEY_PGI_VERSION) && \
-          !defined(JSON_HEDLEY_IAR_VERSION)) || \
-       (JSON_HEDLEY_HAS_EXTENSION(c_generic_selections) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
-       JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
-       JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) || \
-       JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
-       JSON_HEDLEY_ARM_VERSION_CHECK(5,3,0)
-#if defined(__INTPTR_TYPE__)
-    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0), int*: 1, void*: 0)
-#else
-    #include <stdint.h>
-    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((intptr_t) * 0) : (int*) 0), int*: 1, void*: 0)
-#endif
-#  elif \
-       defined(JSON_HEDLEY_GCC_VERSION) || \
-       defined(JSON_HEDLEY_INTEL_VERSION) || \
-       defined(JSON_HEDLEY_TINYC_VERSION) || \
-       defined(JSON_HEDLEY_TI_ARMCL_VERSION) || \
-       JSON_HEDLEY_TI_CL430_VERSION_CHECK(18,12,0) || \
-       defined(JSON_HEDLEY_TI_CL2000_VERSION) || \
-       defined(JSON_HEDLEY_TI_CL6X_VERSION) || \
-       defined(JSON_HEDLEY_TI_CL7X_VERSION) || \
-       defined(JSON_HEDLEY_TI_CLPRU_VERSION) || \
-       defined(__clang__)
-#    define JSON_HEDLEY_IS_CONSTEXPR_(expr) ( \
-        sizeof(void) != \
-        sizeof(*( \
-                  1 ? \
-                  ((void*) ((expr) * 0L) ) : \
-((struct { char v[sizeof(void) * 2]; } *) 1) \
-                ) \
-              ) \
-                                            )
-#  endif
-#endif
-#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
-    #if !defined(JSON_HEDLEY_IS_CONSTANT)
-        #define JSON_HEDLEY_IS_CONSTANT(expr) JSON_HEDLEY_IS_CONSTEXPR_(expr)
-    #endif
-    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (JSON_HEDLEY_IS_CONSTEXPR_(expr) ? (expr) : (-1))
-#else
-    #if !defined(JSON_HEDLEY_IS_CONSTANT)
-        #define JSON_HEDLEY_IS_CONSTANT(expr) (0)
-    #endif
-    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (expr)
-#endif
-
-#if defined(JSON_HEDLEY_BEGIN_C_DECLS)
-    #undef JSON_HEDLEY_BEGIN_C_DECLS
-#endif
-#if defined(JSON_HEDLEY_END_C_DECLS)
-    #undef JSON_HEDLEY_END_C_DECLS
-#endif
-#if defined(JSON_HEDLEY_C_DECL)
-    #undef JSON_HEDLEY_C_DECL
-#endif
-#if defined(__cplusplus)
-    #define JSON_HEDLEY_BEGIN_C_DECLS extern "C" {
-    #define JSON_HEDLEY_END_C_DECLS }
-    #define JSON_HEDLEY_C_DECL extern "C"
-#else
-    #define JSON_HEDLEY_BEGIN_C_DECLS
-    #define JSON_HEDLEY_END_C_DECLS
-    #define JSON_HEDLEY_C_DECL
-#endif
-
-#if defined(JSON_HEDLEY_STATIC_ASSERT)
-    #undef JSON_HEDLEY_STATIC_ASSERT
-#endif
-#if \
-  !defined(__cplusplus) && ( \
-      (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) || \
-      (JSON_HEDLEY_HAS_FEATURE(c_static_assert) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
-      JSON_HEDLEY_GCC_VERSION_CHECK(6,0,0) || \
-      JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
-      defined(_Static_assert) \
-    )
-#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) _Static_assert(expr, message)
-#elif \
-  (defined(__cplusplus) && (__cplusplus >= 201103L)) || \
-  JSON_HEDLEY_MSVC_VERSION_CHECK(16,0,0) || \
-  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(static_assert(expr, message))
-#else
-#  define JSON_HEDLEY_STATIC_ASSERT(expr, message)
-#endif
-
-#if defined(JSON_HEDLEY_NULL)
-    #undef JSON_HEDLEY_NULL
-#endif
-#if defined(__cplusplus)
-    #if __cplusplus >= 201103L
-        #define JSON_HEDLEY_NULL JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(nullptr)
-    #elif defined(NULL)
-        #define JSON_HEDLEY_NULL NULL
-    #else
-        #define JSON_HEDLEY_NULL JSON_HEDLEY_STATIC_CAST(void*, 0)
-    #endif
-#elif defined(NULL)
-    #define JSON_HEDLEY_NULL NULL
-#else
-    #define JSON_HEDLEY_NULL ((void*) 0)
-#endif
-
-#if defined(JSON_HEDLEY_MESSAGE)
-    #undef JSON_HEDLEY_MESSAGE
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
-#  define JSON_HEDLEY_MESSAGE(msg) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
-    JSON_HEDLEY_PRAGMA(message msg) \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#elif \
-  JSON_HEDLEY_GCC_VERSION_CHECK(4,4,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message msg)
-#elif JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0)
-#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(_CRI message msg)
-#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
-#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
-#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,0,0)
-#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
-#else
-#  define JSON_HEDLEY_MESSAGE(msg)
-#endif
-
-#if defined(JSON_HEDLEY_WARNING)
-    #undef JSON_HEDLEY_WARNING
-#endif
-#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
-#  define JSON_HEDLEY_WARNING(msg) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
-    JSON_HEDLEY_PRAGMA(clang warning msg) \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#elif \
-  JSON_HEDLEY_GCC_VERSION_CHECK(4,8,0) || \
-  JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
-  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
-#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(GCC warning msg)
-#elif \
-  JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
-  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(message(msg))
-#else
-#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_MESSAGE(msg)
-#endif
-
-#if defined(JSON_HEDLEY_REQUIRE)
-    #undef JSON_HEDLEY_REQUIRE
-#endif
-#if defined(JSON_HEDLEY_REQUIRE_MSG)
-    #undef JSON_HEDLEY_REQUIRE_MSG
-#endif
-#if JSON_HEDLEY_HAS_ATTRIBUTE(diagnose_if)
-#  if JSON_HEDLEY_HAS_WARNING("-Wgcc-compat")
-#    define JSON_HEDLEY_REQUIRE(expr) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
-    __attribute__((diagnose_if(!(expr), #expr, "error"))) \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) \
-    JSON_HEDLEY_DIAGNOSTIC_PUSH \
-    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
-    __attribute__((diagnose_if(!(expr), msg, "error"))) \
-    JSON_HEDLEY_DIAGNOSTIC_POP
-#  else
-#    define JSON_HEDLEY_REQUIRE(expr) __attribute__((diagnose_if(!(expr), #expr, "error")))
-#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) __attribute__((diagnose_if(!(expr), msg, "error")))
-#  endif
-#else
-#  define JSON_HEDLEY_REQUIRE(expr)
-#  define JSON_HEDLEY_REQUIRE_MSG(expr,msg)
-#endif
-
-#if defined(JSON_HEDLEY_FLAGS)
-    #undef JSON_HEDLEY_FLAGS
-#endif
-#if JSON_HEDLEY_HAS_ATTRIBUTE(flag_enum) && (!defined(__cplusplus) || JSON_HEDLEY_HAS_WARNING("-Wbitfield-enum-conversion"))
-    #define JSON_HEDLEY_FLAGS __attribute__((__flag_enum__))
-#else
-    #define JSON_HEDLEY_FLAGS
-#endif
-
-#if defined(JSON_HEDLEY_FLAGS_CAST)
-    #undef JSON_HEDLEY_FLAGS_CAST
-#endif
-#if JSON_HEDLEY_INTEL_VERSION_CHECK(19,0,0)
-#  define JSON_HEDLEY_FLAGS_CAST(T, expr) (__extension__ ({ \
-        JSON_HEDLEY_DIAGNOSTIC_PUSH \
-        _Pragma("warning(disable:188)") \
-        ((T) (expr)); \
-        JSON_HEDLEY_DIAGNOSTIC_POP \
-    }))
-#else
-#  define JSON_HEDLEY_FLAGS_CAST(T, expr) JSON_HEDLEY_STATIC_CAST(T, expr)
-#endif
-
-#if defined(JSON_HEDLEY_EMPTY_BASES)
-    #undef JSON_HEDLEY_EMPTY_BASES
-#endif
-#if \
-    (JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,23918) && !JSON_HEDLEY_MSVC_VERSION_CHECK(20,0,0)) || \
-    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
-    #define JSON_HEDLEY_EMPTY_BASES __declspec(empty_bases)
-#else
-    #define JSON_HEDLEY_EMPTY_BASES
-#endif
-
-/* Remaining macros are deprecated. */
-
-#if defined(JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK)
-    #undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
-#endif
-#if defined(__clang__)
-    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) (0)
-#else
-    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
-#endif
-
-#if defined(JSON_HEDLEY_CLANG_HAS_ATTRIBUTE)
-    #undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
-#endif
-#define JSON_HEDLEY_CLANG_HAS_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE)
-    #undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
-#endif
-#define JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_BUILTIN)
-    #undef JSON_HEDLEY_CLANG_HAS_BUILTIN
-#endif
-#define JSON_HEDLEY_CLANG_HAS_BUILTIN(builtin) JSON_HEDLEY_HAS_BUILTIN(builtin)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_FEATURE)
-    #undef JSON_HEDLEY_CLANG_HAS_FEATURE
-#endif
-#define JSON_HEDLEY_CLANG_HAS_FEATURE(feature) JSON_HEDLEY_HAS_FEATURE(feature)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_EXTENSION)
-    #undef JSON_HEDLEY_CLANG_HAS_EXTENSION
-#endif
-#define JSON_HEDLEY_CLANG_HAS_EXTENSION(extension) JSON_HEDLEY_HAS_EXTENSION(extension)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE)
-    #undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
-#endif
-#define JSON_HEDLEY_CLANG_HAS_DECLSPEC_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute)
-
-#if defined(JSON_HEDLEY_CLANG_HAS_WARNING)
-    #undef JSON_HEDLEY_CLANG_HAS_WARNING
-#endif
-#define JSON_HEDLEY_CLANG_HAS_WARNING(warning) JSON_HEDLEY_HAS_WARNING(warning)
-
-#endif /* !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < X) */
-
-
-// This file contains all internal macro definitions (except those affecting ABI)
-// You MUST include macro_unscope.hpp at the end of json.hpp to undef all of them
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-// exclude unsupported compilers
-#if !defined(JSON_SKIP_UNSUPPORTED_COMPILER_CHECK)
-    #if defined(__clang__)
-        #if (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) < 30400
-            #error "unsupported Clang version - see https://github.com/nlohmann/json#supported-compilers"
-        #endif
-    #elif defined(__GNUC__) && !(defined(__ICC) || defined(__INTEL_COMPILER))
-        #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) < 40800
-            #error "unsupported GCC version - see https://github.com/nlohmann/json#supported-compilers"
-        #endif
-    #endif
-#endif
-
-// C++ language standard detection
-// if the user manually specified the used c++ version this is skipped
-#if !defined(JSON_HAS_CPP_20) && !defined(JSON_HAS_CPP_17) && !defined(JSON_HAS_CPP_14) && !defined(JSON_HAS_CPP_11)
-    #if (defined(__cplusplus) && __cplusplus >= 202002L) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
-        #define JSON_HAS_CPP_20
-        #define JSON_HAS_CPP_17
-        #define JSON_HAS_CPP_14
-    #elif (defined(__cplusplus) && __cplusplus >= 201703L) || (defined(_HAS_CXX17) && _HAS_CXX17 == 1) // fix for issue #464
-        #define JSON_HAS_CPP_17
-        #define JSON_HAS_CPP_14
-    #elif (defined(__cplusplus) && __cplusplus >= 201402L) || (defined(_HAS_CXX14) && _HAS_CXX14 == 1)
-        #define JSON_HAS_CPP_14
-    #endif
-    // the cpp 11 flag is always specified because it is the minimal required version
-    #define JSON_HAS_CPP_11
-#endif
-
-#ifdef __has_include
-    #if __has_include(<version>)
-        #include <version>
-    #endif
-#endif
-
-#if !defined(JSON_HAS_FILESYSTEM) && !defined(JSON_HAS_EXPERIMENTAL_FILESYSTEM)
-    #ifdef JSON_HAS_CPP_17
-        #if defined(__cpp_lib_filesystem)
-            #define JSON_HAS_FILESYSTEM 1
-        #elif defined(__cpp_lib_experimental_filesystem)
-            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
-        #elif !defined(__has_include)
-            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
-        #elif __has_include(<filesystem>)
-            #define JSON_HAS_FILESYSTEM 1
-        #elif __has_include(<experimental/filesystem>)
-            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
-        #endif
-
-        // std::filesystem does not work on MinGW GCC 8: https://sourceforge.net/p/mingw-w64/bugs/737/
-        #if defined(__MINGW32__) && defined(__GNUC__) && __GNUC__ == 8
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-
-        // no filesystem support before GCC 8: https://en.cppreference.com/w/cpp/compiler_support
-        #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-
-        // no filesystem support before Clang 7: https://en.cppreference.com/w/cpp/compiler_support
-        #if defined(__clang_major__) && __clang_major__ < 7
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-
-        // no filesystem support before MSVC 19.14: https://en.cppreference.com/w/cpp/compiler_support
-        #if defined(_MSC_VER) && _MSC_VER < 1914
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-
-        // no filesystem support before iOS 13
-        #if defined(__IPHONE_OS_VERSION_MIN_REQUIRED) && __IPHONE_OS_VERSION_MIN_REQUIRED < 130000
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-
-        // no filesystem support before macOS Catalina
-        #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 101500
-            #undef JSON_HAS_FILESYSTEM
-            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-        #endif
-    #endif
-#endif
-
-#ifndef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-    #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 0
-#endif
-
-#ifndef JSON_HAS_FILESYSTEM
-    #define JSON_HAS_FILESYSTEM 0
-#endif
-
-#ifndef JSON_HAS_THREE_WAY_COMPARISON
-    #if defined(__cpp_impl_three_way_comparison) && __cpp_impl_three_way_comparison >= 201907L \
-        && defined(__cpp_lib_three_way_comparison) && __cpp_lib_three_way_comparison >= 201907L
-        #define JSON_HAS_THREE_WAY_COMPARISON 1
-    #else
-        #define JSON_HAS_THREE_WAY_COMPARISON 0
-    #endif
-#endif
-
-#ifndef JSON_HAS_RANGES
-    // ranges header shipping in GCC 11.1.0 (released 2021-04-27) has syntax error
-    #if defined(__GLIBCXX__) && __GLIBCXX__ == 20210427
-        #define JSON_HAS_RANGES 0
-    #elif defined(__cpp_lib_ranges)
-        #define JSON_HAS_RANGES 1
-    #else
-        #define JSON_HAS_RANGES 0
-    #endif
-#endif
-
-#ifdef JSON_HAS_CPP_17
-    #define JSON_INLINE_VARIABLE inline
-#else
-    #define JSON_INLINE_VARIABLE
-#endif
-
-#if JSON_HEDLEY_HAS_ATTRIBUTE(no_unique_address)
-    #define JSON_NO_UNIQUE_ADDRESS [[no_unique_address]]
-#else
-    #define JSON_NO_UNIQUE_ADDRESS
-#endif
-
-// disable documentation warnings on clang
-#if defined(__clang__)
-    #pragma clang diagnostic push
-    #pragma clang diagnostic ignored "-Wdocumentation"
-    #pragma clang diagnostic ignored "-Wdocumentation-unknown-command"
-#endif
-
-// allow disabling exceptions
-#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)) && !defined(JSON_NOEXCEPTION)
-    #define JSON_THROW(exception) throw exception
-    #define JSON_TRY try
-    #define JSON_CATCH(exception) catch(exception)
-    #define JSON_INTERNAL_CATCH(exception) catch(exception)
-#else
-    #include <cstdlib>
-    #define JSON_THROW(exception) std::abort()
-    #define JSON_TRY if(true)
-    #define JSON_CATCH(exception) if(false)
-    #define JSON_INTERNAL_CATCH(exception) if(false)
-#endif
-
-// override exception macros
-#if defined(JSON_THROW_USER)
-    #undef JSON_THROW
-    #define JSON_THROW JSON_THROW_USER
-#endif
-#if defined(JSON_TRY_USER)
-    #undef JSON_TRY
-    #define JSON_TRY JSON_TRY_USER
-#endif
-#if defined(JSON_CATCH_USER)
-    #undef JSON_CATCH
-    #define JSON_CATCH JSON_CATCH_USER
-    #undef JSON_INTERNAL_CATCH
-    #define JSON_INTERNAL_CATCH JSON_CATCH_USER
-#endif
-#if defined(JSON_INTERNAL_CATCH_USER)
-    #undef JSON_INTERNAL_CATCH
-    #define JSON_INTERNAL_CATCH JSON_INTERNAL_CATCH_USER
-#endif
-
-// allow overriding assert
-#if !defined(JSON_ASSERT)
-    #include <cassert> // assert
-    #define JSON_ASSERT(x) assert(x)
-#endif
-
-// allow to access some private functions (needed by the test suite)
-#if defined(JSON_TESTS_PRIVATE)
-    #define JSON_PRIVATE_UNLESS_TESTED public
-#else
-    #define JSON_PRIVATE_UNLESS_TESTED private
-#endif
-
-/*!
-@brief macro to briefly define a mapping between an enum and JSON
-@def NLOHMANN_JSON_SERIALIZE_ENUM
-@since version 3.4.0
-*/
-#define NLOHMANN_JSON_SERIALIZE_ENUM(ENUM_TYPE, ...)                                            \
-    template<typename BasicJsonType>                                                            \
-    inline void to_json(BasicJsonType& j, const ENUM_TYPE& e)                                   \
-    {                                                                                           \
-        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
-        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
-        auto it = std::find_if(std::begin(m), std::end(m),                                      \
-                               [e](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool  \
-        {                                                                                       \
-            return ej_pair.first == e;                                                          \
-        });                                                                                     \
-        j = ((it != std::end(m)) ? it : std::begin(m))->second;                                 \
-    }                                                                                           \
-    template<typename BasicJsonType>                                                            \
-    inline void from_json(const BasicJsonType& j, ENUM_TYPE& e)                                 \
-    {                                                                                           \
-        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
-        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
-        auto it = std::find_if(std::begin(m), std::end(m),                                      \
-                               [&j](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool \
-        {                                                                                       \
-            return ej_pair.second == j;                                                         \
-        });                                                                                     \
-        e = ((it != std::end(m)) ? it : std::begin(m))->first;                                  \
-    }
-
-// Ugly macros to avoid uglier copy-paste when specializing basic_json. They
-// may be removed in the future once the class is split.
-
-#define NLOHMANN_BASIC_JSON_TPL_DECLARATION                                \
-    template<template<typename, typename, typename...> class ObjectType,   \
-             template<typename, typename...> class ArrayType,              \
-             class StringType, class BooleanType, class NumberIntegerType, \
-             class NumberUnsignedType, class NumberFloatType,              \
-             template<typename> class AllocatorType,                       \
-             template<typename, typename = void> class JSONSerializer,     \
-             class BinaryType>
-
-#define NLOHMANN_BASIC_JSON_TPL                                            \
-    basic_json<ObjectType, ArrayType, StringType, BooleanType,             \
-    NumberIntegerType, NumberUnsignedType, NumberFloatType,                \
-    AllocatorType, JSONSerializer, BinaryType>
-
-// Macros to simplify conversion from/to types
-
-#define NLOHMANN_JSON_EXPAND( x ) x
-#define NLOHMANN_JSON_GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, _61, _62, _63, _64, NAME,...) NAME
-#define NLOHMANN_JSON_PASTE(...) NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_GET_MACRO(__VA_ARGS__, \
-        NLOHMANN_JSON_PASTE64, \
-        NLOHMANN_JSON_PASTE63, \
-        NLOHMANN_JSON_PASTE62, \
-        NLOHMANN_JSON_PASTE61, \
-        NLOHMANN_JSON_PASTE60, \
-        NLOHMANN_JSON_PASTE59, \
-        NLOHMANN_JSON_PASTE58, \
-        NLOHMANN_JSON_PASTE57, \
-        NLOHMANN_JSON_PASTE56, \
-        NLOHMANN_JSON_PASTE55, \
-        NLOHMANN_JSON_PASTE54, \
-        NLOHMANN_JSON_PASTE53, \
-        NLOHMANN_JSON_PASTE52, \
-        NLOHMANN_JSON_PASTE51, \
-        NLOHMANN_JSON_PASTE50, \
-        NLOHMANN_JSON_PASTE49, \
-        NLOHMANN_JSON_PASTE48, \
-        NLOHMANN_JSON_PASTE47, \
-        NLOHMANN_JSON_PASTE46, \
-        NLOHMANN_JSON_PASTE45, \
-        NLOHMANN_JSON_PASTE44, \
-        NLOHMANN_JSON_PASTE43, \
-        NLOHMANN_JSON_PASTE42, \
-        NLOHMANN_JSON_PASTE41, \
-        NLOHMANN_JSON_PASTE40, \
-        NLOHMANN_JSON_PASTE39, \
-        NLOHMANN_JSON_PASTE38, \
-        NLOHMANN_JSON_PASTE37, \
-        NLOHMANN_JSON_PASTE36, \
-        NLOHMANN_JSON_PASTE35, \
-        NLOHMANN_JSON_PASTE34, \
-        NLOHMANN_JSON_PASTE33, \
-        NLOHMANN_JSON_PASTE32, \
-        NLOHMANN_JSON_PASTE31, \
-        NLOHMANN_JSON_PASTE30, \
-        NLOHMANN_JSON_PASTE29, \
-        NLOHMANN_JSON_PASTE28, \
-        NLOHMANN_JSON_PASTE27, \
-        NLOHMANN_JSON_PASTE26, \
-        NLOHMANN_JSON_PASTE25, \
-        NLOHMANN_JSON_PASTE24, \
-        NLOHMANN_JSON_PASTE23, \
-        NLOHMANN_JSON_PASTE22, \
-        NLOHMANN_JSON_PASTE21, \
-        NLOHMANN_JSON_PASTE20, \
-        NLOHMANN_JSON_PASTE19, \
-        NLOHMANN_JSON_PASTE18, \
-        NLOHMANN_JSON_PASTE17, \
-        NLOHMANN_JSON_PASTE16, \
-        NLOHMANN_JSON_PASTE15, \
-        NLOHMANN_JSON_PASTE14, \
-        NLOHMANN_JSON_PASTE13, \
-        NLOHMANN_JSON_PASTE12, \
-        NLOHMANN_JSON_PASTE11, \
-        NLOHMANN_JSON_PASTE10, \
-        NLOHMANN_JSON_PASTE9, \
-        NLOHMANN_JSON_PASTE8, \
-        NLOHMANN_JSON_PASTE7, \
-        NLOHMANN_JSON_PASTE6, \
-        NLOHMANN_JSON_PASTE5, \
-        NLOHMANN_JSON_PASTE4, \
-        NLOHMANN_JSON_PASTE3, \
-        NLOHMANN_JSON_PASTE2, \
-        NLOHMANN_JSON_PASTE1)(__VA_ARGS__))
-#define NLOHMANN_JSON_PASTE2(func, v1) func(v1)
-#define NLOHMANN_JSON_PASTE3(func, v1, v2) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE2(func, v2)
-#define NLOHMANN_JSON_PASTE4(func, v1, v2, v3) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE3(func, v2, v3)
-#define NLOHMANN_JSON_PASTE5(func, v1, v2, v3, v4) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE4(func, v2, v3, v4)
-#define NLOHMANN_JSON_PASTE6(func, v1, v2, v3, v4, v5) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE5(func, v2, v3, v4, v5)
-#define NLOHMANN_JSON_PASTE7(func, v1, v2, v3, v4, v5, v6) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE6(func, v2, v3, v4, v5, v6)
-#define NLOHMANN_JSON_PASTE8(func, v1, v2, v3, v4, v5, v6, v7) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE7(func, v2, v3, v4, v5, v6, v7)
-#define NLOHMANN_JSON_PASTE9(func, v1, v2, v3, v4, v5, v6, v7, v8) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE8(func, v2, v3, v4, v5, v6, v7, v8)
-#define NLOHMANN_JSON_PASTE10(func, v1, v2, v3, v4, v5, v6, v7, v8, v9) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE9(func, v2, v3, v4, v5, v6, v7, v8, v9)
-#define NLOHMANN_JSON_PASTE11(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE10(func, v2, v3, v4, v5, v6, v7, v8, v9, v10)
-#define NLOHMANN_JSON_PASTE12(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE11(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11)
-#define NLOHMANN_JSON_PASTE13(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE12(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12)
-#define NLOHMANN_JSON_PASTE14(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE13(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13)
-#define NLOHMANN_JSON_PASTE15(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE14(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14)
-#define NLOHMANN_JSON_PASTE16(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE15(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15)
-#define NLOHMANN_JSON_PASTE17(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE16(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16)
-#define NLOHMANN_JSON_PASTE18(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17)
-#define NLOHMANN_JSON_PASTE19(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18)
-#define NLOHMANN_JSON_PASTE20(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19)
-#define NLOHMANN_JSON_PASTE21(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20)
-#define NLOHMANN_JSON_PASTE22(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21)
-#define NLOHMANN_JSON_PASTE23(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22)
-#define NLOHMANN_JSON_PASTE24(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23)
-#define NLOHMANN_JSON_PASTE25(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24)
-#define NLOHMANN_JSON_PASTE26(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25)
-#define NLOHMANN_JSON_PASTE27(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE26(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26)
-#define NLOHMANN_JSON_PASTE28(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE27(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27)
-#define NLOHMANN_JSON_PASTE29(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE28(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28)
-#define NLOHMANN_JSON_PASTE30(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29)
-#define NLOHMANN_JSON_PASTE31(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30)
-#define NLOHMANN_JSON_PASTE32(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31)
-#define NLOHMANN_JSON_PASTE33(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32)
-#define NLOHMANN_JSON_PASTE34(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33)
-#define NLOHMANN_JSON_PASTE35(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34)
-#define NLOHMANN_JSON_PASTE36(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35)
-#define NLOHMANN_JSON_PASTE37(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36)
-#define NLOHMANN_JSON_PASTE38(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37)
-#define NLOHMANN_JSON_PASTE39(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE38(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38)
-#define NLOHMANN_JSON_PASTE40(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE39(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39)
-#define NLOHMANN_JSON_PASTE41(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE40(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40)
-#define NLOHMANN_JSON_PASTE42(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE41(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41)
-#define NLOHMANN_JSON_PASTE43(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE42(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42)
-#define NLOHMANN_JSON_PASTE44(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE43(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43)
-#define NLOHMANN_JSON_PASTE45(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE44(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44)
-#define NLOHMANN_JSON_PASTE46(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE45(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45)
-#define NLOHMANN_JSON_PASTE47(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE46(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46)
-#define NLOHMANN_JSON_PASTE48(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE47(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47)
-#define NLOHMANN_JSON_PASTE49(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE48(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48)
-#define NLOHMANN_JSON_PASTE50(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE49(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49)
-#define NLOHMANN_JSON_PASTE51(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE50(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50)
-#define NLOHMANN_JSON_PASTE52(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE51(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51)
-#define NLOHMANN_JSON_PASTE53(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE52(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52)
-#define NLOHMANN_JSON_PASTE54(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE53(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53)
-#define NLOHMANN_JSON_PASTE55(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE54(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54)
-#define NLOHMANN_JSON_PASTE56(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE55(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55)
-#define NLOHMANN_JSON_PASTE57(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE56(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56)
-#define NLOHMANN_JSON_PASTE58(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE57(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57)
-#define NLOHMANN_JSON_PASTE59(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE58(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58)
-#define NLOHMANN_JSON_PASTE60(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE59(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59)
-#define NLOHMANN_JSON_PASTE61(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE60(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60)
-#define NLOHMANN_JSON_PASTE62(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE61(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61)
-#define NLOHMANN_JSON_PASTE63(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE62(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62)
-#define NLOHMANN_JSON_PASTE64(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE63(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63)
-
-#define NLOHMANN_JSON_TO(v1) nlohmann_json_j[#v1] = nlohmann_json_t.v1;
-#define NLOHMANN_JSON_FROM(v1) nlohmann_json_j.at(#v1).get_to(nlohmann_json_t.v1);
-#define NLOHMANN_JSON_FROM_WITH_DEFAULT(v1) nlohmann_json_t.v1 = nlohmann_json_j.value(#v1, nlohmann_json_default_obj.v1);
-
-/*!
-@brief macro
-@def NLOHMANN_DEFINE_TYPE_INTRUSIVE
-@since version 3.9.0
-*/
-#define NLOHMANN_DEFINE_TYPE_INTRUSIVE(Type, ...)  \
-    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
-    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
-
-#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
-    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
-    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
-
-/*!
-@brief macro
-@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE
-@since version 3.9.0
-*/
-#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE(Type, ...)  \
-    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
-    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
-
-#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
-    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
-    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
-
-
-// inspired from https://stackoverflow.com/a/26745591
-// allows to call any std function as if (e.g. with begin):
-// using std::begin; begin(x);
-//
-// it allows using the detected idiom to retrieve the return type
-// of such an expression
-#define NLOHMANN_CAN_CALL_STD_FUNC_IMPL(std_name)                                 \
-    namespace detail {                                                            \
-    using std::std_name;                                                          \
-    \
-    template<typename... T>                                                       \
-    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
-    }                                                                             \
-    \
-    namespace detail2 {                                                           \
-    struct std_name##_tag                                                         \
-    {                                                                             \
-    };                                                                            \
-    \
-    template<typename... T>                                                       \
-    std_name##_tag std_name(T&&...);                                              \
-    \
-    template<typename... T>                                                       \
-    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
-    \
-    template<typename... T>                                                       \
-    struct would_call_std_##std_name                                              \
-    {                                                                             \
-        static constexpr auto const value = ::nlohmann::detail::                  \
-                                            is_detected_exact<std_name##_tag, result_of_##std_name, T...>::value; \
-    };                                                                            \
-    } /* namespace detail2 */ \
-    \
-    template<typename... T>                                                       \
-    struct would_call_std_##std_name : detail2::would_call_std_##std_name<T...>   \
-    {                                                                             \
-    }
-
-#ifndef JSON_USE_IMPLICIT_CONVERSIONS
-    #define JSON_USE_IMPLICIT_CONVERSIONS 1
-#endif
-
-#if JSON_USE_IMPLICIT_CONVERSIONS
-    #define JSON_EXPLICIT
-#else
-    #define JSON_EXPLICIT explicit
-#endif
-
-#ifndef JSON_DISABLE_ENUM_SERIALIZATION
-    #define JSON_DISABLE_ENUM_SERIALIZATION 0
-#endif
-
-#ifndef JSON_USE_GLOBAL_UDLS
-    #define JSON_USE_GLOBAL_UDLS 1
-#endif
-
-#if JSON_HAS_THREE_WAY_COMPARISON
-    #include <compare> // partial_ordering
-#endif
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-///////////////////////////
-// JSON type enumeration //
-///////////////////////////
-
-/*!
-@brief the JSON type enumeration
-
-This enumeration collects the different JSON types. It is internally used to
-distinguish the stored values, and the functions @ref basic_json::is_null(),
-@ref basic_json::is_object(), @ref basic_json::is_array(),
-@ref basic_json::is_string(), @ref basic_json::is_boolean(),
-@ref basic_json::is_number() (with @ref basic_json::is_number_integer(),
-@ref basic_json::is_number_unsigned(), and @ref basic_json::is_number_float()),
-@ref basic_json::is_discarded(), @ref basic_json::is_primitive(), and
-@ref basic_json::is_structured() rely on it.
-
-@note There are three enumeration entries (number_integer, number_unsigned, and
-number_float), because the library distinguishes these three types for numbers:
-@ref basic_json::number_unsigned_t is used for unsigned integers,
-@ref basic_json::number_integer_t is used for signed integers, and
-@ref basic_json::number_float_t is used for floating-point numbers or to
-approximate integers which do not fit in the limits of their respective type.
-
-@sa see @ref basic_json::basic_json(const value_t value_type) -- create a JSON
-value with the default value for a given type
-
-@since version 1.0.0
-*/
-enum class value_t : std::uint8_t
-{
-    null,             ///< null value
-    object,           ///< object (unordered set of name/value pairs)
-    array,            ///< array (ordered collection of values)
-    string,           ///< string value
-    boolean,          ///< boolean value
-    number_integer,   ///< number value (signed integer)
-    number_unsigned,  ///< number value (unsigned integer)
-    number_float,     ///< number value (floating-point)
-    binary,           ///< binary array (ordered collection of bytes)
-    discarded         ///< discarded by the parser callback function
-};
-
-/*!
-@brief comparison operator for JSON types
-
-Returns an ordering that is similar to Python:
-- order: null < boolean < number < object < array < string < binary
-- furthermore, each type is not smaller than itself
-- discarded values are not comparable
-- binary is represented as a b"" string in python and directly comparable to a
-  string; however, making a binary array directly comparable with a string would
-  be surprising behavior in a JSON file.
-
-@since version 1.0.0
-*/
-#if JSON_HAS_THREE_WAY_COMPARISON
-    inline std::partial_ordering operator<=>(const value_t lhs, const value_t rhs) noexcept // *NOPAD*
-#else
-    inline bool operator<(const value_t lhs, const value_t rhs) noexcept
-#endif
-{
-    static constexpr std::array<std::uint8_t, 9> order = {{
-            0 /* null */, 3 /* object */, 4 /* array */, 5 /* string */,
-            1 /* boolean */, 2 /* integer */, 2 /* unsigned */, 2 /* float */,
-            6 /* binary */
-        }
-    };
-
-    const auto l_index = static_cast<std::size_t>(lhs);
-    const auto r_index = static_cast<std::size_t>(rhs);
-#if JSON_HAS_THREE_WAY_COMPARISON
-    if (l_index < order.size() && r_index < order.size())
-    {
-        return order[l_index] <=> order[r_index]; // *NOPAD*
-    }
-    return std::partial_ordering::unordered;
-#else
-    return l_index < order.size() && r_index < order.size() && order[l_index] < order[r_index];
-#endif
-}
-
-// GCC selects the built-in operator< over an operator rewritten from
-// a user-defined spaceship operator
-// Clang, MSVC, and ICC select the rewritten candidate
-// (see GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105200)
-#if JSON_HAS_THREE_WAY_COMPARISON && defined(__GNUC__)
-inline bool operator<(const value_t lhs, const value_t rhs) noexcept
-{
-    return std::is_lt(lhs <=> rhs); // *NOPAD*
-}
-#endif
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/string_escape.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/*!
-@brief replace all occurrences of a substring by another string
-
-@param[in,out] s  the string to manipulate; changed so that all
-               occurrences of @a f are replaced with @a t
-@param[in]     f  the substring to replace with @a t
-@param[in]     t  the string to replace @a f
-
-@pre The search string @a f must not be empty. **This precondition is
-enforced with an assertion.**
-
-@since version 2.0.0
-*/
-template<typename StringType>
-inline void replace_substring(StringType& s, const StringType& f,
-                              const StringType& t)
-{
-    JSON_ASSERT(!f.empty());
-    for (auto pos = s.find(f);                // find first occurrence of f
-            pos != StringType::npos;          // make sure f was found
-            s.replace(pos, f.size(), t),      // replace with t, and
-            pos = s.find(f, pos + t.size()))  // find next occurrence of f
-    {}
-}
-
-/*!
- * @brief string escaping as described in RFC 6901 (Sect. 4)
- * @param[in] s string to escape
- * @return    escaped string
- *
- * Note the order of escaping "~" to "~0" and "/" to "~1" is important.
- */
-template<typename StringType>
-inline StringType escape(StringType s)
-{
-    replace_substring(s, StringType{"~"}, StringType{"~0"});
-    replace_substring(s, StringType{"/"}, StringType{"~1"});
-    return s;
-}
-
-/*!
- * @brief string unescaping as described in RFC 6901 (Sect. 4)
- * @param[in] s string to unescape
- * @return    unescaped string
- *
- * Note the order of escaping "~1" to "/" and "~0" to "~" is important.
- */
-template<typename StringType>
-static void unescape(StringType& s)
-{
-    replace_substring(s, StringType{"~1"}, StringType{"/"});
-    replace_substring(s, StringType{"~0"}, StringType{"~"});
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/input/position_t.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef> // size_t
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/// struct to capture the start position of the current token
-struct position_t
-{
-    /// the total number of characters read
-    std::size_t chars_read_total = 0;
-    /// the number of characters read in the current line
-    std::size_t chars_read_current_line = 0;
-    /// the number of lines read
-    std::size_t lines_read = 0;
-
-    /// conversion to size_t to preserve SAX interface
-    constexpr operator size_t() const
-    {
-        return chars_read_total;
-    }
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-FileCopyrightText: 2018 The Abseil Authors
-// SPDX-License-Identifier: MIT
-
-
-
-#include <array> // array
-#include <cstddef> // size_t
-#include <type_traits> // conditional, enable_if, false_type, integral_constant, is_constructible, is_integral, is_same, remove_cv, remove_reference, true_type
-#include <utility> // index_sequence, make_index_sequence, index_sequence_for
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename T>
-using uncvref_t = typename std::remove_cv<typename std::remove_reference<T>::type>::type;
-
-#ifdef JSON_HAS_CPP_14
-
-// the following utilities are natively available in C++14
-using std::enable_if_t;
-using std::index_sequence;
-using std::make_index_sequence;
-using std::index_sequence_for;
-
-#else
-
-// alias templates to reduce boilerplate
-template<bool B, typename T = void>
-using enable_if_t = typename std::enable_if<B, T>::type;
-
-// The following code is taken from https://github.com/abseil/abseil-cpp/blob/10cb35e459f5ecca5b2ff107635da0bfa41011b4/absl/utility/utility.h
-// which is part of Google Abseil (https://github.com/abseil/abseil-cpp), licensed under the Apache License 2.0.
-
-//// START OF CODE FROM GOOGLE ABSEIL
-
-// integer_sequence
-//
-// Class template representing a compile-time integer sequence. An instantiation
-// of `integer_sequence<T, Ints...>` has a sequence of integers encoded in its
-// type through its template arguments (which is a common need when
-// working with C++11 variadic templates). `absl::integer_sequence` is designed
-// to be a drop-in replacement for C++14's `std::integer_sequence`.
-//
-// Example:
-//
-//   template< class T, T... Ints >
-//   void user_function(integer_sequence<T, Ints...>);
-//
-//   int main()
-//   {
-//     // user_function's `T` will be deduced to `int` and `Ints...`
-//     // will be deduced to `0, 1, 2, 3, 4`.
-//     user_function(make_integer_sequence<int, 5>());
-//   }
-template <typename T, T... Ints>
-struct integer_sequence
-{
-    using value_type = T;
-    static constexpr std::size_t size() noexcept
-    {
-        return sizeof...(Ints);
-    }
-};
-
-// index_sequence
-//
-// A helper template for an `integer_sequence` of `size_t`,
-// `absl::index_sequence` is designed to be a drop-in replacement for C++14's
-// `std::index_sequence`.
-template <size_t... Ints>
-using index_sequence = integer_sequence<size_t, Ints...>;
-
-namespace utility_internal
-{
-
-template <typename Seq, size_t SeqSize, size_t Rem>
-struct Extend;
-
-// Note that SeqSize == sizeof...(Ints). It's passed explicitly for efficiency.
-template <typename T, T... Ints, size_t SeqSize>
-struct Extend<integer_sequence<T, Ints...>, SeqSize, 0>
-{
-    using type = integer_sequence < T, Ints..., (Ints + SeqSize)... >;
-};
-
-template <typename T, T... Ints, size_t SeqSize>
-struct Extend<integer_sequence<T, Ints...>, SeqSize, 1>
-{
-    using type = integer_sequence < T, Ints..., (Ints + SeqSize)..., 2 * SeqSize >;
-};
-
-// Recursion helper for 'make_integer_sequence<T, N>'.
-// 'Gen<T, N>::type' is an alias for 'integer_sequence<T, 0, 1, ... N-1>'.
-template <typename T, size_t N>
-struct Gen
-{
-    using type =
-        typename Extend < typename Gen < T, N / 2 >::type, N / 2, N % 2 >::type;
-};
-
-template <typename T>
-struct Gen<T, 0>
-{
-    using type = integer_sequence<T>;
-};
-
-}  // namespace utility_internal
-
-// Compile-time sequences of integers
-
-// make_integer_sequence
-//
-// This template alias is equivalent to
-// `integer_sequence<int, 0, 1, ..., N-1>`, and is designed to be a drop-in
-// replacement for C++14's `std::make_integer_sequence`.
-template <typename T, T N>
-using make_integer_sequence = typename utility_internal::Gen<T, N>::type;
-
-// make_index_sequence
-//
-// This template alias is equivalent to `index_sequence<0, 1, ..., N-1>`,
-// and is designed to be a drop-in replacement for C++14's
-// `std::make_index_sequence`.
-template <size_t N>
-using make_index_sequence = make_integer_sequence<size_t, N>;
-
-// index_sequence_for
-//
-// Converts a typename pack into an index sequence of the same length, and
-// is designed to be a drop-in replacement for C++14's
-// `std::index_sequence_for()`
-template <typename... Ts>
-using index_sequence_for = make_index_sequence<sizeof...(Ts)>;
-
-//// END OF CODE FROM GOOGLE ABSEIL
-
-#endif
-
-// dispatch utility (taken from ranges-v3)
-template<unsigned N> struct priority_tag : priority_tag < N - 1 > {};
-template<> struct priority_tag<0> {};
-
-// taken from ranges-v3
-template<typename T>
-struct static_const
-{
-    static JSON_INLINE_VARIABLE constexpr T value{};
-};
-
-#ifndef JSON_HAS_CPP_17
-    template<typename T>
-    constexpr T static_const<T>::value;
-#endif
-
-template<typename T, typename... Args>
-inline constexpr std::array<T, sizeof...(Args)> make_array(Args&& ... args)
-{
-    return std::array<T, sizeof...(Args)> {{static_cast<T>(std::forward<Args>(args))...}};
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <limits> // numeric_limits
-#include <type_traits> // false_type, is_constructible, is_integral, is_same, true_type
-#include <utility> // declval
-#include <tuple> // tuple
-
-// #include <nlohmann/detail/iterators/iterator_traits.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <iterator> // random_access_iterator_tag
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/meta/void_t.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename It, typename = void>
-struct iterator_types {};
-
-template<typename It>
-struct iterator_types <
-    It,
-    void_t<typename It::difference_type, typename It::value_type, typename It::pointer,
-    typename It::reference, typename It::iterator_category >>
-{
-    using difference_type = typename It::difference_type;
-    using value_type = typename It::value_type;
-    using pointer = typename It::pointer;
-    using reference = typename It::reference;
-    using iterator_category = typename It::iterator_category;
-};
-
-// This is required as some compilers implement std::iterator_traits in a way that
-// doesn't work with SFINAE. See https://github.com/nlohmann/json/issues/1341.
-template<typename T, typename = void>
-struct iterator_traits
-{
-};
-
-template<typename T>
-struct iterator_traits < T, enable_if_t < !std::is_pointer<T>::value >>
-            : iterator_types<T>
-{
-};
-
-template<typename T>
-struct iterator_traits<T*, enable_if_t<std::is_object<T>::value>>
-{
-    using iterator_category = std::random_access_iterator_tag;
-    using value_type = T;
-    using difference_type = ptrdiff_t;
-    using pointer = T*;
-    using reference = T&;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/call_std/begin.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-NLOHMANN_CAN_CALL_STD_FUNC_IMPL(begin);
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/call_std/end.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-NLOHMANN_CAN_CALL_STD_FUNC_IMPL(end);
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/detected.hpp>
-
-// #include <nlohmann/json_fwd.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_
-    #define INCLUDE_NLOHMANN_JSON_FWD_HPP_
-
-    #include <cstdint> // int64_t, uint64_t
-    #include <map> // map
-    #include <memory> // allocator
-    #include <string> // string
-    #include <vector> // vector
-
-    // #include <nlohmann/detail/abi_macros.hpp>
-
-
-    /*!
-    @brief namespace for Niels Lohmann
-    @see https://github.com/nlohmann
-    @since version 1.0.0
-    */
-    NLOHMANN_JSON_NAMESPACE_BEGIN
-
-    /*!
-    @brief default JSONSerializer template argument
-
-    This serializer ignores the template arguments and uses ADL
-    ([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl))
-    for serialization.
-    */
-    template<typename T = void, typename SFINAE = void>
-    struct adl_serializer;
-
-    /// a class to store JSON values
-    /// @sa https://json.nlohmann.me/api/basic_json/
-    template<template<typename U, typename V, typename... Args> class ObjectType =
-    std::map,
-    template<typename U, typename... Args> class ArrayType = std::vector,
-    class StringType = std::string, class BooleanType = bool,
-    class NumberIntegerType = std::int64_t,
-    class NumberUnsignedType = std::uint64_t,
-    class NumberFloatType = double,
-    template<typename U> class AllocatorType = std::allocator,
-    template<typename T, typename SFINAE = void> class JSONSerializer =
-    adl_serializer,
-    class BinaryType = std::vector<std::uint8_t>>
-    class basic_json;
-
-    /// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
-    /// @sa https://json.nlohmann.me/api/json_pointer/
-    template<typename RefStringType>
-    class json_pointer;
-
-    /*!
-    @brief default specialization
-    @sa https://json.nlohmann.me/api/json/
-    */
-    using json = basic_json<>;
-
-    /// @brief a minimal map-like container that preserves insertion order
-    /// @sa https://json.nlohmann.me/api/ordered_map/
-    template<class Key, class T, class IgnoredLess, class Allocator>
-    struct ordered_map;
-
-    /// @brief specialization that maintains the insertion order of object keys
-    /// @sa https://json.nlohmann.me/api/ordered_json/
-    using ordered_json = basic_json<nlohmann::ordered_map>;
-
-    NLOHMANN_JSON_NAMESPACE_END
-
-#endif  // INCLUDE_NLOHMANN_JSON_FWD_HPP_
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-/*!
-@brief detail namespace with internal helper functions
-
-This namespace collects functions that should not be exposed,
-implementations of some @ref basic_json methods, and meta-programming helpers.
-
-@since version 2.1.0
-*/
-namespace detail
-{
-
-/////////////
-// helpers //
-/////////////
-
-// Note to maintainers:
-//
-// Every trait in this file expects a non CV-qualified type.
-// The only exceptions are in the 'aliases for detected' section
-// (i.e. those of the form: decltype(T::member_function(std::declval<T>())))
-//
-// In this case, T has to be properly CV-qualified to constraint the function arguments
-// (e.g. to_json(BasicJsonType&, const T&))
-
-template<typename> struct is_basic_json : std::false_type {};
-
-NLOHMANN_BASIC_JSON_TPL_DECLARATION
-struct is_basic_json<NLOHMANN_BASIC_JSON_TPL> : std::true_type {};
-
-// used by exceptions create() member functions
-// true_type for pointer to possibly cv-qualified basic_json or std::nullptr_t
-// false_type otherwise
-template<typename BasicJsonContext>
-struct is_basic_json_context :
-    std::integral_constant < bool,
-    is_basic_json<typename std::remove_cv<typename std::remove_pointer<BasicJsonContext>::type>::type>::value
-    || std::is_same<BasicJsonContext, std::nullptr_t>::value >
-{};
-
-//////////////////////
-// json_ref helpers //
-//////////////////////
-
-template<typename>
-class json_ref;
-
-template<typename>
-struct is_json_ref : std::false_type {};
-
-template<typename T>
-struct is_json_ref<json_ref<T>> : std::true_type {};
-
-//////////////////////////
-// aliases for detected //
-//////////////////////////
-
-template<typename T>
-using mapped_type_t = typename T::mapped_type;
-
-template<typename T>
-using key_type_t = typename T::key_type;
-
-template<typename T>
-using value_type_t = typename T::value_type;
-
-template<typename T>
-using difference_type_t = typename T::difference_type;
-
-template<typename T>
-using pointer_t = typename T::pointer;
-
-template<typename T>
-using reference_t = typename T::reference;
-
-template<typename T>
-using iterator_category_t = typename T::iterator_category;
-
-template<typename T, typename... Args>
-using to_json_function = decltype(T::to_json(std::declval<Args>()...));
-
-template<typename T, typename... Args>
-using from_json_function = decltype(T::from_json(std::declval<Args>()...));
-
-template<typename T, typename U>
-using get_template_function = decltype(std::declval<T>().template get<U>());
-
-// trait checking if JSONSerializer<T>::from_json(json const&, udt&) exists
-template<typename BasicJsonType, typename T, typename = void>
-struct has_from_json : std::false_type {};
-
-// trait checking if j.get<T> is valid
-// use this trait instead of std::is_constructible or std::is_convertible,
-// both rely on, or make use of implicit conversions, and thus fail when T
-// has several constructors/operator= (see https://github.com/nlohmann/json/issues/958)
-template <typename BasicJsonType, typename T>
-struct is_getable
-{
-    static constexpr bool value = is_detected<get_template_function, const BasicJsonType&, T>::value;
-};
-
-template<typename BasicJsonType, typename T>
-struct has_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
-{
-    using serializer = typename BasicJsonType::template json_serializer<T, void>;
-
-    static constexpr bool value =
-        is_detected_exact<void, from_json_function, serializer,
-        const BasicJsonType&, T&>::value;
-};
-
-// This trait checks if JSONSerializer<T>::from_json(json const&) exists
-// this overload is used for non-default-constructible user-defined-types
-template<typename BasicJsonType, typename T, typename = void>
-struct has_non_default_from_json : std::false_type {};
-
-template<typename BasicJsonType, typename T>
-struct has_non_default_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
-{
-    using serializer = typename BasicJsonType::template json_serializer<T, void>;
-
-    static constexpr bool value =
-        is_detected_exact<T, from_json_function, serializer,
-        const BasicJsonType&>::value;
-};
-
-// This trait checks if BasicJsonType::json_serializer<T>::to_json exists
-// Do not evaluate the trait when T is a basic_json type, to avoid template instantiation infinite recursion.
-template<typename BasicJsonType, typename T, typename = void>
-struct has_to_json : std::false_type {};
-
-template<typename BasicJsonType, typename T>
-struct has_to_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
-{
-    using serializer = typename BasicJsonType::template json_serializer<T, void>;
-
-    static constexpr bool value =
-        is_detected_exact<void, to_json_function, serializer, BasicJsonType&,
-        T>::value;
-};
-
-template<typename T>
-using detect_key_compare = typename T::key_compare;
-
-template<typename T>
-struct has_key_compare : std::integral_constant<bool, is_detected<detect_key_compare, T>::value> {};
-
-// obtains the actual object key comparator
-template<typename BasicJsonType>
-struct actual_object_comparator
-{
-    using object_t = typename BasicJsonType::object_t;
-    using object_comparator_t = typename BasicJsonType::default_object_comparator_t;
-    using type = typename std::conditional < has_key_compare<object_t>::value,
-          typename object_t::key_compare, object_comparator_t>::type;
-};
-
-template<typename BasicJsonType>
-using actual_object_comparator_t = typename actual_object_comparator<BasicJsonType>::type;
-
-///////////////////
-// is_ functions //
-///////////////////
-
-// https://en.cppreference.com/w/cpp/types/conjunction
-template<class...> struct conjunction : std::true_type { };
-template<class B> struct conjunction<B> : B { };
-template<class B, class... Bn>
-struct conjunction<B, Bn...>
-: std::conditional<static_cast<bool>(B::value), conjunction<Bn...>, B>::type {};
-
-// https://en.cppreference.com/w/cpp/types/negation
-template<class B> struct negation : std::integral_constant < bool, !B::value > { };
-
-// Reimplementation of is_constructible and is_default_constructible, due to them being broken for
-// std::pair and std::tuple until LWG 2367 fix (see https://cplusplus.github.io/LWG/lwg-defects.html#2367).
-// This causes compile errors in e.g. clang 3.5 or gcc 4.9.
-template <typename T>
-struct is_default_constructible : std::is_default_constructible<T> {};
-
-template <typename T1, typename T2>
-struct is_default_constructible<std::pair<T1, T2>>
-            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
-
-template <typename T1, typename T2>
-struct is_default_constructible<const std::pair<T1, T2>>
-            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
-
-template <typename... Ts>
-struct is_default_constructible<std::tuple<Ts...>>
-            : conjunction<is_default_constructible<Ts>...> {};
-
-template <typename... Ts>
-struct is_default_constructible<const std::tuple<Ts...>>
-            : conjunction<is_default_constructible<Ts>...> {};
-
-
-template <typename T, typename... Args>
-struct is_constructible : std::is_constructible<T, Args...> {};
-
-template <typename T1, typename T2>
-struct is_constructible<std::pair<T1, T2>> : is_default_constructible<std::pair<T1, T2>> {};
-
-template <typename T1, typename T2>
-struct is_constructible<const std::pair<T1, T2>> : is_default_constructible<const std::pair<T1, T2>> {};
-
-template <typename... Ts>
-struct is_constructible<std::tuple<Ts...>> : is_default_constructible<std::tuple<Ts...>> {};
-
-template <typename... Ts>
-struct is_constructible<const std::tuple<Ts...>> : is_default_constructible<const std::tuple<Ts...>> {};
-
-
-template<typename T, typename = void>
-struct is_iterator_traits : std::false_type {};
-
-template<typename T>
-struct is_iterator_traits<iterator_traits<T>>
-{
-  private:
-    using traits = iterator_traits<T>;
-
-  public:
-    static constexpr auto value =
-        is_detected<value_type_t, traits>::value &&
-        is_detected<difference_type_t, traits>::value &&
-        is_detected<pointer_t, traits>::value &&
-        is_detected<iterator_category_t, traits>::value &&
-        is_detected<reference_t, traits>::value;
-};
-
-template<typename T>
-struct is_range
-{
-  private:
-    using t_ref = typename std::add_lvalue_reference<T>::type;
-
-    using iterator = detected_t<result_of_begin, t_ref>;
-    using sentinel = detected_t<result_of_end, t_ref>;
-
-    // to be 100% correct, it should use https://en.cppreference.com/w/cpp/iterator/input_or_output_iterator
-    // and https://en.cppreference.com/w/cpp/iterator/sentinel_for
-    // but reimplementing these would be too much work, as a lot of other concepts are used underneath
-    static constexpr auto is_iterator_begin =
-        is_iterator_traits<iterator_traits<iterator>>::value;
-
-  public:
-    static constexpr bool value = !std::is_same<iterator, nonesuch>::value && !std::is_same<sentinel, nonesuch>::value && is_iterator_begin;
-};
-
-template<typename R>
-using iterator_t = enable_if_t<is_range<R>::value, result_of_begin<decltype(std::declval<R&>())>>;
-
-template<typename T>
-using range_value_t = value_type_t<iterator_traits<iterator_t<T>>>;
-
-// The following implementation of is_complete_type is taken from
-// https://blogs.msdn.microsoft.com/vcblog/2015/12/02/partial-support-for-expression-sfinae-in-vs-2015-update-1/
-// and is written by Xiang Fan who agreed to using it in this library.
-
-template<typename T, typename = void>
-struct is_complete_type : std::false_type {};
-
-template<typename T>
-struct is_complete_type<T, decltype(void(sizeof(T)))> : std::true_type {};
-
-template<typename BasicJsonType, typename CompatibleObjectType,
-         typename = void>
-struct is_compatible_object_type_impl : std::false_type {};
-
-template<typename BasicJsonType, typename CompatibleObjectType>
-struct is_compatible_object_type_impl <
-    BasicJsonType, CompatibleObjectType,
-    enable_if_t < is_detected<mapped_type_t, CompatibleObjectType>::value&&
-    is_detected<key_type_t, CompatibleObjectType>::value >>
-{
-    using object_t = typename BasicJsonType::object_t;
-
-    // macOS's is_constructible does not play well with nonesuch...
-    static constexpr bool value =
-        is_constructible<typename object_t::key_type,
-        typename CompatibleObjectType::key_type>::value &&
-        is_constructible<typename object_t::mapped_type,
-        typename CompatibleObjectType::mapped_type>::value;
-};
-
-template<typename BasicJsonType, typename CompatibleObjectType>
-struct is_compatible_object_type
-    : is_compatible_object_type_impl<BasicJsonType, CompatibleObjectType> {};
-
-template<typename BasicJsonType, typename ConstructibleObjectType,
-         typename = void>
-struct is_constructible_object_type_impl : std::false_type {};
-
-template<typename BasicJsonType, typename ConstructibleObjectType>
-struct is_constructible_object_type_impl <
-    BasicJsonType, ConstructibleObjectType,
-    enable_if_t < is_detected<mapped_type_t, ConstructibleObjectType>::value&&
-    is_detected<key_type_t, ConstructibleObjectType>::value >>
-{
-    using object_t = typename BasicJsonType::object_t;
-
-    static constexpr bool value =
-        (is_default_constructible<ConstructibleObjectType>::value &&
-         (std::is_move_assignable<ConstructibleObjectType>::value ||
-          std::is_copy_assignable<ConstructibleObjectType>::value) &&
-         (is_constructible<typename ConstructibleObjectType::key_type,
-          typename object_t::key_type>::value &&
-          std::is_same <
-          typename object_t::mapped_type,
-          typename ConstructibleObjectType::mapped_type >::value)) ||
-        (has_from_json<BasicJsonType,
-         typename ConstructibleObjectType::mapped_type>::value ||
-         has_non_default_from_json <
-         BasicJsonType,
-         typename ConstructibleObjectType::mapped_type >::value);
-};
-
-template<typename BasicJsonType, typename ConstructibleObjectType>
-struct is_constructible_object_type
-    : is_constructible_object_type_impl<BasicJsonType,
-      ConstructibleObjectType> {};
-
-template<typename BasicJsonType, typename CompatibleStringType>
-struct is_compatible_string_type
-{
-    static constexpr auto value =
-        is_constructible<typename BasicJsonType::string_t, CompatibleStringType>::value;
-};
-
-template<typename BasicJsonType, typename ConstructibleStringType>
-struct is_constructible_string_type
-{
-    // launder type through decltype() to fix compilation failure on ICPC
-#ifdef __INTEL_COMPILER
-    using laundered_type = decltype(std::declval<ConstructibleStringType>());
-#else
-    using laundered_type = ConstructibleStringType;
-#endif
-
-    static constexpr auto value =
-        conjunction <
-        is_constructible<laundered_type, typename BasicJsonType::string_t>,
-        is_detected_exact<typename BasicJsonType::string_t::value_type,
-        value_type_t, laundered_type >>::value;
-};
-
-template<typename BasicJsonType, typename CompatibleArrayType, typename = void>
-struct is_compatible_array_type_impl : std::false_type {};
-
-template<typename BasicJsonType, typename CompatibleArrayType>
-struct is_compatible_array_type_impl <
-    BasicJsonType, CompatibleArrayType,
-    enable_if_t <
-    is_detected<iterator_t, CompatibleArrayType>::value&&
-    is_iterator_traits<iterator_traits<detected_t<iterator_t, CompatibleArrayType>>>::value&&
-// special case for types like std::filesystem::path whose iterator's value_type are themselves
-// c.f. https://github.com/nlohmann/json/pull/3073
-    !std::is_same<CompatibleArrayType, detected_t<range_value_t, CompatibleArrayType>>::value >>
-{
-    static constexpr bool value =
-        is_constructible<BasicJsonType,
-        range_value_t<CompatibleArrayType>>::value;
-};
-
-template<typename BasicJsonType, typename CompatibleArrayType>
-struct is_compatible_array_type
-    : is_compatible_array_type_impl<BasicJsonType, CompatibleArrayType> {};
-
-template<typename BasicJsonType, typename ConstructibleArrayType, typename = void>
-struct is_constructible_array_type_impl : std::false_type {};
-
-template<typename BasicJsonType, typename ConstructibleArrayType>
-struct is_constructible_array_type_impl <
-    BasicJsonType, ConstructibleArrayType,
-    enable_if_t<std::is_same<ConstructibleArrayType,
-    typename BasicJsonType::value_type>::value >>
-            : std::true_type {};
-
-template<typename BasicJsonType, typename ConstructibleArrayType>
-struct is_constructible_array_type_impl <
-    BasicJsonType, ConstructibleArrayType,
-    enable_if_t < !std::is_same<ConstructibleArrayType,
-    typename BasicJsonType::value_type>::value&&
-    !is_compatible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
-    is_default_constructible<ConstructibleArrayType>::value&&
-(std::is_move_assignable<ConstructibleArrayType>::value ||
- std::is_copy_assignable<ConstructibleArrayType>::value)&&
-is_detected<iterator_t, ConstructibleArrayType>::value&&
-is_iterator_traits<iterator_traits<detected_t<iterator_t, ConstructibleArrayType>>>::value&&
-is_detected<range_value_t, ConstructibleArrayType>::value&&
-// special case for types like std::filesystem::path whose iterator's value_type are themselves
-// c.f. https://github.com/nlohmann/json/pull/3073
-!std::is_same<ConstructibleArrayType, detected_t<range_value_t, ConstructibleArrayType>>::value&&
-        is_complete_type <
-        detected_t<range_value_t, ConstructibleArrayType >>::value >>
-{
-    using value_type = range_value_t<ConstructibleArrayType>;
-
-    static constexpr bool value =
-        std::is_same<value_type,
-        typename BasicJsonType::array_t::value_type>::value ||
-        has_from_json<BasicJsonType,
-        value_type>::value ||
-        has_non_default_from_json <
-        BasicJsonType,
-        value_type >::value;
-};
-
-template<typename BasicJsonType, typename ConstructibleArrayType>
-struct is_constructible_array_type
-    : is_constructible_array_type_impl<BasicJsonType, ConstructibleArrayType> {};
-
-template<typename RealIntegerType, typename CompatibleNumberIntegerType,
-         typename = void>
-struct is_compatible_integer_type_impl : std::false_type {};
-
-template<typename RealIntegerType, typename CompatibleNumberIntegerType>
-struct is_compatible_integer_type_impl <
-    RealIntegerType, CompatibleNumberIntegerType,
-    enable_if_t < std::is_integral<RealIntegerType>::value&&
-    std::is_integral<CompatibleNumberIntegerType>::value&&
-    !std::is_same<bool, CompatibleNumberIntegerType>::value >>
-{
-    // is there an assert somewhere on overflows?
-    using RealLimits = std::numeric_limits<RealIntegerType>;
-    using CompatibleLimits = std::numeric_limits<CompatibleNumberIntegerType>;
-
-    static constexpr auto value =
-        is_constructible<RealIntegerType,
-        CompatibleNumberIntegerType>::value &&
-        CompatibleLimits::is_integer &&
-        RealLimits::is_signed == CompatibleLimits::is_signed;
-};
-
-template<typename RealIntegerType, typename CompatibleNumberIntegerType>
-struct is_compatible_integer_type
-    : is_compatible_integer_type_impl<RealIntegerType,
-      CompatibleNumberIntegerType> {};
-
-template<typename BasicJsonType, typename CompatibleType, typename = void>
-struct is_compatible_type_impl: std::false_type {};
-
-template<typename BasicJsonType, typename CompatibleType>
-struct is_compatible_type_impl <
-    BasicJsonType, CompatibleType,
-    enable_if_t<is_complete_type<CompatibleType>::value >>
-{
-    static constexpr bool value =
-        has_to_json<BasicJsonType, CompatibleType>::value;
-};
-
-template<typename BasicJsonType, typename CompatibleType>
-struct is_compatible_type
-    : is_compatible_type_impl<BasicJsonType, CompatibleType> {};
-
-template<typename T1, typename T2>
-struct is_constructible_tuple : std::false_type {};
-
-template<typename T1, typename... Args>
-struct is_constructible_tuple<T1, std::tuple<Args...>> : conjunction<is_constructible<T1, Args>...> {};
-
-template<typename BasicJsonType, typename T>
-struct is_json_iterator_of : std::false_type {};
-
-template<typename BasicJsonType>
-struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::iterator> : std::true_type {};
-
-template<typename BasicJsonType>
-struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::const_iterator> : std::true_type
-{};
-
-// checks if a given type T is a template specialization of Primary
-template<template <typename...> class Primary, typename T>
-struct is_specialization_of : std::false_type {};
-
-template<template <typename...> class Primary, typename... Args>
-struct is_specialization_of<Primary, Primary<Args...>> : std::true_type {};
-
-template<typename T>
-using is_json_pointer = is_specialization_of<::nlohmann::json_pointer, uncvref_t<T>>;
-
-// checks if A and B are comparable using Compare functor
-template<typename Compare, typename A, typename B, typename = void>
-struct is_comparable : std::false_type {};
-
-template<typename Compare, typename A, typename B>
-struct is_comparable<Compare, A, B, void_t<
-decltype(std::declval<Compare>()(std::declval<A>(), std::declval<B>())),
-decltype(std::declval<Compare>()(std::declval<B>(), std::declval<A>()))
->> : std::true_type {};
-
-template<typename T>
-using detect_is_transparent = typename T::is_transparent;
-
-// type trait to check if KeyType can be used as object key (without a BasicJsonType)
-// see is_usable_as_basic_json_key_type below
-template<typename Comparator, typename ObjectKeyType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
-         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
-using is_usable_as_key_type = typename std::conditional <
-                              is_comparable<Comparator, ObjectKeyType, KeyTypeCVRef>::value
-                              && !(ExcludeObjectKeyType && std::is_same<KeyType,
-                                   ObjectKeyType>::value)
-                              && (!RequireTransparentComparator
-                                  || is_detected <detect_is_transparent, Comparator>::value)
-                              && !is_json_pointer<KeyType>::value,
-                              std::true_type,
-                              std::false_type >::type;
-
-// type trait to check if KeyType can be used as object key
-// true if:
-//   - KeyType is comparable with BasicJsonType::object_t::key_type
-//   - if ExcludeObjectKeyType is true, KeyType is not BasicJsonType::object_t::key_type
-//   - the comparator is transparent or RequireTransparentComparator is false
-//   - KeyType is not a JSON iterator or json_pointer
-template<typename BasicJsonType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
-         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
-using is_usable_as_basic_json_key_type = typename std::conditional <
-        is_usable_as_key_type<typename BasicJsonType::object_comparator_t,
-        typename BasicJsonType::object_t::key_type, KeyTypeCVRef,
-        RequireTransparentComparator, ExcludeObjectKeyType>::value
-        && !is_json_iterator_of<BasicJsonType, KeyType>::value,
-        std::true_type,
-        std::false_type >::type;
-
-template<typename ObjectType, typename KeyType>
-using detect_erase_with_key_type = decltype(std::declval<ObjectType&>().erase(std::declval<KeyType>()));
-
-// type trait to check if object_t has an erase() member functions accepting KeyType
-template<typename BasicJsonType, typename KeyType>
-using has_erase_with_key_type = typename std::conditional <
-                                is_detected <
-                                detect_erase_with_key_type,
-                                typename BasicJsonType::object_t, KeyType >::value,
-                                std::true_type,
-                                std::false_type >::type;
-
-// a naive helper to check if a type is an ordered_map (exploits the fact that
-// ordered_map inherits capacity() from std::vector)
-template <typename T>
-struct is_ordered_map
-{
-    using one = char;
-
-    struct two
-    {
-        char x[2]; // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-    };
-
-    template <typename C> static one test( decltype(&C::capacity) ) ;
-    template <typename C> static two test(...);
-
-    enum { value = sizeof(test<T>(nullptr)) == sizeof(char) }; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-};
-
-// to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
-template < typename T, typename U, enable_if_t < !std::is_same<T, U>::value, int > = 0 >
-T conditional_static_cast(U value)
-{
-    return static_cast<T>(value);
-}
-
-template<typename T, typename U, enable_if_t<std::is_same<T, U>::value, int> = 0>
-T conditional_static_cast(U value)
-{
-    return value;
-}
-
-template<typename... Types>
-using all_integral = conjunction<std::is_integral<Types>...>;
-
-template<typename... Types>
-using all_signed = conjunction<std::is_signed<Types>...>;
-
-template<typename... Types>
-using all_unsigned = conjunction<std::is_unsigned<Types>...>;
-
-// there's a disjunction trait in another PR; replace when merged
-template<typename... Types>
-using same_sign = std::integral_constant < bool,
-      all_signed<Types...>::value || all_unsigned<Types...>::value >;
-
-template<typename OfType, typename T>
-using never_out_of_range = std::integral_constant < bool,
-      (std::is_signed<OfType>::value && (sizeof(T) < sizeof(OfType)))
-      || (same_sign<OfType, T>::value && sizeof(OfType) == sizeof(T)) >;
-
-template<typename OfType, typename T,
-         bool OfTypeSigned = std::is_signed<OfType>::value,
-         bool TSigned = std::is_signed<T>::value>
-struct value_in_range_of_impl2;
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl2<OfType, T, false, false>
-{
-    static constexpr bool test(T val)
-    {
-        using CommonType = typename std::common_type<OfType, T>::type;
-        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
-    }
-};
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl2<OfType, T, true, false>
-{
-    static constexpr bool test(T val)
-    {
-        using CommonType = typename std::common_type<OfType, T>::type;
-        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
-    }
-};
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl2<OfType, T, false, true>
-{
-    static constexpr bool test(T val)
-    {
-        using CommonType = typename std::common_type<OfType, T>::type;
-        return val >= 0 && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
-    }
-};
-
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl2<OfType, T, true, true>
-{
-    static constexpr bool test(T val)
-    {
-        using CommonType = typename std::common_type<OfType, T>::type;
-        return static_cast<CommonType>(val) >= static_cast<CommonType>((std::numeric_limits<OfType>::min)())
-               && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
-    }
-};
-
-template<typename OfType, typename T,
-         bool NeverOutOfRange = never_out_of_range<OfType, T>::value,
-         typename = detail::enable_if_t<all_integral<OfType, T>::value>>
-struct value_in_range_of_impl1;
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl1<OfType, T, false>
-{
-    static constexpr bool test(T val)
-    {
-        return value_in_range_of_impl2<OfType, T>::test(val);
-    }
-};
-
-template<typename OfType, typename T>
-struct value_in_range_of_impl1<OfType, T, true>
-{
-    static constexpr bool test(T /*val*/)
-    {
-        return true;
-    }
-};
-
-template<typename OfType, typename T>
-inline constexpr bool value_in_range_of(T val)
-{
-    return value_in_range_of_impl1<OfType, T>::test(val);
-}
-
-template<bool Value>
-using bool_constant = std::integral_constant<bool, Value>;
-
-///////////////////////////////////////////////////////////////////////////////
-// is_c_string
-///////////////////////////////////////////////////////////////////////////////
-
-namespace impl
-{
-
-template<typename T>
-inline constexpr bool is_c_string()
-{
-    using TUnExt = typename std::remove_extent<T>::type;
-    using TUnCVExt = typename std::remove_cv<TUnExt>::type;
-    using TUnPtr = typename std::remove_pointer<T>::type;
-    using TUnCVPtr = typename std::remove_cv<TUnPtr>::type;
-    return
-        (std::is_array<T>::value && std::is_same<TUnCVExt, char>::value)
-        || (std::is_pointer<T>::value && std::is_same<TUnCVPtr, char>::value);
-}
-
-}  // namespace impl
-
-// checks whether T is a [cv] char */[cv] char[] C string
-template<typename T>
-struct is_c_string : bool_constant<impl::is_c_string<T>()> {};
-
-template<typename T>
-using is_c_string_uncvref = is_c_string<uncvref_t<T>>;
-
-///////////////////////////////////////////////////////////////////////////////
-// is_transparent
-///////////////////////////////////////////////////////////////////////////////
-
-namespace impl
-{
-
-template<typename T>
-inline constexpr bool is_transparent()
-{
-    return is_detected<detect_is_transparent, T>::value;
-}
-
-}  // namespace impl
-
-// checks whether T has a member named is_transparent
-template<typename T>
-struct is_transparent : bool_constant<impl::is_transparent<T>()> {};
-
-///////////////////////////////////////////////////////////////////////////////
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/string_concat.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstring> // strlen
-#include <string> // string
-#include <utility> // forward
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/detected.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-inline std::size_t concat_length()
-{
-    return 0;
-}
-
-template<typename... Args>
-inline std::size_t concat_length(const char* cstr, Args&& ... rest);
-
-template<typename StringType, typename... Args>
-inline std::size_t concat_length(const StringType& str, Args&& ... rest);
-
-template<typename... Args>
-inline std::size_t concat_length(const char /*c*/, Args&& ... rest)
-{
-    return 1 + concat_length(std::forward<Args>(rest)...);
-}
-
-template<typename... Args>
-inline std::size_t concat_length(const char* cstr, Args&& ... rest)
-{
-    // cppcheck-suppress ignoredReturnValue
-    return ::strlen(cstr) + concat_length(std::forward<Args>(rest)...);
-}
-
-template<typename StringType, typename... Args>
-inline std::size_t concat_length(const StringType& str, Args&& ... rest)
-{
-    return str.size() + concat_length(std::forward<Args>(rest)...);
-}
-
-template<typename OutStringType>
-inline void concat_into(OutStringType& /*out*/)
-{}
-
-template<typename StringType, typename Arg>
-using string_can_append = decltype(std::declval<StringType&>().append(std::declval < Arg && > ()));
-
-template<typename StringType, typename Arg>
-using detect_string_can_append = is_detected<string_can_append, StringType, Arg>;
-
-template<typename StringType, typename Arg>
-using string_can_append_op = decltype(std::declval<StringType&>() += std::declval < Arg && > ());
-
-template<typename StringType, typename Arg>
-using detect_string_can_append_op = is_detected<string_can_append_op, StringType, Arg>;
-
-template<typename StringType, typename Arg>
-using string_can_append_iter = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().begin(), std::declval<const Arg&>().end()));
-
-template<typename StringType, typename Arg>
-using detect_string_can_append_iter = is_detected<string_can_append_iter, StringType, Arg>;
-
-template<typename StringType, typename Arg>
-using string_can_append_data = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().data(), std::declval<const Arg&>().size()));
-
-template<typename StringType, typename Arg>
-using detect_string_can_append_data = is_detected<string_can_append_data, StringType, Arg>;
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && detect_string_can_append_op<OutStringType, Arg>::value, int > = 0 >
-inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest);
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && !detect_string_can_append_op<OutStringType, Arg>::value
-                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > = 0 >
-inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && !detect_string_can_append_op<OutStringType, Arg>::value
-                         && !detect_string_can_append_iter<OutStringType, Arg>::value
-                         && detect_string_can_append_data<OutStringType, Arg>::value, int > = 0 >
-inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
-
-template<typename OutStringType, typename Arg, typename... Args,
-         enable_if_t<detect_string_can_append<OutStringType, Arg>::value, int> = 0>
-inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest)
-{
-    out.append(std::forward<Arg>(arg));
-    concat_into(out, std::forward<Args>(rest)...);
-}
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && detect_string_can_append_op<OutStringType, Arg>::value, int > >
-inline void concat_into(OutStringType& out, Arg&& arg, Args&& ... rest)
-{
-    out += std::forward<Arg>(arg);
-    concat_into(out, std::forward<Args>(rest)...);
-}
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && !detect_string_can_append_op<OutStringType, Arg>::value
-                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > >
-inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
-{
-    out.append(arg.begin(), arg.end());
-    concat_into(out, std::forward<Args>(rest)...);
-}
-
-template < typename OutStringType, typename Arg, typename... Args,
-           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
-                         && !detect_string_can_append_op<OutStringType, Arg>::value
-                         && !detect_string_can_append_iter<OutStringType, Arg>::value
-                         && detect_string_can_append_data<OutStringType, Arg>::value, int > >
-inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
-{
-    out.append(arg.data(), arg.size());
-    concat_into(out, std::forward<Args>(rest)...);
-}
-
-template<typename OutStringType = std::string, typename... Args>
-inline OutStringType concat(Args && ... args)
-{
-    OutStringType str;
-    str.reserve(concat_length(std::forward<Args>(args)...));
-    concat_into(str, std::forward<Args>(args)...);
-    return str;
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-////////////////
-// exceptions //
-////////////////
-
-/// @brief general exception of the @ref basic_json class
-/// @sa https://json.nlohmann.me/api/basic_json/exception/
-class exception : public std::exception
-{
-  public:
-    /// returns the explanatory string
-    const char* what() const noexcept override
-    {
-        return m.what();
-    }
-
-    /// the id of the exception
-    const int id; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
-
-  protected:
-    JSON_HEDLEY_NON_NULL(3)
-    exception(int id_, const char* what_arg) : id(id_), m(what_arg) {} // NOLINT(bugprone-throw-keyword-missing)
-
-    static std::string name(const std::string& ename, int id_)
-    {
-        return concat("[json.exception.", ename, '.', std::to_string(id_), "] ");
-    }
-
-    static std::string diagnostics(std::nullptr_t /*leaf_element*/)
-    {
-        return "";
-    }
-
-    template<typename BasicJsonType>
-    static std::string diagnostics(const BasicJsonType* leaf_element)
-    {
-#if JSON_DIAGNOSTICS
-        std::vector<std::string> tokens;
-        for (const auto* current = leaf_element; current != nullptr && current->m_parent != nullptr; current = current->m_parent)
-        {
-            switch (current->m_parent->type())
-            {
-                case value_t::array:
-                {
-                    for (std::size_t i = 0; i < current->m_parent->m_value.array->size(); ++i)
-                    {
-                        if (&current->m_parent->m_value.array->operator[](i) == current)
-                        {
-                            tokens.emplace_back(std::to_string(i));
-                            break;
-                        }
-                    }
-                    break;
-                }
-
-                case value_t::object:
-                {
-                    for (const auto& element : *current->m_parent->m_value.object)
-                    {
-                        if (&element.second == current)
-                        {
-                            tokens.emplace_back(element.first.c_str());
-                            break;
-                        }
-                    }
-                    break;
-                }
-
-                case value_t::null: // LCOV_EXCL_LINE
-                case value_t::string: // LCOV_EXCL_LINE
-                case value_t::boolean: // LCOV_EXCL_LINE
-                case value_t::number_integer: // LCOV_EXCL_LINE
-                case value_t::number_unsigned: // LCOV_EXCL_LINE
-                case value_t::number_float: // LCOV_EXCL_LINE
-                case value_t::binary: // LCOV_EXCL_LINE
-                case value_t::discarded: // LCOV_EXCL_LINE
-                default:   // LCOV_EXCL_LINE
-                    break; // LCOV_EXCL_LINE
-            }
-        }
-
-        if (tokens.empty())
-        {
-            return "";
-        }
-
-        auto str = std::accumulate(tokens.rbegin(), tokens.rend(), std::string{},
-                                   [](const std::string & a, const std::string & b)
-        {
-            return concat(a, '/', detail::escape(b));
-        });
-        return concat('(', str, ") ");
-#else
-        static_cast<void>(leaf_element);
-        return "";
-#endif
-    }
-
-  private:
-    /// an exception object as storage for error messages
-    std::runtime_error m;
-};
-
-/// @brief exception indicating a parse error
-/// @sa https://json.nlohmann.me/api/basic_json/parse_error/
-class parse_error : public exception
-{
-  public:
-    /*!
-    @brief create a parse error exception
-    @param[in] id_       the id of the exception
-    @param[in] pos       the position where the error occurred (or with
-                         chars_read_total=0 if the position cannot be
-                         determined)
-    @param[in] what_arg  the explanatory string
-    @return parse_error object
-    */
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static parse_error create(int id_, const position_t& pos, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("parse_error", id_), "parse error",
-                               position_string(pos), ": ", exception::diagnostics(context), what_arg);
-        return {id_, pos.chars_read_total, w.c_str()};
-    }
-
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static parse_error create(int id_, std::size_t byte_, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("parse_error", id_), "parse error",
-                               (byte_ != 0 ? (concat(" at byte ", std::to_string(byte_))) : ""),
-                               ": ", exception::diagnostics(context), what_arg);
-        return {id_, byte_, w.c_str()};
-    }
-
-    /*!
-    @brief byte index of the parse error
-
-    The byte index of the last read character in the input file.
-
-    @note For an input with n bytes, 1 is the index of the first character and
-          n+1 is the index of the terminating null byte or the end of file.
-          This also holds true when reading a byte vector (CBOR or MessagePack).
-    */
-    const std::size_t byte;
-
-  private:
-    parse_error(int id_, std::size_t byte_, const char* what_arg)
-        : exception(id_, what_arg), byte(byte_) {}
-
-    static std::string position_string(const position_t& pos)
-    {
-        return concat(" at line ", std::to_string(pos.lines_read + 1),
-                      ", column ", std::to_string(pos.chars_read_current_line));
-    }
-};
-
-/// @brief exception indicating errors with iterators
-/// @sa https://json.nlohmann.me/api/basic_json/invalid_iterator/
-class invalid_iterator : public exception
-{
-  public:
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static invalid_iterator create(int id_, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("invalid_iterator", id_), exception::diagnostics(context), what_arg);
-        return {id_, w.c_str()};
-    }
-
-  private:
-    JSON_HEDLEY_NON_NULL(3)
-    invalid_iterator(int id_, const char* what_arg)
-        : exception(id_, what_arg) {}
-};
-
-/// @brief exception indicating executing a member function with a wrong type
-/// @sa https://json.nlohmann.me/api/basic_json/type_error/
-class type_error : public exception
-{
-  public:
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static type_error create(int id_, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("type_error", id_), exception::diagnostics(context), what_arg);
-        return {id_, w.c_str()};
-    }
-
-  private:
-    JSON_HEDLEY_NON_NULL(3)
-    type_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
-};
-
-/// @brief exception indicating access out of the defined range
-/// @sa https://json.nlohmann.me/api/basic_json/out_of_range/
-class out_of_range : public exception
-{
-  public:
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static out_of_range create(int id_, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("out_of_range", id_), exception::diagnostics(context), what_arg);
-        return {id_, w.c_str()};
-    }
-
-  private:
-    JSON_HEDLEY_NON_NULL(3)
-    out_of_range(int id_, const char* what_arg) : exception(id_, what_arg) {}
-};
-
-/// @brief exception indicating other library errors
-/// @sa https://json.nlohmann.me/api/basic_json/other_error/
-class other_error : public exception
-{
-  public:
-    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
-    static other_error create(int id_, const std::string& what_arg, BasicJsonContext context)
-    {
-        std::string w = concat(exception::name("other_error", id_), exception::diagnostics(context), what_arg);
-        return {id_, w.c_str()};
-    }
-
-  private:
-    JSON_HEDLEY_NON_NULL(3)
-    other_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/identity_tag.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-// dispatching helper struct
-template <class T> struct identity_tag {};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/std_fs.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-#if JSON_HAS_EXPERIMENTAL_FILESYSTEM
-#include <experimental/filesystem>
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-namespace std_fs = std::experimental::filesystem;
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-#elif JSON_HAS_FILESYSTEM
-#include <filesystem>
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-namespace std_fs = std::filesystem;
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-#endif
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename std::nullptr_t& n)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_null()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be null, but is ", j.type_name()), &j));
-    }
-    n = nullptr;
-}
-
-// overloads for basic_json template parameters
-template < typename BasicJsonType, typename ArithmeticType,
-           enable_if_t < std::is_arithmetic<ArithmeticType>::value&&
-                         !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
-                         int > = 0 >
-void get_arithmetic_value(const BasicJsonType& j, ArithmeticType& val)
-{
-    switch (static_cast<value_t>(j))
-    {
-        case value_t::number_unsigned:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
-            break;
-        }
-        case value_t::number_integer:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
-            break;
-        }
-        case value_t::number_float:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
-            break;
-        }
-
-        case value_t::null:
-        case value_t::object:
-        case value_t::array:
-        case value_t::string:
-        case value_t::boolean:
-        case value_t::binary:
-        case value_t::discarded:
-        default:
-            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
-    }
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::boolean_t& b)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_boolean()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be boolean, but is ", j.type_name()), &j));
-    }
-    b = *j.template get_ptr<const typename BasicJsonType::boolean_t*>();
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::string_t& s)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
-    }
-    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
-}
-
-template <
-    typename BasicJsonType, typename StringType,
-    enable_if_t <
-        std::is_assignable<StringType&, const typename BasicJsonType::string_t>::value
-        && is_detected_exact<typename BasicJsonType::string_t::value_type, value_type_t, StringType>::value
-        && !std::is_same<typename BasicJsonType::string_t, StringType>::value
-        && !is_json_ref<StringType>::value, int > = 0 >
-inline void from_json(const BasicJsonType& j, StringType& s)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
-    }
-
-    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_float_t& val)
-{
-    get_arithmetic_value(j, val);
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_unsigned_t& val)
-{
-    get_arithmetic_value(j, val);
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_integer_t& val)
-{
-    get_arithmetic_value(j, val);
-}
-
-#if !JSON_DISABLE_ENUM_SERIALIZATION
-template<typename BasicJsonType, typename EnumType,
-         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
-inline void from_json(const BasicJsonType& j, EnumType& e)
-{
-    typename std::underlying_type<EnumType>::type val;
-    get_arithmetic_value(j, val);
-    e = static_cast<EnumType>(val);
-}
-#endif  // JSON_DISABLE_ENUM_SERIALIZATION
-
-// forward_list doesn't have an insert method
-template<typename BasicJsonType, typename T, typename Allocator,
-         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
-inline void from_json(const BasicJsonType& j, std::forward_list<T, Allocator>& l)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-    l.clear();
-    std::transform(j.rbegin(), j.rend(),
-                   std::front_inserter(l), [](const BasicJsonType & i)
-    {
-        return i.template get<T>();
-    });
-}
-
-// valarray doesn't have an insert method
-template<typename BasicJsonType, typename T,
-         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
-inline void from_json(const BasicJsonType& j, std::valarray<T>& l)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-    l.resize(j.size());
-    std::transform(j.begin(), j.end(), std::begin(l),
-                   [](const BasicJsonType & elem)
-    {
-        return elem.template get<T>();
-    });
-}
-
-template<typename BasicJsonType, typename T, std::size_t N>
-auto from_json(const BasicJsonType& j, T (&arr)[N])  // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
--> decltype(j.template get<T>(), void())
-{
-    for (std::size_t i = 0; i < N; ++i)
-    {
-        arr[i] = j.at(i).template get<T>();
-    }
-}
-
-template<typename BasicJsonType>
-inline void from_json_array_impl(const BasicJsonType& j, typename BasicJsonType::array_t& arr, priority_tag<3> /*unused*/)
-{
-    arr = *j.template get_ptr<const typename BasicJsonType::array_t*>();
-}
-
-template<typename BasicJsonType, typename T, std::size_t N>
-auto from_json_array_impl(const BasicJsonType& j, std::array<T, N>& arr,
-                          priority_tag<2> /*unused*/)
--> decltype(j.template get<T>(), void())
-{
-    for (std::size_t i = 0; i < N; ++i)
-    {
-        arr[i] = j.at(i).template get<T>();
-    }
-}
-
-template<typename BasicJsonType, typename ConstructibleArrayType,
-         enable_if_t<
-             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
-             int> = 0>
-auto from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr, priority_tag<1> /*unused*/)
--> decltype(
-    arr.reserve(std::declval<typename ConstructibleArrayType::size_type>()),
-    j.template get<typename ConstructibleArrayType::value_type>(),
-    void())
-{
-    using std::end;
-
-    ConstructibleArrayType ret;
-    ret.reserve(j.size());
-    std::transform(j.begin(), j.end(),
-                   std::inserter(ret, end(ret)), [](const BasicJsonType & i)
-    {
-        // get<BasicJsonType>() returns *this, this won't call a from_json
-        // method when value_type is BasicJsonType
-        return i.template get<typename ConstructibleArrayType::value_type>();
-    });
-    arr = std::move(ret);
-}
-
-template<typename BasicJsonType, typename ConstructibleArrayType,
-         enable_if_t<
-             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
-             int> = 0>
-inline void from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr,
-                                 priority_tag<0> /*unused*/)
-{
-    using std::end;
-
-    ConstructibleArrayType ret;
-    std::transform(
-        j.begin(), j.end(), std::inserter(ret, end(ret)),
-        [](const BasicJsonType & i)
-    {
-        // get<BasicJsonType>() returns *this, this won't call a from_json
-        // method when value_type is BasicJsonType
-        return i.template get<typename ConstructibleArrayType::value_type>();
-    });
-    arr = std::move(ret);
-}
-
-template < typename BasicJsonType, typename ConstructibleArrayType,
-           enable_if_t <
-               is_constructible_array_type<BasicJsonType, ConstructibleArrayType>::value&&
-               !is_constructible_object_type<BasicJsonType, ConstructibleArrayType>::value&&
-               !is_constructible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
-               !std::is_same<ConstructibleArrayType, typename BasicJsonType::binary_t>::value&&
-               !is_basic_json<ConstructibleArrayType>::value,
-               int > = 0 >
-auto from_json(const BasicJsonType& j, ConstructibleArrayType& arr)
--> decltype(from_json_array_impl(j, arr, priority_tag<3> {}),
-j.template get<typename ConstructibleArrayType::value_type>(),
-void())
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-
-    from_json_array_impl(j, arr, priority_tag<3> {});
-}
-
-template < typename BasicJsonType, typename T, std::size_t... Idx >
-std::array<T, sizeof...(Idx)> from_json_inplace_array_impl(BasicJsonType&& j,
-        identity_tag<std::array<T, sizeof...(Idx)>> /*unused*/, index_sequence<Idx...> /*unused*/)
-{
-    return { { std::forward<BasicJsonType>(j).at(Idx).template get<T>()... } };
-}
-
-template < typename BasicJsonType, typename T, std::size_t N >
-auto from_json(BasicJsonType&& j, identity_tag<std::array<T, N>> tag)
--> decltype(from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {}))
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-
-    return from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {});
-}
-
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::binary_t& bin)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_binary()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be binary, but is ", j.type_name()), &j));
-    }
-
-    bin = *j.template get_ptr<const typename BasicJsonType::binary_t*>();
-}
-
-template<typename BasicJsonType, typename ConstructibleObjectType,
-         enable_if_t<is_constructible_object_type<BasicJsonType, ConstructibleObjectType>::value, int> = 0>
-inline void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_object()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be object, but is ", j.type_name()), &j));
-    }
-
-    ConstructibleObjectType ret;
-    const auto* inner_object = j.template get_ptr<const typename BasicJsonType::object_t*>();
-    using value_type = typename ConstructibleObjectType::value_type;
-    std::transform(
-        inner_object->begin(), inner_object->end(),
-        std::inserter(ret, ret.begin()),
-        [](typename BasicJsonType::object_t::value_type const & p)
-    {
-        return value_type(p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>());
-    });
-    obj = std::move(ret);
-}
-
-// overload for arithmetic types, not chosen for basic_json template arguments
-// (BooleanType, etc..); note: Is it really necessary to provide explicit
-// overloads for boolean_t etc. in case of a custom BooleanType which is not
-// an arithmetic type?
-template < typename BasicJsonType, typename ArithmeticType,
-           enable_if_t <
-               std::is_arithmetic<ArithmeticType>::value&&
-               !std::is_same<ArithmeticType, typename BasicJsonType::number_unsigned_t>::value&&
-               !std::is_same<ArithmeticType, typename BasicJsonType::number_integer_t>::value&&
-               !std::is_same<ArithmeticType, typename BasicJsonType::number_float_t>::value&&
-               !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
-               int > = 0 >
-inline void from_json(const BasicJsonType& j, ArithmeticType& val)
-{
-    switch (static_cast<value_t>(j))
-    {
-        case value_t::number_unsigned:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
-            break;
-        }
-        case value_t::number_integer:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
-            break;
-        }
-        case value_t::number_float:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
-            break;
-        }
-        case value_t::boolean:
-        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::boolean_t*>());
-            break;
-        }
-
-        case value_t::null:
-        case value_t::object:
-        case value_t::array:
-        case value_t::string:
-        case value_t::binary:
-        case value_t::discarded:
-        default:
-            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
-    }
-}
-
-template<typename BasicJsonType, typename... Args, std::size_t... Idx>
-std::tuple<Args...> from_json_tuple_impl_base(BasicJsonType&& j, index_sequence<Idx...> /*unused*/)
-{
-    return std::make_tuple(std::forward<BasicJsonType>(j).at(Idx).template get<Args>()...);
-}
-
-template < typename BasicJsonType, class A1, class A2 >
-std::pair<A1, A2> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::pair<A1, A2>> /*unused*/, priority_tag<0> /*unused*/)
-{
-    return {std::forward<BasicJsonType>(j).at(0).template get<A1>(),
-            std::forward<BasicJsonType>(j).at(1).template get<A2>()};
-}
-
-template<typename BasicJsonType, typename A1, typename A2>
-inline void from_json_tuple_impl(BasicJsonType&& j, std::pair<A1, A2>& p, priority_tag<1> /*unused*/)
-{
-    p = from_json_tuple_impl(std::forward<BasicJsonType>(j), identity_tag<std::pair<A1, A2>> {}, priority_tag<0> {});
-}
-
-template<typename BasicJsonType, typename... Args>
-std::tuple<Args...> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::tuple<Args...>> /*unused*/, priority_tag<2> /*unused*/)
-{
-    return from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
-}
-
-template<typename BasicJsonType, typename... Args>
-inline void from_json_tuple_impl(BasicJsonType&& j, std::tuple<Args...>& t, priority_tag<3> /*unused*/)
-{
-    t = from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
-}
-
-template<typename BasicJsonType, typename TupleRelated>
-auto from_json(BasicJsonType&& j, TupleRelated&& t)
--> decltype(from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {}))
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-
-    return from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {});
-}
-
-template < typename BasicJsonType, typename Key, typename Value, typename Compare, typename Allocator,
-           typename = enable_if_t < !std::is_constructible <
-                                        typename BasicJsonType::string_t, Key >::value >>
-inline void from_json(const BasicJsonType& j, std::map<Key, Value, Compare, Allocator>& m)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-    m.clear();
-    for (const auto& p : j)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
-        {
-            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
-        }
-        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
-    }
-}
-
-template < typename BasicJsonType, typename Key, typename Value, typename Hash, typename KeyEqual, typename Allocator,
-           typename = enable_if_t < !std::is_constructible <
-                                        typename BasicJsonType::string_t, Key >::value >>
-inline void from_json(const BasicJsonType& j, std::unordered_map<Key, Value, Hash, KeyEqual, Allocator>& m)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
-    }
-    m.clear();
-    for (const auto& p : j)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
-        {
-            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
-        }
-        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
-    }
-}
-
-#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
-template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, std_fs::path& p)
-{
-    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
-    {
-        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
-    }
-    p = *j.template get_ptr<const typename BasicJsonType::string_t*>();
-}
-#endif
-
-struct from_json_fn
-{
-    template<typename BasicJsonType, typename T>
-    auto operator()(const BasicJsonType& j, T&& val) const
-    noexcept(noexcept(from_json(j, std::forward<T>(val))))
-    -> decltype(from_json(j, std::forward<T>(val)))
-    {
-        return from_json(j, std::forward<T>(val));
-    }
-};
-
-}  // namespace detail
-
-#ifndef JSON_HAS_CPP_17
-/// namespace to hold default `from_json` function
-/// to see why this is required:
-/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
-namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
-{
-#endif
-JSON_INLINE_VARIABLE constexpr const auto& from_json = // NOLINT(misc-definitions-in-headers)
-    detail::static_const<detail::from_json_fn>::value;
-#ifndef JSON_HAS_CPP_17
-}  // namespace
-#endif
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/conversions/to_json.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // copy
-#include <iterator> // begin, end
-#include <string> // string
-#include <tuple> // tuple, get
-#include <type_traits> // is_same, is_constructible, is_floating_point, is_enum, underlying_type
-#include <utility> // move, forward, declval, pair
-#include <valarray> // valarray
-#include <vector> // vector
-
-// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef> // size_t
-#include <iterator> // input_iterator_tag
-#include <string> // string, to_string
-#include <tuple> // tuple_size, get, tuple_element
-#include <utility> // move
-
-#if JSON_HAS_RANGES
-    #include <ranges> // enable_borrowed_range
-#endif
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename string_type>
-void int_to_string( string_type& target, std::size_t value )
-{
-    // For ADL
-    using std::to_string;
-    target = to_string(value);
-}
-template<typename IteratorType> class iteration_proxy_value
-{
-  public:
-    using difference_type = std::ptrdiff_t;
-    using value_type = iteration_proxy_value;
-    using pointer = value_type *;
-    using reference = value_type &;
-    using iterator_category = std::input_iterator_tag;
-    using string_type = typename std::remove_cv< typename std::remove_reference<decltype( std::declval<IteratorType>().key() ) >::type >::type;
-
-  private:
-    /// the iterator
-    IteratorType anchor{};
-    /// an index for arrays (used to create key names)
-    std::size_t array_index = 0;
-    /// last stringified array index
-    mutable std::size_t array_index_last = 0;
-    /// a string representation of the array index
-    mutable string_type array_index_str = "0";
-    /// an empty string (to return a reference for primitive values)
-    string_type empty_str{};
-
-  public:
-    explicit iteration_proxy_value() = default;
-    explicit iteration_proxy_value(IteratorType it, std::size_t array_index_ = 0)
-    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
-             && std::is_nothrow_default_constructible<string_type>::value)
-        : anchor(std::move(it))
-        , array_index(array_index_)
-    {}
-
-    iteration_proxy_value(iteration_proxy_value const&) = default;
-    iteration_proxy_value& operator=(iteration_proxy_value const&) = default;
-    // older GCCs are a bit fussy and require explicit noexcept specifiers on defaulted functions
-    iteration_proxy_value(iteration_proxy_value&&)
-    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
-             && std::is_nothrow_move_constructible<string_type>::value) = default;
-    iteration_proxy_value& operator=(iteration_proxy_value&&)
-    noexcept(std::is_nothrow_move_assignable<IteratorType>::value
-             && std::is_nothrow_move_assignable<string_type>::value) = default;
-    ~iteration_proxy_value() = default;
-
-    /// dereference operator (needed for range-based for)
-    const iteration_proxy_value& operator*() const
-    {
-        return *this;
-    }
-
-    /// increment operator (needed for range-based for)
-    iteration_proxy_value& operator++()
-    {
-        ++anchor;
-        ++array_index;
-
-        return *this;
-    }
-
-    iteration_proxy_value operator++(int)& // NOLINT(cert-dcl21-cpp)
-    {
-        auto tmp = iteration_proxy_value(anchor, array_index);
-        ++anchor;
-        ++array_index;
-        return tmp;
-    }
-
-    /// equality operator (needed for InputIterator)
-    bool operator==(const iteration_proxy_value& o) const
-    {
-        return anchor == o.anchor;
-    }
-
-    /// inequality operator (needed for range-based for)
-    bool operator!=(const iteration_proxy_value& o) const
-    {
-        return anchor != o.anchor;
-    }
-
-    /// return key of the iterator
-    const string_type& key() const
-    {
-        JSON_ASSERT(anchor.m_object != nullptr);
-
-        switch (anchor.m_object->type())
-        {
-            // use integer array index as key
-            case value_t::array:
-            {
-                if (array_index != array_index_last)
-                {
-                    int_to_string( array_index_str, array_index );
-                    array_index_last = array_index;
-                }
-                return array_index_str;
-            }
-
-            // use key from the object
-            case value_t::object:
-                return anchor.key();
-
-            // use an empty key for all primitive types
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                return empty_str;
-        }
-    }
-
-    /// return value of the iterator
-    typename IteratorType::reference value() const
-    {
-        return anchor.value();
-    }
-};
-
-/// proxy class for the items() function
-template<typename IteratorType> class iteration_proxy
-{
-  private:
-    /// the container to iterate
-    typename IteratorType::pointer container = nullptr;
-
-  public:
-    explicit iteration_proxy() = default;
-
-    /// construct iteration proxy from a container
-    explicit iteration_proxy(typename IteratorType::reference cont) noexcept
-        : container(&cont) {}
-
-    iteration_proxy(iteration_proxy const&) = default;
-    iteration_proxy& operator=(iteration_proxy const&) = default;
-    iteration_proxy(iteration_proxy&&) noexcept = default;
-    iteration_proxy& operator=(iteration_proxy&&) noexcept = default;
-    ~iteration_proxy() = default;
-
-    /// return iterator begin (needed for range-based for)
-    iteration_proxy_value<IteratorType> begin() const noexcept
-    {
-        return iteration_proxy_value<IteratorType>(container->begin());
-    }
-
-    /// return iterator end (needed for range-based for)
-    iteration_proxy_value<IteratorType> end() const noexcept
-    {
-        return iteration_proxy_value<IteratorType>(container->end());
-    }
-};
-
-// Structured Bindings Support
-// For further reference see https://blog.tartanllama.xyz/structured-bindings/
-// And see https://github.com/nlohmann/json/pull/1391
-template<std::size_t N, typename IteratorType, enable_if_t<N == 0, int> = 0>
-auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.key())
-{
-    return i.key();
-}
-// Structured Bindings Support
-// For further reference see https://blog.tartanllama.xyz/structured-bindings/
-// And see https://github.com/nlohmann/json/pull/1391
-template<std::size_t N, typename IteratorType, enable_if_t<N == 1, int> = 0>
-auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.value())
-{
-    return i.value();
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// The Addition to the STD Namespace is required to add
-// Structured Bindings Support to the iteration_proxy_value class
-// For further reference see https://blog.tartanllama.xyz/structured-bindings/
-// And see https://github.com/nlohmann/json/pull/1391
-namespace std
-{
-
-#if defined(__clang__)
-    // Fix: https://github.com/nlohmann/json/issues/1401
-    #pragma clang diagnostic push
-    #pragma clang diagnostic ignored "-Wmismatched-tags"
-#endif
-template<typename IteratorType>
-class tuple_size<::nlohmann::detail::iteration_proxy_value<IteratorType>>
-            : public std::integral_constant<std::size_t, 2> {};
-
-template<std::size_t N, typename IteratorType>
-class tuple_element<N, ::nlohmann::detail::iteration_proxy_value<IteratorType >>
-{
-  public:
-    using type = decltype(
-                     get<N>(std::declval <
-                            ::nlohmann::detail::iteration_proxy_value<IteratorType >> ()));
-};
-#if defined(__clang__)
-    #pragma clang diagnostic pop
-#endif
-
-}  // namespace std
-
-#if JSON_HAS_RANGES
-    template <typename IteratorType>
-    inline constexpr bool ::std::ranges::enable_borrowed_range<::nlohmann::detail::iteration_proxy<IteratorType>> = true;
-#endif
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/std_fs.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-//////////////////
-// constructors //
-//////////////////
-
-/*
- * Note all external_constructor<>::construct functions need to call
- * j.m_value.destroy(j.m_type) to avoid a memory leak in case j contains an
- * allocated value (e.g., a string). See bug issue
- * https://github.com/nlohmann/json/issues/2865 for more information.
- */
-
-template<value_t> struct external_constructor;
-
-template<>
-struct external_constructor<value_t::boolean>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::boolean_t b) noexcept
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::boolean;
-        j.m_value = b;
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::string>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, const typename BasicJsonType::string_t& s)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::string;
-        j.m_value = s;
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::string_t&& s)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::string;
-        j.m_value = std::move(s);
-        j.assert_invariant();
-    }
-
-    template < typename BasicJsonType, typename CompatibleStringType,
-               enable_if_t < !std::is_same<CompatibleStringType, typename BasicJsonType::string_t>::value,
-                             int > = 0 >
-    static void construct(BasicJsonType& j, const CompatibleStringType& str)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::string;
-        j.m_value.string = j.template create<typename BasicJsonType::string_t>(str);
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::binary>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, const typename BasicJsonType::binary_t& b)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::binary;
-        j.m_value = typename BasicJsonType::binary_t(b);
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::binary_t&& b)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::binary;
-        j.m_value = typename BasicJsonType::binary_t(std::move(b));
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::number_float>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::number_float_t val) noexcept
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::number_float;
-        j.m_value = val;
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::number_unsigned>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::number_unsigned_t val) noexcept
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::number_unsigned;
-        j.m_value = val;
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::number_integer>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::number_integer_t val) noexcept
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::number_integer;
-        j.m_value = val;
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::array>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, const typename BasicJsonType::array_t& arr)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::array;
-        j.m_value = arr;
-        j.set_parents();
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::array;
-        j.m_value = std::move(arr);
-        j.set_parents();
-        j.assert_invariant();
-    }
-
-    template < typename BasicJsonType, typename CompatibleArrayType,
-               enable_if_t < !std::is_same<CompatibleArrayType, typename BasicJsonType::array_t>::value,
-                             int > = 0 >
-    static void construct(BasicJsonType& j, const CompatibleArrayType& arr)
-    {
-        using std::begin;
-        using std::end;
-
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::array;
-        j.m_value.array = j.template create<typename BasicJsonType::array_t>(begin(arr), end(arr));
-        j.set_parents();
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, const std::vector<bool>& arr)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::array;
-        j.m_value = value_t::array;
-        j.m_value.array->reserve(arr.size());
-        for (const bool x : arr)
-        {
-            j.m_value.array->push_back(x);
-            j.set_parent(j.m_value.array->back());
-        }
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType, typename T,
-             enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
-    static void construct(BasicJsonType& j, const std::valarray<T>& arr)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::array;
-        j.m_value = value_t::array;
-        j.m_value.array->resize(arr.size());
-        if (arr.size() > 0)
-        {
-            std::copy(std::begin(arr), std::end(arr), j.m_value.array->begin());
-        }
-        j.set_parents();
-        j.assert_invariant();
-    }
-};
-
-template<>
-struct external_constructor<value_t::object>
-{
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, const typename BasicJsonType::object_t& obj)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::object;
-        j.m_value = obj;
-        j.set_parents();
-        j.assert_invariant();
-    }
-
-    template<typename BasicJsonType>
-    static void construct(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
-    {
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::object;
-        j.m_value = std::move(obj);
-        j.set_parents();
-        j.assert_invariant();
-    }
-
-    template < typename BasicJsonType, typename CompatibleObjectType,
-               enable_if_t < !std::is_same<CompatibleObjectType, typename BasicJsonType::object_t>::value, int > = 0 >
-    static void construct(BasicJsonType& j, const CompatibleObjectType& obj)
-    {
-        using std::begin;
-        using std::end;
-
-        j.m_value.destroy(j.m_type);
-        j.m_type = value_t::object;
-        j.m_value.object = j.template create<typename BasicJsonType::object_t>(begin(obj), end(obj));
-        j.set_parents();
-        j.assert_invariant();
-    }
-};
-
-/////////////
-// to_json //
-/////////////
-
-template<typename BasicJsonType, typename T,
-         enable_if_t<std::is_same<T, typename BasicJsonType::boolean_t>::value, int> = 0>
-inline void to_json(BasicJsonType& j, T b) noexcept
-{
-    external_constructor<value_t::boolean>::construct(j, b);
-}
-
-template < typename BasicJsonType, typename BoolRef,
-           enable_if_t <
-               ((std::is_same<std::vector<bool>::reference, BoolRef>::value
-                 && !std::is_same <std::vector<bool>::reference, typename BasicJsonType::boolean_t&>::value)
-                || (std::is_same<std::vector<bool>::const_reference, BoolRef>::value
-                    && !std::is_same <detail::uncvref_t<std::vector<bool>::const_reference>,
-                                      typename BasicJsonType::boolean_t >::value))
-               && std::is_convertible<const BoolRef&, typename BasicJsonType::boolean_t>::value, int > = 0 >
-inline void to_json(BasicJsonType& j, const BoolRef& b) noexcept
-{
-    external_constructor<value_t::boolean>::construct(j, static_cast<typename BasicJsonType::boolean_t>(b));
-}
-
-template<typename BasicJsonType, typename CompatibleString,
-         enable_if_t<std::is_constructible<typename BasicJsonType::string_t, CompatibleString>::value, int> = 0>
-inline void to_json(BasicJsonType& j, const CompatibleString& s)
-{
-    external_constructor<value_t::string>::construct(j, s);
-}
-
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, typename BasicJsonType::string_t&& s)
-{
-    external_constructor<value_t::string>::construct(j, std::move(s));
-}
-
-template<typename BasicJsonType, typename FloatType,
-         enable_if_t<std::is_floating_point<FloatType>::value, int> = 0>
-inline void to_json(BasicJsonType& j, FloatType val) noexcept
-{
-    external_constructor<value_t::number_float>::construct(j, static_cast<typename BasicJsonType::number_float_t>(val));
-}
-
-template<typename BasicJsonType, typename CompatibleNumberUnsignedType,
-         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_unsigned_t, CompatibleNumberUnsignedType>::value, int> = 0>
-inline void to_json(BasicJsonType& j, CompatibleNumberUnsignedType val) noexcept
-{
-    external_constructor<value_t::number_unsigned>::construct(j, static_cast<typename BasicJsonType::number_unsigned_t>(val));
-}
-
-template<typename BasicJsonType, typename CompatibleNumberIntegerType,
-         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType>::value, int> = 0>
-inline void to_json(BasicJsonType& j, CompatibleNumberIntegerType val) noexcept
-{
-    external_constructor<value_t::number_integer>::construct(j, static_cast<typename BasicJsonType::number_integer_t>(val));
-}
-
-#if !JSON_DISABLE_ENUM_SERIALIZATION
-template<typename BasicJsonType, typename EnumType,
-         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
-inline void to_json(BasicJsonType& j, EnumType e) noexcept
-{
-    using underlying_type = typename std::underlying_type<EnumType>::type;
-    external_constructor<value_t::number_integer>::construct(j, static_cast<underlying_type>(e));
-}
-#endif  // JSON_DISABLE_ENUM_SERIALIZATION
-
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, const std::vector<bool>& e)
-{
-    external_constructor<value_t::array>::construct(j, e);
-}
-
-template < typename BasicJsonType, typename CompatibleArrayType,
-           enable_if_t < is_compatible_array_type<BasicJsonType,
-                         CompatibleArrayType>::value&&
-                         !is_compatible_object_type<BasicJsonType, CompatibleArrayType>::value&&
-                         !is_compatible_string_type<BasicJsonType, CompatibleArrayType>::value&&
-                         !std::is_same<typename BasicJsonType::binary_t, CompatibleArrayType>::value&&
-                         !is_basic_json<CompatibleArrayType>::value,
-                         int > = 0 >
-inline void to_json(BasicJsonType& j, const CompatibleArrayType& arr)
-{
-    external_constructor<value_t::array>::construct(j, arr);
-}
-
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, const typename BasicJsonType::binary_t& bin)
-{
-    external_constructor<value_t::binary>::construct(j, bin);
-}
-
-template<typename BasicJsonType, typename T,
-         enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
-inline void to_json(BasicJsonType& j, const std::valarray<T>& arr)
-{
-    external_constructor<value_t::array>::construct(j, std::move(arr));
-}
-
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
-{
-    external_constructor<value_t::array>::construct(j, std::move(arr));
-}
-
-template < typename BasicJsonType, typename CompatibleObjectType,
-           enable_if_t < is_compatible_object_type<BasicJsonType, CompatibleObjectType>::value&& !is_basic_json<CompatibleObjectType>::value, int > = 0 >
-inline void to_json(BasicJsonType& j, const CompatibleObjectType& obj)
-{
-    external_constructor<value_t::object>::construct(j, obj);
-}
-
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
-{
-    external_constructor<value_t::object>::construct(j, std::move(obj));
-}
-
-template <
-    typename BasicJsonType, typename T, std::size_t N,
-    enable_if_t < !std::is_constructible<typename BasicJsonType::string_t,
-                  const T(&)[N]>::value, // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-                  int > = 0 >
-inline void to_json(BasicJsonType& j, const T(&arr)[N]) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-{
-    external_constructor<value_t::array>::construct(j, arr);
-}
-
-template < typename BasicJsonType, typename T1, typename T2, enable_if_t < std::is_constructible<BasicJsonType, T1>::value&& std::is_constructible<BasicJsonType, T2>::value, int > = 0 >
-inline void to_json(BasicJsonType& j, const std::pair<T1, T2>& p)
-{
-    j = { p.first, p.second };
-}
-
-// for https://github.com/nlohmann/json/pull/1134
-template<typename BasicJsonType, typename T,
-         enable_if_t<std::is_same<T, iteration_proxy_value<typename BasicJsonType::iterator>>::value, int> = 0>
-inline void to_json(BasicJsonType& j, const T& b)
-{
-    j = { {b.key(), b.value()} };
-}
-
-template<typename BasicJsonType, typename Tuple, std::size_t... Idx>
-inline void to_json_tuple_impl(BasicJsonType& j, const Tuple& t, index_sequence<Idx...> /*unused*/)
-{
-    j = { std::get<Idx>(t)... };
-}
-
-template<typename BasicJsonType, typename T, enable_if_t<is_constructible_tuple<BasicJsonType, T>::value, int > = 0>
-inline void to_json(BasicJsonType& j, const T& t)
-{
-    to_json_tuple_impl(j, t, make_index_sequence<std::tuple_size<T>::value> {});
-}
-
-#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
-template<typename BasicJsonType>
-inline void to_json(BasicJsonType& j, const std_fs::path& p)
-{
-    j = p.string();
-}
-#endif
-
-struct to_json_fn
-{
-    template<typename BasicJsonType, typename T>
-    auto operator()(BasicJsonType& j, T&& val) const noexcept(noexcept(to_json(j, std::forward<T>(val))))
-    -> decltype(to_json(j, std::forward<T>(val)), void())
-    {
-        return to_json(j, std::forward<T>(val));
-    }
-};
-}  // namespace detail
-
-#ifndef JSON_HAS_CPP_17
-/// namespace to hold default `to_json` function
-/// to see why this is required:
-/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
-namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
-{
-#endif
-JSON_INLINE_VARIABLE constexpr const auto& to_json = // NOLINT(misc-definitions-in-headers)
-    detail::static_const<detail::to_json_fn>::value;
-#ifndef JSON_HAS_CPP_17
-}  // namespace
-#endif
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/identity_tag.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/// @sa https://json.nlohmann.me/api/adl_serializer/
-template<typename ValueType, typename>
-struct adl_serializer
-{
-    /// @brief convert a JSON value to any value type
-    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
-    template<typename BasicJsonType, typename TargetType = ValueType>
-    static auto from_json(BasicJsonType && j, TargetType& val) noexcept(
-        noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), val)))
-    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), val), void())
-    {
-        ::nlohmann::from_json(std::forward<BasicJsonType>(j), val);
-    }
-
-    /// @brief convert a JSON value to any value type
-    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
-    template<typename BasicJsonType, typename TargetType = ValueType>
-    static auto from_json(BasicJsonType && j) noexcept(
-    noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {})))
-    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {}))
-    {
-        return ::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {});
-    }
-
-    /// @brief convert any value type to a JSON value
-    /// @sa https://json.nlohmann.me/api/adl_serializer/to_json/
-    template<typename BasicJsonType, typename TargetType = ValueType>
-    static auto to_json(BasicJsonType& j, TargetType && val) noexcept(
-        noexcept(::nlohmann::to_json(j, std::forward<TargetType>(val))))
-    -> decltype(::nlohmann::to_json(j, std::forward<TargetType>(val)), void())
-    {
-        ::nlohmann::to_json(j, std::forward<TargetType>(val));
-    }
-};
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/byte_container_with_subtype.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstdint> // uint8_t, uint64_t
-#include <tuple> // tie
-#include <utility> // move
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/// @brief an internal type for a backed binary type
-/// @sa https://json.nlohmann.me/api/byte_container_with_subtype/
-template<typename BinaryType>
-class byte_container_with_subtype : public BinaryType
-{
-  public:
-    using container_type = BinaryType;
-    using subtype_type = std::uint64_t;
-
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
-    byte_container_with_subtype() noexcept(noexcept(container_type()))
-        : container_type()
-    {}
-
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
-    byte_container_with_subtype(const container_type& b) noexcept(noexcept(container_type(b)))
-        : container_type(b)
-    {}
-
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
-    byte_container_with_subtype(container_type&& b) noexcept(noexcept(container_type(std::move(b))))
-        : container_type(std::move(b))
-    {}
-
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
-    byte_container_with_subtype(const container_type& b, subtype_type subtype_) noexcept(noexcept(container_type(b)))
-        : container_type(b)
-        , m_subtype(subtype_)
-        , m_has_subtype(true)
-    {}
-
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
-    byte_container_with_subtype(container_type&& b, subtype_type subtype_) noexcept(noexcept(container_type(std::move(b))))
-        : container_type(std::move(b))
-        , m_subtype(subtype_)
-        , m_has_subtype(true)
-    {}
-
-    bool operator==(const byte_container_with_subtype& rhs) const
-    {
-        return std::tie(static_cast<const BinaryType&>(*this), m_subtype, m_has_subtype) ==
-               std::tie(static_cast<const BinaryType&>(rhs), rhs.m_subtype, rhs.m_has_subtype);
-    }
-
-    bool operator!=(const byte_container_with_subtype& rhs) const
-    {
-        return !(rhs == *this);
-    }
-
-    /// @brief sets the binary subtype
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/set_subtype/
-    void set_subtype(subtype_type subtype_) noexcept
-    {
-        m_subtype = subtype_;
-        m_has_subtype = true;
-    }
-
-    /// @brief return the binary subtype
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/subtype/
-    constexpr subtype_type subtype() const noexcept
-    {
-        return m_has_subtype ? m_subtype : static_cast<subtype_type>(-1);
-    }
-
-    /// @brief return whether the value has a subtype
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/has_subtype/
-    constexpr bool has_subtype() const noexcept
-    {
-        return m_has_subtype;
-    }
-
-    /// @brief clears the binary subtype
-    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/clear_subtype/
-    void clear_subtype() noexcept
-    {
-        m_subtype = 0;
-        m_has_subtype = false;
-    }
-
-  private:
-    subtype_type m_subtype = 0;
-    bool m_has_subtype = false;
-};
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/conversions/from_json.hpp>
-
-// #include <nlohmann/detail/conversions/to_json.hpp>
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/hash.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstdint> // uint8_t
-#include <cstddef> // size_t
-#include <functional> // hash
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-// boost::hash_combine
-inline std::size_t combine(std::size_t seed, std::size_t h) noexcept
-{
-    seed ^= h + 0x9e3779b9 + (seed << 6U) + (seed >> 2U);
-    return seed;
-}
-
-/*!
-@brief hash a JSON value
-
-The hash function tries to rely on std::hash where possible. Furthermore, the
-type of the JSON value is taken into account to have different hash values for
-null, 0, 0U, and false, etc.
-
-@tparam BasicJsonType basic_json specialization
-@param j JSON value to hash
-@return hash value of j
-*/
-template<typename BasicJsonType>
-std::size_t hash(const BasicJsonType& j)
-{
-    using string_t = typename BasicJsonType::string_t;
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-
-    const auto type = static_cast<std::size_t>(j.type());
-    switch (j.type())
-    {
-        case BasicJsonType::value_t::null:
-        case BasicJsonType::value_t::discarded:
-        {
-            return combine(type, 0);
-        }
-
-        case BasicJsonType::value_t::object:
-        {
-            auto seed = combine(type, j.size());
-            for (const auto& element : j.items())
-            {
-                const auto h = std::hash<string_t> {}(element.key());
-                seed = combine(seed, h);
-                seed = combine(seed, hash(element.value()));
-            }
-            return seed;
-        }
-
-        case BasicJsonType::value_t::array:
-        {
-            auto seed = combine(type, j.size());
-            for (const auto& element : j)
-            {
-                seed = combine(seed, hash(element));
-            }
-            return seed;
-        }
-
-        case BasicJsonType::value_t::string:
-        {
-            const auto h = std::hash<string_t> {}(j.template get_ref<const string_t&>());
-            return combine(type, h);
-        }
-
-        case BasicJsonType::value_t::boolean:
-        {
-            const auto h = std::hash<bool> {}(j.template get<bool>());
-            return combine(type, h);
-        }
-
-        case BasicJsonType::value_t::number_integer:
-        {
-            const auto h = std::hash<number_integer_t> {}(j.template get<number_integer_t>());
-            return combine(type, h);
-        }
-
-        case BasicJsonType::value_t::number_unsigned:
-        {
-            const auto h = std::hash<number_unsigned_t> {}(j.template get<number_unsigned_t>());
-            return combine(type, h);
-        }
-
-        case BasicJsonType::value_t::number_float:
-        {
-            const auto h = std::hash<number_float_t> {}(j.template get<number_float_t>());
-            return combine(type, h);
-        }
-
-        case BasicJsonType::value_t::binary:
-        {
-            auto seed = combine(type, j.get_binary().size());
-            const auto h = std::hash<bool> {}(j.get_binary().has_subtype());
-            seed = combine(seed, h);
-            seed = combine(seed, static_cast<std::size_t>(j.get_binary().subtype()));
-            for (const auto byte : j.get_binary())
-            {
-                seed = combine(seed, std::hash<std::uint8_t> {}(byte));
-            }
-            return seed;
-        }
-
-        default:                   // LCOV_EXCL_LINE
-            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-            return 0;              // LCOV_EXCL_LINE
-    }
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/input/binary_reader.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // generate_n
-#include <array> // array
-#include <cmath> // ldexp
-#include <cstddef> // size_t
-#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
-#include <cstdio> // snprintf
-#include <cstring> // memcpy
-#include <iterator> // back_inserter
-#include <limits> // numeric_limits
-#include <string> // char_traits, string
-#include <utility> // make_pair, move
-#include <vector> // vector
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/input/input_adapters.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <array> // array
-#include <cstddef> // size_t
-#include <cstring> // strlen
-#include <iterator> // begin, end, iterator_traits, random_access_iterator_tag, distance, next
-#include <memory> // shared_ptr, make_shared, addressof
-#include <numeric> // accumulate
-#include <string> // string, char_traits
-#include <type_traits> // enable_if, is_base_of, is_pointer, is_integral, remove_pointer
-#include <utility> // pair, declval
-
-#ifndef JSON_NO_IO
-    #include <cstdio>   // FILE *
-    #include <istream>  // istream
-#endif                  // JSON_NO_IO
-
-// #include <nlohmann/detail/iterators/iterator_traits.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/// the supported input formats
-enum class input_format_t { json, cbor, msgpack, ubjson, bson, bjdata };
-
-////////////////////
-// input adapters //
-////////////////////
-
-#ifndef JSON_NO_IO
-/*!
-Input adapter for stdio file access. This adapter read only 1 byte and do not use any
- buffer. This adapter is a very low level adapter.
-*/
-class file_input_adapter
-{
-  public:
-    using char_type = char;
-
-    JSON_HEDLEY_NON_NULL(2)
-    explicit file_input_adapter(std::FILE* f) noexcept
-        : m_file(f)
-    {
-        JSON_ASSERT(m_file != nullptr);
-    }
-
-    // make class move-only
-    file_input_adapter(const file_input_adapter&) = delete;
-    file_input_adapter(file_input_adapter&&) noexcept = default;
-    file_input_adapter& operator=(const file_input_adapter&) = delete;
-    file_input_adapter& operator=(file_input_adapter&&) = delete;
-    ~file_input_adapter() = default;
-
-    std::char_traits<char>::int_type get_character() noexcept
-    {
-        return std::fgetc(m_file);
-    }
-
-  private:
-    /// the file pointer to read from
-    std::FILE* m_file;
-};
-
-
-/*!
-Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at
-beginning of input. Does not support changing the underlying std::streambuf
-in mid-input. Maintains underlying std::istream and std::streambuf to support
-subsequent use of standard std::istream operations to process any input
-characters following those used in parsing the JSON input.  Clears the
-std::istream flags; any input errors (e.g., EOF) will be detected by the first
-subsequent call for input from the std::istream.
-*/
-class input_stream_adapter
-{
-  public:
-    using char_type = char;
-
-    ~input_stream_adapter()
-    {
-        // clear stream flags; we use underlying streambuf I/O, do not
-        // maintain ifstream flags, except eof
-        if (is != nullptr)
-        {
-            is->clear(is->rdstate() & std::ios::eofbit);
-        }
-    }
-
-    explicit input_stream_adapter(std::istream& i)
-        : is(&i), sb(i.rdbuf())
-    {}
-
-    // delete because of pointer members
-    input_stream_adapter(const input_stream_adapter&) = delete;
-    input_stream_adapter& operator=(input_stream_adapter&) = delete;
-    input_stream_adapter& operator=(input_stream_adapter&&) = delete;
-
-    input_stream_adapter(input_stream_adapter&& rhs) noexcept
-        : is(rhs.is), sb(rhs.sb)
-    {
-        rhs.is = nullptr;
-        rhs.sb = nullptr;
-    }
-
-    // std::istream/std::streambuf use std::char_traits<char>::to_int_type, to
-    // ensure that std::char_traits<char>::eof() and the character 0xFF do not
-    // end up as the same value, e.g. 0xFFFFFFFF.
-    std::char_traits<char>::int_type get_character()
-    {
-        auto res = sb->sbumpc();
-        // set eof manually, as we don't use the istream interface.
-        if (JSON_HEDLEY_UNLIKELY(res == std::char_traits<char>::eof()))
-        {
-            is->clear(is->rdstate() | std::ios::eofbit);
-        }
-        return res;
-    }
-
-  private:
-    /// the associated input stream
-    std::istream* is = nullptr;
-    std::streambuf* sb = nullptr;
-};
-#endif  // JSON_NO_IO
-
-// General-purpose iterator-based adapter. It might not be as fast as
-// theoretically possible for some containers, but it is extremely versatile.
-template<typename IteratorType>
-class iterator_input_adapter
-{
-  public:
-    using char_type = typename std::iterator_traits<IteratorType>::value_type;
-
-    iterator_input_adapter(IteratorType first, IteratorType last)
-        : current(std::move(first)), end(std::move(last))
-    {}
-
-    typename std::char_traits<char_type>::int_type get_character()
-    {
-        if (JSON_HEDLEY_LIKELY(current != end))
-        {
-            auto result = std::char_traits<char_type>::to_int_type(*current);
-            std::advance(current, 1);
-            return result;
-        }
-
-        return std::char_traits<char_type>::eof();
-    }
-
-  private:
-    IteratorType current;
-    IteratorType end;
-
-    template<typename BaseInputAdapter, size_t T>
-    friend struct wide_string_input_helper;
-
-    bool empty() const
-    {
-        return current == end;
-    }
-};
-
-
-template<typename BaseInputAdapter, size_t T>
-struct wide_string_input_helper;
-
-template<typename BaseInputAdapter>
-struct wide_string_input_helper<BaseInputAdapter, 4>
-{
-    // UTF-32
-    static void fill_buffer(BaseInputAdapter& input,
-                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
-                            size_t& utf8_bytes_index,
-                            size_t& utf8_bytes_filled)
-    {
-        utf8_bytes_index = 0;
-
-        if (JSON_HEDLEY_UNLIKELY(input.empty()))
-        {
-            utf8_bytes[0] = std::char_traits<char>::eof();
-            utf8_bytes_filled = 1;
-        }
-        else
-        {
-            // get the current character
-            const auto wc = input.get_character();
-
-            // UTF-32 to UTF-8 encoding
-            if (wc < 0x80)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
-                utf8_bytes_filled = 1;
-            }
-            else if (wc <= 0x7FF)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u) & 0x1Fu));
-                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
-                utf8_bytes_filled = 2;
-            }
-            else if (wc <= 0xFFFF)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u) & 0x0Fu));
-                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
-                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
-                utf8_bytes_filled = 3;
-            }
-            else if (wc <= 0x10FFFF)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | ((static_cast<unsigned int>(wc) >> 18u) & 0x07u));
-                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 12u) & 0x3Fu));
-                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
-                utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
-                utf8_bytes_filled = 4;
-            }
-            else
-            {
-                // unknown character
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
-                utf8_bytes_filled = 1;
-            }
-        }
-    }
-};
-
-template<typename BaseInputAdapter>
-struct wide_string_input_helper<BaseInputAdapter, 2>
-{
-    // UTF-16
-    static void fill_buffer(BaseInputAdapter& input,
-                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
-                            size_t& utf8_bytes_index,
-                            size_t& utf8_bytes_filled)
-    {
-        utf8_bytes_index = 0;
-
-        if (JSON_HEDLEY_UNLIKELY(input.empty()))
-        {
-            utf8_bytes[0] = std::char_traits<char>::eof();
-            utf8_bytes_filled = 1;
-        }
-        else
-        {
-            // get the current character
-            const auto wc = input.get_character();
-
-            // UTF-16 to UTF-8 encoding
-            if (wc < 0x80)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
-                utf8_bytes_filled = 1;
-            }
-            else if (wc <= 0x7FF)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u)));
-                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
-                utf8_bytes_filled = 2;
-            }
-            else if (0xD800 > wc || wc >= 0xE000)
-            {
-                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u)));
-                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
-                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
-                utf8_bytes_filled = 3;
-            }
-            else
-            {
-                if (JSON_HEDLEY_UNLIKELY(!input.empty()))
-                {
-                    const auto wc2 = static_cast<unsigned int>(input.get_character());
-                    const auto charcode = 0x10000u + (((static_cast<unsigned int>(wc) & 0x3FFu) << 10u) | (wc2 & 0x3FFu));
-                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | (charcode >> 18u));
-                    utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 12u) & 0x3Fu));
-                    utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 6u) & 0x3Fu));
-                    utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (charcode & 0x3Fu));
-                    utf8_bytes_filled = 4;
-                }
-                else
-                {
-                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
-                    utf8_bytes_filled = 1;
-                }
-            }
-        }
-    }
-};
-
-// Wraps another input apdater to convert wide character types into individual bytes.
-template<typename BaseInputAdapter, typename WideCharType>
-class wide_string_input_adapter
-{
-  public:
-    using char_type = char;
-
-    wide_string_input_adapter(BaseInputAdapter base)
-        : base_adapter(base) {}
-
-    typename std::char_traits<char>::int_type get_character() noexcept
-    {
-        // check if buffer needs to be filled
-        if (utf8_bytes_index == utf8_bytes_filled)
-        {
-            fill_buffer<sizeof(WideCharType)>();
-
-            JSON_ASSERT(utf8_bytes_filled > 0);
-            JSON_ASSERT(utf8_bytes_index == 0);
-        }
-
-        // use buffer
-        JSON_ASSERT(utf8_bytes_filled > 0);
-        JSON_ASSERT(utf8_bytes_index < utf8_bytes_filled);
-        return utf8_bytes[utf8_bytes_index++];
-    }
-
-  private:
-    BaseInputAdapter base_adapter;
-
-    template<size_t T>
-    void fill_buffer()
-    {
-        wide_string_input_helper<BaseInputAdapter, T>::fill_buffer(base_adapter, utf8_bytes, utf8_bytes_index, utf8_bytes_filled);
-    }
-
-    /// a buffer for UTF-8 bytes
-    std::array<std::char_traits<char>::int_type, 4> utf8_bytes = {{0, 0, 0, 0}};
-
-    /// index to the utf8_codes array for the next valid byte
-    std::size_t utf8_bytes_index = 0;
-    /// number of valid bytes in the utf8_codes array
-    std::size_t utf8_bytes_filled = 0;
-};
-
-
-template<typename IteratorType, typename Enable = void>
-struct iterator_input_adapter_factory
-{
-    using iterator_type = IteratorType;
-    using char_type = typename std::iterator_traits<iterator_type>::value_type;
-    using adapter_type = iterator_input_adapter<iterator_type>;
-
-    static adapter_type create(IteratorType first, IteratorType last)
-    {
-        return adapter_type(std::move(first), std::move(last));
-    }
-};
-
-template<typename T>
-struct is_iterator_of_multibyte
-{
-    using value_type = typename std::iterator_traits<T>::value_type;
-    enum
-    {
-        value = sizeof(value_type) > 1
-    };
-};
-
-template<typename IteratorType>
-struct iterator_input_adapter_factory<IteratorType, enable_if_t<is_iterator_of_multibyte<IteratorType>::value>>
-{
-    using iterator_type = IteratorType;
-    using char_type = typename std::iterator_traits<iterator_type>::value_type;
-    using base_adapter_type = iterator_input_adapter<iterator_type>;
-    using adapter_type = wide_string_input_adapter<base_adapter_type, char_type>;
-
-    static adapter_type create(IteratorType first, IteratorType last)
-    {
-        return adapter_type(base_adapter_type(std::move(first), std::move(last)));
-    }
-};
-
-// General purpose iterator-based input
-template<typename IteratorType>
-typename iterator_input_adapter_factory<IteratorType>::adapter_type input_adapter(IteratorType first, IteratorType last)
-{
-    using factory_type = iterator_input_adapter_factory<IteratorType>;
-    return factory_type::create(first, last);
-}
-
-// Convenience shorthand from container to iterator
-// Enables ADL on begin(container) and end(container)
-// Encloses the using declarations in namespace for not to leak them to outside scope
-
-namespace container_input_adapter_factory_impl
-{
-
-using std::begin;
-using std::end;
-
-template<typename ContainerType, typename Enable = void>
-struct container_input_adapter_factory {};
-
-template<typename ContainerType>
-struct container_input_adapter_factory< ContainerType,
-       void_t<decltype(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>()))>>
-       {
-           using adapter_type = decltype(input_adapter(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>())));
-
-           static adapter_type create(const ContainerType& container)
-{
-    return input_adapter(begin(container), end(container));
-}
-       };
-
-}  // namespace container_input_adapter_factory_impl
-
-template<typename ContainerType>
-typename container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::adapter_type input_adapter(const ContainerType& container)
-{
-    return container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::create(container);
-}
-
-#ifndef JSON_NO_IO
-// Special cases with fast paths
-inline file_input_adapter input_adapter(std::FILE* file)
-{
-    return file_input_adapter(file);
-}
-
-inline input_stream_adapter input_adapter(std::istream& stream)
-{
-    return input_stream_adapter(stream);
-}
-
-inline input_stream_adapter input_adapter(std::istream&& stream)
-{
-    return input_stream_adapter(stream);
-}
-#endif  // JSON_NO_IO
-
-using contiguous_bytes_input_adapter = decltype(input_adapter(std::declval<const char*>(), std::declval<const char*>()));
-
-// Null-delimited strings, and the like.
-template < typename CharT,
-           typename std::enable_if <
-               std::is_pointer<CharT>::value&&
-               !std::is_array<CharT>::value&&
-               std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
-               sizeof(typename std::remove_pointer<CharT>::type) == 1,
-               int >::type = 0 >
-contiguous_bytes_input_adapter input_adapter(CharT b)
-{
-    auto length = std::strlen(reinterpret_cast<const char*>(b));
-    const auto* ptr = reinterpret_cast<const char*>(b);
-    return input_adapter(ptr, ptr + length);
-}
-
-template<typename T, std::size_t N>
-auto input_adapter(T (&array)[N]) -> decltype(input_adapter(array, array + N)) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-{
-    return input_adapter(array, array + N);
-}
-
-// This class only handles inputs of input_buffer_adapter type.
-// It's required so that expressions like {ptr, len} can be implicitly cast
-// to the correct adapter.
-class span_input_adapter
-{
-  public:
-    template < typename CharT,
-               typename std::enable_if <
-                   std::is_pointer<CharT>::value&&
-                   std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
-                   sizeof(typename std::remove_pointer<CharT>::type) == 1,
-                   int >::type = 0 >
-    span_input_adapter(CharT b, std::size_t l)
-        : ia(reinterpret_cast<const char*>(b), reinterpret_cast<const char*>(b) + l) {}
-
-    template<class IteratorType,
-             typename std::enable_if<
-                 std::is_same<typename iterator_traits<IteratorType>::iterator_category, std::random_access_iterator_tag>::value,
-                 int>::type = 0>
-    span_input_adapter(IteratorType first, IteratorType last)
-        : ia(input_adapter(first, last)) {}
-
-    contiguous_bytes_input_adapter&& get()
-    {
-        return std::move(ia); // NOLINT(hicpp-move-const-arg,performance-move-const-arg)
-    }
-
-  private:
-    contiguous_bytes_input_adapter ia;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/input/json_sax.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef>
-#include <string> // string
-#include <utility> // move
-#include <vector> // vector
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/*!
-@brief SAX interface
-
-This class describes the SAX interface used by @ref nlohmann::json::sax_parse.
-Each function is called in different situations while the input is parsed. The
-boolean return value informs the parser whether to continue processing the
-input.
-*/
-template<typename BasicJsonType>
-struct json_sax
-{
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-
-    /*!
-    @brief a null value was read
-    @return whether parsing should proceed
-    */
-    virtual bool null() = 0;
-
-    /*!
-    @brief a boolean value was read
-    @param[in] val  boolean value
-    @return whether parsing should proceed
-    */
-    virtual bool boolean(bool val) = 0;
-
-    /*!
-    @brief an integer number was read
-    @param[in] val  integer value
-    @return whether parsing should proceed
-    */
-    virtual bool number_integer(number_integer_t val) = 0;
-
-    /*!
-    @brief an unsigned integer number was read
-    @param[in] val  unsigned integer value
-    @return whether parsing should proceed
-    */
-    virtual bool number_unsigned(number_unsigned_t val) = 0;
-
-    /*!
-    @brief a floating-point number was read
-    @param[in] val  floating-point value
-    @param[in] s    raw token value
-    @return whether parsing should proceed
-    */
-    virtual bool number_float(number_float_t val, const string_t& s) = 0;
-
-    /*!
-    @brief a string value was read
-    @param[in] val  string value
-    @return whether parsing should proceed
-    @note It is safe to move the passed string value.
-    */
-    virtual bool string(string_t& val) = 0;
-
-    /*!
-    @brief a binary value was read
-    @param[in] val  binary value
-    @return whether parsing should proceed
-    @note It is safe to move the passed binary value.
-    */
-    virtual bool binary(binary_t& val) = 0;
-
-    /*!
-    @brief the beginning of an object was read
-    @param[in] elements  number of object elements or -1 if unknown
-    @return whether parsing should proceed
-    @note binary formats may report the number of elements
-    */
-    virtual bool start_object(std::size_t elements) = 0;
-
-    /*!
-    @brief an object key was read
-    @param[in] val  object key
-    @return whether parsing should proceed
-    @note It is safe to move the passed string.
-    */
-    virtual bool key(string_t& val) = 0;
-
-    /*!
-    @brief the end of an object was read
-    @return whether parsing should proceed
-    */
-    virtual bool end_object() = 0;
-
-    /*!
-    @brief the beginning of an array was read
-    @param[in] elements  number of array elements or -1 if unknown
-    @return whether parsing should proceed
-    @note binary formats may report the number of elements
-    */
-    virtual bool start_array(std::size_t elements) = 0;
-
-    /*!
-    @brief the end of an array was read
-    @return whether parsing should proceed
-    */
-    virtual bool end_array() = 0;
-
-    /*!
-    @brief a parse error occurred
-    @param[in] position    the position in the input where the error occurs
-    @param[in] last_token  the last read token
-    @param[in] ex          an exception object describing the error
-    @return whether parsing should proceed (must return false)
-    */
-    virtual bool parse_error(std::size_t position,
-                             const std::string& last_token,
-                             const detail::exception& ex) = 0;
-
-    json_sax() = default;
-    json_sax(const json_sax&) = default;
-    json_sax(json_sax&&) noexcept = default;
-    json_sax& operator=(const json_sax&) = default;
-    json_sax& operator=(json_sax&&) noexcept = default;
-    virtual ~json_sax() = default;
-};
-
-
-namespace detail
-{
-/*!
-@brief SAX implementation to create a JSON value from SAX events
-
-This class implements the @ref json_sax interface and processes the SAX events
-to create a JSON value which makes it basically a DOM parser. The structure or
-hierarchy of the JSON value is managed by the stack `ref_stack` which contains
-a pointer to the respective array or object for each recursion depth.
-
-After successful parsing, the value that is passed by reference to the
-constructor contains the parsed value.
-
-@tparam BasicJsonType  the JSON type
-*/
-template<typename BasicJsonType>
-class json_sax_dom_parser
-{
-  public:
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-
-    /*!
-    @param[in,out] r  reference to a JSON value that is manipulated while
-                       parsing
-    @param[in] allow_exceptions_  whether parse errors yield exceptions
-    */
-    explicit json_sax_dom_parser(BasicJsonType& r, const bool allow_exceptions_ = true)
-        : root(r), allow_exceptions(allow_exceptions_)
-    {}
-
-    // make class move-only
-    json_sax_dom_parser(const json_sax_dom_parser&) = delete;
-    json_sax_dom_parser(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    json_sax_dom_parser& operator=(const json_sax_dom_parser&) = delete;
-    json_sax_dom_parser& operator=(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    ~json_sax_dom_parser() = default;
-
-    bool null()
-    {
-        handle_value(nullptr);
-        return true;
-    }
-
-    bool boolean(bool val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_integer(number_integer_t val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_unsigned(number_unsigned_t val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_float(number_float_t val, const string_t& /*unused*/)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool string(string_t& val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool binary(binary_t& val)
-    {
-        handle_value(std::move(val));
-        return true;
-    }
-
-    bool start_object(std::size_t len)
-    {
-        ref_stack.push_back(handle_value(BasicJsonType::value_t::object));
-
-        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
-        {
-            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
-        }
-
-        return true;
-    }
-
-    bool key(string_t& val)
-    {
-        JSON_ASSERT(!ref_stack.empty());
-        JSON_ASSERT(ref_stack.back()->is_object());
-
-        // add null at given key and store the reference for later
-        object_element = &(ref_stack.back()->m_value.object->operator[](val));
-        return true;
-    }
-
-    bool end_object()
-    {
-        JSON_ASSERT(!ref_stack.empty());
-        JSON_ASSERT(ref_stack.back()->is_object());
-
-        ref_stack.back()->set_parents();
-        ref_stack.pop_back();
-        return true;
-    }
-
-    bool start_array(std::size_t len)
-    {
-        ref_stack.push_back(handle_value(BasicJsonType::value_t::array));
-
-        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
-        {
-            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
-        }
-
-        return true;
-    }
-
-    bool end_array()
-    {
-        JSON_ASSERT(!ref_stack.empty());
-        JSON_ASSERT(ref_stack.back()->is_array());
-
-        ref_stack.back()->set_parents();
-        ref_stack.pop_back();
-        return true;
-    }
-
-    template<class Exception>
-    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
-                     const Exception& ex)
-    {
-        errored = true;
-        static_cast<void>(ex);
-        if (allow_exceptions)
-        {
-            JSON_THROW(ex);
-        }
-        return false;
-    }
-
-    constexpr bool is_errored() const
-    {
-        return errored;
-    }
-
-  private:
-    /*!
-    @invariant If the ref stack is empty, then the passed value will be the new
-               root.
-    @invariant If the ref stack contains a value, then it is an array or an
-               object to which we can add elements
-    */
-    template<typename Value>
-    JSON_HEDLEY_RETURNS_NON_NULL
-    BasicJsonType* handle_value(Value&& v)
-    {
-        if (ref_stack.empty())
-        {
-            root = BasicJsonType(std::forward<Value>(v));
-            return &root;
-        }
-
-        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
-
-        if (ref_stack.back()->is_array())
-        {
-            ref_stack.back()->m_value.array->emplace_back(std::forward<Value>(v));
-            return &(ref_stack.back()->m_value.array->back());
-        }
-
-        JSON_ASSERT(ref_stack.back()->is_object());
-        JSON_ASSERT(object_element);
-        *object_element = BasicJsonType(std::forward<Value>(v));
-        return object_element;
-    }
-
-    /// the parsed JSON value
-    BasicJsonType& root;
-    /// stack to model hierarchy of values
-    std::vector<BasicJsonType*> ref_stack {};
-    /// helper to hold the reference for the next object element
-    BasicJsonType* object_element = nullptr;
-    /// whether a syntax error occurred
-    bool errored = false;
-    /// whether to throw exceptions in case of errors
-    const bool allow_exceptions = true;
-};
-
-template<typename BasicJsonType>
-class json_sax_dom_callback_parser
-{
-  public:
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-    using parser_callback_t = typename BasicJsonType::parser_callback_t;
-    using parse_event_t = typename BasicJsonType::parse_event_t;
-
-    json_sax_dom_callback_parser(BasicJsonType& r,
-                                 const parser_callback_t cb,
-                                 const bool allow_exceptions_ = true)
-        : root(r), callback(cb), allow_exceptions(allow_exceptions_)
-    {
-        keep_stack.push_back(true);
-    }
-
-    // make class move-only
-    json_sax_dom_callback_parser(const json_sax_dom_callback_parser&) = delete;
-    json_sax_dom_callback_parser(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    json_sax_dom_callback_parser& operator=(const json_sax_dom_callback_parser&) = delete;
-    json_sax_dom_callback_parser& operator=(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    ~json_sax_dom_callback_parser() = default;
-
-    bool null()
-    {
-        handle_value(nullptr);
-        return true;
-    }
-
-    bool boolean(bool val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_integer(number_integer_t val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_unsigned(number_unsigned_t val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool number_float(number_float_t val, const string_t& /*unused*/)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool string(string_t& val)
-    {
-        handle_value(val);
-        return true;
-    }
-
-    bool binary(binary_t& val)
-    {
-        handle_value(std::move(val));
-        return true;
-    }
-
-    bool start_object(std::size_t len)
-    {
-        // check callback for object start
-        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::object_start, discarded);
-        keep_stack.push_back(keep);
-
-        auto val = handle_value(BasicJsonType::value_t::object, true);
-        ref_stack.push_back(val.second);
-
-        // check object limit
-        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
-        {
-            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
-        }
-
-        return true;
-    }
-
-    bool key(string_t& val)
-    {
-        BasicJsonType k = BasicJsonType(val);
-
-        // check callback for key
-        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::key, k);
-        key_keep_stack.push_back(keep);
-
-        // add discarded value at given key and store the reference for later
-        if (keep && ref_stack.back())
-        {
-            object_element = &(ref_stack.back()->m_value.object->operator[](val) = discarded);
-        }
-
-        return true;
-    }
-
-    bool end_object()
-    {
-        if (ref_stack.back())
-        {
-            if (!callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::object_end, *ref_stack.back()))
-            {
-                // discard object
-                *ref_stack.back() = discarded;
-            }
-            else
-            {
-                ref_stack.back()->set_parents();
-            }
-        }
-
-        JSON_ASSERT(!ref_stack.empty());
-        JSON_ASSERT(!keep_stack.empty());
-        ref_stack.pop_back();
-        keep_stack.pop_back();
-
-        if (!ref_stack.empty() && ref_stack.back() && ref_stack.back()->is_structured())
-        {
-            // remove discarded value
-            for (auto it = ref_stack.back()->begin(); it != ref_stack.back()->end(); ++it)
-            {
-                if (it->is_discarded())
-                {
-                    ref_stack.back()->erase(it);
-                    break;
-                }
-            }
-        }
-
-        return true;
-    }
-
-    bool start_array(std::size_t len)
-    {
-        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::array_start, discarded);
-        keep_stack.push_back(keep);
-
-        auto val = handle_value(BasicJsonType::value_t::array, true);
-        ref_stack.push_back(val.second);
-
-        // check array limit
-        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
-        {
-            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
-        }
-
-        return true;
-    }
-
-    bool end_array()
-    {
-        bool keep = true;
-
-        if (ref_stack.back())
-        {
-            keep = callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::array_end, *ref_stack.back());
-            if (keep)
-            {
-                ref_stack.back()->set_parents();
-            }
-            else
-            {
-                // discard array
-                *ref_stack.back() = discarded;
-            }
-        }
-
-        JSON_ASSERT(!ref_stack.empty());
-        JSON_ASSERT(!keep_stack.empty());
-        ref_stack.pop_back();
-        keep_stack.pop_back();
-
-        // remove discarded value
-        if (!keep && !ref_stack.empty() && ref_stack.back()->is_array())
-        {
-            ref_stack.back()->m_value.array->pop_back();
-        }
-
-        return true;
-    }
-
-    template<class Exception>
-    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
-                     const Exception& ex)
-    {
-        errored = true;
-        static_cast<void>(ex);
-        if (allow_exceptions)
-        {
-            JSON_THROW(ex);
-        }
-        return false;
-    }
-
-    constexpr bool is_errored() const
-    {
-        return errored;
-    }
-
-  private:
-    /*!
-    @param[in] v  value to add to the JSON value we build during parsing
-    @param[in] skip_callback  whether we should skip calling the callback
-               function; this is required after start_array() and
-               start_object() SAX events, because otherwise we would call the
-               callback function with an empty array or object, respectively.
-
-    @invariant If the ref stack is empty, then the passed value will be the new
-               root.
-    @invariant If the ref stack contains a value, then it is an array or an
-               object to which we can add elements
-
-    @return pair of boolean (whether value should be kept) and pointer (to the
-            passed value in the ref_stack hierarchy; nullptr if not kept)
-    */
-    template<typename Value>
-    std::pair<bool, BasicJsonType*> handle_value(Value&& v, const bool skip_callback = false)
-    {
-        JSON_ASSERT(!keep_stack.empty());
-
-        // do not handle this value if we know it would be added to a discarded
-        // container
-        if (!keep_stack.back())
-        {
-            return {false, nullptr};
-        }
-
-        // create value
-        auto value = BasicJsonType(std::forward<Value>(v));
-
-        // check callback
-        const bool keep = skip_callback || callback(static_cast<int>(ref_stack.size()), parse_event_t::value, value);
-
-        // do not handle this value if we just learnt it shall be discarded
-        if (!keep)
-        {
-            return {false, nullptr};
-        }
-
-        if (ref_stack.empty())
-        {
-            root = std::move(value);
-            return {true, &root};
-        }
-
-        // skip this value if we already decided to skip the parent
-        // (https://github.com/nlohmann/json/issues/971#issuecomment-413678360)
-        if (!ref_stack.back())
-        {
-            return {false, nullptr};
-        }
-
-        // we now only expect arrays and objects
-        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
-
-        // array
-        if (ref_stack.back()->is_array())
-        {
-            ref_stack.back()->m_value.array->emplace_back(std::move(value));
-            return {true, &(ref_stack.back()->m_value.array->back())};
-        }
-
-        // object
-        JSON_ASSERT(ref_stack.back()->is_object());
-        // check if we should store an element for the current key
-        JSON_ASSERT(!key_keep_stack.empty());
-        const bool store_element = key_keep_stack.back();
-        key_keep_stack.pop_back();
-
-        if (!store_element)
-        {
-            return {false, nullptr};
-        }
-
-        JSON_ASSERT(object_element);
-        *object_element = std::move(value);
-        return {true, object_element};
-    }
-
-    /// the parsed JSON value
-    BasicJsonType& root;
-    /// stack to model hierarchy of values
-    std::vector<BasicJsonType*> ref_stack {};
-    /// stack to manage which values to keep
-    std::vector<bool> keep_stack {};
-    /// stack to manage which object keys to keep
-    std::vector<bool> key_keep_stack {};
-    /// helper to hold the reference for the next object element
-    BasicJsonType* object_element = nullptr;
-    /// whether a syntax error occurred
-    bool errored = false;
-    /// callback function
-    const parser_callback_t callback = nullptr;
-    /// whether to throw exceptions in case of errors
-    const bool allow_exceptions = true;
-    /// a discarded value for the callback
-    BasicJsonType discarded = BasicJsonType::value_t::discarded;
-};
-
-template<typename BasicJsonType>
-class json_sax_acceptor
-{
-  public:
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-
-    bool null()
-    {
-        return true;
-    }
-
-    bool boolean(bool /*unused*/)
-    {
-        return true;
-    }
-
-    bool number_integer(number_integer_t /*unused*/)
-    {
-        return true;
-    }
-
-    bool number_unsigned(number_unsigned_t /*unused*/)
-    {
-        return true;
-    }
-
-    bool number_float(number_float_t /*unused*/, const string_t& /*unused*/)
-    {
-        return true;
-    }
-
-    bool string(string_t& /*unused*/)
-    {
-        return true;
-    }
-
-    bool binary(binary_t& /*unused*/)
-    {
-        return true;
-    }
-
-    bool start_object(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
-    {
-        return true;
-    }
-
-    bool key(string_t& /*unused*/)
-    {
-        return true;
-    }
-
-    bool end_object()
-    {
-        return true;
-    }
-
-    bool start_array(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
-    {
-        return true;
-    }
-
-    bool end_array()
-    {
-        return true;
-    }
-
-    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/, const detail::exception& /*unused*/)
-    {
-        return false;
-    }
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/input/lexer.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <array> // array
-#include <clocale> // localeconv
-#include <cstddef> // size_t
-#include <cstdio> // snprintf
-#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
-#include <initializer_list> // initializer_list
-#include <string> // char_traits, string
-#include <utility> // move
-#include <vector> // vector
-
-// #include <nlohmann/detail/input/input_adapters.hpp>
-
-// #include <nlohmann/detail/input/position_t.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-///////////
-// lexer //
-///////////
-
-template<typename BasicJsonType>
-class lexer_base
-{
-  public:
-    /// token types for the parser
-    enum class token_type
-    {
-        uninitialized,    ///< indicating the scanner is uninitialized
-        literal_true,     ///< the `true` literal
-        literal_false,    ///< the `false` literal
-        literal_null,     ///< the `null` literal
-        value_string,     ///< a string -- use get_string() for actual value
-        value_unsigned,   ///< an unsigned integer -- use get_number_unsigned() for actual value
-        value_integer,    ///< a signed integer -- use get_number_integer() for actual value
-        value_float,      ///< an floating point number -- use get_number_float() for actual value
-        begin_array,      ///< the character for array begin `[`
-        begin_object,     ///< the character for object begin `{`
-        end_array,        ///< the character for array end `]`
-        end_object,       ///< the character for object end `}`
-        name_separator,   ///< the name separator `:`
-        value_separator,  ///< the value separator `,`
-        parse_error,      ///< indicating a parse error
-        end_of_input,     ///< indicating the end of the input buffer
-        literal_or_value  ///< a literal or the begin of a value (only for diagnostics)
-    };
-
-    /// return name of values of type token_type (only used for errors)
-    JSON_HEDLEY_RETURNS_NON_NULL
-    JSON_HEDLEY_CONST
-    static const char* token_type_name(const token_type t) noexcept
-    {
-        switch (t)
-        {
-            case token_type::uninitialized:
-                return "<uninitialized>";
-            case token_type::literal_true:
-                return "true literal";
-            case token_type::literal_false:
-                return "false literal";
-            case token_type::literal_null:
-                return "null literal";
-            case token_type::value_string:
-                return "string literal";
-            case token_type::value_unsigned:
-            case token_type::value_integer:
-            case token_type::value_float:
-                return "number literal";
-            case token_type::begin_array:
-                return "'['";
-            case token_type::begin_object:
-                return "'{'";
-            case token_type::end_array:
-                return "']'";
-            case token_type::end_object:
-                return "'}'";
-            case token_type::name_separator:
-                return "':'";
-            case token_type::value_separator:
-                return "','";
-            case token_type::parse_error:
-                return "<parse error>";
-            case token_type::end_of_input:
-                return "end of input";
-            case token_type::literal_or_value:
-                return "'[', '{', or a literal";
-            // LCOV_EXCL_START
-            default: // catch non-enum values
-                return "unknown token";
-                // LCOV_EXCL_STOP
-        }
-    }
-};
-/*!
-@brief lexical analysis
-
-This class organizes the lexical analysis during JSON deserialization.
-*/
-template<typename BasicJsonType, typename InputAdapterType>
-class lexer : public lexer_base<BasicJsonType>
-{
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using char_type = typename InputAdapterType::char_type;
-    using char_int_type = typename std::char_traits<char_type>::int_type;
-
-  public:
-    using token_type = typename lexer_base<BasicJsonType>::token_type;
-
-    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false) noexcept
-        : ia(std::move(adapter))
-        , ignore_comments(ignore_comments_)
-        , decimal_point_char(static_cast<char_int_type>(get_decimal_point()))
-    {}
-
-    // delete because of pointer members
-    lexer(const lexer&) = delete;
-    lexer(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    lexer& operator=(lexer&) = delete;
-    lexer& operator=(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    ~lexer() = default;
-
-  private:
-    /////////////////////
-    // locales
-    /////////////////////
-
-    /// return the locale-dependent decimal point
-    JSON_HEDLEY_PURE
-    static char get_decimal_point() noexcept
-    {
-        const auto* loc = localeconv();
-        JSON_ASSERT(loc != nullptr);
-        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
-    }
-
-    /////////////////////
-    // scan functions
-    /////////////////////
-
-    /*!
-    @brief get codepoint from 4 hex characters following `\u`
-
-    For input "\u c1 c2 c3 c4" the codepoint is:
-      (c1 * 0x1000) + (c2 * 0x0100) + (c3 * 0x0010) + c4
-    = (c1 << 12) + (c2 << 8) + (c3 << 4) + (c4 << 0)
-
-    Furthermore, the possible characters '0'..'9', 'A'..'F', and 'a'..'f'
-    must be converted to the integers 0x0..0x9, 0xA..0xF, 0xA..0xF, resp. The
-    conversion is done by subtracting the offset (0x30, 0x37, and 0x57)
-    between the ASCII value of the character and the desired integer value.
-
-    @return codepoint (0x0000..0xFFFF) or -1 in case of an error (e.g. EOF or
-            non-hex character)
-    */
-    int get_codepoint()
-    {
-        // this function only makes sense after reading `\u`
-        JSON_ASSERT(current == 'u');
-        int codepoint = 0;
-
-        const auto factors = { 12u, 8u, 4u, 0u };
-        for (const auto factor : factors)
-        {
-            get();
-
-            if (current >= '0' && current <= '9')
-            {
-                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x30u) << factor);
-            }
-            else if (current >= 'A' && current <= 'F')
-            {
-                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x37u) << factor);
-            }
-            else if (current >= 'a' && current <= 'f')
-            {
-                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x57u) << factor);
-            }
-            else
-            {
-                return -1;
-            }
-        }
-
-        JSON_ASSERT(0x0000 <= codepoint && codepoint <= 0xFFFF);
-        return codepoint;
-    }
-
-    /*!
-    @brief check if the next byte(s) are inside a given range
-
-    Adds the current byte and, for each passed range, reads a new byte and
-    checks if it is inside the range. If a violation was detected, set up an
-    error message and return false. Otherwise, return true.
-
-    @param[in] ranges  list of integers; interpreted as list of pairs of
-                       inclusive lower and upper bound, respectively
-
-    @pre The passed list @a ranges must have 2, 4, or 6 elements; that is,
-         1, 2, or 3 pairs. This precondition is enforced by an assertion.
-
-    @return true if and only if no range violation was detected
-    */
-    bool next_byte_in_range(std::initializer_list<char_int_type> ranges)
-    {
-        JSON_ASSERT(ranges.size() == 2 || ranges.size() == 4 || ranges.size() == 6);
-        add(current);
-
-        for (auto range = ranges.begin(); range != ranges.end(); ++range)
-        {
-            get();
-            if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range)))
-            {
-                add(current);
-            }
-            else
-            {
-                error_message = "invalid string: ill-formed UTF-8 byte";
-                return false;
-            }
-        }
-
-        return true;
-    }
-
-    /*!
-    @brief scan a string literal
-
-    This function scans a string according to Sect. 7 of RFC 8259. While
-    scanning, bytes are escaped and copied into buffer token_buffer. Then the
-    function returns successfully, token_buffer is *not* null-terminated (as it
-    may contain \0 bytes), and token_buffer.size() is the number of bytes in the
-    string.
-
-    @return token_type::value_string if string could be successfully scanned,
-            token_type::parse_error otherwise
-
-    @note In case of errors, variable error_message contains a textual
-          description.
-    */
-    token_type scan_string()
-    {
-        // reset token_buffer (ignore opening quote)
-        reset();
-
-        // we entered the function by reading an open quote
-        JSON_ASSERT(current == '\"');
-
-        while (true)
-        {
-            // get next character
-            switch (get())
-            {
-                // end of file while parsing string
-                case std::char_traits<char_type>::eof():
-                {
-                    error_message = "invalid string: missing closing quote";
-                    return token_type::parse_error;
-                }
-
-                // closing quote
-                case '\"':
-                {
-                    return token_type::value_string;
-                }
-
-                // escapes
-                case '\\':
-                {
-                    switch (get())
-                    {
-                        // quotation mark
-                        case '\"':
-                            add('\"');
-                            break;
-                        // reverse solidus
-                        case '\\':
-                            add('\\');
-                            break;
-                        // solidus
-                        case '/':
-                            add('/');
-                            break;
-                        // backspace
-                        case 'b':
-                            add('\b');
-                            break;
-                        // form feed
-                        case 'f':
-                            add('\f');
-                            break;
-                        // line feed
-                        case 'n':
-                            add('\n');
-                            break;
-                        // carriage return
-                        case 'r':
-                            add('\r');
-                            break;
-                        // tab
-                        case 't':
-                            add('\t');
-                            break;
-
-                        // unicode escapes
-                        case 'u':
-                        {
-                            const int codepoint1 = get_codepoint();
-                            int codepoint = codepoint1; // start with codepoint1
-
-                            if (JSON_HEDLEY_UNLIKELY(codepoint1 == -1))
-                            {
-                                error_message = "invalid string: '\\u' must be followed by 4 hex digits";
-                                return token_type::parse_error;
-                            }
-
-                            // check if code point is a high surrogate
-                            if (0xD800 <= codepoint1 && codepoint1 <= 0xDBFF)
-                            {
-                                // expect next \uxxxx entry
-                                if (JSON_HEDLEY_LIKELY(get() == '\\' && get() == 'u'))
-                                {
-                                    const int codepoint2 = get_codepoint();
-
-                                    if (JSON_HEDLEY_UNLIKELY(codepoint2 == -1))
-                                    {
-                                        error_message = "invalid string: '\\u' must be followed by 4 hex digits";
-                                        return token_type::parse_error;
-                                    }
-
-                                    // check if codepoint2 is a low surrogate
-                                    if (JSON_HEDLEY_LIKELY(0xDC00 <= codepoint2 && codepoint2 <= 0xDFFF))
-                                    {
-                                        // overwrite codepoint
-                                        codepoint = static_cast<int>(
-                                                        // high surrogate occupies the most significant 22 bits
-                                                        (static_cast<unsigned int>(codepoint1) << 10u)
-                                                        // low surrogate occupies the least significant 15 bits
-                                                        + static_cast<unsigned int>(codepoint2)
-                                                        // there is still the 0xD800, 0xDC00 and 0x10000 noise
-                                                        // in the result, so we have to subtract with:
-                                                        // (0xD800 << 10) + DC00 - 0x10000 = 0x35FDC00
-                                                        - 0x35FDC00u);
-                                    }
-                                    else
-                                    {
-                                        error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
-                                        return token_type::parse_error;
-                                    }
-                                }
-                                else
-                                {
-                                    error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
-                                    return token_type::parse_error;
-                                }
-                            }
-                            else
-                            {
-                                if (JSON_HEDLEY_UNLIKELY(0xDC00 <= codepoint1 && codepoint1 <= 0xDFFF))
-                                {
-                                    error_message = "invalid string: surrogate U+DC00..U+DFFF must follow U+D800..U+DBFF";
-                                    return token_type::parse_error;
-                                }
-                            }
-
-                            // result of the above calculation yields a proper codepoint
-                            JSON_ASSERT(0x00 <= codepoint && codepoint <= 0x10FFFF);
-
-                            // translate codepoint into bytes
-                            if (codepoint < 0x80)
-                            {
-                                // 1-byte characters: 0xxxxxxx (ASCII)
-                                add(static_cast<char_int_type>(codepoint));
-                            }
-                            else if (codepoint <= 0x7FF)
-                            {
-                                // 2-byte characters: 110xxxxx 10xxxxxx
-                                add(static_cast<char_int_type>(0xC0u | (static_cast<unsigned int>(codepoint) >> 6u)));
-                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
-                            }
-                            else if (codepoint <= 0xFFFF)
-                            {
-                                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
-                                add(static_cast<char_int_type>(0xE0u | (static_cast<unsigned int>(codepoint) >> 12u)));
-                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
-                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
-                            }
-                            else
-                            {
-                                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
-                                add(static_cast<char_int_type>(0xF0u | (static_cast<unsigned int>(codepoint) >> 18u)));
-                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 12u) & 0x3Fu)));
-                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
-                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
-                            }
-
-                            break;
-                        }
-
-                        // other characters after escape
-                        default:
-                            error_message = "invalid string: forbidden character after backslash";
-                            return token_type::parse_error;
-                    }
-
-                    break;
-                }
-
-                // invalid control characters
-                case 0x00:
-                {
-                    error_message = "invalid string: control character U+0000 (NUL) must be escaped to \\u0000";
-                    return token_type::parse_error;
-                }
-
-                case 0x01:
-                {
-                    error_message = "invalid string: control character U+0001 (SOH) must be escaped to \\u0001";
-                    return token_type::parse_error;
-                }
-
-                case 0x02:
-                {
-                    error_message = "invalid string: control character U+0002 (STX) must be escaped to \\u0002";
-                    return token_type::parse_error;
-                }
-
-                case 0x03:
-                {
-                    error_message = "invalid string: control character U+0003 (ETX) must be escaped to \\u0003";
-                    return token_type::parse_error;
-                }
-
-                case 0x04:
-                {
-                    error_message = "invalid string: control character U+0004 (EOT) must be escaped to \\u0004";
-                    return token_type::parse_error;
-                }
-
-                case 0x05:
-                {
-                    error_message = "invalid string: control character U+0005 (ENQ) must be escaped to \\u0005";
-                    return token_type::parse_error;
-                }
-
-                case 0x06:
-                {
-                    error_message = "invalid string: control character U+0006 (ACK) must be escaped to \\u0006";
-                    return token_type::parse_error;
-                }
-
-                case 0x07:
-                {
-                    error_message = "invalid string: control character U+0007 (BEL) must be escaped to \\u0007";
-                    return token_type::parse_error;
-                }
-
-                case 0x08:
-                {
-                    error_message = "invalid string: control character U+0008 (BS) must be escaped to \\u0008 or \\b";
-                    return token_type::parse_error;
-                }
-
-                case 0x09:
-                {
-                    error_message = "invalid string: control character U+0009 (HT) must be escaped to \\u0009 or \\t";
-                    return token_type::parse_error;
-                }
-
-                case 0x0A:
-                {
-                    error_message = "invalid string: control character U+000A (LF) must be escaped to \\u000A or \\n";
-                    return token_type::parse_error;
-                }
-
-                case 0x0B:
-                {
-                    error_message = "invalid string: control character U+000B (VT) must be escaped to \\u000B";
-                    return token_type::parse_error;
-                }
-
-                case 0x0C:
-                {
-                    error_message = "invalid string: control character U+000C (FF) must be escaped to \\u000C or \\f";
-                    return token_type::parse_error;
-                }
-
-                case 0x0D:
-                {
-                    error_message = "invalid string: control character U+000D (CR) must be escaped to \\u000D or \\r";
-                    return token_type::parse_error;
-                }
-
-                case 0x0E:
-                {
-                    error_message = "invalid string: control character U+000E (SO) must be escaped to \\u000E";
-                    return token_type::parse_error;
-                }
-
-                case 0x0F:
-                {
-                    error_message = "invalid string: control character U+000F (SI) must be escaped to \\u000F";
-                    return token_type::parse_error;
-                }
-
-                case 0x10:
-                {
-                    error_message = "invalid string: control character U+0010 (DLE) must be escaped to \\u0010";
-                    return token_type::parse_error;
-                }
-
-                case 0x11:
-                {
-                    error_message = "invalid string: control character U+0011 (DC1) must be escaped to \\u0011";
-                    return token_type::parse_error;
-                }
-
-                case 0x12:
-                {
-                    error_message = "invalid string: control character U+0012 (DC2) must be escaped to \\u0012";
-                    return token_type::parse_error;
-                }
-
-                case 0x13:
-                {
-                    error_message = "invalid string: control character U+0013 (DC3) must be escaped to \\u0013";
-                    return token_type::parse_error;
-                }
-
-                case 0x14:
-                {
-                    error_message = "invalid string: control character U+0014 (DC4) must be escaped to \\u0014";
-                    return token_type::parse_error;
-                }
-
-                case 0x15:
-                {
-                    error_message = "invalid string: control character U+0015 (NAK) must be escaped to \\u0015";
-                    return token_type::parse_error;
-                }
-
-                case 0x16:
-                {
-                    error_message = "invalid string: control character U+0016 (SYN) must be escaped to \\u0016";
-                    return token_type::parse_error;
-                }
-
-                case 0x17:
-                {
-                    error_message = "invalid string: control character U+0017 (ETB) must be escaped to \\u0017";
-                    return token_type::parse_error;
-                }
-
-                case 0x18:
-                {
-                    error_message = "invalid string: control character U+0018 (CAN) must be escaped to \\u0018";
-                    return token_type::parse_error;
-                }
-
-                case 0x19:
-                {
-                    error_message = "invalid string: control character U+0019 (EM) must be escaped to \\u0019";
-                    return token_type::parse_error;
-                }
-
-                case 0x1A:
-                {
-                    error_message = "invalid string: control character U+001A (SUB) must be escaped to \\u001A";
-                    return token_type::parse_error;
-                }
-
-                case 0x1B:
-                {
-                    error_message = "invalid string: control character U+001B (ESC) must be escaped to \\u001B";
-                    return token_type::parse_error;
-                }
-
-                case 0x1C:
-                {
-                    error_message = "invalid string: control character U+001C (FS) must be escaped to \\u001C";
-                    return token_type::parse_error;
-                }
-
-                case 0x1D:
-                {
-                    error_message = "invalid string: control character U+001D (GS) must be escaped to \\u001D";
-                    return token_type::parse_error;
-                }
-
-                case 0x1E:
-                {
-                    error_message = "invalid string: control character U+001E (RS) must be escaped to \\u001E";
-                    return token_type::parse_error;
-                }
-
-                case 0x1F:
-                {
-                    error_message = "invalid string: control character U+001F (US) must be escaped to \\u001F";
-                    return token_type::parse_error;
-                }
-
-                // U+0020..U+007F (except U+0022 (quote) and U+005C (backspace))
-                case 0x20:
-                case 0x21:
-                case 0x23:
-                case 0x24:
-                case 0x25:
-                case 0x26:
-                case 0x27:
-                case 0x28:
-                case 0x29:
-                case 0x2A:
-                case 0x2B:
-                case 0x2C:
-                case 0x2D:
-                case 0x2E:
-                case 0x2F:
-                case 0x30:
-                case 0x31:
-                case 0x32:
-                case 0x33:
-                case 0x34:
-                case 0x35:
-                case 0x36:
-                case 0x37:
-                case 0x38:
-                case 0x39:
-                case 0x3A:
-                case 0x3B:
-                case 0x3C:
-                case 0x3D:
-                case 0x3E:
-                case 0x3F:
-                case 0x40:
-                case 0x41:
-                case 0x42:
-                case 0x43:
-                case 0x44:
-                case 0x45:
-                case 0x46:
-                case 0x47:
-                case 0x48:
-                case 0x49:
-                case 0x4A:
-                case 0x4B:
-                case 0x4C:
-                case 0x4D:
-                case 0x4E:
-                case 0x4F:
-                case 0x50:
-                case 0x51:
-                case 0x52:
-                case 0x53:
-                case 0x54:
-                case 0x55:
-                case 0x56:
-                case 0x57:
-                case 0x58:
-                case 0x59:
-                case 0x5A:
-                case 0x5B:
-                case 0x5D:
-                case 0x5E:
-                case 0x5F:
-                case 0x60:
-                case 0x61:
-                case 0x62:
-                case 0x63:
-                case 0x64:
-                case 0x65:
-                case 0x66:
-                case 0x67:
-                case 0x68:
-                case 0x69:
-                case 0x6A:
-                case 0x6B:
-                case 0x6C:
-                case 0x6D:
-                case 0x6E:
-                case 0x6F:
-                case 0x70:
-                case 0x71:
-                case 0x72:
-                case 0x73:
-                case 0x74:
-                case 0x75:
-                case 0x76:
-                case 0x77:
-                case 0x78:
-                case 0x79:
-                case 0x7A:
-                case 0x7B:
-                case 0x7C:
-                case 0x7D:
-                case 0x7E:
-                case 0x7F:
-                {
-                    add(current);
-                    break;
-                }
-
-                // U+0080..U+07FF: bytes C2..DF 80..BF
-                case 0xC2:
-                case 0xC3:
-                case 0xC4:
-                case 0xC5:
-                case 0xC6:
-                case 0xC7:
-                case 0xC8:
-                case 0xC9:
-                case 0xCA:
-                case 0xCB:
-                case 0xCC:
-                case 0xCD:
-                case 0xCE:
-                case 0xCF:
-                case 0xD0:
-                case 0xD1:
-                case 0xD2:
-                case 0xD3:
-                case 0xD4:
-                case 0xD5:
-                case 0xD6:
-                case 0xD7:
-                case 0xD8:
-                case 0xD9:
-                case 0xDA:
-                case 0xDB:
-                case 0xDC:
-                case 0xDD:
-                case 0xDE:
-                case 0xDF:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!next_byte_in_range({0x80, 0xBF})))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+0800..U+0FFF: bytes E0 A0..BF 80..BF
-                case 0xE0:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0xA0, 0xBF, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+1000..U+CFFF: bytes E1..EC 80..BF 80..BF
-                // U+E000..U+FFFF: bytes EE..EF 80..BF 80..BF
-                case 0xE1:
-                case 0xE2:
-                case 0xE3:
-                case 0xE4:
-                case 0xE5:
-                case 0xE6:
-                case 0xE7:
-                case 0xE8:
-                case 0xE9:
-                case 0xEA:
-                case 0xEB:
-                case 0xEC:
-                case 0xEE:
-                case 0xEF:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+D000..U+D7FF: bytes ED 80..9F 80..BF
-                case 0xED:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x9F, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+10000..U+3FFFF F0 90..BF 80..BF 80..BF
-                case 0xF0:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x90, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+40000..U+FFFFF F1..F3 80..BF 80..BF 80..BF
-                case 0xF1:
-                case 0xF2:
-                case 0xF3:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // U+100000..U+10FFFF F4 80..8F 80..BF 80..BF
-                case 0xF4:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x8F, 0x80, 0xBF, 0x80, 0xBF}))))
-                    {
-                        return token_type::parse_error;
-                    }
-                    break;
-                }
-
-                // remaining bytes (80..C1 and F5..FF) are ill-formed
-                default:
-                {
-                    error_message = "invalid string: ill-formed UTF-8 byte";
-                    return token_type::parse_error;
-                }
-            }
-        }
-    }
-
-    /*!
-     * @brief scan a comment
-     * @return whether comment could be scanned successfully
-     */
-    bool scan_comment()
-    {
-        switch (get())
-        {
-            // single-line comments skip input until a newline or EOF is read
-            case '/':
-            {
-                while (true)
-                {
-                    switch (get())
-                    {
-                        case '\n':
-                        case '\r':
-                        case std::char_traits<char_type>::eof():
-                        case '\0':
-                            return true;
-
-                        default:
-                            break;
-                    }
-                }
-            }
-
-            // multi-line comments skip input until */ is read
-            case '*':
-            {
-                while (true)
-                {
-                    switch (get())
-                    {
-                        case std::char_traits<char_type>::eof():
-                        case '\0':
-                        {
-                            error_message = "invalid comment; missing closing '*/'";
-                            return false;
-                        }
-
-                        case '*':
-                        {
-                            switch (get())
-                            {
-                                case '/':
-                                    return true;
-
-                                default:
-                                {
-                                    unget();
-                                    continue;
-                                }
-                            }
-                        }
-
-                        default:
-                            continue;
-                    }
-                }
-            }
-
-            // unexpected character after reading '/'
-            default:
-            {
-                error_message = "invalid comment; expecting '/' or '*' after '/'";
-                return false;
-            }
-        }
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(float& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtof(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtod(str, endptr);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    static void strtof(long double& f, const char* str, char** endptr) noexcept
-    {
-        f = std::strtold(str, endptr);
-    }
-
-    /*!
-    @brief scan a number literal
-
-    This function scans a string according to Sect. 6 of RFC 8259.
-
-    The function is realized with a deterministic finite state machine derived
-    from the grammar described in RFC 8259. Starting in state "init", the
-    input is read and used to determined the next state. Only state "done"
-    accepts the number. State "error" is a trap state to model errors. In the
-    table below, "anything" means any character but the ones listed before.
-
-    state    | 0        | 1-9      | e E      | +       | -       | .        | anything
-    ---------|----------|----------|----------|---------|---------|----------|-----------
-    init     | zero     | any1     | [error]  | [error] | minus   | [error]  | [error]
-    minus    | zero     | any1     | [error]  | [error] | [error] | [error]  | [error]
-    zero     | done     | done     | exponent | done    | done    | decimal1 | done
-    any1     | any1     | any1     | exponent | done    | done    | decimal1 | done
-    decimal1 | decimal2 | decimal2 | [error]  | [error] | [error] | [error]  | [error]
-    decimal2 | decimal2 | decimal2 | exponent | done    | done    | done     | done
-    exponent | any2     | any2     | [error]  | sign    | sign    | [error]  | [error]
-    sign     | any2     | any2     | [error]  | [error] | [error] | [error]  | [error]
-    any2     | any2     | any2     | done     | done    | done    | done     | done
-
-    The state machine is realized with one label per state (prefixed with
-    "scan_number_") and `goto` statements between them. The state machine
-    contains cycles, but any cycle can be left when EOF is read. Therefore,
-    the function is guaranteed to terminate.
-
-    During scanning, the read bytes are stored in token_buffer. This string is
-    then converted to a signed integer, an unsigned integer, or a
-    floating-point number.
-
-    @return token_type::value_unsigned, token_type::value_integer, or
-            token_type::value_float if number could be successfully scanned,
-            token_type::parse_error otherwise
-
-    @note The scanner is independent of the current locale. Internally, the
-          locale's decimal point is used instead of `.` to work with the
-          locale-dependent converters.
-    */
-    token_type scan_number()  // lgtm [cpp/use-of-goto]
-    {
-        // reset token_buffer to store the number's bytes
-        reset();
-
-        // the type of the parsed number; initially set to unsigned; will be
-        // changed if minus sign, decimal point or exponent is read
-        token_type number_type = token_type::value_unsigned;
-
-        // state (init): we just found out we need to scan a number
-        switch (current)
-        {
-            case '-':
-            {
-                add(current);
-                goto scan_number_minus;
-            }
-
-            case '0':
-            {
-                add(current);
-                goto scan_number_zero;
-            }
-
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any1;
-            }
-
-            // all other characters are rejected outside scan_number()
-            default:            // LCOV_EXCL_LINE
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        }
-
-scan_number_minus:
-        // state: we just parsed a leading minus sign
-        number_type = token_type::value_integer;
-        switch (get())
-        {
-            case '0':
-            {
-                add(current);
-                goto scan_number_zero;
-            }
-
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any1;
-            }
-
-            default:
-            {
-                error_message = "invalid number; expected digit after '-'";
-                return token_type::parse_error;
-            }
-        }
-
-scan_number_zero:
-        // state: we just parse a zero (maybe with a leading minus sign)
-        switch (get())
-        {
-            case '.':
-            {
-                add(decimal_point_char);
-                goto scan_number_decimal1;
-            }
-
-            case 'e':
-            case 'E':
-            {
-                add(current);
-                goto scan_number_exponent;
-            }
-
-            default:
-                goto scan_number_done;
-        }
-
-scan_number_any1:
-        // state: we just parsed a number 0-9 (maybe with a leading minus sign)
-        switch (get())
-        {
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any1;
-            }
-
-            case '.':
-            {
-                add(decimal_point_char);
-                goto scan_number_decimal1;
-            }
-
-            case 'e':
-            case 'E':
-            {
-                add(current);
-                goto scan_number_exponent;
-            }
-
-            default:
-                goto scan_number_done;
-        }
-
-scan_number_decimal1:
-        // state: we just parsed a decimal point
-        number_type = token_type::value_float;
-        switch (get())
-        {
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_decimal2;
-            }
-
-            default:
-            {
-                error_message = "invalid number; expected digit after '.'";
-                return token_type::parse_error;
-            }
-        }
-
-scan_number_decimal2:
-        // we just parsed at least one number after a decimal point
-        switch (get())
-        {
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_decimal2;
-            }
-
-            case 'e':
-            case 'E':
-            {
-                add(current);
-                goto scan_number_exponent;
-            }
-
-            default:
-                goto scan_number_done;
-        }
-
-scan_number_exponent:
-        // we just parsed an exponent
-        number_type = token_type::value_float;
-        switch (get())
-        {
-            case '+':
-            case '-':
-            {
-                add(current);
-                goto scan_number_sign;
-            }
-
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any2;
-            }
-
-            default:
-            {
-                error_message =
-                    "invalid number; expected '+', '-', or digit after exponent";
-                return token_type::parse_error;
-            }
-        }
-
-scan_number_sign:
-        // we just parsed an exponent sign
-        switch (get())
-        {
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any2;
-            }
-
-            default:
-            {
-                error_message = "invalid number; expected digit after exponent sign";
-                return token_type::parse_error;
-            }
-        }
-
-scan_number_any2:
-        // we just parsed a number after the exponent or exponent sign
-        switch (get())
-        {
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-            {
-                add(current);
-                goto scan_number_any2;
-            }
-
-            default:
-                goto scan_number_done;
-        }
-
-scan_number_done:
-        // unget the character after the number (we only read it to know that
-        // we are done scanning a number)
-        unget();
-
-        char* endptr = nullptr; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-        errno = 0;
-
-        // try to parse integers first and fall back to floats
-        if (number_type == token_type::value_unsigned)
-        {
-            const auto x = std::strtoull(token_buffer.data(), &endptr, 10);
-
-            // we checked the number format before
-            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
-
-            if (errno == 0)
-            {
-                value_unsigned = static_cast<number_unsigned_t>(x);
-                if (value_unsigned == x)
-                {
-                    return token_type::value_unsigned;
-                }
-            }
-        }
-        else if (number_type == token_type::value_integer)
-        {
-            const auto x = std::strtoll(token_buffer.data(), &endptr, 10);
-
-            // we checked the number format before
-            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
-
-            if (errno == 0)
-            {
-                value_integer = static_cast<number_integer_t>(x);
-                if (value_integer == x)
-                {
-                    return token_type::value_integer;
-                }
-            }
-        }
-
-        // this code is reached if we parse a floating-point number or if an
-        // integer conversion above failed
-        strtof(value_float, token_buffer.data(), &endptr);
-
-        // we checked the number format before
-        JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
-
-        return token_type::value_float;
-    }
-
-    /*!
-    @param[in] literal_text  the literal text to expect
-    @param[in] length        the length of the passed literal text
-    @param[in] return_type   the token type to return on success
-    */
-    JSON_HEDLEY_NON_NULL(2)
-    token_type scan_literal(const char_type* literal_text, const std::size_t length,
-                            token_type return_type)
-    {
-        JSON_ASSERT(std::char_traits<char_type>::to_char_type(current) == literal_text[0]);
-        for (std::size_t i = 1; i < length; ++i)
-        {
-            if (JSON_HEDLEY_UNLIKELY(std::char_traits<char_type>::to_char_type(get()) != literal_text[i]))
-            {
-                error_message = "invalid literal";
-                return token_type::parse_error;
-            }
-        }
-        return return_type;
-    }
-
-    /////////////////////
-    // input management
-    /////////////////////
-
-    /// reset token_buffer; current character is beginning of token
-    void reset() noexcept
-    {
-        token_buffer.clear();
-        token_string.clear();
-        token_string.push_back(std::char_traits<char_type>::to_char_type(current));
-    }
-
-    /*
-    @brief get next character from the input
-
-    This function provides the interface to the used input adapter. It does
-    not throw in case the input reached EOF, but returns a
-    `std::char_traits<char>::eof()` in that case.  Stores the scanned characters
-    for use in error messages.
-
-    @return character read from the input
-    */
-    char_int_type get()
-    {
-        ++position.chars_read_total;
-        ++position.chars_read_current_line;
-
-        if (next_unget)
-        {
-            // just reset the next_unget variable and work with current
-            next_unget = false;
-        }
-        else
-        {
-            current = ia.get_character();
-        }
-
-        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
-        {
-            token_string.push_back(std::char_traits<char_type>::to_char_type(current));
-        }
-
-        if (current == '\n')
-        {
-            ++position.lines_read;
-            position.chars_read_current_line = 0;
-        }
-
-        return current;
-    }
-
-    /*!
-    @brief unget current character (read it again on next get)
-
-    We implement unget by setting variable next_unget to true. The input is not
-    changed - we just simulate ungetting by modifying chars_read_total,
-    chars_read_current_line, and token_string. The next call to get() will
-    behave as if the unget character is read again.
-    */
-    void unget()
-    {
-        next_unget = true;
-
-        --position.chars_read_total;
-
-        // in case we "unget" a newline, we have to also decrement the lines_read
-        if (position.chars_read_current_line == 0)
-        {
-            if (position.lines_read > 0)
-            {
-                --position.lines_read;
-            }
-        }
-        else
-        {
-            --position.chars_read_current_line;
-        }
-
-        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
-        {
-            JSON_ASSERT(!token_string.empty());
-            token_string.pop_back();
-        }
-    }
-
-    /// add a character to token_buffer
-    void add(char_int_type c)
-    {
-        token_buffer.push_back(static_cast<typename string_t::value_type>(c));
-    }
-
-  public:
-    /////////////////////
-    // value getters
-    /////////////////////
-
-    /// return integer value
-    constexpr number_integer_t get_number_integer() const noexcept
-    {
-        return value_integer;
-    }
-
-    /// return unsigned integer value
-    constexpr number_unsigned_t get_number_unsigned() const noexcept
-    {
-        return value_unsigned;
-    }
-
-    /// return floating-point value
-    constexpr number_float_t get_number_float() const noexcept
-    {
-        return value_float;
-    }
-
-    /// return current string value (implicitly resets the token; useful only once)
-    string_t& get_string()
-    {
-        return token_buffer;
-    }
-
-    /////////////////////
-    // diagnostics
-    /////////////////////
-
-    /// return position of last read token
-    constexpr position_t get_position() const noexcept
-    {
-        return position;
-    }
-
-    /// return the last read token (for errors only).  Will never contain EOF
-    /// (an arbitrary value that is not a valid char value, often -1), because
-    /// 255 may legitimately occur.  May contain NUL, which should be escaped.
-    std::string get_token_string() const
-    {
-        // escape control characters
-        std::string result;
-        for (const auto c : token_string)
-        {
-            if (static_cast<unsigned char>(c) <= '\x1F')
-            {
-                // escape control characters
-                std::array<char, 9> cs{{}};
-                static_cast<void>((std::snprintf)(cs.data(), cs.size(), "<U+%.4X>", static_cast<unsigned char>(c))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-                result += cs.data();
-            }
-            else
-            {
-                // add character as is
-                result.push_back(static_cast<std::string::value_type>(c));
-            }
-        }
-
-        return result;
-    }
-
-    /// return syntax error message
-    JSON_HEDLEY_RETURNS_NON_NULL
-    constexpr const char* get_error_message() const noexcept
-    {
-        return error_message;
-    }
-
-    /////////////////////
-    // actual scanner
-    /////////////////////
-
-    /*!
-    @brief skip the UTF-8 byte order mark
-    @return true iff there is no BOM or the correct BOM has been skipped
-    */
-    bool skip_bom()
-    {
-        if (get() == 0xEF)
-        {
-            // check if we completely parse the BOM
-            return get() == 0xBB && get() == 0xBF;
-        }
-
-        // the first character is not the beginning of the BOM; unget it to
-        // process is later
-        unget();
-        return true;
-    }
-
-    void skip_whitespace()
-    {
-        do
-        {
-            get();
-        }
-        while (current == ' ' || current == '\t' || current == '\n' || current == '\r');
-    }
-
-    token_type scan()
-    {
-        // initially, skip the BOM
-        if (position.chars_read_total == 0 && !skip_bom())
-        {
-            error_message = "invalid BOM; must be 0xEF 0xBB 0xBF if given";
-            return token_type::parse_error;
-        }
-
-        // read next character and ignore whitespace
-        skip_whitespace();
-
-        // ignore comments
-        while (ignore_comments && current == '/')
-        {
-            if (!scan_comment())
-            {
-                return token_type::parse_error;
-            }
-
-            // skip following whitespace
-            skip_whitespace();
-        }
-
-        switch (current)
-        {
-            // structural characters
-            case '[':
-                return token_type::begin_array;
-            case ']':
-                return token_type::end_array;
-            case '{':
-                return token_type::begin_object;
-            case '}':
-                return token_type::end_object;
-            case ':':
-                return token_type::name_separator;
-            case ',':
-                return token_type::value_separator;
-
-            // literals
-            case 't':
-            {
-                std::array<char_type, 4> true_literal = {{static_cast<char_type>('t'), static_cast<char_type>('r'), static_cast<char_type>('u'), static_cast<char_type>('e')}};
-                return scan_literal(true_literal.data(), true_literal.size(), token_type::literal_true);
-            }
-            case 'f':
-            {
-                std::array<char_type, 5> false_literal = {{static_cast<char_type>('f'), static_cast<char_type>('a'), static_cast<char_type>('l'), static_cast<char_type>('s'), static_cast<char_type>('e')}};
-                return scan_literal(false_literal.data(), false_literal.size(), token_type::literal_false);
-            }
-            case 'n':
-            {
-                std::array<char_type, 4> null_literal = {{static_cast<char_type>('n'), static_cast<char_type>('u'), static_cast<char_type>('l'), static_cast<char_type>('l')}};
-                return scan_literal(null_literal.data(), null_literal.size(), token_type::literal_null);
-            }
-
-            // string
-            case '\"':
-                return scan_string();
-
-            // number
-            case '-':
-            case '0':
-            case '1':
-            case '2':
-            case '3':
-            case '4':
-            case '5':
-            case '6':
-            case '7':
-            case '8':
-            case '9':
-                return scan_number();
-
-            // end of input (the null byte is needed when parsing from
-            // string literals)
-            case '\0':
-            case std::char_traits<char_type>::eof():
-                return token_type::end_of_input;
-
-            // error
-            default:
-                error_message = "invalid literal";
-                return token_type::parse_error;
-        }
-    }
-
-  private:
-    /// input adapter
-    InputAdapterType ia;
-
-    /// whether comments should be ignored (true) or signaled as errors (false)
-    const bool ignore_comments = false;
-
-    /// the current character
-    char_int_type current = std::char_traits<char_type>::eof();
-
-    /// whether the next get() call should just return current
-    bool next_unget = false;
-
-    /// the start position of the current token
-    position_t position {};
-
-    /// raw input token string (for error messages)
-    std::vector<char_type> token_string {};
-
-    /// buffer for variable-length tokens (numbers, strings)
-    string_t token_buffer {};
-
-    /// a description of occurred lexer errors
-    const char* error_message = "";
-
-    // number values
-    number_integer_t value_integer = 0;
-    number_unsigned_t value_unsigned = 0;
-    number_float_t value_float = 0;
-
-    /// the decimal point
-    const char_int_type decimal_point_char = '.';
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/is_sax.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstdint> // size_t
-#include <utility> // declval
-#include <string> // string
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/meta/detected.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename T>
-using null_function_t = decltype(std::declval<T&>().null());
-
-template<typename T>
-using boolean_function_t =
-    decltype(std::declval<T&>().boolean(std::declval<bool>()));
-
-template<typename T, typename Integer>
-using number_integer_function_t =
-    decltype(std::declval<T&>().number_integer(std::declval<Integer>()));
-
-template<typename T, typename Unsigned>
-using number_unsigned_function_t =
-    decltype(std::declval<T&>().number_unsigned(std::declval<Unsigned>()));
-
-template<typename T, typename Float, typename String>
-using number_float_function_t = decltype(std::declval<T&>().number_float(
-                                    std::declval<Float>(), std::declval<const String&>()));
-
-template<typename T, typename String>
-using string_function_t =
-    decltype(std::declval<T&>().string(std::declval<String&>()));
-
-template<typename T, typename Binary>
-using binary_function_t =
-    decltype(std::declval<T&>().binary(std::declval<Binary&>()));
-
-template<typename T>
-using start_object_function_t =
-    decltype(std::declval<T&>().start_object(std::declval<std::size_t>()));
-
-template<typename T, typename String>
-using key_function_t =
-    decltype(std::declval<T&>().key(std::declval<String&>()));
-
-template<typename T>
-using end_object_function_t = decltype(std::declval<T&>().end_object());
-
-template<typename T>
-using start_array_function_t =
-    decltype(std::declval<T&>().start_array(std::declval<std::size_t>()));
-
-template<typename T>
-using end_array_function_t = decltype(std::declval<T&>().end_array());
-
-template<typename T, typename Exception>
-using parse_error_function_t = decltype(std::declval<T&>().parse_error(
-        std::declval<std::size_t>(), std::declval<const std::string&>(),
-        std::declval<const Exception&>()));
-
-template<typename SAX, typename BasicJsonType>
-struct is_sax
-{
-  private:
-    static_assert(is_basic_json<BasicJsonType>::value,
-                  "BasicJsonType must be of type basic_json<...>");
-
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-    using exception_t = typename BasicJsonType::exception;
-
-  public:
-    static constexpr bool value =
-        is_detected_exact<bool, null_function_t, SAX>::value &&
-        is_detected_exact<bool, boolean_function_t, SAX>::value &&
-        is_detected_exact<bool, number_integer_function_t, SAX, number_integer_t>::value &&
-        is_detected_exact<bool, number_unsigned_function_t, SAX, number_unsigned_t>::value &&
-        is_detected_exact<bool, number_float_function_t, SAX, number_float_t, string_t>::value &&
-        is_detected_exact<bool, string_function_t, SAX, string_t>::value &&
-        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value &&
-        is_detected_exact<bool, start_object_function_t, SAX>::value &&
-        is_detected_exact<bool, key_function_t, SAX, string_t>::value &&
-        is_detected_exact<bool, end_object_function_t, SAX>::value &&
-        is_detected_exact<bool, start_array_function_t, SAX>::value &&
-        is_detected_exact<bool, end_array_function_t, SAX>::value &&
-        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value;
-};
-
-template<typename SAX, typename BasicJsonType>
-struct is_sax_static_asserts
-{
-  private:
-    static_assert(is_basic_json<BasicJsonType>::value,
-                  "BasicJsonType must be of type basic_json<...>");
-
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-    using exception_t = typename BasicJsonType::exception;
-
-  public:
-    static_assert(is_detected_exact<bool, null_function_t, SAX>::value,
-                  "Missing/invalid function: bool null()");
-    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
-                  "Missing/invalid function: bool boolean(bool)");
-    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
-                  "Missing/invalid function: bool boolean(bool)");
-    static_assert(
-        is_detected_exact<bool, number_integer_function_t, SAX,
-        number_integer_t>::value,
-        "Missing/invalid function: bool number_integer(number_integer_t)");
-    static_assert(
-        is_detected_exact<bool, number_unsigned_function_t, SAX,
-        number_unsigned_t>::value,
-        "Missing/invalid function: bool number_unsigned(number_unsigned_t)");
-    static_assert(is_detected_exact<bool, number_float_function_t, SAX,
-                  number_float_t, string_t>::value,
-                  "Missing/invalid function: bool number_float(number_float_t, const string_t&)");
-    static_assert(
-        is_detected_exact<bool, string_function_t, SAX, string_t>::value,
-        "Missing/invalid function: bool string(string_t&)");
-    static_assert(
-        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value,
-        "Missing/invalid function: bool binary(binary_t&)");
-    static_assert(is_detected_exact<bool, start_object_function_t, SAX>::value,
-                  "Missing/invalid function: bool start_object(std::size_t)");
-    static_assert(is_detected_exact<bool, key_function_t, SAX, string_t>::value,
-                  "Missing/invalid function: bool key(string_t&)");
-    static_assert(is_detected_exact<bool, end_object_function_t, SAX>::value,
-                  "Missing/invalid function: bool end_object()");
-    static_assert(is_detected_exact<bool, start_array_function_t, SAX>::value,
-                  "Missing/invalid function: bool start_array(std::size_t)");
-    static_assert(is_detected_exact<bool, end_array_function_t, SAX>::value,
-                  "Missing/invalid function: bool end_array()");
-    static_assert(
-        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value,
-        "Missing/invalid function: bool parse_error(std::size_t, const "
-        "std::string&, const exception&)");
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/// how to treat CBOR tags
-enum class cbor_tag_handler_t
-{
-    error,   ///< throw a parse_error exception in case of a tag
-    ignore,  ///< ignore tags
-    store    ///< store tags as binary type
-};
-
-/*!
-@brief determine system byte order
-
-@return true if and only if system's byte order is little endian
-
-@note from https://stackoverflow.com/a/1001328/266378
-*/
-static inline bool little_endianness(int num = 1) noexcept
-{
-    return *reinterpret_cast<char*>(&num) == 1;
-}
-
-
-///////////////////
-// binary reader //
-///////////////////
-
-/*!
-@brief deserialization of CBOR, MessagePack, and UBJSON values
-*/
-template<typename BasicJsonType, typename InputAdapterType, typename SAX = json_sax_dom_parser<BasicJsonType>>
-class binary_reader
-{
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-    using json_sax_t = SAX;
-    using char_type = typename InputAdapterType::char_type;
-    using char_int_type = typename std::char_traits<char_type>::int_type;
-
-  public:
-    /*!
-    @brief create a binary reader
-
-    @param[in] adapter  input adapter to read from
-    */
-    explicit binary_reader(InputAdapterType&& adapter, const input_format_t format = input_format_t::json) noexcept : ia(std::move(adapter)), input_format(format)
-    {
-        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
-    }
-
-    // make class move-only
-    binary_reader(const binary_reader&) = delete;
-    binary_reader(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    binary_reader& operator=(const binary_reader&) = delete;
-    binary_reader& operator=(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
-    ~binary_reader() = default;
-
-    /*!
-    @param[in] format  the binary format to parse
-    @param[in] sax_    a SAX event processor
-    @param[in] strict  whether to expect the input to be consumed completed
-    @param[in] tag_handler  how to treat CBOR tags
-
-    @return whether parsing was successful
-    */
-    JSON_HEDLEY_NON_NULL(3)
-    bool sax_parse(const input_format_t format,
-                   json_sax_t* sax_,
-                   const bool strict = true,
-                   const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
-    {
-        sax = sax_;
-        bool result = false;
-
-        switch (format)
-        {
-            case input_format_t::bson:
-                result = parse_bson_internal();
-                break;
-
-            case input_format_t::cbor:
-                result = parse_cbor_internal(true, tag_handler);
-                break;
-
-            case input_format_t::msgpack:
-                result = parse_msgpack_internal();
-                break;
-
-            case input_format_t::ubjson:
-            case input_format_t::bjdata:
-                result = parse_ubjson_internal();
-                break;
-
-            case input_format_t::json: // LCOV_EXCL_LINE
-            default:            // LCOV_EXCL_LINE
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        }
-
-        // strict mode: next byte must be EOF
-        if (result && strict)
-        {
-            if (input_format == input_format_t::ubjson || input_format == input_format_t::bjdata)
-            {
-                get_ignore_noop();
-            }
-            else
-            {
-                get();
-            }
-
-            if (JSON_HEDLEY_UNLIKELY(current != std::char_traits<char_type>::eof()))
-            {
-                return sax->parse_error(chars_read, get_token_string(), parse_error::create(110, chars_read,
-                                        exception_message(input_format, concat("expected end of input; last byte: 0x", get_token_string()), "value"), nullptr));
-            }
-        }
-
-        return result;
-    }
-
-  private:
-    //////////
-    // BSON //
-    //////////
-
-    /*!
-    @brief Reads in a BSON-object and passes it to the SAX-parser.
-    @return whether a valid BSON-value was passed to the SAX parser
-    */
-    bool parse_bson_internal()
-    {
-        std::int32_t document_size{};
-        get_number<std::int32_t, true>(input_format_t::bson, document_size);
-
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
-        {
-            return false;
-        }
-
-        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/false)))
-        {
-            return false;
-        }
-
-        return sax->end_object();
-    }
-
-    /*!
-    @brief Parses a C-style string from the BSON input.
-    @param[in,out] result  A reference to the string variable where the read
-                            string is to be stored.
-    @return `true` if the \x00-byte indicating the end of the string was
-             encountered before the EOF; false` indicates an unexpected EOF.
-    */
-    bool get_bson_cstr(string_t& result)
-    {
-        auto out = std::back_inserter(result);
-        while (true)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "cstring")))
-            {
-                return false;
-            }
-            if (current == 0x00)
-            {
-                return true;
-            }
-            *out++ = static_cast<typename string_t::value_type>(current);
-        }
-    }
-
-    /*!
-    @brief Parses a zero-terminated string of length @a len from the BSON
-           input.
-    @param[in] len  The length (including the zero-byte at the end) of the
-                    string to be read.
-    @param[in,out] result  A reference to the string variable where the read
-                            string is to be stored.
-    @tparam NumberType The type of the length @a len
-    @pre len >= 1
-    @return `true` if the string was successfully parsed
-    */
-    template<typename NumberType>
-    bool get_bson_string(const NumberType len, string_t& result)
-    {
-        if (JSON_HEDLEY_UNLIKELY(len < 1))
-        {
-            auto last_token = get_token_string();
-            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                    exception_message(input_format_t::bson, concat("string length must be at least 1, is ", std::to_string(len)), "string"), nullptr));
-        }
-
-        return get_string(input_format_t::bson, len - static_cast<NumberType>(1), result) && get() != std::char_traits<char_type>::eof();
-    }
-
-    /*!
-    @brief Parses a byte array input of length @a len from the BSON input.
-    @param[in] len  The length of the byte array to be read.
-    @param[in,out] result  A reference to the binary variable where the read
-                            array is to be stored.
-    @tparam NumberType The type of the length @a len
-    @pre len >= 0
-    @return `true` if the byte array was successfully parsed
-    */
-    template<typename NumberType>
-    bool get_bson_binary(const NumberType len, binary_t& result)
-    {
-        if (JSON_HEDLEY_UNLIKELY(len < 0))
-        {
-            auto last_token = get_token_string();
-            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                    exception_message(input_format_t::bson, concat("byte array length cannot be negative, is ", std::to_string(len)), "binary"), nullptr));
-        }
-
-        // All BSON binary values have a subtype
-        std::uint8_t subtype{};
-        get_number<std::uint8_t>(input_format_t::bson, subtype);
-        result.set_subtype(subtype);
-
-        return get_binary(input_format_t::bson, len, result);
-    }
-
-    /*!
-    @brief Read a BSON document element of the given @a element_type.
-    @param[in] element_type The BSON element type, c.f. http://bsonspec.org/spec.html
-    @param[in] element_type_parse_position The position in the input stream,
-               where the `element_type` was read.
-    @warning Not all BSON element types are supported yet. An unsupported
-             @a element_type will give rise to a parse_error.114:
-             Unsupported BSON record type 0x...
-    @return whether a valid BSON-object/array was passed to the SAX parser
-    */
-    bool parse_bson_element_internal(const char_int_type element_type,
-                                     const std::size_t element_type_parse_position)
-    {
-        switch (element_type)
-        {
-            case 0x01: // double
-            {
-                double number{};
-                return get_number<double, true>(input_format_t::bson, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 0x02: // string
-            {
-                std::int32_t len{};
-                string_t value;
-                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_string(len, value) && sax->string(value);
-            }
-
-            case 0x03: // object
-            {
-                return parse_bson_internal();
-            }
-
-            case 0x04: // array
-            {
-                return parse_bson_array();
-            }
-
-            case 0x05: // binary
-            {
-                std::int32_t len{};
-                binary_t value;
-                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_binary(len, value) && sax->binary(value);
-            }
-
-            case 0x08: // boolean
-            {
-                return sax->boolean(get() != 0);
-            }
-
-            case 0x0A: // null
-            {
-                return sax->null();
-            }
-
-            case 0x10: // int32
-            {
-                std::int32_t value{};
-                return get_number<std::int32_t, true>(input_format_t::bson, value) && sax->number_integer(value);
-            }
-
-            case 0x12: // int64
-            {
-                std::int64_t value{};
-                return get_number<std::int64_t, true>(input_format_t::bson, value) && sax->number_integer(value);
-            }
-
-            default: // anything else not supported (yet)
-            {
-                std::array<char, 3> cr{{}};
-                static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(element_type))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-                std::string cr_str{cr.data()};
-                return sax->parse_error(element_type_parse_position, cr_str,
-                                        parse_error::create(114, element_type_parse_position, concat("Unsupported BSON record type 0x", cr_str), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @brief Read a BSON element list (as specified in the BSON-spec)
-
-    The same binary layout is used for objects and arrays, hence it must be
-    indicated with the argument @a is_array which one is expected
-    (true --> array, false --> object).
-
-    @param[in] is_array Determines if the element list being read is to be
-                        treated as an object (@a is_array == false), or as an
-                        array (@a is_array == true).
-    @return whether a valid BSON-object/array was passed to the SAX parser
-    */
-    bool parse_bson_element_list(const bool is_array)
-    {
-        string_t key;
-
-        while (auto element_type = get())
-        {
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "element list")))
-            {
-                return false;
-            }
-
-            const std::size_t element_type_parse_position = chars_read;
-            if (JSON_HEDLEY_UNLIKELY(!get_bson_cstr(key)))
-            {
-                return false;
-            }
-
-            if (!is_array && !sax->key(key))
-            {
-                return false;
-            }
-
-            if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_internal(element_type, element_type_parse_position)))
-            {
-                return false;
-            }
-
-            // get_bson_cstr only appends
-            key.clear();
-        }
-
-        return true;
-    }
-
-    /*!
-    @brief Reads an array from the BSON input and passes it to the SAX-parser.
-    @return whether a valid BSON-array was passed to the SAX parser
-    */
-    bool parse_bson_array()
-    {
-        std::int32_t document_size{};
-        get_number<std::int32_t, true>(input_format_t::bson, document_size);
-
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
-        {
-            return false;
-        }
-
-        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/true)))
-        {
-            return false;
-        }
-
-        return sax->end_array();
-    }
-
-    //////////
-    // CBOR //
-    //////////
-
-    /*!
-    @param[in] get_char  whether a new character should be retrieved from the
-                         input (true) or whether the last read character should
-                         be considered instead (false)
-    @param[in] tag_handler how CBOR tags should be treated
-
-    @return whether a valid CBOR value was passed to the SAX parser
-    */
-    bool parse_cbor_internal(const bool get_char,
-                             const cbor_tag_handler_t tag_handler)
-    {
-        switch (get_char ? get() : current)
-        {
-            // EOF
-            case std::char_traits<char_type>::eof():
-                return unexpect_eof(input_format_t::cbor, "value");
-
-            // Integer 0x00..0x17 (0..23)
-            case 0x00:
-            case 0x01:
-            case 0x02:
-            case 0x03:
-            case 0x04:
-            case 0x05:
-            case 0x06:
-            case 0x07:
-            case 0x08:
-            case 0x09:
-            case 0x0A:
-            case 0x0B:
-            case 0x0C:
-            case 0x0D:
-            case 0x0E:
-            case 0x0F:
-            case 0x10:
-            case 0x11:
-            case 0x12:
-            case 0x13:
-            case 0x14:
-            case 0x15:
-            case 0x16:
-            case 0x17:
-                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
-
-            case 0x18: // Unsigned integer (one-byte uint8_t follows)
-            {
-                std::uint8_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
-            }
-
-            case 0x19: // Unsigned integer (two-byte uint16_t follows)
-            {
-                std::uint16_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
-            }
-
-            case 0x1A: // Unsigned integer (four-byte uint32_t follows)
-            {
-                std::uint32_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
-            }
-
-            case 0x1B: // Unsigned integer (eight-byte uint64_t follows)
-            {
-                std::uint64_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
-            }
-
-            // Negative integer -1-0x00..-1-0x17 (-1..-24)
-            case 0x20:
-            case 0x21:
-            case 0x22:
-            case 0x23:
-            case 0x24:
-            case 0x25:
-            case 0x26:
-            case 0x27:
-            case 0x28:
-            case 0x29:
-            case 0x2A:
-            case 0x2B:
-            case 0x2C:
-            case 0x2D:
-            case 0x2E:
-            case 0x2F:
-            case 0x30:
-            case 0x31:
-            case 0x32:
-            case 0x33:
-            case 0x34:
-            case 0x35:
-            case 0x36:
-            case 0x37:
-                return sax->number_integer(static_cast<std::int8_t>(0x20 - 1 - current));
-
-            case 0x38: // Negative integer (one-byte uint8_t follows)
-            {
-                std::uint8_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
-            }
-
-            case 0x39: // Negative integer -1-n (two-byte uint16_t follows)
-            {
-                std::uint16_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
-            }
-
-            case 0x3A: // Negative integer -1-n (four-byte uint32_t follows)
-            {
-                std::uint32_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
-            }
-
-            case 0x3B: // Negative integer -1-n (eight-byte uint64_t follows)
-            {
-                std::uint64_t number{};
-                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1)
-                        - static_cast<number_integer_t>(number));
-            }
-
-            // Binary data (0x00..0x17 bytes follow)
-            case 0x40:
-            case 0x41:
-            case 0x42:
-            case 0x43:
-            case 0x44:
-            case 0x45:
-            case 0x46:
-            case 0x47:
-            case 0x48:
-            case 0x49:
-            case 0x4A:
-            case 0x4B:
-            case 0x4C:
-            case 0x4D:
-            case 0x4E:
-            case 0x4F:
-            case 0x50:
-            case 0x51:
-            case 0x52:
-            case 0x53:
-            case 0x54:
-            case 0x55:
-            case 0x56:
-            case 0x57:
-            case 0x58: // Binary data (one-byte uint8_t for n follows)
-            case 0x59: // Binary data (two-byte uint16_t for n follow)
-            case 0x5A: // Binary data (four-byte uint32_t for n follow)
-            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
-            case 0x5F: // Binary data (indefinite length)
-            {
-                binary_t b;
-                return get_cbor_binary(b) && sax->binary(b);
-            }
-
-            // UTF-8 string (0x00..0x17 bytes follow)
-            case 0x60:
-            case 0x61:
-            case 0x62:
-            case 0x63:
-            case 0x64:
-            case 0x65:
-            case 0x66:
-            case 0x67:
-            case 0x68:
-            case 0x69:
-            case 0x6A:
-            case 0x6B:
-            case 0x6C:
-            case 0x6D:
-            case 0x6E:
-            case 0x6F:
-            case 0x70:
-            case 0x71:
-            case 0x72:
-            case 0x73:
-            case 0x74:
-            case 0x75:
-            case 0x76:
-            case 0x77:
-            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
-            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
-            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
-            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
-            case 0x7F: // UTF-8 string (indefinite length)
-            {
-                string_t s;
-                return get_cbor_string(s) && sax->string(s);
-            }
-
-            // array (0x00..0x17 data items follow)
-            case 0x80:
-            case 0x81:
-            case 0x82:
-            case 0x83:
-            case 0x84:
-            case 0x85:
-            case 0x86:
-            case 0x87:
-            case 0x88:
-            case 0x89:
-            case 0x8A:
-            case 0x8B:
-            case 0x8C:
-            case 0x8D:
-            case 0x8E:
-            case 0x8F:
-            case 0x90:
-            case 0x91:
-            case 0x92:
-            case 0x93:
-            case 0x94:
-            case 0x95:
-            case 0x96:
-            case 0x97:
-                return get_cbor_array(
-                           conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
-
-            case 0x98: // array (one-byte uint8_t for n follows)
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0x99: // array (two-byte uint16_t for n follow)
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0x9A: // array (four-byte uint32_t for n follow)
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0x9B: // array (eight-byte uint64_t for n follow)
-            {
-                std::uint64_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0x9F: // array (indefinite length)
-                return get_cbor_array(static_cast<std::size_t>(-1), tag_handler);
-
-            // map (0x00..0x17 pairs of data items follow)
-            case 0xA0:
-            case 0xA1:
-            case 0xA2:
-            case 0xA3:
-            case 0xA4:
-            case 0xA5:
-            case 0xA6:
-            case 0xA7:
-            case 0xA8:
-            case 0xA9:
-            case 0xAA:
-            case 0xAB:
-            case 0xAC:
-            case 0xAD:
-            case 0xAE:
-            case 0xAF:
-            case 0xB0:
-            case 0xB1:
-            case 0xB2:
-            case 0xB3:
-            case 0xB4:
-            case 0xB5:
-            case 0xB6:
-            case 0xB7:
-                return get_cbor_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
-
-            case 0xB8: // map (one-byte uint8_t for n follows)
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0xB9: // map (two-byte uint16_t for n follow)
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0xBA: // map (four-byte uint32_t for n follow)
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0xBB: // map (eight-byte uint64_t for n follow)
-            {
-                std::uint64_t len{};
-                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
-            }
-
-            case 0xBF: // map (indefinite length)
-                return get_cbor_object(static_cast<std::size_t>(-1), tag_handler);
-
-            case 0xC6: // tagged item
-            case 0xC7:
-            case 0xC8:
-            case 0xC9:
-            case 0xCA:
-            case 0xCB:
-            case 0xCC:
-            case 0xCD:
-            case 0xCE:
-            case 0xCF:
-            case 0xD0:
-            case 0xD1:
-            case 0xD2:
-            case 0xD3:
-            case 0xD4:
-            case 0xD8: // tagged item (1 bytes follow)
-            case 0xD9: // tagged item (2 bytes follow)
-            case 0xDA: // tagged item (4 bytes follow)
-            case 0xDB: // tagged item (8 bytes follow)
-            {
-                switch (tag_handler)
-                {
-                    case cbor_tag_handler_t::error:
-                    {
-                        auto last_token = get_token_string();
-                        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                                exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
-                    }
-
-                    case cbor_tag_handler_t::ignore:
-                    {
-                        // ignore binary subtype
-                        switch (current)
-                        {
-                            case 0xD8:
-                            {
-                                std::uint8_t subtype_to_ignore{};
-                                get_number(input_format_t::cbor, subtype_to_ignore);
-                                break;
-                            }
-                            case 0xD9:
-                            {
-                                std::uint16_t subtype_to_ignore{};
-                                get_number(input_format_t::cbor, subtype_to_ignore);
-                                break;
-                            }
-                            case 0xDA:
-                            {
-                                std::uint32_t subtype_to_ignore{};
-                                get_number(input_format_t::cbor, subtype_to_ignore);
-                                break;
-                            }
-                            case 0xDB:
-                            {
-                                std::uint64_t subtype_to_ignore{};
-                                get_number(input_format_t::cbor, subtype_to_ignore);
-                                break;
-                            }
-                            default:
-                                break;
-                        }
-                        return parse_cbor_internal(true, tag_handler);
-                    }
-
-                    case cbor_tag_handler_t::store:
-                    {
-                        binary_t b;
-                        // use binary subtype and store in binary container
-                        switch (current)
-                        {
-                            case 0xD8:
-                            {
-                                std::uint8_t subtype{};
-                                get_number(input_format_t::cbor, subtype);
-                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
-                                break;
-                            }
-                            case 0xD9:
-                            {
-                                std::uint16_t subtype{};
-                                get_number(input_format_t::cbor, subtype);
-                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
-                                break;
-                            }
-                            case 0xDA:
-                            {
-                                std::uint32_t subtype{};
-                                get_number(input_format_t::cbor, subtype);
-                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
-                                break;
-                            }
-                            case 0xDB:
-                            {
-                                std::uint64_t subtype{};
-                                get_number(input_format_t::cbor, subtype);
-                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
-                                break;
-                            }
-                            default:
-                                return parse_cbor_internal(true, tag_handler);
-                        }
-                        get();
-                        return get_cbor_binary(b) && sax->binary(b);
-                    }
-
-                    default:                 // LCOV_EXCL_LINE
-                        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-                        return false;        // LCOV_EXCL_LINE
-                }
-            }
-
-            case 0xF4: // false
-                return sax->boolean(false);
-
-            case 0xF5: // true
-                return sax->boolean(true);
-
-            case 0xF6: // null
-                return sax->null();
-
-            case 0xF9: // Half-Precision Float (two-byte IEEE 754)
-            {
-                const auto byte1_raw = get();
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
-                {
-                    return false;
-                }
-                const auto byte2_raw = get();
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
-                {
-                    return false;
-                }
-
-                const auto byte1 = static_cast<unsigned char>(byte1_raw);
-                const auto byte2 = static_cast<unsigned char>(byte2_raw);
-
-                // code from RFC 7049, Appendix D, Figure 3:
-                // As half-precision floating-point numbers were only added
-                // to IEEE 754 in 2008, today's programming platforms often
-                // still only have limited support for them. It is very
-                // easy to include at least decoding support for them even
-                // without such support. An example of a small decoder for
-                // half-precision floating-point numbers in the C language
-                // is shown in Fig. 3.
-                const auto half = static_cast<unsigned int>((byte1 << 8u) + byte2);
-                const double val = [&half]
-                {
-                    const int exp = (half >> 10u) & 0x1Fu;
-                    const unsigned int mant = half & 0x3FFu;
-                    JSON_ASSERT(0 <= exp&& exp <= 32);
-                    JSON_ASSERT(mant <= 1024);
-                    switch (exp)
-                    {
-                        case 0:
-                            return std::ldexp(mant, -24);
-                        case 31:
-                            return (mant == 0)
-                            ? std::numeric_limits<double>::infinity()
-                            : std::numeric_limits<double>::quiet_NaN();
-                        default:
-                            return std::ldexp(mant + 1024, exp - 25);
-                    }
-                }();
-                return sax->number_float((half & 0x8000u) != 0
-                                         ? static_cast<number_float_t>(-val)
-                                         : static_cast<number_float_t>(val), "");
-            }
-
-            case 0xFA: // Single-Precision Float (four-byte IEEE 754)
-            {
-                float number{};
-                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 0xFB: // Double-Precision Float (eight-byte IEEE 754)
-            {
-                double number{};
-                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            default: // anything else (0xFF is handled inside the other types)
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                        exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @brief reads a CBOR string
-
-    This function first reads starting bytes to determine the expected
-    string length and then copies this number of bytes into a string.
-    Additionally, CBOR's strings with indefinite lengths are supported.
-
-    @param[out] result  created string
-
-    @return whether string creation completed
-    */
-    bool get_cbor_string(string_t& result)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "string")))
-        {
-            return false;
-        }
-
-        switch (current)
-        {
-            // UTF-8 string (0x00..0x17 bytes follow)
-            case 0x60:
-            case 0x61:
-            case 0x62:
-            case 0x63:
-            case 0x64:
-            case 0x65:
-            case 0x66:
-            case 0x67:
-            case 0x68:
-            case 0x69:
-            case 0x6A:
-            case 0x6B:
-            case 0x6C:
-            case 0x6D:
-            case 0x6E:
-            case 0x6F:
-            case 0x70:
-            case 0x71:
-            case 0x72:
-            case 0x73:
-            case 0x74:
-            case 0x75:
-            case 0x76:
-            case 0x77:
-            {
-                return get_string(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
-            }
-
-            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
-            }
-
-            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
-            }
-
-            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
-            }
-
-            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
-            {
-                std::uint64_t len{};
-                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
-            }
-
-            case 0x7F: // UTF-8 string (indefinite length)
-            {
-                while (get() != 0xFF)
-                {
-                    string_t chunk;
-                    if (!get_cbor_string(chunk))
-                    {
-                        return false;
-                    }
-                    result.append(chunk);
-                }
-                return true;
-            }
-
-            default:
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
-                                        exception_message(input_format_t::cbor, concat("expected length specification (0x60-0x7B) or indefinite string type (0x7F); last byte: 0x", last_token), "string"), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @brief reads a CBOR byte array
-
-    This function first reads starting bytes to determine the expected
-    byte array length and then copies this number of bytes into the byte array.
-    Additionally, CBOR's byte arrays with indefinite lengths are supported.
-
-    @param[out] result  created byte array
-
-    @return whether byte array creation completed
-    */
-    bool get_cbor_binary(binary_t& result)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "binary")))
-        {
-            return false;
-        }
-
-        switch (current)
-        {
-            // Binary data (0x00..0x17 bytes follow)
-            case 0x40:
-            case 0x41:
-            case 0x42:
-            case 0x43:
-            case 0x44:
-            case 0x45:
-            case 0x46:
-            case 0x47:
-            case 0x48:
-            case 0x49:
-            case 0x4A:
-            case 0x4B:
-            case 0x4C:
-            case 0x4D:
-            case 0x4E:
-            case 0x4F:
-            case 0x50:
-            case 0x51:
-            case 0x52:
-            case 0x53:
-            case 0x54:
-            case 0x55:
-            case 0x56:
-            case 0x57:
-            {
-                return get_binary(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
-            }
-
-            case 0x58: // Binary data (one-byte uint8_t for n follows)
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::cbor, len) &&
-                       get_binary(input_format_t::cbor, len, result);
-            }
-
-            case 0x59: // Binary data (two-byte uint16_t for n follow)
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::cbor, len) &&
-                       get_binary(input_format_t::cbor, len, result);
-            }
-
-            case 0x5A: // Binary data (four-byte uint32_t for n follow)
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::cbor, len) &&
-                       get_binary(input_format_t::cbor, len, result);
-            }
-
-            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
-            {
-                std::uint64_t len{};
-                return get_number(input_format_t::cbor, len) &&
-                       get_binary(input_format_t::cbor, len, result);
-            }
-
-            case 0x5F: // Binary data (indefinite length)
-            {
-                while (get() != 0xFF)
-                {
-                    binary_t chunk;
-                    if (!get_cbor_binary(chunk))
-                    {
-                        return false;
-                    }
-                    result.insert(result.end(), chunk.begin(), chunk.end());
-                }
-                return true;
-            }
-
-            default:
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
-                                        exception_message(input_format_t::cbor, concat("expected length specification (0x40-0x5B) or indefinite binary array type (0x5F); last byte: 0x", last_token), "binary"), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @param[in] len  the length of the array or static_cast<std::size_t>(-1) for an
-                    array of indefinite size
-    @param[in] tag_handler how CBOR tags should be treated
-    @return whether array creation completed
-    */
-    bool get_cbor_array(const std::size_t len,
-                        const cbor_tag_handler_t tag_handler)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
-        {
-            return false;
-        }
-
-        if (len != static_cast<std::size_t>(-1))
-        {
-            for (std::size_t i = 0; i < len; ++i)
-            {
-                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
-                {
-                    return false;
-                }
-            }
-        }
-        else
-        {
-            while (get() != 0xFF)
-            {
-                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(false, tag_handler)))
-                {
-                    return false;
-                }
-            }
-        }
-
-        return sax->end_array();
-    }
-
-    /*!
-    @param[in] len  the length of the object or static_cast<std::size_t>(-1) for an
-                    object of indefinite size
-    @param[in] tag_handler how CBOR tags should be treated
-    @return whether object creation completed
-    */
-    bool get_cbor_object(const std::size_t len,
-                         const cbor_tag_handler_t tag_handler)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
-        {
-            return false;
-        }
-
-        if (len != 0)
-        {
-            string_t key;
-            if (len != static_cast<std::size_t>(-1))
-            {
-                for (std::size_t i = 0; i < len; ++i)
-                {
-                    get();
-                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
-                    {
-                        return false;
-                    }
-
-                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
-                    {
-                        return false;
-                    }
-                    key.clear();
-                }
-            }
-            else
-            {
-                while (get() != 0xFF)
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
-                    {
-                        return false;
-                    }
-
-                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
-                    {
-                        return false;
-                    }
-                    key.clear();
-                }
-            }
-        }
-
-        return sax->end_object();
-    }
-
-    /////////////
-    // MsgPack //
-    /////////////
-
-    /*!
-    @return whether a valid MessagePack value was passed to the SAX parser
-    */
-    bool parse_msgpack_internal()
-    {
-        switch (get())
-        {
-            // EOF
-            case std::char_traits<char_type>::eof():
-                return unexpect_eof(input_format_t::msgpack, "value");
-
-            // positive fixint
-            case 0x00:
-            case 0x01:
-            case 0x02:
-            case 0x03:
-            case 0x04:
-            case 0x05:
-            case 0x06:
-            case 0x07:
-            case 0x08:
-            case 0x09:
-            case 0x0A:
-            case 0x0B:
-            case 0x0C:
-            case 0x0D:
-            case 0x0E:
-            case 0x0F:
-            case 0x10:
-            case 0x11:
-            case 0x12:
-            case 0x13:
-            case 0x14:
-            case 0x15:
-            case 0x16:
-            case 0x17:
-            case 0x18:
-            case 0x19:
-            case 0x1A:
-            case 0x1B:
-            case 0x1C:
-            case 0x1D:
-            case 0x1E:
-            case 0x1F:
-            case 0x20:
-            case 0x21:
-            case 0x22:
-            case 0x23:
-            case 0x24:
-            case 0x25:
-            case 0x26:
-            case 0x27:
-            case 0x28:
-            case 0x29:
-            case 0x2A:
-            case 0x2B:
-            case 0x2C:
-            case 0x2D:
-            case 0x2E:
-            case 0x2F:
-            case 0x30:
-            case 0x31:
-            case 0x32:
-            case 0x33:
-            case 0x34:
-            case 0x35:
-            case 0x36:
-            case 0x37:
-            case 0x38:
-            case 0x39:
-            case 0x3A:
-            case 0x3B:
-            case 0x3C:
-            case 0x3D:
-            case 0x3E:
-            case 0x3F:
-            case 0x40:
-            case 0x41:
-            case 0x42:
-            case 0x43:
-            case 0x44:
-            case 0x45:
-            case 0x46:
-            case 0x47:
-            case 0x48:
-            case 0x49:
-            case 0x4A:
-            case 0x4B:
-            case 0x4C:
-            case 0x4D:
-            case 0x4E:
-            case 0x4F:
-            case 0x50:
-            case 0x51:
-            case 0x52:
-            case 0x53:
-            case 0x54:
-            case 0x55:
-            case 0x56:
-            case 0x57:
-            case 0x58:
-            case 0x59:
-            case 0x5A:
-            case 0x5B:
-            case 0x5C:
-            case 0x5D:
-            case 0x5E:
-            case 0x5F:
-            case 0x60:
-            case 0x61:
-            case 0x62:
-            case 0x63:
-            case 0x64:
-            case 0x65:
-            case 0x66:
-            case 0x67:
-            case 0x68:
-            case 0x69:
-            case 0x6A:
-            case 0x6B:
-            case 0x6C:
-            case 0x6D:
-            case 0x6E:
-            case 0x6F:
-            case 0x70:
-            case 0x71:
-            case 0x72:
-            case 0x73:
-            case 0x74:
-            case 0x75:
-            case 0x76:
-            case 0x77:
-            case 0x78:
-            case 0x79:
-            case 0x7A:
-            case 0x7B:
-            case 0x7C:
-            case 0x7D:
-            case 0x7E:
-            case 0x7F:
-                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
-
-            // fixmap
-            case 0x80:
-            case 0x81:
-            case 0x82:
-            case 0x83:
-            case 0x84:
-            case 0x85:
-            case 0x86:
-            case 0x87:
-            case 0x88:
-            case 0x89:
-            case 0x8A:
-            case 0x8B:
-            case 0x8C:
-            case 0x8D:
-            case 0x8E:
-            case 0x8F:
-                return get_msgpack_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
-
-            // fixarray
-            case 0x90:
-            case 0x91:
-            case 0x92:
-            case 0x93:
-            case 0x94:
-            case 0x95:
-            case 0x96:
-            case 0x97:
-            case 0x98:
-            case 0x99:
-            case 0x9A:
-            case 0x9B:
-            case 0x9C:
-            case 0x9D:
-            case 0x9E:
-            case 0x9F:
-                return get_msgpack_array(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
-
-            // fixstr
-            case 0xA0:
-            case 0xA1:
-            case 0xA2:
-            case 0xA3:
-            case 0xA4:
-            case 0xA5:
-            case 0xA6:
-            case 0xA7:
-            case 0xA8:
-            case 0xA9:
-            case 0xAA:
-            case 0xAB:
-            case 0xAC:
-            case 0xAD:
-            case 0xAE:
-            case 0xAF:
-            case 0xB0:
-            case 0xB1:
-            case 0xB2:
-            case 0xB3:
-            case 0xB4:
-            case 0xB5:
-            case 0xB6:
-            case 0xB7:
-            case 0xB8:
-            case 0xB9:
-            case 0xBA:
-            case 0xBB:
-            case 0xBC:
-            case 0xBD:
-            case 0xBE:
-            case 0xBF:
-            case 0xD9: // str 8
-            case 0xDA: // str 16
-            case 0xDB: // str 32
-            {
-                string_t s;
-                return get_msgpack_string(s) && sax->string(s);
-            }
-
-            case 0xC0: // nil
-                return sax->null();
-
-            case 0xC2: // false
-                return sax->boolean(false);
-
-            case 0xC3: // true
-                return sax->boolean(true);
-
-            case 0xC4: // bin 8
-            case 0xC5: // bin 16
-            case 0xC6: // bin 32
-            case 0xC7: // ext 8
-            case 0xC8: // ext 16
-            case 0xC9: // ext 32
-            case 0xD4: // fixext 1
-            case 0xD5: // fixext 2
-            case 0xD6: // fixext 4
-            case 0xD7: // fixext 8
-            case 0xD8: // fixext 16
-            {
-                binary_t b;
-                return get_msgpack_binary(b) && sax->binary(b);
-            }
-
-            case 0xCA: // float 32
-            {
-                float number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 0xCB: // float 64
-            {
-                double number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 0xCC: // uint 8
-            {
-                std::uint8_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
-            }
-
-            case 0xCD: // uint 16
-            {
-                std::uint16_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
-            }
-
-            case 0xCE: // uint 32
-            {
-                std::uint32_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
-            }
-
-            case 0xCF: // uint 64
-            {
-                std::uint64_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
-            }
-
-            case 0xD0: // int 8
-            {
-                std::int8_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
-            }
-
-            case 0xD1: // int 16
-            {
-                std::int16_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
-            }
-
-            case 0xD2: // int 32
-            {
-                std::int32_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
-            }
-
-            case 0xD3: // int 64
-            {
-                std::int64_t number{};
-                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
-            }
-
-            case 0xDC: // array 16
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::msgpack, len) && get_msgpack_array(static_cast<std::size_t>(len));
-            }
-
-            case 0xDD: // array 32
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::msgpack, len) && get_msgpack_array(conditional_static_cast<std::size_t>(len));
-            }
-
-            case 0xDE: // map 16
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::msgpack, len) && get_msgpack_object(static_cast<std::size_t>(len));
-            }
-
-            case 0xDF: // map 32
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::msgpack, len) && get_msgpack_object(conditional_static_cast<std::size_t>(len));
-            }
-
-            // negative fixint
-            case 0xE0:
-            case 0xE1:
-            case 0xE2:
-            case 0xE3:
-            case 0xE4:
-            case 0xE5:
-            case 0xE6:
-            case 0xE7:
-            case 0xE8:
-            case 0xE9:
-            case 0xEA:
-            case 0xEB:
-            case 0xEC:
-            case 0xED:
-            case 0xEE:
-            case 0xEF:
-            case 0xF0:
-            case 0xF1:
-            case 0xF2:
-            case 0xF3:
-            case 0xF4:
-            case 0xF5:
-            case 0xF6:
-            case 0xF7:
-            case 0xF8:
-            case 0xF9:
-            case 0xFA:
-            case 0xFB:
-            case 0xFC:
-            case 0xFD:
-            case 0xFE:
-            case 0xFF:
-                return sax->number_integer(static_cast<std::int8_t>(current));
-
-            default: // anything else
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                        exception_message(input_format_t::msgpack, concat("invalid byte: 0x", last_token), "value"), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @brief reads a MessagePack string
-
-    This function first reads starting bytes to determine the expected
-    string length and then copies this number of bytes into a string.
-
-    @param[out] result  created string
-
-    @return whether string creation completed
-    */
-    bool get_msgpack_string(string_t& result)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::msgpack, "string")))
-        {
-            return false;
-        }
-
-        switch (current)
-        {
-            // fixstr
-            case 0xA0:
-            case 0xA1:
-            case 0xA2:
-            case 0xA3:
-            case 0xA4:
-            case 0xA5:
-            case 0xA6:
-            case 0xA7:
-            case 0xA8:
-            case 0xA9:
-            case 0xAA:
-            case 0xAB:
-            case 0xAC:
-            case 0xAD:
-            case 0xAE:
-            case 0xAF:
-            case 0xB0:
-            case 0xB1:
-            case 0xB2:
-            case 0xB3:
-            case 0xB4:
-            case 0xB5:
-            case 0xB6:
-            case 0xB7:
-            case 0xB8:
-            case 0xB9:
-            case 0xBA:
-            case 0xBB:
-            case 0xBC:
-            case 0xBD:
-            case 0xBE:
-            case 0xBF:
-            {
-                return get_string(input_format_t::msgpack, static_cast<unsigned int>(current) & 0x1Fu, result);
-            }
-
-            case 0xD9: // str 8
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
-            }
-
-            case 0xDA: // str 16
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
-            }
-
-            case 0xDB: // str 32
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
-            }
-
-            default:
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
-                                        exception_message(input_format_t::msgpack, concat("expected length specification (0xA0-0xBF, 0xD9-0xDB); last byte: 0x", last_token), "string"), nullptr));
-            }
-        }
-    }
-
-    /*!
-    @brief reads a MessagePack byte array
-
-    This function first reads starting bytes to determine the expected
-    byte array length and then copies this number of bytes into a byte array.
-
-    @param[out] result  created byte array
-
-    @return whether byte array creation completed
-    */
-    bool get_msgpack_binary(binary_t& result)
-    {
-        // helper function to set the subtype
-        auto assign_and_return_true = [&result](std::int8_t subtype)
-        {
-            result.set_subtype(static_cast<std::uint8_t>(subtype));
-            return true;
-        };
-
-        switch (current)
-        {
-            case 0xC4: // bin 8
-            {
-                std::uint8_t len{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_binary(input_format_t::msgpack, len, result);
-            }
-
-            case 0xC5: // bin 16
-            {
-                std::uint16_t len{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_binary(input_format_t::msgpack, len, result);
-            }
-
-            case 0xC6: // bin 32
-            {
-                std::uint32_t len{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_binary(input_format_t::msgpack, len, result);
-            }
-
-            case 0xC7: // ext 8
-            {
-                std::uint8_t len{};
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, len, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xC8: // ext 16
-            {
-                std::uint16_t len{};
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, len, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xC9: // ext 32
-            {
-                std::uint32_t len{};
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, len) &&
-                       get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, len, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xD4: // fixext 1
-            {
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, 1, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xD5: // fixext 2
-            {
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, 2, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xD6: // fixext 4
-            {
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, 4, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xD7: // fixext 8
-            {
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, 8, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            case 0xD8: // fixext 16
-            {
-                std::int8_t subtype{};
-                return get_number(input_format_t::msgpack, subtype) &&
-                       get_binary(input_format_t::msgpack, 16, result) &&
-                       assign_and_return_true(subtype);
-            }
-
-            default:           // LCOV_EXCL_LINE
-                return false;  // LCOV_EXCL_LINE
-        }
-    }
-
-    /*!
-    @param[in] len  the length of the array
-    @return whether array creation completed
-    */
-    bool get_msgpack_array(const std::size_t len)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
-        {
-            return false;
-        }
-
-        for (std::size_t i = 0; i < len; ++i)
-        {
-            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
-            {
-                return false;
-            }
-        }
-
-        return sax->end_array();
-    }
-
-    /*!
-    @param[in] len  the length of the object
-    @return whether object creation completed
-    */
-    bool get_msgpack_object(const std::size_t len)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
-        {
-            return false;
-        }
-
-        string_t key;
-        for (std::size_t i = 0; i < len; ++i)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!get_msgpack_string(key) || !sax->key(key)))
-            {
-                return false;
-            }
-
-            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
-            {
-                return false;
-            }
-            key.clear();
-        }
-
-        return sax->end_object();
-    }
-
-    ////////////
-    // UBJSON //
-    ////////////
-
-    /*!
-    @param[in] get_char  whether a new character should be retrieved from the
-                         input (true, default) or whether the last read
-                         character should be considered instead
-
-    @return whether a valid UBJSON value was passed to the SAX parser
-    */
-    bool parse_ubjson_internal(const bool get_char = true)
-    {
-        return get_ubjson_value(get_char ? get_ignore_noop() : current);
-    }
-
-    /*!
-    @brief reads a UBJSON string
-
-    This function is either called after reading the 'S' byte explicitly
-    indicating a string, or in case of an object key where the 'S' byte can be
-    left out.
-
-    @param[out] result   created string
-    @param[in] get_char  whether a new character should be retrieved from the
-                         input (true, default) or whether the last read
-                         character should be considered instead
-
-    @return whether string creation completed
-    */
-    bool get_ubjson_string(string_t& result, const bool get_char = true)
-    {
-        if (get_char)
-        {
-            get();  // TODO(niels): may we ignore N here?
-        }
-
-        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
-        {
-            return false;
-        }
-
-        switch (current)
-        {
-            case 'U':
-            {
-                std::uint8_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'i':
-            {
-                std::int8_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'I':
-            {
-                std::int16_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'l':
-            {
-                std::int32_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'L':
-            {
-                std::int64_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'u':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint16_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'm':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint32_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            case 'M':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint64_t len{};
-                return get_number(input_format, len) && get_string(input_format, len, result);
-            }
-
-            default:
-                break;
-        }
-        auto last_token = get_token_string();
-        std::string message;
-
-        if (input_format != input_format_t::bjdata)
-        {
-            message = "expected length type specification (U, i, I, l, L); last byte: 0x" + last_token;
-        }
-        else
-        {
-            message = "expected length type specification (U, i, u, I, m, l, M, L); last byte: 0x" + last_token;
-        }
-        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "string"), nullptr));
-    }
-
-    /*!
-    @param[out] dim  an integer vector storing the ND array dimensions
-    @return whether reading ND array size vector is successful
-    */
-    bool get_ubjson_ndarray_size(std::vector<size_t>& dim)
-    {
-        std::pair<std::size_t, char_int_type> size_and_type;
-        size_t dimlen = 0;
-        bool no_ndarray = true;
-
-        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type, no_ndarray)))
-        {
-            return false;
-        }
-
-        if (size_and_type.first != npos)
-        {
-            if (size_and_type.second != 0)
-            {
-                if (size_and_type.second != 'N')
-                {
-                    for (std::size_t i = 0; i < size_and_type.first; ++i)
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, size_and_type.second)))
-                        {
-                            return false;
-                        }
-                        dim.push_back(dimlen);
-                    }
-                }
-            }
-            else
-            {
-                for (std::size_t i = 0; i < size_and_type.first; ++i)
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray)))
-                    {
-                        return false;
-                    }
-                    dim.push_back(dimlen);
-                }
-            }
-        }
-        else
-        {
-            while (current != ']')
-            {
-                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, current)))
-                {
-                    return false;
-                }
-                dim.push_back(dimlen);
-                get_ignore_noop();
-            }
-        }
-        return true;
-    }
-
-    /*!
-    @param[out] result  determined size
-    @param[in,out] is_ndarray  for input, `true` means already inside an ndarray vector
-                               or ndarray dimension is not allowed; `false` means ndarray
-                               is allowed; for output, `true` means an ndarray is found;
-                               is_ndarray can only return `true` when its initial value
-                               is `false`
-    @param[in] prefix  type marker if already read, otherwise set to 0
-
-    @return whether size determination completed
-    */
-    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0)
-    {
-        if (prefix == 0)
-        {
-            prefix = get_ignore_noop();
-        }
-
-        switch (prefix)
-        {
-            case 'U':
-            {
-                std::uint8_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                result = static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'i':
-            {
-                std::int8_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                if (number < 0)
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
-                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
-                }
-                result = static_cast<std::size_t>(number); // NOLINT(bugprone-signed-char-misuse,cert-str34-c): number is not a char
-                return true;
-            }
-
-            case 'I':
-            {
-                std::int16_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                if (number < 0)
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
-                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
-                }
-                result = static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'l':
-            {
-                std::int32_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                if (number < 0)
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
-                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
-                }
-                result = static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'L':
-            {
-                std::int64_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                if (number < 0)
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
-                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
-                }
-                if (!value_in_range_of<std::size_t>(number))
-                {
-                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
-                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
-                }
-                result = static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'u':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint16_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                result = static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'm':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint32_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                result = conditional_static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case 'M':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint64_t number{};
-                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
-                {
-                    return false;
-                }
-                if (!value_in_range_of<std::size_t>(number))
-                {
-                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
-                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
-                }
-                result = detail::conditional_static_cast<std::size_t>(number);
-                return true;
-            }
-
-            case '[':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimensional vector is not allowed", "size"), nullptr));
-                }
-                std::vector<size_t> dim;
-                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))
-                {
-                    return false;
-                }
-                if (dim.size() == 1 || (dim.size() == 2 && dim.at(0) == 1)) // return normal array size if 1D row vector
-                {
-                    result = dim.at(dim.size() - 1);
-                    return true;
-                }
-                if (!dim.empty())  // if ndarray, convert to an object in JData annotated array format
-                {
-                    for (auto i : dim) // test if any dimension in an ndarray is 0, if so, return a 1D empty container
-                    {
-                        if ( i == 0 )
-                        {
-                            result = 0;
-                            return true;
-                        }
-                    }
-
-                    string_t key = "_ArraySize_";
-                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3) || !sax->key(key) || !sax->start_array(dim.size())))
-                    {
-                        return false;
-                    }
-                    result = 1;
-                    for (auto i : dim)
-                    {
-                        result *= i;
-                        if (result == 0 || result == npos) // because dim elements shall not have zeros, result = 0 means overflow happened; it also can't be npos as it is used to initialize size in get_ubjson_size_type()
-                        {
-                            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408, exception_message(input_format, "excessive ndarray size caused overflow", "size"), nullptr));
-                        }
-                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(static_cast<number_unsigned_t>(i))))
-                        {
-                            return false;
-                        }
-                    }
-                    is_ndarray = true;
-                    return sax->end_array();
-                }
-                result = 0;
-                return true;
-            }
-
-            default:
-                break;
-        }
-        auto last_token = get_token_string();
-        std::string message;
-
-        if (input_format != input_format_t::bjdata)
-        {
-            message = "expected length type specification (U, i, I, l, L) after '#'; last byte: 0x" + last_token;
-        }
-        else
-        {
-            message = "expected length type specification (U, i, u, I, m, l, M, L) after '#'; last byte: 0x" + last_token;
-        }
-        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "size"), nullptr));
-    }
-
-    /*!
-    @brief determine the type and size for a container
-
-    In the optimized UBJSON format, a type and a size can be provided to allow
-    for a more compact representation.
-
-    @param[out] result  pair of the size and the type
-    @param[in] inside_ndarray  whether the parser is parsing an ND array dimensional vector
-
-    @return whether pair creation completed
-    */
-    bool get_ubjson_size_type(std::pair<std::size_t, char_int_type>& result, bool inside_ndarray = false)
-    {
-        result.first = npos; // size
-        result.second = 0; // type
-        bool is_ndarray = false;
-
-        get_ignore_noop();
-
-        if (current == '$')
-        {
-            result.second = get();  // must not ignore 'N', because 'N' maybe the type
-            if (input_format == input_format_t::bjdata
-                    && JSON_HEDLEY_UNLIKELY(std::binary_search(bjd_optimized_type_markers.begin(), bjd_optimized_type_markers.end(), result.second)))
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                        exception_message(input_format, concat("marker 0x", last_token, " is not a permitted optimized array type"), "type"), nullptr));
-            }
-
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "type")))
-            {
-                return false;
-            }
-
-            get_ignore_noop();
-            if (JSON_HEDLEY_UNLIKELY(current != '#'))
-            {
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
-                {
-                    return false;
-                }
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                        exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr));
-            }
-
-            bool is_error = get_ubjson_size_value(result.first, is_ndarray);
-            if (input_format == input_format_t::bjdata && is_ndarray)
-            {
-                if (inside_ndarray)
-                {
-                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
-                                            exception_message(input_format, "ndarray can not be recursive", "size"), nullptr));
-                }
-                result.second |= (1 << 8); // use bit 8 to indicate ndarray, all UBJSON and BJData markers should be ASCII letters
-            }
-            return is_error;
-        }
-
-        if (current == '#')
-        {
-            bool is_error = get_ubjson_size_value(result.first, is_ndarray);
-            if (input_format == input_format_t::bjdata && is_ndarray)
-            {
-                return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
-                                        exception_message(input_format, "ndarray requires both type and size", "size"), nullptr));
-            }
-            return is_error;
-        }
-
-        return true;
-    }
-
-    /*!
-    @param prefix  the previously read or set type prefix
-    @return whether value creation completed
-    */
-    bool get_ubjson_value(const char_int_type prefix)
-    {
-        switch (prefix)
-        {
-            case std::char_traits<char_type>::eof():  // EOF
-                return unexpect_eof(input_format, "value");
-
-            case 'T':  // true
-                return sax->boolean(true);
-            case 'F':  // false
-                return sax->boolean(false);
-
-            case 'Z':  // null
-                return sax->null();
-
-            case 'U':
-            {
-                std::uint8_t number{};
-                return get_number(input_format, number) && sax->number_unsigned(number);
-            }
-
-            case 'i':
-            {
-                std::int8_t number{};
-                return get_number(input_format, number) && sax->number_integer(number);
-            }
-
-            case 'I':
-            {
-                std::int16_t number{};
-                return get_number(input_format, number) && sax->number_integer(number);
-            }
-
-            case 'l':
-            {
-                std::int32_t number{};
-                return get_number(input_format, number) && sax->number_integer(number);
-            }
-
-            case 'L':
-            {
-                std::int64_t number{};
-                return get_number(input_format, number) && sax->number_integer(number);
-            }
-
-            case 'u':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint16_t number{};
-                return get_number(input_format, number) && sax->number_unsigned(number);
-            }
-
-            case 'm':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint32_t number{};
-                return get_number(input_format, number) && sax->number_unsigned(number);
-            }
-
-            case 'M':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                std::uint64_t number{};
-                return get_number(input_format, number) && sax->number_unsigned(number);
-            }
-
-            case 'h':
-            {
-                if (input_format != input_format_t::bjdata)
-                {
-                    break;
-                }
-                const auto byte1_raw = get();
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
-                {
-                    return false;
-                }
-                const auto byte2_raw = get();
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
-                {
-                    return false;
-                }
-
-                const auto byte1 = static_cast<unsigned char>(byte1_raw);
-                const auto byte2 = static_cast<unsigned char>(byte2_raw);
-
-                // code from RFC 7049, Appendix D, Figure 3:
-                // As half-precision floating-point numbers were only added
-                // to IEEE 754 in 2008, today's programming platforms often
-                // still only have limited support for them. It is very
-                // easy to include at least decoding support for them even
-                // without such support. An example of a small decoder for
-                // half-precision floating-point numbers in the C language
-                // is shown in Fig. 3.
-                const auto half = static_cast<unsigned int>((byte2 << 8u) + byte1);
-                const double val = [&half]
-                {
-                    const int exp = (half >> 10u) & 0x1Fu;
-                    const unsigned int mant = half & 0x3FFu;
-                    JSON_ASSERT(0 <= exp&& exp <= 32);
-                    JSON_ASSERT(mant <= 1024);
-                    switch (exp)
-                    {
-                        case 0:
-                            return std::ldexp(mant, -24);
-                        case 31:
-                            return (mant == 0)
-                            ? std::numeric_limits<double>::infinity()
-                            : std::numeric_limits<double>::quiet_NaN();
-                        default:
-                            return std::ldexp(mant + 1024, exp - 25);
-                    }
-                }();
-                return sax->number_float((half & 0x8000u) != 0
-                                         ? static_cast<number_float_t>(-val)
-                                         : static_cast<number_float_t>(val), "");
-            }
-
-            case 'd':
-            {
-                float number{};
-                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 'D':
-            {
-                double number{};
-                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
-            }
-
-            case 'H':
-            {
-                return get_ubjson_high_precision_number();
-            }
-
-            case 'C':  // char
-            {
-                get();
-                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "char")))
-                {
-                    return false;
-                }
-                if (JSON_HEDLEY_UNLIKELY(current > 127))
-                {
-                    auto last_token = get_token_string();
-                    return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
-                                            exception_message(input_format, concat("byte after 'C' must be in range 0x00..0x7F; last byte: 0x", last_token), "char"), nullptr));
-                }
-                string_t s(1, static_cast<typename string_t::value_type>(current));
-                return sax->string(s);
-            }
-
-            case 'S':  // string
-            {
-                string_t s;
-                return get_ubjson_string(s) && sax->string(s);
-            }
-
-            case '[':  // array
-                return get_ubjson_array();
-
-            case '{':  // object
-                return get_ubjson_object();
-
-            default: // anything else
-                break;
-        }
-        auto last_token = get_token_string();
-        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format, "invalid byte: 0x" + last_token, "value"), nullptr));
-    }
-
-    /*!
-    @return whether array creation completed
-    */
-    bool get_ubjson_array()
-    {
-        std::pair<std::size_t, char_int_type> size_and_type;
-        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
-        {
-            return false;
-        }
-
-        // if bit-8 of size_and_type.second is set to 1, encode bjdata ndarray as an object in JData annotated array format (https://github.com/NeuroJSON/jdata):
-        // {"_ArrayType_" : "typeid", "_ArraySize_" : [n1, n2, ...], "_ArrayData_" : [v1, v2, ...]}
-
-        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
-        {
-            size_and_type.second &= ~(static_cast<char_int_type>(1) << 8);  // use bit 8 to indicate ndarray, here we remove the bit to restore the type marker
-            auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), size_and_type.second, [](const bjd_type & p, char_int_type t)
-            {
-                return p.first < t;
-            });
-            string_t key = "_ArrayType_";
-            if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != size_and_type.second))
-            {
-                auto last_token = get_token_string();
-                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                        exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
-            }
-
-            string_t type = it->second; // sax->string() takes a reference
-            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->string(type)))
-            {
-                return false;
-            }
-
-            if (size_and_type.second == 'C')
-            {
-                size_and_type.second = 'U';
-            }
-
-            key = "_ArrayData_";
-            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(size_and_type.first) ))
-            {
-                return false;
-            }
-
-            for (std::size_t i = 0; i < size_and_type.first; ++i)
-            {
-                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
-                {
-                    return false;
-                }
-            }
-
-            return (sax->end_array() && sax->end_object());
-        }
-
-        if (size_and_type.first != npos)
-        {
-            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(size_and_type.first)))
-            {
-                return false;
-            }
-
-            if (size_and_type.second != 0)
-            {
-                if (size_and_type.second != 'N')
-                {
-                    for (std::size_t i = 0; i < size_and_type.first; ++i)
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
-                        {
-                            return false;
-                        }
-                    }
-                }
-            }
-            else
-            {
-                for (std::size_t i = 0; i < size_and_type.first; ++i)
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
-                    {
-                        return false;
-                    }
-                }
-            }
-        }
-        else
-        {
-            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
-            {
-                return false;
-            }
-
-            while (current != ']')
-            {
-                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal(false)))
-                {
-                    return false;
-                }
-                get_ignore_noop();
-            }
-        }
-
-        return sax->end_array();
-    }
-
-    /*!
-    @return whether object creation completed
-    */
-    bool get_ubjson_object()
-    {
-        std::pair<std::size_t, char_int_type> size_and_type;
-        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
-        {
-            return false;
-        }
-
-        // do not accept ND-array size in objects in BJData
-        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
-        {
-            auto last_token = get_token_string();
-            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
-                                    exception_message(input_format, "BJData object does not support ND-array size in optimized format", "object"), nullptr));
-        }
-
-        string_t key;
-        if (size_and_type.first != npos)
-        {
-            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(size_and_type.first)))
-            {
-                return false;
-            }
-
-            if (size_and_type.second != 0)
-            {
-                for (std::size_t i = 0; i < size_and_type.first; ++i)
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
-                    {
-                        return false;
-                    }
-                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
-                    {
-                        return false;
-                    }
-                    key.clear();
-                }
-            }
-            else
-            {
-                for (std::size_t i = 0; i < size_and_type.first; ++i)
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
-                    {
-                        return false;
-                    }
-                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
-                    {
-                        return false;
-                    }
-                    key.clear();
-                }
-            }
-        }
-        else
-        {
-            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
-            {
-                return false;
-            }
-
-            while (current != '}')
-            {
-                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key, false) || !sax->key(key)))
-                {
-                    return false;
-                }
-                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
-                {
-                    return false;
-                }
-                get_ignore_noop();
-                key.clear();
-            }
-        }
-
-        return sax->end_object();
-    }
-
-    // Note, no reader for UBJSON binary types is implemented because they do
-    // not exist
-
-    bool get_ubjson_high_precision_number()
-    {
-        // get size of following number string
-        std::size_t size{};
-        bool no_ndarray = true;
-        auto res = get_ubjson_size_value(size, no_ndarray);
-        if (JSON_HEDLEY_UNLIKELY(!res))
-        {
-            return res;
-        }
-
-        // get number string
-        std::vector<char> number_vector;
-        for (std::size_t i = 0; i < size; ++i)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
-            {
-                return false;
-            }
-            number_vector.push_back(static_cast<char>(current));
-        }
-
-        // parse number string
-        using ia_type = decltype(detail::input_adapter(number_vector));
-        auto number_lexer = detail::lexer<BasicJsonType, ia_type>(detail::input_adapter(number_vector), false);
-        const auto result_number = number_lexer.scan();
-        const auto number_string = number_lexer.get_token_string();
-        const auto result_remainder = number_lexer.scan();
-
-        using token_type = typename detail::lexer_base<BasicJsonType>::token_type;
-
-        if (JSON_HEDLEY_UNLIKELY(result_remainder != token_type::end_of_input))
-        {
-            return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
-                                    exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
-        }
-
-        switch (result_number)
-        {
-            case token_type::value_integer:
-                return sax->number_integer(number_lexer.get_number_integer());
-            case token_type::value_unsigned:
-                return sax->number_unsigned(number_lexer.get_number_unsigned());
-            case token_type::value_float:
-                return sax->number_float(number_lexer.get_number_float(), std::move(number_string));
-            case token_type::uninitialized:
-            case token_type::literal_true:
-            case token_type::literal_false:
-            case token_type::literal_null:
-            case token_type::value_string:
-            case token_type::begin_array:
-            case token_type::begin_object:
-            case token_type::end_array:
-            case token_type::end_object:
-            case token_type::name_separator:
-            case token_type::value_separator:
-            case token_type::parse_error:
-            case token_type::end_of_input:
-            case token_type::literal_or_value:
-            default:
-                return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
-                                        exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
-        }
-    }
-
-    ///////////////////////
-    // Utility functions //
-    ///////////////////////
-
-    /*!
-    @brief get next character from the input
-
-    This function provides the interface to the used input adapter. It does
-    not throw in case the input reached EOF, but returns a -'ve valued
-    `std::char_traits<char_type>::eof()` in that case.
-
-    @return character read from the input
-    */
-    char_int_type get()
-    {
-        ++chars_read;
-        return current = ia.get_character();
-    }
-
-    /*!
-    @return character read from the input after ignoring all 'N' entries
-    */
-    char_int_type get_ignore_noop()
-    {
-        do
-        {
-            get();
-        }
-        while (current == 'N');
-
-        return current;
-    }
-
-    /*
-    @brief read a number from the input
-
-    @tparam NumberType the type of the number
-    @param[in] format   the current format (for diagnostics)
-    @param[out] result  number of type @a NumberType
-
-    @return whether conversion completed
-
-    @note This function needs to respect the system's endianness, because
-          bytes in CBOR, MessagePack, and UBJSON are stored in network order
-          (big endian) and therefore need reordering on little endian systems.
-          On the other hand, BSON and BJData use little endian and should reorder
-          on big endian systems.
-    */
-    template<typename NumberType, bool InputIsLittleEndian = false>
-    bool get_number(const input_format_t format, NumberType& result)
-    {
-        // step 1: read input into array with system's byte order
-        std::array<std::uint8_t, sizeof(NumberType)> vec{};
-        for (std::size_t i = 0; i < sizeof(NumberType); ++i)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "number")))
-            {
-                return false;
-            }
-
-            // reverse byte order prior to conversion if necessary
-            if (is_little_endian != (InputIsLittleEndian || format == input_format_t::bjdata))
-            {
-                vec[sizeof(NumberType) - i - 1] = static_cast<std::uint8_t>(current);
-            }
-            else
-            {
-                vec[i] = static_cast<std::uint8_t>(current); // LCOV_EXCL_LINE
-            }
-        }
-
-        // step 2: convert array into number of type T and return
-        std::memcpy(&result, vec.data(), sizeof(NumberType));
-        return true;
-    }
-
-    /*!
-    @brief create a string by reading characters from the input
-
-    @tparam NumberType the type of the number
-    @param[in] format the current format (for diagnostics)
-    @param[in] len number of characters to read
-    @param[out] result string created by reading @a len bytes
-
-    @return whether string creation completed
-
-    @note We can not reserve @a len bytes for the result, because @a len
-          may be too large. Usually, @ref unexpect_eof() detects the end of
-          the input before we run out of string memory.
-    */
-    template<typename NumberType>
-    bool get_string(const input_format_t format,
-                    const NumberType len,
-                    string_t& result)
-    {
-        bool success = true;
-        for (NumberType i = 0; i < len; i++)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "string")))
-            {
-                success = false;
-                break;
-            }
-            result.push_back(static_cast<typename string_t::value_type>(current));
-        }
-        return success;
-    }
-
-    /*!
-    @brief create a byte array by reading bytes from the input
-
-    @tparam NumberType the type of the number
-    @param[in] format the current format (for diagnostics)
-    @param[in] len number of bytes to read
-    @param[out] result byte array created by reading @a len bytes
-
-    @return whether byte array creation completed
-
-    @note We can not reserve @a len bytes for the result, because @a len
-          may be too large. Usually, @ref unexpect_eof() detects the end of
-          the input before we run out of memory.
-    */
-    template<typename NumberType>
-    bool get_binary(const input_format_t format,
-                    const NumberType len,
-                    binary_t& result)
-    {
-        bool success = true;
-        for (NumberType i = 0; i < len; i++)
-        {
-            get();
-            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "binary")))
-            {
-                success = false;
-                break;
-            }
-            result.push_back(static_cast<std::uint8_t>(current));
-        }
-        return success;
-    }
-
-    /*!
-    @param[in] format   the current format (for diagnostics)
-    @param[in] context  further context information (for diagnostics)
-    @return whether the last read character is not EOF
-    */
-    JSON_HEDLEY_NON_NULL(3)
-    bool unexpect_eof(const input_format_t format, const char* context) const
-    {
-        if (JSON_HEDLEY_UNLIKELY(current == std::char_traits<char_type>::eof()))
-        {
-            return sax->parse_error(chars_read, "<end of file>",
-                                    parse_error::create(110, chars_read, exception_message(format, "unexpected end of input", context), nullptr));
-        }
-        return true;
-    }
-
-    /*!
-    @return a string representation of the last read byte
-    */
-    std::string get_token_string() const
-    {
-        std::array<char, 3> cr{{}};
-        static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(current))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-        return std::string{cr.data()};
-    }
-
-    /*!
-    @param[in] format   the current format
-    @param[in] detail   a detailed error message
-    @param[in] context  further context information
-    @return a message string to use in the parse_error exceptions
-    */
-    std::string exception_message(const input_format_t format,
-                                  const std::string& detail,
-                                  const std::string& context) const
-    {
-        std::string error_msg = "syntax error while parsing ";
-
-        switch (format)
-        {
-            case input_format_t::cbor:
-                error_msg += "CBOR";
-                break;
-
-            case input_format_t::msgpack:
-                error_msg += "MessagePack";
-                break;
-
-            case input_format_t::ubjson:
-                error_msg += "UBJSON";
-                break;
-
-            case input_format_t::bson:
-                error_msg += "BSON";
-                break;
-
-            case input_format_t::bjdata:
-                error_msg += "BJData";
-                break;
-
-            case input_format_t::json: // LCOV_EXCL_LINE
-            default:            // LCOV_EXCL_LINE
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        }
-
-        return concat(error_msg, ' ', context, ": ", detail);
-    }
-
-  private:
-    static JSON_INLINE_VARIABLE constexpr std::size_t npos = static_cast<std::size_t>(-1);
-
-    /// input adapter
-    InputAdapterType ia;
-
-    /// the current character
-    char_int_type current = std::char_traits<char_type>::eof();
-
-    /// the number of characters read
-    std::size_t chars_read = 0;
-
-    /// whether we can assume little endianness
-    const bool is_little_endian = little_endianness();
-
-    /// input format
-    const input_format_t input_format = input_format_t::json;
-
-    /// the SAX parser
-    json_sax_t* sax = nullptr;
-
-    // excluded markers in bjdata optimized type
-#define JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_ \
-    make_array<char_int_type>('F', 'H', 'N', 'S', 'T', 'Z', '[', '{')
-
-#define JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_ \
-    make_array<bjd_type>(                      \
-    bjd_type{'C', "char"},                     \
-    bjd_type{'D', "double"},                   \
-    bjd_type{'I', "int16"},                    \
-    bjd_type{'L', "int64"},                    \
-    bjd_type{'M', "uint64"},                   \
-    bjd_type{'U', "uint8"},                    \
-    bjd_type{'d', "single"},                   \
-    bjd_type{'i', "int8"},                     \
-    bjd_type{'l', "int32"},                    \
-    bjd_type{'m', "uint32"},                   \
-    bjd_type{'u', "uint16"})
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    // lookup tables
-    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
-    const decltype(JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_) bjd_optimized_type_markers =
-        JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_;
-
-    using bjd_type = std::pair<char_int_type, string_t>;
-    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
-    const decltype(JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_) bjd_types_map =
-        JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_;
-
-#undef JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_
-#undef JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_
-};
-
-#ifndef JSON_HAS_CPP_17
-    template<typename BasicJsonType, typename InputAdapterType, typename SAX>
-    constexpr std::size_t binary_reader<BasicJsonType, InputAdapterType, SAX>::npos;
-#endif
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/input/input_adapters.hpp>
-
-// #include <nlohmann/detail/input/lexer.hpp>
-
-// #include <nlohmann/detail/input/parser.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cmath> // isfinite
-#include <cstdint> // uint8_t
-#include <functional> // function
-#include <string> // string
-#include <utility> // move
-#include <vector> // vector
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/input/input_adapters.hpp>
-
-// #include <nlohmann/detail/input/json_sax.hpp>
-
-// #include <nlohmann/detail/input/lexer.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/is_sax.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-////////////
-// parser //
-////////////
-
-enum class parse_event_t : std::uint8_t
-{
-    /// the parser read `{` and started to process a JSON object
-    object_start,
-    /// the parser read `}` and finished processing a JSON object
-    object_end,
-    /// the parser read `[` and started to process a JSON array
-    array_start,
-    /// the parser read `]` and finished processing a JSON array
-    array_end,
-    /// the parser read a key of a value in an object
-    key,
-    /// the parser finished reading a JSON value
-    value
-};
-
-template<typename BasicJsonType>
-using parser_callback_t =
-    std::function<bool(int /*depth*/, parse_event_t /*event*/, BasicJsonType& /*parsed*/)>;
-
-/*!
-@brief syntax analysis
-
-This class implements a recursive descent parser.
-*/
-template<typename BasicJsonType, typename InputAdapterType>
-class parser
-{
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using string_t = typename BasicJsonType::string_t;
-    using lexer_t = lexer<BasicJsonType, InputAdapterType>;
-    using token_type = typename lexer_t::token_type;
-
-  public:
-    /// a parser reading from an input adapter
-    explicit parser(InputAdapterType&& adapter,
-                    const parser_callback_t<BasicJsonType> cb = nullptr,
-                    const bool allow_exceptions_ = true,
-                    const bool skip_comments = false)
-        : callback(cb)
-        , m_lexer(std::move(adapter), skip_comments)
-        , allow_exceptions(allow_exceptions_)
-    {
-        // read first token
-        get_token();
-    }
-
-    /*!
-    @brief public parser interface
-
-    @param[in] strict      whether to expect the last token to be EOF
-    @param[in,out] result  parsed JSON value
-
-    @throw parse_error.101 in case of an unexpected token
-    @throw parse_error.102 if to_unicode fails or surrogate error
-    @throw parse_error.103 if to_unicode fails
-    */
-    void parse(const bool strict, BasicJsonType& result)
-    {
-        if (callback)
-        {
-            json_sax_dom_callback_parser<BasicJsonType> sdp(result, callback, allow_exceptions);
-            sax_parse_internal(&sdp);
-
-            // in strict mode, input must be completely read
-            if (strict && (get_token() != token_type::end_of_input))
-            {
-                sdp.parse_error(m_lexer.get_position(),
-                                m_lexer.get_token_string(),
-                                parse_error::create(101, m_lexer.get_position(),
-                                                    exception_message(token_type::end_of_input, "value"), nullptr));
-            }
-
-            // in case of an error, return discarded value
-            if (sdp.is_errored())
-            {
-                result = value_t::discarded;
-                return;
-            }
-
-            // set top-level value to null if it was discarded by the callback
-            // function
-            if (result.is_discarded())
-            {
-                result = nullptr;
-            }
-        }
-        else
-        {
-            json_sax_dom_parser<BasicJsonType> sdp(result, allow_exceptions);
-            sax_parse_internal(&sdp);
-
-            // in strict mode, input must be completely read
-            if (strict && (get_token() != token_type::end_of_input))
-            {
-                sdp.parse_error(m_lexer.get_position(),
-                                m_lexer.get_token_string(),
-                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
-            }
-
-            // in case of an error, return discarded value
-            if (sdp.is_errored())
-            {
-                result = value_t::discarded;
-                return;
-            }
-        }
-
-        result.assert_invariant();
-    }
-
-    /*!
-    @brief public accept interface
-
-    @param[in] strict  whether to expect the last token to be EOF
-    @return whether the input is a proper JSON text
-    */
-    bool accept(const bool strict = true)
-    {
-        json_sax_acceptor<BasicJsonType> sax_acceptor;
-        return sax_parse(&sax_acceptor, strict);
-    }
-
-    template<typename SAX>
-    JSON_HEDLEY_NON_NULL(2)
-    bool sax_parse(SAX* sax, const bool strict = true)
-    {
-        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
-        const bool result = sax_parse_internal(sax);
-
-        // strict mode: next byte must be EOF
-        if (result && strict && (get_token() != token_type::end_of_input))
-        {
-            return sax->parse_error(m_lexer.get_position(),
-                                    m_lexer.get_token_string(),
-                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
-        }
-
-        return result;
-    }
-
-  private:
-    template<typename SAX>
-    JSON_HEDLEY_NON_NULL(2)
-    bool sax_parse_internal(SAX* sax)
-    {
-        // stack to remember the hierarchy of structured values we are parsing
-        // true = array; false = object
-        std::vector<bool> states;
-        // value to avoid a goto (see comment where set to true)
-        bool skip_to_state_evaluation = false;
-
-        while (true)
-        {
-            if (!skip_to_state_evaluation)
-            {
-                // invariant: get_token() was called before each iteration
-                switch (last_token)
-                {
-                    case token_type::begin_object:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
-                        {
-                            return false;
-                        }
-
-                        // closing } -> we are done
-                        if (get_token() == token_type::end_object)
-                        {
-                            if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
-                            {
-                                return false;
-                            }
-                            break;
-                        }
-
-                        // parse key
-                        if (JSON_HEDLEY_UNLIKELY(last_token != token_type::value_string))
-                        {
-                            return sax->parse_error(m_lexer.get_position(),
-                                                    m_lexer.get_token_string(),
-                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
-                        }
-                        if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
-                        {
-                            return false;
-                        }
-
-                        // parse separator (:)
-                        if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
-                        {
-                            return sax->parse_error(m_lexer.get_position(),
-                                                    m_lexer.get_token_string(),
-                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
-                        }
-
-                        // remember we are now inside an object
-                        states.push_back(false);
-
-                        // parse values
-                        get_token();
-                        continue;
-                    }
-
-                    case token_type::begin_array:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
-                        {
-                            return false;
-                        }
-
-                        // closing ] -> we are done
-                        if (get_token() == token_type::end_array)
-                        {
-                            if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
-                            {
-                                return false;
-                            }
-                            break;
-                        }
-
-                        // remember we are now inside an array
-                        states.push_back(true);
-
-                        // parse values (no need to call get_token)
-                        continue;
-                    }
-
-                    case token_type::value_float:
-                    {
-                        const auto res = m_lexer.get_number_float();
-
-                        if (JSON_HEDLEY_UNLIKELY(!std::isfinite(res)))
-                        {
-                            return sax->parse_error(m_lexer.get_position(),
-                                                    m_lexer.get_token_string(),
-                                                    out_of_range::create(406, concat("number overflow parsing '", m_lexer.get_token_string(), '\''), nullptr));
-                        }
-
-                        if (JSON_HEDLEY_UNLIKELY(!sax->number_float(res, m_lexer.get_string())))
-                        {
-                            return false;
-                        }
-
-                        break;
-                    }
-
-                    case token_type::literal_false:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(false)))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::literal_null:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->null()))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::literal_true:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(true)))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::value_integer:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->number_integer(m_lexer.get_number_integer())))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::value_string:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->string(m_lexer.get_string())))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::value_unsigned:
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(m_lexer.get_number_unsigned())))
-                        {
-                            return false;
-                        }
-                        break;
-                    }
-
-                    case token_type::parse_error:
-                    {
-                        // using "uninitialized" to avoid "expected" message
-                        return sax->parse_error(m_lexer.get_position(),
-                                                m_lexer.get_token_string(),
-                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::uninitialized, "value"), nullptr));
-                    }
-
-                    case token_type::uninitialized:
-                    case token_type::end_array:
-                    case token_type::end_object:
-                    case token_type::name_separator:
-                    case token_type::value_separator:
-                    case token_type::end_of_input:
-                    case token_type::literal_or_value:
-                    default: // the last token was unexpected
-                    {
-                        return sax->parse_error(m_lexer.get_position(),
-                                                m_lexer.get_token_string(),
-                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), nullptr));
-                    }
-                }
-            }
-            else
-            {
-                skip_to_state_evaluation = false;
-            }
-
-            // we reached this line after we successfully parsed a value
-            if (states.empty())
-            {
-                // empty stack: we reached the end of the hierarchy: done
-                return true;
-            }
-
-            if (states.back())  // array
-            {
-                // comma -> next value
-                if (get_token() == token_type::value_separator)
-                {
-                    // parse a new value
-                    get_token();
-                    continue;
-                }
-
-                // closing ]
-                if (JSON_HEDLEY_LIKELY(last_token == token_type::end_array))
-                {
-                    if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
-                    {
-                        return false;
-                    }
-
-                    // We are done with this array. Before we can parse a
-                    // new value, we need to evaluate the new state first.
-                    // By setting skip_to_state_evaluation to false, we
-                    // are effectively jumping to the beginning of this if.
-                    JSON_ASSERT(!states.empty());
-                    states.pop_back();
-                    skip_to_state_evaluation = true;
-                    continue;
-                }
-
-                return sax->parse_error(m_lexer.get_position(),
-                                        m_lexer.get_token_string(),
-                                        parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_array, "array"), nullptr));
-            }
-
-            // states.back() is false -> object
-
-            // comma -> next value
-            if (get_token() == token_type::value_separator)
-            {
-                // parse key
-                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::value_string))
-                {
-                    return sax->parse_error(m_lexer.get_position(),
-                                            m_lexer.get_token_string(),
-                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
-                }
-
-                if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
-                {
-                    return false;
-                }
-
-                // parse separator (:)
-                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
-                {
-                    return sax->parse_error(m_lexer.get_position(),
-                                            m_lexer.get_token_string(),
-                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
-                }
-
-                // parse values
-                get_token();
-                continue;
-            }
-
-            // closing }
-            if (JSON_HEDLEY_LIKELY(last_token == token_type::end_object))
-            {
-                if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
-                {
-                    return false;
-                }
-
-                // We are done with this object. Before we can parse a
-                // new value, we need to evaluate the new state first.
-                // By setting skip_to_state_evaluation to false, we
-                // are effectively jumping to the beginning of this if.
-                JSON_ASSERT(!states.empty());
-                states.pop_back();
-                skip_to_state_evaluation = true;
-                continue;
-            }
-
-            return sax->parse_error(m_lexer.get_position(),
-                                    m_lexer.get_token_string(),
-                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_object, "object"), nullptr));
-        }
-    }
-
-    /// get next token from lexer
-    token_type get_token()
-    {
-        return last_token = m_lexer.scan();
-    }
-
-    std::string exception_message(const token_type expected, const std::string& context)
-    {
-        std::string error_msg = "syntax error ";
-
-        if (!context.empty())
-        {
-            error_msg += concat("while parsing ", context, ' ');
-        }
-
-        error_msg += "- ";
-
-        if (last_token == token_type::parse_error)
-        {
-            error_msg += concat(m_lexer.get_error_message(), "; last read: '",
-                                m_lexer.get_token_string(), '\'');
-        }
-        else
-        {
-            error_msg += concat("unexpected ", lexer_t::token_type_name(last_token));
-        }
-
-        if (expected != token_type::uninitialized)
-        {
-            error_msg += concat("; expected ", lexer_t::token_type_name(expected));
-        }
-
-        return error_msg;
-    }
-
-  private:
-    /// callback function
-    const parser_callback_t<BasicJsonType> callback = nullptr;
-    /// the type of the last read token
-    token_type last_token = token_type::uninitialized;
-    /// the lexer
-    lexer_t m_lexer;
-    /// whether to throw exceptions in case of errors
-    const bool allow_exceptions = true;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/iterators/internal_iterator.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef> // ptrdiff_t
-#include <limits>  // numeric_limits
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/*
-@brief an iterator for primitive JSON types
-
-This class models an iterator for primitive JSON types (boolean, number,
-string). It's only purpose is to allow the iterator/const_iterator classes
-to "iterate" over primitive values. Internally, the iterator is modeled by
-a `difference_type` variable. Value begin_value (`0`) models the begin,
-end_value (`1`) models past the end.
-*/
-class primitive_iterator_t
-{
-  private:
-    using difference_type = std::ptrdiff_t;
-    static constexpr difference_type begin_value = 0;
-    static constexpr difference_type end_value = begin_value + 1;
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    /// iterator as signed integer type
-    difference_type m_it = (std::numeric_limits<std::ptrdiff_t>::min)();
-
-  public:
-    constexpr difference_type get_value() const noexcept
-    {
-        return m_it;
-    }
-
-    /// set iterator to a defined beginning
-    void set_begin() noexcept
-    {
-        m_it = begin_value;
-    }
-
-    /// set iterator to a defined past the end
-    void set_end() noexcept
-    {
-        m_it = end_value;
-    }
-
-    /// return whether the iterator can be dereferenced
-    constexpr bool is_begin() const noexcept
-    {
-        return m_it == begin_value;
-    }
-
-    /// return whether the iterator is at end
-    constexpr bool is_end() const noexcept
-    {
-        return m_it == end_value;
-    }
-
-    friend constexpr bool operator==(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
-    {
-        return lhs.m_it == rhs.m_it;
-    }
-
-    friend constexpr bool operator<(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
-    {
-        return lhs.m_it < rhs.m_it;
-    }
-
-    primitive_iterator_t operator+(difference_type n) noexcept
-    {
-        auto result = *this;
-        result += n;
-        return result;
-    }
-
-    friend constexpr difference_type operator-(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
-    {
-        return lhs.m_it - rhs.m_it;
-    }
-
-    primitive_iterator_t& operator++() noexcept
-    {
-        ++m_it;
-        return *this;
-    }
-
-    primitive_iterator_t operator++(int)& noexcept // NOLINT(cert-dcl21-cpp)
-    {
-        auto result = *this;
-        ++m_it;
-        return result;
-    }
-
-    primitive_iterator_t& operator--() noexcept
-    {
-        --m_it;
-        return *this;
-    }
-
-    primitive_iterator_t operator--(int)& noexcept // NOLINT(cert-dcl21-cpp)
-    {
-        auto result = *this;
-        --m_it;
-        return result;
-    }
-
-    primitive_iterator_t& operator+=(difference_type n) noexcept
-    {
-        m_it += n;
-        return *this;
-    }
-
-    primitive_iterator_t& operator-=(difference_type n) noexcept
-    {
-        m_it -= n;
-        return *this;
-    }
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/*!
-@brief an iterator value
-
-@note This structure could easily be a union, but MSVC currently does not allow
-unions members with complex constructors, see https://github.com/nlohmann/json/pull/105.
-*/
-template<typename BasicJsonType> struct internal_iterator
-{
-    /// iterator for JSON objects
-    typename BasicJsonType::object_t::iterator object_iterator {};
-    /// iterator for JSON arrays
-    typename BasicJsonType::array_t::iterator array_iterator {};
-    /// generic iterator for all other types
-    primitive_iterator_t primitive_iterator {};
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/iterators/iter_impl.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <iterator> // iterator, random_access_iterator_tag, bidirectional_iterator_tag, advance, next
-#include <type_traits> // conditional, is_const, remove_const
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/iterators/internal_iterator.hpp>
-
-// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-// forward declare, to be able to friend it later on
-template<typename IteratorType> class iteration_proxy;
-template<typename IteratorType> class iteration_proxy_value;
-
-/*!
-@brief a template for a bidirectional iterator for the @ref basic_json class
-This class implements a both iterators (iterator and const_iterator) for the
-@ref basic_json class.
-@note An iterator is called *initialized* when a pointer to a JSON value has
-      been set (e.g., by a constructor or a copy assignment). If the iterator is
-      default-constructed, it is *uninitialized* and most methods are undefined.
-      **The library uses assertions to detect calls on uninitialized iterators.**
-@requirement The class satisfies the following concept requirements:
--
-[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
-  The iterator that can be moved can be moved in both directions (i.e.
-  incremented and decremented).
-@since version 1.0.0, simplified in version 2.0.9, change to bidirectional
-       iterators in version 3.0.0 (see https://github.com/nlohmann/json/issues/593)
-*/
-template<typename BasicJsonType>
-class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
-{
-    /// the iterator with BasicJsonType of different const-ness
-    using other_iter_impl = iter_impl<typename std::conditional<std::is_const<BasicJsonType>::value, typename std::remove_const<BasicJsonType>::type, const BasicJsonType>::type>;
-    /// allow basic_json to access private members
-    friend other_iter_impl;
-    friend BasicJsonType;
-    friend iteration_proxy<iter_impl>;
-    friend iteration_proxy_value<iter_impl>;
-
-    using object_t = typename BasicJsonType::object_t;
-    using array_t = typename BasicJsonType::array_t;
-    // make sure BasicJsonType is basic_json or const basic_json
-    static_assert(is_basic_json<typename std::remove_const<BasicJsonType>::type>::value,
-                  "iter_impl only accepts (const) basic_json");
-    // superficial check for the LegacyBidirectionalIterator named requirement
-    static_assert(std::is_base_of<std::bidirectional_iterator_tag, std::bidirectional_iterator_tag>::value
-                  &&  std::is_base_of<std::bidirectional_iterator_tag, typename std::iterator_traits<typename array_t::iterator>::iterator_category>::value,
-                  "basic_json iterator assumes array and object type iterators satisfy the LegacyBidirectionalIterator named requirement.");
-
-  public:
-    /// The std::iterator class template (used as a base class to provide typedefs) is deprecated in C++17.
-    /// The C++ Standard has never required user-defined iterators to derive from std::iterator.
-    /// A user-defined iterator should provide publicly accessible typedefs named
-    /// iterator_category, value_type, difference_type, pointer, and reference.
-    /// Note that value_type is required to be non-const, even for constant iterators.
-    using iterator_category = std::bidirectional_iterator_tag;
-
-    /// the type of the values when the iterator is dereferenced
-    using value_type = typename BasicJsonType::value_type;
-    /// a type to represent differences between iterators
-    using difference_type = typename BasicJsonType::difference_type;
-    /// defines a pointer to the type iterated over (value_type)
-    using pointer = typename std::conditional<std::is_const<BasicJsonType>::value,
-          typename BasicJsonType::const_pointer,
-          typename BasicJsonType::pointer>::type;
-    /// defines a reference to the type iterated over (value_type)
-    using reference =
-        typename std::conditional<std::is_const<BasicJsonType>::value,
-        typename BasicJsonType::const_reference,
-        typename BasicJsonType::reference>::type;
-
-    iter_impl() = default;
-    ~iter_impl() = default;
-    iter_impl(iter_impl&&) noexcept = default;
-    iter_impl& operator=(iter_impl&&) noexcept = default;
-
-    /*!
-    @brief constructor for a given JSON instance
-    @param[in] object  pointer to a JSON object for this iterator
-    @pre object != nullptr
-    @post The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    explicit iter_impl(pointer object) noexcept : m_object(object)
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                m_it.object_iterator = typename object_t::iterator();
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_it.array_iterator = typename array_t::iterator();
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                m_it.primitive_iterator = primitive_iterator_t();
-                break;
-            }
-        }
-    }
-
-    /*!
-    @note The conventional copy constructor and copy assignment are implicitly
-          defined. Combined with the following converting constructor and
-          assignment, they support: (1) copy from iterator to iterator, (2)
-          copy from const iterator to const iterator, and (3) conversion from
-          iterator to const iterator. However conversion from const iterator
-          to iterator is not defined.
-    */
-
-    /*!
-    @brief const copy constructor
-    @param[in] other const iterator to copy from
-    @note This copy constructor had to be defined explicitly to circumvent a bug
-          occurring on msvc v19.0 compiler (VS 2015) debug build. For more
-          information refer to: https://github.com/nlohmann/json/issues/1608
-    */
-    iter_impl(const iter_impl<const BasicJsonType>& other) noexcept
-        : m_object(other.m_object), m_it(other.m_it)
-    {}
-
-    /*!
-    @brief converting assignment
-    @param[in] other const iterator to copy from
-    @return const/non-const iterator
-    @note It is not checked whether @a other is initialized.
-    */
-    iter_impl& operator=(const iter_impl<const BasicJsonType>& other) noexcept
-    {
-        if (&other != this)
-        {
-            m_object = other.m_object;
-            m_it = other.m_it;
-        }
-        return *this;
-    }
-
-    /*!
-    @brief converting constructor
-    @param[in] other  non-const iterator to copy from
-    @note It is not checked whether @a other is initialized.
-    */
-    iter_impl(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept
-        : m_object(other.m_object), m_it(other.m_it)
-    {}
-
-    /*!
-    @brief converting assignment
-    @param[in] other  non-const iterator to copy from
-    @return const/non-const iterator
-    @note It is not checked whether @a other is initialized.
-    */
-    iter_impl& operator=(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept // NOLINT(cert-oop54-cpp)
-    {
-        m_object = other.m_object;
-        m_it = other.m_it;
-        return *this;
-    }
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    /*!
-    @brief set the iterator to the first value
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    void set_begin() noexcept
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                m_it.object_iterator = m_object->m_value.object->begin();
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_it.array_iterator = m_object->m_value.array->begin();
-                break;
-            }
-
-            case value_t::null:
-            {
-                // set to end so begin()==end() is true: null is empty
-                m_it.primitive_iterator.set_end();
-                break;
-            }
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                m_it.primitive_iterator.set_begin();
-                break;
-            }
-        }
-    }
-
-    /*!
-    @brief set the iterator past the last value
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    void set_end() noexcept
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                m_it.object_iterator = m_object->m_value.object->end();
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_it.array_iterator = m_object->m_value.array->end();
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                m_it.primitive_iterator.set_end();
-                break;
-            }
-        }
-    }
-
-  public:
-    /*!
-    @brief return a reference to the value pointed to by the iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    reference operator*() const
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
-                return m_it.object_iterator->second;
-            }
-
-            case value_t::array:
-            {
-                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
-                return *m_it.array_iterator;
-            }
-
-            case value_t::null:
-                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
-                {
-                    return *m_object;
-                }
-
-                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
-            }
-        }
-    }
-
-    /*!
-    @brief dereference the iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    pointer operator->() const
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
-                return &(m_it.object_iterator->second);
-            }
-
-            case value_t::array:
-            {
-                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
-                return &*m_it.array_iterator;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
-                {
-                    return m_object;
-                }
-
-                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
-            }
-        }
-    }
-
-    /*!
-    @brief post-increment (it++)
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl operator++(int)& // NOLINT(cert-dcl21-cpp)
-    {
-        auto result = *this;
-        ++(*this);
-        return result;
-    }
-
-    /*!
-    @brief pre-increment (++it)
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl& operator++()
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                std::advance(m_it.object_iterator, 1);
-                break;
-            }
-
-            case value_t::array:
-            {
-                std::advance(m_it.array_iterator, 1);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                ++m_it.primitive_iterator;
-                break;
-            }
-        }
-
-        return *this;
-    }
-
-    /*!
-    @brief post-decrement (it--)
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl operator--(int)& // NOLINT(cert-dcl21-cpp)
-    {
-        auto result = *this;
-        --(*this);
-        return result;
-    }
-
-    /*!
-    @brief pre-decrement (--it)
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl& operator--()
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-            {
-                std::advance(m_it.object_iterator, -1);
-                break;
-            }
-
-            case value_t::array:
-            {
-                std::advance(m_it.array_iterator, -1);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                --m_it.primitive_iterator;
-                break;
-            }
-        }
-
-        return *this;
-    }
-
-    /*!
-    @brief comparison: equal
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
-    bool operator==(const IterImpl& other) const
-    {
-        // if objects are not the same, the comparison is undefined
-        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
-        }
-
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-                return (m_it.object_iterator == other.m_it.object_iterator);
-
-            case value_t::array:
-                return (m_it.array_iterator == other.m_it.array_iterator);
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                return (m_it.primitive_iterator == other.m_it.primitive_iterator);
-        }
-    }
-
-    /*!
-    @brief comparison: not equal
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
-    bool operator!=(const IterImpl& other) const
-    {
-        return !operator==(other);
-    }
-
-    /*!
-    @brief comparison: smaller
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    bool operator<(const iter_impl& other) const
-    {
-        // if objects are not the same, the comparison is undefined
-        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
-        }
-
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-                JSON_THROW(invalid_iterator::create(213, "cannot compare order of object iterators", m_object));
-
-            case value_t::array:
-                return (m_it.array_iterator < other.m_it.array_iterator);
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                return (m_it.primitive_iterator < other.m_it.primitive_iterator);
-        }
-    }
-
-    /*!
-    @brief comparison: less than or equal
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    bool operator<=(const iter_impl& other) const
-    {
-        return !other.operator < (*this);
-    }
-
-    /*!
-    @brief comparison: greater than
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    bool operator>(const iter_impl& other) const
-    {
-        return !operator<=(other);
-    }
-
-    /*!
-    @brief comparison: greater than or equal
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    bool operator>=(const iter_impl& other) const
-    {
-        return !operator<(other);
-    }
-
-    /*!
-    @brief add to iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl& operator+=(difference_type i)
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
-
-            case value_t::array:
-            {
-                std::advance(m_it.array_iterator, i);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                m_it.primitive_iterator += i;
-                break;
-            }
-        }
-
-        return *this;
-    }
-
-    /*!
-    @brief subtract from iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl& operator-=(difference_type i)
-    {
-        return operator+=(-i);
-    }
-
-    /*!
-    @brief add to iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl operator+(difference_type i) const
-    {
-        auto result = *this;
-        result += i;
-        return result;
-    }
-
-    /*!
-    @brief addition of distance and iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    friend iter_impl operator+(difference_type i, const iter_impl& it)
-    {
-        auto result = it;
-        result += i;
-        return result;
-    }
-
-    /*!
-    @brief subtract from iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    iter_impl operator-(difference_type i) const
-    {
-        auto result = *this;
-        result -= i;
-        return result;
-    }
-
-    /*!
-    @brief return difference
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    difference_type operator-(const iter_impl& other) const
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
-
-            case value_t::array:
-                return m_it.array_iterator - other.m_it.array_iterator;
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                return m_it.primitive_iterator - other.m_it.primitive_iterator;
-        }
-    }
-
-    /*!
-    @brief access to successor
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    reference operator[](difference_type n) const
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        switch (m_object->m_type)
-        {
-            case value_t::object:
-                JSON_THROW(invalid_iterator::create(208, "cannot use operator[] for object iterators", m_object));
-
-            case value_t::array:
-                return *std::next(m_it.array_iterator, n);
-
-            case value_t::null:
-                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.get_value() == -n))
-                {
-                    return *m_object;
-                }
-
-                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
-            }
-        }
-    }
-
-    /*!
-    @brief return the key of an object iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    const typename object_t::key_type& key() const
-    {
-        JSON_ASSERT(m_object != nullptr);
-
-        if (JSON_HEDLEY_LIKELY(m_object->is_object()))
-        {
-            return m_it.object_iterator->first;
-        }
-
-        JSON_THROW(invalid_iterator::create(207, "cannot use key() for non-object iterators", m_object));
-    }
-
-    /*!
-    @brief return the value of an iterator
-    @pre The iterator is initialized; i.e. `m_object != nullptr`.
-    */
-    reference value() const
-    {
-        return operator*();
-    }
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    /// associated JSON instance
-    pointer m_object = nullptr;
-    /// the actual iterator of the associated instance
-    internal_iterator<typename std::remove_const<BasicJsonType>::type> m_it {};
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
-
-// #include <nlohmann/detail/iterators/json_reverse_iterator.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <cstddef> // ptrdiff_t
-#include <iterator> // reverse_iterator
-#include <utility> // declval
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-//////////////////////
-// reverse_iterator //
-//////////////////////
-
-/*!
-@brief a template for a reverse iterator class
-
-@tparam Base the base iterator type to reverse. Valid types are @ref
-iterator (to create @ref reverse_iterator) and @ref const_iterator (to
-create @ref const_reverse_iterator).
-
-@requirement The class satisfies the following concept requirements:
--
-[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
-  The iterator that can be moved can be moved in both directions (i.e.
-  incremented and decremented).
-- [OutputIterator](https://en.cppreference.com/w/cpp/named_req/OutputIterator):
-  It is possible to write to the pointed-to element (only if @a Base is
-  @ref iterator).
-
-@since version 1.0.0
-*/
-template<typename Base>
-class json_reverse_iterator : public std::reverse_iterator<Base>
-{
-  public:
-    using difference_type = std::ptrdiff_t;
-    /// shortcut to the reverse iterator adapter
-    using base_iterator = std::reverse_iterator<Base>;
-    /// the reference type for the pointed-to element
-    using reference = typename Base::reference;
-
-    /// create reverse iterator from iterator
-    explicit json_reverse_iterator(const typename base_iterator::iterator_type& it) noexcept
-        : base_iterator(it) {}
-
-    /// create reverse iterator from base class
-    explicit json_reverse_iterator(const base_iterator& it) noexcept : base_iterator(it) {}
-
-    /// post-increment (it++)
-    json_reverse_iterator operator++(int)& // NOLINT(cert-dcl21-cpp)
-    {
-        return static_cast<json_reverse_iterator>(base_iterator::operator++(1));
-    }
-
-    /// pre-increment (++it)
-    json_reverse_iterator& operator++()
-    {
-        return static_cast<json_reverse_iterator&>(base_iterator::operator++());
-    }
-
-    /// post-decrement (it--)
-    json_reverse_iterator operator--(int)& // NOLINT(cert-dcl21-cpp)
-    {
-        return static_cast<json_reverse_iterator>(base_iterator::operator--(1));
-    }
-
-    /// pre-decrement (--it)
-    json_reverse_iterator& operator--()
-    {
-        return static_cast<json_reverse_iterator&>(base_iterator::operator--());
-    }
-
-    /// add to iterator
-    json_reverse_iterator& operator+=(difference_type i)
-    {
-        return static_cast<json_reverse_iterator&>(base_iterator::operator+=(i));
-    }
-
-    /// add to iterator
-    json_reverse_iterator operator+(difference_type i) const
-    {
-        return static_cast<json_reverse_iterator>(base_iterator::operator+(i));
-    }
-
-    /// subtract from iterator
-    json_reverse_iterator operator-(difference_type i) const
-    {
-        return static_cast<json_reverse_iterator>(base_iterator::operator-(i));
-    }
-
-    /// return difference
-    difference_type operator-(const json_reverse_iterator& other) const
-    {
-        return base_iterator(*this) - base_iterator(other);
-    }
-
-    /// access to successor
-    reference operator[](difference_type n) const
-    {
-        return *(this->operator+(n));
-    }
-
-    /// return the key of an object iterator
-    auto key() const -> decltype(std::declval<Base>().key())
-    {
-        auto it = --this->base();
-        return it.key();
-    }
-
-    /// return the value of an iterator
-    reference value() const
-    {
-        auto it = --this->base();
-        return it.operator * ();
-    }
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
-
-// #include <nlohmann/detail/json_pointer.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // all_of
-#include <cctype> // isdigit
-#include <cerrno> // errno, ERANGE
-#include <cstdlib> // strtoull
-#ifndef JSON_NO_IO
-    #include <iosfwd> // ostream
-#endif  // JSON_NO_IO
-#include <limits> // max
-#include <numeric> // accumulate
-#include <string> // string
-#include <utility> // move
-#include <vector> // vector
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/string_escape.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
-/// @sa https://json.nlohmann.me/api/json_pointer/
-template<typename RefStringType>
-class json_pointer
-{
-    // allow basic_json to access private members
-    NLOHMANN_BASIC_JSON_TPL_DECLARATION
-    friend class basic_json;
-
-    template<typename>
-    friend class json_pointer;
-
-    template<typename T>
-    struct string_t_helper
-    {
-        using type = T;
-    };
-
-    NLOHMANN_BASIC_JSON_TPL_DECLARATION
-    struct string_t_helper<NLOHMANN_BASIC_JSON_TPL>
-    {
-        using type = StringType;
-    };
-
-  public:
-    // for backwards compatibility accept BasicJsonType
-    using string_t = typename string_t_helper<RefStringType>::type;
-
-    /// @brief create JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/json_pointer/
-    explicit json_pointer(const string_t& s = "")
-        : reference_tokens(split(s))
-    {}
-
-    /// @brief return a string representation of the JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/to_string/
-    string_t to_string() const
-    {
-        return std::accumulate(reference_tokens.begin(), reference_tokens.end(),
-                               string_t{},
-                               [](const string_t& a, const string_t& b)
-        {
-            return detail::concat(a, '/', detail::escape(b));
-        });
-    }
-
-    /// @brief return a string representation of the JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_string/
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, to_string())
-    operator string_t() const
-    {
-        return to_string();
-    }
-
-#ifndef JSON_NO_IO
-    /// @brief write string representation of the JSON pointer to stream
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
-    friend std::ostream& operator<<(std::ostream& o, const json_pointer& ptr)
-    {
-        o << ptr.to_string();
-        return o;
-    }
-#endif
-
-    /// @brief append another JSON pointer at the end of this JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
-    json_pointer& operator/=(const json_pointer& ptr)
-    {
-        reference_tokens.insert(reference_tokens.end(),
-                                ptr.reference_tokens.begin(),
-                                ptr.reference_tokens.end());
-        return *this;
-    }
-
-    /// @brief append an unescaped reference token at the end of this JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
-    json_pointer& operator/=(string_t token)
-    {
-        push_back(std::move(token));
-        return *this;
-    }
-
-    /// @brief append an array index at the end of this JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
-    json_pointer& operator/=(std::size_t array_idx)
-    {
-        return *this /= std::to_string(array_idx);
-    }
-
-    /// @brief create a new JSON pointer by appending the right JSON pointer at the end of the left JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
-    friend json_pointer operator/(const json_pointer& lhs,
-                                  const json_pointer& rhs)
-    {
-        return json_pointer(lhs) /= rhs;
-    }
-
-    /// @brief create a new JSON pointer by appending the unescaped token at the end of the JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
-    friend json_pointer operator/(const json_pointer& lhs, string_t token) // NOLINT(performance-unnecessary-value-param)
-    {
-        return json_pointer(lhs) /= std::move(token);
-    }
-
-    /// @brief create a new JSON pointer by appending the array-index-token at the end of the JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
-    friend json_pointer operator/(const json_pointer& lhs, std::size_t array_idx)
-    {
-        return json_pointer(lhs) /= array_idx;
-    }
-
-    /// @brief returns the parent of this JSON pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/parent_pointer/
-    json_pointer parent_pointer() const
-    {
-        if (empty())
-        {
-            return *this;
-        }
-
-        json_pointer res = *this;
-        res.pop_back();
-        return res;
-    }
-
-    /// @brief remove last reference token
-    /// @sa https://json.nlohmann.me/api/json_pointer/pop_back/
-    void pop_back()
-    {
-        if (JSON_HEDLEY_UNLIKELY(empty()))
-        {
-            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
-        }
-
-        reference_tokens.pop_back();
-    }
-
-    /// @brief return last reference token
-    /// @sa https://json.nlohmann.me/api/json_pointer/back/
-    const string_t& back() const
-    {
-        if (JSON_HEDLEY_UNLIKELY(empty()))
-        {
-            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
-        }
-
-        return reference_tokens.back();
-    }
-
-    /// @brief append an unescaped token at the end of the reference pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
-    void push_back(const string_t& token)
-    {
-        reference_tokens.push_back(token);
-    }
-
-    /// @brief append an unescaped token at the end of the reference pointer
-    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
-    void push_back(string_t&& token)
-    {
-        reference_tokens.push_back(std::move(token));
-    }
-
-    /// @brief return whether pointer points to the root document
-    /// @sa https://json.nlohmann.me/api/json_pointer/empty/
-    bool empty() const noexcept
-    {
-        return reference_tokens.empty();
-    }
-
-  private:
-    /*!
-    @param[in] s  reference token to be converted into an array index
-
-    @return integer representation of @a s
-
-    @throw parse_error.106  if an array index begins with '0'
-    @throw parse_error.109  if an array index begins not with a digit
-    @throw out_of_range.404 if string @a s could not be converted to an integer
-    @throw out_of_range.410 if an array index exceeds size_type
-    */
-    template<typename BasicJsonType>
-    static typename BasicJsonType::size_type array_index(const string_t& s)
-    {
-        using size_type = typename BasicJsonType::size_type;
-
-        // error condition (cf. RFC 6901, Sect. 4)
-        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && s[0] == '0'))
-        {
-            JSON_THROW(detail::parse_error::create(106, 0, detail::concat("array index '", s, "' must not begin with '0'"), nullptr));
-        }
-
-        // error condition (cf. RFC 6901, Sect. 4)
-        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && !(s[0] >= '1' && s[0] <= '9')))
-        {
-            JSON_THROW(detail::parse_error::create(109, 0, detail::concat("array index '", s, "' is not a number"), nullptr));
-        }
-
-        const char* p = s.c_str();
-        char* p_end = nullptr;
-        errno = 0; // strtoull doesn't reset errno
-        unsigned long long res = std::strtoull(p, &p_end, 10); // NOLINT(runtime/int)
-        if (p == p_end // invalid input or empty string
-                || errno == ERANGE // out of range
-                || JSON_HEDLEY_UNLIKELY(static_cast<std::size_t>(p_end - p) != s.size())) // incomplete read
-        {
-            JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", s, "'"), nullptr));
-        }
-
-        // only triggered on special platforms (like 32bit), see also
-        // https://github.com/nlohmann/json/pull/2203
-        if (res >= static_cast<unsigned long long>((std::numeric_limits<size_type>::max)()))  // NOLINT(runtime/int)
-        {
-            JSON_THROW(detail::out_of_range::create(410, detail::concat("array index ", s, " exceeds size_type"), nullptr));   // LCOV_EXCL_LINE
-        }
-
-        return static_cast<size_type>(res);
-    }
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    json_pointer top() const
-    {
-        if (JSON_HEDLEY_UNLIKELY(empty()))
-        {
-            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
-        }
-
-        json_pointer result = *this;
-        result.reference_tokens = {reference_tokens[0]};
-        return result;
-    }
-
-  private:
-    /*!
-    @brief create and return a reference to the pointed to value
-
-    @complexity Linear in the number of reference tokens.
-
-    @throw parse_error.109 if array index is not a number
-    @throw type_error.313 if value cannot be unflattened
-    */
-    template<typename BasicJsonType>
-    BasicJsonType& get_and_create(BasicJsonType& j) const
-    {
-        auto* result = &j;
-
-        // in case no reference tokens exist, return a reference to the JSON value
-        // j which will be overwritten by a primitive value
-        for (const auto& reference_token : reference_tokens)
-        {
-            switch (result->type())
-            {
-                case detail::value_t::null:
-                {
-                    if (reference_token == "0")
-                    {
-                        // start a new array if reference token is 0
-                        result = &result->operator[](0);
-                    }
-                    else
-                    {
-                        // start a new object otherwise
-                        result = &result->operator[](reference_token);
-                    }
-                    break;
-                }
-
-                case detail::value_t::object:
-                {
-                    // create an entry in the object
-                    result = &result->operator[](reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    // create an entry in the array
-                    result = &result->operator[](array_index<BasicJsonType>(reference_token));
-                    break;
-                }
-
-                /*
-                The following code is only reached if there exists a reference
-                token _and_ the current value is primitive. In this case, we have
-                an error situation, because primitive values may only occur as
-                single value; that is, with an empty list of reference tokens.
-                */
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                    JSON_THROW(detail::type_error::create(313, "invalid value to unflatten", &j));
-            }
-        }
-
-        return *result;
-    }
-
-    /*!
-    @brief return a reference to the pointed to value
-
-    @note This version does not throw if a value is not present, but tries to
-          create nested values instead. For instance, calling this function
-          with pointer `"/this/that"` on a null value is equivalent to calling
-          `operator[]("this").operator[]("that")` on that value, effectively
-          changing the null value to an object.
-
-    @param[in] ptr  a JSON value
-
-    @return reference to the JSON value pointed to by the JSON pointer
-
-    @complexity Linear in the length of the JSON pointer.
-
-    @throw parse_error.106   if an array index begins with '0'
-    @throw parse_error.109   if an array index was not a number
-    @throw out_of_range.404  if the JSON pointer can not be resolved
-    */
-    template<typename BasicJsonType>
-    BasicJsonType& get_unchecked(BasicJsonType* ptr) const
-    {
-        for (const auto& reference_token : reference_tokens)
-        {
-            // convert null values to arrays or objects before continuing
-            if (ptr->is_null())
-            {
-                // check if reference token is a number
-                const bool nums =
-                    std::all_of(reference_token.begin(), reference_token.end(),
-                                [](const unsigned char x)
-                {
-                    return std::isdigit(x);
-                });
-
-                // change value to array for numbers or "-" or to object otherwise
-                *ptr = (nums || reference_token == "-")
-                       ? detail::value_t::array
-                       : detail::value_t::object;
-            }
-
-            switch (ptr->type())
-            {
-                case detail::value_t::object:
-                {
-                    // use unchecked object access
-                    ptr = &ptr->operator[](reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    if (reference_token == "-")
-                    {
-                        // explicitly treat "-" as index beyond the end
-                        ptr = &ptr->operator[](ptr->m_value.array->size());
-                    }
-                    else
-                    {
-                        // convert array index to number; unchecked access
-                        ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
-                    }
-                    break;
-                }
-
-                case detail::value_t::null:
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
-            }
-        }
-
-        return *ptr;
-    }
-
-    /*!
-    @throw parse_error.106   if an array index begins with '0'
-    @throw parse_error.109   if an array index was not a number
-    @throw out_of_range.402  if the array index '-' is used
-    @throw out_of_range.404  if the JSON pointer can not be resolved
-    */
-    template<typename BasicJsonType>
-    BasicJsonType& get_checked(BasicJsonType* ptr) const
-    {
-        for (const auto& reference_token : reference_tokens)
-        {
-            switch (ptr->type())
-            {
-                case detail::value_t::object:
-                {
-                    // note: at performs range check
-                    ptr = &ptr->at(reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
-                    {
-                        // "-" always fails the range check
-                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
-                                "array index '-' (", std::to_string(ptr->m_value.array->size()),
-                                ") is out of range"), ptr));
-                    }
-
-                    // note: at performs range check
-                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
-                    break;
-                }
-
-                case detail::value_t::null:
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
-            }
-        }
-
-        return *ptr;
-    }
-
-    /*!
-    @brief return a const reference to the pointed to value
-
-    @param[in] ptr  a JSON value
-
-    @return const reference to the JSON value pointed to by the JSON
-    pointer
-
-    @throw parse_error.106   if an array index begins with '0'
-    @throw parse_error.109   if an array index was not a number
-    @throw out_of_range.402  if the array index '-' is used
-    @throw out_of_range.404  if the JSON pointer can not be resolved
-    */
-    template<typename BasicJsonType>
-    const BasicJsonType& get_unchecked(const BasicJsonType* ptr) const
-    {
-        for (const auto& reference_token : reference_tokens)
-        {
-            switch (ptr->type())
-            {
-                case detail::value_t::object:
-                {
-                    // use unchecked object access
-                    ptr = &ptr->operator[](reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
-                    {
-                        // "-" cannot be used for const access
-                        JSON_THROW(detail::out_of_range::create(402, detail::concat("array index '-' (", std::to_string(ptr->m_value.array->size()), ") is out of range"), ptr));
-                    }
-
-                    // use unchecked array access
-                    ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
-                    break;
-                }
-
-                case detail::value_t::null:
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
-            }
-        }
-
-        return *ptr;
-    }
-
-    /*!
-    @throw parse_error.106   if an array index begins with '0'
-    @throw parse_error.109   if an array index was not a number
-    @throw out_of_range.402  if the array index '-' is used
-    @throw out_of_range.404  if the JSON pointer can not be resolved
-    */
-    template<typename BasicJsonType>
-    const BasicJsonType& get_checked(const BasicJsonType* ptr) const
-    {
-        for (const auto& reference_token : reference_tokens)
-        {
-            switch (ptr->type())
-            {
-                case detail::value_t::object:
-                {
-                    // note: at performs range check
-                    ptr = &ptr->at(reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
-                    {
-                        // "-" always fails the range check
-                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
-                                "array index '-' (", std::to_string(ptr->m_value.array->size()),
-                                ") is out of range"), ptr));
-                    }
-
-                    // note: at performs range check
-                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
-                    break;
-                }
-
-                case detail::value_t::null:
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
-            }
-        }
-
-        return *ptr;
-    }
-
-    /*!
-    @throw parse_error.106   if an array index begins with '0'
-    @throw parse_error.109   if an array index was not a number
-    */
-    template<typename BasicJsonType>
-    bool contains(const BasicJsonType* ptr) const
-    {
-        for (const auto& reference_token : reference_tokens)
-        {
-            switch (ptr->type())
-            {
-                case detail::value_t::object:
-                {
-                    if (!ptr->contains(reference_token))
-                    {
-                        // we did not find the key in the object
-                        return false;
-                    }
-
-                    ptr = &ptr->operator[](reference_token);
-                    break;
-                }
-
-                case detail::value_t::array:
-                {
-                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
-                    {
-                        // "-" always fails the range check
-                        return false;
-                    }
-                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() == 1 && !("0" <= reference_token && reference_token <= "9")))
-                    {
-                        // invalid char
-                        return false;
-                    }
-                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() > 1))
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(!('1' <= reference_token[0] && reference_token[0] <= '9')))
-                        {
-                            // first char should be between '1' and '9'
-                            return false;
-                        }
-                        for (std::size_t i = 1; i < reference_token.size(); i++)
-                        {
-                            if (JSON_HEDLEY_UNLIKELY(!('0' <= reference_token[i] && reference_token[i] <= '9')))
-                            {
-                                // other char should be between '0' and '9'
-                                return false;
-                            }
-                        }
-                    }
-
-                    const auto idx = array_index<BasicJsonType>(reference_token);
-                    if (idx >= ptr->size())
-                    {
-                        // index out of range
-                        return false;
-                    }
-
-                    ptr = &ptr->operator[](idx);
-                    break;
-                }
-
-                case detail::value_t::null:
-                case detail::value_t::string:
-                case detail::value_t::boolean:
-                case detail::value_t::number_integer:
-                case detail::value_t::number_unsigned:
-                case detail::value_t::number_float:
-                case detail::value_t::binary:
-                case detail::value_t::discarded:
-                default:
-                {
-                    // we do not expect primitive values if there is still a
-                    // reference token to process
-                    return false;
-                }
-            }
-        }
-
-        // no reference token left means we found a primitive value
-        return true;
-    }
-
-    /*!
-    @brief split the string input to reference tokens
-
-    @note This function is only called by the json_pointer constructor.
-          All exceptions below are documented there.
-
-    @throw parse_error.107  if the pointer is not empty or begins with '/'
-    @throw parse_error.108  if character '~' is not followed by '0' or '1'
-    */
-    static std::vector<string_t> split(const string_t& reference_string)
-    {
-        std::vector<string_t> result;
-
-        // special case: empty reference string -> no reference tokens
-        if (reference_string.empty())
-        {
-            return result;
-        }
-
-        // check if nonempty reference string begins with slash
-        if (JSON_HEDLEY_UNLIKELY(reference_string[0] != '/'))
-        {
-            JSON_THROW(detail::parse_error::create(107, 1, detail::concat("JSON pointer must be empty or begin with '/' - was: '", reference_string, "'"), nullptr));
-        }
-
-        // extract the reference tokens:
-        // - slash: position of the last read slash (or end of string)
-        // - start: position after the previous slash
-        for (
-            // search for the first slash after the first character
-            std::size_t slash = reference_string.find_first_of('/', 1),
-            // set the beginning of the first reference token
-            start = 1;
-            // we can stop if start == 0 (if slash == string_t::npos)
-            start != 0;
-            // set the beginning of the next reference token
-            // (will eventually be 0 if slash == string_t::npos)
-            start = (slash == string_t::npos) ? 0 : slash + 1,
-            // find next slash
-            slash = reference_string.find_first_of('/', start))
-        {
-            // use the text between the beginning of the reference token
-            // (start) and the last slash (slash).
-            auto reference_token = reference_string.substr(start, slash - start);
-
-            // check reference tokens are properly escaped
-            for (std::size_t pos = reference_token.find_first_of('~');
-                    pos != string_t::npos;
-                    pos = reference_token.find_first_of('~', pos + 1))
-            {
-                JSON_ASSERT(reference_token[pos] == '~');
-
-                // ~ must be followed by 0 or 1
-                if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
-                                         (reference_token[pos + 1] != '0' &&
-                                          reference_token[pos + 1] != '1')))
-                {
-                    JSON_THROW(detail::parse_error::create(108, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
-                }
-            }
-
-            // finally, store the reference token
-            detail::unescape(reference_token);
-            result.push_back(reference_token);
-        }
-
-        return result;
-    }
-
-  private:
-    /*!
-    @param[in] reference_string  the reference string to the current value
-    @param[in] value             the value to consider
-    @param[in,out] result        the result object to insert values to
-
-    @note Empty objects or arrays are flattened to `null`.
-    */
-    template<typename BasicJsonType>
-    static void flatten(const string_t& reference_string,
-                        const BasicJsonType& value,
-                        BasicJsonType& result)
-    {
-        switch (value.type())
-        {
-            case detail::value_t::array:
-            {
-                if (value.m_value.array->empty())
-                {
-                    // flatten empty array as null
-                    result[reference_string] = nullptr;
-                }
-                else
-                {
-                    // iterate array and use index as reference string
-                    for (std::size_t i = 0; i < value.m_value.array->size(); ++i)
-                    {
-                        flatten(detail::concat(reference_string, '/', std::to_string(i)),
-                                value.m_value.array->operator[](i), result);
-                    }
-                }
-                break;
-            }
-
-            case detail::value_t::object:
-            {
-                if (value.m_value.object->empty())
-                {
-                    // flatten empty object as null
-                    result[reference_string] = nullptr;
-                }
-                else
-                {
-                    // iterate object and use keys as reference string
-                    for (const auto& element : *value.m_value.object)
-                    {
-                        flatten(detail::concat(reference_string, '/', detail::escape(element.first)), element.second, result);
-                    }
-                }
-                break;
-            }
-
-            case detail::value_t::null:
-            case detail::value_t::string:
-            case detail::value_t::boolean:
-            case detail::value_t::number_integer:
-            case detail::value_t::number_unsigned:
-            case detail::value_t::number_float:
-            case detail::value_t::binary:
-            case detail::value_t::discarded:
-            default:
-            {
-                // add primitive value with its reference string
-                result[reference_string] = value;
-                break;
-            }
-        }
-    }
-
-    /*!
-    @param[in] value  flattened JSON
-
-    @return unflattened JSON
-
-    @throw parse_error.109 if array index is not a number
-    @throw type_error.314  if value is not an object
-    @throw type_error.315  if object values are not primitive
-    @throw type_error.313  if value cannot be unflattened
-    */
-    template<typename BasicJsonType>
-    static BasicJsonType
-    unflatten(const BasicJsonType& value)
-    {
-        if (JSON_HEDLEY_UNLIKELY(!value.is_object()))
-        {
-            JSON_THROW(detail::type_error::create(314, "only objects can be unflattened", &value));
-        }
-
-        BasicJsonType result;
-
-        // iterate the JSON object values
-        for (const auto& element : *value.m_value.object)
-        {
-            if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive()))
-            {
-                JSON_THROW(detail::type_error::create(315, "values in object must be primitive", &element.second));
-            }
-
-            // assign value to reference pointed to by JSON pointer; Note that if
-            // the JSON pointer is "" (i.e., points to the whole value), function
-            // get_and_create returns a reference to result itself. An assignment
-            // will then create a primitive value.
-            json_pointer(element.first).get_and_create(result) = element.second;
-        }
-
-        return result;
-    }
-
-    // can't use conversion operator because of ambiguity
-    json_pointer<string_t> convert() const&
-    {
-        json_pointer<string_t> result;
-        result.reference_tokens = reference_tokens;
-        return result;
-    }
-
-    json_pointer<string_t> convert()&&
-    {
-        json_pointer<string_t> result;
-        result.reference_tokens = std::move(reference_tokens);
-        return result;
-    }
-
-  public:
-#if JSON_HAS_THREE_WAY_COMPARISON
-    /// @brief compares two JSON pointers for equality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
-    template<typename RefStringTypeRhs>
-    bool operator==(const json_pointer<RefStringTypeRhs>& rhs) const noexcept
-    {
-        return reference_tokens == rhs.reference_tokens;
-    }
-
-    /// @brief compares JSON pointer and string for equality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer))
-    bool operator==(const string_t& rhs) const
-    {
-        return *this == json_pointer(rhs);
-    }
-
-    /// @brief 3-way compares two JSON pointers
-    template<typename RefStringTypeRhs>
-    std::strong_ordering operator<=>(const json_pointer<RefStringTypeRhs>& rhs) const noexcept // *NOPAD*
-    {
-        return  reference_tokens <=> rhs.reference_tokens; // *NOPAD*
-    }
-#else
-    /// @brief compares two JSON pointers for equality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
-    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
-                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
-
-    /// @brief compares JSON pointer and string for equality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
-    template<typename RefStringTypeLhs, typename StringType>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
-                           const StringType& rhs);
-
-    /// @brief compares string and JSON pointer for equality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
-    template<typename RefStringTypeRhs, typename StringType>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator==(const StringType& lhs,
-                           const json_pointer<RefStringTypeRhs>& rhs);
-
-    /// @brief compares two JSON pointers for inequality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
-    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
-                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
-
-    /// @brief compares JSON pointer and string for inequality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
-    template<typename RefStringTypeLhs, typename StringType>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
-                           const StringType& rhs);
-
-    /// @brief compares string and JSON pointer for inequality
-    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
-    template<typename RefStringTypeRhs, typename StringType>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator!=(const StringType& lhs,
-                           const json_pointer<RefStringTypeRhs>& rhs);
-
-    /// @brief compares two JSON pointer for less-than
-    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-    // NOLINTNEXTLINE(readability-redundant-declaration)
-    friend bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
-                          const json_pointer<RefStringTypeRhs>& rhs) noexcept;
-#endif
-
-  private:
-    /// the reference tokens
-    std::vector<string_t> reference_tokens;
-};
-
-#if !JSON_HAS_THREE_WAY_COMPARISON
-// functions cannot be defined inside class due to ODR violations
-template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
-                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
-{
-    return lhs.reference_tokens == rhs.reference_tokens;
-}
-
-template<typename RefStringTypeLhs,
-         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
-JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
-inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
-                       const StringType& rhs)
-{
-    return lhs == json_pointer<RefStringTypeLhs>(rhs);
-}
-
-template<typename RefStringTypeRhs,
-         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
-JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
-inline bool operator==(const StringType& lhs,
-                       const json_pointer<RefStringTypeRhs>& rhs)
-{
-    return json_pointer<RefStringTypeRhs>(lhs) == rhs;
-}
-
-template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
-                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
-{
-    return !(lhs == rhs);
-}
-
-template<typename RefStringTypeLhs,
-         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
-JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
-inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
-                       const StringType& rhs)
-{
-    return !(lhs == rhs);
-}
-
-template<typename RefStringTypeRhs,
-         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
-JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
-inline bool operator!=(const StringType& lhs,
-                       const json_pointer<RefStringTypeRhs>& rhs)
-{
-    return !(lhs == rhs);
-}
-
-template<typename RefStringTypeLhs, typename RefStringTypeRhs>
-inline bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
-                      const json_pointer<RefStringTypeRhs>& rhs) noexcept
-{
-    return lhs.reference_tokens < rhs.reference_tokens;
-}
-#endif
-
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/json_ref.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <initializer_list>
-#include <utility>
-
-// #include <nlohmann/detail/abi_macros.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-template<typename BasicJsonType>
-class json_ref
-{
-  public:
-    using value_type = BasicJsonType;
-
-    json_ref(value_type&& value)
-        : owned_value(std::move(value))
-    {}
-
-    json_ref(const value_type& value)
-        : value_ref(&value)
-    {}
-
-    json_ref(std::initializer_list<json_ref> init)
-        : owned_value(init)
-    {}
-
-    template <
-        class... Args,
-        enable_if_t<std::is_constructible<value_type, Args...>::value, int> = 0 >
-    json_ref(Args && ... args)
-        : owned_value(std::forward<Args>(args)...)
-    {}
-
-    // class should be movable only
-    json_ref(json_ref&&) noexcept = default;
-    json_ref(const json_ref&) = delete;
-    json_ref& operator=(const json_ref&) = delete;
-    json_ref& operator=(json_ref&&) = delete;
-    ~json_ref() = default;
-
-    value_type moved_or_copied() const
-    {
-        if (value_ref == nullptr)
-        {
-            return std::move(owned_value);
-        }
-        return *value_ref;
-    }
-
-    value_type const& operator*() const
-    {
-        return value_ref ? *value_ref : owned_value;
-    }
-
-    value_type const* operator->() const
-    {
-        return &** this;
-    }
-
-  private:
-    mutable value_type owned_value = nullptr;
-    value_type const* value_ref = nullptr;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/string_escape.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-// #include <nlohmann/detail/output/binary_writer.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // reverse
-#include <array> // array
-#include <map> // map
-#include <cmath> // isnan, isinf
-#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
-#include <cstring> // memcpy
-#include <limits> // numeric_limits
-#include <string> // string
-#include <utility> // move
-#include <vector> // vector
-
-// #include <nlohmann/detail/input/binary_reader.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/output/output_adapters.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // copy
-#include <cstddef> // size_t
-#include <iterator> // back_inserter
-#include <memory> // shared_ptr, make_shared
-#include <string> // basic_string
-#include <vector> // vector
-
-#ifndef JSON_NO_IO
-    #include <ios>      // streamsize
-    #include <ostream>  // basic_ostream
-#endif  // JSON_NO_IO
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/// abstract output adapter interface
-template<typename CharType> struct output_adapter_protocol
-{
-    virtual void write_character(CharType c) = 0;
-    virtual void write_characters(const CharType* s, std::size_t length) = 0;
-    virtual ~output_adapter_protocol() = default;
-
-    output_adapter_protocol() = default;
-    output_adapter_protocol(const output_adapter_protocol&) = default;
-    output_adapter_protocol(output_adapter_protocol&&) noexcept = default;
-    output_adapter_protocol& operator=(const output_adapter_protocol&) = default;
-    output_adapter_protocol& operator=(output_adapter_protocol&&) noexcept = default;
-};
-
-/// a type to simplify interfaces
-template<typename CharType>
-using output_adapter_t = std::shared_ptr<output_adapter_protocol<CharType>>;
-
-/// output adapter for byte vectors
-template<typename CharType, typename AllocatorType = std::allocator<CharType>>
-class output_vector_adapter : public output_adapter_protocol<CharType>
-{
-  public:
-    explicit output_vector_adapter(std::vector<CharType, AllocatorType>& vec) noexcept
-        : v(vec)
-    {}
-
-    void write_character(CharType c) override
-    {
-        v.push_back(c);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    void write_characters(const CharType* s, std::size_t length) override
-    {
-        v.insert(v.end(), s, s + length);
-    }
-
-  private:
-    std::vector<CharType, AllocatorType>& v;
-};
-
-#ifndef JSON_NO_IO
-/// output adapter for output streams
-template<typename CharType>
-class output_stream_adapter : public output_adapter_protocol<CharType>
-{
-  public:
-    explicit output_stream_adapter(std::basic_ostream<CharType>& s) noexcept
-        : stream(s)
-    {}
-
-    void write_character(CharType c) override
-    {
-        stream.put(c);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    void write_characters(const CharType* s, std::size_t length) override
-    {
-        stream.write(s, static_cast<std::streamsize>(length));
-    }
-
-  private:
-    std::basic_ostream<CharType>& stream;
-};
-#endif  // JSON_NO_IO
-
-/// output adapter for basic_string
-template<typename CharType, typename StringType = std::basic_string<CharType>>
-class output_string_adapter : public output_adapter_protocol<CharType>
-{
-  public:
-    explicit output_string_adapter(StringType& s) noexcept
-        : str(s)
-    {}
-
-    void write_character(CharType c) override
-    {
-        str.push_back(c);
-    }
-
-    JSON_HEDLEY_NON_NULL(2)
-    void write_characters(const CharType* s, std::size_t length) override
-    {
-        str.append(s, length);
-    }
-
-  private:
-    StringType& str;
-};
-
-template<typename CharType, typename StringType = std::basic_string<CharType>>
-class output_adapter
-{
-  public:
-    template<typename AllocatorType = std::allocator<CharType>>
-    output_adapter(std::vector<CharType, AllocatorType>& vec)
-        : oa(std::make_shared<output_vector_adapter<CharType, AllocatorType>>(vec)) {}
-
-#ifndef JSON_NO_IO
-    output_adapter(std::basic_ostream<CharType>& s)
-        : oa(std::make_shared<output_stream_adapter<CharType>>(s)) {}
-#endif  // JSON_NO_IO
-
-    output_adapter(StringType& s)
-        : oa(std::make_shared<output_string_adapter<CharType, StringType>>(s)) {}
-
-    operator output_adapter_t<CharType>()
-    {
-        return oa;
-    }
-
-  private:
-    output_adapter_t<CharType> oa = nullptr;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-///////////////////
-// binary writer //
-///////////////////
-
-/*!
-@brief serialization to CBOR and MessagePack values
-*/
-template<typename BasicJsonType, typename CharType>
-class binary_writer
-{
-    using string_t = typename BasicJsonType::string_t;
-    using binary_t = typename BasicJsonType::binary_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-
-  public:
-    /*!
-    @brief create a binary writer
-
-    @param[in] adapter  output adapter to write to
-    */
-    explicit binary_writer(output_adapter_t<CharType> adapter) : oa(std::move(adapter))
-    {
-        JSON_ASSERT(oa);
-    }
-
-    /*!
-    @param[in] j  JSON value to serialize
-    @pre       j.type() == value_t::object
-    */
-    void write_bson(const BasicJsonType& j)
-    {
-        switch (j.type())
-        {
-            case value_t::object:
-            {
-                write_bson_object(*j.m_value.object);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::array:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                JSON_THROW(type_error::create(317, concat("to serialize to BSON, top-level type must be object, but is ", j.type_name()), &j));
-            }
-        }
-    }
-
-    /*!
-    @param[in] j  JSON value to serialize
-    */
-    void write_cbor(const BasicJsonType& j)
-    {
-        switch (j.type())
-        {
-            case value_t::null:
-            {
-                oa->write_character(to_char_type(0xF6));
-                break;
-            }
-
-            case value_t::boolean:
-            {
-                oa->write_character(j.m_value.boolean
-                                    ? to_char_type(0xF5)
-                                    : to_char_type(0xF4));
-                break;
-            }
-
-            case value_t::number_integer:
-            {
-                if (j.m_value.number_integer >= 0)
-                {
-                    // CBOR does not differentiate between positive signed
-                    // integers and unsigned integers. Therefore, we used the
-                    // code from the value_t::number_unsigned case here.
-                    if (j.m_value.number_integer <= 0x17)
-                    {
-                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x18));
-                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x19));
-                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x1A));
-                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
-                    }
-                    else
-                    {
-                        oa->write_character(to_char_type(0x1B));
-                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
-                    }
-                }
-                else
-                {
-                    // The conversions below encode the sign in the first
-                    // byte, and the value is converted to a positive number.
-                    const auto positive_number = -1 - j.m_value.number_integer;
-                    if (j.m_value.number_integer >= -24)
-                    {
-                        write_number(static_cast<std::uint8_t>(0x20 + positive_number));
-                    }
-                    else if (positive_number <= (std::numeric_limits<std::uint8_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x38));
-                        write_number(static_cast<std::uint8_t>(positive_number));
-                    }
-                    else if (positive_number <= (std::numeric_limits<std::uint16_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x39));
-                        write_number(static_cast<std::uint16_t>(positive_number));
-                    }
-                    else if (positive_number <= (std::numeric_limits<std::uint32_t>::max)())
-                    {
-                        oa->write_character(to_char_type(0x3A));
-                        write_number(static_cast<std::uint32_t>(positive_number));
-                    }
-                    else
-                    {
-                        oa->write_character(to_char_type(0x3B));
-                        write_number(static_cast<std::uint64_t>(positive_number));
-                    }
-                }
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                if (j.m_value.number_unsigned <= 0x17)
-                {
-                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x18));
-                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x19));
-                    write_number(static_cast<std::uint16_t>(j.m_value.number_unsigned));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x1A));
-                    write_number(static_cast<std::uint32_t>(j.m_value.number_unsigned));
-                }
-                else
-                {
-                    oa->write_character(to_char_type(0x1B));
-                    write_number(static_cast<std::uint64_t>(j.m_value.number_unsigned));
-                }
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                if (std::isnan(j.m_value.number_float))
-                {
-                    // NaN is 0xf97e00 in CBOR
-                    oa->write_character(to_char_type(0xF9));
-                    oa->write_character(to_char_type(0x7E));
-                    oa->write_character(to_char_type(0x00));
-                }
-                else if (std::isinf(j.m_value.number_float))
-                {
-                    // Infinity is 0xf97c00, -Infinity is 0xf9fc00
-                    oa->write_character(to_char_type(0xf9));
-                    oa->write_character(j.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC));
-                    oa->write_character(to_char_type(0x00));
-                }
-                else
-                {
-                    write_compact_float(j.m_value.number_float, detail::input_format_t::cbor);
-                }
-                break;
-            }
-
-            case value_t::string:
-            {
-                // step 1: write control byte and the string length
-                const auto N = j.m_value.string->size();
-                if (N <= 0x17)
-                {
-                    write_number(static_cast<std::uint8_t>(0x60 + N));
-                }
-                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x78));
-                    write_number(static_cast<std::uint8_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x79));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x7A));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-                // LCOV_EXCL_START
-                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x7B));
-                    write_number(static_cast<std::uint64_t>(N));
-                }
-                // LCOV_EXCL_STOP
-
-                // step 2: write the string
-                oa->write_characters(
-                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
-                    j.m_value.string->size());
-                break;
-            }
-
-            case value_t::array:
-            {
-                // step 1: write control byte and the array size
-                const auto N = j.m_value.array->size();
-                if (N <= 0x17)
-                {
-                    write_number(static_cast<std::uint8_t>(0x80 + N));
-                }
-                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x98));
-                    write_number(static_cast<std::uint8_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x99));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x9A));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-                // LCOV_EXCL_START
-                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x9B));
-                    write_number(static_cast<std::uint64_t>(N));
-                }
-                // LCOV_EXCL_STOP
-
-                // step 2: write each element
-                for (const auto& el : *j.m_value.array)
-                {
-                    write_cbor(el);
-                }
-                break;
-            }
-
-            case value_t::binary:
-            {
-                if (j.m_value.binary->has_subtype())
-                {
-                    if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint8_t>::max)())
-                    {
-                        write_number(static_cast<std::uint8_t>(0xd8));
-                        write_number(static_cast<std::uint8_t>(j.m_value.binary->subtype()));
-                    }
-                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint16_t>::max)())
-                    {
-                        write_number(static_cast<std::uint8_t>(0xd9));
-                        write_number(static_cast<std::uint16_t>(j.m_value.binary->subtype()));
-                    }
-                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint32_t>::max)())
-                    {
-                        write_number(static_cast<std::uint8_t>(0xda));
-                        write_number(static_cast<std::uint32_t>(j.m_value.binary->subtype()));
-                    }
-                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint64_t>::max)())
-                    {
-                        write_number(static_cast<std::uint8_t>(0xdb));
-                        write_number(static_cast<std::uint64_t>(j.m_value.binary->subtype()));
-                    }
-                }
-
-                // step 1: write control byte and the binary array size
-                const auto N = j.m_value.binary->size();
-                if (N <= 0x17)
-                {
-                    write_number(static_cast<std::uint8_t>(0x40 + N));
-                }
-                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x58));
-                    write_number(static_cast<std::uint8_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x59));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x5A));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-                // LCOV_EXCL_START
-                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    oa->write_character(to_char_type(0x5B));
-                    write_number(static_cast<std::uint64_t>(N));
-                }
-                // LCOV_EXCL_STOP
-
-                // step 2: write each element
-                oa->write_characters(
-                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
-                    N);
-
-                break;
-            }
-
-            case value_t::object:
-            {
-                // step 1: write control byte and the object size
-                const auto N = j.m_value.object->size();
-                if (N <= 0x17)
-                {
-                    write_number(static_cast<std::uint8_t>(0xA0 + N));
-                }
-                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    oa->write_character(to_char_type(0xB8));
-                    write_number(static_cast<std::uint8_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    oa->write_character(to_char_type(0xB9));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    oa->write_character(to_char_type(0xBA));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-                // LCOV_EXCL_START
-                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    oa->write_character(to_char_type(0xBB));
-                    write_number(static_cast<std::uint64_t>(N));
-                }
-                // LCOV_EXCL_STOP
-
-                // step 2: write each element
-                for (const auto& el : *j.m_value.object)
-                {
-                    write_cbor(el.first);
-                    write_cbor(el.second);
-                }
-                break;
-            }
-
-            case value_t::discarded:
-            default:
-                break;
-        }
-    }
-
-    /*!
-    @param[in] j  JSON value to serialize
-    */
-    void write_msgpack(const BasicJsonType& j)
-    {
-        switch (j.type())
-        {
-            case value_t::null: // nil
-            {
-                oa->write_character(to_char_type(0xC0));
-                break;
-            }
-
-            case value_t::boolean: // true and false
-            {
-                oa->write_character(j.m_value.boolean
-                                    ? to_char_type(0xC3)
-                                    : to_char_type(0xC2));
-                break;
-            }
-
-            case value_t::number_integer:
-            {
-                if (j.m_value.number_integer >= 0)
-                {
-                    // MessagePack does not differentiate between positive
-                    // signed integers and unsigned integers. Therefore, we used
-                    // the code from the value_t::number_unsigned case here.
-                    if (j.m_value.number_unsigned < 128)
-                    {
-                        // positive fixnum
-                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
-                    {
-                        // uint 8
-                        oa->write_character(to_char_type(0xCC));
-                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
-                    {
-                        // uint 16
-                        oa->write_character(to_char_type(0xCD));
-                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
-                    {
-                        // uint 32
-                        oa->write_character(to_char_type(0xCE));
-                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
-                    {
-                        // uint 64
-                        oa->write_character(to_char_type(0xCF));
-                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
-                    }
-                }
-                else
-                {
-                    if (j.m_value.number_integer >= -32)
-                    {
-                        // negative fixnum
-                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int8_t>::min)() &&
-                             j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
-                    {
-                        // int 8
-                        oa->write_character(to_char_type(0xD0));
-                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int16_t>::min)() &&
-                             j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
-                    {
-                        // int 16
-                        oa->write_character(to_char_type(0xD1));
-                        write_number(static_cast<std::int16_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int32_t>::min)() &&
-                             j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
-                    {
-                        // int 32
-                        oa->write_character(to_char_type(0xD2));
-                        write_number(static_cast<std::int32_t>(j.m_value.number_integer));
-                    }
-                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int64_t>::min)() &&
-                             j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
-                    {
-                        // int 64
-                        oa->write_character(to_char_type(0xD3));
-                        write_number(static_cast<std::int64_t>(j.m_value.number_integer));
-                    }
-                }
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                if (j.m_value.number_unsigned < 128)
-                {
-                    // positive fixnum
-                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    // uint 8
-                    oa->write_character(to_char_type(0xCC));
-                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    // uint 16
-                    oa->write_character(to_char_type(0xCD));
-                    write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    // uint 32
-                    oa->write_character(to_char_type(0xCE));
-                    write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
-                }
-                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    // uint 64
-                    oa->write_character(to_char_type(0xCF));
-                    write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
-                }
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                write_compact_float(j.m_value.number_float, detail::input_format_t::msgpack);
-                break;
-            }
-
-            case value_t::string:
-            {
-                // step 1: write control byte and the string length
-                const auto N = j.m_value.string->size();
-                if (N <= 31)
-                {
-                    // fixstr
-                    write_number(static_cast<std::uint8_t>(0xA0 | N));
-                }
-                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    // str 8
-                    oa->write_character(to_char_type(0xD9));
-                    write_number(static_cast<std::uint8_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    // str 16
-                    oa->write_character(to_char_type(0xDA));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    // str 32
-                    oa->write_character(to_char_type(0xDB));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-
-                // step 2: write the string
-                oa->write_characters(
-                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
-                    j.m_value.string->size());
-                break;
-            }
-
-            case value_t::array:
-            {
-                // step 1: write control byte and the array size
-                const auto N = j.m_value.array->size();
-                if (N <= 15)
-                {
-                    // fixarray
-                    write_number(static_cast<std::uint8_t>(0x90 | N));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    // array 16
-                    oa->write_character(to_char_type(0xDC));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    // array 32
-                    oa->write_character(to_char_type(0xDD));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-
-                // step 2: write each element
-                for (const auto& el : *j.m_value.array)
-                {
-                    write_msgpack(el);
-                }
-                break;
-            }
-
-            case value_t::binary:
-            {
-                // step 0: determine if the binary type has a set subtype to
-                // determine whether or not to use the ext or fixext types
-                const bool use_ext = j.m_value.binary->has_subtype();
-
-                // step 1: write control byte and the byte string length
-                const auto N = j.m_value.binary->size();
-                if (N <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    std::uint8_t output_type{};
-                    bool fixed = true;
-                    if (use_ext)
-                    {
-                        switch (N)
-                        {
-                            case 1:
-                                output_type = 0xD4; // fixext 1
-                                break;
-                            case 2:
-                                output_type = 0xD5; // fixext 2
-                                break;
-                            case 4:
-                                output_type = 0xD6; // fixext 4
-                                break;
-                            case 8:
-                                output_type = 0xD7; // fixext 8
-                                break;
-                            case 16:
-                                output_type = 0xD8; // fixext 16
-                                break;
-                            default:
-                                output_type = 0xC7; // ext 8
-                                fixed = false;
-                                break;
-                        }
-
-                    }
-                    else
-                    {
-                        output_type = 0xC4; // bin 8
-                        fixed = false;
-                    }
-
-                    oa->write_character(to_char_type(output_type));
-                    if (!fixed)
-                    {
-                        write_number(static_cast<std::uint8_t>(N));
-                    }
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    std::uint8_t output_type = use_ext
-                                               ? 0xC8 // ext 16
-                                               : 0xC5; // bin 16
-
-                    oa->write_character(to_char_type(output_type));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    std::uint8_t output_type = use_ext
-                                               ? 0xC9 // ext 32
-                                               : 0xC6; // bin 32
-
-                    oa->write_character(to_char_type(output_type));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-
-                // step 1.5: if this is an ext type, write the subtype
-                if (use_ext)
-                {
-                    write_number(static_cast<std::int8_t>(j.m_value.binary->subtype()));
-                }
-
-                // step 2: write the byte string
-                oa->write_characters(
-                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
-                    N);
-
-                break;
-            }
-
-            case value_t::object:
-            {
-                // step 1: write control byte and the object size
-                const auto N = j.m_value.object->size();
-                if (N <= 15)
-                {
-                    // fixmap
-                    write_number(static_cast<std::uint8_t>(0x80 | (N & 0xF)));
-                }
-                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
-                {
-                    // map 16
-                    oa->write_character(to_char_type(0xDE));
-                    write_number(static_cast<std::uint16_t>(N));
-                }
-                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
-                {
-                    // map 32
-                    oa->write_character(to_char_type(0xDF));
-                    write_number(static_cast<std::uint32_t>(N));
-                }
-
-                // step 2: write each element
-                for (const auto& el : *j.m_value.object)
-                {
-                    write_msgpack(el.first);
-                    write_msgpack(el.second);
-                }
-                break;
-            }
-
-            case value_t::discarded:
-            default:
-                break;
-        }
-    }
-
-    /*!
-    @param[in] j  JSON value to serialize
-    @param[in] use_count   whether to use '#' prefixes (optimized format)
-    @param[in] use_type    whether to use '$' prefixes (optimized format)
-    @param[in] add_prefix  whether prefixes need to be used for this value
-    @param[in] use_bjdata  whether write in BJData format, default is false
-    */
-    void write_ubjson(const BasicJsonType& j, const bool use_count,
-                      const bool use_type, const bool add_prefix = true,
-                      const bool use_bjdata = false)
-    {
-        switch (j.type())
-        {
-            case value_t::null:
-            {
-                if (add_prefix)
-                {
-                    oa->write_character(to_char_type('Z'));
-                }
-                break;
-            }
-
-            case value_t::boolean:
-            {
-                if (add_prefix)
-                {
-                    oa->write_character(j.m_value.boolean
-                                        ? to_char_type('T')
-                                        : to_char_type('F'));
-                }
-                break;
-            }
-
-            case value_t::number_integer:
-            {
-                write_number_with_ubjson_prefix(j.m_value.number_integer, add_prefix, use_bjdata);
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                write_number_with_ubjson_prefix(j.m_value.number_unsigned, add_prefix, use_bjdata);
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                write_number_with_ubjson_prefix(j.m_value.number_float, add_prefix, use_bjdata);
-                break;
-            }
-
-            case value_t::string:
-            {
-                if (add_prefix)
-                {
-                    oa->write_character(to_char_type('S'));
-                }
-                write_number_with_ubjson_prefix(j.m_value.string->size(), true, use_bjdata);
-                oa->write_characters(
-                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
-                    j.m_value.string->size());
-                break;
-            }
-
-            case value_t::array:
-            {
-                if (add_prefix)
-                {
-                    oa->write_character(to_char_type('['));
-                }
-
-                bool prefix_required = true;
-                if (use_type && !j.m_value.array->empty())
-                {
-                    JSON_ASSERT(use_count);
-                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
-                    const bool same_prefix = std::all_of(j.begin() + 1, j.end(),
-                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
-                    {
-                        return ubjson_prefix(v, use_bjdata) == first_prefix;
-                    });
-
-                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
-
-                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
-                    {
-                        prefix_required = false;
-                        oa->write_character(to_char_type('$'));
-                        oa->write_character(first_prefix);
-                    }
-                }
-
-                if (use_count)
-                {
-                    oa->write_character(to_char_type('#'));
-                    write_number_with_ubjson_prefix(j.m_value.array->size(), true, use_bjdata);
-                }
-
-                for (const auto& el : *j.m_value.array)
-                {
-                    write_ubjson(el, use_count, use_type, prefix_required, use_bjdata);
-                }
-
-                if (!use_count)
-                {
-                    oa->write_character(to_char_type(']'));
-                }
-
-                break;
-            }
-
-            case value_t::binary:
-            {
-                if (add_prefix)
-                {
-                    oa->write_character(to_char_type('['));
-                }
-
-                if (use_type && !j.m_value.binary->empty())
-                {
-                    JSON_ASSERT(use_count);
-                    oa->write_character(to_char_type('$'));
-                    oa->write_character('U');
-                }
-
-                if (use_count)
-                {
-                    oa->write_character(to_char_type('#'));
-                    write_number_with_ubjson_prefix(j.m_value.binary->size(), true, use_bjdata);
-                }
-
-                if (use_type)
-                {
-                    oa->write_characters(
-                        reinterpret_cast<const CharType*>(j.m_value.binary->data()),
-                        j.m_value.binary->size());
-                }
-                else
-                {
-                    for (size_t i = 0; i < j.m_value.binary->size(); ++i)
-                    {
-                        oa->write_character(to_char_type('U'));
-                        oa->write_character(j.m_value.binary->data()[i]);
-                    }
-                }
-
-                if (!use_count)
-                {
-                    oa->write_character(to_char_type(']'));
-                }
-
-                break;
-            }
-
-            case value_t::object:
-            {
-                if (use_bjdata && j.m_value.object->size() == 3 && j.m_value.object->find("_ArrayType_") != j.m_value.object->end() && j.m_value.object->find("_ArraySize_") != j.m_value.object->end() && j.m_value.object->find("_ArrayData_") != j.m_value.object->end())
-                {
-                    if (!write_bjdata_ndarray(*j.m_value.object, use_count, use_type))  // decode bjdata ndarray in the JData format (https://github.com/NeuroJSON/jdata)
-                    {
-                        break;
-                    }
-                }
-
-                if (add_prefix)
-                {
-                    oa->write_character(to_char_type('{'));
-                }
-
-                bool prefix_required = true;
-                if (use_type && !j.m_value.object->empty())
-                {
-                    JSON_ASSERT(use_count);
-                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
-                    const bool same_prefix = std::all_of(j.begin(), j.end(),
-                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
-                    {
-                        return ubjson_prefix(v, use_bjdata) == first_prefix;
-                    });
-
-                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
-
-                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
-                    {
-                        prefix_required = false;
-                        oa->write_character(to_char_type('$'));
-                        oa->write_character(first_prefix);
-                    }
-                }
-
-                if (use_count)
-                {
-                    oa->write_character(to_char_type('#'));
-                    write_number_with_ubjson_prefix(j.m_value.object->size(), true, use_bjdata);
-                }
-
-                for (const auto& el : *j.m_value.object)
-                {
-                    write_number_with_ubjson_prefix(el.first.size(), true, use_bjdata);
-                    oa->write_characters(
-                        reinterpret_cast<const CharType*>(el.first.c_str()),
-                        el.first.size());
-                    write_ubjson(el.second, use_count, use_type, prefix_required, use_bjdata);
-                }
-
-                if (!use_count)
-                {
-                    oa->write_character(to_char_type('}'));
-                }
-
-                break;
-            }
-
-            case value_t::discarded:
-            default:
-                break;
-        }
-    }
-
-  private:
-    //////////
-    // BSON //
-    //////////
-
-    /*!
-    @return The size of a BSON document entry header, including the id marker
-            and the entry name size (and its null-terminator).
-    */
-    static std::size_t calc_bson_entry_header_size(const string_t& name, const BasicJsonType& j)
-    {
-        const auto it = name.find(static_cast<typename string_t::value_type>(0));
-        if (JSON_HEDLEY_UNLIKELY(it != BasicJsonType::string_t::npos))
-        {
-            JSON_THROW(out_of_range::create(409, concat("BSON key cannot contain code point U+0000 (at byte ", std::to_string(it), ")"), &j));
-            static_cast<void>(j);
-        }
-
-        return /*id*/ 1ul + name.size() + /*zero-terminator*/1u;
-    }
-
-    /*!
-    @brief Writes the given @a element_type and @a name to the output adapter
-    */
-    void write_bson_entry_header(const string_t& name,
-                                 const std::uint8_t element_type)
-    {
-        oa->write_character(to_char_type(element_type)); // boolean
-        oa->write_characters(
-            reinterpret_cast<const CharType*>(name.c_str()),
-            name.size() + 1u);
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and boolean value @a value
-    */
-    void write_bson_boolean(const string_t& name,
-                            const bool value)
-    {
-        write_bson_entry_header(name, 0x08);
-        oa->write_character(value ? to_char_type(0x01) : to_char_type(0x00));
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and double value @a value
-    */
-    void write_bson_double(const string_t& name,
-                           const double value)
-    {
-        write_bson_entry_header(name, 0x01);
-        write_number<double>(value, true);
-    }
-
-    /*!
-    @return The size of the BSON-encoded string in @a value
-    */
-    static std::size_t calc_bson_string_size(const string_t& value)
-    {
-        return sizeof(std::int32_t) + value.size() + 1ul;
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and string value @a value
-    */
-    void write_bson_string(const string_t& name,
-                           const string_t& value)
-    {
-        write_bson_entry_header(name, 0x02);
-
-        write_number<std::int32_t>(static_cast<std::int32_t>(value.size() + 1ul), true);
-        oa->write_characters(
-            reinterpret_cast<const CharType*>(value.c_str()),
-            value.size() + 1);
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and null value
-    */
-    void write_bson_null(const string_t& name)
-    {
-        write_bson_entry_header(name, 0x0A);
-    }
-
-    /*!
-    @return The size of the BSON-encoded integer @a value
-    */
-    static std::size_t calc_bson_integer_size(const std::int64_t value)
-    {
-        return (std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)()
-               ? sizeof(std::int32_t)
-               : sizeof(std::int64_t);
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and integer @a value
-    */
-    void write_bson_integer(const string_t& name,
-                            const std::int64_t value)
-    {
-        if ((std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)())
-        {
-            write_bson_entry_header(name, 0x10); // int32
-            write_number<std::int32_t>(static_cast<std::int32_t>(value), true);
-        }
-        else
-        {
-            write_bson_entry_header(name, 0x12); // int64
-            write_number<std::int64_t>(static_cast<std::int64_t>(value), true);
-        }
-    }
-
-    /*!
-    @return The size of the BSON-encoded unsigned integer in @a j
-    */
-    static constexpr std::size_t calc_bson_unsigned_size(const std::uint64_t value) noexcept
-    {
-        return (value <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
-               ? sizeof(std::int32_t)
-               : sizeof(std::int64_t);
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and unsigned @a value
-    */
-    void write_bson_unsigned(const string_t& name,
-                             const BasicJsonType& j)
-    {
-        if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
-        {
-            write_bson_entry_header(name, 0x10 /* int32 */);
-            write_number<std::int32_t>(static_cast<std::int32_t>(j.m_value.number_unsigned), true);
-        }
-        else if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
-        {
-            write_bson_entry_header(name, 0x12 /* int64 */);
-            write_number<std::int64_t>(static_cast<std::int64_t>(j.m_value.number_unsigned), true);
-        }
-        else
-        {
-            JSON_THROW(out_of_range::create(407, concat("integer number ", std::to_string(j.m_value.number_unsigned), " cannot be represented by BSON as it does not fit int64"), &j));
-        }
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and object @a value
-    */
-    void write_bson_object_entry(const string_t& name,
-                                 const typename BasicJsonType::object_t& value)
-    {
-        write_bson_entry_header(name, 0x03); // object
-        write_bson_object(value);
-    }
-
-    /*!
-    @return The size of the BSON-encoded array @a value
-    */
-    static std::size_t calc_bson_array_size(const typename BasicJsonType::array_t& value)
-    {
-        std::size_t array_index = 0ul;
-
-        const std::size_t embedded_document_size = std::accumulate(std::begin(value), std::end(value), static_cast<std::size_t>(0), [&array_index](std::size_t result, const typename BasicJsonType::array_t::value_type & el)
-        {
-            return result + calc_bson_element_size(std::to_string(array_index++), el);
-        });
-
-        return sizeof(std::int32_t) + embedded_document_size + 1ul;
-    }
-
-    /*!
-    @return The size of the BSON-encoded binary array @a value
-    */
-    static std::size_t calc_bson_binary_size(const typename BasicJsonType::binary_t& value)
-    {
-        return sizeof(std::int32_t) + value.size() + 1ul;
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and array @a value
-    */
-    void write_bson_array(const string_t& name,
-                          const typename BasicJsonType::array_t& value)
-    {
-        write_bson_entry_header(name, 0x04); // array
-        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_array_size(value)), true);
-
-        std::size_t array_index = 0ul;
-
-        for (const auto& el : value)
-        {
-            write_bson_element(std::to_string(array_index++), el);
-        }
-
-        oa->write_character(to_char_type(0x00));
-    }
-
-    /*!
-    @brief Writes a BSON element with key @a name and binary value @a value
-    */
-    void write_bson_binary(const string_t& name,
-                           const binary_t& value)
-    {
-        write_bson_entry_header(name, 0x05);
-
-        write_number<std::int32_t>(static_cast<std::int32_t>(value.size()), true);
-        write_number(value.has_subtype() ? static_cast<std::uint8_t>(value.subtype()) : static_cast<std::uint8_t>(0x00));
-
-        oa->write_characters(reinterpret_cast<const CharType*>(value.data()), value.size());
-    }
-
-    /*!
-    @brief Calculates the size necessary to serialize the JSON value @a j with its @a name
-    @return The calculated size for the BSON document entry for @a j with the given @a name.
-    */
-    static std::size_t calc_bson_element_size(const string_t& name,
-            const BasicJsonType& j)
-    {
-        const auto header_size = calc_bson_entry_header_size(name, j);
-        switch (j.type())
-        {
-            case value_t::object:
-                return header_size + calc_bson_object_size(*j.m_value.object);
-
-            case value_t::array:
-                return header_size + calc_bson_array_size(*j.m_value.array);
-
-            case value_t::binary:
-                return header_size + calc_bson_binary_size(*j.m_value.binary);
-
-            case value_t::boolean:
-                return header_size + 1ul;
-
-            case value_t::number_float:
-                return header_size + 8ul;
-
-            case value_t::number_integer:
-                return header_size + calc_bson_integer_size(j.m_value.number_integer);
-
-            case value_t::number_unsigned:
-                return header_size + calc_bson_unsigned_size(j.m_value.number_unsigned);
-
-            case value_t::string:
-                return header_size + calc_bson_string_size(*j.m_value.string);
-
-            case value_t::null:
-                return header_size + 0ul;
-
-            // LCOV_EXCL_START
-            case value_t::discarded:
-            default:
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
-                return 0ul;
-                // LCOV_EXCL_STOP
-        }
-    }
-
-    /*!
-    @brief Serializes the JSON value @a j to BSON and associates it with the
-           key @a name.
-    @param name The name to associate with the JSON entity @a j within the
-                current BSON document
-    */
-    void write_bson_element(const string_t& name,
-                            const BasicJsonType& j)
-    {
-        switch (j.type())
-        {
-            case value_t::object:
-                return write_bson_object_entry(name, *j.m_value.object);
-
-            case value_t::array:
-                return write_bson_array(name, *j.m_value.array);
-
-            case value_t::binary:
-                return write_bson_binary(name, *j.m_value.binary);
-
-            case value_t::boolean:
-                return write_bson_boolean(name, j.m_value.boolean);
-
-            case value_t::number_float:
-                return write_bson_double(name, j.m_value.number_float);
-
-            case value_t::number_integer:
-                return write_bson_integer(name, j.m_value.number_integer);
-
-            case value_t::number_unsigned:
-                return write_bson_unsigned(name, j);
-
-            case value_t::string:
-                return write_bson_string(name, *j.m_value.string);
-
-            case value_t::null:
-                return write_bson_null(name);
-
-            // LCOV_EXCL_START
-            case value_t::discarded:
-            default:
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
-                return;
-                // LCOV_EXCL_STOP
-        }
-    }
-
-    /*!
-    @brief Calculates the size of the BSON serialization of the given
-           JSON-object @a j.
-    @param[in] value  JSON value to serialize
-    @pre       value.type() == value_t::object
-    */
-    static std::size_t calc_bson_object_size(const typename BasicJsonType::object_t& value)
-    {
-        std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast<std::size_t>(0),
-                                    [](size_t result, const typename BasicJsonType::object_t::value_type & el)
-        {
-            return result += calc_bson_element_size(el.first, el.second);
-        });
-
-        return sizeof(std::int32_t) + document_size + 1ul;
-    }
-
-    /*!
-    @param[in] value  JSON value to serialize
-    @pre       value.type() == value_t::object
-    */
-    void write_bson_object(const typename BasicJsonType::object_t& value)
-    {
-        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_object_size(value)), true);
-
-        for (const auto& el : value)
-        {
-            write_bson_element(el.first, el.second);
-        }
-
-        oa->write_character(to_char_type(0x00));
-    }
-
-    //////////
-    // CBOR //
-    //////////
-
-    static constexpr CharType get_cbor_float_prefix(float /*unused*/)
-    {
-        return to_char_type(0xFA);  // Single-Precision Float
-    }
-
-    static constexpr CharType get_cbor_float_prefix(double /*unused*/)
-    {
-        return to_char_type(0xFB);  // Double-Precision Float
-    }
-
-    /////////////
-    // MsgPack //
-    /////////////
-
-    static constexpr CharType get_msgpack_float_prefix(float /*unused*/)
-    {
-        return to_char_type(0xCA);  // float 32
-    }
-
-    static constexpr CharType get_msgpack_float_prefix(double /*unused*/)
-    {
-        return to_char_type(0xCB);  // float 64
-    }
-
-    ////////////
-    // UBJSON //
-    ////////////
-
-    // UBJSON: write number (floating point)
-    template<typename NumberType, typename std::enable_if<
-                 std::is_floating_point<NumberType>::value, int>::type = 0>
-    void write_number_with_ubjson_prefix(const NumberType n,
-                                         const bool add_prefix,
-                                         const bool use_bjdata)
-    {
-        if (add_prefix)
-        {
-            oa->write_character(get_ubjson_float_prefix(n));
-        }
-        write_number(n, use_bjdata);
-    }
-
-    // UBJSON: write number (unsigned integer)
-    template<typename NumberType, typename std::enable_if<
-                 std::is_unsigned<NumberType>::value, int>::type = 0>
-    void write_number_with_ubjson_prefix(const NumberType n,
-                                         const bool add_prefix,
-                                         const bool use_bjdata)
-    {
-        if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('i'));  // int8
-            }
-            write_number(static_cast<std::uint8_t>(n), use_bjdata);
-        }
-        else if (n <= (std::numeric_limits<std::uint8_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('U'));  // uint8
-            }
-            write_number(static_cast<std::uint8_t>(n), use_bjdata);
-        }
-        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('I'));  // int16
-            }
-            write_number(static_cast<std::int16_t>(n), use_bjdata);
-        }
-        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint16_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
-            }
-            write_number(static_cast<std::uint16_t>(n), use_bjdata);
-        }
-        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('l'));  // int32
-            }
-            write_number(static_cast<std::int32_t>(n), use_bjdata);
-        }
-        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint32_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
-            }
-            write_number(static_cast<std::uint32_t>(n), use_bjdata);
-        }
-        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('L'));  // int64
-            }
-            write_number(static_cast<std::int64_t>(n), use_bjdata);
-        }
-        else if (use_bjdata && n <= (std::numeric_limits<uint64_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('M'));  // uint64 - bjdata only
-            }
-            write_number(static_cast<std::uint64_t>(n), use_bjdata);
-        }
-        else
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('H'));  // high-precision number
-            }
-
-            const auto number = BasicJsonType(n).dump();
-            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
-            for (std::size_t i = 0; i < number.size(); ++i)
-            {
-                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
-            }
-        }
-    }
-
-    // UBJSON: write number (signed integer)
-    template < typename NumberType, typename std::enable_if <
-                   std::is_signed<NumberType>::value&&
-                   !std::is_floating_point<NumberType>::value, int >::type = 0 >
-    void write_number_with_ubjson_prefix(const NumberType n,
-                                         const bool add_prefix,
-                                         const bool use_bjdata)
-    {
-        if ((std::numeric_limits<std::int8_t>::min)() <= n && n <= (std::numeric_limits<std::int8_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('i'));  // int8
-            }
-            write_number(static_cast<std::int8_t>(n), use_bjdata);
-        }
-        else if (static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::max)()))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('U'));  // uint8
-            }
-            write_number(static_cast<std::uint8_t>(n), use_bjdata);
-        }
-        else if ((std::numeric_limits<std::int16_t>::min)() <= n && n <= (std::numeric_limits<std::int16_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('I'));  // int16
-            }
-            write_number(static_cast<std::int16_t>(n), use_bjdata);
-        }
-        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::max)())))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
-            }
-            write_number(static_cast<uint16_t>(n), use_bjdata);
-        }
-        else if ((std::numeric_limits<std::int32_t>::min)() <= n && n <= (std::numeric_limits<std::int32_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('l'));  // int32
-            }
-            write_number(static_cast<std::int32_t>(n), use_bjdata);
-        }
-        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::max)())))
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
-            }
-            write_number(static_cast<uint32_t>(n), use_bjdata);
-        }
-        else if ((std::numeric_limits<std::int64_t>::min)() <= n && n <= (std::numeric_limits<std::int64_t>::max)())
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('L'));  // int64
-            }
-            write_number(static_cast<std::int64_t>(n), use_bjdata);
-        }
-        // LCOV_EXCL_START
-        else
-        {
-            if (add_prefix)
-            {
-                oa->write_character(to_char_type('H'));  // high-precision number
-            }
-
-            const auto number = BasicJsonType(n).dump();
-            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
-            for (std::size_t i = 0; i < number.size(); ++i)
-            {
-                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
-            }
-        }
-        // LCOV_EXCL_STOP
-    }
-
-    /*!
-    @brief determine the type prefix of container values
-    */
-    CharType ubjson_prefix(const BasicJsonType& j, const bool use_bjdata) const noexcept
-    {
-        switch (j.type())
-        {
-            case value_t::null:
-                return 'Z';
-
-            case value_t::boolean:
-                return j.m_value.boolean ? 'T' : 'F';
-
-            case value_t::number_integer:
-            {
-                if ((std::numeric_limits<std::int8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
-                {
-                    return 'i';
-                }
-                if ((std::numeric_limits<std::uint8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
-                {
-                    return 'U';
-                }
-                if ((std::numeric_limits<std::int16_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
-                {
-                    return 'I';
-                }
-                if (use_bjdata && ((std::numeric_limits<std::uint16_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)()))
-                {
-                    return 'u';
-                }
-                if ((std::numeric_limits<std::int32_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
-                {
-                    return 'l';
-                }
-                if (use_bjdata && ((std::numeric_limits<std::uint32_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)()))
-                {
-                    return 'm';
-                }
-                if ((std::numeric_limits<std::int64_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
-                {
-                    return 'L';
-                }
-                // anything else is treated as high-precision number
-                return 'H'; // LCOV_EXCL_LINE
-            }
-
-            case value_t::number_unsigned:
-            {
-                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
-                {
-                    return 'i';
-                }
-                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint8_t>::max)()))
-                {
-                    return 'U';
-                }
-                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
-                {
-                    return 'I';
-                }
-                if (use_bjdata && j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint16_t>::max)()))
-                {
-                    return 'u';
-                }
-                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
-                {
-                    return 'l';
-                }
-                if (use_bjdata && j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint32_t>::max)()))
-                {
-                    return 'm';
-                }
-                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
-                {
-                    return 'L';
-                }
-                if (use_bjdata && j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
-                {
-                    return 'M';
-                }
-                // anything else is treated as high-precision number
-                return 'H'; // LCOV_EXCL_LINE
-            }
-
-            case value_t::number_float:
-                return get_ubjson_float_prefix(j.m_value.number_float);
-
-            case value_t::string:
-                return 'S';
-
-            case value_t::array: // fallthrough
-            case value_t::binary:
-                return '[';
-
-            case value_t::object:
-                return '{';
-
-            case value_t::discarded:
-            default:  // discarded values
-                return 'N';
-        }
-    }
-
-    static constexpr CharType get_ubjson_float_prefix(float /*unused*/)
-    {
-        return 'd';  // float 32
-    }
-
-    static constexpr CharType get_ubjson_float_prefix(double /*unused*/)
-    {
-        return 'D';  // float 64
-    }
-
-    /*!
-    @return false if the object is successfully converted to a bjdata ndarray, true if the type or size is invalid
-    */
-    bool write_bjdata_ndarray(const typename BasicJsonType::object_t& value, const bool use_count, const bool use_type)
-    {
-        std::map<string_t, CharType> bjdtype = {{"uint8", 'U'},  {"int8", 'i'},  {"uint16", 'u'}, {"int16", 'I'},
-            {"uint32", 'm'}, {"int32", 'l'}, {"uint64", 'M'}, {"int64", 'L'}, {"single", 'd'}, {"double", 'D'}, {"char", 'C'}
-        };
-
-        string_t key = "_ArrayType_";
-        auto it = bjdtype.find(static_cast<string_t>(value.at(key)));
-        if (it == bjdtype.end())
-        {
-            return true;
-        }
-        CharType dtype = it->second;
-
-        key = "_ArraySize_";
-        std::size_t len = (value.at(key).empty() ? 0 : 1);
-        for (const auto& el : value.at(key))
-        {
-            len *= static_cast<std::size_t>(el.m_value.number_unsigned);
-        }
-
-        key = "_ArrayData_";
-        if (value.at(key).size() != len)
-        {
-            return true;
-        }
-
-        oa->write_character('[');
-        oa->write_character('$');
-        oa->write_character(dtype);
-        oa->write_character('#');
-
-        key = "_ArraySize_";
-        write_ubjson(value.at(key), use_count, use_type, true,  true);
-
-        key = "_ArrayData_";
-        if (dtype == 'U' || dtype == 'C')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::uint8_t>(el.m_value.number_unsigned), true);
-            }
-        }
-        else if (dtype == 'i')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::int8_t>(el.m_value.number_integer), true);
-            }
-        }
-        else if (dtype == 'u')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::uint16_t>(el.m_value.number_unsigned), true);
-            }
-        }
-        else if (dtype == 'I')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::int16_t>(el.m_value.number_integer), true);
-            }
-        }
-        else if (dtype == 'm')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::uint32_t>(el.m_value.number_unsigned), true);
-            }
-        }
-        else if (dtype == 'l')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::int32_t>(el.m_value.number_integer), true);
-            }
-        }
-        else if (dtype == 'M')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::uint64_t>(el.m_value.number_unsigned), true);
-            }
-        }
-        else if (dtype == 'L')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<std::int64_t>(el.m_value.number_integer), true);
-            }
-        }
-        else if (dtype == 'd')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<float>(el.m_value.number_float), true);
-            }
-        }
-        else if (dtype == 'D')
-        {
-            for (const auto& el : value.at(key))
-            {
-                write_number(static_cast<double>(el.m_value.number_float), true);
-            }
-        }
-        return false;
-    }
-
-    ///////////////////////
-    // Utility functions //
-    ///////////////////////
-
-    /*
-    @brief write a number to output input
-    @param[in] n number of type @a NumberType
-    @param[in] OutputIsLittleEndian Set to true if output data is
-                                 required to be little endian
-    @tparam NumberType the type of the number
-
-    @note This function needs to respect the system's endianness, because bytes
-          in CBOR, MessagePack, and UBJSON are stored in network order (big
-          endian) and therefore need reordering on little endian systems.
-          On the other hand, BSON and BJData use little endian and should reorder
-          on big endian systems.
-    */
-    template<typename NumberType>
-    void write_number(const NumberType n, const bool OutputIsLittleEndian = false)
-    {
-        // step 1: write number to array of length NumberType
-        std::array<CharType, sizeof(NumberType)> vec{};
-        std::memcpy(vec.data(), &n, sizeof(NumberType));
-
-        // step 2: write array to output (with possible reordering)
-        if (is_little_endian != OutputIsLittleEndian)
-        {
-            // reverse byte order prior to conversion if necessary
-            std::reverse(vec.begin(), vec.end());
-        }
-
-        oa->write_characters(vec.data(), sizeof(NumberType));
-    }
-
-    void write_compact_float(const number_float_t n, detail::input_format_t format)
-    {
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wfloat-equal"
-#endif
-        if (static_cast<double>(n) >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
-                static_cast<double>(n) <= static_cast<double>((std::numeric_limits<float>::max)()) &&
-                static_cast<double>(static_cast<float>(n)) == static_cast<double>(n))
-        {
-            oa->write_character(format == detail::input_format_t::cbor
-                                ? get_cbor_float_prefix(static_cast<float>(n))
-                                : get_msgpack_float_prefix(static_cast<float>(n)));
-            write_number(static_cast<float>(n));
-        }
-        else
-        {
-            oa->write_character(format == detail::input_format_t::cbor
-                                ? get_cbor_float_prefix(n)
-                                : get_msgpack_float_prefix(n));
-            write_number(n);
-        }
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-    }
-
-  public:
-    // The following to_char_type functions are implement the conversion
-    // between uint8_t and CharType. In case CharType is not unsigned,
-    // such a conversion is required to allow values greater than 128.
-    // See <https://github.com/nlohmann/json/issues/1286> for a discussion.
-    template < typename C = CharType,
-               enable_if_t < std::is_signed<C>::value && std::is_signed<char>::value > * = nullptr >
-    static constexpr CharType to_char_type(std::uint8_t x) noexcept
-    {
-        return *reinterpret_cast<char*>(&x);
-    }
-
-    template < typename C = CharType,
-               enable_if_t < std::is_signed<C>::value && std::is_unsigned<char>::value > * = nullptr >
-    static CharType to_char_type(std::uint8_t x) noexcept
-    {
-        static_assert(sizeof(std::uint8_t) == sizeof(CharType), "size of CharType must be equal to std::uint8_t");
-        static_assert(std::is_trivial<CharType>::value, "CharType must be trivial");
-        CharType result;
-        std::memcpy(&result, &x, sizeof(x));
-        return result;
-    }
-
-    template<typename C = CharType,
-             enable_if_t<std::is_unsigned<C>::value>* = nullptr>
-    static constexpr CharType to_char_type(std::uint8_t x) noexcept
-    {
-        return x;
-    }
-
-    template < typename InputCharType, typename C = CharType,
-               enable_if_t <
-                   std::is_signed<C>::value &&
-                   std::is_signed<char>::value &&
-                   std::is_same<char, typename std::remove_cv<InputCharType>::type>::value
-                   > * = nullptr >
-    static constexpr CharType to_char_type(InputCharType x) noexcept
-    {
-        return x;
-    }
-
-  private:
-    /// whether we can assume little endianness
-    const bool is_little_endian = little_endianness();
-
-    /// the output
-    output_adapter_t<CharType> oa = nullptr;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/output/output_adapters.hpp>
-
-// #include <nlohmann/detail/output/serializer.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2008-2009 Björn Hoehrmann <bjoern@hoehrmann.de>
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <algorithm> // reverse, remove, fill, find, none_of
-#include <array> // array
-#include <clocale> // localeconv, lconv
-#include <cmath> // labs, isfinite, isnan, signbit
-#include <cstddef> // size_t, ptrdiff_t
-#include <cstdint> // uint8_t
-#include <cstdio> // snprintf
-#include <limits> // numeric_limits
-#include <string> // string, char_traits
-#include <iomanip> // setfill, setw
-#include <type_traits> // is_same
-#include <utility> // move
-
-// #include <nlohmann/detail/conversions/to_chars.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2009 Florian Loitsch <https://florian.loitsch.com/>
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <array> // array
-#include <cmath>   // signbit, isfinite
-#include <cstdint> // intN_t, uintN_t
-#include <cstring> // memcpy, memmove
-#include <limits> // numeric_limits
-#include <type_traits> // conditional
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-/*!
-@brief implements the Grisu2 algorithm for binary to decimal floating-point
-conversion.
-
-This implementation is a slightly modified version of the reference
-implementation which may be obtained from
-http://florian.loitsch.com/publications (bench.tar.gz).
-
-The code is distributed under the MIT license, Copyright (c) 2009 Florian Loitsch.
-
-For a detailed description of the algorithm see:
-
-[1] Loitsch, "Printing Floating-Point Numbers Quickly and Accurately with
-    Integers", Proceedings of the ACM SIGPLAN 2010 Conference on Programming
-    Language Design and Implementation, PLDI 2010
-[2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and Accurately",
-    Proceedings of the ACM SIGPLAN 1996 Conference on Programming Language
-    Design and Implementation, PLDI 1996
-*/
-namespace dtoa_impl
-{
-
-template<typename Target, typename Source>
-Target reinterpret_bits(const Source source)
-{
-    static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
-
-    Target target;
-    std::memcpy(&target, &source, sizeof(Source));
-    return target;
-}
-
-struct diyfp // f * 2^e
-{
-    static constexpr int kPrecision = 64; // = q
-
-    std::uint64_t f = 0;
-    int e = 0;
-
-    constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
-
-    /*!
-    @brief returns x - y
-    @pre x.e == y.e and x.f >= y.f
-    */
-    static diyfp sub(const diyfp& x, const diyfp& y) noexcept
-    {
-        JSON_ASSERT(x.e == y.e);
-        JSON_ASSERT(x.f >= y.f);
-
-        return {x.f - y.f, x.e};
-    }
-
-    /*!
-    @brief returns x * y
-    @note The result is rounded. (Only the upper q bits are returned.)
-    */
-    static diyfp mul(const diyfp& x, const diyfp& y) noexcept
-    {
-        static_assert(kPrecision == 64, "internal error");
-
-        // Computes:
-        //  f = round((x.f * y.f) / 2^q)
-        //  e = x.e + y.e + q
-
-        // Emulate the 64-bit * 64-bit multiplication:
-        //
-        // p = u * v
-        //   = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
-        //   = (u_lo v_lo         ) + 2^32 ((u_lo v_hi         ) + (u_hi v_lo         )) + 2^64 (u_hi v_hi         )
-        //   = (p0                ) + 2^32 ((p1                ) + (p2                )) + 2^64 (p3                )
-        //   = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3                )
-        //   = (p0_lo             ) + 2^32 (p0_hi + p1_lo + p2_lo                      ) + 2^64 (p1_hi + p2_hi + p3)
-        //   = (p0_lo             ) + 2^32 (Q                                          ) + 2^64 (H                 )
-        //   = (p0_lo             ) + 2^32 (Q_lo + 2^32 Q_hi                           ) + 2^64 (H                 )
-        //
-        // (Since Q might be larger than 2^32 - 1)
-        //
-        //   = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
-        //
-        // (Q_hi + H does not overflow a 64-bit int)
-        //
-        //   = p_lo + 2^64 p_hi
-
-        const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
-        const std::uint64_t u_hi = x.f >> 32u;
-        const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
-        const std::uint64_t v_hi = y.f >> 32u;
-
-        const std::uint64_t p0 = u_lo * v_lo;
-        const std::uint64_t p1 = u_lo * v_hi;
-        const std::uint64_t p2 = u_hi * v_lo;
-        const std::uint64_t p3 = u_hi * v_hi;
-
-        const std::uint64_t p0_hi = p0 >> 32u;
-        const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
-        const std::uint64_t p1_hi = p1 >> 32u;
-        const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
-        const std::uint64_t p2_hi = p2 >> 32u;
-
-        std::uint64_t Q = p0_hi + p1_lo + p2_lo;
-
-        // The full product might now be computed as
-        //
-        // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
-        // p_lo = p0_lo + (Q << 32)
-        //
-        // But in this particular case here, the full p_lo is not required.
-        // Effectively we only need to add the highest bit in p_lo to p_hi (and
-        // Q_hi + 1 does not overflow).
-
-        Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
-
-        const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
-
-        return {h, x.e + y.e + 64};
-    }
-
-    /*!
-    @brief normalize x such that the significand is >= 2^(q-1)
-    @pre x.f != 0
-    */
-    static diyfp normalize(diyfp x) noexcept
-    {
-        JSON_ASSERT(x.f != 0);
-
-        while ((x.f >> 63u) == 0)
-        {
-            x.f <<= 1u;
-            x.e--;
-        }
-
-        return x;
-    }
-
-    /*!
-    @brief normalize x such that the result has the exponent E
-    @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
-    */
-    static diyfp normalize_to(const diyfp& x, const int target_exponent) noexcept
-    {
-        const int delta = x.e - target_exponent;
-
-        JSON_ASSERT(delta >= 0);
-        JSON_ASSERT(((x.f << delta) >> delta) == x.f);
-
-        return {x.f << delta, target_exponent};
-    }
-};
-
-struct boundaries
-{
-    diyfp w;
-    diyfp minus;
-    diyfp plus;
-};
-
-/*!
-Compute the (normalized) diyfp representing the input number 'value' and its
-boundaries.
-
-@pre value must be finite and positive
-*/
-template<typename FloatType>
-boundaries compute_boundaries(FloatType value)
-{
-    JSON_ASSERT(std::isfinite(value));
-    JSON_ASSERT(value > 0);
-
-    // Convert the IEEE representation into a diyfp.
-    //
-    // If v is denormal:
-    //      value = 0.F * 2^(1 - bias) = (          F) * 2^(1 - bias - (p-1))
-    // If v is normalized:
-    //      value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
-
-    static_assert(std::numeric_limits<FloatType>::is_iec559,
-                  "internal error: dtoa_short requires an IEEE-754 floating-point implementation");
-
-    constexpr int      kPrecision = std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
-    constexpr int      kBias      = std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
-    constexpr int      kMinExp    = 1 - kBias;
-    constexpr std::uint64_t kHiddenBit = std::uint64_t{1} << (kPrecision - 1); // = 2^(p-1)
-
-    using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, std::uint64_t >::type;
-
-    const auto bits = static_cast<std::uint64_t>(reinterpret_bits<bits_type>(value));
-    const std::uint64_t E = bits >> (kPrecision - 1);
-    const std::uint64_t F = bits & (kHiddenBit - 1);
-
-    const bool is_denormal = E == 0;
-    const diyfp v = is_denormal
-                    ? diyfp(F, kMinExp)
-                    : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
-
-    // Compute the boundaries m- and m+ of the floating-point value
-    // v = f * 2^e.
-    //
-    // Determine v- and v+, the floating-point predecessor and successor if v,
-    // respectively.
-    //
-    //      v- = v - 2^e        if f != 2^(p-1) or e == e_min                (A)
-    //         = v - 2^(e-1)    if f == 2^(p-1) and e > e_min                (B)
-    //
-    //      v+ = v + 2^e
-    //
-    // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
-    // between m- and m+ round to v, regardless of how the input rounding
-    // algorithm breaks ties.
-    //
-    //      ---+-------------+-------------+-------------+-------------+---  (A)
-    //         v-            m-            v             m+            v+
-    //
-    //      -----------------+------+------+-------------+-------------+---  (B)
-    //                       v-     m-     v             m+            v+
-
-    const bool lower_boundary_is_closer = F == 0 && E > 1;
-    const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
-    const diyfp m_minus = lower_boundary_is_closer
-                          ? diyfp(4 * v.f - 1, v.e - 2)  // (B)
-                          : diyfp(2 * v.f - 1, v.e - 1); // (A)
-
-    // Determine the normalized w+ = m+.
-    const diyfp w_plus = diyfp::normalize(m_plus);
-
-    // Determine w- = m- such that e_(w-) = e_(w+).
-    const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
-
-    return {diyfp::normalize(v), w_minus, w_plus};
-}
-
-// Given normalized diyfp w, Grisu needs to find a (normalized) cached
-// power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
-// within a certain range [alpha, gamma] (Definition 3.2 from [1])
-//
-//      alpha <= e = e_c + e_w + q <= gamma
-//
-// or
-//
-//      f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
-//                          <= f_c * f_w * 2^gamma
-//
-// Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
-//
-//      2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
-//
-// or
-//
-//      2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
-//
-// The choice of (alpha,gamma) determines the size of the table and the form of
-// the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
-// in practice:
-//
-// The idea is to cut the number c * w = f * 2^e into two parts, which can be
-// processed independently: An integral part p1, and a fractional part p2:
-//
-//      f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
-//              = (f div 2^-e) + (f mod 2^-e) * 2^e
-//              = p1 + p2 * 2^e
-//
-// The conversion of p1 into decimal form requires a series of divisions and
-// modulos by (a power of) 10. These operations are faster for 32-bit than for
-// 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
-// achieved by choosing
-//
-//      -e >= 32   or   e <= -32 := gamma
-//
-// In order to convert the fractional part
-//
-//      p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
-//
-// into decimal form, the fraction is repeatedly multiplied by 10 and the digits
-// d[-i] are extracted in order:
-//
-//      (10 * p2) div 2^-e = d[-1]
-//      (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
-//
-// The multiplication by 10 must not overflow. It is sufficient to choose
-//
-//      10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
-//
-// Since p2 = f mod 2^-e < 2^-e,
-//
-//      -e <= 60   or   e >= -60 := alpha
-
-constexpr int kAlpha = -60;
-constexpr int kGamma = -32;
-
-struct cached_power // c = f * 2^e ~= 10^k
-{
-    std::uint64_t f;
-    int e;
-    int k;
-};
-
-/*!
-For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
-power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
-satisfies (Definition 3.2 from [1])
-
-     alpha <= e_c + e + q <= gamma.
-*/
-inline cached_power get_cached_power_for_binary_exponent(int e)
-{
-    // Now
-    //
-    //      alpha <= e_c + e + q <= gamma                                    (1)
-    //      ==> f_c * 2^alpha <= c * 2^e * 2^q
-    //
-    // and since the c's are normalized, 2^(q-1) <= f_c,
-    //
-    //      ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
-    //      ==> 2^(alpha - e - 1) <= c
-    //
-    // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
-    //
-    //      k = ceil( log_10( 2^(alpha - e - 1) ) )
-    //        = ceil( (alpha - e - 1) * log_10(2) )
-    //
-    // From the paper:
-    // "In theory the result of the procedure could be wrong since c is rounded,
-    //  and the computation itself is approximated [...]. In practice, however,
-    //  this simple function is sufficient."
-    //
-    // For IEEE double precision floating-point numbers converted into
-    // normalized diyfp's w = f * 2^e, with q = 64,
-    //
-    //      e >= -1022      (min IEEE exponent)
-    //           -52        (p - 1)
-    //           -52        (p - 1, possibly normalize denormal IEEE numbers)
-    //           -11        (normalize the diyfp)
-    //         = -1137
-    //
-    // and
-    //
-    //      e <= +1023      (max IEEE exponent)
-    //           -52        (p - 1)
-    //           -11        (normalize the diyfp)
-    //         = 960
-    //
-    // This binary exponent range [-1137,960] results in a decimal exponent
-    // range [-307,324]. One does not need to store a cached power for each
-    // k in this range. For each such k it suffices to find a cached power
-    // such that the exponent of the product lies in [alpha,gamma].
-    // This implies that the difference of the decimal exponents of adjacent
-    // table entries must be less than or equal to
-    //
-    //      floor( (gamma - alpha) * log_10(2) ) = 8.
-    //
-    // (A smaller distance gamma-alpha would require a larger table.)
-
-    // NB:
-    // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
-
-    constexpr int kCachedPowersMinDecExp = -300;
-    constexpr int kCachedPowersDecStep = 8;
-
-    static constexpr std::array<cached_power, 79> kCachedPowers =
-    {
-        {
-            { 0xAB70FE17C79AC6CA, -1060, -300 },
-            { 0xFF77B1FCBEBCDC4F, -1034, -292 },
-            { 0xBE5691EF416BD60C, -1007, -284 },
-            { 0x8DD01FAD907FFC3C,  -980, -276 },
-            { 0xD3515C2831559A83,  -954, -268 },
-            { 0x9D71AC8FADA6C9B5,  -927, -260 },
-            { 0xEA9C227723EE8BCB,  -901, -252 },
-            { 0xAECC49914078536D,  -874, -244 },
-            { 0x823C12795DB6CE57,  -847, -236 },
-            { 0xC21094364DFB5637,  -821, -228 },
-            { 0x9096EA6F3848984F,  -794, -220 },
-            { 0xD77485CB25823AC7,  -768, -212 },
-            { 0xA086CFCD97BF97F4,  -741, -204 },
-            { 0xEF340A98172AACE5,  -715, -196 },
-            { 0xB23867FB2A35B28E,  -688, -188 },
-            { 0x84C8D4DFD2C63F3B,  -661, -180 },
-            { 0xC5DD44271AD3CDBA,  -635, -172 },
-            { 0x936B9FCEBB25C996,  -608, -164 },
-            { 0xDBAC6C247D62A584,  -582, -156 },
-            { 0xA3AB66580D5FDAF6,  -555, -148 },
-            { 0xF3E2F893DEC3F126,  -529, -140 },
-            { 0xB5B5ADA8AAFF80B8,  -502, -132 },
-            { 0x87625F056C7C4A8B,  -475, -124 },
-            { 0xC9BCFF6034C13053,  -449, -116 },
-            { 0x964E858C91BA2655,  -422, -108 },
-            { 0xDFF9772470297EBD,  -396, -100 },
-            { 0xA6DFBD9FB8E5B88F,  -369,  -92 },
-            { 0xF8A95FCF88747D94,  -343,  -84 },
-            { 0xB94470938FA89BCF,  -316,  -76 },
-            { 0x8A08F0F8BF0F156B,  -289,  -68 },
-            { 0xCDB02555653131B6,  -263,  -60 },
-            { 0x993FE2C6D07B7FAC,  -236,  -52 },
-            { 0xE45C10C42A2B3B06,  -210,  -44 },
-            { 0xAA242499697392D3,  -183,  -36 },
-            { 0xFD87B5F28300CA0E,  -157,  -28 },
-            { 0xBCE5086492111AEB,  -130,  -20 },
-            { 0x8CBCCC096F5088CC,  -103,  -12 },
-            { 0xD1B71758E219652C,   -77,   -4 },
-            { 0x9C40000000000000,   -50,    4 },
-            { 0xE8D4A51000000000,   -24,   12 },
-            { 0xAD78EBC5AC620000,     3,   20 },
-            { 0x813F3978F8940984,    30,   28 },
-            { 0xC097CE7BC90715B3,    56,   36 },
-            { 0x8F7E32CE7BEA5C70,    83,   44 },
-            { 0xD5D238A4ABE98068,   109,   52 },
-            { 0x9F4F2726179A2245,   136,   60 },
-            { 0xED63A231D4C4FB27,   162,   68 },
-            { 0xB0DE65388CC8ADA8,   189,   76 },
-            { 0x83C7088E1AAB65DB,   216,   84 },
-            { 0xC45D1DF942711D9A,   242,   92 },
-            { 0x924D692CA61BE758,   269,  100 },
-            { 0xDA01EE641A708DEA,   295,  108 },
-            { 0xA26DA3999AEF774A,   322,  116 },
-            { 0xF209787BB47D6B85,   348,  124 },
-            { 0xB454E4A179DD1877,   375,  132 },
-            { 0x865B86925B9BC5C2,   402,  140 },
-            { 0xC83553C5C8965D3D,   428,  148 },
-            { 0x952AB45CFA97A0B3,   455,  156 },
-            { 0xDE469FBD99A05FE3,   481,  164 },
-            { 0xA59BC234DB398C25,   508,  172 },
-            { 0xF6C69A72A3989F5C,   534,  180 },
-            { 0xB7DCBF5354E9BECE,   561,  188 },
-            { 0x88FCF317F22241E2,   588,  196 },
-            { 0xCC20CE9BD35C78A5,   614,  204 },
-            { 0x98165AF37B2153DF,   641,  212 },
-            { 0xE2A0B5DC971F303A,   667,  220 },
-            { 0xA8D9D1535CE3B396,   694,  228 },
-            { 0xFB9B7CD9A4A7443C,   720,  236 },
-            { 0xBB764C4CA7A44410,   747,  244 },
-            { 0x8BAB8EEFB6409C1A,   774,  252 },
-            { 0xD01FEF10A657842C,   800,  260 },
-            { 0x9B10A4E5E9913129,   827,  268 },
-            { 0xE7109BFBA19C0C9D,   853,  276 },
-            { 0xAC2820D9623BF429,   880,  284 },
-            { 0x80444B5E7AA7CF85,   907,  292 },
-            { 0xBF21E44003ACDD2D,   933,  300 },
-            { 0x8E679C2F5E44FF8F,   960,  308 },
-            { 0xD433179D9C8CB841,   986,  316 },
-            { 0x9E19DB92B4E31BA9,  1013,  324 },
-        }
-    };
-
-    // This computation gives exactly the same results for k as
-    //      k = ceil((kAlpha - e - 1) * 0.30102999566398114)
-    // for |e| <= 1500, but doesn't require floating-point operations.
-    // NB: log_10(2) ~= 78913 / 2^18
-    JSON_ASSERT(e >= -1500);
-    JSON_ASSERT(e <=  1500);
-    const int f = kAlpha - e - 1;
-    const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
-
-    const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / kCachedPowersDecStep;
-    JSON_ASSERT(index >= 0);
-    JSON_ASSERT(static_cast<std::size_t>(index) < kCachedPowers.size());
-
-    const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
-    JSON_ASSERT(kAlpha <= cached.e + e + 64);
-    JSON_ASSERT(kGamma >= cached.e + e + 64);
-
-    return cached;
-}
-
-/*!
-For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
-For n == 0, returns 1 and sets pow10 := 1.
-*/
-inline int find_largest_pow10(const std::uint32_t n, std::uint32_t& pow10)
-{
-    // LCOV_EXCL_START
-    if (n >= 1000000000)
-    {
-        pow10 = 1000000000;
-        return 10;
-    }
-    // LCOV_EXCL_STOP
-    if (n >= 100000000)
-    {
-        pow10 = 100000000;
-        return  9;
-    }
-    if (n >= 10000000)
-    {
-        pow10 = 10000000;
-        return  8;
-    }
-    if (n >= 1000000)
-    {
-        pow10 = 1000000;
-        return  7;
-    }
-    if (n >= 100000)
-    {
-        pow10 = 100000;
-        return  6;
-    }
-    if (n >= 10000)
-    {
-        pow10 = 10000;
-        return  5;
-    }
-    if (n >= 1000)
-    {
-        pow10 = 1000;
-        return  4;
-    }
-    if (n >= 100)
-    {
-        pow10 = 100;
-        return  3;
-    }
-    if (n >= 10)
-    {
-        pow10 = 10;
-        return  2;
-    }
-
-    pow10 = 1;
-    return 1;
-}
-
-inline void grisu2_round(char* buf, int len, std::uint64_t dist, std::uint64_t delta,
-                         std::uint64_t rest, std::uint64_t ten_k)
-{
-    JSON_ASSERT(len >= 1);
-    JSON_ASSERT(dist <= delta);
-    JSON_ASSERT(rest <= delta);
-    JSON_ASSERT(ten_k > 0);
-
-    //               <--------------------------- delta ---->
-    //                                  <---- dist --------->
-    // --------------[------------------+-------------------]--------------
-    //               M-                 w                   M+
-    //
-    //                                  ten_k
-    //                                <------>
-    //                                       <---- rest ---->
-    // --------------[------------------+----+--------------]--------------
-    //                                  w    V
-    //                                       = buf * 10^k
-    //
-    // ten_k represents a unit-in-the-last-place in the decimal representation
-    // stored in buf.
-    // Decrement buf by ten_k while this takes buf closer to w.
-
-    // The tests are written in this order to avoid overflow in unsigned
-    // integer arithmetic.
-
-    while (rest < dist
-            && delta - rest >= ten_k
-            && (rest + ten_k < dist || dist - rest > rest + ten_k - dist))
-    {
-        JSON_ASSERT(buf[len - 1] != '0');
-        buf[len - 1]--;
-        rest += ten_k;
-    }
-}
-
-/*!
-Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
-M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
-*/
-inline void grisu2_digit_gen(char* buffer, int& length, int& decimal_exponent,
-                             diyfp M_minus, diyfp w, diyfp M_plus)
-{
-    static_assert(kAlpha >= -60, "internal error");
-    static_assert(kGamma <= -32, "internal error");
-
-    // Generates the digits (and the exponent) of a decimal floating-point
-    // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
-    // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= gamma.
-    //
-    //               <--------------------------- delta ---->
-    //                                  <---- dist --------->
-    // --------------[------------------+-------------------]--------------
-    //               M-                 w                   M+
-    //
-    // Grisu2 generates the digits of M+ from left to right and stops as soon as
-    // V is in [M-,M+].
-
-    JSON_ASSERT(M_plus.e >= kAlpha);
-    JSON_ASSERT(M_plus.e <= kGamma);
-
-    std::uint64_t delta = diyfp::sub(M_plus, M_minus).f; // (significand of (M+ - M-), implicit exponent is e)
-    std::uint64_t dist  = diyfp::sub(M_plus, w      ).f; // (significand of (M+ - w ), implicit exponent is e)
-
-    // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
-    //
-    //      M+ = f * 2^e
-    //         = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
-    //         = ((p1        ) * 2^-e + (p2        )) * 2^e
-    //         = p1 + p2 * 2^e
-
-    const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
-
-    auto p1 = static_cast<std::uint32_t>(M_plus.f >> -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
-    std::uint64_t p2 = M_plus.f & (one.f - 1);                    // p2 = f mod 2^-e
-
-    // 1)
-    //
-    // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
-
-    JSON_ASSERT(p1 > 0);
-
-    std::uint32_t pow10{};
-    const int k = find_largest_pow10(p1, pow10);
-
-    //      10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
-    //
-    //      p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
-    //         = (d[k-1]         ) * 10^(k-1) + (p1 mod 10^(k-1))
-    //
-    //      M+ = p1                                             + p2 * 2^e
-    //         = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1))          + p2 * 2^e
-    //         = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
-    //         = d[k-1] * 10^(k-1) + (                         rest) * 2^e
-    //
-    // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
-    //
-    //      p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
-    //
-    // but stop as soon as
-    //
-    //      rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
-
-    int n = k;
-    while (n > 0)
-    {
-        // Invariants:
-        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)    (buffer = 0 for n = k)
-        //      pow10 = 10^(n-1) <= p1 < 10^n
-        //
-        const std::uint32_t d = p1 / pow10;  // d = p1 div 10^(n-1)
-        const std::uint32_t r = p1 % pow10;  // r = p1 mod 10^(n-1)
-        //
-        //      M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
-        //         = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
-        //
-        JSON_ASSERT(d <= 9);
-        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
-        //
-        //      M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
-        //
-        p1 = r;
-        n--;
-        //
-        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)
-        //      pow10 = 10^n
-        //
-
-        // Now check if enough digits have been generated.
-        // Compute
-        //
-        //      p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
-        //
-        // Note:
-        // Since rest and delta share the same exponent e, it suffices to
-        // compare the significands.
-        const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
-        if (rest <= delta)
-        {
-            // V = buffer * 10^n, with M- <= V <= M+.
-
-            decimal_exponent += n;
-
-            // We may now just stop. But instead look if the buffer could be
-            // decremented to bring V closer to w.
-            //
-            // pow10 = 10^n is now 1 ulp in the decimal representation V.
-            // The rounding procedure works with diyfp's with an implicit
-            // exponent of e.
-            //
-            //      10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
-            //
-            const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
-            grisu2_round(buffer, length, dist, delta, rest, ten_n);
-
-            return;
-        }
-
-        pow10 /= 10;
-        //
-        //      pow10 = 10^(n-1) <= p1 < 10^n
-        // Invariants restored.
-    }
-
-    // 2)
-    //
-    // The digits of the integral part have been generated:
-    //
-    //      M+ = d[k-1]...d[1]d[0] + p2 * 2^e
-    //         = buffer            + p2 * 2^e
-    //
-    // Now generate the digits of the fractional part p2 * 2^e.
-    //
-    // Note:
-    // No decimal point is generated: the exponent is adjusted instead.
-    //
-    // p2 actually represents the fraction
-    //
-    //      p2 * 2^e
-    //          = p2 / 2^-e
-    //          = d[-1] / 10^1 + d[-2] / 10^2 + ...
-    //
-    // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
-    //
-    //      p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
-    //                      + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
-    //
-    // using
-    //
-    //      10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
-    //                = (                   d) * 2^-e + (                   r)
-    //
-    // or
-    //      10^m * p2 * 2^e = d + r * 2^e
-    //
-    // i.e.
-    //
-    //      M+ = buffer + p2 * 2^e
-    //         = buffer + 10^-m * (d + r * 2^e)
-    //         = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
-    //
-    // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
-
-    JSON_ASSERT(p2 > delta);
-
-    int m = 0;
-    for (;;)
-    {
-        // Invariant:
-        //      M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) * 2^e
-        //         = buffer * 10^-m + 10^-m * (p2                                 ) * 2^e
-        //         = buffer * 10^-m + 10^-m * (1/10 * (10 * p2)                   ) * 2^e
-        //         = buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + (10*p2 mod 2^-e)) * 2^e
-        //
-        JSON_ASSERT(p2 <= (std::numeric_limits<std::uint64_t>::max)() / 10);
-        p2 *= 10;
-        const std::uint64_t d = p2 >> -one.e;     // d = (10 * p2) div 2^-e
-        const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
-        //
-        //      M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
-        //         = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
-        //         = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
-        //
-        JSON_ASSERT(d <= 9);
-        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
-        //
-        //      M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
-        //
-        p2 = r;
-        m++;
-        //
-        //      M+ = buffer * 10^-m + 10^-m * p2 * 2^e
-        // Invariant restored.
-
-        // Check if enough digits have been generated.
-        //
-        //      10^-m * p2 * 2^e <= delta * 2^e
-        //              p2 * 2^e <= 10^m * delta * 2^e
-        //                    p2 <= 10^m * delta
-        delta *= 10;
-        dist  *= 10;
-        if (p2 <= delta)
-        {
-            break;
-        }
-    }
-
-    // V = buffer * 10^-m, with M- <= V <= M+.
-
-    decimal_exponent -= m;
-
-    // 1 ulp in the decimal representation is now 10^-m.
-    // Since delta and dist are now scaled by 10^m, we need to do the
-    // same with ulp in order to keep the units in sync.
-    //
-    //      10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
-    //
-    const std::uint64_t ten_m = one.f;
-    grisu2_round(buffer, length, dist, delta, p2, ten_m);
-
-    // By construction this algorithm generates the shortest possible decimal
-    // number (Loitsch, Theorem 6.2) which rounds back to w.
-    // For an input number of precision p, at least
-    //
-    //      N = 1 + ceil(p * log_10(2))
-    //
-    // decimal digits are sufficient to identify all binary floating-point
-    // numbers (Matula, "In-and-Out conversions").
-    // This implies that the algorithm does not produce more than N decimal
-    // digits.
-    //
-    //      N = 17 for p = 53 (IEEE double precision)
-    //      N = 9  for p = 24 (IEEE single precision)
-}
-
-/*!
-v = buf * 10^decimal_exponent
-len is the length of the buffer (number of decimal digits)
-The buffer must be large enough, i.e. >= max_digits10.
-*/
-JSON_HEDLEY_NON_NULL(1)
-inline void grisu2(char* buf, int& len, int& decimal_exponent,
-                   diyfp m_minus, diyfp v, diyfp m_plus)
-{
-    JSON_ASSERT(m_plus.e == m_minus.e);
-    JSON_ASSERT(m_plus.e == v.e);
-
-    //  --------(-----------------------+-----------------------)--------    (A)
-    //          m-                      v                       m+
-    //
-    //  --------------------(-----------+-----------------------)--------    (B)
-    //                      m-          v                       m+
-    //
-    // First scale v (and m- and m+) such that the exponent is in the range
-    // [alpha, gamma].
-
-    const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
-
-    const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
-
-    // The exponent of the products is = v.e + c_minus_k.e + q and is in the range [alpha,gamma]
-    const diyfp w       = diyfp::mul(v,       c_minus_k);
-    const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
-    const diyfp w_plus  = diyfp::mul(m_plus,  c_minus_k);
-
-    //  ----(---+---)---------------(---+---)---------------(---+---)----
-    //          w-                      w                       w+
-    //          = c*m-                  = c*v                   = c*m+
-    //
-    // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
-    // w+ are now off by a small amount.
-    // In fact:
-    //
-    //      w - v * 10^k < 1 ulp
-    //
-    // To account for this inaccuracy, add resp. subtract 1 ulp.
-    //
-    //  --------+---[---------------(---+---)---------------]---+--------
-    //          w-  M-                  w                   M+  w+
-    //
-    // Now any number in [M-, M+] (bounds included) will round to w when input,
-    // regardless of how the input rounding algorithm breaks ties.
-    //
-    // And digit_gen generates the shortest possible such number in [M-, M+].
-    // Note that this does not mean that Grisu2 always generates the shortest
-    // possible number in the interval (m-, m+).
-    const diyfp M_minus(w_minus.f + 1, w_minus.e);
-    const diyfp M_plus (w_plus.f  - 1, w_plus.e );
-
-    decimal_exponent = -cached.k; // = -(-k) = k
-
-    grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
-}
-
-/*!
-v = buf * 10^decimal_exponent
-len is the length of the buffer (number of decimal digits)
-The buffer must be large enough, i.e. >= max_digits10.
-*/
-template<typename FloatType>
-JSON_HEDLEY_NON_NULL(1)
-void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
-{
-    static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
-                  "internal error: not enough precision");
-
-    JSON_ASSERT(std::isfinite(value));
-    JSON_ASSERT(value > 0);
-
-    // If the neighbors (and boundaries) of 'value' are always computed for double-precision
-    // numbers, all float's can be recovered using strtod (and strtof). However, the resulting
-    // decimal representations are not exactly "short".
-    //
-    // The documentation for 'std::to_chars' (https://en.cppreference.com/w/cpp/utility/to_chars)
-    // says "value is converted to a string as if by std::sprintf in the default ("C") locale"
-    // and since sprintf promotes floats to doubles, I think this is exactly what 'std::to_chars'
-    // does.
-    // On the other hand, the documentation for 'std::to_chars' requires that "parsing the
-    // representation using the corresponding std::from_chars function recovers value exactly". That
-    // indicates that single precision floating-point numbers should be recovered using
-    // 'std::strtof'.
-    //
-    // NB: If the neighbors are computed for single-precision numbers, there is a single float
-    //     (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision
-    //     value is off by 1 ulp.
-#if 0
-    const boundaries w = compute_boundaries(static_cast<double>(value));
-#else
-    const boundaries w = compute_boundaries(value);
-#endif
-
-    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
-}
-
-/*!
-@brief appends a decimal representation of e to buf
-@return a pointer to the element following the exponent.
-@pre -1000 < e < 1000
-*/
-JSON_HEDLEY_NON_NULL(1)
-JSON_HEDLEY_RETURNS_NON_NULL
-inline char* append_exponent(char* buf, int e)
-{
-    JSON_ASSERT(e > -1000);
-    JSON_ASSERT(e <  1000);
-
-    if (e < 0)
-    {
-        e = -e;
-        *buf++ = '-';
-    }
-    else
-    {
-        *buf++ = '+';
-    }
-
-    auto k = static_cast<std::uint32_t>(e);
-    if (k < 10)
-    {
-        // Always print at least two digits in the exponent.
-        // This is for compatibility with printf("%g").
-        *buf++ = '0';
-        *buf++ = static_cast<char>('0' + k);
-    }
-    else if (k < 100)
-    {
-        *buf++ = static_cast<char>('0' + k / 10);
-        k %= 10;
-        *buf++ = static_cast<char>('0' + k);
-    }
-    else
-    {
-        *buf++ = static_cast<char>('0' + k / 100);
-        k %= 100;
-        *buf++ = static_cast<char>('0' + k / 10);
-        k %= 10;
-        *buf++ = static_cast<char>('0' + k);
-    }
-
-    return buf;
-}
-
-/*!
-@brief prettify v = buf * 10^decimal_exponent
-
-If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
-notation. Otherwise it will be printed in exponential notation.
-
-@pre min_exp < 0
-@pre max_exp > 0
-*/
-JSON_HEDLEY_NON_NULL(1)
-JSON_HEDLEY_RETURNS_NON_NULL
-inline char* format_buffer(char* buf, int len, int decimal_exponent,
-                           int min_exp, int max_exp)
-{
-    JSON_ASSERT(min_exp < 0);
-    JSON_ASSERT(max_exp > 0);
-
-    const int k = len;
-    const int n = len + decimal_exponent;
-
-    // v = buf * 10^(n-k)
-    // k is the length of the buffer (number of decimal digits)
-    // n is the position of the decimal point relative to the start of the buffer.
-
-    if (k <= n && n <= max_exp)
-    {
-        // digits[000]
-        // len <= max_exp + 2
-
-        std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
-        // Make it look like a floating-point number (#362, #378)
-        buf[n + 0] = '.';
-        buf[n + 1] = '0';
-        return buf + (static_cast<size_t>(n) + 2);
-    }
-
-    if (0 < n && n <= max_exp)
-    {
-        // dig.its
-        // len <= max_digits10 + 1
-
-        JSON_ASSERT(k > n);
-
-        std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n, static_cast<size_t>(k) - static_cast<size_t>(n));
-        buf[n] = '.';
-        return buf + (static_cast<size_t>(k) + 1U);
-    }
-
-    if (min_exp < n && n <= 0)
-    {
-        // 0.[000]digits
-        // len <= 2 + (-min_exp - 1) + max_digits10
-
-        std::memmove(buf + (2 + static_cast<size_t>(-n)), buf, static_cast<size_t>(k));
-        buf[0] = '0';
-        buf[1] = '.';
-        std::memset(buf + 2, '0', static_cast<size_t>(-n));
-        return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
-    }
-
-    if (k == 1)
-    {
-        // dE+123
-        // len <= 1 + 5
-
-        buf += 1;
-    }
-    else
-    {
-        // d.igitsE+123
-        // len <= max_digits10 + 1 + 5
-
-        std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
-        buf[1] = '.';
-        buf += 1 + static_cast<size_t>(k);
-    }
-
-    *buf++ = 'e';
-    return append_exponent(buf, n - 1);
-}
-
-}  // namespace dtoa_impl
-
-/*!
-@brief generates a decimal representation of the floating-point number value in [first, last).
-
-The format of the resulting decimal representation is similar to printf's %g
-format. Returns an iterator pointing past-the-end of the decimal representation.
-
-@note The input number must be finite, i.e. NaN's and Inf's are not supported.
-@note The buffer must be large enough.
-@note The result is NOT null-terminated.
-*/
-template<typename FloatType>
-JSON_HEDLEY_NON_NULL(1, 2)
-JSON_HEDLEY_RETURNS_NON_NULL
-char* to_chars(char* first, const char* last, FloatType value)
-{
-    static_cast<void>(last); // maybe unused - fix warning
-    JSON_ASSERT(std::isfinite(value));
-
-    // Use signbit(value) instead of (value < 0) since signbit works for -0.
-    if (std::signbit(value))
-    {
-        value = -value;
-        *first++ = '-';
-    }
-
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wfloat-equal"
-#endif
-    if (value == 0) // +-0
-    {
-        *first++ = '0';
-        // Make it look like a floating-point number (#362, #378)
-        *first++ = '.';
-        *first++ = '0';
-        return first;
-    }
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-
-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
-
-    // Compute v = buffer * 10^decimal_exponent.
-    // The decimal digits are stored in the buffer, which needs to be interpreted
-    // as an unsigned decimal integer.
-    // len is the length of the buffer, i.e. the number of decimal digits.
-    int len = 0;
-    int decimal_exponent = 0;
-    dtoa_impl::grisu2(first, len, decimal_exponent, value);
-
-    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
-
-    // Format the buffer like printf("%.*g", prec, value)
-    constexpr int kMinExp = -4;
-    // Use digits10 here to increase compatibility with version 2.
-    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
-
-    JSON_ASSERT(last - first >= kMaxExp + 2);
-    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
-
-    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
-}
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/exceptions.hpp>
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/cpp_future.hpp>
-
-// #include <nlohmann/detail/output/binary_writer.hpp>
-
-// #include <nlohmann/detail/output/output_adapters.hpp>
-
-// #include <nlohmann/detail/string_concat.hpp>
-
-// #include <nlohmann/detail/value_t.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-namespace detail
-{
-
-///////////////////
-// serialization //
-///////////////////
-
-/// how to treat decoding errors
-enum class error_handler_t
-{
-    strict,  ///< throw a type_error exception in case of invalid UTF-8
-    replace, ///< replace invalid UTF-8 sequences with U+FFFD
-    ignore   ///< ignore invalid UTF-8 sequences
-};
-
-template<typename BasicJsonType>
-class serializer
-{
-    using string_t = typename BasicJsonType::string_t;
-    using number_float_t = typename BasicJsonType::number_float_t;
-    using number_integer_t = typename BasicJsonType::number_integer_t;
-    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-    using binary_char_t = typename BasicJsonType::binary_t::value_type;
-    static constexpr std::uint8_t UTF8_ACCEPT = 0;
-    static constexpr std::uint8_t UTF8_REJECT = 1;
-
-  public:
-    /*!
-    @param[in] s  output stream to serialize to
-    @param[in] ichar  indentation character to use
-    @param[in] error_handler_  how to react on decoding errors
-    */
-    serializer(output_adapter_t<char> s, const char ichar,
-               error_handler_t error_handler_ = error_handler_t::strict)
-        : o(std::move(s))
-        , loc(std::localeconv())
-        , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
-        , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
-        , indent_char(ichar)
-        , indent_string(512, indent_char)
-        , error_handler(error_handler_)
-    {}
-
-    // delete because of pointer members
-    serializer(const serializer&) = delete;
-    serializer& operator=(const serializer&) = delete;
-    serializer(serializer&&) = delete;
-    serializer& operator=(serializer&&) = delete;
-    ~serializer() = default;
-
-    /*!
-    @brief internal implementation of the serialization function
-
-    This function is called by the public member function dump and organizes
-    the serialization internally. The indentation level is propagated as
-    additional parameter. In case of arrays and objects, the function is
-    called recursively.
-
-    - strings and object keys are escaped using `escape_string()`
-    - integer numbers are converted implicitly via `operator<<`
-    - floating-point numbers are converted to a string using `"%g"` format
-    - binary values are serialized as objects containing the subtype and the
-      byte array
-
-    @param[in] val               value to serialize
-    @param[in] pretty_print      whether the output shall be pretty-printed
-    @param[in] ensure_ascii If @a ensure_ascii is true, all non-ASCII characters
-    in the output are escaped with `\uXXXX` sequences, and the result consists
-    of ASCII characters only.
-    @param[in] indent_step       the indent level
-    @param[in] current_indent    the current indent level (only used internally)
-    */
-    void dump(const BasicJsonType& val,
-              const bool pretty_print,
-              const bool ensure_ascii,
-              const unsigned int indent_step,
-              const unsigned int current_indent = 0)
-    {
-        switch (val.m_type)
-        {
-            case value_t::object:
-            {
-                if (val.m_value.object->empty())
-                {
-                    o->write_characters("{}", 2);
-                    return;
-                }
-
-                if (pretty_print)
-                {
-                    o->write_characters("{\n", 2);
-
-                    // variable to hold indentation for recursive calls
-                    const auto new_indent = current_indent + indent_step;
-                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
-                    {
-                        indent_string.resize(indent_string.size() * 2, ' ');
-                    }
-
-                    // first n-1 elements
-                    auto i = val.m_value.object->cbegin();
-                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
-                    {
-                        o->write_characters(indent_string.c_str(), new_indent);
-                        o->write_character('\"');
-                        dump_escaped(i->first, ensure_ascii);
-                        o->write_characters("\": ", 3);
-                        dump(i->second, true, ensure_ascii, indent_step, new_indent);
-                        o->write_characters(",\n", 2);
-                    }
-
-                    // last element
-                    JSON_ASSERT(i != val.m_value.object->cend());
-                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
-                    o->write_characters(indent_string.c_str(), new_indent);
-                    o->write_character('\"');
-                    dump_escaped(i->first, ensure_ascii);
-                    o->write_characters("\": ", 3);
-                    dump(i->second, true, ensure_ascii, indent_step, new_indent);
-
-                    o->write_character('\n');
-                    o->write_characters(indent_string.c_str(), current_indent);
-                    o->write_character('}');
-                }
-                else
-                {
-                    o->write_character('{');
-
-                    // first n-1 elements
-                    auto i = val.m_value.object->cbegin();
-                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
-                    {
-                        o->write_character('\"');
-                        dump_escaped(i->first, ensure_ascii);
-                        o->write_characters("\":", 2);
-                        dump(i->second, false, ensure_ascii, indent_step, current_indent);
-                        o->write_character(',');
-                    }
-
-                    // last element
-                    JSON_ASSERT(i != val.m_value.object->cend());
-                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
-                    o->write_character('\"');
-                    dump_escaped(i->first, ensure_ascii);
-                    o->write_characters("\":", 2);
-                    dump(i->second, false, ensure_ascii, indent_step, current_indent);
-
-                    o->write_character('}');
-                }
-
-                return;
-            }
-
-            case value_t::array:
-            {
-                if (val.m_value.array->empty())
-                {
-                    o->write_characters("[]", 2);
-                    return;
-                }
-
-                if (pretty_print)
-                {
-                    o->write_characters("[\n", 2);
-
-                    // variable to hold indentation for recursive calls
-                    const auto new_indent = current_indent + indent_step;
-                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
-                    {
-                        indent_string.resize(indent_string.size() * 2, ' ');
-                    }
-
-                    // first n-1 elements
-                    for (auto i = val.m_value.array->cbegin();
-                            i != val.m_value.array->cend() - 1; ++i)
-                    {
-                        o->write_characters(indent_string.c_str(), new_indent);
-                        dump(*i, true, ensure_ascii, indent_step, new_indent);
-                        o->write_characters(",\n", 2);
-                    }
-
-                    // last element
-                    JSON_ASSERT(!val.m_value.array->empty());
-                    o->write_characters(indent_string.c_str(), new_indent);
-                    dump(val.m_value.array->back(), true, ensure_ascii, indent_step, new_indent);
-
-                    o->write_character('\n');
-                    o->write_characters(indent_string.c_str(), current_indent);
-                    o->write_character(']');
-                }
-                else
-                {
-                    o->write_character('[');
-
-                    // first n-1 elements
-                    for (auto i = val.m_value.array->cbegin();
-                            i != val.m_value.array->cend() - 1; ++i)
-                    {
-                        dump(*i, false, ensure_ascii, indent_step, current_indent);
-                        o->write_character(',');
-                    }
-
-                    // last element
-                    JSON_ASSERT(!val.m_value.array->empty());
-                    dump(val.m_value.array->back(), false, ensure_ascii, indent_step, current_indent);
-
-                    o->write_character(']');
-                }
-
-                return;
-            }
-
-            case value_t::string:
-            {
-                o->write_character('\"');
-                dump_escaped(*val.m_value.string, ensure_ascii);
-                o->write_character('\"');
-                return;
-            }
-
-            case value_t::binary:
-            {
-                if (pretty_print)
-                {
-                    o->write_characters("{\n", 2);
-
-                    // variable to hold indentation for recursive calls
-                    const auto new_indent = current_indent + indent_step;
-                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
-                    {
-                        indent_string.resize(indent_string.size() * 2, ' ');
-                    }
-
-                    o->write_characters(indent_string.c_str(), new_indent);
-
-                    o->write_characters("\"bytes\": [", 10);
-
-                    if (!val.m_value.binary->empty())
-                    {
-                        for (auto i = val.m_value.binary->cbegin();
-                                i != val.m_value.binary->cend() - 1; ++i)
-                        {
-                            dump_integer(*i);
-                            o->write_characters(", ", 2);
-                        }
-                        dump_integer(val.m_value.binary->back());
-                    }
-
-                    o->write_characters("],\n", 3);
-                    o->write_characters(indent_string.c_str(), new_indent);
-
-                    o->write_characters("\"subtype\": ", 11);
-                    if (val.m_value.binary->has_subtype())
-                    {
-                        dump_integer(val.m_value.binary->subtype());
-                    }
-                    else
-                    {
-                        o->write_characters("null", 4);
-                    }
-                    o->write_character('\n');
-                    o->write_characters(indent_string.c_str(), current_indent);
-                    o->write_character('}');
-                }
-                else
-                {
-                    o->write_characters("{\"bytes\":[", 10);
-
-                    if (!val.m_value.binary->empty())
-                    {
-                        for (auto i = val.m_value.binary->cbegin();
-                                i != val.m_value.binary->cend() - 1; ++i)
-                        {
-                            dump_integer(*i);
-                            o->write_character(',');
-                        }
-                        dump_integer(val.m_value.binary->back());
-                    }
-
-                    o->write_characters("],\"subtype\":", 12);
-                    if (val.m_value.binary->has_subtype())
-                    {
-                        dump_integer(val.m_value.binary->subtype());
-                        o->write_character('}');
-                    }
-                    else
-                    {
-                        o->write_characters("null}", 5);
-                    }
-                }
-                return;
-            }
-
-            case value_t::boolean:
-            {
-                if (val.m_value.boolean)
-                {
-                    o->write_characters("true", 4);
-                }
-                else
-                {
-                    o->write_characters("false", 5);
-                }
-                return;
-            }
-
-            case value_t::number_integer:
-            {
-                dump_integer(val.m_value.number_integer);
-                return;
-            }
-
-            case value_t::number_unsigned:
-            {
-                dump_integer(val.m_value.number_unsigned);
-                return;
-            }
-
-            case value_t::number_float:
-            {
-                dump_float(val.m_value.number_float);
-                return;
-            }
-
-            case value_t::discarded:
-            {
-                o->write_characters("<discarded>", 11);
-                return;
-            }
-
-            case value_t::null:
-            {
-                o->write_characters("null", 4);
-                return;
-            }
-
-            default:            // LCOV_EXCL_LINE
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        }
-    }
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    /*!
-    @brief dump escaped string
-
-    Escape a string by replacing certain special characters by a sequence of an
-    escape character (backslash) and another character and other control
-    characters by a sequence of "\u" followed by a four-digit hex
-    representation. The escaped string is written to output stream @a o.
-
-    @param[in] s  the string to escape
-    @param[in] ensure_ascii  whether to escape non-ASCII characters with
-                             \uXXXX sequences
-
-    @complexity Linear in the length of string @a s.
-    */
-    void dump_escaped(const string_t& s, const bool ensure_ascii)
-    {
-        std::uint32_t codepoint{};
-        std::uint8_t state = UTF8_ACCEPT;
-        std::size_t bytes = 0;  // number of bytes written to string_buffer
-
-        // number of bytes written at the point of the last valid byte
-        std::size_t bytes_after_last_accept = 0;
-        std::size_t undumped_chars = 0;
-
-        for (std::size_t i = 0; i < s.size(); ++i)
-        {
-            const auto byte = static_cast<std::uint8_t>(s[i]);
-
-            switch (decode(state, codepoint, byte))
-            {
-                case UTF8_ACCEPT:  // decode found a new code point
-                {
-                    switch (codepoint)
-                    {
-                        case 0x08: // backspace
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'b';
-                            break;
-                        }
-
-                        case 0x09: // horizontal tab
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 't';
-                            break;
-                        }
-
-                        case 0x0A: // newline
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'n';
-                            break;
-                        }
-
-                        case 0x0C: // formfeed
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'f';
-                            break;
-                        }
-
-                        case 0x0D: // carriage return
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'r';
-                            break;
-                        }
-
-                        case 0x22: // quotation mark
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = '\"';
-                            break;
-                        }
-
-                        case 0x5C: // reverse solidus
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = '\\';
-                            break;
-                        }
-
-                        default:
-                        {
-                            // escape control characters (0x00..0x1F) or, if
-                            // ensure_ascii parameter is used, non-ASCII characters
-                            if ((codepoint <= 0x1F) || (ensure_ascii && (codepoint >= 0x7F)))
-                            {
-                                if (codepoint <= 0xFFFF)
-                                {
-                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 7, "\\u%04x",
-                                                                      static_cast<std::uint16_t>(codepoint)));
-                                    bytes += 6;
-                                }
-                                else
-                                {
-                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 13, "\\u%04x\\u%04x",
-                                                                      static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)),
-                                                                      static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu))));
-                                    bytes += 12;
-                                }
-                            }
-                            else
-                            {
-                                // copy byte to buffer (all previous bytes
-                                // been copied have in default case above)
-                                string_buffer[bytes++] = s[i];
-                            }
-                            break;
-                        }
-                    }
-
-                    // write buffer and reset index; there must be 13 bytes
-                    // left, as this is the maximal number of bytes to be
-                    // written ("\uxxxx\uxxxx\0") for one code point
-                    if (string_buffer.size() - bytes < 13)
-                    {
-                        o->write_characters(string_buffer.data(), bytes);
-                        bytes = 0;
-                    }
-
-                    // remember the byte position of this accept
-                    bytes_after_last_accept = bytes;
-                    undumped_chars = 0;
-                    break;
-                }
-
-                case UTF8_REJECT:  // decode found invalid UTF-8 byte
-                {
-                    switch (error_handler)
-                    {
-                        case error_handler_t::strict:
-                        {
-                            JSON_THROW(type_error::create(316, concat("invalid UTF-8 byte at index ", std::to_string(i), ": 0x", hex_bytes(byte | 0)), nullptr));
-                        }
-
-                        case error_handler_t::ignore:
-                        case error_handler_t::replace:
-                        {
-                            // in case we saw this character the first time, we
-                            // would like to read it again, because the byte
-                            // may be OK for itself, but just not OK for the
-                            // previous sequence
-                            if (undumped_chars > 0)
-                            {
-                                --i;
-                            }
-
-                            // reset length buffer to the last accepted index;
-                            // thus removing/ignoring the invalid characters
-                            bytes = bytes_after_last_accept;
-
-                            if (error_handler == error_handler_t::replace)
-                            {
-                                // add a replacement character
-                                if (ensure_ascii)
-                                {
-                                    string_buffer[bytes++] = '\\';
-                                    string_buffer[bytes++] = 'u';
-                                    string_buffer[bytes++] = 'f';
-                                    string_buffer[bytes++] = 'f';
-                                    string_buffer[bytes++] = 'f';
-                                    string_buffer[bytes++] = 'd';
-                                }
-                                else
-                                {
-                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xEF');
-                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBF');
-                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBD');
-                                }
-
-                                // write buffer and reset index; there must be 13 bytes
-                                // left, as this is the maximal number of bytes to be
-                                // written ("\uxxxx\uxxxx\0") for one code point
-                                if (string_buffer.size() - bytes < 13)
-                                {
-                                    o->write_characters(string_buffer.data(), bytes);
-                                    bytes = 0;
-                                }
-
-                                bytes_after_last_accept = bytes;
-                            }
-
-                            undumped_chars = 0;
-
-                            // continue processing the string
-                            state = UTF8_ACCEPT;
-                            break;
-                        }
-
-                        default:            // LCOV_EXCL_LINE
-                            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-                    }
-                    break;
-                }
-
-                default:  // decode found yet incomplete multi-byte code point
-                {
-                    if (!ensure_ascii)
-                    {
-                        // code point will not be escaped - copy byte to buffer
-                        string_buffer[bytes++] = s[i];
-                    }
-                    ++undumped_chars;
-                    break;
-                }
-            }
-        }
-
-        // we finished processing the string
-        if (JSON_HEDLEY_LIKELY(state == UTF8_ACCEPT))
-        {
-            // write buffer
-            if (bytes > 0)
-            {
-                o->write_characters(string_buffer.data(), bytes);
-            }
-        }
-        else
-        {
-            // we finish reading, but do not accept: string was incomplete
-            switch (error_handler)
-            {
-                case error_handler_t::strict:
-                {
-                    JSON_THROW(type_error::create(316, concat("incomplete UTF-8 string; last byte: 0x", hex_bytes(static_cast<std::uint8_t>(s.back() | 0))), nullptr));
-                }
-
-                case error_handler_t::ignore:
-                {
-                    // write all accepted bytes
-                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
-                    break;
-                }
-
-                case error_handler_t::replace:
-                {
-                    // write all accepted bytes
-                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
-                    // add a replacement character
-                    if (ensure_ascii)
-                    {
-                        o->write_characters("\\ufffd", 6);
-                    }
-                    else
-                    {
-                        o->write_characters("\xEF\xBF\xBD", 3);
-                    }
-                    break;
-                }
-
-                default:            // LCOV_EXCL_LINE
-                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-            }
-        }
-    }
-
-  private:
-    /*!
-    @brief count digits
-
-    Count the number of decimal (base 10) digits for an input unsigned integer.
-
-    @param[in] x  unsigned integer number to count its digits
-    @return    number of decimal digits
-    */
-    inline unsigned int count_digits(number_unsigned_t x) noexcept
-    {
-        unsigned int n_digits = 1;
-        for (;;)
-        {
-            if (x < 10)
-            {
-                return n_digits;
-            }
-            if (x < 100)
-            {
-                return n_digits + 1;
-            }
-            if (x < 1000)
-            {
-                return n_digits + 2;
-            }
-            if (x < 10000)
-            {
-                return n_digits + 3;
-            }
-            x = x / 10000u;
-            n_digits += 4;
-        }
-    }
-
-    /*!
-     * @brief convert a byte to a uppercase hex representation
-     * @param[in] byte byte to represent
-     * @return representation ("00".."FF")
-     */
-    static std::string hex_bytes(std::uint8_t byte)
-    {
-        std::string result = "FF";
-        constexpr const char* nibble_to_hex = "0123456789ABCDEF";
-        result[0] = nibble_to_hex[byte / 16];
-        result[1] = nibble_to_hex[byte % 16];
-        return result;
-    }
-
-    // templates to avoid warnings about useless casts
-    template <typename NumberType, enable_if_t<std::is_signed<NumberType>::value, int> = 0>
-    bool is_negative_number(NumberType x)
-    {
-        return x < 0;
-    }
-
-    template < typename NumberType, enable_if_t <std::is_unsigned<NumberType>::value, int > = 0 >
-    bool is_negative_number(NumberType /*unused*/)
-    {
-        return false;
-    }
-
-    /*!
-    @brief dump an integer
-
-    Dump a given integer to output stream @a o. Works internally with
-    @a number_buffer.
-
-    @param[in] x  integer number (signed or unsigned) to dump
-    @tparam NumberType either @a number_integer_t or @a number_unsigned_t
-    */
-    template < typename NumberType, detail::enable_if_t <
-                   std::is_integral<NumberType>::value ||
-                   std::is_same<NumberType, number_unsigned_t>::value ||
-                   std::is_same<NumberType, number_integer_t>::value ||
-                   std::is_same<NumberType, binary_char_t>::value,
-                   int > = 0 >
-    void dump_integer(NumberType x)
-    {
-        static constexpr std::array<std::array<char, 2>, 100> digits_to_99
-        {
-            {
-                {{'0', '0'}}, {{'0', '1'}}, {{'0', '2'}}, {{'0', '3'}}, {{'0', '4'}}, {{'0', '5'}}, {{'0', '6'}}, {{'0', '7'}}, {{'0', '8'}}, {{'0', '9'}},
-                {{'1', '0'}}, {{'1', '1'}}, {{'1', '2'}}, {{'1', '3'}}, {{'1', '4'}}, {{'1', '5'}}, {{'1', '6'}}, {{'1', '7'}}, {{'1', '8'}}, {{'1', '9'}},
-                {{'2', '0'}}, {{'2', '1'}}, {{'2', '2'}}, {{'2', '3'}}, {{'2', '4'}}, {{'2', '5'}}, {{'2', '6'}}, {{'2', '7'}}, {{'2', '8'}}, {{'2', '9'}},
-                {{'3', '0'}}, {{'3', '1'}}, {{'3', '2'}}, {{'3', '3'}}, {{'3', '4'}}, {{'3', '5'}}, {{'3', '6'}}, {{'3', '7'}}, {{'3', '8'}}, {{'3', '9'}},
-                {{'4', '0'}}, {{'4', '1'}}, {{'4', '2'}}, {{'4', '3'}}, {{'4', '4'}}, {{'4', '5'}}, {{'4', '6'}}, {{'4', '7'}}, {{'4', '8'}}, {{'4', '9'}},
-                {{'5', '0'}}, {{'5', '1'}}, {{'5', '2'}}, {{'5', '3'}}, {{'5', '4'}}, {{'5', '5'}}, {{'5', '6'}}, {{'5', '7'}}, {{'5', '8'}}, {{'5', '9'}},
-                {{'6', '0'}}, {{'6', '1'}}, {{'6', '2'}}, {{'6', '3'}}, {{'6', '4'}}, {{'6', '5'}}, {{'6', '6'}}, {{'6', '7'}}, {{'6', '8'}}, {{'6', '9'}},
-                {{'7', '0'}}, {{'7', '1'}}, {{'7', '2'}}, {{'7', '3'}}, {{'7', '4'}}, {{'7', '5'}}, {{'7', '6'}}, {{'7', '7'}}, {{'7', '8'}}, {{'7', '9'}},
-                {{'8', '0'}}, {{'8', '1'}}, {{'8', '2'}}, {{'8', '3'}}, {{'8', '4'}}, {{'8', '5'}}, {{'8', '6'}}, {{'8', '7'}}, {{'8', '8'}}, {{'8', '9'}},
-                {{'9', '0'}}, {{'9', '1'}}, {{'9', '2'}}, {{'9', '3'}}, {{'9', '4'}}, {{'9', '5'}}, {{'9', '6'}}, {{'9', '7'}}, {{'9', '8'}}, {{'9', '9'}},
-            }
-        };
-
-        // special case for "0"
-        if (x == 0)
-        {
-            o->write_character('0');
-            return;
-        }
-
-        // use a pointer to fill the buffer
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-
-        number_unsigned_t abs_value;
-
-        unsigned int n_chars{};
-
-        if (is_negative_number(x))
-        {
-            *buffer_ptr = '-';
-            abs_value = remove_sign(static_cast<number_integer_t>(x));
-
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
-        }
-        else
-        {
-            abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
-        }
-
-        // spare 1 byte for '\0'
-        JSON_ASSERT(n_chars < number_buffer.size() - 1);
-
-        // jump to the end to generate the string from backward,
-        // so we later avoid reversing the result
-        buffer_ptr += n_chars;
-
-        // Fast int2ascii implementation inspired by "Fastware" talk by Andrei Alexandrescu
-        // See: https://www.youtube.com/watch?v=o4-CwDo2zpg
-        while (abs_value >= 100)
-        {
-            const auto digits_index = static_cast<unsigned>((abs_value % 100));
-            abs_value /= 100;
-            *(--buffer_ptr) = digits_to_99[digits_index][1];
-            *(--buffer_ptr) = digits_to_99[digits_index][0];
-        }
-
-        if (abs_value >= 10)
-        {
-            const auto digits_index = static_cast<unsigned>(abs_value);
-            *(--buffer_ptr) = digits_to_99[digits_index][1];
-            *(--buffer_ptr) = digits_to_99[digits_index][0];
-        }
-        else
-        {
-            *(--buffer_ptr) = static_cast<char>('0' + abs_value);
-        }
-
-        o->write_characters(number_buffer.data(), n_chars);
-    }
-
-    /*!
-    @brief dump a floating-point number
-
-    Dump a given floating-point number to output stream @a o. Works internally
-    with @a number_buffer.
-
-    @param[in] x  floating-point number to dump
-    */
-    void dump_float(number_float_t x)
-    {
-        // NaN / inf
-        if (!std::isfinite(x))
-        {
-            o->write_characters("null", 4);
-            return;
-        }
-
-        // If number_float_t is an IEEE-754 single or double precision number,
-        // use the Grisu2 algorithm to produce short numbers which are
-        // guaranteed to round-trip, using strtof and strtod, resp.
-        //
-        // NB: The test below works if <long double> == <double>.
-        static constexpr bool is_ieee_single_or_double
-            = (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 24 && std::numeric_limits<number_float_t>::max_exponent == 128) ||
-              (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 53 && std::numeric_limits<number_float_t>::max_exponent == 1024);
-
-        dump_float(x, std::integral_constant<bool, is_ieee_single_or_double>());
-    }
-
-    void dump_float(number_float_t x, std::true_type /*is_ieee_single_or_double*/)
-    {
-        auto* begin = number_buffer.data();
-        auto* end = ::nlohmann::detail::to_chars(begin, begin + number_buffer.size(), x);
-
-        o->write_characters(begin, static_cast<size_t>(end - begin));
-    }
-
-    void dump_float(number_float_t x, std::false_type /*is_ieee_single_or_double*/)
-    {
-        // get number of digits for a float -> text -> float round-trip
-        static constexpr auto d = std::numeric_limits<number_float_t>::max_digits10;
-
-        // the actual conversion
-        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
-        std::ptrdiff_t len = (std::snprintf)(number_buffer.data(), number_buffer.size(), "%.*g", d, x);
-
-        // negative value indicates an error
-        JSON_ASSERT(len > 0);
-        // check if buffer was large enough
-        JSON_ASSERT(static_cast<std::size_t>(len) < number_buffer.size());
-
-        // erase thousands separator
-        if (thousands_sep != '\0')
-        {
-            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::remove returns an iterator, see https://github.com/nlohmann/json/issues/3081
-            const auto end = std::remove(number_buffer.begin(), number_buffer.begin() + len, thousands_sep);
-            std::fill(end, number_buffer.end(), '\0');
-            JSON_ASSERT((end - number_buffer.begin()) <= len);
-            len = (end - number_buffer.begin());
-        }
-
-        // convert decimal point to '.'
-        if (decimal_point != '\0' && decimal_point != '.')
-        {
-            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::find returns an iterator, see https://github.com/nlohmann/json/issues/3081
-            const auto dec_pos = std::find(number_buffer.begin(), number_buffer.end(), decimal_point);
-            if (dec_pos != number_buffer.end())
-            {
-                *dec_pos = '.';
-            }
-        }
-
-        o->write_characters(number_buffer.data(), static_cast<std::size_t>(len));
-
-        // determine if we need to append ".0"
-        const bool value_is_int_like =
-            std::none_of(number_buffer.begin(), number_buffer.begin() + len + 1,
-                         [](char c)
-        {
-            return c == '.' || c == 'e';
-        });
-
-        if (value_is_int_like)
-        {
-            o->write_characters(".0", 2);
-        }
-    }
-
-    /*!
-    @brief check whether a string is UTF-8 encoded
-
-    The function checks each byte of a string whether it is UTF-8 encoded. The
-    result of the check is stored in the @a state parameter. The function must
-    be called initially with state 0 (accept). State 1 means the string must
-    be rejected, because the current byte is not allowed. If the string is
-    completely processed, but the state is non-zero, the string ended
-    prematurely; that is, the last byte indicated more bytes should have
-    followed.
-
-    @param[in,out] state  the state of the decoding
-    @param[in,out] codep  codepoint (valid only if resulting state is UTF8_ACCEPT)
-    @param[in] byte       next byte to decode
-    @return               new state
-
-    @note The function has been edited: a std::array is used.
-
-    @copyright Copyright (c) 2008-2009 Bjoern Hoehrmann <bjoern@hoehrmann.de>
-    @sa http://bjoern.hoehrmann.de/utf-8/decoder/dfa/
-    */
-    static std::uint8_t decode(std::uint8_t& state, std::uint32_t& codep, const std::uint8_t byte) noexcept
-    {
-        static const std::array<std::uint8_t, 400> utf8d =
-        {
-            {
-                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 00..1F
-                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 20..3F
-                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 40..5F
-                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 60..7F
-                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, // 80..9F
-                7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // A0..BF
-                8, 8, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // C0..DF
-                0xA, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x4, 0x3, 0x3, // E0..EF
-                0xB, 0x6, 0x6, 0x6, 0x5, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, // F0..FF
-                0x0, 0x1, 0x2, 0x3, 0x5, 0x8, 0x7, 0x1, 0x1, 0x1, 0x4, 0x6, 0x1, 0x1, 0x1, 0x1, // s0..s0
-                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, // s1..s2
-                1, 2, 1, 1, 1, 1, 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, // s3..s4
-                1, 2, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, // s5..s6
-                1, 3, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 // s7..s8
-            }
-        };
-
-        JSON_ASSERT(byte < utf8d.size());
-        const std::uint8_t type = utf8d[byte];
-
-        codep = (state != UTF8_ACCEPT)
-                ? (byte & 0x3fu) | (codep << 6u)
-                : (0xFFu >> type) & (byte);
-
-        std::size_t index = 256u + static_cast<size_t>(state) * 16u + static_cast<size_t>(type);
-        JSON_ASSERT(index < 400);
-        state = utf8d[index];
-        return state;
-    }
-
-    /*
-     * Overload to make the compiler happy while it is instantiating
-     * dump_integer for number_unsigned_t.
-     * Must never be called.
-     */
-    number_unsigned_t remove_sign(number_unsigned_t x)
-    {
-        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        return x; // LCOV_EXCL_LINE
-    }
-
-    /*
-     * Helper function for dump_integer
-     *
-     * This function takes a negative signed integer and returns its absolute
-     * value as unsigned integer. The plus/minus shuffling is necessary as we can
-     * not directly remove the sign of an arbitrary signed integer as the
-     * absolute values of INT_MIN and INT_MAX are usually not the same. See
-     * #1708 for details.
-     */
-    inline number_unsigned_t remove_sign(number_integer_t x) noexcept
-    {
-        JSON_ASSERT(x < 0 && x < (std::numeric_limits<number_integer_t>::max)()); // NOLINT(misc-redundant-expression)
-        return static_cast<number_unsigned_t>(-(x + 1)) + 1;
-    }
-
-  private:
-    /// the output of the serializer
-    output_adapter_t<char> o = nullptr;
-
-    /// a (hopefully) large enough character buffer
-    std::array<char, 64> number_buffer{{}};
-
-    /// the locale
-    const std::lconv* loc = nullptr;
-    /// the locale's thousand separator character
-    const char thousands_sep = '\0';
-    /// the locale's decimal point character
-    const char decimal_point = '\0';
-
-    /// string buffer
-    std::array<char, 512> string_buffer{{}};
-
-    /// the indentation character
-    const char indent_char;
-    /// the indentation string
-    string_t indent_string;
-
-    /// error_handler how to react on decoding errors
-    const error_handler_t error_handler;
-};
-
-}  // namespace detail
-NLOHMANN_JSON_NAMESPACE_END
-
-// #include <nlohmann/detail/value_t.hpp>
-
-// #include <nlohmann/json_fwd.hpp>
-
-// #include <nlohmann/ordered_map.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#include <functional> // equal_to, less
-#include <initializer_list> // initializer_list
-#include <iterator> // input_iterator_tag, iterator_traits
-#include <memory> // allocator
-#include <stdexcept> // for out_of_range
-#include <type_traits> // enable_if, is_convertible
-#include <utility> // pair
-#include <vector> // vector
-
-// #include <nlohmann/detail/macro_scope.hpp>
-
-// #include <nlohmann/detail/meta/type_traits.hpp>
-
-
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/// ordered_map: a minimal map-like container that preserves insertion order
-/// for use within nlohmann::basic_json<ordered_map>
-template <class Key, class T, class IgnoredLess = std::less<Key>,
-          class Allocator = std::allocator<std::pair<const Key, T>>>
-                  struct ordered_map : std::vector<std::pair<const Key, T>, Allocator>
-{
-    using key_type = Key;
-    using mapped_type = T;
-    using Container = std::vector<std::pair<const Key, T>, Allocator>;
-    using iterator = typename Container::iterator;
-    using const_iterator = typename Container::const_iterator;
-    using size_type = typename Container::size_type;
-    using value_type = typename Container::value_type;
-#ifdef JSON_HAS_CPP_14
-    using key_compare = std::equal_to<>;
-#else
-    using key_compare = std::equal_to<Key>;
-#endif
-
-    // Explicit constructors instead of `using Container::Container`
-    // otherwise older compilers choke on it (GCC <= 5.5, xcode <= 9.4)
-    ordered_map() noexcept(noexcept(Container())) : Container{} {}
-    explicit ordered_map(const Allocator& alloc) noexcept(noexcept(Container(alloc))) : Container{alloc} {}
-    template <class It>
-    ordered_map(It first, It last, const Allocator& alloc = Allocator())
-        : Container{first, last, alloc} {}
-    ordered_map(std::initializer_list<value_type> init, const Allocator& alloc = Allocator() )
-        : Container{init, alloc} {}
-
-    std::pair<iterator, bool> emplace(const key_type& key, T&& t)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return {it, false};
-            }
-        }
-        Container::emplace_back(key, std::forward<T>(t));
-        return {std::prev(this->end()), true};
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    std::pair<iterator, bool> emplace(KeyType && key, T && t)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return {it, false};
-            }
-        }
-        Container::emplace_back(std::forward<KeyType>(key), std::forward<T>(t));
-        return {std::prev(this->end()), true};
-    }
-
-    T& operator[](const key_type& key)
-    {
-        return emplace(key, T{}).first->second;
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    T & operator[](KeyType && key)
-    {
-        return emplace(std::forward<KeyType>(key), T{}).first->second;
-    }
-
-    const T& operator[](const key_type& key) const
-    {
-        return at(key);
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    const T & operator[](KeyType && key) const
-    {
-        return at(std::forward<KeyType>(key));
-    }
-
-    T& at(const key_type& key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it->second;
-            }
-        }
-
-        JSON_THROW(std::out_of_range("key not found"));
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    T & at(KeyType && key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it->second;
-            }
-        }
-
-        JSON_THROW(std::out_of_range("key not found"));
-    }
-
-    const T& at(const key_type& key) const
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it->second;
-            }
-        }
-
-        JSON_THROW(std::out_of_range("key not found"));
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    const T & at(KeyType && key) const
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it->second;
-            }
-        }
-
-        JSON_THROW(std::out_of_range("key not found"));
-    }
-
-    size_type erase(const key_type& key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                // Since we cannot move const Keys, re-construct them in place
-                for (auto next = it; ++next != this->end(); ++it)
-                {
-                    it->~value_type(); // Destroy but keep allocation
-                    new (&*it) value_type{std::move(*next)};
-                }
-                Container::pop_back();
-                return 1;
-            }
-        }
-        return 0;
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    size_type erase(KeyType && key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                // Since we cannot move const Keys, re-construct them in place
-                for (auto next = it; ++next != this->end(); ++it)
-                {
-                    it->~value_type(); // Destroy but keep allocation
-                    new (&*it) value_type{std::move(*next)};
-                }
-                Container::pop_back();
-                return 1;
-            }
-        }
-        return 0;
-    }
-
-    iterator erase(iterator pos)
-    {
-        return erase(pos, std::next(pos));
-    }
-
-    iterator erase(iterator first, iterator last)
-    {
-        if (first == last)
-        {
-            return first;
-        }
-
-        const auto elements_affected = std::distance(first, last);
-        const auto offset = std::distance(Container::begin(), first);
-
-        // This is the start situation. We need to delete elements_affected
-        // elements (3 in this example: e, f, g), and need to return an
-        // iterator past the last deleted element (h in this example).
-        // Note that offset is the distance from the start of the vector
-        // to first. We will need this later.
-
-        // [ a, b, c, d, e, f, g, h, i, j ]
-        //               ^        ^
-        //             first    last
-
-        // Since we cannot move const Keys, we re-construct them in place.
-        // We start at first and re-construct (viz. copy) the elements from
-        // the back of the vector. Example for first iteration:
-
-        //               ,--------.
-        //               v        |   destroy e and re-construct with h
-        // [ a, b, c, d, e, f, g, h, i, j ]
-        //               ^        ^
-        //               it       it + elements_affected
-
-        for (auto it = first; std::next(it, elements_affected) != Container::end(); ++it)
-        {
-            it->~value_type(); // destroy but keep allocation
-            new (&*it) value_type{std::move(*std::next(it, elements_affected))}; // "move" next element to it
-        }
-
-        // [ a, b, c, d, h, i, j, h, i, j ]
-        //               ^        ^
-        //             first    last
-
-        // remove the unneeded elements at the end of the vector
-        Container::resize(this->size() - static_cast<size_type>(elements_affected));
-
-        // [ a, b, c, d, h, i, j ]
-        //               ^        ^
-        //             first    last
-
-        // first is now pointing past the last deleted element, but we cannot
-        // use this iterator, because it may have been invalidated by the
-        // resize call. Instead, we can return begin() + offset.
-        return Container::begin() + offset;
-    }
-
-    size_type count(const key_type& key) const
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return 1;
-            }
-        }
-        return 0;
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    size_type count(KeyType && key) const
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return 1;
-            }
-        }
-        return 0;
-    }
-
-    iterator find(const key_type& key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it;
-            }
-        }
-        return Container::end();
-    }
-
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
-    iterator find(KeyType && key)
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it;
-            }
-        }
-        return Container::end();
-    }
-
-    const_iterator find(const key_type& key) const
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, key))
-            {
-                return it;
-            }
-        }
-        return Container::end();
-    }
-
-    std::pair<iterator, bool> insert( value_type&& value )
-    {
-        return emplace(value.first, std::move(value.second));
-    }
-
-    std::pair<iterator, bool> insert( const value_type& value )
-    {
-        for (auto it = this->begin(); it != this->end(); ++it)
-        {
-            if (m_compare(it->first, value.first))
-            {
-                return {it, false};
-            }
-        }
-        Container::push_back(value);
-        return {--this->end(), true};
-    }
-
-    template<typename InputIt>
-    using require_input_iter = typename std::enable_if<std::is_convertible<typename std::iterator_traits<InputIt>::iterator_category,
-            std::input_iterator_tag>::value>::type;
-
-    template<typename InputIt, typename = require_input_iter<InputIt>>
-    void insert(InputIt first, InputIt last)
-    {
-        for (auto it = first; it != last; ++it)
-        {
-            insert(*it);
-        }
-    }
-
-private:
-    JSON_NO_UNIQUE_ADDRESS key_compare m_compare = key_compare();
-};
-
-NLOHMANN_JSON_NAMESPACE_END
-
-
-#if defined(JSON_HAS_CPP_17)
-    #include <any>
-    #include <string_view>
-#endif
-
-/*!
-@brief namespace for Niels Lohmann
-@see https://github.com/nlohmann
-@since version 1.0.0
-*/
-NLOHMANN_JSON_NAMESPACE_BEGIN
-
-/*!
-@brief a class to store JSON values
-
-@internal
-@invariant The member variables @a m_value and @a m_type have the following
-relationship:
-- If `m_type == value_t::object`, then `m_value.object != nullptr`.
-- If `m_type == value_t::array`, then `m_value.array != nullptr`.
-- If `m_type == value_t::string`, then `m_value.string != nullptr`.
-The invariants are checked by member function assert_invariant().
-
-@note ObjectType trick from https://stackoverflow.com/a/9860911
-@endinternal
-
-@since version 1.0.0
-
-@nosubgrouping
-*/
-NLOHMANN_BASIC_JSON_TPL_DECLARATION
-class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
-{
-  private:
-    template<detail::value_t> friend struct detail::external_constructor;
-
-    template<typename>
-    friend class ::nlohmann::json_pointer;
-    // can be restored when json_pointer backwards compatibility is removed
-    // friend ::nlohmann::json_pointer<StringType>;
-
-    template<typename BasicJsonType, typename InputType>
-    friend class ::nlohmann::detail::parser;
-    friend ::nlohmann::detail::serializer<basic_json>;
-    template<typename BasicJsonType>
-    friend class ::nlohmann::detail::iter_impl;
-    template<typename BasicJsonType, typename CharType>
-    friend class ::nlohmann::detail::binary_writer;
-    template<typename BasicJsonType, typename InputType, typename SAX>
-    friend class ::nlohmann::detail::binary_reader;
-    template<typename BasicJsonType>
-    friend class ::nlohmann::detail::json_sax_dom_parser;
-    template<typename BasicJsonType>
-    friend class ::nlohmann::detail::json_sax_dom_callback_parser;
-    friend class ::nlohmann::detail::exception;
-
-    /// workaround type for MSVC
-    using basic_json_t = NLOHMANN_BASIC_JSON_TPL;
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    // convenience aliases for types residing in namespace detail;
-    using lexer = ::nlohmann::detail::lexer_base<basic_json>;
-
-    template<typename InputAdapterType>
-    static ::nlohmann::detail::parser<basic_json, InputAdapterType> parser(
-        InputAdapterType adapter,
-        detail::parser_callback_t<basic_json>cb = nullptr,
-        const bool allow_exceptions = true,
-        const bool ignore_comments = false
-                                 )
-    {
-        return ::nlohmann::detail::parser<basic_json, InputAdapterType>(std::move(adapter),
-                std::move(cb), allow_exceptions, ignore_comments);
-    }
-
-  private:
-    using primitive_iterator_t = ::nlohmann::detail::primitive_iterator_t;
-    template<typename BasicJsonType>
-    using internal_iterator = ::nlohmann::detail::internal_iterator<BasicJsonType>;
-    template<typename BasicJsonType>
-    using iter_impl = ::nlohmann::detail::iter_impl<BasicJsonType>;
-    template<typename Iterator>
-    using iteration_proxy = ::nlohmann::detail::iteration_proxy<Iterator>;
-    template<typename Base> using json_reverse_iterator = ::nlohmann::detail::json_reverse_iterator<Base>;
-
-    template<typename CharType>
-    using output_adapter_t = ::nlohmann::detail::output_adapter_t<CharType>;
-
-    template<typename InputType>
-    using binary_reader = ::nlohmann::detail::binary_reader<basic_json, InputType>;
-    template<typename CharType> using binary_writer = ::nlohmann::detail::binary_writer<basic_json, CharType>;
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    using serializer = ::nlohmann::detail::serializer<basic_json>;
-
-  public:
-    using value_t = detail::value_t;
-    /// JSON Pointer, see @ref nlohmann::json_pointer
-    using json_pointer = ::nlohmann::json_pointer<StringType>;
-    template<typename T, typename SFINAE>
-    using json_serializer = JSONSerializer<T, SFINAE>;
-    /// how to treat decoding errors
-    using error_handler_t = detail::error_handler_t;
-    /// how to treat CBOR tags
-    using cbor_tag_handler_t = detail::cbor_tag_handler_t;
-    /// helper type for initializer lists of basic_json values
-    using initializer_list_t = std::initializer_list<detail::json_ref<basic_json>>;
-
-    using input_format_t = detail::input_format_t;
-    /// SAX interface type, see @ref nlohmann::json_sax
-    using json_sax_t = json_sax<basic_json>;
-
-    ////////////////
-    // exceptions //
-    ////////////////
-
-    /// @name exceptions
-    /// Classes to implement user-defined exceptions.
-    /// @{
-
-    using exception = detail::exception;
-    using parse_error = detail::parse_error;
-    using invalid_iterator = detail::invalid_iterator;
-    using type_error = detail::type_error;
-    using out_of_range = detail::out_of_range;
-    using other_error = detail::other_error;
-
-    /// @}
-
-
-    /////////////////////
-    // container types //
-    /////////////////////
-
-    /// @name container types
-    /// The canonic container types to use @ref basic_json like any other STL
-    /// container.
-    /// @{
-
-    /// the type of elements in a basic_json container
-    using value_type = basic_json;
-
-    /// the type of an element reference
-    using reference = value_type&;
-    /// the type of an element const reference
-    using const_reference = const value_type&;
-
-    /// a type to represent differences between iterators
-    using difference_type = std::ptrdiff_t;
-    /// a type to represent container sizes
-    using size_type = std::size_t;
-
-    /// the allocator type
-    using allocator_type = AllocatorType<basic_json>;
-
-    /// the type of an element pointer
-    using pointer = typename std::allocator_traits<allocator_type>::pointer;
-    /// the type of an element const pointer
-    using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
-
-    /// an iterator for a basic_json container
-    using iterator = iter_impl<basic_json>;
-    /// a const iterator for a basic_json container
-    using const_iterator = iter_impl<const basic_json>;
-    /// a reverse iterator for a basic_json container
-    using reverse_iterator = json_reverse_iterator<typename basic_json::iterator>;
-    /// a const reverse iterator for a basic_json container
-    using const_reverse_iterator = json_reverse_iterator<typename basic_json::const_iterator>;
-
-    /// @}
-
-
-    /// @brief returns the allocator associated with the container
-    /// @sa https://json.nlohmann.me/api/basic_json/get_allocator/
-    static allocator_type get_allocator()
-    {
-        return allocator_type();
-    }
-
-    /// @brief returns version information on the library
-    /// @sa https://json.nlohmann.me/api/basic_json/meta/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json meta()
-    {
-        basic_json result;
-
-        result["copyright"] = "(C) 2013-2022 Niels Lohmann";
-        result["name"] = "JSON for Modern C++";
-        result["url"] = "https://github.com/nlohmann/json";
-        result["version"]["string"] =
-            detail::concat(std::to_string(NLOHMANN_JSON_VERSION_MAJOR), '.',
-                           std::to_string(NLOHMANN_JSON_VERSION_MINOR), '.',
-                           std::to_string(NLOHMANN_JSON_VERSION_PATCH));
-        result["version"]["major"] = NLOHMANN_JSON_VERSION_MAJOR;
-        result["version"]["minor"] = NLOHMANN_JSON_VERSION_MINOR;
-        result["version"]["patch"] = NLOHMANN_JSON_VERSION_PATCH;
-
-#ifdef _WIN32
-        result["platform"] = "win32";
-#elif defined __linux__
-        result["platform"] = "linux";
-#elif defined __APPLE__
-        result["platform"] = "apple";
-#elif defined __unix__
-        result["platform"] = "unix";
-#else
-        result["platform"] = "unknown";
-#endif
-
-#if defined(__ICC) || defined(__INTEL_COMPILER)
-        result["compiler"] = {{"family", "icc"}, {"version", __INTEL_COMPILER}};
-#elif defined(__clang__)
-        result["compiler"] = {{"family", "clang"}, {"version", __clang_version__}};
-#elif defined(__GNUC__) || defined(__GNUG__)
-        result["compiler"] = {{"family", "gcc"}, {"version", detail::concat(
-                    std::to_string(__GNUC__), '.',
-                    std::to_string(__GNUC_MINOR__), '.',
-                    std::to_string(__GNUC_PATCHLEVEL__))
-            }
-        };
-#elif defined(__HP_cc) || defined(__HP_aCC)
-        result["compiler"] = "hp"
-#elif defined(__IBMCPP__)
-        result["compiler"] = {{"family", "ilecpp"}, {"version", __IBMCPP__}};
-#elif defined(_MSC_VER)
-        result["compiler"] = {{"family", "msvc"}, {"version", _MSC_VER}};
-#elif defined(__PGI)
-        result["compiler"] = {{"family", "pgcpp"}, {"version", __PGI}};
-#elif defined(__SUNPRO_CC)
-        result["compiler"] = {{"family", "sunpro"}, {"version", __SUNPRO_CC}};
-#else
-        result["compiler"] = {{"family", "unknown"}, {"version", "unknown"}};
-#endif
-
-
-#if defined(_MSVC_LANG)
-        result["compiler"]["c++"] = std::to_string(_MSVC_LANG);
-#elif defined(__cplusplus)
-        result["compiler"]["c++"] = std::to_string(__cplusplus);
-#else
-        result["compiler"]["c++"] = "unknown";
-#endif
-        return result;
-    }
-
-
-    ///////////////////////////
-    // JSON value data types //
-    ///////////////////////////
-
-    /// @name JSON value data types
-    /// The data types to store a JSON value. These types are derived from
-    /// the template arguments passed to class @ref basic_json.
-    /// @{
-
-    /// @brief default object key comparator type
-    /// The actual object key comparator type (@ref object_comparator_t) may be
-    /// different.
-    /// @sa https://json.nlohmann.me/api/basic_json/default_object_comparator_t/
-#if defined(JSON_HAS_CPP_14)
-    // use of transparent comparator avoids unnecessary repeated construction of temporaries
-    // in functions involving lookup by key with types other than object_t::key_type (aka. StringType)
-    using default_object_comparator_t = std::less<>;
-#else
-    using default_object_comparator_t = std::less<StringType>;
-#endif
-
-    /// @brief a type for an object
-    /// @sa https://json.nlohmann.me/api/basic_json/object_t/
-    using object_t = ObjectType<StringType,
-          basic_json,
-          default_object_comparator_t,
-          AllocatorType<std::pair<const StringType,
-          basic_json>>>;
-
-    /// @brief a type for an array
-    /// @sa https://json.nlohmann.me/api/basic_json/array_t/
-    using array_t = ArrayType<basic_json, AllocatorType<basic_json>>;
-
-    /// @brief a type for a string
-    /// @sa https://json.nlohmann.me/api/basic_json/string_t/
-    using string_t = StringType;
-
-    /// @brief a type for a boolean
-    /// @sa https://json.nlohmann.me/api/basic_json/boolean_t/
-    using boolean_t = BooleanType;
-
-    /// @brief a type for a number (integer)
-    /// @sa https://json.nlohmann.me/api/basic_json/number_integer_t/
-    using number_integer_t = NumberIntegerType;
-
-    /// @brief a type for a number (unsigned)
-    /// @sa https://json.nlohmann.me/api/basic_json/number_unsigned_t/
-    using number_unsigned_t = NumberUnsignedType;
-
-    /// @brief a type for a number (floating-point)
-    /// @sa https://json.nlohmann.me/api/basic_json/number_float_t/
-    using number_float_t = NumberFloatType;
-
-    /// @brief a type for a packed binary type
-    /// @sa https://json.nlohmann.me/api/basic_json/binary_t/
-    using binary_t = nlohmann::byte_container_with_subtype<BinaryType>;
-
-    /// @brief object key comparator type
-    /// @sa https://json.nlohmann.me/api/basic_json/object_comparator_t/
-    using object_comparator_t = detail::actual_object_comparator_t<basic_json>;
-
-    /// @}
-
-  private:
-
-    /// helper for exception-safe object creation
-    template<typename T, typename... Args>
-    JSON_HEDLEY_RETURNS_NON_NULL
-    static T* create(Args&& ... args)
-    {
-        AllocatorType<T> alloc;
-        using AllocatorTraits = std::allocator_traits<AllocatorType<T>>;
-
-        auto deleter = [&](T * obj)
-        {
-            AllocatorTraits::deallocate(alloc, obj, 1);
-        };
-        std::unique_ptr<T, decltype(deleter)> obj(AllocatorTraits::allocate(alloc, 1), deleter);
-        AllocatorTraits::construct(alloc, obj.get(), std::forward<Args>(args)...);
-        JSON_ASSERT(obj != nullptr);
-        return obj.release();
-    }
-
-    ////////////////////////
-    // JSON value storage //
-    ////////////////////////
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    /*!
-    @brief a JSON value
-
-    The actual storage for a JSON value of the @ref basic_json class. This
-    union combines the different storage types for the JSON value types
-    defined in @ref value_t.
-
-    JSON type | value_t type    | used type
-    --------- | --------------- | ------------------------
-    object    | object          | pointer to @ref object_t
-    array     | array           | pointer to @ref array_t
-    string    | string          | pointer to @ref string_t
-    boolean   | boolean         | @ref boolean_t
-    number    | number_integer  | @ref number_integer_t
-    number    | number_unsigned | @ref number_unsigned_t
-    number    | number_float    | @ref number_float_t
-    binary    | binary          | pointer to @ref binary_t
-    null      | null            | *no value is stored*
-
-    @note Variable-length types (objects, arrays, and strings) are stored as
-    pointers. The size of the union should not exceed 64 bits if the default
-    value types are used.
-
-    @since version 1.0.0
-    */
-    union json_value
-    {
-        /// object (stored with pointer to save storage)
-        object_t* object;
-        /// array (stored with pointer to save storage)
-        array_t* array;
-        /// string (stored with pointer to save storage)
-        string_t* string;
-        /// binary (stored with pointer to save storage)
-        binary_t* binary;
-        /// boolean
-        boolean_t boolean;
-        /// number (integer)
-        number_integer_t number_integer;
-        /// number (unsigned integer)
-        number_unsigned_t number_unsigned;
-        /// number (floating-point)
-        number_float_t number_float;
-
-        /// default constructor (for null values)
-        json_value() = default;
-        /// constructor for booleans
-        json_value(boolean_t v) noexcept : boolean(v) {}
-        /// constructor for numbers (integer)
-        json_value(number_integer_t v) noexcept : number_integer(v) {}
-        /// constructor for numbers (unsigned)
-        json_value(number_unsigned_t v) noexcept : number_unsigned(v) {}
-        /// constructor for numbers (floating-point)
-        json_value(number_float_t v) noexcept : number_float(v) {}
-        /// constructor for empty values of a given type
-        json_value(value_t t)
-        {
-            switch (t)
-            {
-                case value_t::object:
-                {
-                    object = create<object_t>();
-                    break;
-                }
-
-                case value_t::array:
-                {
-                    array = create<array_t>();
-                    break;
-                }
-
-                case value_t::string:
-                {
-                    string = create<string_t>("");
-                    break;
-                }
-
-                case value_t::binary:
-                {
-                    binary = create<binary_t>();
-                    break;
-                }
-
-                case value_t::boolean:
-                {
-                    boolean = static_cast<boolean_t>(false);
-                    break;
-                }
-
-                case value_t::number_integer:
-                {
-                    number_integer = static_cast<number_integer_t>(0);
-                    break;
-                }
-
-                case value_t::number_unsigned:
-                {
-                    number_unsigned = static_cast<number_unsigned_t>(0);
-                    break;
-                }
-
-                case value_t::number_float:
-                {
-                    number_float = static_cast<number_float_t>(0.0);
-                    break;
-                }
-
-                case value_t::null:
-                {
-                    object = nullptr;  // silence warning, see #821
-                    break;
-                }
-
-                case value_t::discarded:
-                default:
-                {
-                    object = nullptr;  // silence warning, see #821
-                    if (JSON_HEDLEY_UNLIKELY(t == value_t::null))
-                    {
-                        JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.11.2", nullptr)); // LCOV_EXCL_LINE
-                    }
-                    break;
-                }
-            }
-        }
-
-        /// constructor for strings
-        json_value(const string_t& value) : string(create<string_t>(value)) {}
-
-        /// constructor for rvalue strings
-        json_value(string_t&& value) : string(create<string_t>(std::move(value))) {}
-
-        /// constructor for objects
-        json_value(const object_t& value) : object(create<object_t>(value)) {}
-
-        /// constructor for rvalue objects
-        json_value(object_t&& value) : object(create<object_t>(std::move(value))) {}
-
-        /// constructor for arrays
-        json_value(const array_t& value) : array(create<array_t>(value)) {}
-
-        /// constructor for rvalue arrays
-        json_value(array_t&& value) : array(create<array_t>(std::move(value))) {}
-
-        /// constructor for binary arrays
-        json_value(const typename binary_t::container_type& value) : binary(create<binary_t>(value)) {}
-
-        /// constructor for rvalue binary arrays
-        json_value(typename binary_t::container_type&& value) : binary(create<binary_t>(std::move(value))) {}
-
-        /// constructor for binary arrays (internal type)
-        json_value(const binary_t& value) : binary(create<binary_t>(value)) {}
-
-        /// constructor for rvalue binary arrays (internal type)
-        json_value(binary_t&& value) : binary(create<binary_t>(std::move(value))) {}
-
-        void destroy(value_t t)
-        {
-            if (t == value_t::array || t == value_t::object)
-            {
-                // flatten the current json_value to a heap-allocated stack
-                std::vector<basic_json> stack;
-
-                // move the top-level items to stack
-                if (t == value_t::array)
-                {
-                    stack.reserve(array->size());
-                    std::move(array->begin(), array->end(), std::back_inserter(stack));
-                }
-                else
-                {
-                    stack.reserve(object->size());
-                    for (auto&& it : *object)
-                    {
-                        stack.push_back(std::move(it.second));
-                    }
-                }
-
-                while (!stack.empty())
-                {
-                    // move the last item to local variable to be processed
-                    basic_json current_item(std::move(stack.back()));
-                    stack.pop_back();
-
-                    // if current_item is array/object, move
-                    // its children to the stack to be processed later
-                    if (current_item.is_array())
-                    {
-                        std::move(current_item.m_value.array->begin(), current_item.m_value.array->end(), std::back_inserter(stack));
-
-                        current_item.m_value.array->clear();
-                    }
-                    else if (current_item.is_object())
-                    {
-                        for (auto&& it : *current_item.m_value.object)
-                        {
-                            stack.push_back(std::move(it.second));
-                        }
-
-                        current_item.m_value.object->clear();
-                    }
-
-                    // it's now safe that current_item get destructed
-                    // since it doesn't have any children
-                }
-            }
-
-            switch (t)
-            {
-                case value_t::object:
-                {
-                    AllocatorType<object_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, object);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, object, 1);
-                    break;
-                }
-
-                case value_t::array:
-                {
-                    AllocatorType<array_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, array);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, array, 1);
-                    break;
-                }
-
-                case value_t::string:
-                {
-                    AllocatorType<string_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, string);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, string, 1);
-                    break;
-                }
-
-                case value_t::binary:
-                {
-                    AllocatorType<binary_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, binary);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, binary, 1);
-                    break;
-                }
-
-                case value_t::null:
-                case value_t::boolean:
-                case value_t::number_integer:
-                case value_t::number_unsigned:
-                case value_t::number_float:
-                case value_t::discarded:
-                default:
-                {
-                    break;
-                }
-            }
-        }
-    };
-
-  private:
-    /*!
-    @brief checks the class invariants
-
-    This function asserts the class invariants. It needs to be called at the
-    end of every constructor to make sure that created objects respect the
-    invariant. Furthermore, it has to be called each time the type of a JSON
-    value is changed, because the invariant expresses a relationship between
-    @a m_type and @a m_value.
-
-    Furthermore, the parent relation is checked for arrays and objects: If
-    @a check_parents true and the value is an array or object, then the
-    container's elements must have the current value as parent.
-
-    @param[in] check_parents  whether the parent relation should be checked.
-               The value is true by default and should only be set to false
-               during destruction of objects when the invariant does not
-               need to hold.
-    */
-    void assert_invariant(bool check_parents = true) const noexcept
-    {
-        JSON_ASSERT(m_type != value_t::object || m_value.object != nullptr);
-        JSON_ASSERT(m_type != value_t::array || m_value.array != nullptr);
-        JSON_ASSERT(m_type != value_t::string || m_value.string != nullptr);
-        JSON_ASSERT(m_type != value_t::binary || m_value.binary != nullptr);
-
-#if JSON_DIAGNOSTICS
-        JSON_TRY
-        {
-            // cppcheck-suppress assertWithSideEffect
-            JSON_ASSERT(!check_parents || !is_structured() || std::all_of(begin(), end(), [this](const basic_json & j)
-            {
-                return j.m_parent == this;
-            }));
-        }
-        JSON_CATCH(...) {} // LCOV_EXCL_LINE
-#endif
-        static_cast<void>(check_parents);
-    }
-
-    void set_parents()
-    {
-#if JSON_DIAGNOSTICS
-        switch (m_type)
-        {
-            case value_t::array:
-            {
-                for (auto& element : *m_value.array)
-                {
-                    element.m_parent = this;
-                }
-                break;
-            }
-
-            case value_t::object:
-            {
-                for (auto& element : *m_value.object)
-                {
-                    element.second.m_parent = this;
-                }
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                break;
-        }
-#endif
-    }
-
-    iterator set_parents(iterator it, typename iterator::difference_type count_set_parents)
-    {
-#if JSON_DIAGNOSTICS
-        for (typename iterator::difference_type i = 0; i < count_set_parents; ++i)
-        {
-            (it + i)->m_parent = this;
-        }
-#else
-        static_cast<void>(count_set_parents);
-#endif
-        return it;
-    }
-
-    reference set_parent(reference j, std::size_t old_capacity = static_cast<std::size_t>(-1))
-    {
-#if JSON_DIAGNOSTICS
-        if (old_capacity != static_cast<std::size_t>(-1))
-        {
-            // see https://github.com/nlohmann/json/issues/2838
-            JSON_ASSERT(type() == value_t::array);
-            if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
-            {
-                // capacity has changed: update all parents
-                set_parents();
-                return j;
-            }
-        }
-
-        // ordered_json uses a vector internally, so pointers could have
-        // been invalidated; see https://github.com/nlohmann/json/issues/2962
-#ifdef JSON_HEDLEY_MSVC_VERSION
-#pragma warning(push )
-#pragma warning(disable : 4127) // ignore warning to replace if with if constexpr
-#endif
-        if (detail::is_ordered_map<object_t>::value)
-        {
-            set_parents();
-            return j;
-        }
-#ifdef JSON_HEDLEY_MSVC_VERSION
-#pragma warning( pop )
-#endif
-
-        j.m_parent = this;
-#else
-        static_cast<void>(j);
-        static_cast<void>(old_capacity);
-#endif
-        return j;
-    }
-
-  public:
-    //////////////////////////
-    // JSON parser callback //
-    //////////////////////////
-
-    /// @brief parser event types
-    /// @sa https://json.nlohmann.me/api/basic_json/parse_event_t/
-    using parse_event_t = detail::parse_event_t;
-
-    /// @brief per-element parser callback type
-    /// @sa https://json.nlohmann.me/api/basic_json/parser_callback_t/
-    using parser_callback_t = detail::parser_callback_t<basic_json>;
-
-    //////////////////
-    // constructors //
-    //////////////////
-
-    /// @name constructors and destructors
-    /// Constructors of class @ref basic_json, copy/move constructor, copy
-    /// assignment, static functions creating objects, and the destructor.
-    /// @{
-
-    /// @brief create an empty value with a given type
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(const value_t v)
-        : m_type(v), m_value(v)
-    {
-        assert_invariant();
-    }
-
-    /// @brief create a null object
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(std::nullptr_t = nullptr) noexcept // NOLINT(bugprone-exception-escape)
-        : basic_json(value_t::null)
-    {
-        assert_invariant();
-    }
-
-    /// @brief create a JSON value from compatible types
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    template < typename CompatibleType,
-               typename U = detail::uncvref_t<CompatibleType>,
-               detail::enable_if_t <
-                   !detail::is_basic_json<U>::value && detail::is_compatible_type<basic_json_t, U>::value, int > = 0 >
-    basic_json(CompatibleType && val) noexcept(noexcept( // NOLINT(bugprone-forwarding-reference-overload,bugprone-exception-escape)
-                JSONSerializer<U>::to_json(std::declval<basic_json_t&>(),
-                                           std::forward<CompatibleType>(val))))
-    {
-        JSONSerializer<U>::to_json(*this, std::forward<CompatibleType>(val));
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief create a JSON value from an existing one
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    template < typename BasicJsonType,
-               detail::enable_if_t <
-                   detail::is_basic_json<BasicJsonType>::value&& !std::is_same<basic_json, BasicJsonType>::value, int > = 0 >
-    basic_json(const BasicJsonType& val)
-    {
-        using other_boolean_t = typename BasicJsonType::boolean_t;
-        using other_number_float_t = typename BasicJsonType::number_float_t;
-        using other_number_integer_t = typename BasicJsonType::number_integer_t;
-        using other_number_unsigned_t = typename BasicJsonType::number_unsigned_t;
-        using other_string_t = typename BasicJsonType::string_t;
-        using other_object_t = typename BasicJsonType::object_t;
-        using other_array_t = typename BasicJsonType::array_t;
-        using other_binary_t = typename BasicJsonType::binary_t;
-
-        switch (val.type())
-        {
-            case value_t::boolean:
-                JSONSerializer<other_boolean_t>::to_json(*this, val.template get<other_boolean_t>());
-                break;
-            case value_t::number_float:
-                JSONSerializer<other_number_float_t>::to_json(*this, val.template get<other_number_float_t>());
-                break;
-            case value_t::number_integer:
-                JSONSerializer<other_number_integer_t>::to_json(*this, val.template get<other_number_integer_t>());
-                break;
-            case value_t::number_unsigned:
-                JSONSerializer<other_number_unsigned_t>::to_json(*this, val.template get<other_number_unsigned_t>());
-                break;
-            case value_t::string:
-                JSONSerializer<other_string_t>::to_json(*this, val.template get_ref<const other_string_t&>());
-                break;
-            case value_t::object:
-                JSONSerializer<other_object_t>::to_json(*this, val.template get_ref<const other_object_t&>());
-                break;
-            case value_t::array:
-                JSONSerializer<other_array_t>::to_json(*this, val.template get_ref<const other_array_t&>());
-                break;
-            case value_t::binary:
-                JSONSerializer<other_binary_t>::to_json(*this, val.template get_ref<const other_binary_t&>());
-                break;
-            case value_t::null:
-                *this = nullptr;
-                break;
-            case value_t::discarded:
-                m_type = value_t::discarded;
-                break;
-            default:            // LCOV_EXCL_LINE
-                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        }
-        JSON_ASSERT(m_type == val.type());
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief create a container (array or object) from an initializer list
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(initializer_list_t init,
-               bool type_deduction = true,
-               value_t manual_type = value_t::array)
-    {
-        // check if each element is an array with two elements whose first
-        // element is a string
-        bool is_an_object = std::all_of(init.begin(), init.end(),
-                                        [](const detail::json_ref<basic_json>& element_ref)
-        {
-            return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[0].is_string();
-        });
-
-        // adjust type if type deduction is not wanted
-        if (!type_deduction)
-        {
-            // if array is wanted, do not create an object though possible
-            if (manual_type == value_t::array)
-            {
-                is_an_object = false;
-            }
-
-            // if object is wanted but impossible, throw an exception
-            if (JSON_HEDLEY_UNLIKELY(manual_type == value_t::object && !is_an_object))
-            {
-                JSON_THROW(type_error::create(301, "cannot create object from initializer list", nullptr));
-            }
-        }
-
-        if (is_an_object)
-        {
-            // the initializer list is a list of pairs -> create object
-            m_type = value_t::object;
-            m_value = value_t::object;
-
-            for (auto& element_ref : init)
-            {
-                auto element = element_ref.moved_or_copied();
-                m_value.object->emplace(
-                    std::move(*((*element.m_value.array)[0].m_value.string)),
-                    std::move((*element.m_value.array)[1]));
-            }
-        }
-        else
-        {
-            // the initializer list describes an array -> create array
-            m_type = value_t::array;
-            m_value.array = create<array_t>(init.begin(), init.end());
-        }
-
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief explicitly create a binary array (without subtype)
-    /// @sa https://json.nlohmann.me/api/basic_json/binary/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json binary(const typename binary_t::container_type& init)
-    {
-        auto res = basic_json();
-        res.m_type = value_t::binary;
-        res.m_value = init;
-        return res;
-    }
-
-    /// @brief explicitly create a binary array (with subtype)
-    /// @sa https://json.nlohmann.me/api/basic_json/binary/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json binary(const typename binary_t::container_type& init, typename binary_t::subtype_type subtype)
-    {
-        auto res = basic_json();
-        res.m_type = value_t::binary;
-        res.m_value = binary_t(init, subtype);
-        return res;
-    }
-
-    /// @brief explicitly create a binary array
-    /// @sa https://json.nlohmann.me/api/basic_json/binary/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json binary(typename binary_t::container_type&& init)
-    {
-        auto res = basic_json();
-        res.m_type = value_t::binary;
-        res.m_value = std::move(init);
-        return res;
-    }
-
-    /// @brief explicitly create a binary array (with subtype)
-    /// @sa https://json.nlohmann.me/api/basic_json/binary/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json binary(typename binary_t::container_type&& init, typename binary_t::subtype_type subtype)
-    {
-        auto res = basic_json();
-        res.m_type = value_t::binary;
-        res.m_value = binary_t(std::move(init), subtype);
-        return res;
-    }
-
-    /// @brief explicitly create an array from an initializer list
-    /// @sa https://json.nlohmann.me/api/basic_json/array/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json array(initializer_list_t init = {})
-    {
-        return basic_json(init, false, value_t::array);
-    }
-
-    /// @brief explicitly create an object from an initializer list
-    /// @sa https://json.nlohmann.me/api/basic_json/object/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json object(initializer_list_t init = {})
-    {
-        return basic_json(init, false, value_t::object);
-    }
-
-    /// @brief construct an array with count copies of given value
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(size_type cnt, const basic_json& val)
-        : m_type(value_t::array)
-    {
-        m_value.array = create<array_t>(cnt, val);
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief construct a JSON container given an iterator range
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    template < class InputIT, typename std::enable_if <
-                   std::is_same<InputIT, typename basic_json_t::iterator>::value ||
-                   std::is_same<InputIT, typename basic_json_t::const_iterator>::value, int >::type = 0 >
-    basic_json(InputIT first, InputIT last)
-    {
-        JSON_ASSERT(first.m_object != nullptr);
-        JSON_ASSERT(last.m_object != nullptr);
-
-        // make sure iterator fits the current value
-        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(201, "iterators are not compatible", nullptr));
-        }
-
-        // copy type from first iterator
-        m_type = first.m_object->m_type;
-
-        // check if iterator range is complete for primitive values
-        switch (m_type)
-        {
-            case value_t::boolean:
-            case value_t::number_float:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::string:
-            {
-                if (JSON_HEDLEY_UNLIKELY(!first.m_it.primitive_iterator.is_begin()
-                                         || !last.m_it.primitive_iterator.is_end()))
-                {
-                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", first.m_object));
-                }
-                break;
-            }
-
-            case value_t::null:
-            case value_t::object:
-            case value_t::array:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                break;
-        }
-
-        switch (m_type)
-        {
-            case value_t::number_integer:
-            {
-                m_value.number_integer = first.m_object->m_value.number_integer;
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                m_value.number_unsigned = first.m_object->m_value.number_unsigned;
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                m_value.number_float = first.m_object->m_value.number_float;
-                break;
-            }
-
-            case value_t::boolean:
-            {
-                m_value.boolean = first.m_object->m_value.boolean;
-                break;
-            }
-
-            case value_t::string:
-            {
-                m_value = *first.m_object->m_value.string;
-                break;
-            }
-
-            case value_t::object:
-            {
-                m_value.object = create<object_t>(first.m_it.object_iterator,
-                                                  last.m_it.object_iterator);
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_value.array = create<array_t>(first.m_it.array_iterator,
-                                                last.m_it.array_iterator);
-                break;
-            }
-
-            case value_t::binary:
-            {
-                m_value = *first.m_object->m_value.binary;
-                break;
-            }
-
-            case value_t::null:
-            case value_t::discarded:
-            default:
-                JSON_THROW(invalid_iterator::create(206, detail::concat("cannot construct with iterators from ", first.m_object->type_name()), first.m_object));
-        }
-
-        set_parents();
-        assert_invariant();
-    }
-
-
-    ///////////////////////////////////////
-    // other constructors and destructor //
-    ///////////////////////////////////////
-
-    template<typename JsonRef,
-             detail::enable_if_t<detail::conjunction<detail::is_json_ref<JsonRef>,
-                                 std::is_same<typename JsonRef::value_type, basic_json>>::value, int> = 0 >
-    basic_json(const JsonRef& ref) : basic_json(ref.moved_or_copied()) {}
-
-    /// @brief copy constructor
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(const basic_json& other)
-        : m_type(other.m_type)
-    {
-        // check of passed value is valid
-        other.assert_invariant();
-
-        switch (m_type)
-        {
-            case value_t::object:
-            {
-                m_value = *other.m_value.object;
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_value = *other.m_value.array;
-                break;
-            }
-
-            case value_t::string:
-            {
-                m_value = *other.m_value.string;
-                break;
-            }
-
-            case value_t::boolean:
-            {
-                m_value = other.m_value.boolean;
-                break;
-            }
-
-            case value_t::number_integer:
-            {
-                m_value = other.m_value.number_integer;
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                m_value = other.m_value.number_unsigned;
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                m_value = other.m_value.number_float;
-                break;
-            }
-
-            case value_t::binary:
-            {
-                m_value = *other.m_value.binary;
-                break;
-            }
-
-            case value_t::null:
-            case value_t::discarded:
-            default:
-                break;
-        }
-
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief move constructor
-    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
-    basic_json(basic_json&& other) noexcept
-        : m_type(std::move(other.m_type)),
-          m_value(std::move(other.m_value))
-    {
-        // check that passed value is valid
-        other.assert_invariant(false);
-
-        // invalidate payload
-        other.m_type = value_t::null;
-        other.m_value = {};
-
-        set_parents();
-        assert_invariant();
-    }
-
-    /// @brief copy assignment
-    /// @sa https://json.nlohmann.me/api/basic_json/operator=/
-    basic_json& operator=(basic_json other) noexcept (
-        std::is_nothrow_move_constructible<value_t>::value&&
-        std::is_nothrow_move_assignable<value_t>::value&&
-        std::is_nothrow_move_constructible<json_value>::value&&
-        std::is_nothrow_move_assignable<json_value>::value
-    )
-    {
-        // check that passed value is valid
-        other.assert_invariant();
-
-        using std::swap;
-        swap(m_type, other.m_type);
-        swap(m_value, other.m_value);
-
-        set_parents();
-        assert_invariant();
-        return *this;
-    }
-
-    /// @brief destructor
-    /// @sa https://json.nlohmann.me/api/basic_json/~basic_json/
-    ~basic_json() noexcept
-    {
-        assert_invariant(false);
-        m_value.destroy(m_type);
-    }
-
-    /// @}
-
-  public:
-    ///////////////////////
-    // object inspection //
-    ///////////////////////
-
-    /// @name object inspection
-    /// Functions to inspect the type of a JSON value.
-    /// @{
-
-    /// @brief serialization
-    /// @sa https://json.nlohmann.me/api/basic_json/dump/
-    string_t dump(const int indent = -1,
-                  const char indent_char = ' ',
-                  const bool ensure_ascii = false,
-                  const error_handler_t error_handler = error_handler_t::strict) const
-    {
-        string_t result;
-        serializer s(detail::output_adapter<char, string_t>(result), indent_char, error_handler);
-
-        if (indent >= 0)
-        {
-            s.dump(*this, true, ensure_ascii, static_cast<unsigned int>(indent));
-        }
-        else
-        {
-            s.dump(*this, false, ensure_ascii, 0);
-        }
-
-        return result;
-    }
-
-    /// @brief return the type of the JSON value (explicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/type/
-    constexpr value_t type() const noexcept
-    {
-        return m_type;
-    }
-
-    /// @brief return whether type is primitive
-    /// @sa https://json.nlohmann.me/api/basic_json/is_primitive/
-    constexpr bool is_primitive() const noexcept
-    {
-        return is_null() || is_string() || is_boolean() || is_number() || is_binary();
-    }
-
-    /// @brief return whether type is structured
-    /// @sa https://json.nlohmann.me/api/basic_json/is_structured/
-    constexpr bool is_structured() const noexcept
-    {
-        return is_array() || is_object();
-    }
-
-    /// @brief return whether value is null
-    /// @sa https://json.nlohmann.me/api/basic_json/is_null/
-    constexpr bool is_null() const noexcept
-    {
-        return m_type == value_t::null;
-    }
-
-    /// @brief return whether value is a boolean
-    /// @sa https://json.nlohmann.me/api/basic_json/is_boolean/
-    constexpr bool is_boolean() const noexcept
-    {
-        return m_type == value_t::boolean;
-    }
-
-    /// @brief return whether value is a number
-    /// @sa https://json.nlohmann.me/api/basic_json/is_number/
-    constexpr bool is_number() const noexcept
-    {
-        return is_number_integer() || is_number_float();
-    }
-
-    /// @brief return whether value is an integer number
-    /// @sa https://json.nlohmann.me/api/basic_json/is_number_integer/
-    constexpr bool is_number_integer() const noexcept
-    {
-        return m_type == value_t::number_integer || m_type == value_t::number_unsigned;
-    }
-
-    /// @brief return whether value is an unsigned integer number
-    /// @sa https://json.nlohmann.me/api/basic_json/is_number_unsigned/
-    constexpr bool is_number_unsigned() const noexcept
-    {
-        return m_type == value_t::number_unsigned;
-    }
-
-    /// @brief return whether value is a floating-point number
-    /// @sa https://json.nlohmann.me/api/basic_json/is_number_float/
-    constexpr bool is_number_float() const noexcept
-    {
-        return m_type == value_t::number_float;
-    }
-
-    /// @brief return whether value is an object
-    /// @sa https://json.nlohmann.me/api/basic_json/is_object/
-    constexpr bool is_object() const noexcept
-    {
-        return m_type == value_t::object;
-    }
-
-    /// @brief return whether value is an array
-    /// @sa https://json.nlohmann.me/api/basic_json/is_array/
-    constexpr bool is_array() const noexcept
-    {
-        return m_type == value_t::array;
-    }
-
-    /// @brief return whether value is a string
-    /// @sa https://json.nlohmann.me/api/basic_json/is_string/
-    constexpr bool is_string() const noexcept
-    {
-        return m_type == value_t::string;
-    }
-
-    /// @brief return whether value is a binary array
-    /// @sa https://json.nlohmann.me/api/basic_json/is_binary/
-    constexpr bool is_binary() const noexcept
-    {
-        return m_type == value_t::binary;
-    }
-
-    /// @brief return whether value is discarded
-    /// @sa https://json.nlohmann.me/api/basic_json/is_discarded/
-    constexpr bool is_discarded() const noexcept
-    {
-        return m_type == value_t::discarded;
-    }
-
-    /// @brief return the type of the JSON value (implicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_value_t/
-    constexpr operator value_t() const noexcept
-    {
-        return m_type;
-    }
-
-    /// @}
-
-  private:
-    //////////////////
-    // value access //
-    //////////////////
-
-    /// get a boolean (explicit)
-    boolean_t get_impl(boolean_t* /*unused*/) const
-    {
-        if (JSON_HEDLEY_LIKELY(is_boolean()))
-        {
-            return m_value.boolean;
-        }
-
-        JSON_THROW(type_error::create(302, detail::concat("type must be boolean, but is ", type_name()), this));
-    }
-
-    /// get a pointer to the value (object)
-    object_t* get_impl_ptr(object_t* /*unused*/) noexcept
-    {
-        return is_object() ? m_value.object : nullptr;
-    }
-
-    /// get a pointer to the value (object)
-    constexpr const object_t* get_impl_ptr(const object_t* /*unused*/) const noexcept
-    {
-        return is_object() ? m_value.object : nullptr;
-    }
-
-    /// get a pointer to the value (array)
-    array_t* get_impl_ptr(array_t* /*unused*/) noexcept
-    {
-        return is_array() ? m_value.array : nullptr;
-    }
-
-    /// get a pointer to the value (array)
-    constexpr const array_t* get_impl_ptr(const array_t* /*unused*/) const noexcept
-    {
-        return is_array() ? m_value.array : nullptr;
-    }
-
-    /// get a pointer to the value (string)
-    string_t* get_impl_ptr(string_t* /*unused*/) noexcept
-    {
-        return is_string() ? m_value.string : nullptr;
-    }
-
-    /// get a pointer to the value (string)
-    constexpr const string_t* get_impl_ptr(const string_t* /*unused*/) const noexcept
-    {
-        return is_string() ? m_value.string : nullptr;
-    }
-
-    /// get a pointer to the value (boolean)
-    boolean_t* get_impl_ptr(boolean_t* /*unused*/) noexcept
-    {
-        return is_boolean() ? &m_value.boolean : nullptr;
-    }
-
-    /// get a pointer to the value (boolean)
-    constexpr const boolean_t* get_impl_ptr(const boolean_t* /*unused*/) const noexcept
-    {
-        return is_boolean() ? &m_value.boolean : nullptr;
-    }
-
-    /// get a pointer to the value (integer number)
-    number_integer_t* get_impl_ptr(number_integer_t* /*unused*/) noexcept
-    {
-        return is_number_integer() ? &m_value.number_integer : nullptr;
-    }
-
-    /// get a pointer to the value (integer number)
-    constexpr const number_integer_t* get_impl_ptr(const number_integer_t* /*unused*/) const noexcept
-    {
-        return is_number_integer() ? &m_value.number_integer : nullptr;
-    }
-
-    /// get a pointer to the value (unsigned number)
-    number_unsigned_t* get_impl_ptr(number_unsigned_t* /*unused*/) noexcept
-    {
-        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
-    }
-
-    /// get a pointer to the value (unsigned number)
-    constexpr const number_unsigned_t* get_impl_ptr(const number_unsigned_t* /*unused*/) const noexcept
-    {
-        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
-    }
-
-    /// get a pointer to the value (floating-point number)
-    number_float_t* get_impl_ptr(number_float_t* /*unused*/) noexcept
-    {
-        return is_number_float() ? &m_value.number_float : nullptr;
-    }
-
-    /// get a pointer to the value (floating-point number)
-    constexpr const number_float_t* get_impl_ptr(const number_float_t* /*unused*/) const noexcept
-    {
-        return is_number_float() ? &m_value.number_float : nullptr;
-    }
-
-    /// get a pointer to the value (binary)
-    binary_t* get_impl_ptr(binary_t* /*unused*/) noexcept
-    {
-        return is_binary() ? m_value.binary : nullptr;
-    }
-
-    /// get a pointer to the value (binary)
-    constexpr const binary_t* get_impl_ptr(const binary_t* /*unused*/) const noexcept
-    {
-        return is_binary() ? m_value.binary : nullptr;
-    }
-
-    /*!
-    @brief helper function to implement get_ref()
-
-    This function helps to implement get_ref() without code duplication for
-    const and non-const overloads
-
-    @tparam ThisType will be deduced as `basic_json` or `const basic_json`
-
-    @throw type_error.303 if ReferenceType does not match underlying value
-    type of the current JSON
-    */
-    template<typename ReferenceType, typename ThisType>
-    static ReferenceType get_ref_impl(ThisType& obj)
-    {
-        // delegate the call to get_ptr<>()
-        auto* ptr = obj.template get_ptr<typename std::add_pointer<ReferenceType>::type>();
-
-        if (JSON_HEDLEY_LIKELY(ptr != nullptr))
-        {
-            return *ptr;
-        }
-
-        JSON_THROW(type_error::create(303, detail::concat("incompatible ReferenceType for get_ref, actual type is ", obj.type_name()), &obj));
-    }
-
-  public:
-    /// @name value access
-    /// Direct access to the stored value of a JSON value.
-    /// @{
-
-    /// @brief get a pointer value (implicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
-    template<typename PointerType, typename std::enable_if<
-                 std::is_pointer<PointerType>::value, int>::type = 0>
-    auto get_ptr() noexcept -> decltype(std::declval<basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
-    {
-        // delegate the call to get_impl_ptr<>()
-        return get_impl_ptr(static_cast<PointerType>(nullptr));
-    }
-
-    /// @brief get a pointer value (implicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
-    template < typename PointerType, typename std::enable_if <
-                   std::is_pointer<PointerType>::value&&
-                   std::is_const<typename std::remove_pointer<PointerType>::type>::value, int >::type = 0 >
-    constexpr auto get_ptr() const noexcept -> decltype(std::declval<const basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
-    {
-        // delegate the call to get_impl_ptr<>() const
-        return get_impl_ptr(static_cast<PointerType>(nullptr));
-    }
-
-  private:
-    /*!
-    @brief get a value (explicit)
-
-    Explicit type conversion between the JSON value and a compatible value
-    which is [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
-    and [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
-    The value is converted by calling the @ref json_serializer<ValueType>
-    `from_json()` method.
-
-    The function is equivalent to executing
-    @code {.cpp}
-    ValueType ret;
-    JSONSerializer<ValueType>::from_json(*this, ret);
-    return ret;
-    @endcode
-
-    This overloads is chosen if:
-    - @a ValueType is not @ref basic_json,
-    - @ref json_serializer<ValueType> has a `from_json()` method of the form
-      `void from_json(const basic_json&, ValueType&)`, and
-    - @ref json_serializer<ValueType> does not have a `from_json()` method of
-      the form `ValueType from_json(const basic_json&)`
-
-    @tparam ValueType the returned value type
-
-    @return copy of the JSON value, converted to @a ValueType
-
-    @throw what @ref json_serializer<ValueType> `from_json()` method throws
-
-    @liveexample{The example below shows several conversions from JSON values
-    to other types. There a few things to note: (1) Floating-point numbers can
-    be converted to integers\, (2) A JSON array can be converted to a standard
-    `std::vector<short>`\, (3) A JSON object can be converted to C++
-    associative containers such as `std::unordered_map<std::string\,
-    json>`.,get__ValueType_const}
-
-    @since version 2.1.0
-    */
-    template < typename ValueType,
-               detail::enable_if_t <
-                   detail::is_default_constructible<ValueType>::value&&
-                   detail::has_from_json<basic_json_t, ValueType>::value,
-                   int > = 0 >
-    ValueType get_impl(detail::priority_tag<0> /*unused*/) const noexcept(noexcept(
-                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), std::declval<ValueType&>())))
-    {
-        auto ret = ValueType();
-        JSONSerializer<ValueType>::from_json(*this, ret);
-        return ret;
-    }
-
-    /*!
-    @brief get a value (explicit); special case
-
-    Explicit type conversion between the JSON value and a compatible value
-    which is **not** [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
-    and **not** [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
-    The value is converted by calling the @ref json_serializer<ValueType>
-    `from_json()` method.
-
-    The function is equivalent to executing
-    @code {.cpp}
-    return JSONSerializer<ValueType>::from_json(*this);
-    @endcode
-
-    This overloads is chosen if:
-    - @a ValueType is not @ref basic_json and
-    - @ref json_serializer<ValueType> has a `from_json()` method of the form
-      `ValueType from_json(const basic_json&)`
-
-    @note If @ref json_serializer<ValueType> has both overloads of
-    `from_json()`, this one is chosen.
-
-    @tparam ValueType the returned value type
-
-    @return copy of the JSON value, converted to @a ValueType
-
-    @throw what @ref json_serializer<ValueType> `from_json()` method throws
-
-    @since version 2.1.0
-    */
-    template < typename ValueType,
-               detail::enable_if_t <
-                   detail::has_non_default_from_json<basic_json_t, ValueType>::value,
-                   int > = 0 >
-    ValueType get_impl(detail::priority_tag<1> /*unused*/) const noexcept(noexcept(
-                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>())))
-    {
-        return JSONSerializer<ValueType>::from_json(*this);
-    }
-
-    /*!
-    @brief get special-case overload
-
-    This overloads converts the current @ref basic_json in a different
-    @ref basic_json type
-
-    @tparam BasicJsonType == @ref basic_json
-
-    @return a copy of *this, converted into @a BasicJsonType
-
-    @complexity Depending on the implementation of the called `from_json()`
-                method.
-
-    @since version 3.2.0
-    */
-    template < typename BasicJsonType,
-               detail::enable_if_t <
-                   detail::is_basic_json<BasicJsonType>::value,
-                   int > = 0 >
-    BasicJsonType get_impl(detail::priority_tag<2> /*unused*/) const
-    {
-        return *this;
-    }
-
-    /*!
-    @brief get special-case overload
-
-    This overloads avoids a lot of template boilerplate, it can be seen as the
-    identity method
-
-    @tparam BasicJsonType == @ref basic_json
-
-    @return a copy of *this
-
-    @complexity Constant.
-
-    @since version 2.1.0
-    */
-    template<typename BasicJsonType,
-             detail::enable_if_t<
-                 std::is_same<BasicJsonType, basic_json_t>::value,
-                 int> = 0>
-    basic_json get_impl(detail::priority_tag<3> /*unused*/) const
-    {
-        return *this;
-    }
-
-    /*!
-    @brief get a pointer value (explicit)
-    @copydoc get()
-    */
-    template<typename PointerType,
-             detail::enable_if_t<
-                 std::is_pointer<PointerType>::value,
-                 int> = 0>
-    constexpr auto get_impl(detail::priority_tag<4> /*unused*/) const noexcept
-    -> decltype(std::declval<const basic_json_t&>().template get_ptr<PointerType>())
-    {
-        // delegate the call to get_ptr
-        return get_ptr<PointerType>();
-    }
-
-  public:
-    /*!
-    @brief get a (pointer) value (explicit)
-
-    Performs explicit type conversion between the JSON value and a compatible value if required.
-
-    - If the requested type is a pointer to the internally stored JSON value that pointer is returned.
-    No copies are made.
-
-    - If the requested type is the current @ref basic_json, or a different @ref basic_json convertible
-    from the current @ref basic_json.
-
-    - Otherwise the value is converted by calling the @ref json_serializer<ValueType> `from_json()`
-    method.
-
-    @tparam ValueTypeCV the provided value type
-    @tparam ValueType the returned value type
-
-    @return copy of the JSON value, converted to @tparam ValueType if necessary
-
-    @throw what @ref json_serializer<ValueType> `from_json()` method throws if conversion is required
-
-    @since version 2.1.0
-    */
-    template < typename ValueTypeCV, typename ValueType = detail::uncvref_t<ValueTypeCV>>
-#if defined(JSON_HAS_CPP_14)
-    constexpr
-#endif
-    auto get() const noexcept(
-    noexcept(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {})))
-    -> decltype(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {}))
-    {
-        // we cannot static_assert on ValueTypeCV being non-const, because
-        // there is support for get<const basic_json_t>(), which is why we
-        // still need the uncvref
-        static_assert(!std::is_reference<ValueTypeCV>::value,
-                      "get() cannot be used with reference types, you might want to use get_ref()");
-        return get_impl<ValueType>(detail::priority_tag<4> {});
-    }
-
-    /*!
-    @brief get a pointer value (explicit)
-
-    Explicit pointer access to the internally stored JSON value. No copies are
-    made.
-
-    @warning The pointer becomes invalid if the underlying JSON object
-    changes.
-
-    @tparam PointerType pointer type; must be a pointer to @ref array_t, @ref
-    object_t, @ref string_t, @ref boolean_t, @ref number_integer_t,
-    @ref number_unsigned_t, or @ref number_float_t.
-
-    @return pointer to the internally stored JSON value if the requested
-    pointer type @a PointerType fits to the JSON value; `nullptr` otherwise
-
-    @complexity Constant.
-
-    @liveexample{The example below shows how pointers to internal values of a
-    JSON value can be requested. Note that no type conversions are made and a
-    `nullptr` is returned if the value and the requested pointer type does not
-    match.,get__PointerType}
-
-    @sa see @ref get_ptr() for explicit pointer-member access
-
-    @since version 1.0.0
-    */
-    template<typename PointerType, typename std::enable_if<
-                 std::is_pointer<PointerType>::value, int>::type = 0>
-    auto get() noexcept -> decltype(std::declval<basic_json_t&>().template get_ptr<PointerType>())
-    {
-        // delegate the call to get_ptr
-        return get_ptr<PointerType>();
-    }
-
-    /// @brief get a value (explicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/get_to/
-    template < typename ValueType,
-               detail::enable_if_t <
-                   !detail::is_basic_json<ValueType>::value&&
-                   detail::has_from_json<basic_json_t, ValueType>::value,
-                   int > = 0 >
-    ValueType & get_to(ValueType& v) const noexcept(noexcept(
-                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), v)))
-    {
-        JSONSerializer<ValueType>::from_json(*this, v);
-        return v;
-    }
-
-    // specialization to allow calling get_to with a basic_json value
-    // see https://github.com/nlohmann/json/issues/2175
-    template<typename ValueType,
-             detail::enable_if_t <
-                 detail::is_basic_json<ValueType>::value,
-                 int> = 0>
-    ValueType & get_to(ValueType& v) const
-    {
-        v = *this;
-        return v;
-    }
-
-    template <
-        typename T, std::size_t N,
-        typename Array = T (&)[N], // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-        detail::enable_if_t <
-            detail::has_from_json<basic_json_t, Array>::value, int > = 0 >
-    Array get_to(T (&v)[N]) const // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
-    noexcept(noexcept(JSONSerializer<Array>::from_json(
-                          std::declval<const basic_json_t&>(), v)))
-    {
-        JSONSerializer<Array>::from_json(*this, v);
-        return v;
-    }
-
-    /// @brief get a reference value (implicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
-    template<typename ReferenceType, typename std::enable_if<
-                 std::is_reference<ReferenceType>::value, int>::type = 0>
-    ReferenceType get_ref()
-    {
-        // delegate call to get_ref_impl
-        return get_ref_impl<ReferenceType>(*this);
-    }
-
-    /// @brief get a reference value (implicit)
-    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
-    template < typename ReferenceType, typename std::enable_if <
-                   std::is_reference<ReferenceType>::value&&
-                   std::is_const<typename std::remove_reference<ReferenceType>::type>::value, int >::type = 0 >
-    ReferenceType get_ref() const
-    {
-        // delegate call to get_ref_impl
-        return get_ref_impl<ReferenceType>(*this);
-    }
-
-    /*!
-    @brief get a value (implicit)
-
-    Implicit type conversion between the JSON value and a compatible value.
-    The call is realized by calling @ref get() const.
-
-    @tparam ValueType non-pointer type compatible to the JSON value, for
-    instance `int` for JSON integer numbers, `bool` for JSON booleans, or
-    `std::vector` types for JSON arrays. The character type of @ref string_t
-    as well as an initializer list of this type is excluded to avoid
-    ambiguities as these types implicitly convert to `std::string`.
-
-    @return copy of the JSON value, converted to type @a ValueType
-
-    @throw type_error.302 in case passed type @a ValueType is incompatible
-    to the JSON value type (e.g., the JSON value is of type boolean, but a
-    string is requested); see example below
-
-    @complexity Linear in the size of the JSON value.
-
-    @liveexample{The example below shows several conversions from JSON values
-    to other types. There a few things to note: (1) Floating-point numbers can
-    be converted to integers\, (2) A JSON array can be converted to a standard
-    `std::vector<short>`\, (3) A JSON object can be converted to C++
-    associative containers such as `std::unordered_map<std::string\,
-    json>`.,operator__ValueType}
-
-    @since version 1.0.0
-    */
-    template < typename ValueType, typename std::enable_if <
-                   detail::conjunction <
-                       detail::negation<std::is_pointer<ValueType>>,
-                       detail::negation<std::is_same<ValueType, std::nullptr_t>>,
-                       detail::negation<std::is_same<ValueType, detail::json_ref<basic_json>>>,
-                                        detail::negation<std::is_same<ValueType, typename string_t::value_type>>,
-                                        detail::negation<detail::is_basic_json<ValueType>>,
-                                        detail::negation<std::is_same<ValueType, std::initializer_list<typename string_t::value_type>>>,
-#if defined(JSON_HAS_CPP_17) && (defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1910 && _MSC_VER <= 1914))
-                                                detail::negation<std::is_same<ValueType, std::string_view>>,
-#endif
-#if defined(JSON_HAS_CPP_17)
-                                                detail::negation<std::is_same<ValueType, std::any>>,
-#endif
-                                                detail::is_detected_lazy<detail::get_template_function, const basic_json_t&, ValueType>
-                                                >::value, int >::type = 0 >
-                                        JSON_EXPLICIT operator ValueType() const
-    {
-        // delegate the call to get<>() const
-        return get<ValueType>();
-    }
-
-    /// @brief get a binary value
-    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
-    binary_t& get_binary()
-    {
-        if (!is_binary())
-        {
-            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
-        }
-
-        return *get_ptr<binary_t*>();
-    }
-
-    /// @brief get a binary value
-    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
-    const binary_t& get_binary() const
-    {
-        if (!is_binary())
-        {
-            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
-        }
-
-        return *get_ptr<const binary_t*>();
-    }
-
-    /// @}
-
-
-    ////////////////////
-    // element access //
-    ////////////////////
-
-    /// @name element access
-    /// Access to the JSON value.
-    /// @{
-
-    /// @brief access specified array element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    reference at(size_type idx)
-    {
-        // at only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            JSON_TRY
-            {
-                return set_parent(m_value.array->at(idx));
-            }
-            JSON_CATCH (std::out_of_range&)
-            {
-                // create better exception explanation
-                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
-            }
-        }
-        else
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-    }
-
-    /// @brief access specified array element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    const_reference at(size_type idx) const
-    {
-        // at only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            JSON_TRY
-            {
-                return m_value.array->at(idx);
-            }
-            JSON_CATCH (std::out_of_range&)
-            {
-                // create better exception explanation
-                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
-            }
-        }
-        else
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-    }
-
-    /// @brief access specified object element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    reference at(const typename object_t::key_type& key)
-    {
-        // at only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-
-        auto it = m_value.object->find(key);
-        if (it == m_value.object->end())
-        {
-            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
-        }
-        return set_parent(it->second);
-    }
-
-    /// @brief access specified object element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    reference at(KeyType && key)
-    {
-        // at only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-
-        auto it = m_value.object->find(std::forward<KeyType>(key));
-        if (it == m_value.object->end())
-        {
-            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
-        }
-        return set_parent(it->second);
-    }
-
-    /// @brief access specified object element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    const_reference at(const typename object_t::key_type& key) const
-    {
-        // at only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-
-        auto it = m_value.object->find(key);
-        if (it == m_value.object->end())
-        {
-            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
-        }
-        return it->second;
-    }
-
-    /// @brief access specified object element with bounds checking
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    const_reference at(KeyType && key) const
-    {
-        // at only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
-        }
-
-        auto it = m_value.object->find(std::forward<KeyType>(key));
-        if (it == m_value.object->end())
-        {
-            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
-        }
-        return it->second;
-    }
-
-    /// @brief access specified array element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    reference operator[](size_type idx)
-    {
-        // implicitly convert null value to an empty array
-        if (is_null())
-        {
-            m_type = value_t::array;
-            m_value.array = create<array_t>();
-            assert_invariant();
-        }
-
-        // operator[] only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            // fill up array with null values if given idx is outside range
-            if (idx >= m_value.array->size())
-            {
-#if JSON_DIAGNOSTICS
-                // remember array size & capacity before resizing
-                const auto old_size = m_value.array->size();
-                const auto old_capacity = m_value.array->capacity();
-#endif
-                m_value.array->resize(idx + 1);
-
-#if JSON_DIAGNOSTICS
-                if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
-                {
-                    // capacity has changed: update all parents
-                    set_parents();
-                }
-                else
-                {
-                    // set parent for values added above
-                    set_parents(begin() + static_cast<typename iterator::difference_type>(old_size), static_cast<typename iterator::difference_type>(idx + 1 - old_size));
-                }
-#endif
-                assert_invariant();
-            }
-
-            return m_value.array->operator[](idx);
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
-    }
-
-    /// @brief access specified array element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    const_reference operator[](size_type idx) const
-    {
-        // const operator[] only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            return m_value.array->operator[](idx);
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
-    }
-
-    /// @brief access specified object element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    reference operator[](typename object_t::key_type key)
-    {
-        // implicitly convert null value to an empty object
-        if (is_null())
-        {
-            m_type = value_t::object;
-            m_value.object = create<object_t>();
-            assert_invariant();
-        }
-
-        // operator[] only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            auto result = m_value.object->emplace(std::move(key), nullptr);
-            return set_parent(result.first->second);
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
-    }
-
-    /// @brief access specified object element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    const_reference operator[](const typename object_t::key_type& key) const
-    {
-        // const operator[] only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            auto it = m_value.object->find(key);
-            JSON_ASSERT(it != m_value.object->end());
-            return it->second;
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
-    }
-
-    // these two functions resolve a (const) char * ambiguity affecting Clang and MSVC
-    // (they seemingly cannot be constrained to resolve the ambiguity)
-    template<typename T>
-    reference operator[](T* key)
-    {
-        return operator[](typename object_t::key_type(key));
-    }
-
-    template<typename T>
-    const_reference operator[](T* key) const
-    {
-        return operator[](typename object_t::key_type(key));
-    }
-
-    /// @brief access specified object element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
-    reference operator[](KeyType && key)
-    {
-        // implicitly convert null value to an empty object
-        if (is_null())
-        {
-            m_type = value_t::object;
-            m_value.object = create<object_t>();
-            assert_invariant();
-        }
-
-        // operator[] only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            auto result = m_value.object->emplace(std::forward<KeyType>(key), nullptr);
-            return set_parent(result.first->second);
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
-    }
-
-    /// @brief access specified object element
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
-    const_reference operator[](KeyType && key) const
-    {
-        // const operator[] only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            auto it = m_value.object->find(std::forward<KeyType>(key));
-            JSON_ASSERT(it != m_value.object->end());
-            return it->second;
-        }
-
-        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
-    }
-
-  private:
-    template<typename KeyType>
-    using is_comparable_with_object_key = detail::is_comparable <
-        object_comparator_t, const typename object_t::key_type&, KeyType >;
-
-    template<typename ValueType>
-    using value_return_type = std::conditional <
-        detail::is_c_string_uncvref<ValueType>::value,
-        string_t, typename std::decay<ValueType>::type >;
-
-  public:
-    /// @brief access specified object element with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, detail::enable_if_t <
-                   !detail::is_transparent<object_comparator_t>::value
-                   && detail::is_getable<basic_json_t, ValueType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ValueType value(const typename object_t::key_type& key, const ValueType& default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if key is found, return value and given default value otherwise
-            const auto it = find(key);
-            if (it != end())
-            {
-                return it->template get<ValueType>();
-            }
-
-            return default_value;
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    /// @brief access specified object element with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
-               detail::enable_if_t <
-                   !detail::is_transparent<object_comparator_t>::value
-                   && detail::is_getable<basic_json_t, ReturnType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ReturnType value(const typename object_t::key_type& key, ValueType && default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if key is found, return value and given default value otherwise
-            const auto it = find(key);
-            if (it != end())
-            {
-                return it->template get<ReturnType>();
-            }
-
-            return std::forward<ValueType>(default_value);
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    /// @brief access specified object element with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, class KeyType, detail::enable_if_t <
-                   detail::is_transparent<object_comparator_t>::value
-                   && !detail::is_json_pointer<KeyType>::value
-                   && is_comparable_with_object_key<KeyType>::value
-                   && detail::is_getable<basic_json_t, ValueType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ValueType value(KeyType && key, const ValueType& default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if key is found, return value and given default value otherwise
-            const auto it = find(std::forward<KeyType>(key));
-            if (it != end())
-            {
-                return it->template get<ValueType>();
-            }
-
-            return default_value;
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    /// @brief access specified object element via JSON Pointer with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, class KeyType, class ReturnType = typename value_return_type<ValueType>::type,
-               detail::enable_if_t <
-                   detail::is_transparent<object_comparator_t>::value
-                   && !detail::is_json_pointer<KeyType>::value
-                   && is_comparable_with_object_key<KeyType>::value
-                   && detail::is_getable<basic_json_t, ReturnType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ReturnType value(KeyType && key, ValueType && default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if key is found, return value and given default value otherwise
-            const auto it = find(std::forward<KeyType>(key));
-            if (it != end())
-            {
-                return it->template get<ReturnType>();
-            }
-
-            return std::forward<ValueType>(default_value);
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    /// @brief access specified object element via JSON Pointer with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, detail::enable_if_t <
-                   detail::is_getable<basic_json_t, ValueType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ValueType value(const json_pointer& ptr, const ValueType& default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if pointer resolves a value, return it or use default value
-            JSON_TRY
-            {
-                return ptr.get_checked(this).template get<ValueType>();
-            }
-            JSON_INTERNAL_CATCH (out_of_range&)
-            {
-                return default_value;
-            }
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    /// @brief access specified object element via JSON Pointer with default value
-    /// @sa https://json.nlohmann.me/api/basic_json/value/
-    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
-               detail::enable_if_t <
-                   detail::is_getable<basic_json_t, ReturnType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    ReturnType value(const json_pointer& ptr, ValueType && default_value) const
-    {
-        // value only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            // if pointer resolves a value, return it or use default value
-            JSON_TRY
-            {
-                return ptr.get_checked(this).template get<ReturnType>();
-            }
-            JSON_INTERNAL_CATCH (out_of_range&)
-            {
-                return std::forward<ValueType>(default_value);
-            }
-        }
-
-        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
-    }
-
-    template < class ValueType, class BasicJsonType, detail::enable_if_t <
-                   detail::is_basic_json<BasicJsonType>::value
-                   && detail::is_getable<basic_json_t, ValueType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    ValueType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, const ValueType& default_value) const
-    {
-        return value(ptr.convert(), default_value);
-    }
-
-    template < class ValueType, class BasicJsonType, class ReturnType = typename value_return_type<ValueType>::type,
-               detail::enable_if_t <
-                   detail::is_basic_json<BasicJsonType>::value
-                   && detail::is_getable<basic_json_t, ReturnType>::value
-                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    ReturnType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, ValueType && default_value) const
-    {
-        return value(ptr.convert(), std::forward<ValueType>(default_value));
-    }
-
-    /// @brief access the first element
-    /// @sa https://json.nlohmann.me/api/basic_json/front/
-    reference front()
-    {
-        return *begin();
-    }
-
-    /// @brief access the first element
-    /// @sa https://json.nlohmann.me/api/basic_json/front/
-    const_reference front() const
-    {
-        return *cbegin();
-    }
-
-    /// @brief access the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/back/
-    reference back()
-    {
-        auto tmp = end();
-        --tmp;
-        return *tmp;
-    }
-
-    /// @brief access the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/back/
-    const_reference back() const
-    {
-        auto tmp = cend();
-        --tmp;
-        return *tmp;
-    }
-
-    /// @brief remove element given an iterator
-    /// @sa https://json.nlohmann.me/api/basic_json/erase/
-    template < class IteratorType, detail::enable_if_t <
-                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
-                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
-    IteratorType erase(IteratorType pos)
-    {
-        // make sure iterator fits the current value
-        if (JSON_HEDLEY_UNLIKELY(this != pos.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
-        }
-
-        IteratorType result = end();
-
-        switch (m_type)
-        {
-            case value_t::boolean:
-            case value_t::number_float:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::string:
-            case value_t::binary:
-            {
-                if (JSON_HEDLEY_UNLIKELY(!pos.m_it.primitive_iterator.is_begin()))
-                {
-                    JSON_THROW(invalid_iterator::create(205, "iterator out of range", this));
-                }
-
-                if (is_string())
-                {
-                    AllocatorType<string_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
-                    m_value.string = nullptr;
-                }
-                else if (is_binary())
-                {
-                    AllocatorType<binary_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
-                    m_value.binary = nullptr;
-                }
-
-                m_type = value_t::null;
-                assert_invariant();
-                break;
-            }
-
-            case value_t::object:
-            {
-                result.m_it.object_iterator = m_value.object->erase(pos.m_it.object_iterator);
-                break;
-            }
-
-            case value_t::array:
-            {
-                result.m_it.array_iterator = m_value.array->erase(pos.m_it.array_iterator);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::discarded:
-            default:
-                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
-        }
-
-        return result;
-    }
-
-    /// @brief remove elements given an iterator range
-    /// @sa https://json.nlohmann.me/api/basic_json/erase/
-    template < class IteratorType, detail::enable_if_t <
-                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
-                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
-    IteratorType erase(IteratorType first, IteratorType last)
-    {
-        // make sure iterator fits the current value
-        if (JSON_HEDLEY_UNLIKELY(this != first.m_object || this != last.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(203, "iterators do not fit current value", this));
-        }
-
-        IteratorType result = end();
-
-        switch (m_type)
-        {
-            case value_t::boolean:
-            case value_t::number_float:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::string:
-            case value_t::binary:
-            {
-                if (JSON_HEDLEY_LIKELY(!first.m_it.primitive_iterator.is_begin()
-                                       || !last.m_it.primitive_iterator.is_end()))
-                {
-                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", this));
-                }
-
-                if (is_string())
-                {
-                    AllocatorType<string_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
-                    m_value.string = nullptr;
-                }
-                else if (is_binary())
-                {
-                    AllocatorType<binary_t> alloc;
-                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
-                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
-                    m_value.binary = nullptr;
-                }
-
-                m_type = value_t::null;
-                assert_invariant();
-                break;
-            }
-
-            case value_t::object:
-            {
-                result.m_it.object_iterator = m_value.object->erase(first.m_it.object_iterator,
-                                              last.m_it.object_iterator);
-                break;
-            }
-
-            case value_t::array:
-            {
-                result.m_it.array_iterator = m_value.array->erase(first.m_it.array_iterator,
-                                             last.m_it.array_iterator);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::discarded:
-            default:
-                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
-        }
-
-        return result;
-    }
-
-  private:
-    template < typename KeyType, detail::enable_if_t <
-                   detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
-    size_type erase_internal(KeyType && key)
-    {
-        // this erase only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
-        }
-
-        return m_value.object->erase(std::forward<KeyType>(key));
-    }
-
-    template < typename KeyType, detail::enable_if_t <
-                   !detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
-    size_type erase_internal(KeyType && key)
-    {
-        // this erase only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
-        }
-
-        const auto it = m_value.object->find(std::forward<KeyType>(key));
-        if (it != m_value.object->end())
-        {
-            m_value.object->erase(it);
-            return 1;
-        }
-        return 0;
-    }
-
-  public:
-
-    /// @brief remove element from a JSON object given a key
-    /// @sa https://json.nlohmann.me/api/basic_json/erase/
-    size_type erase(const typename object_t::key_type& key)
-    {
-        // the indirection via erase_internal() is added to avoid making this
-        // function a template and thus de-rank it during overload resolution
-        return erase_internal(key);
-    }
-
-    /// @brief remove element from a JSON object given a key
-    /// @sa https://json.nlohmann.me/api/basic_json/erase/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    size_type erase(KeyType && key)
-    {
-        return erase_internal(std::forward<KeyType>(key));
-    }
-
-    /// @brief remove element from a JSON array given an index
-    /// @sa https://json.nlohmann.me/api/basic_json/erase/
-    void erase(const size_type idx)
-    {
-        // this erase only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            if (JSON_HEDLEY_UNLIKELY(idx >= size()))
-            {
-                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
-            }
-
-            m_value.array->erase(m_value.array->begin() + static_cast<difference_type>(idx));
-        }
-        else
-        {
-            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
-        }
-    }
-
-    /// @}
-
-
-    ////////////
-    // lookup //
-    ////////////
-
-    /// @name lookup
-    /// @{
-
-    /// @brief find an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/find/
-    iterator find(const typename object_t::key_type& key)
-    {
-        auto result = end();
-
-        if (is_object())
-        {
-            result.m_it.object_iterator = m_value.object->find(key);
-        }
-
-        return result;
-    }
-
-    /// @brief find an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/find/
-    const_iterator find(const typename object_t::key_type& key) const
-    {
-        auto result = cend();
-
-        if (is_object())
-        {
-            result.m_it.object_iterator = m_value.object->find(key);
-        }
-
-        return result;
-    }
-
-    /// @brief find an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/find/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    iterator find(KeyType && key)
-    {
-        auto result = end();
-
-        if (is_object())
-        {
-            result.m_it.object_iterator = m_value.object->find(std::forward<KeyType>(key));
-        }
-
-        return result;
-    }
-
-    /// @brief find an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/find/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    const_iterator find(KeyType && key) const
-    {
-        auto result = cend();
-
-        if (is_object())
-        {
-            result.m_it.object_iterator = m_value.object->find(std::forward<KeyType>(key));
-        }
-
-        return result;
-    }
-
-    /// @brief returns the number of occurrences of a key in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/count/
-    size_type count(const typename object_t::key_type& key) const
-    {
-        // return 0 for all nonobject types
-        return is_object() ? m_value.object->count(key) : 0;
-    }
-
-    /// @brief returns the number of occurrences of a key in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/count/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    size_type count(KeyType && key) const
-    {
-        // return 0 for all nonobject types
-        return is_object() ? m_value.object->count(std::forward<KeyType>(key)) : 0;
-    }
-
-    /// @brief check the existence of an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/contains/
-    bool contains(const typename object_t::key_type& key) const
-    {
-        return is_object() && m_value.object->find(key) != m_value.object->end();
-    }
-
-    /// @brief check the existence of an element in a JSON object
-    /// @sa https://json.nlohmann.me/api/basic_json/contains/
-    template<class KeyType, detail::enable_if_t<
-                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
-    bool contains(KeyType && key) const
-    {
-        return is_object() && m_value.object->find(std::forward<KeyType>(key)) != m_value.object->end();
-    }
-
-    /// @brief check the existence of an element in a JSON object given a JSON pointer
-    /// @sa https://json.nlohmann.me/api/basic_json/contains/
-    bool contains(const json_pointer& ptr) const
-    {
-        return ptr.contains(this);
-    }
-
-    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    bool contains(const typename ::nlohmann::json_pointer<BasicJsonType>& ptr) const
-    {
-        return ptr.contains(this);
-    }
-
-    /// @}
-
-
-    ///////////////
-    // iterators //
-    ///////////////
-
-    /// @name iterators
-    /// @{
-
-    /// @brief returns an iterator to the first element
-    /// @sa https://json.nlohmann.me/api/basic_json/begin/
-    iterator begin() noexcept
-    {
-        iterator result(this);
-        result.set_begin();
-        return result;
-    }
-
-    /// @brief returns an iterator to the first element
-    /// @sa https://json.nlohmann.me/api/basic_json/begin/
-    const_iterator begin() const noexcept
-    {
-        return cbegin();
-    }
-
-    /// @brief returns a const iterator to the first element
-    /// @sa https://json.nlohmann.me/api/basic_json/cbegin/
-    const_iterator cbegin() const noexcept
-    {
-        const_iterator result(this);
-        result.set_begin();
-        return result;
-    }
-
-    /// @brief returns an iterator to one past the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/end/
-    iterator end() noexcept
-    {
-        iterator result(this);
-        result.set_end();
-        return result;
-    }
-
-    /// @brief returns an iterator to one past the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/end/
-    const_iterator end() const noexcept
-    {
-        return cend();
-    }
-
-    /// @brief returns an iterator to one past the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/cend/
-    const_iterator cend() const noexcept
-    {
-        const_iterator result(this);
-        result.set_end();
-        return result;
-    }
-
-    /// @brief returns an iterator to the reverse-beginning
-    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
-    reverse_iterator rbegin() noexcept
-    {
-        return reverse_iterator(end());
-    }
-
-    /// @brief returns an iterator to the reverse-beginning
-    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
-    const_reverse_iterator rbegin() const noexcept
-    {
-        return crbegin();
-    }
-
-    /// @brief returns an iterator to the reverse-end
-    /// @sa https://json.nlohmann.me/api/basic_json/rend/
-    reverse_iterator rend() noexcept
-    {
-        return reverse_iterator(begin());
-    }
-
-    /// @brief returns an iterator to the reverse-end
-    /// @sa https://json.nlohmann.me/api/basic_json/rend/
-    const_reverse_iterator rend() const noexcept
-    {
-        return crend();
-    }
-
-    /// @brief returns a const reverse iterator to the last element
-    /// @sa https://json.nlohmann.me/api/basic_json/crbegin/
-    const_reverse_iterator crbegin() const noexcept
-    {
-        return const_reverse_iterator(cend());
-    }
-
-    /// @brief returns a const reverse iterator to one before the first
-    /// @sa https://json.nlohmann.me/api/basic_json/crend/
-    const_reverse_iterator crend() const noexcept
-    {
-        return const_reverse_iterator(cbegin());
-    }
-
-  public:
-    /// @brief wrapper to access iterator member functions in range-based for
-    /// @sa https://json.nlohmann.me/api/basic_json/items/
-    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
-    ///             version 4.0.0 of the library. Please use @ref items() instead;
-    ///             that is, replace `json::iterator_wrapper(j)` with `j.items()`.
-    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
-    static iteration_proxy<iterator> iterator_wrapper(reference ref) noexcept
-    {
-        return ref.items();
-    }
-
-    /// @brief wrapper to access iterator member functions in range-based for
-    /// @sa https://json.nlohmann.me/api/basic_json/items/
-    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
-    ///         version 4.0.0 of the library. Please use @ref items() instead;
-    ///         that is, replace `json::iterator_wrapper(j)` with `j.items()`.
-    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
-    static iteration_proxy<const_iterator> iterator_wrapper(const_reference ref) noexcept
-    {
-        return ref.items();
-    }
-
-    /// @brief helper to access iterator member functions in range-based for
-    /// @sa https://json.nlohmann.me/api/basic_json/items/
-    iteration_proxy<iterator> items() noexcept
-    {
-        return iteration_proxy<iterator>(*this);
-    }
-
-    /// @brief helper to access iterator member functions in range-based for
-    /// @sa https://json.nlohmann.me/api/basic_json/items/
-    iteration_proxy<const_iterator> items() const noexcept
-    {
-        return iteration_proxy<const_iterator>(*this);
-    }
-
-    /// @}
-
-
-    //////////////
-    // capacity //
-    //////////////
-
-    /// @name capacity
-    /// @{
-
-    /// @brief checks whether the container is empty.
-    /// @sa https://json.nlohmann.me/api/basic_json/empty/
-    bool empty() const noexcept
-    {
-        switch (m_type)
-        {
-            case value_t::null:
-            {
-                // null values are empty
-                return true;
-            }
-
-            case value_t::array:
-            {
-                // delegate call to array_t::empty()
-                return m_value.array->empty();
-            }
-
-            case value_t::object:
-            {
-                // delegate call to object_t::empty()
-                return m_value.object->empty();
-            }
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                // all other types are nonempty
-                return false;
-            }
-        }
-    }
-
-    /// @brief returns the number of elements
-    /// @sa https://json.nlohmann.me/api/basic_json/size/
-    size_type size() const noexcept
-    {
-        switch (m_type)
-        {
-            case value_t::null:
-            {
-                // null values are empty
-                return 0;
-            }
-
-            case value_t::array:
-            {
-                // delegate call to array_t::size()
-                return m_value.array->size();
-            }
-
-            case value_t::object:
-            {
-                // delegate call to object_t::size()
-                return m_value.object->size();
-            }
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                // all other types have size 1
-                return 1;
-            }
-        }
-    }
-
-    /// @brief returns the maximum possible number of elements
-    /// @sa https://json.nlohmann.me/api/basic_json/max_size/
-    size_type max_size() const noexcept
-    {
-        switch (m_type)
-        {
-            case value_t::array:
-            {
-                // delegate call to array_t::max_size()
-                return m_value.array->max_size();
-            }
-
-            case value_t::object:
-            {
-                // delegate call to object_t::max_size()
-                return m_value.object->max_size();
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                // all other types have max_size() == size()
-                return size();
-            }
-        }
-    }
-
-    /// @}
-
-
-    ///////////////
-    // modifiers //
-    ///////////////
-
-    /// @name modifiers
-    /// @{
-
-    /// @brief clears the contents
-    /// @sa https://json.nlohmann.me/api/basic_json/clear/
-    void clear() noexcept
-    {
-        switch (m_type)
-        {
-            case value_t::number_integer:
-            {
-                m_value.number_integer = 0;
-                break;
-            }
-
-            case value_t::number_unsigned:
-            {
-                m_value.number_unsigned = 0;
-                break;
-            }
-
-            case value_t::number_float:
-            {
-                m_value.number_float = 0.0;
-                break;
-            }
-
-            case value_t::boolean:
-            {
-                m_value.boolean = false;
-                break;
-            }
-
-            case value_t::string:
-            {
-                m_value.string->clear();
-                break;
-            }
-
-            case value_t::binary:
-            {
-                m_value.binary->clear();
-                break;
-            }
-
-            case value_t::array:
-            {
-                m_value.array->clear();
-                break;
-            }
-
-            case value_t::object:
-            {
-                m_value.object->clear();
-                break;
-            }
-
-            case value_t::null:
-            case value_t::discarded:
-            default:
-                break;
-        }
-    }
-
-    /// @brief add an object to an array
-    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
-    void push_back(basic_json&& val)
-    {
-        // push_back only works for null objects or arrays
-        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
-        {
-            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
-        }
-
-        // transform null object into an array
-        if (is_null())
-        {
-            m_type = value_t::array;
-            m_value = value_t::array;
-            assert_invariant();
-        }
-
-        // add element to array (move semantics)
-        const auto old_capacity = m_value.array->capacity();
-        m_value.array->push_back(std::move(val));
-        set_parent(m_value.array->back(), old_capacity);
-        // if val is moved from, basic_json move constructor marks it null, so we do not call the destructor
-    }
-
-    /// @brief add an object to an array
-    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
-    reference operator+=(basic_json&& val)
-    {
-        push_back(std::move(val));
-        return *this;
-    }
-
-    /// @brief add an object to an array
-    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
-    void push_back(const basic_json& val)
-    {
-        // push_back only works for null objects or arrays
-        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
-        {
-            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
-        }
-
-        // transform null object into an array
-        if (is_null())
-        {
-            m_type = value_t::array;
-            m_value = value_t::array;
-            assert_invariant();
-        }
-
-        // add element to array
-        const auto old_capacity = m_value.array->capacity();
-        m_value.array->push_back(val);
-        set_parent(m_value.array->back(), old_capacity);
-    }
-
-    /// @brief add an object to an array
-    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
-    reference operator+=(const basic_json& val)
-    {
-        push_back(val);
-        return *this;
-    }
-
-    /// @brief add an object to an object
-    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
-    void push_back(const typename object_t::value_type& val)
-    {
-        // push_back only works for null objects or objects
-        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
-        {
-            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
-        }
-
-        // transform null object into an object
-        if (is_null())
-        {
-            m_type = value_t::object;
-            m_value = value_t::object;
-            assert_invariant();
-        }
-
-        // add element to object
-        auto res = m_value.object->insert(val);
-        set_parent(res.first->second);
-    }
-
-    /// @brief add an object to an object
-    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
-    reference operator+=(const typename object_t::value_type& val)
-    {
-        push_back(val);
-        return *this;
-    }
-
-    /// @brief add an object to an object
-    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
-    void push_back(initializer_list_t init)
-    {
-        if (is_object() && init.size() == 2 && (*init.begin())->is_string())
-        {
-            basic_json&& key = init.begin()->moved_or_copied();
-            push_back(typename object_t::value_type(
-                          std::move(key.get_ref<string_t&>()), (init.begin() + 1)->moved_or_copied()));
-        }
-        else
-        {
-            push_back(basic_json(init));
-        }
-    }
-
-    /// @brief add an object to an object
-    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
-    reference operator+=(initializer_list_t init)
-    {
-        push_back(init);
-        return *this;
-    }
-
-    /// @brief add an object to an array
-    /// @sa https://json.nlohmann.me/api/basic_json/emplace_back/
-    template<class... Args>
-    reference emplace_back(Args&& ... args)
-    {
-        // emplace_back only works for null objects or arrays
-        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
-        {
-            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace_back() with ", type_name()), this));
-        }
-
-        // transform null object into an array
-        if (is_null())
-        {
-            m_type = value_t::array;
-            m_value = value_t::array;
-            assert_invariant();
-        }
-
-        // add element to array (perfect forwarding)
-        const auto old_capacity = m_value.array->capacity();
-        m_value.array->emplace_back(std::forward<Args>(args)...);
-        return set_parent(m_value.array->back(), old_capacity);
-    }
-
-    /// @brief add an object to an object if key does not exist
-    /// @sa https://json.nlohmann.me/api/basic_json/emplace/
-    template<class... Args>
-    std::pair<iterator, bool> emplace(Args&& ... args)
-    {
-        // emplace only works for null objects or arrays
-        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
-        {
-            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace() with ", type_name()), this));
-        }
-
-        // transform null object into an object
-        if (is_null())
-        {
-            m_type = value_t::object;
-            m_value = value_t::object;
-            assert_invariant();
-        }
-
-        // add element to array (perfect forwarding)
-        auto res = m_value.object->emplace(std::forward<Args>(args)...);
-        set_parent(res.first->second);
-
-        // create result iterator and set iterator to the result of emplace
-        auto it = begin();
-        it.m_it.object_iterator = res.first;
-
-        // return pair of iterator and boolean
-        return {it, res.second};
-    }
-
-    /// Helper for insertion of an iterator
-    /// @note: This uses std::distance to support GCC 4.8,
-    ///        see https://github.com/nlohmann/json/pull/1257
-    template<typename... Args>
-    iterator insert_iterator(const_iterator pos, Args&& ... args)
-    {
-        iterator result(this);
-        JSON_ASSERT(m_value.array != nullptr);
-
-        auto insert_pos = std::distance(m_value.array->begin(), pos.m_it.array_iterator);
-        m_value.array->insert(pos.m_it.array_iterator, std::forward<Args>(args)...);
-        result.m_it.array_iterator = m_value.array->begin() + insert_pos;
-
-        // This could have been written as:
-        // result.m_it.array_iterator = m_value.array->insert(pos.m_it.array_iterator, cnt, val);
-        // but the return value of insert is missing in GCC 4.8, so it is written this way instead.
-
-        set_parents();
-        return result;
-    }
-
-    /// @brief inserts element into array
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    iterator insert(const_iterator pos, const basic_json& val)
-    {
-        // insert only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            // check if iterator pos fits to this JSON value
-            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
-            {
-                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
-            }
-
-            // insert to array and return iterator
-            return insert_iterator(pos, val);
-        }
-
-        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
-    }
-
-    /// @brief inserts element into array
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    iterator insert(const_iterator pos, basic_json&& val)
-    {
-        return insert(pos, val);
-    }
-
-    /// @brief inserts copies of element into array
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    iterator insert(const_iterator pos, size_type cnt, const basic_json& val)
-    {
-        // insert only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            // check if iterator pos fits to this JSON value
-            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
-            {
-                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
-            }
-
-            // insert to array and return iterator
-            return insert_iterator(pos, cnt, val);
-        }
-
-        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
-    }
-
-    /// @brief inserts range of elements into array
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    iterator insert(const_iterator pos, const_iterator first, const_iterator last)
-    {
-        // insert only works for arrays
-        if (JSON_HEDLEY_UNLIKELY(!is_array()))
-        {
-            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
-        }
-
-        // check if iterator pos fits to this JSON value
-        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
-        {
-            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
-        }
-
-        // check if range iterators belong to the same JSON object
-        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
-        }
-
-        if (JSON_HEDLEY_UNLIKELY(first.m_object == this))
-        {
-            JSON_THROW(invalid_iterator::create(211, "passed iterators may not belong to container", this));
-        }
-
-        // insert to array and return iterator
-        return insert_iterator(pos, first.m_it.array_iterator, last.m_it.array_iterator);
-    }
-
-    /// @brief inserts elements from initializer list into array
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    iterator insert(const_iterator pos, initializer_list_t ilist)
-    {
-        // insert only works for arrays
-        if (JSON_HEDLEY_UNLIKELY(!is_array()))
-        {
-            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
-        }
-
-        // check if iterator pos fits to this JSON value
-        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
-        {
-            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
-        }
-
-        // insert to array and return iterator
-        return insert_iterator(pos, ilist.begin(), ilist.end());
-    }
-
-    /// @brief inserts range of elements into object
-    /// @sa https://json.nlohmann.me/api/basic_json/insert/
-    void insert(const_iterator first, const_iterator last)
-    {
-        // insert only works for objects
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
-        }
-
-        // check if range iterators belong to the same JSON object
-        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
-        }
-
-        // passed iterators must belong to objects
-        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
-        {
-            JSON_THROW(invalid_iterator::create(202, "iterators first and last must point to objects", this));
-        }
-
-        m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator);
-    }
-
-    /// @brief updates a JSON object from another object, overwriting existing keys
-    /// @sa https://json.nlohmann.me/api/basic_json/update/
-    void update(const_reference j, bool merge_objects = false)
-    {
-        update(j.begin(), j.end(), merge_objects);
-    }
-
-    /// @brief updates a JSON object from another object, overwriting existing keys
-    /// @sa https://json.nlohmann.me/api/basic_json/update/
-    void update(const_iterator first, const_iterator last, bool merge_objects = false)
-    {
-        // implicitly convert null value to an empty object
-        if (is_null())
-        {
-            m_type = value_t::object;
-            m_value.object = create<object_t>();
-            assert_invariant();
-        }
-
-        if (JSON_HEDLEY_UNLIKELY(!is_object()))
-        {
-            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", type_name()), this));
-        }
-
-        // check if range iterators belong to the same JSON object
-        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
-        {
-            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
-        }
-
-        // passed iterators must belong to objects
-        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
-        {
-            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", first.m_object->type_name()), first.m_object));
-        }
-
-        for (auto it = first; it != last; ++it)
-        {
-            if (merge_objects && it.value().is_object())
-            {
-                auto it2 = m_value.object->find(it.key());
-                if (it2 != m_value.object->end())
-                {
-                    it2->second.update(it.value(), true);
-                    continue;
-                }
-            }
-            m_value.object->operator[](it.key()) = it.value();
-#if JSON_DIAGNOSTICS
-            m_value.object->operator[](it.key()).m_parent = this;
-#endif
-        }
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(reference other) noexcept (
-        std::is_nothrow_move_constructible<value_t>::value&&
-        std::is_nothrow_move_assignable<value_t>::value&&
-        std::is_nothrow_move_constructible<json_value>::value&&
-        std::is_nothrow_move_assignable<json_value>::value
-    )
-    {
-        std::swap(m_type, other.m_type);
-        std::swap(m_value, other.m_value);
-
-        set_parents();
-        other.set_parents();
-        assert_invariant();
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    friend void swap(reference left, reference right) noexcept (
-        std::is_nothrow_move_constructible<value_t>::value&&
-        std::is_nothrow_move_assignable<value_t>::value&&
-        std::is_nothrow_move_constructible<json_value>::value&&
-        std::is_nothrow_move_assignable<json_value>::value
-    )
-    {
-        left.swap(right);
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(array_t& other) // NOLINT(bugprone-exception-escape)
-    {
-        // swap only works for arrays
-        if (JSON_HEDLEY_LIKELY(is_array()))
-        {
-            using std::swap;
-            swap(*(m_value.array), other);
-        }
-        else
-        {
-            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(array_t&) with ", type_name()), this));
-        }
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(object_t& other) // NOLINT(bugprone-exception-escape)
-    {
-        // swap only works for objects
-        if (JSON_HEDLEY_LIKELY(is_object()))
-        {
-            using std::swap;
-            swap(*(m_value.object), other);
-        }
-        else
-        {
-            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(object_t&) with ", type_name()), this));
-        }
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(string_t& other) // NOLINT(bugprone-exception-escape)
-    {
-        // swap only works for strings
-        if (JSON_HEDLEY_LIKELY(is_string()))
-        {
-            using std::swap;
-            swap(*(m_value.string), other);
-        }
-        else
-        {
-            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(string_t&) with ", type_name()), this));
-        }
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(binary_t& other) // NOLINT(bugprone-exception-escape)
-    {
-        // swap only works for strings
-        if (JSON_HEDLEY_LIKELY(is_binary()))
-        {
-            using std::swap;
-            swap(*(m_value.binary), other);
-        }
-        else
-        {
-            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t&) with ", type_name()), this));
-        }
-    }
-
-    /// @brief exchanges the values
-    /// @sa https://json.nlohmann.me/api/basic_json/swap/
-    void swap(typename binary_t::container_type& other) // NOLINT(bugprone-exception-escape)
-    {
-        // swap only works for strings
-        if (JSON_HEDLEY_LIKELY(is_binary()))
-        {
-            using std::swap;
-            swap(*(m_value.binary), other);
-        }
-        else
-        {
-            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t::container_type&) with ", type_name()), this));
-        }
-    }
-
-    /// @}
-
-    //////////////////////////////////////////
-    // lexicographical comparison operators //
-    //////////////////////////////////////////
-
-    /// @name lexicographical comparison operators
-    /// @{
-
-    // note parentheses around operands are necessary; see
-    // https://github.com/nlohmann/json/issues/1530
-#define JSON_IMPLEMENT_OPERATOR(op, null_result, unordered_result, default_result)                       \
-    const auto lhs_type = lhs.type();                                                                    \
-    const auto rhs_type = rhs.type();                                                                    \
-    \
-    if (lhs_type == rhs_type) /* NOLINT(readability/braces) */                                           \
-    {                                                                                                    \
-        switch (lhs_type)                                                                                \
-        {                                                                                                \
-            case value_t::array:                                                                         \
-                return (*lhs.m_value.array) op (*rhs.m_value.array);                                     \
-                \
-            case value_t::object:                                                                        \
-                return (*lhs.m_value.object) op (*rhs.m_value.object);                                   \
-                \
-            case value_t::null:                                                                          \
-                return (null_result);                                                                    \
-                \
-            case value_t::string:                                                                        \
-                return (*lhs.m_value.string) op (*rhs.m_value.string);                                   \
-                \
-            case value_t::boolean:                                                                       \
-                return (lhs.m_value.boolean) op (rhs.m_value.boolean);                                   \
-                \
-            case value_t::number_integer:                                                                \
-                return (lhs.m_value.number_integer) op (rhs.m_value.number_integer);                     \
-                \
-            case value_t::number_unsigned:                                                               \
-                return (lhs.m_value.number_unsigned) op (rhs.m_value.number_unsigned);                   \
-                \
-            case value_t::number_float:                                                                  \
-                return (lhs.m_value.number_float) op (rhs.m_value.number_float);                         \
-                \
-            case value_t::binary:                                                                        \
-                return (*lhs.m_value.binary) op (*rhs.m_value.binary);                                   \
-                \
-            case value_t::discarded:                                                                     \
-            default:                                                                                     \
-                return (unordered_result);                                                               \
-        }                                                                                                \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float)                   \
-    {                                                                                                    \
-        return static_cast<number_float_t>(lhs.m_value.number_integer) op rhs.m_value.number_float;      \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer)                   \
-    {                                                                                                    \
-        return lhs.m_value.number_float op static_cast<number_float_t>(rhs.m_value.number_integer);      \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float)                  \
-    {                                                                                                    \
-        return static_cast<number_float_t>(lhs.m_value.number_unsigned) op rhs.m_value.number_float;     \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned)                  \
-    {                                                                                                    \
-        return lhs.m_value.number_float op static_cast<number_float_t>(rhs.m_value.number_unsigned);     \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer)                \
-    {                                                                                                    \
-        return static_cast<number_integer_t>(lhs.m_value.number_unsigned) op rhs.m_value.number_integer; \
-    }                                                                                                    \
-    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned)                \
-    {                                                                                                    \
-        return lhs.m_value.number_integer op static_cast<number_integer_t>(rhs.m_value.number_unsigned); \
-    }                                                                                                    \
-    else if(compares_unordered(lhs, rhs))\
-    {\
-        return (unordered_result);\
-    }\
-    \
-    return (default_result);
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    // returns true if:
-    // - any operand is NaN and the other operand is of number type
-    // - any operand is discarded
-    // in legacy mode, discarded values are considered ordered if
-    // an operation is computed as an odd number of inverses of others
-    static bool compares_unordered(const_reference lhs, const_reference rhs, bool inverse = false) noexcept
-    {
-        if ((lhs.is_number_float() && std::isnan(lhs.m_value.number_float) && rhs.is_number())
-                || (rhs.is_number_float() && std::isnan(rhs.m_value.number_float) && lhs.is_number()))
-        {
-            return true;
-        }
-#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
-        return (lhs.is_discarded() || rhs.is_discarded()) && !inverse;
-#else
-        static_cast<void>(inverse);
-        return lhs.is_discarded() || rhs.is_discarded();
-#endif
-    }
-
-  private:
-    bool compares_unordered(const_reference rhs, bool inverse = false) const noexcept
-    {
-        return compares_unordered(*this, rhs, inverse);
-    }
-
-  public:
-#if JSON_HAS_THREE_WAY_COMPARISON
-    /// @brief comparison: equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
-    bool operator==(const_reference rhs) const noexcept
-    {
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wfloat-equal"
-#endif
-        const_reference lhs = *this;
-        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-    }
-
-    /// @brief comparison: equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
-    template<typename ScalarType>
-    requires std::is_scalar_v<ScalarType>
-    bool operator==(ScalarType rhs) const noexcept
-    {
-        return *this == basic_json(rhs);
-    }
-
-    /// @brief comparison: not equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
-    bool operator!=(const_reference rhs) const noexcept
-    {
-        if (compares_unordered(rhs, true))
-        {
-            return false;
-        }
-        return !operator==(rhs);
-    }
-
-    /// @brief comparison: 3-way
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
-    std::partial_ordering operator<=>(const_reference rhs) const noexcept // *NOPAD*
-    {
-        const_reference lhs = *this;
-        // default_result is used if we cannot compare values. In that case,
-        // we compare types.
-        JSON_IMPLEMENT_OPERATOR(<=>, // *NOPAD*
-                                std::partial_ordering::equivalent,
-                                std::partial_ordering::unordered,
-                                lhs_type <=> rhs_type) // *NOPAD*
-    }
-
-    /// @brief comparison: 3-way
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
-    template<typename ScalarType>
-    requires std::is_scalar_v<ScalarType>
-    std::partial_ordering operator<=>(ScalarType rhs) const noexcept // *NOPAD*
-    {
-        return *this <=> basic_json(rhs); // *NOPAD*
-    }
-
-#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
-    // all operators that are computed as an odd number of inverses of others
-    // need to be overloaded to emulate the legacy comparison behavior
-
-    /// @brief comparison: less than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
-    bool operator<=(const_reference rhs) const noexcept
-    {
-        if (compares_unordered(rhs, true))
-        {
-            return false;
-        }
-        return !(rhs < *this);
-    }
-
-    /// @brief comparison: less than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
-    template<typename ScalarType>
-    requires std::is_scalar_v<ScalarType>
-    bool operator<=(ScalarType rhs) const noexcept
-    {
-        return *this <= basic_json(rhs);
-    }
-
-    /// @brief comparison: greater than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
-    bool operator>=(const_reference rhs) const noexcept
-    {
-        if (compares_unordered(rhs, true))
-        {
-            return false;
-        }
-        return !(*this < rhs);
-    }
-
-    /// @brief comparison: greater than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
-    template<typename ScalarType>
-    requires std::is_scalar_v<ScalarType>
-    bool operator>=(ScalarType rhs) const noexcept
-    {
-        return *this >= basic_json(rhs);
-    }
-#endif
-#else
-    /// @brief comparison: equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
-    friend bool operator==(const_reference lhs, const_reference rhs) noexcept
-    {
-#ifdef __GNUC__
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wfloat-equal"
-#endif
-        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
-#ifdef __GNUC__
-#pragma GCC diagnostic pop
-#endif
-    }
-
-    /// @brief comparison: equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator==(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs == basic_json(rhs);
-    }
-
-    /// @brief comparison: equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator==(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) == rhs;
-    }
-
-    /// @brief comparison: not equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
-    friend bool operator!=(const_reference lhs, const_reference rhs) noexcept
-    {
-        if (compares_unordered(lhs, rhs, true))
-        {
-            return false;
-        }
-        return !(lhs == rhs);
-    }
-
-    /// @brief comparison: not equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator!=(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs != basic_json(rhs);
-    }
-
-    /// @brief comparison: not equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator!=(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) != rhs;
-    }
-
-    /// @brief comparison: less than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
-    friend bool operator<(const_reference lhs, const_reference rhs) noexcept
-    {
-        // default_result is used if we cannot compare values. In that case,
-        // we compare types. Note we have to call the operator explicitly,
-        // because MSVC has problems otherwise.
-        JSON_IMPLEMENT_OPERATOR( <, false, false, operator<(lhs_type, rhs_type))
-    }
-
-    /// @brief comparison: less than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator<(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs < basic_json(rhs);
-    }
-
-    /// @brief comparison: less than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator<(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) < rhs;
-    }
-
-    /// @brief comparison: less than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
-    friend bool operator<=(const_reference lhs, const_reference rhs) noexcept
-    {
-        if (compares_unordered(lhs, rhs, true))
-        {
-            return false;
-        }
-        return !(rhs < lhs);
-    }
-
-    /// @brief comparison: less than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator<=(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs <= basic_json(rhs);
-    }
-
-    /// @brief comparison: less than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator<=(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) <= rhs;
-    }
-
-    /// @brief comparison: greater than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
-    friend bool operator>(const_reference lhs, const_reference rhs) noexcept
-    {
-        // double inverse
-        if (compares_unordered(lhs, rhs))
-        {
-            return false;
-        }
-        return !(lhs <= rhs);
-    }
-
-    /// @brief comparison: greater than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator>(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs > basic_json(rhs);
-    }
-
-    /// @brief comparison: greater than
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator>(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) > rhs;
-    }
-
-    /// @brief comparison: greater than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
-    friend bool operator>=(const_reference lhs, const_reference rhs) noexcept
-    {
-        if (compares_unordered(lhs, rhs, true))
-        {
-            return false;
-        }
-        return !(lhs < rhs);
-    }
-
-    /// @brief comparison: greater than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator>=(const_reference lhs, ScalarType rhs) noexcept
-    {
-        return lhs >= basic_json(rhs);
-    }
-
-    /// @brief comparison: greater than or equal
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
-    template<typename ScalarType, typename std::enable_if<
-                 std::is_scalar<ScalarType>::value, int>::type = 0>
-    friend bool operator>=(ScalarType lhs, const_reference rhs) noexcept
-    {
-        return basic_json(lhs) >= rhs;
-    }
-#endif
-
-#undef JSON_IMPLEMENT_OPERATOR
-
-    /// @}
-
-    ///////////////////
-    // serialization //
-    ///////////////////
-
-    /// @name serialization
-    /// @{
-#ifndef JSON_NO_IO
-    /// @brief serialize to stream
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
-    friend std::ostream& operator<<(std::ostream& o, const basic_json& j)
-    {
-        // read width member and use it as indentation parameter if nonzero
-        const bool pretty_print = o.width() > 0;
-        const auto indentation = pretty_print ? o.width() : 0;
-
-        // reset width to 0 for subsequent calls to this stream
-        o.width(0);
-
-        // do the actual serialization
-        serializer s(detail::output_adapter<char>(o), o.fill());
-        s.dump(j, pretty_print, false, static_cast<unsigned int>(indentation));
-        return o;
-    }
-
-    /// @brief serialize to stream
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
-    /// @deprecated This function is deprecated since 3.0.0 and will be removed in
-    ///             version 4.0.0 of the library. Please use
-    ///             operator<<(std::ostream&, const basic_json&) instead; that is,
-    ///             replace calls like `j >> o;` with `o << j;`.
-    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator<<(std::ostream&, const basic_json&))
-    friend std::ostream& operator>>(const basic_json& j, std::ostream& o)
-    {
-        return o << j;
-    }
-#endif  // JSON_NO_IO
-    /// @}
-
-
-    /////////////////////
-    // deserialization //
-    /////////////////////
-
-    /// @name deserialization
-    /// @{
-
-    /// @brief deserialize from a compatible input
-    /// @sa https://json.nlohmann.me/api/basic_json/parse/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json parse(InputType&& i,
-                            const parser_callback_t cb = nullptr,
-                            const bool allow_exceptions = true,
-                            const bool ignore_comments = false)
-    {
-        basic_json result;
-        parser(detail::input_adapter(std::forward<InputType>(i)), cb, allow_exceptions, ignore_comments).parse(true, result);
-        return result;
-    }
-
-    /// @brief deserialize from a pair of character iterators
-    /// @sa https://json.nlohmann.me/api/basic_json/parse/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json parse(IteratorType first,
-                            IteratorType last,
-                            const parser_callback_t cb = nullptr,
-                            const bool allow_exceptions = true,
-                            const bool ignore_comments = false)
-    {
-        basic_json result;
-        parser(detail::input_adapter(std::move(first), std::move(last)), cb, allow_exceptions, ignore_comments).parse(true, result);
-        return result;
-    }
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, parse(ptr, ptr + len))
-    static basic_json parse(detail::span_input_adapter&& i,
-                            const parser_callback_t cb = nullptr,
-                            const bool allow_exceptions = true,
-                            const bool ignore_comments = false)
-    {
-        basic_json result;
-        parser(i.get(), cb, allow_exceptions, ignore_comments).parse(true, result);
-        return result;
-    }
-
-    /// @brief check if the input is valid JSON
-    /// @sa https://json.nlohmann.me/api/basic_json/accept/
-    template<typename InputType>
-    static bool accept(InputType&& i,
-                       const bool ignore_comments = false)
-    {
-        return parser(detail::input_adapter(std::forward<InputType>(i)), nullptr, false, ignore_comments).accept(true);
-    }
-
-    /// @brief check if the input is valid JSON
-    /// @sa https://json.nlohmann.me/api/basic_json/accept/
-    template<typename IteratorType>
-    static bool accept(IteratorType first, IteratorType last,
-                       const bool ignore_comments = false)
-    {
-        return parser(detail::input_adapter(std::move(first), std::move(last)), nullptr, false, ignore_comments).accept(true);
-    }
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, accept(ptr, ptr + len))
-    static bool accept(detail::span_input_adapter&& i,
-                       const bool ignore_comments = false)
-    {
-        return parser(i.get(), nullptr, false, ignore_comments).accept(true);
-    }
-
-    /// @brief generate SAX events
-    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
-    template <typename InputType, typename SAX>
-    JSON_HEDLEY_NON_NULL(2)
-    static bool sax_parse(InputType&& i, SAX* sax,
-                          input_format_t format = input_format_t::json,
-                          const bool strict = true,
-                          const bool ignore_comments = false)
-    {
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        return format == input_format_t::json
-               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
-               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
-    }
-
-    /// @brief generate SAX events
-    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
-    template<class IteratorType, class SAX>
-    JSON_HEDLEY_NON_NULL(3)
-    static bool sax_parse(IteratorType first, IteratorType last, SAX* sax,
-                          input_format_t format = input_format_t::json,
-                          const bool strict = true,
-                          const bool ignore_comments = false)
-    {
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        return format == input_format_t::json
-               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
-               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
-    }
-
-    /// @brief generate SAX events
-    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
-    /// @deprecated This function is deprecated since 3.8.0 and will be removed in
-    ///             version 4.0.0 of the library. Please use
-    ///             sax_parse(ptr, ptr + len) instead.
-    template <typename SAX>
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, sax_parse(ptr, ptr + len, ...))
-    JSON_HEDLEY_NON_NULL(2)
-    static bool sax_parse(detail::span_input_adapter&& i, SAX* sax,
-                          input_format_t format = input_format_t::json,
-                          const bool strict = true,
-                          const bool ignore_comments = false)
-    {
-        auto ia = i.get();
-        return format == input_format_t::json
-               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
-               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
-    }
-#ifndef JSON_NO_IO
-    /// @brief deserialize from stream
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
-    /// @deprecated This stream operator is deprecated since 3.0.0 and will be removed in
-    ///             version 4.0.0 of the library. Please use
-    ///             operator>>(std::istream&, basic_json&) instead; that is,
-    ///             replace calls like `j << i;` with `i >> j;`.
-    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator>>(std::istream&, basic_json&))
-    friend std::istream& operator<<(basic_json& j, std::istream& i)
-    {
-        return operator>>(i, j);
-    }
-
-    /// @brief deserialize from stream
-    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
-    friend std::istream& operator>>(std::istream& i, basic_json& j)
-    {
-        parser(detail::input_adapter(i)).parse(false, j);
-        return i;
-    }
-#endif  // JSON_NO_IO
-    /// @}
-
-    ///////////////////////////
-    // convenience functions //
-    ///////////////////////////
-
-    /// @brief return the type as string
-    /// @sa https://json.nlohmann.me/api/basic_json/type_name/
-    JSON_HEDLEY_RETURNS_NON_NULL
-    const char* type_name() const noexcept
-    {
-        switch (m_type)
-        {
-            case value_t::null:
-                return "null";
-            case value_t::object:
-                return "object";
-            case value_t::array:
-                return "array";
-            case value_t::string:
-                return "string";
-            case value_t::boolean:
-                return "boolean";
-            case value_t::binary:
-                return "binary";
-            case value_t::discarded:
-                return "discarded";
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            default:
-                return "number";
-        }
-    }
-
-
-  JSON_PRIVATE_UNLESS_TESTED:
-    //////////////////////
-    // member variables //
-    //////////////////////
-
-    /// the type of the current element
-    value_t m_type = value_t::null;
-
-    /// the value of the current element
-    json_value m_value = {};
-
-#if JSON_DIAGNOSTICS
-    /// a pointer to a parent value (for debugging purposes)
-    basic_json* m_parent = nullptr;
-#endif
-
-    //////////////////////////////////////////
-    // binary serialization/deserialization //
-    //////////////////////////////////////////
-
-    /// @name binary serialization/deserialization support
-    /// @{
-
-  public:
-    /// @brief create a CBOR serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
-    static std::vector<std::uint8_t> to_cbor(const basic_json& j)
-    {
-        std::vector<std::uint8_t> result;
-        to_cbor(j, result);
-        return result;
-    }
-
-    /// @brief create a CBOR serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
-    static void to_cbor(const basic_json& j, detail::output_adapter<std::uint8_t> o)
-    {
-        binary_writer<std::uint8_t>(o).write_cbor(j);
-    }
-
-    /// @brief create a CBOR serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
-    static void to_cbor(const basic_json& j, detail::output_adapter<char> o)
-    {
-        binary_writer<char>(o).write_cbor(j);
-    }
-
-    /// @brief create a MessagePack serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
-    static std::vector<std::uint8_t> to_msgpack(const basic_json& j)
-    {
-        std::vector<std::uint8_t> result;
-        to_msgpack(j, result);
-        return result;
-    }
-
-    /// @brief create a MessagePack serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
-    static void to_msgpack(const basic_json& j, detail::output_adapter<std::uint8_t> o)
-    {
-        binary_writer<std::uint8_t>(o).write_msgpack(j);
-    }
-
-    /// @brief create a MessagePack serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
-    static void to_msgpack(const basic_json& j, detail::output_adapter<char> o)
-    {
-        binary_writer<char>(o).write_msgpack(j);
-    }
-
-    /// @brief create a UBJSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
-    static std::vector<std::uint8_t> to_ubjson(const basic_json& j,
-            const bool use_size = false,
-            const bool use_type = false)
-    {
-        std::vector<std::uint8_t> result;
-        to_ubjson(j, result, use_size, use_type);
-        return result;
-    }
-
-    /// @brief create a UBJSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
-    static void to_ubjson(const basic_json& j, detail::output_adapter<std::uint8_t> o,
-                          const bool use_size = false, const bool use_type = false)
-    {
-        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type);
-    }
-
-    /// @brief create a UBJSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
-    static void to_ubjson(const basic_json& j, detail::output_adapter<char> o,
-                          const bool use_size = false, const bool use_type = false)
-    {
-        binary_writer<char>(o).write_ubjson(j, use_size, use_type);
-    }
-
-    /// @brief create a BJData serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
-    static std::vector<std::uint8_t> to_bjdata(const basic_json& j,
-            const bool use_size = false,
-            const bool use_type = false)
-    {
-        std::vector<std::uint8_t> result;
-        to_bjdata(j, result, use_size, use_type);
-        return result;
-    }
-
-    /// @brief create a BJData serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
-    static void to_bjdata(const basic_json& j, detail::output_adapter<std::uint8_t> o,
-                          const bool use_size = false, const bool use_type = false)
-    {
-        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type, true, true);
-    }
-
-    /// @brief create a BJData serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
-    static void to_bjdata(const basic_json& j, detail::output_adapter<char> o,
-                          const bool use_size = false, const bool use_type = false)
-    {
-        binary_writer<char>(o).write_ubjson(j, use_size, use_type, true, true);
-    }
-
-    /// @brief create a BSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
-    static std::vector<std::uint8_t> to_bson(const basic_json& j)
-    {
-        std::vector<std::uint8_t> result;
-        to_bson(j, result);
-        return result;
-    }
-
-    /// @brief create a BSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
-    static void to_bson(const basic_json& j, detail::output_adapter<std::uint8_t> o)
-    {
-        binary_writer<std::uint8_t>(o).write_bson(j);
-    }
-
-    /// @brief create a BSON serialization of a given JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
-    static void to_bson(const basic_json& j, detail::output_adapter<char> o)
-    {
-        binary_writer<char>(o).write_bson(j);
-    }
-
-    /// @brief create a JSON value from an input in CBOR format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_cbor(InputType&& i,
-                                const bool strict = true,
-                                const bool allow_exceptions = true,
-                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in CBOR format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_cbor(IteratorType first, IteratorType last,
-                                const bool strict = true,
-                                const bool allow_exceptions = true,
-                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    template<typename T>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
-    static basic_json from_cbor(const T* ptr, std::size_t len,
-                                const bool strict = true,
-                                const bool allow_exceptions = true,
-                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
-    {
-        return from_cbor(ptr, ptr + len, strict, allow_exceptions, tag_handler);
-    }
-
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
-    static basic_json from_cbor(detail::span_input_adapter&& i,
-                                const bool strict = true,
-                                const bool allow_exceptions = true,
-                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = i.get();
-        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in MessagePack format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_msgpack(InputType&& i,
-                                   const bool strict = true,
-                                   const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in MessagePack format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_msgpack(IteratorType first, IteratorType last,
-                                   const bool strict = true,
-                                   const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    template<typename T>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
-    static basic_json from_msgpack(const T* ptr, std::size_t len,
-                                   const bool strict = true,
-                                   const bool allow_exceptions = true)
-    {
-        return from_msgpack(ptr, ptr + len, strict, allow_exceptions);
-    }
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
-    static basic_json from_msgpack(detail::span_input_adapter&& i,
-                                   const bool strict = true,
-                                   const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = i.get();
-        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in UBJSON format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_ubjson(InputType&& i,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in UBJSON format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_ubjson(IteratorType first, IteratorType last,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    template<typename T>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
-    static basic_json from_ubjson(const T* ptr, std::size_t len,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        return from_ubjson(ptr, ptr + len, strict, allow_exceptions);
-    }
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
-    static basic_json from_ubjson(detail::span_input_adapter&& i,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = i.get();
-        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-
-    /// @brief create a JSON value from an input in BJData format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_bjdata(InputType&& i,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in BJData format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_bjdata(IteratorType first, IteratorType last,
-                                  const bool strict = true,
-                                  const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in BSON format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
-    template<typename InputType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_bson(InputType&& i,
-                                const bool strict = true,
-                                const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::forward<InputType>(i));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    /// @brief create a JSON value from an input in BSON format
-    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
-    template<typename IteratorType>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json from_bson(IteratorType first, IteratorType last,
-                                const bool strict = true,
-                                const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = detail::input_adapter(std::move(first), std::move(last));
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-
-    template<typename T>
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
-    static basic_json from_bson(const T* ptr, std::size_t len,
-                                const bool strict = true,
-                                const bool allow_exceptions = true)
-    {
-        return from_bson(ptr, ptr + len, strict, allow_exceptions);
-    }
-
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
-    static basic_json from_bson(detail::span_input_adapter&& i,
-                                const bool strict = true,
-                                const bool allow_exceptions = true)
-    {
-        basic_json result;
-        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
-        auto ia = i.get();
-        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
-        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
-        return res ? result : basic_json(value_t::discarded);
-    }
-    /// @}
-
-    //////////////////////////
-    // JSON Pointer support //
-    //////////////////////////
-
-    /// @name JSON Pointer functions
-    /// @{
-
-    /// @brief access specified element via JSON Pointer
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    reference operator[](const json_pointer& ptr)
-    {
-        return ptr.get_unchecked(this);
-    }
-
-    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr)
-    {
-        return ptr.get_unchecked(this);
-    }
-
-    /// @brief access specified element via JSON Pointer
-    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
-    const_reference operator[](const json_pointer& ptr) const
-    {
-        return ptr.get_unchecked(this);
-    }
-
-    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    const_reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
-    {
-        return ptr.get_unchecked(this);
-    }
-
-    /// @brief access specified element via JSON Pointer
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    reference at(const json_pointer& ptr)
-    {
-        return ptr.get_checked(this);
-    }
-
-    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr)
-    {
-        return ptr.get_checked(this);
-    }
-
-    /// @brief access specified element via JSON Pointer
-    /// @sa https://json.nlohmann.me/api/basic_json/at/
-    const_reference at(const json_pointer& ptr) const
-    {
-        return ptr.get_checked(this);
-    }
-
-    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
-    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
-    const_reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
-    {
-        return ptr.get_checked(this);
-    }
-
-    /// @brief return flattened JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/flatten/
-    basic_json flatten() const
-    {
-        basic_json result(value_t::object);
-        json_pointer::flatten("", *this, result);
-        return result;
-    }
-
-    /// @brief unflatten a previously flattened JSON value
-    /// @sa https://json.nlohmann.me/api/basic_json/unflatten/
-    basic_json unflatten() const
-    {
-        return json_pointer::unflatten(*this);
-    }
-
-    /// @}
-
-    //////////////////////////
-    // JSON Patch functions //
-    //////////////////////////
-
-    /// @name JSON Patch functions
-    /// @{
-
-    /// @brief applies a JSON patch in-place without copying the object
-    /// @sa https://json.nlohmann.me/api/basic_json/patch/
-    void patch_inplace(const basic_json& json_patch)
-    {
-        basic_json& result = *this;
-        // the valid JSON Patch operations
-        enum class patch_operations {add, remove, replace, move, copy, test, invalid};
-
-        const auto get_op = [](const std::string & op)
-        {
-            if (op == "add")
-            {
-                return patch_operations::add;
-            }
-            if (op == "remove")
-            {
-                return patch_operations::remove;
-            }
-            if (op == "replace")
-            {
-                return patch_operations::replace;
-            }
-            if (op == "move")
-            {
-                return patch_operations::move;
-            }
-            if (op == "copy")
-            {
-                return patch_operations::copy;
-            }
-            if (op == "test")
-            {
-                return patch_operations::test;
-            }
-
-            return patch_operations::invalid;
-        };
-
-        // wrapper for "add" operation; add value at ptr
-        const auto operation_add = [&result](json_pointer & ptr, basic_json val)
-        {
-            // adding to the root of the target document means replacing it
-            if (ptr.empty())
-            {
-                result = val;
-                return;
-            }
-
-            // make sure the top element of the pointer exists
-            json_pointer top_pointer = ptr.top();
-            if (top_pointer != ptr)
-            {
-                result.at(top_pointer);
-            }
-
-            // get reference to parent of JSON pointer ptr
-            const auto last_path = ptr.back();
-            ptr.pop_back();
-            // parent must exist when performing patch add per RFC6902 specs
-            basic_json& parent = result.at(ptr);
-
-            switch (parent.m_type)
-            {
-                case value_t::null:
-                case value_t::object:
-                {
-                    // use operator[] to add value
-                    parent[last_path] = val;
-                    break;
-                }
-
-                case value_t::array:
-                {
-                    if (last_path == "-")
-                    {
-                        // special case: append to back
-                        parent.push_back(val);
-                    }
-                    else
-                    {
-                        const auto idx = json_pointer::template array_index<basic_json_t>(last_path);
-                        if (JSON_HEDLEY_UNLIKELY(idx > parent.size()))
-                        {
-                            // avoid undefined behavior
-                            JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), &parent));
-                        }
-
-                        // default case: insert add offset
-                        parent.insert(parent.begin() + static_cast<difference_type>(idx), val);
-                    }
-                    break;
-                }
-
-                // if there exists a parent it cannot be primitive
-                case value_t::string: // LCOV_EXCL_LINE
-                case value_t::boolean: // LCOV_EXCL_LINE
-                case value_t::number_integer: // LCOV_EXCL_LINE
-                case value_t::number_unsigned: // LCOV_EXCL_LINE
-                case value_t::number_float: // LCOV_EXCL_LINE
-                case value_t::binary: // LCOV_EXCL_LINE
-                case value_t::discarded: // LCOV_EXCL_LINE
-                default:            // LCOV_EXCL_LINE
-                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-            }
-        };
-
-        // wrapper for "remove" operation; remove value at ptr
-        const auto operation_remove = [this, &result](json_pointer & ptr)
-        {
-            // get reference to parent of JSON pointer ptr
-            const auto last_path = ptr.back();
-            ptr.pop_back();
-            basic_json& parent = result.at(ptr);
-
-            // remove child
-            if (parent.is_object())
-            {
-                // perform range check
-                auto it = parent.find(last_path);
-                if (JSON_HEDLEY_LIKELY(it != parent.end()))
-                {
-                    parent.erase(it);
-                }
-                else
-                {
-                    JSON_THROW(out_of_range::create(403, detail::concat("key '", last_path, "' not found"), this));
-                }
-            }
-            else if (parent.is_array())
-            {
-                // note erase performs range check
-                parent.erase(json_pointer::template array_index<basic_json_t>(last_path));
-            }
-        };
-
-        // type check: top level value must be an array
-        if (JSON_HEDLEY_UNLIKELY(!json_patch.is_array()))
-        {
-            JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &json_patch));
-        }
-
-        // iterate and apply the operations
-        for (const auto& val : json_patch)
-        {
-            // wrapper to get a value for an operation
-            const auto get_value = [&val](const std::string & op,
-                                          const std::string & member,
-                                          bool string_type) -> basic_json &
-            {
-                // find value
-                auto it = val.m_value.object->find(member);
-
-                // context-sensitive error message
-                const auto error_msg = (op == "op") ? "operation" : detail::concat("operation '", op, '\'');
-
-                // check if desired value is present
-                if (JSON_HEDLEY_UNLIKELY(it == val.m_value.object->end()))
-                {
-                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
-                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have member '", member, "'"), &val));
-                }
-
-                // check if result is of type string
-                if (JSON_HEDLEY_UNLIKELY(string_type && !it->second.is_string()))
-                {
-                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
-                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have string member '", member, "'"), &val));
-                }
-
-                // no error: return value
-                return it->second;
-            };
-
-            // type check: every element of the array must be an object
-            if (JSON_HEDLEY_UNLIKELY(!val.is_object()))
-            {
-                JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &val));
-            }
-
-            // collect mandatory members
-            const auto op = get_value("op", "op", true).template get<std::string>();
-            const auto path = get_value(op, "path", true).template get<std::string>();
-            json_pointer ptr(path);
-
-            switch (get_op(op))
-            {
-                case patch_operations::add:
-                {
-                    operation_add(ptr, get_value("add", "value", false));
-                    break;
-                }
-
-                case patch_operations::remove:
-                {
-                    operation_remove(ptr);
-                    break;
-                }
-
-                case patch_operations::replace:
-                {
-                    // the "path" location must exist - use at()
-                    result.at(ptr) = get_value("replace", "value", false);
-                    break;
-                }
-
-                case patch_operations::move:
-                {
-                    const auto from_path = get_value("move", "from", true).template get<std::string>();
-                    json_pointer from_ptr(from_path);
-
-                    // the "from" location must exist - use at()
-                    basic_json v = result.at(from_ptr);
-
-                    // The move operation is functionally identical to a
-                    // "remove" operation on the "from" location, followed
-                    // immediately by an "add" operation at the target
-                    // location with the value that was just removed.
-                    operation_remove(from_ptr);
-                    operation_add(ptr, v);
-                    break;
-                }
-
-                case patch_operations::copy:
-                {
-                    const auto from_path = get_value("copy", "from", true).template get<std::string>();
-                    const json_pointer from_ptr(from_path);
-
-                    // the "from" location must exist - use at()
-                    basic_json v = result.at(from_ptr);
-
-                    // The copy is functionally identical to an "add"
-                    // operation at the target location using the value
-                    // specified in the "from" member.
-                    operation_add(ptr, v);
-                    break;
-                }
-
-                case patch_operations::test:
-                {
-                    bool success = false;
-                    JSON_TRY
-                    {
-                        // check if "value" matches the one at "path"
-                        // the "path" location must exist - use at()
-                        success = (result.at(ptr) == get_value("test", "value", false));
-                    }
-                    JSON_INTERNAL_CATCH (out_of_range&)
-                    {
-                        // ignore out of range errors: success remains false
-                    }
-
-                    // throw an exception if test fails
-                    if (JSON_HEDLEY_UNLIKELY(!success))
-                    {
-                        JSON_THROW(other_error::create(501, detail::concat("unsuccessful: ", val.dump()), &val));
-                    }
-
-                    break;
-                }
-
-                case patch_operations::invalid:
-                default:
-                {
-                    // op must be "add", "remove", "replace", "move", "copy", or
-                    // "test"
-                    JSON_THROW(parse_error::create(105, 0, detail::concat("operation value '", op, "' is invalid"), &val));
-                }
-            }
-        }
-    }
-
-    /// @brief applies a JSON patch to a copy of the current object
-    /// @sa https://json.nlohmann.me/api/basic_json/patch/
-    basic_json patch(const basic_json& json_patch) const
-    {
-        basic_json result = *this;
-        result.patch_inplace(json_patch);
-        return result;
-    }
-
-    /// @brief creates a diff as a JSON patch
-    /// @sa https://json.nlohmann.me/api/basic_json/diff/
-    JSON_HEDLEY_WARN_UNUSED_RESULT
-    static basic_json diff(const basic_json& source, const basic_json& target,
-                           const std::string& path = "")
-    {
-        // the patch
-        basic_json result(value_t::array);
-
-        // if the values are the same, return empty patch
-        if (source == target)
-        {
-            return result;
-        }
-
-        if (source.type() != target.type())
-        {
-            // different types: replace value
-            result.push_back(
-            {
-                {"op", "replace"}, {"path", path}, {"value", target}
-            });
-            return result;
-        }
-
-        switch (source.type())
-        {
-            case value_t::array:
-            {
-                // first pass: traverse common elements
-                std::size_t i = 0;
-                while (i < source.size() && i < target.size())
-                {
-                    // recursive call to compare array values at index i
-                    auto temp_diff = diff(source[i], target[i], detail::concat(path, '/', std::to_string(i)));
-                    result.insert(result.end(), temp_diff.begin(), temp_diff.end());
-                    ++i;
-                }
-
-                // We now reached the end of at least one array
-                // in a second pass, traverse the remaining elements
-
-                // remove my remaining elements
-                const auto end_index = static_cast<difference_type>(result.size());
-                while (i < source.size())
-                {
-                    // add operations in reverse order to avoid invalid
-                    // indices
-                    result.insert(result.begin() + end_index, object(
-                    {
-                        {"op", "remove"},
-                        {"path", detail::concat(path, '/', std::to_string(i))}
-                    }));
-                    ++i;
-                }
-
-                // add other remaining elements
-                while (i < target.size())
-                {
-                    result.push_back(
-                    {
-                        {"op", "add"},
-                        {"path", detail::concat(path, "/-")},
-                        {"value", target[i]}
-                    });
-                    ++i;
-                }
-
-                break;
-            }
-
-            case value_t::object:
-            {
-                // first pass: traverse this object's elements
-                for (auto it = source.cbegin(); it != source.cend(); ++it)
-                {
-                    // escape the key name to be used in a JSON patch
-                    const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
-
-                    if (target.find(it.key()) != target.end())
-                    {
-                        // recursive call to compare object values at key it
-                        auto temp_diff = diff(it.value(), target[it.key()], path_key);
-                        result.insert(result.end(), temp_diff.begin(), temp_diff.end());
-                    }
-                    else
-                    {
-                        // found a key that is not in o -> remove it
-                        result.push_back(object(
-                        {
-                            {"op", "remove"}, {"path", path_key}
-                        }));
-                    }
-                }
-
-                // second pass: traverse other object's elements
-                for (auto it = target.cbegin(); it != target.cend(); ++it)
-                {
-                    if (source.find(it.key()) == source.end())
-                    {
-                        // found a key that is not in this -> add it
-                        const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
-                        result.push_back(
-                        {
-                            {"op", "add"}, {"path", path_key},
-                            {"value", it.value()}
-                        });
-                    }
-                }
-
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                // both primitive type: replace value
-                result.push_back(
-                {
-                    {"op", "replace"}, {"path", path}, {"value", target}
-                });
-                break;
-            }
-        }
-
-        return result;
-    }
-    /// @}
-
-    ////////////////////////////////
-    // JSON Merge Patch functions //
-    ////////////////////////////////
-
-    /// @name JSON Merge Patch functions
-    /// @{
-
-    /// @brief applies a JSON Merge Patch
-    /// @sa https://json.nlohmann.me/api/basic_json/merge_patch/
-    void merge_patch(const basic_json& apply_patch)
-    {
-        if (apply_patch.is_object())
-        {
-            if (!is_object())
-            {
-                *this = object();
-            }
-            for (auto it = apply_patch.begin(); it != apply_patch.end(); ++it)
-            {
-                if (it.value().is_null())
-                {
-                    erase(it.key());
-                }
-                else
-                {
-                    operator[](it.key()).merge_patch(it.value());
-                }
-            }
-        }
-        else
-        {
-            *this = apply_patch;
-        }
-    }
-
-    /// @}
-};
-
-/// @brief user-defined to_string function for JSON values
-/// @sa https://json.nlohmann.me/api/basic_json/to_string/
-NLOHMANN_BASIC_JSON_TPL_DECLARATION
-std::string to_string(const NLOHMANN_BASIC_JSON_TPL& j)
-{
-    return j.dump();
-}
-
-inline namespace literals
-{
-inline namespace json_literals
-{
-
-/// @brief user-defined string literal for JSON values
-/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/
-JSON_HEDLEY_NON_NULL(1)
-inline nlohmann::json operator "" _json(const char* s, std::size_t n)
-{
-    return nlohmann::json::parse(s, s + n);
-}
-
-/// @brief user-defined string literal for JSON pointer
-/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/
-JSON_HEDLEY_NON_NULL(1)
-inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n)
-{
-    return nlohmann::json::json_pointer(std::string(s, n));
-}
-
-}  // namespace json_literals
-}  // namespace literals
-NLOHMANN_JSON_NAMESPACE_END
-
-///////////////////////
-// nonmember support //
-///////////////////////
-
-namespace std // NOLINT(cert-dcl58-cpp)
-{
-
-/// @brief hash value for JSON objects
-/// @sa https://json.nlohmann.me/api/basic_json/std_hash/
-NLOHMANN_BASIC_JSON_TPL_DECLARATION
-struct hash<nlohmann::NLOHMANN_BASIC_JSON_TPL>
-{
-    std::size_t operator()(const nlohmann::NLOHMANN_BASIC_JSON_TPL& j) const
-    {
-        return nlohmann::detail::hash(j);
-    }
-};
-
-// specialization for std::less<value_t>
-template<>
-struct less< ::nlohmann::detail::value_t> // do not remove the space after '<', see https://github.com/nlohmann/json/pull/679
-{
-    /*!
-    @brief compare two value_t enum values
-    @since version 3.0.0
-    */
-    bool operator()(::nlohmann::detail::value_t lhs,
-                    ::nlohmann::detail::value_t rhs) const noexcept
-    {
-#if JSON_HAS_THREE_WAY_COMPARISON
-        return std::is_lt(lhs <=> rhs); // *NOPAD*
-#else
-        return ::nlohmann::detail::operator<(lhs, rhs);
-#endif
-    }
-};
-
-// C++20 prohibit function specialization in the std namespace.
-#ifndef JSON_HAS_CPP_20
-
-/// @brief exchanges the values of two JSON objects
-/// @sa https://json.nlohmann.me/api/basic_json/std_swap/
-NLOHMANN_BASIC_JSON_TPL_DECLARATION
-inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept(  // NOLINT(readability-inconsistent-declaration-parameter-name)
-    is_nothrow_move_constructible<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value&&                          // NOLINT(misc-redundant-expression)
-    is_nothrow_move_assignable<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value)
-{
-    j1.swap(j2);
-}
-
-#endif
-
-}  // namespace std
-
-#if JSON_USE_GLOBAL_UDLS
-    using nlohmann::literals::json_literals::operator "" _json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers)
-    using nlohmann::literals::json_literals::operator "" _json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers)
-#endif
-
-// #include <nlohmann/detail/macro_unscope.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-// restore clang diagnostic settings
-#if defined(__clang__)
-    #pragma clang diagnostic pop
-#endif
-
-// clean up
-#undef JSON_ASSERT
-#undef JSON_INTERNAL_CATCH
-#undef JSON_THROW
-#undef JSON_PRIVATE_UNLESS_TESTED
-#undef NLOHMANN_BASIC_JSON_TPL_DECLARATION
-#undef NLOHMANN_BASIC_JSON_TPL
-#undef JSON_EXPLICIT
-#undef NLOHMANN_CAN_CALL_STD_FUNC_IMPL
-#undef JSON_INLINE_VARIABLE
-#undef JSON_NO_UNIQUE_ADDRESS
-#undef JSON_DISABLE_ENUM_SERIALIZATION
-#undef JSON_USE_GLOBAL_UDLS
-
-#ifndef JSON_TEST_KEEP_MACROS
-    #undef JSON_CATCH
-    #undef JSON_TRY
-    #undef JSON_HAS_CPP_11
-    #undef JSON_HAS_CPP_14
-    #undef JSON_HAS_CPP_17
-    #undef JSON_HAS_CPP_20
-    #undef JSON_HAS_FILESYSTEM
-    #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
-    #undef JSON_HAS_THREE_WAY_COMPARISON
-    #undef JSON_HAS_RANGES
-    #undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
-#endif
-
-// #include <nlohmann/thirdparty/hedley/hedley_undef.hpp>
-//     __ _____ _____ _____
-//  __|  |   __|     |   | |  JSON for Modern C++
-// |  |  |__   |  |  | | | |  version 3.11.2
-// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
-//
-// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
-// SPDX-License-Identifier: MIT
-
-
-
-#undef JSON_HEDLEY_ALWAYS_INLINE
-#undef JSON_HEDLEY_ARM_VERSION
-#undef JSON_HEDLEY_ARM_VERSION_CHECK
-#undef JSON_HEDLEY_ARRAY_PARAM
-#undef JSON_HEDLEY_ASSUME
-#undef JSON_HEDLEY_BEGIN_C_DECLS
-#undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
-#undef JSON_HEDLEY_CLANG_HAS_BUILTIN
-#undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
-#undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
-#undef JSON_HEDLEY_CLANG_HAS_EXTENSION
-#undef JSON_HEDLEY_CLANG_HAS_FEATURE
-#undef JSON_HEDLEY_CLANG_HAS_WARNING
-#undef JSON_HEDLEY_COMPCERT_VERSION
-#undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
-#undef JSON_HEDLEY_CONCAT
-#undef JSON_HEDLEY_CONCAT3
-#undef JSON_HEDLEY_CONCAT3_EX
-#undef JSON_HEDLEY_CONCAT_EX
-#undef JSON_HEDLEY_CONST
-#undef JSON_HEDLEY_CONSTEXPR
-#undef JSON_HEDLEY_CONST_CAST
-#undef JSON_HEDLEY_CPP_CAST
-#undef JSON_HEDLEY_CRAY_VERSION
-#undef JSON_HEDLEY_CRAY_VERSION_CHECK
-#undef JSON_HEDLEY_C_DECL
-#undef JSON_HEDLEY_DEPRECATED
-#undef JSON_HEDLEY_DEPRECATED_FOR
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
-#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
-#undef JSON_HEDLEY_DIAGNOSTIC_POP
-#undef JSON_HEDLEY_DIAGNOSTIC_PUSH
-#undef JSON_HEDLEY_DMC_VERSION
-#undef JSON_HEDLEY_DMC_VERSION_CHECK
-#undef JSON_HEDLEY_EMPTY_BASES
-#undef JSON_HEDLEY_EMSCRIPTEN_VERSION
-#undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
-#undef JSON_HEDLEY_END_C_DECLS
-#undef JSON_HEDLEY_FLAGS
-#undef JSON_HEDLEY_FLAGS_CAST
-#undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
-#undef JSON_HEDLEY_GCC_HAS_BUILTIN
-#undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
-#undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
-#undef JSON_HEDLEY_GCC_HAS_EXTENSION
-#undef JSON_HEDLEY_GCC_HAS_FEATURE
-#undef JSON_HEDLEY_GCC_HAS_WARNING
-#undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
-#undef JSON_HEDLEY_GCC_VERSION
-#undef JSON_HEDLEY_GCC_VERSION_CHECK
-#undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
-#undef JSON_HEDLEY_GNUC_HAS_BUILTIN
-#undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
-#undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
-#undef JSON_HEDLEY_GNUC_HAS_EXTENSION
-#undef JSON_HEDLEY_GNUC_HAS_FEATURE
-#undef JSON_HEDLEY_GNUC_HAS_WARNING
-#undef JSON_HEDLEY_GNUC_VERSION
-#undef JSON_HEDLEY_GNUC_VERSION_CHECK
-#undef JSON_HEDLEY_HAS_ATTRIBUTE
-#undef JSON_HEDLEY_HAS_BUILTIN
-#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
-#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
-#undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
-#undef JSON_HEDLEY_HAS_EXTENSION
-#undef JSON_HEDLEY_HAS_FEATURE
-#undef JSON_HEDLEY_HAS_WARNING
-#undef JSON_HEDLEY_IAR_VERSION
-#undef JSON_HEDLEY_IAR_VERSION_CHECK
-#undef JSON_HEDLEY_IBM_VERSION
-#undef JSON_HEDLEY_IBM_VERSION_CHECK
-#undef JSON_HEDLEY_IMPORT
-#undef JSON_HEDLEY_INLINE
-#undef JSON_HEDLEY_INTEL_CL_VERSION
-#undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
-#undef JSON_HEDLEY_INTEL_VERSION
-#undef JSON_HEDLEY_INTEL_VERSION_CHECK
-#undef JSON_HEDLEY_IS_CONSTANT
-#undef JSON_HEDLEY_IS_CONSTEXPR_
-#undef JSON_HEDLEY_LIKELY
-#undef JSON_HEDLEY_MALLOC
-#undef JSON_HEDLEY_MCST_LCC_VERSION
-#undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
-#undef JSON_HEDLEY_MESSAGE
-#undef JSON_HEDLEY_MSVC_VERSION
-#undef JSON_HEDLEY_MSVC_VERSION_CHECK
-#undef JSON_HEDLEY_NEVER_INLINE
-#undef JSON_HEDLEY_NON_NULL
-#undef JSON_HEDLEY_NO_ESCAPE
-#undef JSON_HEDLEY_NO_RETURN
-#undef JSON_HEDLEY_NO_THROW
-#undef JSON_HEDLEY_NULL
-#undef JSON_HEDLEY_PELLES_VERSION
-#undef JSON_HEDLEY_PELLES_VERSION_CHECK
-#undef JSON_HEDLEY_PGI_VERSION
-#undef JSON_HEDLEY_PGI_VERSION_CHECK
-#undef JSON_HEDLEY_PREDICT
-#undef JSON_HEDLEY_PRINTF_FORMAT
-#undef JSON_HEDLEY_PRIVATE
-#undef JSON_HEDLEY_PUBLIC
-#undef JSON_HEDLEY_PURE
-#undef JSON_HEDLEY_REINTERPRET_CAST
-#undef JSON_HEDLEY_REQUIRE
-#undef JSON_HEDLEY_REQUIRE_CONSTEXPR
-#undef JSON_HEDLEY_REQUIRE_MSG
-#undef JSON_HEDLEY_RESTRICT
-#undef JSON_HEDLEY_RETURNS_NON_NULL
-#undef JSON_HEDLEY_SENTINEL
-#undef JSON_HEDLEY_STATIC_ASSERT
-#undef JSON_HEDLEY_STATIC_CAST
-#undef JSON_HEDLEY_STRINGIFY
-#undef JSON_HEDLEY_STRINGIFY_EX
-#undef JSON_HEDLEY_SUNPRO_VERSION
-#undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
-#undef JSON_HEDLEY_TINYC_VERSION
-#undef JSON_HEDLEY_TINYC_VERSION_CHECK
-#undef JSON_HEDLEY_TI_ARMCL_VERSION
-#undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
-#undef JSON_HEDLEY_TI_CL2000_VERSION
-#undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
-#undef JSON_HEDLEY_TI_CL430_VERSION
-#undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
-#undef JSON_HEDLEY_TI_CL6X_VERSION
-#undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
-#undef JSON_HEDLEY_TI_CL7X_VERSION
-#undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
-#undef JSON_HEDLEY_TI_CLPRU_VERSION
-#undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
-#undef JSON_HEDLEY_TI_VERSION
-#undef JSON_HEDLEY_TI_VERSION_CHECK
-#undef JSON_HEDLEY_UNAVAILABLE
-#undef JSON_HEDLEY_UNLIKELY
-#undef JSON_HEDLEY_UNPREDICTABLE
-#undef JSON_HEDLEY_UNREACHABLE
-#undef JSON_HEDLEY_UNREACHABLE_RETURN
-#undef JSON_HEDLEY_VERSION
-#undef JSON_HEDLEY_VERSION_DECODE_MAJOR
-#undef JSON_HEDLEY_VERSION_DECODE_MINOR
-#undef JSON_HEDLEY_VERSION_DECODE_REVISION
-#undef JSON_HEDLEY_VERSION_ENCODE
-#undef JSON_HEDLEY_WARNING
-#undef JSON_HEDLEY_WARN_UNUSED_RESULT
-#undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
-#undef JSON_HEDLEY_FALL_THROUGH
-
-
-
-#endif  // INCLUDE_NLOHMANN_JSON_HPP_
diff --git a/llm/ext_server/server.cpp b/llm/ext_server/server.cpp
index 3bc012521..6ce457ae4 100644
--- a/llm/ext_server/server.cpp
+++ b/llm/ext_server/server.cpp
@@ -41,8 +41,10 @@
 
 #if defined(_WIN32)
 #include <windows.h>
+#include <errhandlingapi.h>
 #endif
 
+#include <algorithm>
 #include <cstddef>
 #include <thread>
 #include <chrono>
@@ -260,7 +262,7 @@ struct server_slot {
        char buffer[512];
         double t_token = t_prompt_processing / n_prompt_tokens_processed;
         double n_tokens_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed;
-        sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
+        snprintf(buffer, sizeof(buffer), "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
                 t_prompt_processing, n_prompt_tokens_processed,
                 t_token, n_tokens_second);
         LOG_DEBUG(buffer, {
@@ -274,7 +276,7 @@ struct server_slot {
 
         t_token = t_token_generation / n_decoded;
         n_tokens_second = 1e3 / t_token_generation * n_decoded;
-        sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
+        snprintf(buffer, sizeof(buffer), "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
                 t_token_generation, n_decoded,
                 t_token, n_tokens_second);
         LOG_DEBUG(buffer, {
@@ -286,7 +288,7 @@ struct server_slot {
             {"n_tokens_second",    n_tokens_second},
         });
 
-        sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
+        snprintf(buffer, sizeof(buffer), "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
         LOG_DEBUG(buffer, {
             {"slot_id",             id},
             {"task_id",             task_id},
@@ -401,7 +403,9 @@ struct llama_server_context
             }
         }
 
-        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        auto init_result = llama_init_from_gpt_params(params);
+        model = init_result.model;
+        ctx = init_result.context;
         if (model == nullptr)
         {
             LOG_ERROR("unable to load model", {{"model", params.model}});
@@ -421,7 +425,7 @@ struct llama_server_context
 
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_should_add_bos_token(model);
+        add_bos_token = llama_add_bos_token(model);
 
         return true;
     }
@@ -909,7 +913,9 @@ struct llama_server_context
         slot.sampled = result.tok;
 
         // search stop word and delete it
-        slot.generated_text += token_str;
+        if (!llama_token_is_eog(model, result.tok))
+            slot.generated_text += token_str;
+
         slot.has_next_token = true;
 
         if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
@@ -950,30 +956,36 @@ struct llama_server_context
         if (!incomplete)
         {
             size_t pos = std::min(slot.n_sent_text, slot.generated_text.size());
-            const std::string str_test = slot.generated_text.substr(pos);
-            bool is_stop_full = false;
-            size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
-            if (stop_pos != std::string::npos)
-            {
-                is_stop_full = true;
-                slot.generated_text.erase(
-                    slot.generated_text.begin() + pos + stop_pos,
-                    slot.generated_text.end());
-                pos = std::min(slot.n_sent_text, slot.generated_text.size());
-            }
-            else
-            {
-                is_stop_full = false;
-                stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
-            }
 
-            // check if there is any token to predict
-            if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
-            {
-                // no send the stop word in the response
-                result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
-                slot.n_sent_text += result.text_to_send.size();
-                // add the token to slot queue and cache
+            if (!llama_token_is_eog(model, result.tok)) {
+                const std::string str_test = slot.generated_text.substr(pos);
+                bool is_stop_full = false;
+                size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
+                if (stop_pos != std::string::npos)
+                {
+                    is_stop_full = true;
+                    slot.generated_text.erase(
+                        slot.generated_text.begin() + pos + stop_pos,
+                        slot.generated_text.end());
+                    pos = std::min(slot.n_sent_text, slot.generated_text.size());
+                }
+                else
+                {
+                    is_stop_full = false;
+                    stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
+                }
+
+                // check if there is any token to predict
+                if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
+                {
+                    // no send the stop word in the response
+                    result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
+                    slot.n_sent_text += result.text_to_send.size();
+                    // add the token to slot queue and cache
+                }
+            } else {
+                    result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
+                    slot.n_sent_text += result.text_to_send.size();
             }
 
             if (slot.params.stream)
@@ -1027,7 +1039,7 @@ struct llama_server_context
                 continue;
             }
 
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
+            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
                 LOG_TEE("Error processing the given image");
                 return false;
             }
@@ -1113,9 +1125,7 @@ struct llama_server_context
             {"multimodal", multimodal}
         };
 
-        if (!llama_token_is_eog(model, tkn.tok)) {
-            res.result_json["content"] = tkn.text_to_send;
-        }
+        res.result_json["content"] = tkn.text_to_send;
 
         if (slot.sparams.n_probs > 0)
         {
@@ -1382,12 +1392,56 @@ struct llama_server_context
         }
     }
 
+    std::string common_prefix(const std::string& str1, const std::string& str2) {
+        auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
+        return std::string(str1.begin(), mismatch_pair.first);
+    }
+
+    // Find the slot that has the greatest common prefix
+    server_slot *prefix_slot(const json &prompt) {
+        if (!prompt.is_string()) {
+            return nullptr;
+        }
+
+        std::string prompt_str = prompt.get<std::string>();
+        server_slot *slot = nullptr;
+        size_t longest = 0;
+
+        for (server_slot &s : slots) {
+            if (s.available() && s.prompt.is_string()) {
+                std::string s_prompt = s.prompt.get<std::string>();
+                std::string prefix = common_prefix(s_prompt, prompt_str);
+
+                if (prefix.size() > longest) {
+                    slot = &s;
+                    longest = prefix.size();
+                }
+            }
+        }
+
+        if (!slot) {
+            return get_slot(-1);
+        }
+
+        LOG_DEBUG("slot with common prefix found", {{
+            "slot_id", slot->id,
+            "characters", longest
+        }});
+        return slot;
+    }
+
     void process_single_task(task_server& task)
     {
         switch (task.type)
         {
             case TASK_TYPE_COMPLETION: {
-                server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
+                server_slot *slot = nullptr;
+                if (task.embedding_mode) {
+                    // Embedding seq_id (aka slot id) must always be <= token length, so always use slot 0
+                    slot = slots[0].available() ? &slots[0] : nullptr;
+                } else {
+                    slot = prefix_slot(task.data["prompt"]);
+                }
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
@@ -1650,22 +1704,8 @@ struct llama_server_context
                     }
                     slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
 
-                    char buf[256];
-                    llama_model_meta_val_str(model, "general.architecture", buf, 256);
-                    bool gemma2 = strcmp(buf, "gemma2") == 0;
-
-                    int32_t truncate_at = slot.n_ctx;
-
-                    // truncate at 2/3 of the context length for gemma2 models
-                    // as they do not support context shifts (from the sliding window implementation).
-                    // this way, prompts that almost fit the context length can still generate a full
-                    // response without a sudden stop from hitting the context limit
-                    if (gemma2) {
-                        truncate_at = 2 * slot.n_ctx / 3;
-                    }
-
                     // if input prompt is too big, truncate it, if group attention self-extend is disabled
-                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at)
+                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
                         const int n_shift = n_left / 2;
@@ -1693,19 +1733,6 @@ struct llama_server_context
                         GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                     }
 
-                    // Models with sliding window attention do not work with context shifts, so
-                    // limit their prediction to the context length
-                    if (gemma2) {
-                        int32_t limit = slot.n_ctx - slot.n_prompt_tokens;
-                        slot.n_predict = limit;
-                        slot.params.n_predict = limit;
-                        LOG_INFO("model does not support sliding window, limiting generation", {
-                            {"n_ctx", slot.n_ctx},
-                            {"n_prompt_tokens", slot.n_prompt_tokens},
-                            {"n_predict", slot.n_predict}
-                        });
-                    }
-
                     if (!slot.params.cache_prompt)
                     {
                         llama_sampling_reset(slot.ctx_sampling);
@@ -1732,7 +1759,7 @@ struct llama_server_context
                             slot.n_past -= 1;
                         }
 
-                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
+                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens;
 
                         if (slot.ga_n != 1)
                         {
@@ -1993,7 +2020,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("options:\n");
     printf("  -h, --help                show this help message and exit\n");
     printf("  -v, --verbose             verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -t N, --threads N         number of threads to use during computation (default: %d)\n", params.cpuparams.n_threads);
     printf("  -tb N, --threads-batch N  number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  --threads-http N          number of threads in the http server pool to process requests (default: max(hardware concurrency - 1, --parallel N + 2))\n");
     printf("  -c N, --ctx-size N        size of the prompt context (default: %d)\n", params.n_ctx);
@@ -2266,7 +2293,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.n_threads = std::stoi(argv[i]);
+            params.cpuparams.n_threads = std::stoi(argv[i]);
         }
         else if (arg == "--grp-attn-n" || arg == "-gan")
         {
@@ -2294,7 +2321,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.n_threads_batch = std::stoi(argv[i]);
+            params.cpuparams_batch.n_threads = std::stoi(argv[i]);
         }
         else if (arg == "--threads-http")
         {
@@ -2407,7 +2434,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(argv[i], 1.0f);
+            params.lora_adapters.push_back({
+                std::string(argv[i]),
+                1.0,
+            });
             params.use_mmap = false;
         }
         else if (arg == "--lora-scaled")
@@ -2423,18 +2453,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
+            params.lora_adapters.push_back({
+                lora_adapter,
+                std::stof(argv[i])
+            });
             params.use_mmap = false;
         }
-        else if (arg == "--lora-base")
-        {
-            if (++i >= argc)
-            {
-                invalid_param = true;
-                break;
-            }
-            params.lora_base = argv[i];
-        }
         else if (arg == "-v" || arg == "--verbose")
         {
             server_verbose = true;
@@ -2608,6 +2632,11 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
         params.kv_overrides.back().key[0] = 0;
     }
 
+    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
     if (invalid_param)
     {
         fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
@@ -2726,6 +2755,9 @@ int wmain(int argc, wchar_t **wargv) {
     for (int i = 0; i < argc; ++i) {
         argv[i] = wchar_to_char(wargv[i]);
     }
+
+    // Adjust error mode to avoid error dialog after we start.
+    SetErrorMode(SEM_FAILCRITICALERRORS);
 #else
 int main(int argc, char **argv) {
 #endif
@@ -2754,8 +2786,8 @@ int main(int argc, char **argv) {
                             {"commit", LLAMA_COMMIT}});
 
     LOG_INFO("system info", {
-                                {"n_threads", params.n_threads},
-                                {"n_threads_batch", params.n_threads_batch},
+                                {"n_threads", params.cpuparams.n_threads},
+                                {"n_threads_batch", params.cpuparams_batch.n_threads},
                                 {"total_threads", std::thread::hardware_concurrency()},
                                 {"system_info", llama_print_system_info()},
                             });
@@ -3177,19 +3209,10 @@ int main(int argc, char **argv) {
                     prompt = "";
                 }
 
-                json image_data;
-                if (body.count("image_data") != 0) {
-                    image_data = body["image_data"];
-                }
-                else
-                {
-                    image_data = "";
-                }
-
                 // create and queue the task
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);
-                llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
+                llama.request_completion(task_id, {{"prompt", prompt}}, true, -1);
 
                 // get the result
                 task_result result = llama.queue_results.recv(task_id);
diff --git a/llm/generate/gen_common.sh b/llm/generate/gen_common.sh
index da1b06882..3825c155a 100644
--- a/llm/generate/gen_common.sh
+++ b/llm/generate/gen_common.sh
@@ -9,11 +9,14 @@ init_vars() {
         ARCH="arm64"
         ;;
     *)
-        ARCH=$(uname -m | sed -e "s/aarch64/arm64/g")
+        echo "GOARCH must be set"
+        echo "this script is meant to be run from within go generate"
+        exit 1
+        ;;
     esac
 
     LLAMACPP_DIR=../llama.cpp
-    CMAKE_DEFS=""
+    CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on"
     CMAKE_TARGETS="--target ollama_llama_server"
     if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
@@ -27,6 +30,8 @@ init_vars() {
         WHOLE_ARCHIVE="-Wl,-force_load"
         NO_WHOLE_ARCHIVE=""
         GCC_ARCH="-arch ${ARCH}"
+        DIST_BASE=../../dist/darwin-${GOARCH}/
+        PAYLOAD_BASE=../../build/darwin/${GOARCH}
         ;;
     "Linux")
         LIB_EXT="so"
@@ -35,6 +40,8 @@ init_vars() {
 
         # Cross compiling not supported on linux - Use docker
         GCC_ARCH=""
+        DIST_BASE=../../dist/linux-${GOARCH}/
+        PAYLOAD_BASE=../../build/linux/${GOARCH}
         ;;
     *)
         ;;
@@ -42,6 +49,8 @@ init_vars() {
     if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
         CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
     fi
+    GZIP=$(command -v pigz 2>/dev/null || echo "gzip")
+    RUNNER_BASE="${DIST_BASE}/lib/ollama/runners"
 }
 
 git_module_setup() {
@@ -60,51 +69,68 @@ git_module_setup() {
 }
 
 apply_patches() {
-    # Wire up our CMakefile
-    if ! grep ollama ${LLAMACPP_DIR}/CMakeLists.txt; then
-        echo 'add_subdirectory(../ext_server ext_server) # ollama' >>${LLAMACPP_DIR}/CMakeLists.txt
-    fi
-
-    if [ -n "$(ls -A ../patches/*.diff)" ]; then
-        # apply temporary patches until fix is upstream
-        for patch in ../patches/*.diff; do
-            for file in $(grep "^+++ " ${patch} | cut -f2 -d' ' | cut -f2- -d/); do
-                (cd ${LLAMACPP_DIR}; git checkout ${file})
-            done
-        done
-        for patch in ../patches/*.diff; do
-            (cd ${LLAMACPP_DIR} && git apply ${patch})
-        done
-    fi
+    # apply temporary patches until fix is upstream
+    for patch in ../patches/*.patch; do
+        git -c 'user.name=nobody' -c 'user.email=<>' -C ${LLAMACPP_DIR} am ${patch}
+    done
 }
 
 build() {
     cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
     cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
+    # remove unnecessary build artifacts
+    rm -f ${BUILD_DIR}/bin/ggml-common.h ${BUILD_DIR}/bin/ggml-metal.metal
 }
 
-compress() {
-    echo "Compressing payloads to reduce overall binary size..."
-    pids=""
-    rm -rf ${BUILD_DIR}/bin/*.gz
+dist() {
+    [ -z "${RUNNER}" ] && exit 1
+    mkdir -p ${RUNNER_BASE}/${RUNNER}/
     for f in ${BUILD_DIR}/bin/* ; do
-        gzip -n --best -f ${f} &
-        pids+=" $!"
+        cp ${f} ${RUNNER_BASE}/${RUNNER}/
     done
     # check for lib directory
     if [ -d ${BUILD_DIR}/lib ]; then
         for f in ${BUILD_DIR}/lib/* ; do
-            gzip -n --best -f ${f} &
-            pids+=" $!"
+            cp ${f} ${RUNNER_BASE}/${RUNNER}/
+        done
+    fi
+}
+
+# Compress from the build $BUILD_DIR into the $PAYLOAD_BASE/$RUNNER dir
+compress() {
+    [ -z "${RUNNER}" ] && exit 1
+    echo "Compressing payloads with ${GZIP} to reduce overall binary size..."
+    rm -rf "${PAYLOAD_BASE}/${RUNNER}/"
+    mkdir -p "${PAYLOAD_BASE}/${RUNNER}/"
+    for f in ${BUILD_DIR}/bin/* ; do
+        ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
+        compress_pids+=" $!"
+    done
+    # check for lib directory
+    if [ -d ${BUILD_DIR}/lib ]; then
+        for f in ${BUILD_DIR}/lib/* ; do
+            ${GZIP} -c --best ${f} > "${PAYLOAD_BASE}/${RUNNER}/$(basename ${f}).gz" &
+            compress_pids+=" $!"
         done
     fi
     echo
-    for pid in ${pids}; do
+}
+
+wait_for_compress() {
+    for pid in ${compress_pids}; do
         wait $pid
     done
     echo "Finished compression"
 }
 
+install() {
+    echo "Installing libraries to bin dir ${BUILD_DIR}/bin/"
+    for lib in $(find ${BUILD_DIR} -name \*.${LIB_EXT} | grep -v "${BUILD_DIR}/bin/" ); do
+        rm -f "${BUILD_DIR}/bin/$(basename ${lib})"
+        cp -af "${lib}" "${BUILD_DIR}/bin/"
+    done
+}
+
 # Keep the local tree clean after we're done with the build
 cleanup() {
     (cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)
diff --git a/llm/generate/gen_darwin.sh b/llm/generate/gen_darwin.sh
index 721a9ae80..c37366f3e 100755
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@@ -6,6 +6,7 @@
 
 set -ex
 set -o pipefail
+compress_pids=""
 echo "Starting darwin generate script"
 source $(dirname $0)/gen_common.sh
 init_vars
@@ -18,16 +19,16 @@ sign() {
     fi
 }
 
-COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_OPENMP=off"
+COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DGGML_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
 
 case "${GOARCH}" in
 "amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
 
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_BLAS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
@@ -37,8 +38,9 @@ case "${GOARCH}" in
         # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+        RUNNER=cpu
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
         echo "Building LCD CPU"
         build
         sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -49,8 +51,9 @@ case "${GOARCH}" in
         # Approximately 400% faster than LCD on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+        RUNNER=cpu_avx
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
         echo "Building AVX CPU"
         build
         sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -61,8 +64,9 @@ case "${GOARCH}" in
         # Approximately 10% faster than AVX on same CPU
         #
         init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
+        RUNNER=cpu_avx2
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
         echo "Building AVX2 CPU"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
         build
@@ -75,15 +79,16 @@ case "${GOARCH}" in
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_BLAS=off -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
     BUILD_DIR="../build/darwin/${ARCH}_static"
     echo "Building static library"
     build
 
     if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
         init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/darwin/${ARCH}/metal"
+        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
+        RUNNER="metal"
+        BUILD_DIR="../build/darwin/${GOARCH}/${RUNNER}"
         EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
         build
         sign ${BUILD_DIR}/bin/ollama_llama_server
@@ -98,4 +103,5 @@ case "${GOARCH}" in
 esac
 
 cleanup
+wait_for_compress
 echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
diff --git a/llm/generate/gen_linux.sh b/llm/generate/gen_linux.sh
index 0e98e1635..7e5f531e1 100755
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@@ -13,6 +13,7 @@
 
 set -ex
 set -o pipefail
+compress_pids=""
 
 # See https://llvm.org/docs/AMDGPUUsage.html#processors for reference
 amdGPUs() {
@@ -51,7 +52,7 @@ if [ -z "${CUDACXX}" ]; then
         export CUDACXX=$(command -v nvcc)
     fi
 fi
-COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off"
+COMMON_CMAKE_DEFS="-DCMAKE_SKIP_RPATH=on -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@@ -64,7 +65,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
     # Static build for linking into the Go binary
     init_vars
     CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}_static"
     echo "Building static library"
     build
@@ -77,32 +78,38 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
-        BUILD_DIR="../build/linux/${ARCH}/cpu"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        RUNNER="cpu"
+        BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
         echo "Building custom CPU"
         build
+        install
+        dist
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
         # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
+        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
 
-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_OPENMP=off"
+        COMMON_CPU_DEFS="-DBUILD_SHARED_LIBS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
         if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
             #
             # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
             #
             init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-            BUILD_DIR="../build/linux/${ARCH}/cpu"
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+            RUNNER=cpu
+            BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
             echo "Building LCD CPU"
             build
+            install
+            dist
             compress
         fi
 
@@ -116,10 +123,13 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 400% faster than LCD on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
+                RUNNER=cpu_avx
+                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
                 echo "Building AVX CPU"
                 build
+                install
+                dist
                 compress
             fi
 
@@ -129,10 +139,13 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                 # Approximately 10% faster than AVX on same CPU
                 #
                 init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
-                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
+                RUNNER=cpu_avx2
+                BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
                 echo "Building AVX2 CPU"
                 build
+                install
+                dist
                 compress
             fi
         fi
@@ -160,7 +173,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
     echo "CUDA libraries detected - building dynamic CUDA library"
     init_vars
     CUDA_MAJOR=$(ls "${CUDA_LIB_DIR}"/libcudart.so.* | head -1 | cut -f3 -d. || true)
-    if [ -n "${CUDA_MAJOR}" ]; then
+    if [ -n "${CUDA_MAJOR}" -a -z "${CUDA_VARIANT}" ]; then
         CUDA_VARIANT=_v${CUDA_MAJOR}
     fi
     if [ "${ARCH}" == "arm64" ]; then
@@ -170,37 +183,29 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         #
         # CUDA compute < 6.0 lacks proper FP16 support on ARM.
         # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
     fi
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
-    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
-    EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    export CUDAFLAGS="-t8"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS} -DGGML_STATIC=off"
+    RUNNER=cuda${CUDA_VARIANT}
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
+    export LLAMA_SERVER_LDFLAGS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
+    CUDA_DIST_DIR="${CUDA_DIST_DIR:-${DIST_BASE}/lib/ollama}"
     build
-
-    # Carry the CUDA libs as payloads to help reduce dependency burden on users
-    #
-    # TODO - in the future we may shift to packaging these separately and conditionally
-    #        downloading them in the install script.
-    DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
-    for lib in libcudart.so libcublas.so libcublasLt.so ; do
-        DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
-        if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
-            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
-            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
-        elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
-            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
-        else
-            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
-        fi
+    install
+    dist
+    echo "Installing CUDA dependencies in ${CUDA_DIST_DIR}"
+    mkdir -p "${CUDA_DIST_DIR}"
+    for lib in ${CUDA_LIB_DIR}/libcudart.so* ${CUDA_LIB_DIR}/libcublas.so* ${CUDA_LIB_DIR}/libcublasLt.so* ; do
+        cp -a "${lib}" "${CUDA_DIST_DIR}"
     done
     compress
 
@@ -244,23 +249,28 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
     init_vars
     source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
     CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
-    BUILD_DIR="../build/linux/${ARCH}/oneapi"
-    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
+    RUNNER=oneapi
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
+    ONEAPI_DIST_DIR="${DIST_BASE}/lib/ollama"
+    export LLAMA_SERVER_LDFLAGS="-fsycl -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
     DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
     build
 
     # copy oneAPI dependencies
+    mkdir -p "${ONEAPI_DIST_DIR}"
     for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e sycl -e mkl -e tbb); do
-        cp "${dep}" "${BUILD_DIR}/bin/"
+        cp -a "${dep}" "${ONEAPI_DIST_DIR}"
     done
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${BUILD_DIR}/bin/"
-    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${BUILD_DIR}/bin/"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libOpenCL.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libimf.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libintlc.so.5" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libirng.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libpi_level_zero.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libsvml.so" "${ONEAPI_DIST_DIR}"
+    cp "${ONEAPI_ROOT}/compiler/latest/lib/libur_loader.so.0" "${ONEAPI_DIST_DIR}"
+    install
+    dist
     compress
 fi
 
@@ -282,31 +292,35 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
         ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
     fi
     init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DGGML_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
     if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
         echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
         CMAKE_DEFS="${CMAKE_DEFS} ${OLLAMA_CUSTOM_ROCM_DEFS}"
         echo "Building custom ROCM GPU"
     fi
-    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    RUNNER=rocm${ROCM_VARIANT}
+    BUILD_DIR="../build/linux/${GOARCH}/${RUNNER}"
+    # ROCm dependencies are too large to fit into a unified bundle
+    ROCM_DIST_DIR="${DIST_BASE}/../linux-${GOARCH}-rocm/lib/ollama"
+    # TODO figure out how to disable runpath (rpath)
+    # export CMAKE_HIP_FLAGS="-fno-rtlib-add-rpath" # doesn't work
+    export LLAMA_SERVER_LDFLAGS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
 
-    # Record the ROCM dependencies
-    rm -f "${BUILD_DIR}/bin/deps.txt"
-    touch "${BUILD_DIR}/bin/deps.txt"
-    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
-        echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
+    # copy the ROCM dependencies
+    mkdir -p "${ROCM_DIST_DIR}"
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -v "${GOARCH}/rocm${ROCM_VARIANT}" | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf ); do
+        cp -a "${dep}"* "${ROCM_DIST_DIR}"
+        if [ $(readlink -f "${dep}") != "${dep}" ] ; then
+            cp $(readlink -f "${dep}") "${ROCM_DIST_DIR}"
+        fi
     done
-    # bomb out if for some reason we didn't get a few deps
-    if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
-        cat "${BUILD_DIR}/bin/deps.txt"
-        echo "ERROR: deps file short"
-        exit 1
-    fi
+    install
+    dist
     compress
 fi
 
 cleanup
-echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"
+wait_for_compress
+echo "go generate completed.  LLM runners: $(cd ${PAYLOAD_BASE}; echo *)"
diff --git a/llm/generate/gen_windows.ps1 b/llm/generate/gen_windows.ps1
index e217a0382..29ff5ff62 100644
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@@ -6,18 +6,9 @@ function amdGPUs {
     if ($env:AMDGPU_TARGETS) {
         return $env:AMDGPU_TARGETS
     }
-    # TODO - load from some common data file for linux + windows build consistency
+    # Current supported rocblas list from ROCm v6.1.2 on windows
+    # https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html#windows-supported-gpus
     $GPU_LIST = @(
-        "gfx900"
-        "gfx906:xnack-"
-        "gfx908:xnack-"
-        "gfx90a:xnack+"
-        "gfx90a:xnack-"
-        "gfx940"
-        "gfx941"
-        "gfx942"
-        "gfx1010"
-        "gfx1012"
         "gfx1030"
         "gfx1100"
         "gfx1101"
@@ -28,6 +19,19 @@ function amdGPUs {
 
 
 function init_vars {
+    write-host "Checking for cmake..."
+    get-command cmake
+    write-host "Checking for ninja..."
+    $d=(get-command -ea 'silentlycontinue' ninja).path
+    if ($null -eq $d) {
+        $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
+        $matches=(gci -path $MSVC_INSTALL -r -fi ninja.exe)
+        if ($matches.count -eq 0) {
+            throw "Unable to locate ninja"
+        }
+        $ninjaDir=($matches[0].FullName | split-path -parent)
+        $env:PATH="$env:PATH;$ninjaDir"
+    }
     if (!$script:SRC_DIR) {
         $script:SRC_DIR = $(resolve-path "..\..\")
     }
@@ -39,12 +43,12 @@ function init_vars {
     }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off",
-        "-DLLAMA_OPENMP=off"
+        "-DGGML_NATIVE=off",
+        "-DGGML_OPENMP=off"
         )
     $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
     $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
-    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\ollama_runners"
+    $script:DIST_BASE = "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\runners"
     md "$script:DIST_BASE" -ea 0 > $null
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
@@ -92,29 +96,9 @@ function git_module_setup {
 }
 
 function apply_patches {
-    # Wire up our CMakefile
-    if (!(Select-String -Path "${script:llamacppDir}/CMakeLists.txt" -Pattern 'ollama')) {
-        Add-Content -Path "${script:llamacppDir}/CMakeLists.txt" -Value 'add_subdirectory(../ext_server ext_server) # ollama'
-    }
-
     # Apply temporary patches until fix is upstream
-    $patches = Get-ChildItem "../patches/*.diff"
-    foreach ($patch in $patches) {
-        # Extract file paths from the patch file
-        $filePaths = Get-Content $patch.FullName | Where-Object { $_ -match '^\+\+\+ ' } | ForEach-Object {
-            $parts = $_ -split ' '
-            ($parts[1] -split '/', 2)[1]
-        }
-
-        # Checkout each file
-        foreach ($file in $filePaths) {
-            git -C "${script:llamacppDir}" checkout $file
-        }
-    }
-
-    # Apply each patch
-    foreach ($patch in $patches) {
-        git -C "${script:llamacppDir}" apply $patch.FullName
+    foreach ($patch in $(Get-ChildItem "../patches/*.patch")) {
+        git -c 'user.name=nobody' -c 'user.email=<>' -C "${script:llamacppDir}" am $patch.FullName
     }
 }
 
@@ -126,7 +110,7 @@ function build {
     if ($cmakeDefs -contains "-G") {
         $extra=@("-j8")
     } else {
-        $extra= @("--", "/p:CL_MPcount=8")
+        $extra= @("--", "/maxCpuCount:8")
     }
     write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ }) $extra"
     & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ }) $extra
@@ -174,7 +158,7 @@ function cleanup {
         }
 
         # Checkout each file
-        foreach ($file in $filePaths) {            
+        foreach ($file in $filePaths) {
             git -C "${script:llamacppDir}" checkout $file
         }
         git -C "${script:llamacppDir}" checkout CMakeLists.txt
@@ -182,21 +166,21 @@ function cleanup {
 }
 
 
-# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
 
 function build_static() {
     if ((-not "${env:OLLAMA_SKIP_STATIC_GENERATE}") -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "static"))) {
         # GCC build for direct linking into the Go binary
         init_vars
-        # cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
-        # as we need this to be compiled by gcc for golang to be able to link with itx
-        write-host "Checking for MinGW..."
-        # error action ensures we exit on failure
-        get-command gcc
-        get-command mingw32-make
+
+        # cmake will silently fallback to msvc compilers if gcc isn't in the path, so detect and fail fast
+        # as we need this to be compiled by gcc for golang to be able to link with it
+        write-host "Checking for gcc..."
+        get-command  gcc
+        get-command  mingw32-make
         $oldTargets = $script:cmakeTargets
         $script:cmakeTargets = @("llama", "ggml")
         $script:cmakeDefs = @(
@@ -204,13 +188,13 @@ function build_static() {
             "-DCMAKE_C_COMPILER=gcc.exe",
             "-DCMAKE_CXX_COMPILER=g++.exe",
             "-DBUILD_SHARED_LIBS=off",
-            "-DLLAMA_NATIVE=off",
-            "-DLLAMA_AVX=off",
-            "-DLLAMA_AVX2=off",
-            "-DLLAMA_AVX512=off",
-            "-DLLAMA_F16C=off",
-            "-DLLAMA_FMA=off",
-            "-DLLAMA_OPENMP=off")
+            "-DGGML_NATIVE=off",
+            "-DGGML_AVX=off",
+            "-DGGML_AVX2=off",
+            "-DGGML_AVX512=off",
+            "-DGGML_F16C=off",
+            "-DGGML_FMA=off",
+            "-DGGML_OPENMP=off")
         $script:buildDir="../build/windows/${script:ARCH}_static"
         write-host "Building static library"
         build
@@ -220,11 +204,10 @@ function build_static() {
     }
 }
 
-function build_cpu($gen_arch) {
+function build_cpu_x64 {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
-        # remaining llama.cpp builds use MSVC 
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu"
         $script:distDir="$script:DIST_BASE\cpu"
         write-host "Building LCD CPU"
@@ -236,10 +219,36 @@ function build_cpu($gen_arch) {
     }
 }
 
+function build_cpu_arm64 {
+    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
+        init_vars
+        write-host "Checking for clang..."
+        get-command clang
+        $env:CFLAGS="-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only"
+        $env:CXXFLAGS="$env:CFLAGS"
+        $env:LDFLAGS="-static-libstdc++"
+        $script:cmakeDefs = $script:commonCpuDefs + @(
+            "-DCMAKE_VERBOSE_MAKEFILE=on",
+            "-DCMAKE_C_COMPILER=clang.exe",
+            "-DCMAKE_CXX_COMPILER=clang++.exe",
+            "-DMSVC_RUNTIME_LIBRARY=MultiThreaded"
+        ) + $script:cmakeDefs
+        $script:buildDir="../build/windows/${script:ARCH}/cpu"
+        $script:distDir="$script:DIST_BASE\cpu"
+        write-host "Building LCD CPU"
+        build
+        sign
+        install
+    } else {
+        write-host "Skipping CPU generation step as requested"
+    }
+}
+
+
 function build_cpu_avx() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
         $script:distDir="$script:DIST_BASE\cpu_avx"
         write-host "Building AVX CPU"
@@ -254,7 +263,7 @@ function build_cpu_avx() {
 function build_cpu_avx2() {
     if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
         init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
         $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
         $script:distDir="$script:DIST_BASE\cpu_avx2"
         write-host "Building AVX2 CPU"
@@ -270,7 +279,7 @@ function build_cuda() {
     if ((-not "${env:OLLAMA_SKIP_CUDA_GENERATE}") -and ("${script:CUDA_LIB_DIR}")) {
         # Then build cuda as a dynamically loaded library
         $nvcc = "$script:CUDA_LIB_DIR\nvcc.exe"
-        $script:CUDA_VERSION=(get-item ($nvcc | split-path | split-path)).Basename
+        $script:CUDA_VERSION=((get-item ($nvcc | split-path | split-path)).Basename -Split "\.")[0]
         if ($null -ne $script:CUDA_VERSION) {
             $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
         }
@@ -279,12 +288,12 @@ function build_cuda() {
         $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
         $script:cmakeDefs += @(
             "-A", "x64",
-            "-DLLAMA_CUDA=ON",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
-            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
-            "-DCMAKE_CUDA_FLAGS=-t8",
-            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
+            "-DGGML_CUDA=ON",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
+            "-DCMAKE_CUDA_FLAGS=-t6",
+            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}",
+            "-DCMAKE_CUDA_COMPILER_TOOLKIT_ROOT=$env:CUDA_PATH"
             )
         if ($null -ne $env:OLLAMA_CUSTOM_CUDA_DEFS) {
             write-host "OLLAMA_CUSTOM_CUDA_DEFS=`"${env:OLLAMA_CUSTOM_CUDA_DEFS}`""
@@ -295,12 +304,11 @@ function build_cuda() {
         sign
         install
 
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\" -ea 0 > $null
-        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
-        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\cuda\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+        write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
     } else {
         write-host "Skipping CUDA generation step"
     }
@@ -319,7 +327,7 @@ function build_oneapi() {
     $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
     $script:cmakeDefs += @(
       "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
+      "-DGGML_SYCL=ON",
       "-DCMAKE_C_COMPILER=icx",
       "-DCMAKE_CXX_COMPILER=icx",
       "-DCMAKE_BUILD_TYPE=Release"
@@ -334,18 +342,17 @@ function build_oneapi() {
     sign
     install
 
-    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\" -ea 0 > $null
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
-    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\oneapi\"
+    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ea 0 > $null
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libirngmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\libmmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_level_zero.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_unified_runtime.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\pi_win_proxy_loader.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\svml_dispmd.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\compiler\latest\bin\sycl7.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_core.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_sycl_blas.4.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+    cp "${env:ONEAPI_ROOT}\mkl\latest\bin\mkl_tbb_thread.2.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
   } else {
     Write-Host "Skipping oneAPI generation step"
   }
@@ -362,13 +369,14 @@ function build_rocm() {
         $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
         $script:distDir="$script:DIST_BASE\rocm$script:ROCM_VARIANT"
         $script:cmakeDefs += @(
-            "-G", "Ninja", 
+            "-G", "Ninja",
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
+            "-DGGML_HIPBLAS=on",
+            "-DGGML_CUDA_NO_PEER_COPY=on",
             "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
             "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
             "-DAMDGPU_TARGETS=$(amdGPUs)",
             "-DGPU_TARGETS=$(amdGPUs)"
@@ -394,13 +402,11 @@ function build_rocm() {
         sign
         install
 
-        # Assumes v5.7, may need adjustments for v6
-        rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+        md "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\" -ea 0 > $null
+        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
+        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\"
         # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
+        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\rocblas\library\"
     } else {
         write-host "Skipping ROCm generation step"
     }
@@ -412,9 +418,9 @@ if ($($args.count) -eq 0) {
     apply_patches
     build_static
     if ($script:ARCH -eq "arm64") {
-        build_cpu("ARM64")
+        build_cpu_arm64
     } else { # amd64
-        build_cpu("x64")
+        build_cpu_x64
         build_cpu_avx
         build_cpu_avx2
         build_cuda
@@ -428,5 +434,5 @@ if ($($args.count) -eq 0) {
     for ( $i = 0; $i -lt $args.count; $i++ ) {
         write-host "performing $($args[$i])"
         & $($args[$i])
-    } 
+    }
 }
\ No newline at end of file
diff --git a/llm/ggla.go b/llm/ggla.go
index 34c4f6ca3..831f60712 100644
--- a/llm/ggla.go
+++ b/llm/ggla.go
@@ -36,6 +36,8 @@ type ggla struct {
 
 	kv      KV
 	tensors []*Tensor
+
+	tensorOffset uint64
 }
 
 func newGGLA(container *containerGGLA) *ggla {
@@ -50,7 +52,10 @@ func (llm *ggla) KV() KV {
 }
 
 func (llm *ggla) Tensors() Tensors {
-	return llm.tensors
+	return Tensors{
+		Items:  llm.tensors,
+		Offset: llm.tensorOffset,
+	}
 }
 
 func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
@@ -66,6 +71,13 @@ func (llm *ggla) decode(rs io.ReadSeeker) (retErr error) {
 	}
 	llm.kv["alpha"] = alpha
 
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	llm.tensorOffset = uint64(offset)
+
 	for {
 		var dims uint32
 		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
diff --git a/llm/ggml.go b/llm/ggml.go
index cfead450d..c4475a947 100644
--- a/llm/ggml.go
+++ b/llm/ggml.go
@@ -43,6 +43,14 @@ func (kv KV) Architecture() string {
 	return "unknown"
 }
 
+func (kv KV) Kind() string {
+	if s, ok := kv["general.type"].(string); ok {
+		return s
+	}
+
+	return "unknown"
+}
+
 func (kv KV) ParameterCount() uint64 {
 	return kv.u64("general.parameter_count")
 }
@@ -112,11 +120,14 @@ func (kv KV) ChatTemplate() string {
 	return s
 }
 
-type Tensors []*Tensor
+type Tensors struct {
+	Items  []*Tensor
+	Offset uint64
+}
 
 func (ts Tensors) Layers() map[string]Layer {
 	layers := make(map[string]Layer)
-	for _, t := range ts {
+	for _, t := range ts.Items {
 		parts := strings.Split(t.Name, ".")
 		if parts[0] == "blk" {
 			// join first and second part, e.g. blk.%d
@@ -154,6 +165,14 @@ type Tensor struct {
 	io.WriterTo `json:"-"`
 }
 
+func (t Tensor) block() (n int) {
+	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
+		return -1
+	}
+
+	return
+}
+
 func (t Tensor) blockSize() uint64 {
 	switch t.Kind {
 	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
@@ -341,11 +360,13 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 	switch llm.KV().Architecture() {
 	case "llama":
-		fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
+		fullOffload = max(
+			4*batch*(1+4*embedding+context*(1+heads)),
+			4*batch*(embedding+vocab),
+		)
 
 		partialOffload = 4 * batch * embedding
 		partialOffload += max(
-			// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
 			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
@@ -424,6 +445,32 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
 			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
 		)
+	case "chatglm":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(embedding+vocab) + embedding*vocab*105/128
+		if qkvBias, ok := layers["blk.0"]["attn_qkv.bias"]; ok {
+			fullOffload = max(
+				fullOffload,
+				4*batch*(2+
+					2*embedding+
+					context+
+					context*heads+
+					embeddingHeadsK*heads+
+					qkvBias.Shape[0]),
+			)
+
+			partialOffload = max(
+				partialOffload,
+				4*batch*(1+
+					2*embedding+
+					embeddingHeadsK*heads+
+					context+
+					context*heads)+
+					4*embeddingHeadsK*context+
+					4*context*embeddingHeadsK+
+					4*qkvBias.Shape[0],
+			)
+		}
 	}
 
 	return
diff --git a/llm/gguf.go b/llm/gguf.go
index 4d343a1bd..2e6bc542a 100644
--- a/llm/gguf.go
+++ b/llm/gguf.go
@@ -2,11 +2,16 @@ package llm
 
 import (
 	"bytes"
+	"cmp"
 	"encoding/binary"
 	"encoding/json"
 	"fmt"
 	"io"
+	"log/slog"
+	"slices"
 	"strings"
+
+	"golang.org/x/exp/maps"
 )
 
 type containerGGUF struct {
@@ -88,7 +93,8 @@ type gguf struct {
 	kv      KV
 	tensors []*Tensor
 
-	parameters uint64
+	parameters   uint64
+	tensorOffset uint64
 
 	scratch [16 << 10]byte
 }
@@ -100,16 +106,15 @@ func newGGUF(container *containerGGUF) *gguf {
 	}
 }
 
-func NewGGUFV3(bo binary.ByteOrder) *gguf {
-	return newGGUF(&containerGGUF{ByteOrder: bo, Version: 3})
-}
-
 func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
 func (llm *gguf) Tensors() Tensors {
-	return llm.tensors
+	return Tensors{
+		Items:  llm.tensors,
+		Offset: llm.tensorOffset,
+	}
 }
 
 func (llm *gguf) numTensor() uint64 {
@@ -199,7 +204,7 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 			return fmt.Errorf("failed to read tensor dimensions: %w", err)
 		}
 
-		shape := [4]uint64{1, 1, 1, 1}
+		shape := make([]uint64, dims)
 		for i := 0; uint32(i) < dims; i++ {
 			shape[i], err = readGGUF[uint64](llm, rs)
 			if err != nil {
@@ -236,13 +241,21 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 		alignment = 32
 	}
 
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	padding := ggufPadding(offset, int64(alignment))
+	llm.tensorOffset = uint64(offset + padding)
+
 	for _, tensor := range llm.tensors {
 		offset, err := rs.Seek(0, io.SeekCurrent)
 		if err != nil {
 			return fmt.Errorf("failed to get current offset: %w", err)
 		}
 
-		padding := llm.padding(offset, int64(alignment))
+		padding := ggufPadding(offset, int64(alignment))
 		if _, err := rs.Seek(padding, io.SeekCurrent); err != nil {
 			return fmt.Errorf("failed to seek to init padding: %w", err)
 		}
@@ -261,12 +274,12 @@ func readGGUF[T any](llm *gguf, r io.Reader) (T, error) {
 	return t, err
 }
 
-func writeGGUF[V any](llm *gguf, w io.Writer, t uint32, v V) error {
-	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+func writeGGUF[V any](w io.Writer, t uint32, v V) error {
+	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
 		return err
 	}
 
-	return binary.Write(w, llm.ByteOrder, v)
+	return binary.Write(w, binary.LittleEndian, v)
 }
 
 func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
@@ -330,12 +343,12 @@ func readGGUFString(llm *gguf, r io.Reader) (string, error) {
 	return string(buf), nil
 }
 
-func writeGGUFString(llm *gguf, w io.Writer, s string) error {
-	if err := binary.Write(w, llm.ByteOrder, ggufTypeString); err != nil {
+func writeGGUFString(w io.Writer, s string) error {
+	if err := binary.Write(w, binary.LittleEndian, ggufTypeString); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
 		return err
 	}
 
@@ -476,215 +489,71 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 	return a, nil
 }
 
-func writeGGUFArray[S ~[]E, E any](llm *gguf, w io.Writer, t uint32, s S) error {
-	if err := binary.Write(w, llm.ByteOrder, ggufTypeArray); err != nil {
+// writeGGUFArray writes a slice s of type E to the write with a gguf type of t
+func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
+	if err := binary.Write(w, binary.LittleEndian, ggufTypeArray); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, t); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, t); err != nil {
 		return err
 	}
 
-	if err := binary.Write(w, llm.ByteOrder, uint64(len(s))); err != nil {
+	if err := binary.Write(w, binary.LittleEndian, uint64(len(s))); err != nil {
 		return err
 	}
 
-	for _, e := range s {
-		if err := binary.Write(w, llm.ByteOrder, e); err != nil {
-			return err
-		}
-	}
-
-	return nil
+	return binary.Write(w, binary.LittleEndian, s)
 }
 
-var ggufKVOrder = map[string][]string{
-	"llama": {
-		"general.architecture",
-		"general.name",
-		"llama.vocab_size",
-		"llama.context_length",
-		"llama.embedding_length",
-		"llama.block_count",
-		"llama.feed_forward_length",
-		"llama.attention.head_count",
-		"llama.attention.head_count_kv",
-		"llama.attention.layer_norm_rms_epsilon",
-		"llama.rope.freq_base",
-		"llama.rope.dimension_count",
-		"llama.expert_count",
-		"llama.expert_used_count",
-		"gemma.context_length",
-		"gemma.embedding_length",
-		"gemma.block_count",
-		"gemma.feed_forward_length",
-		"gemma.attention.head_count",
-		"gemma.attention.head_count_kv",
-		"gemma.attention.layer_norm_rms_epsilon",
-		"gemma.attention.key_length",
-		"gemma.attention.value_length",
-		"general.file_type",
-		"tokenizer.ggml.pre",
-		"tokenizer.ggml.model",
-		"tokenizer.ggml.tokens",
-		"tokenizer.ggml.scores",
-		"tokenizer.ggml.merges",
-		"tokenizer.ggml.token_type",
-		"tokenizer.ggml.bos_token_id",
-		"tokenizer.ggml.eos_token_id",
-		"tokenizer.ggml.unknown_token_id",
-		"tokenizer.ggml.padding_token_id",
-		"tokenizer.ggml.add_bos_token",
-		"tokenizer.ggml.add_eos_token",
-		"tokenizer.chat_template",
-	},
-}
-
-func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
-	switch llm.Version {
-	case 3:
-		llm.V3.NumTensor = uint64(len(tensors))
-		llm.V3.NumKV = uint64(len(kv))
-	default:
-		return fmt.Errorf("not implemented: ggufv%d", llm.Version)
-	}
-
-	if err := binary.Write(ws, llm.ByteOrder, []byte("GGUF")); err != nil {
+func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.Version); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint32(3)); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.numTensor()); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(ts))); err != nil {
 		return err
 	}
 
-	if err := binary.Write(ws, llm.ByteOrder, llm.numKV()); err != nil {
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(kv))); err != nil {
 		return err
 	}
 
-	kvCheck := make(map[string]bool)
-	for k := range kv {
-		kvCheck[k] = false
-	}
+	keys := maps.Keys(kv)
+	slices.Sort(keys)
 
-	for _, k := range ggufKVOrder["llama"] {
-		v, ok := kv[k]
-		if !ok {
-			continue
-		}
-		kvCheck[k] = true
-
-		if err := binary.Write(ws, llm.ByteOrder, uint64(len(k))); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, []byte(k)); err != nil {
-			return err
-		}
-
-		var err error
-		switch v := v.(type) {
-		case uint32:
-			err = writeGGUF(llm, ws, ggufTypeUint32, v)
-		case float32:
-			err = writeGGUF(llm, ws, ggufTypeFloat32, v)
-		case bool:
-			err = writeGGUF(llm, ws, ggufTypeBool, v)
-		case string:
-			err = writeGGUFString(llm, ws, v)
-		case []int32:
-			err = writeGGUFArray(llm, ws, ggufTypeInt32, v)
-		case []uint32:
-			err = writeGGUFArray(llm, ws, ggufTypeUint32, v)
-		case []float32:
-			err = writeGGUFArray(llm, ws, ggufTypeFloat32, v)
-		case []string:
-			if err := binary.Write(ws, llm.ByteOrder, ggufTypeArray); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, llm.ByteOrder, ggufTypeString); err != nil {
-				return err
-			}
-
-			if err := binary.Write(ws, llm.ByteOrder, uint64(len(v))); err != nil {
-				return err
-			}
-
-			for _, e := range v {
-				if err := binary.Write(ws, llm.ByteOrder, uint64(len(e))); err != nil {
-					return err
-				}
-
-				if err := binary.Write(ws, llm.ByteOrder, []byte(e)); err != nil {
-					return err
-				}
-			}
-		default:
-			return fmt.Errorf("improper type for '%s'", k)
-		}
-		if err != nil {
+	for _, key := range keys {
+		if err := ggufWriteKV(ws, key, kv[key]); err != nil {
 			return err
 		}
 	}
 
-	for k, v := range kvCheck {
-		if !v {
-			return fmt.Errorf("Didn't know how to write kv %s", k)
+	slices.SortStableFunc(ts, func(a, b Tensor) int {
+		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+			return 1
+		} else if i > 0 && j < 0 {
+			return -1
+		} else {
+			return cmp.Compare(i, j)
 		}
-	}
+	})
 
-	for _, tensor := range tensors {
-		if err := binary.Write(ws, llm.ByteOrder, uint64(len(tensor.Name))); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, []byte(tensor.Name)); err != nil {
-			return err
-		}
-
-		var dims int
-		for cnt := range len(tensor.Shape) {
-			if tensor.Shape[cnt] > 0 {
-				dims++
-			}
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, uint32(dims)); err != nil {
-			return err
-		}
-
-		for i := range dims {
-			if err := binary.Write(ws, llm.ByteOrder, tensor.Shape[dims-1-i]); err != nil {
-				return err
-			}
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, tensor.Kind); err != nil {
-			return err
-		}
-
-		if err := binary.Write(ws, llm.ByteOrder, tensor.Offset); err != nil {
+	var s uint64
+	for _, t := range ts {
+		t.Offset = s
+		if err := ggufWriteTensorInfo(ws, t); err != nil {
 			return err
 		}
+		s += t.Size()
 	}
 
 	var alignment int64 = 32
-	for _, tensor := range tensors {
-		offset, err := ws.Seek(0, io.SeekCurrent)
-		if err != nil {
-			return err
-		}
-
-		padding := llm.padding(offset, alignment)
-		if err := binary.Write(ws, llm.ByteOrder, bytes.Repeat([]byte{0}, int(padding))); err != nil {
-			return err
-		}
-
-		if _, err := tensor.WriteTo(ws); err != nil {
+	for _, t := range ts {
+		if err := ggufWriteTensor(ws, t, alignment); err != nil {
 			return err
 		}
 	}
@@ -692,6 +561,102 @@ func (llm *gguf) Encode(ws io.WriteSeeker, kv KV, tensors []Tensor) error {
 	return nil
 }
 
-func (gguf) padding(offset, align int64) int64 {
+func ggufWriteKV(ws io.WriteSeeker, k string, v any) error {
+	slog.Debug(k, "type", fmt.Sprintf("%T", v))
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(k))); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, []byte(k)); err != nil {
+		return err
+	}
+
+	var err error
+	switch v := v.(type) {
+	case uint32:
+		err = writeGGUF(ws, ggufTypeUint32, v)
+	case float32:
+		err = writeGGUF(ws, ggufTypeFloat32, v)
+	case bool:
+		err = writeGGUF(ws, ggufTypeBool, v)
+	case string:
+		err = writeGGUFString(ws, v)
+	case []int32:
+		err = writeGGUFArray(ws, ggufTypeInt32, v)
+	case []uint32:
+		err = writeGGUFArray(ws, ggufTypeUint32, v)
+	case []float32:
+		err = writeGGUFArray(ws, ggufTypeFloat32, v)
+	case []string:
+		if err := binary.Write(ws, binary.LittleEndian, ggufTypeArray); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, binary.LittleEndian, ggufTypeString); err != nil {
+			return err
+		}
+
+		if err := binary.Write(ws, binary.LittleEndian, uint64(len(v))); err != nil {
+			return err
+		}
+
+		for _, e := range v {
+			if err := binary.Write(ws, binary.LittleEndian, uint64(len(e))); err != nil {
+				return err
+			}
+
+			if err := binary.Write(ws, binary.LittleEndian, []byte(e)); err != nil {
+				return err
+			}
+		}
+	default:
+		return fmt.Errorf("improper type for '%s'", k)
+	}
+
+	return err
+}
+
+func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
+	slog.Debug(t.Name, "kind", t.Kind, "shape", t.Shape, "offset", t.Offset)
+	if err := binary.Write(ws, binary.LittleEndian, uint64(len(t.Name))); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, []byte(t.Name)); err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, uint32(len(t.Shape))); err != nil {
+		return err
+	}
+
+	for i := range len(t.Shape) {
+		if err := binary.Write(ws, binary.LittleEndian, t.Shape[len(t.Shape)-i-1]); err != nil {
+			return err
+		}
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, t.Kind); err != nil {
+		return err
+	}
+
+	return binary.Write(ws, binary.LittleEndian, t.Offset)
+}
+
+func ggufWriteTensor(ws io.WriteSeeker, t Tensor, alignment int64) error {
+	offset, err := ws.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return err
+	}
+
+	if err := binary.Write(ws, binary.LittleEndian, bytes.Repeat([]byte{0}, int(ggufPadding(offset, alignment)))); err != nil {
+		return err
+	}
+
+	_, err = t.WriteTo(ws)
+	return err
+}
+
+func ggufPadding(offset, align int64) int64 {
 	return (align - offset%align) % align
 }
diff --git a/llm/llama.cpp b/llm/llama.cpp
index 7c26775ad..8962422b1 160000
--- a/llm/llama.cpp
+++ b/llm/llama.cpp
@@ -1 +1 @@
-Subproject commit 7c26775adb579e92b59c82e8084c07a1d0f75e9c
+Subproject commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177
diff --git a/llm/llm.go b/llm/llm.go
index 2a0c4b91a..6c6958894 100644
--- a/llm/llm.go
+++ b/llm/llm.go
@@ -1,17 +1,19 @@
 package llm
 
-// #cgo CFLAGS: -Illama.cpp
-// #cgo darwin,arm64 LDFLAGS: ${SRCDIR}/build/darwin/arm64_static/libllama.a -lstdc++
-// #cgo darwin,amd64 LDFLAGS: ${SRCDIR}/build/darwin/x86_64_static/libllama.a -lstdc++
-// #cgo windows,amd64 LDFLAGS: ${SRCDIR}/build/windows/amd64_static/libllama.a -static -lstdc++
-// #cgo windows,arm64 LDFLAGS: ${SRCDIR}/build/windows/arm64_static/libllama.a -static -lstdc++
-// #cgo linux,amd64 LDFLAGS: ${SRCDIR}/build/linux/x86_64_static/libllama.a -lstdc++
-// #cgo linux,arm64 LDFLAGS: ${SRCDIR}/build/linux/arm64_static/libllama.a -lstdc++
+// #cgo CFLAGS: -Illama.cpp -Illama.cpp/include -Illama.cpp/ggml/include
+// #cgo LDFLAGS: -lllama -lggml -lstdc++ -lpthread
+// #cgo darwin,arm64 LDFLAGS: -L${SRCDIR}/build/darwin/arm64_static -L${SRCDIR}/build/darwin/arm64_static/src -L${SRCDIR}/build/darwin/arm64_static/ggml/src -framework Accelerate -framework Metal
+// #cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/darwin/x86_64_static -L${SRCDIR}/build/darwin/x86_64_static/src -L${SRCDIR}/build/darwin/x86_64_static/ggml/src
+// #cgo windows,amd64 LDFLAGS: -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
+// #cgo windows,arm64 LDFLAGS: -lllama -lggml -static-libstdc++ -static-libgcc -static -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
+// #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/linux/x86_64_static -L${SRCDIR}/build/linux/x86_64_static/src -L${SRCDIR}/build/linux/x86_64_static/ggml/src
+// #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/linux/arm64_static -L${SRCDIR}/build/linux/arm64_static/src -L${SRCDIR}/build/linux/arm64_static/ggml/src
 // #include <stdlib.h>
 // #include "llama.h"
 import "C"
+
 import (
-	"fmt"
+	"errors"
 	"unsafe"
 )
 
@@ -32,7 +34,7 @@ func Quantize(infile, outfile string, ftype fileType) error {
 	params.ftype = ftype.Value()
 
 	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
-		return fmt.Errorf("llama_model_quantize: %d", rc)
+		return errors.New("failed to quantize model. This model architecture may not be supported, or you may need to upgrade Ollama to the latest version")
 	}
 
 	return nil
diff --git a/llm/llm_darwin.go b/llm/llm_darwin.go
new file mode 100644
index 000000000..60837ed06
--- /dev/null
+++ b/llm/llm_darwin.go
@@ -0,0 +1,7 @@
+package llm
+
+import (
+	"syscall"
+)
+
+var LlamaServerSysProcAttr = &syscall.SysProcAttr{}
diff --git a/llm/llm_darwin_amd64.go b/llm/llm_darwin_amd64.go
deleted file mode 100644
index 3093e1ad2..000000000
--- a/llm/llm_darwin_amd64.go
+++ /dev/null
@@ -1,8 +0,0 @@
-package llm
-
-import (
-	"embed"
-)
-
-//go:embed build/darwin/x86_64/*/bin/*
-var libEmbed embed.FS
diff --git a/llm/llm_darwin_arm64.go b/llm/llm_darwin_arm64.go
deleted file mode 100644
index 928f0b824..000000000
--- a/llm/llm_darwin_arm64.go
+++ /dev/null
@@ -1,8 +0,0 @@
-package llm
-
-import (
-	"embed"
-)
-
-//go:embed build/darwin/arm64/*/bin/*
-var libEmbed embed.FS
diff --git a/llm/llm_linux.go b/llm/llm_linux.go
index c2c5c4cb9..60837ed06 100644
--- a/llm/llm_linux.go
+++ b/llm/llm_linux.go
@@ -1,6 +1,7 @@
 package llm
 
-import "embed"
+import (
+	"syscall"
+)
 
-//go:embed build/linux/*/*/bin/*
-var libEmbed embed.FS
+var LlamaServerSysProcAttr = &syscall.SysProcAttr{}
diff --git a/llm/llm_windows.go b/llm/llm_windows.go
index e44f4b951..915355a25 100644
--- a/llm/llm_windows.go
+++ b/llm/llm_windows.go
@@ -1,6 +1,22 @@
 package llm
 
-import "embed"
+import (
+	"syscall"
+)
 
-// unused on windows
-var libEmbed embed.FS
+const (
+	CREATE_DEFAULT_ERROR_MODE   = 0x04000000
+	ABOVE_NORMAL_PRIORITY_CLASS = 0x00008000
+)
+
+var LlamaServerSysProcAttr = &syscall.SysProcAttr{
+	// Wire up the default error handling logic If for some reason a DLL is
+	// missing in the path this will pop up a GUI Dialog explaining the fault so
+	// the user can either fix their PATH, or report a bug. Without this
+	// setting, the process exits immediately with a generic exit status but no
+	// way to (easily) figure out what the actual missing DLL was.
+	//
+	// Setting Above Normal priority class ensures when running as a "background service"
+	// with "programs" given best priority, we aren't starved of cpu cycles
+	CreationFlags: CREATE_DEFAULT_ERROR_MODE | ABOVE_NORMAL_PRIORITY_CLASS,
+}
diff --git a/llm/memory.go b/llm/memory.go
index 19b12cbfc..99db76299 100644
--- a/llm/memory.go
+++ b/llm/memory.go
@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 )
@@ -94,6 +95,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	// Overflow that didn't fit into the GPU
 	var overflow uint64
 
+	overhead := envconfig.GpuOverhead()
 	availableList := make([]string, len(gpus))
 	for i, gpu := range gpus {
 		availableList[i] = format.HumanBytes2(gpu.FreeMemory)
@@ -164,8 +166,22 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 			gzo = gpuZeroOverhead
 		}
 		// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
-		if gpus[i].FreeMemory < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
-			slog.Debug("gpu has too little memory to allocate any layers", "gpu", gpus[i])
+		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
+			slog.Debug("gpu has too little memory to allocate any layers",
+				"id", gpus[i].ID,
+				"library", gpus[i].Library,
+				"variant", gpus[i].Variant,
+				"compute", gpus[i].Compute,
+				"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
+				"name", gpus[i].Name,
+				"total", format.HumanBytes2(gpus[i].TotalMemory),
+				"available", format.HumanBytes2(gpus[i].FreeMemory),
+				"minimum_memory", gpus[i].MinimumMemory,
+				"layer_size", format.HumanBytes2(layerSize),
+				"gpu_zer_overhead", format.HumanBytes2(gzo),
+				"partial_offload", format.HumanBytes2(graphPartialOffload),
+				"full_offload", format.HumanBytes2(graphFullOffload),
+			)
 			continue
 		}
 		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
@@ -196,7 +212,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		for j := len(gpusWithSpace); j > 0; j-- {
 			g := gpusWithSpace[i%j]
 			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
-			if g.g.FreeMemory > used+layerSize {
+			if (g.g.FreeMemory - overhead) > used+layerSize {
 				gpuAllocations[g.i] += layerSize
 				layerCounts[g.i]++
 				layerCount++
@@ -219,7 +235,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		for j := len(gpusWithSpace); j > 0; j-- {
 			g := gpusWithSpace[layerCount%j]
 			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
-			if g.g.FreeMemory > used+memoryLayerOutput {
+			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
 				gpuAllocations[g.i] += memoryLayerOutput
 				layerCounts[g.i]++
 				layerCount++
@@ -306,6 +322,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 }
 
 func (m MemoryEstimate) log() {
+	overhead := envconfig.GpuOverhead()
 	slog.Info(
 		"offload to "+m.inferenceLibrary,
 		slog.Group(
@@ -323,6 +340,7 @@ func (m MemoryEstimate) log() {
 			"memory",
 			// memory available by GPU for offloading
 			"available", m.availableList,
+			"gpu_overhead", format.HumanBytes2(overhead),
 			slog.Group(
 				"required",
 				// memory required for full offloading
diff --git a/llm/memory_test.go b/llm/memory_test.go
index f972f9275..ffb14286b 100644
--- a/llm/memory_test.go
+++ b/llm/memory_test.go
@@ -2,25 +2,24 @@ package llm
 
 import (
 	"bytes"
-	"encoding/binary"
 	"fmt"
 	"os"
 	"testing"
 
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
-	"github.com/ollama/ollama/gpu"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/gpu"
 )
 
 func TestEstimateGPULayers(t *testing.T) {
-	envconfig.Debug = true
+	t.Setenv("OLLAMA_DEBUG", "1")
+
 	modelName := "dummy"
 	f, err := os.CreateTemp(t.TempDir(), modelName)
 	require.NoError(t, err)
 	defer f.Close()
-	gguf := NewGGUFV3(binary.LittleEndian)
 	inputLayerCount := 5
 
 	tensors := []Tensor{
@@ -32,9 +31,8 @@ func TestEstimateGPULayers(t *testing.T) {
 		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 	}
 	assert.Len(t, tensors, inputLayerCount+1)
-	err = gguf.Encode(f, KV{
+	err = WriteGGUF(f, KV{
 		"general.architecture":          "llama",
-		"general.name":                  "name",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
 		"llama.block_count":             uint32(inputLayerCount),
diff --git a/llm/patches/0000-cmakelist.patch b/llm/patches/0000-cmakelist.patch
new file mode 100644
index 000000000..54e9b602d
--- /dev/null
+++ b/llm/patches/0000-cmakelist.patch
@@ -0,0 +1,22 @@
+From 8b8d83ffca775840acc5dc700f3b3703e9f5cfe4 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Fri, 23 Aug 2024 11:27:48 -0700
+Subject: [PATCH] patch cmakelist
+
+---
+ CMakeLists.txt | 2 ++
+ 1 file changed, 2 insertions(+)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index a3132063..6a2a9912 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -199,3 +199,5 @@ if (LLAMA_BUILD_EXAMPLES)
+     add_subdirectory(examples)
+     add_subdirectory(pocs)
+ endif()
++
++add_subdirectory(../ext_server ext_server) # ollama
+-- 
+2.45.2
+
diff --git a/llm/patches/01-load-progress.diff b/llm/patches/0001-load-progress.patch
similarity index 62%
rename from llm/patches/01-load-progress.diff
rename to llm/patches/0001-load-progress.patch
index be5286091..0ddabc801 100644
--- a/llm/patches/01-load-progress.diff
+++ b/llm/patches/0001-load-progress.patch
@@ -1,8 +1,18 @@
+From 2cfaa0a04faa9c87ba8f1ac8527eb953e69c6cde Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:10 -0700
+Subject: [PATCH] 01-load-progress.diff
+
+---
+ common/common.cpp | 2 ++
+ common/common.h   | 7 +++++++
+ 2 files changed, 9 insertions(+)
+
 diff --git a/common/common.cpp b/common/common.cpp
-index 73ff0e85..6adb1a92 100644
+index 9fa18472..48ff41e9 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -2447,6 +2447,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
+@@ -2573,6 +2573,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
      mparams.use_mmap        = params.use_mmap;
      mparams.use_mlock       = params.use_mlock;
      mparams.check_tensors   = params.check_tensors;
@@ -12,10 +22,10 @@ index 73ff0e85..6adb1a92 100644
          mparams.kv_overrides = NULL;
      } else {
 diff --git a/common/common.h b/common/common.h
-index 58ed72f4..0bb2605e 100644
+index cb5e7f6d..d8f043f7 100644
 --- a/common/common.h
 +++ b/common/common.h
-@@ -180,6 +180,13 @@ struct gpt_params {
+@@ -204,6 +204,13 @@ struct gpt_params {
      std::string mmproj = "";        // path to multimodal projector
      std::vector<std::string> image; // path to image file(s)
  
@@ -26,6 +36,9 @@ index 58ed72f4..0bb2605e 100644
 +    // context pointer passed to the progress callback
 +    void * progress_callback_user_data;
 +
-     // server params
-     int32_t port           = 8080;         // server listens on this network port
-     int32_t timeout_read   = 600;          // http read timeout in seconds
+     // embedding
+     bool embedding         = false; // get only sentence embedding
+     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+-- 
+2.46.0
+
diff --git a/llm/patches/02-clip-log.diff b/llm/patches/0002-clip-log.patch
similarity index 60%
rename from llm/patches/02-clip-log.diff
rename to llm/patches/0002-clip-log.patch
index 34a018e8b..8df0da17b 100644
--- a/llm/patches/02-clip-log.diff
+++ b/llm/patches/0002-clip-log.patch
@@ -1,5 +1,14 @@
+From ba4bba80a744f76ac67b8234451c259a3c5da83b Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:11 -0700
+Subject: [PATCH] 02-clip-log.diff
+
+---
+ examples/llava/clip.cpp | 1 +
+ 1 file changed, 1 insertion(+)
+
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index e431c7f7..f077e688 100644
+index 9b890571..cb51793d 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
 @@ -3,6 +3,7 @@
@@ -10,3 +19,6 @@ index e431c7f7..f077e688 100644
  #include "log.h"
  #include "ggml.h"
  #include "ggml-alloc.h"
+-- 
+2.46.0
+
diff --git a/llm/patches/03-load_exception.diff b/llm/patches/0003-load_exception.patch
similarity index 75%
rename from llm/patches/03-load_exception.diff
rename to llm/patches/0003-load_exception.patch
index eb245c2a9..3d858ebb8 100644
--- a/llm/patches/03-load_exception.diff
+++ b/llm/patches/0003-load_exception.patch
@@ -1,17 +1,17 @@
-From 544a2d2e646d39e878d87dfbb3398a356bc560ab Mon Sep 17 00:00:00 2001
+From e43bfd3f607a6dfcaba2d490d35f412a52e55e30 Mon Sep 17 00:00:00 2001
 From: Michael Yang <mxyng@pm.me>
-Date: Thu, 23 May 2024 11:18:45 -0700
-Subject: [PATCH] throw exception on load errors
+Date: Mon, 16 Sep 2024 15:53:12 -0700
+Subject: [PATCH] 03-load_exception.diff
 
 ---
- llama.cpp | 25 ++++++++++++++++---------
+ src/llama.cpp | 25 ++++++++++++++++---------
  1 file changed, 16 insertions(+), 9 deletions(-)
 
-diff --git a/llama.cpp b/llama.cpp
-index 15c66077..8ba90b6a 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -6346,7 +6346,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 88355971..926bb71a 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -8635,7 +8635,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
          }
      } catch (const std::exception & err) {
          LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
@@ -20,7 +20,7 @@ index 15c66077..8ba90b6a 100644
      }
  
      return 0;
-@@ -15600,16 +15600,23 @@ struct llama_model * llama_load_model_from_file(
+@@ -18022,16 +18022,23 @@ struct llama_model * llama_load_model_from_file(
          }
          model->rpc_servers.push_back(servers);
      }
@@ -53,5 +53,5 @@ index 15c66077..8ba90b6a 100644
  
      return model;
 -- 
-2.45.1
+2.46.0
 
diff --git a/llm/patches/04-metal.diff b/llm/patches/0004-metal.patch
similarity index 83%
rename from llm/patches/04-metal.diff
rename to llm/patches/0004-metal.patch
index f8fa7db76..4cfa407e4 100644
--- a/llm/patches/04-metal.diff
+++ b/llm/patches/0004-metal.patch
@@ -1,8 +1,17 @@
-diff --git a/ggml-metal.m b/ggml-metal.m
-index 0207b787..b5e9884b 100644
---- a/ggml-metal.m
-+++ b/ggml-metal.m
-@@ -1396,27 +1396,23 @@ static enum ggml_status ggml_metal_graph_compute(
+From 29411d9a9d2b6a0af6425ffe88498f17f71f7d5d Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:12 -0700
+Subject: [PATCH] 04-metal.diff
+
+---
+ ggml/src/ggml-metal.m | 30 +++++++++++++-----------------
+ 1 file changed, 13 insertions(+), 17 deletions(-)
+
+diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
+index 91b5e61b..9cfa72ac 100644
+--- a/ggml/src/ggml-metal.m
++++ b/ggml/src/ggml-metal.m
+@@ -1734,27 +1734,23 @@ static enum ggml_status ggml_metal_graph_compute(
                          // to the matrix-vector kernel
                          int ne11_mm_min = 1;
  
@@ -43,3 +52,6 @@ index 0207b787..b5e9884b 100644
  
                          // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                          // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+-- 
+2.46.0
+
diff --git a/llm/patches/05-default-pretokenizer.diff b/llm/patches/0005-default-pretokenizer.patch
similarity index 65%
rename from llm/patches/05-default-pretokenizer.diff
rename to llm/patches/0005-default-pretokenizer.patch
index 2a2e7306e..6ad0ee97b 100644
--- a/llm/patches/05-default-pretokenizer.diff
+++ b/llm/patches/0005-default-pretokenizer.patch
@@ -1,11 +1,20 @@
-diff --git a/llama.cpp b/llama.cpp
-index 61948751..4b72a293 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -4824,16 +4824,7 @@ static void llm_load_vocab(
- 
-         // for now, only BPE models have pre-tokenizers
+From b298ac8614d1e38da28f760eb1d2ae8af0fbbe62 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:13 -0700
+Subject: [PATCH] 05-default-pretokenizer.diff
+
+---
+ src/llama.cpp | 14 +++-----------
+ 1 file changed, 3 insertions(+), 11 deletions(-)
+
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 926bb71a..d1e959fc 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -6083,16 +6083,7 @@ static void llm_load_vocab(
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
+             vocab.tokenizer_add_space_prefix = false;
+             vocab.tokenizer_clean_spaces = true;
 -            if (tokenizer_pre.empty()) {
 -                LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__);
 -                LLAMA_LOG_WARN("%s:                                             \n", __func__);
@@ -20,13 +29,16 @@ index 61948751..4b72a293 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -4888,7 +4879,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "poro-chat") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
+@@ -6188,7 +6179,8 @@ static void llm_load_vocab(
+                 tokenizer_pre == "exaone") {
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
 +                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              }
-         } else {
+         } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
              vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+-- 
+2.46.0
+
diff --git a/llm/patches/0006-embeddings.patch b/llm/patches/0006-embeddings.patch
new file mode 100644
index 000000000..8f89ffeb9
--- /dev/null
+++ b/llm/patches/0006-embeddings.patch
@@ -0,0 +1,55 @@
+From c9a6ca9fc039233dee746a4da9705762cd9e515d Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:14 -0700
+Subject: [PATCH] 06-embeddings.diff
+
+---
+ src/llama.cpp | 17 ++++++++++-------
+ 1 file changed, 10 insertions(+), 7 deletions(-)
+
+diff --git a/src/llama.cpp b/src/llama.cpp
+index d1e959fc..f79bd782 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -15898,7 +15898,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+     const auto n_embd  = hparams.n_embd;
+ 
+     // TODO: use a per-batch flag for logits presence instead
+-    const bool has_logits = !cparams.embeddings;
++    const bool has_logits =  cparams.causal_attn;
+     const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
+ 
+     const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
+@@ -16167,20 +16167,23 @@ static int llama_decode_internal(
+             // no output
+             res  = nullptr;
+             embd = nullptr;
+-        } else if (cparams.embeddings) {
+-            res  = nullptr; // do not extract logits for embedding case
+-            embd = nullptr;
++        }
++
++        if (cparams.embeddings) {
+             for (int i = gf->n_nodes - 1; i >= 0; --i) {
+-                if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
+-                    embd = gf->nodes[i];
++                embd = gf->nodes[i];
++                if (strcmp(embd->name, "result_embd_pooled") == 0) {
+                     break;
+                 }
+             }
+-            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
+         } else {
+             embd = nullptr; // do not extract embeddings when not needed
+             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
+         }
++
++        if (!cparams.causal_attn) {
++            res = nullptr; // do not extract logits when not needed
++        }
+         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+ 
+         ggml_backend_sched_alloc_graph(lctx.sched, gf);
+-- 
+2.46.0
+
diff --git a/llm/patches/0007-clip-unicode.patch b/llm/patches/0007-clip-unicode.patch
new file mode 100644
index 000000000..72c061cb4
--- /dev/null
+++ b/llm/patches/0007-clip-unicode.patch
@@ -0,0 +1,54 @@
+From ae2b188a679c83ce105aa1e823499441dfab3c57 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:15 -0700
+Subject: [PATCH] 07-clip-unicode.diff
+
+---
+ examples/llava/clip.cpp | 23 +++++++++++++++++++++++
+ 1 file changed, 23 insertions(+)
+
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index cb51793d..8716472b 100644
+--- a/examples/llava/clip.cpp
++++ b/examples/llava/clip.cpp
+@@ -41,6 +41,14 @@
+ #include <cinttypes>
+ #include <limits>
+ 
++#if defined(_WIN32)
++#define WIN32_LEAN_AND_MEAN
++#ifndef NOMINMAX
++    #define NOMINMAX
++#endif
++#include <windows.h>
++#endif
++
+ //#define CLIP_DEBUG_FUNCTIONS
+ 
+ // RGB uint8 image
+@@ -1223,7 +1231,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+             return nullptr;
+         }
+ 
++#ifdef _WIN32
++        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
++        if (!wlen) {
++            return NULL;
++        }
++        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
++        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
++        if (!wlen) {
++            free(wbuf);
++            return NULL;
++        }
++        auto fin = std::ifstream(wbuf, std::ios::binary);
++        free(wbuf);
++#else
+         auto fin = std::ifstream(fname, std::ios::binary);
++#endif
+         if (!fin) {
+             LOG_TEE("cannot open model file for loading tensors\n");
+             clip_free(new_clip);
+-- 
+2.46.0
+
diff --git a/llm/patches/0008-solar-pro.patch b/llm/patches/0008-solar-pro.patch
new file mode 100644
index 000000000..54f184577
--- /dev/null
+++ b/llm/patches/0008-solar-pro.patch
@@ -0,0 +1,402 @@
+From 8313ce5f43f11f3d84f352f97f3802792e90e18c Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Mon, 16 Sep 2024 15:53:16 -0700
+Subject: [PATCH] add solar-pro support
+
+solar-pro introduces block skip connections where blocks are connected
+to other, non-sequential blocks with a scale multiple
+
+this change adds 4 new keys to store the skip connections and one new
+tensor to store the scalar. the scalar is implemented a 1-dimensional
+tensor with 2 elements dervied from the model's bskcn_tv configuration.
+in general, the values are (bskcn_tv, 1 - bskcn_tv)
+---
+ src/llama.cpp | 267 +++++++++++++++++++++++++++++++++++++++++++++++---
+ 1 file changed, 254 insertions(+), 13 deletions(-)
+
+diff --git a/src/llama.cpp b/src/llama.cpp
+index f79bd782..b7771f53 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -213,6 +213,7 @@ enum llm_arch {
+     LLM_ARCH_NEMOTRON,
+     LLM_ARCH_EXAONE,
+     LLM_ARCH_RWKV6,
++    LLM_ARCH_SOLAR,
+     LLM_ARCH_UNKNOWN,
+ };
+ 
+@@ -261,6 +262,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_NEMOTRON,        "nemotron"     },
+     { LLM_ARCH_EXAONE,          "exaone"       },
+     { LLM_ARCH_RWKV6,           "rwkv6"        },
++    { LLM_ARCH_SOLAR,           "solar"        },
+     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
+ };
+ 
+@@ -314,6 +316,7 @@ enum llm_kv {
+     LLM_KV_ATTENTION_KV_LORA_RANK,
+     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
+     LLM_KV_ATTENTION_SLIDING_WINDOW,
++    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+ 
+     LLM_KV_ROPE_DIMENSION_COUNT,
+     LLM_KV_ROPE_FREQ_BASE,
+@@ -405,19 +408,20 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
+     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
+ 
+-    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
+-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
+-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"         },
+-    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"              },
+-    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"             },
+-    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"           },
+-    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"     },
+-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon" },
+-    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                 },
+-    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"            },
+-    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
+-    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
+-    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
++    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"               },
++    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"            },
++    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,         "%s.attention.max_alibi_bias"           },
++    { LLM_KV_ATTENTION_CLAMP_KQV,              "%s.attention.clamp_kqv"                },
++    { LLM_KV_ATTENTION_KEY_LENGTH,             "%s.attention.key_length"               },
++    { LLM_KV_ATTENTION_VALUE_LENGTH,           "%s.attention.value_length"             },
++    { LLM_KV_ATTENTION_LAYERNORM_EPS,          "%s.attention.layer_norm_epsilon"       },
++    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      "%s.attention.layer_norm_rms_epsilon"   },
++    { LLM_KV_ATTENTION_CAUSAL,                 "%s.attention.causal"                   },
++    { LLM_KV_ATTENTION_Q_LORA_RANK,            "%s.attention.q_lora_rank"              },
++    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"             },
++    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count"   },
++    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
++    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
+ 
+     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+@@ -589,6 +593,7 @@ enum llm_tensor {
+     LLM_TENSOR_ENC_FFN_DOWN,
+     LLM_TENSOR_ENC_FFN_UP,
+     LLM_TENSOR_ENC_OUTPUT_NORM,
++    LLM_TENSOR_BSKCN_TV,
+ };
+ 
+ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+@@ -1408,6 +1413,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
+         },
+     },
++    {
++        LLM_ARCH_SOLAR,
++        {
++            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
++            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
++            { LLM_TENSOR_OUTPUT,          "output" },
++            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
++            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
++            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
++            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
++            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
++            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
++            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
++            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
++            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
++            { LLM_TENSOR_BSKCN_TV,        "bskcn_tv" },
++        },
++    },
+     {
+         LLM_ARCH_UNKNOWN,
+         {
+@@ -2237,6 +2260,7 @@ enum e_model {
+     MODEL_15B,
+     MODEL_16B,
+     MODEL_20B,
++    MODEL_22B,
+     MODEL_30B,
+     MODEL_34B,
+     MODEL_35B,
+@@ -2284,6 +2308,8 @@ struct llama_hparams {
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ 
++    std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
++
+     uint32_t n_layer_dense_lead = 0;
+     uint32_t n_lora_q = 0;
+     uint32_t n_lora_kv = 0;
+@@ -2349,6 +2375,7 @@ struct llama_hparams {
+         if (this->n_head_arr    != other.n_head_arr)    return true;
+         if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
+         if (this->n_ff_arr      != other.n_ff_arr)      return true;
++        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+ 
+         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
+         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+@@ -2455,6 +2482,14 @@ struct llama_hparams {
+             return ssm_d_state * ssm_d_inner;
+         }
+     }
++
++    bool n_bskcn(uint32_t n, uint32_t il = 0) const {
++        if (il < n_layer) {
++            return n_bskcn_arr[n][il] > 0;
++        }
++
++        GGML_ABORT("fatal error");
++    }
+ };
+ 
+ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+@@ -2635,6 +2670,8 @@ struct llama_layer {
+     struct ggml_tensor * ffn_gate_scale;
+     struct ggml_tensor * ffn_up_scale;
+     struct ggml_tensor * ffn_down_scale;
++
++    struct ggml_tensor * bskcn_tv;
+ };
+ 
+ // very similar to llama_batch,
+@@ -5937,6 +5974,21 @@ static void llm_load_hparams(
+                     default: model.type = e_model::MODEL_UNKNOWN;
+                 }
+             } break;
++        case LLM_ARCH_SOLAR:
++            {
++                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++
++                for (int i = 0; i < hparams.n_bskcn_arr.max_size(); ++i) {
++                    auto & bskcn = hparams.n_bskcn_arr.at(i);
++                    bskcn.fill(0);
++                    ml.get_key_or_arr(::format(LLM_KV_NAMES.at(LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION), LLM_ARCH_NAMES.at(ml.llm_kv.arch), i), bskcn, hparams.n_layer, false);
++                }
++
++                switch (hparams.n_layer) {
++                    case 64: model.type = e_model::MODEL_22B; break;
++                    default: model.type = e_model::MODEL_UNKNOWN;
++                }
++            }
+         default: (void)0;
+     }
+ 
+@@ -8420,6 +8472,38 @@ static bool llm_load_tensors(
+                     }
+ 
+                 } break;
++            case LLM_ARCH_SOLAR:
++                {
++                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
++
++                    // output
++                    {
++                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
++                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
++                    }
++
++                    for (int i = 0; i < n_layer; ++i) {
++                        ggml_context * ctx_layer = ctx_for_layer(i);
++                        ggml_context * ctx_split = ctx_for_layer_split(i);
++
++                        auto & layer = model.layers[i];
++
++                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
++
++                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
++                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
++                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
++                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
++
++                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
++
++                        layer.bskcn_tv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_BSKCN_TV, "weight"), {2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
++
++                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
++                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
++                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
++                    }
++                } break;
+             default:
+                 throw std::runtime_error("unknown architecture");
+         }
+@@ -15173,6 +15257,158 @@ struct llm_build_context {
+ 
+         return gf;
+     }
++
++    ggml_cgraph * build_solar() {
++        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
++
++        // mutable variable, needed during the last layer of the computation to skip unused tokens
++        int32_t n_tokens = this->n_tokens;
++
++        const int64_t n_embd_head = hparams.n_embd_head_v;
++        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
++        GGML_ASSERT(n_embd_head == hparams.n_rot);
++
++        struct ggml_tensor * cur;
++        struct ggml_tensor * inpL;
++
++        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
++
++        // inp_pos - contains the positions
++        struct ggml_tensor * inp_pos = build_inp_pos();
++
++        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
++        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
++
++        struct ggml_tensor * bskcn_1;
++        struct ggml_tensor * bskcn_2;
++
++        for (int il = 0; il < n_layer; ++il) {
++            struct ggml_tensor * inpSA = inpL;
++
++            if (hparams.n_bskcn(0, il)) {
++                bskcn_1 = inpSA;
++            }
++
++            if (hparams.n_bskcn(1, il)) {
++                bskcn_2 = inpSA;
++            }
++
++            if (hparams.n_bskcn(2, il)) {
++                inpSA = ggml_add(
++                   ctx0,
++                   ggml_mul(ctx0, bskcn_1, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
++                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
++            }
++
++            if (hparams.n_bskcn(3, il)) {
++                inpSA = ggml_add(
++                   ctx0,
++                   ggml_mul(ctx0, bskcn_2, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, 0)),
++                   ggml_mul(ctx0, inpSA, ggml_view_1d(ctx0, model.layers[il].bskcn_tv, 1, ggml_element_size(model.layers[il].bskcn_tv))));
++            }
++
++            // norm
++            cur = llm_build_norm(ctx0, inpL, hparams,
++                    model.layers[il].attn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "attn_norm", il);
++
++            // self-attention
++            {
++                // rope freq factors for llama3; may return nullptr for llama2 and other models
++                struct ggml_tensor * rope_factors = build_rope_factors(il);
++
++                // compute Q and K and RoPE them
++                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
++                cb(Qcur, "Qcur", il);
++                if (model.layers[il].bq) {
++                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
++                    cb(Qcur, "Qcur", il);
++                }
++
++                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
++                cb(Kcur, "Kcur", il);
++                if (model.layers[il].bk) {
++                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
++                    cb(Kcur, "Kcur", il);
++                }
++
++                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
++                cb(Vcur, "Vcur", il);
++                if (model.layers[il].bv) {
++                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
++                    cb(Vcur, "Vcur", il);
++                }
++
++                Qcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Qcur, "Qcur", il);
++
++                Kcur = ggml_rope_ext(
++                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
++                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                    ext_factor, attn_factor, beta_fast, beta_slow
++                );
++                cb(Kcur, "Kcur", il);
++
++                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
++                        model.layers[il].wo, model.layers[il].bo,
++                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
++            }
++
++            if (il == n_layer - 1) {
++                // skip computing output for unused tokens
++                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
++                n_tokens = n_outputs;
++                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
++                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
++            }
++
++            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++            cb(ffn_inp, "ffn_inp", il);
++
++            // feed-forward network
++            cur = llm_build_norm(ctx0, ffn_inp, hparams,
++                    model.layers[il].ffn_norm, NULL,
++                    LLM_NORM_RMS, cb, il);
++            cb(cur, "ffn_norm", il);
++
++            cur = llm_build_ffn(ctx0, lctx, cur,
++                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
++                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
++                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
++                    NULL,
++                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
++            cb(cur, "ffn_out", il);
++
++            cur = ggml_add(ctx0, cur, ffn_inp);
++            cb(cur, "ffn_out", il);
++
++            cur = lctx.cvec.apply_to(ctx0, cur, il);
++            cb(cur, "l_out", il);
++
++            // input for next layer
++            inpL = cur;
++        }
++
++        cur = inpL;
++
++        cur = llm_build_norm(ctx0, cur, hparams,
++                model.output_norm, NULL,
++                LLM_NORM_RMS, cb, -1);
++        cb(cur, "result_norm", -1);
++
++        // lm_head
++        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
++        cb(cur, "result_output", -1);
++
++        ggml_build_forward_expand(gf, cur);
++
++        return gf;
++    }
+ };
+ 
+ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
+@@ -15423,6 +15659,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_rwkv6();
+             } break;
++        case LLM_ARCH_SOLAR:
++            {
++                result = llm.build_solar();
++            } break;
+         default:
+             GGML_ABORT("fatal error");
+     }
+@@ -18503,6 +18743,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+         case LLM_ARCH_ARCTIC:
+         case LLM_ARCH_DEEPSEEK2:
+         case LLM_ARCH_CHATGLM:
++        case LLM_ARCH_SOLAR:
+             return LLAMA_ROPE_TYPE_NORM;
+ 
+         // the pairs of head values are offset by n_rot/2
+-- 
+2.46.0
+
diff --git a/llm/patches/06-qwen2.diff b/llm/patches/06-qwen2.diff
deleted file mode 100644
index d7b0c1555..000000000
--- a/llm/patches/06-qwen2.diff
+++ /dev/null
@@ -1,13 +0,0 @@
-diff --git a/llama.cpp b/llama.cpp
-index 40d2ec2c..f34eb79a 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -6943,7 +6943,7 @@ static struct ggml_tensor * llm_build_kqv(
-         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
-         cb(kq, "kq", il);
- 
--        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
-+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
-             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
-             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
-             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
diff --git a/llm/patches/07-gemma.diff b/llm/patches/07-gemma.diff
deleted file mode 100644
index 86eac3d17..000000000
--- a/llm/patches/07-gemma.diff
+++ /dev/null
@@ -1,305 +0,0 @@
-From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001
-From: Ollama maintainers <hello@ollama.com>
-Date: Wed, 26 Jun 2024 16:18:09 -0700
-Subject: [PATCH] Architecture support
-
----
- llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
- 1 file changed, 193 insertions(+), 1 deletion(-)
-
-diff --git a/llama.cpp b/llama.cpp
-index 61948751..3b4196f5 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -217,6 +217,7 @@ enum llm_arch {
-     LLM_ARCH_INTERNLM2,
-     LLM_ARCH_MINICPM,
-     LLM_ARCH_GEMMA,
-+    LLM_ARCH_GEMMA2,
-     LLM_ARCH_STARCODER2,
-     LLM_ARCH_MAMBA,
-     LLM_ARCH_XVERSE,
-@@ -255,6 +256,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
-     { LLM_ARCH_INTERNLM2,       "internlm2"    },
-     { LLM_ARCH_MINICPM,         "minicpm"      },
-     { LLM_ARCH_GEMMA,           "gemma"        },
-+    { LLM_ARCH_GEMMA2,          "gemma2"       },
-     { LLM_ARCH_STARCODER2,      "starcoder2"   },
-     { LLM_ARCH_MAMBA,           "mamba"        },
-     { LLM_ARCH_XVERSE,          "xverse"       },
-@@ -464,10 +466,12 @@ enum llm_tensor {
-     LLM_TENSOR_ATTN_NORM,
-     LLM_TENSOR_ATTN_NORM_2,
-     LLM_TENSOR_ATTN_OUT_NORM,
-+    LLM_TENSOR_ATTN_POST_NORM,
-     LLM_TENSOR_ATTN_ROT_EMBD,
-     LLM_TENSOR_FFN_GATE_INP,
-     LLM_TENSOR_FFN_GATE_INP_SHEXP,
-     LLM_TENSOR_FFN_NORM,
-+    LLM_TENSOR_FFN_POST_NORM,
-     LLM_TENSOR_FFN_GATE,
-     LLM_TENSOR_FFN_DOWN,
-     LLM_TENSOR_FFN_UP,
-@@ -960,6 +964,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
-             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-         },
-     },
-+    {
-+        LLM_ARCH_GEMMA2,
-+        {
-+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
-+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
-+        },
-+    },
-     {
-         LLM_ARCH_STARCODER2,
-         {
-@@ -1941,6 +1963,8 @@ enum e_model {
-     MODEL_8x22B,
-     MODEL_16x12B,
-     MODEL_10B_128x3_66B,
-+    MODEL_9B,
-+    MODEL_27B,
- };
- 
- static const size_t kiB = 1024;
-@@ -2114,6 +2138,7 @@ struct llama_layer {
-     struct ggml_tensor * attn_out_norm_b;
-     struct ggml_tensor * attn_q_a_norm;
-     struct ggml_tensor * attn_kv_a_norm;
-+    struct ggml_tensor * attn_post_norm;
- 
-     // attention
-     struct ggml_tensor * wq;
-@@ -2136,6 +2161,7 @@ struct llama_layer {
-     // normalization
-     struct ggml_tensor * ffn_norm;
-     struct ggml_tensor * ffn_norm_b;
-+    struct ggml_tensor * ffn_post_norm;
-     struct ggml_tensor * layer_out_norm;
-     struct ggml_tensor * layer_out_norm_b;
-     struct ggml_tensor * ffn_norm_exps;
-@@ -4529,6 +4555,16 @@ static void llm_load_hparams(
-                 }
-             } break;
-         case LLM_ARCH_GEMMA:
-+            {
-+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-+
-+                switch (hparams.n_layer) {
-+                    case 18: model.type = e_model::MODEL_9B; break;
-+                    case 28: model.type = e_model::MODEL_27B; break;
-+                    default: model.type = e_model::MODEL_UNKNOWN;
-+               }
-+            } break;
-+        case LLM_ARCH_GEMMA2:
-             {
-                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
- 
-@@ -6305,6 +6341,40 @@ static bool llm_load_tensors(
-                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                     }
-                 } break;
-+            case LLM_ARCH_GEMMA2:
-+                {
-+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-+
-+                    // output
-+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
-+
-+                    const int64_t n_ff          = hparams.n_ff;
-+                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
-+                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-+
-+                    for (uint32_t i = 0; i < n_layer; ++i) {
-+                        ggml_context * ctx_layer = ctx_for_layer(i);
-+                        ggml_context * ctx_split = ctx_for_layer_split(i);
-+
-+                        auto & layer = model.layers[i];
-+
-+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-+
-+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
-+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
-+                        layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
-+
-+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-+                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
-+                    }
-+                } break;
-             case LLM_ARCH_STARCODER2:
-                 {
-                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-@@ -10614,6 +10684,123 @@ struct llm_build_context {
-         return gf;
-     }
- 
-+    struct ggml_cgraph * build_gemma2() {
-+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
-+
-+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+
-+        struct ggml_tensor * cur;
-+        struct ggml_tensor * inpL;
-+
-+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
-+
-+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-+        cb(inpL, "inp_scaled", -1);
-+
-+        // inp_pos - contains the positions
-+        struct ggml_tensor * inp_pos = build_inp_pos();
-+
-+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-+
-+        for (int il = 0; il < n_layer; ++il) {
-+            // norm
-+            cur = llm_build_norm(ctx0, inpL, hparams,
-+                    model.layers[il].attn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_norm", il);
-+
-+            // self-attention
-+            {
-+                // compute Q and K and RoPE them
-+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-+                cb(Qcur, "Qcur", il);
-+
-+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-+                cb(Kcur, "Kcur", il);
-+
-+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-+                cb(Vcur, "Vcur", il);
-+
-+                Qcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Qcur, "Qcur", il);
-+
-+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
-+                cb(Qcur, "Qcur_scaled", il);
-+
-+                Kcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Kcur, "Kcur", il);
-+
-+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
-+                        model.layers[il].wo, NULL,
-+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-+            }
-+
-+            if (il == n_layer - 1) {
-+                // skip computing output for unused tokens
-+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                    model.layers[il].attn_post_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_post_norm", il);
-+
-+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-+            cb(sa_out, "sa_out", il);
-+
-+            cur = llm_build_norm(ctx0, sa_out, hparams,
-+                    model.layers[il].ffn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "ffn_norm", il);
-+
-+            // feed-forward network
-+            {
-+                cur = llm_build_ffn(ctx0, cur,
-+                        model.layers[il].ffn_up, NULL,
-+                        model.layers[il].ffn_gate, NULL,
-+                        model.layers[il].ffn_down, NULL,
-+                        NULL,
-+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-+                cb(cur, "ffn_out", il);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                model.layers[il].ffn_post_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+            cb(cur, "ffn_post_norm", -1);
-+
-+            cur = ggml_add(ctx0, cur, sa_out);
-+            cb(cur, "l_out", il);
-+
-+            // input for next layer
-+            inpL = cur;
-+        }
-+
-+        cur = inpL;
-+
-+        cur = llm_build_norm(ctx0, cur, hparams,
-+                model.output_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+        cb(cur, "result_norm", -1);
-+
-+        // lm_head
-+        cur = ggml_mul_mat(ctx0, model.output, cur);
-+        cb(cur, "result_output", -1);
-+
-+        ggml_build_forward_expand(gf, cur);
-+
-+        return gf;
-+    }
-+
-     struct ggml_cgraph * build_starcoder2() {
-         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
- 
-@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph(
-             {
-                 result = llm.build_gemma();
-             } break;
-+        case LLM_ARCH_GEMMA2:
-+            {
-+                result = llm.build_gemma2();
-+            } break;
-         case LLM_ARCH_STARCODER2:
-             {
-                 result = llm.build_starcoder2();
-@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
-         case LLM_ARCH_PHI2:
-         case LLM_ARCH_PHI3:
-         case LLM_ARCH_GEMMA:
-+        case LLM_ARCH_GEMMA2:
-         case LLM_ARCH_STARCODER2:
-         case LLM_ARCH_GPTNEOX:
-             return LLAMA_ROPE_TYPE_NEOX;
-@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal(
-         if (add_ass) {
-             ss << "<s>assistant\n";
-         }
--    } else if (tmpl == "gemma" || tmpl.find("<start_of_turn>") != std::string::npos) {
-+    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("<start_of_turn>") != std::string::npos) {
-         // google/gemma-7b-it
-         std::string system_prompt = "";
-         for (auto message : chat) {
--- 
-2.45.2
-
diff --git a/llm/payload.go b/llm/payload.go
deleted file mode 100644
index 9296db336..000000000
--- a/llm/payload.go
+++ /dev/null
@@ -1,231 +0,0 @@
-package llm
-
-import (
-	"compress/gzip"
-	"errors"
-	"fmt"
-	"io"
-	"io/fs"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"runtime"
-	"slices"
-	"strings"
-
-	"golang.org/x/sync/errgroup"
-
-	"github.com/ollama/ollama/gpu"
-)
-
-var errPayloadMissing = errors.New("expected payloads not included in this build of ollama")
-
-func Init() error {
-	payloadsDir, err := gpu.PayloadsDir()
-	if err != nil {
-		return err
-	}
-
-	if runtime.GOOS != "windows" {
-		slog.Info("extracting embedded files", "dir", payloadsDir)
-		binGlob := "build/*/*/*/bin/*"
-
-		// extract server libraries
-		err = extractFiles(payloadsDir, binGlob)
-		if err != nil {
-			return fmt.Errorf("extract binaries: %v", err)
-		}
-	}
-
-	var variants []string
-	for v := range availableServers() {
-		variants = append(variants, v)
-	}
-	slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
-	slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
-
-	return nil
-}
-
-// binary names may contain an optional variant separated by '_'
-// For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
-// Any library without a variant is the lowest common denominator
-func availableServers() map[string]string {
-	payloadsDir, err := gpu.PayloadsDir()
-	if err != nil {
-		slog.Error("payload lookup error", "error", err)
-		return nil
-	}
-
-	// glob payloadsDir for files that start with ollama_
-	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
-
-	files, err := filepath.Glob(pattern)
-	if err != nil {
-		slog.Debug("could not glob", "pattern", pattern, "error", err)
-		return nil
-	}
-
-	servers := make(map[string]string)
-	for _, file := range files {
-		slog.Debug("availableServers : found", "file", file)
-		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
-	}
-
-	return servers
-}
-
-// serversForGpu returns a list of compatible servers give the provided GPU
-// info, ordered by performance. assumes Init() has been called
-// TODO - switch to metadata based mapping
-func serversForGpu(info gpu.GpuInfo) []string {
-	// glob workDir for files that start with ollama_
-	availableServers := availableServers()
-	requested := info.Library
-	if info.Variant != gpu.CPUCapabilityNone {
-		requested += "_" + info.Variant.String()
-	}
-
-	servers := []string{}
-
-	// exact match first
-	for a := range availableServers {
-		if a == requested {
-			servers = []string{a}
-
-			if a == "metal" {
-				return servers
-			}
-
-			break
-		}
-	}
-
-	alt := []string{}
-
-	// Then for GPUs load alternates and sort the list for consistent load ordering
-	if info.Library != "cpu" {
-		for a := range availableServers {
-			if info.Library == strings.Split(a, "_")[0] && a != requested {
-				alt = append(alt, a)
-			}
-		}
-
-		slices.Sort(alt)
-		servers = append(servers, alt...)
-	}
-
-	// Load up the best CPU variant if not primary requested
-	if info.Library != "cpu" {
-		variant := gpu.GetCPUCapability()
-		// If no variant, then we fall back to default
-		// If we have a variant, try that if we find an exact match
-		// Attempting to run the wrong CPU instructions will panic the
-		// process
-		if variant != gpu.CPUCapabilityNone {
-			for cmp := range availableServers {
-				if cmp == "cpu_"+variant.String() {
-					servers = append(servers, cmp)
-					break
-				}
-			}
-		} else {
-			servers = append(servers, "cpu")
-		}
-	}
-
-	if len(servers) == 0 {
-		servers = []string{"cpu"}
-	}
-
-	return servers
-}
-
-// Return the optimal server for this CPU architecture
-func serverForCpu() string {
-	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
-		return "metal"
-	}
-	variant := gpu.GetCPUCapability()
-	availableServers := availableServers()
-	if variant != gpu.CPUCapabilityNone {
-		for cmp := range availableServers {
-			if cmp == "cpu_"+variant.String() {
-				return cmp
-			}
-		}
-	}
-	return "cpu"
-}
-
-// extract extracts the embedded files to the target directory
-func extractFiles(targetDir string, glob string) error {
-	files, err := fs.Glob(libEmbed, glob)
-	if err != nil || len(files) == 0 {
-		return errPayloadMissing
-	}
-
-	if err := os.MkdirAll(targetDir, 0o755); err != nil {
-		return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
-	}
-
-	g := new(errgroup.Group)
-
-	// build/$OS/$GOARCH/$VARIANT/{bin,lib}/$FILE
-	for _, file := range files {
-		filename := file
-
-		variant := filepath.Base(filepath.Dir(filepath.Dir(filename)))
-
-		slog.Debug("extracting", "variant", variant, "file", filename)
-
-		g.Go(func() error {
-			srcf, err := libEmbed.Open(filename)
-			if err != nil {
-				return err
-			}
-			defer srcf.Close()
-
-			src := io.Reader(srcf)
-			if strings.HasSuffix(filename, ".gz") {
-				src, err = gzip.NewReader(src)
-				if err != nil {
-					return fmt.Errorf("decompress payload %s: %v", filename, err)
-				}
-				filename = strings.TrimSuffix(filename, ".gz")
-			}
-
-			variantDir := filepath.Join(targetDir, variant)
-			if err := os.MkdirAll(variantDir, 0o755); err != nil {
-				return fmt.Errorf("extractFiles could not mkdir %s: %v", variantDir, err)
-			}
-
-			base := filepath.Base(filename)
-			destFilename := filepath.Join(variantDir, base)
-
-			_, err = os.Stat(destFilename)
-			switch {
-			case errors.Is(err, os.ErrNotExist):
-				destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
-				if err != nil {
-					return fmt.Errorf("write payload %s: %v", filename, err)
-				}
-				defer destFile.Close()
-				if _, err := io.Copy(destFile, src); err != nil {
-					return fmt.Errorf("copy payload %s: %v", filename, err)
-				}
-			case err != nil:
-				return fmt.Errorf("stat payload %s: %v", filename, err)
-			}
-			return nil
-		})
-	}
-
-	err = g.Wait()
-	if err != nil {
-		// If we fail to extract, the payload dir is unusable, so cleanup whatever we extracted
-		gpu.Cleanup()
-		return err
-	}
-	return nil
-}
diff --git a/llm/server.go b/llm/server.go
index 8b63cfbd5..6c504f149 100644
--- a/llm/server.go
+++ b/llm/server.go
@@ -24,16 +24,18 @@ import (
 	"golang.org/x/sync/semaphore"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/build"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/runners"
 )
 
 type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
 	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embedding(ctx context.Context, prompt string) ([]float64, error)
+	Embedding(ctx context.Context, input string) ([]float32, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -44,11 +46,12 @@ type LlamaServer interface {
 
 // llmServer is an instance of the llama.cpp server
 type llmServer struct {
-	port    int
-	cmd     *exec.Cmd
-	done    chan error // Channel to signal when the process exits
-	status  *StatusWriter
-	options api.Options
+	port        int
+	cmd         *exec.Cmd
+	done        chan error // Channel to signal when the process exits
+	status      *StatusWriter
+	options     api.Options
+	numParallel int
 
 	estimate    MemoryEstimate
 	totalLayers uint64
@@ -88,6 +91,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	var estimate MemoryEstimate
 	var systemTotalMemory uint64
 	var systemFreeMemory uint64
+	var systemSwapFreeMemory uint64
 
 	systemMemInfo, err := gpu.GetCPUMem()
 	if err != nil {
@@ -95,7 +99,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	} else {
 		systemTotalMemory = systemMemInfo.TotalMemory
 		systemFreeMemory = systemMemInfo.FreeMemory
-		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", systemFreeMemory)
+		systemSwapFreeMemory = systemMemInfo.FreeSwap
+		slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
 	}
 
 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
@@ -103,7 +108,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		gpus = gpu.GetCPUInfo()
 	}
 	if len(gpus) == 1 && gpus[0].Library == "cpu" {
-		cpuRunner = serverForCpu()
+		cpuRunner = runners.ServerForCpu()
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
 	} else {
 		estimate = EstimateGPULayers(gpus, ggml, projectors, opts)
@@ -115,13 +120,24 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			opts.NumGPU = 0
 		case gpus[0].Library != "metal" && estimate.Layers == 0:
 			// Don't bother loading into the GPU if no layers can fit
-			cpuRunner = serverForCpu()
+			cpuRunner = runners.ServerForCpu()
 			gpus = gpu.GetCPUInfo()
 		case opts.NumGPU < 0 && estimate.Layers > 0 && gpus[0].Library != "cpu":
 			opts.NumGPU = estimate.Layers
 		}
 	}
 
+	// On linux and windows, over-allocating CPU memory will almost always result in an error
+	// Darwin has fully dynamic swap so has no direct concept of free swap space
+	if runtime.GOOS != "darwin" {
+		systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
+		available := systemFreeMemory + systemSwapFreeMemory
+		if systemMemoryRequired > available {
+			slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "available", available, "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "swap", format.HumanBytes2(systemSwapFreeMemory))
+			return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(systemMemoryRequired), format.HumanBytes2(available))
+		}
+	}
+
 	estimate.log()
 
 	// Loop through potential servers
@@ -131,14 +147,22 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}
 
-	availableServers := availableServers()
+	rDir, err := runners.Refresh(build.EmbedFS)
+	if err != nil {
+		return nil, err
+	}
+
+	availableServers := runners.GetAvailableServers(rDir)
+	if len(availableServers) == 0 {
+		return nil, finalErr
+	}
 	var servers []string
 	if cpuRunner != "" {
 		servers = []string{cpuRunner}
 	} else {
-		servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
+		servers = runners.ServersForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
 	}
-	demandLib := envconfig.LLMLibrary
+	demandLib := envconfig.LLMLibrary()
 	if demandLib != "" {
 		serverPath := availableServers[demandLib]
 		if serverPath == "" {
@@ -159,23 +183,23 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 	params := []string{
 		"--model", model,
-		"--ctx-size", fmt.Sprintf("%d", opts.NumCtx),
-		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
+		"--ctx-size", strconv.Itoa(opts.NumCtx),
+		"--batch-size", strconv.Itoa(opts.NumBatch),
 		"--embedding",
 	}
 
 	params = append(params, "--log-disable")
 
 	if opts.NumGPU >= 0 {
-		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU))
+		params = append(params, "--n-gpu-layers", strconv.Itoa(opts.NumGPU))
 	}
 
-	if envconfig.Debug {
+	if envconfig.Debug() {
 		params = append(params, "--verbose")
 	}
 
 	if opts.MainGPU > 0 {
-		params = append(params, "--main-gpu", fmt.Sprintf("%d", opts.MainGPU))
+		params = append(params, "--main-gpu", strconv.Itoa(opts.MainGPU))
 	}
 
 	if len(adapters) > 0 {
@@ -189,14 +213,14 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	}
 
 	if opts.NumThread > 0 {
-		params = append(params, "--threads", fmt.Sprintf("%d", opts.NumThread))
+		params = append(params, "--threads", strconv.Itoa(opts.NumThread))
 	}
 
 	if !opts.F16KV {
 		params = append(params, "--memory-f32")
 	}
 
-	flashAttnEnabled := envconfig.FlashAttention
+	flashAttnEnabled := envconfig.FlashAttention()
 
 	for _, g := range gpus {
 		// only cuda (compute capability 7+) and metal support flash attention
@@ -208,7 +232,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if g.Library == "metal" &&
 			uint64(opts.NumGPU) > 0 &&
 			uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
-			opts.UseMMap = api.TriStateFalse
+			opts.UseMMap = new(bool)
+			*opts.UseMMap = false
 		}
 	}
 
@@ -219,10 +244,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	// Windows CUDA should not use mmap for best performance
 	// Linux  with a model larger than free space, mmap leads to thrashing
 	// For CPU loads we want the memory to be allocated, not FS cache
-	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) ||
-		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) ||
-		(gpus[0].Library == "cpu" && opts.UseMMap == api.TriStateUndefined) ||
-		opts.UseMMap == api.TriStateFalse {
+	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == nil) ||
+		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == nil) ||
+		(gpus[0].Library == "cpu" && opts.UseMMap == nil) ||
+		(opts.UseMMap != nil && !*opts.UseMMap) {
 		params = append(params, "--no-mmap")
 	}
 
@@ -230,21 +255,23 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--mlock")
 	}
 
-	if opts.UseNUMA {
-		params = append(params, "--numa")
+	if gpu.IsNUMA() && gpus[0].Library == "cpu" {
+		numaMode := "distribute"
+		if runtime.GOOS == "linux" {
+			if _, err := exec.LookPath("numactl"); err == nil {
+				numaMode = "numactl"
+			}
+		}
+		params = append(params, "--numa", numaMode)
 	}
 
-	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))
+	params = append(params, "--parallel", strconv.Itoa(numParallel))
 
 	if estimate.TensorSplit != "" {
 		params = append(params, "--tensor-split", estimate.TensorSplit)
 	}
 
-	if estimate.TensorSplit != "" {
-		params = append(params, "--tensor-split", estimate.TensorSplit)
-	}
-
-	for i := range len(servers) {
+	for i := range servers {
 		dir := availableServers[servers[i]]
 		if dir == "" {
 			// Shouldn't happen
@@ -276,20 +303,18 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if runtime.GOOS == "windows" {
 			pathEnv = "PATH"
 		}
-		// prepend the server directory to LD_LIBRARY_PATH/PATH and the parent dir for common dependencies
-		libraryPaths := []string{dir, filepath.Dir(dir)}
+		// Start with the server directory for the LD_LIBRARY_PATH/PATH
+		libraryPaths := []string{dir}
 
 		if libraryPath, ok := os.LookupEnv(pathEnv); ok {
-			// Append our runner directory to the path
-			// This will favor system libraries over our bundled library dependencies
+			// favor our bundled library dependencies over system libraries
 			libraryPaths = append(libraryPaths, filepath.SplitList(libraryPath)...)
 		}
 
 		// Note: we always put the dependency path first
-		// since this was the exact version we verified for AMD GPUs
-		// and we favor what the user had in their path
+		// since this was the exact version we compiled/linked against
 		if gpus[0].DependencyPath != "" {
-			// TODO refine for multi-gpu support
+			// assume gpus from the same library have the same dependency path
 			libraryPaths = append([]string{gpus[0].DependencyPath}, libraryPaths...)
 		}
 
@@ -302,7 +327,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		_, err := os.Stat(server)
 		if errors.Is(err, os.ErrNotExist) {
 			slog.Warn("llama server disappeared, reinitializing payloads", "path", server, "error", err)
-			err = Init()
+			_, err = runners.Refresh(build.EmbedFS)
 			if err != nil {
 				slog.Warn("failed to reinitialize payloads", "error", err)
 				return nil, err
@@ -315,6 +340,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			status:      NewStatusWriter(os.Stderr),
 			options:     opts,
 			estimate:    estimate,
+			numParallel: numParallel,
 			sem:         semaphore.NewWeighted(int64(numParallel)),
 			totalLayers: ggml.KV().BlockCount() + 1,
 			gpus:        gpus,
@@ -324,6 +350,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		s.cmd.Env = os.Environ()
 		s.cmd.Stdout = os.Stdout
 		s.cmd.Stderr = s.status
+		s.cmd.SysProcAttr = LlamaServerSysProcAttr
 
 		envWorkarounds := [][2]string{}
 		for _, gpu := range gpus {
@@ -359,12 +386,14 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 
 		slog.Info("starting llama server", "cmd", s.cmd.String())
-		if envconfig.Debug {
+		if envconfig.Debug() {
 			filteredEnv := []string{}
 			for _, ev := range s.cmd.Env {
 				if strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCR_") ||
 					strings.HasPrefix(ev, "ROCM_") ||
 					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "GPU_") ||
 					strings.HasPrefix(ev, "HSA_") ||
 					strings.HasPrefix(ev, "GGML_") ||
 					strings.HasPrefix(ev, "PATH=") ||
@@ -377,7 +406,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 
 		if err = s.cmd.Start(); err != nil {
-			// Detect permission denied and augment them essage about noexec
+			// Detect permission denied and augment the message about noexec
 			if errors.Is(err, os.ErrPermission) {
 				finalErr = fmt.Errorf("unable to start server %w.  %s may have noexec set.  Set OLLAMA_TMPDIR for server to a writable executable directory", err, dir)
 				continue
@@ -393,7 +422,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 		// reap subprocess when it exits
 		go func() {
-			s.done <- s.cmd.Wait()
+			err := s.cmd.Wait()
+			// Favor a more detailed message over the process exit status
+			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
+				slog.Debug("llama runner terminated", "error", err)
+				if strings.Contains(s.status.LastErrMsg, "unknown model") {
+					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
+				}
+				s.done <- errors.New(s.status.LastErrMsg)
+			} else {
+				s.done <- err
+			}
 		}()
 
 		return s, nil
@@ -542,8 +581,7 @@ func (s *llmServer) Ping(ctx context.Context) error {
 
 func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 	start := time.Now()
-	stallDuration := 5 * time.Minute            // If no progress happens
-	finalLoadDuration := 5 * time.Minute        // After we hit 100%, give the runner more time to come online
+	stallDuration := envconfig.LoadTimeout()    // If no progress happens
 	stallTimer := time.Now().Add(stallDuration) // give up if we stall
 
 	slog.Info("waiting for llama runner to start responding")
@@ -556,14 +594,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			slog.Warn("client connection closed before server finished loading, aborting load")
 			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
 		case err := <-s.done:
-			msg := ""
-			if s.status != nil && s.status.LastErrMsg != "" {
-				msg = s.status.LastErrMsg
-			}
-			if strings.Contains(msg, "unknown model") {
-				return fmt.Errorf("this model is not supported by your version of Ollama. You may need to upgrade")
-			}
-			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
+			return fmt.Errorf("llama runner process has terminated: %w", err)
 		default:
 		}
 		if time.Now().After(stallTimer) {
@@ -602,7 +633,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 				stallTimer = time.Now().Add(stallDuration)
 			} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
 				slog.Debug("model load completed, waiting for server to become available", "status", status.ToString())
-				stallTimer = time.Now().Add(finalLoadDuration)
+				stallTimer = time.Now().Add(stallDuration)
 				fullyLoaded = true
 			}
 			time.Sleep(time.Millisecond * 250)
@@ -665,7 +696,7 @@ type CompletionRequest struct {
 	Prompt  string
 	Format  string
 	Images  []ImageData
-	Options api.Options
+	Options *api.Options
 }
 
 type CompletionResponse struct {
@@ -685,10 +716,9 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	}
 	defer s.sem.Release(1)
 
-	// only allow maximum 10 "context shifts" to avoid infinite generation
+	// put an upper limit on num_predict to avoid the model running on forever
 	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
 		req.Options.NumPredict = 10 * s.options.NumCtx
-		slog.Debug("setting token limit to 10x num_ctx", "num_ctx", s.options.NumCtx, "num_predict", req.Options.NumPredict)
 	}
 
 	request := map[string]any{
@@ -700,6 +730,7 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 		"temperature":       req.Options.Temperature,
 		"top_k":             req.Options.TopK,
 		"top_p":             req.Options.TopP,
+		"min_p":             req.Options.MinP,
 		"tfs_z":             req.Options.TFSZ,
 		"typical_p":         req.Options.TypicalP,
 		"repeat_last_n":     req.Options.RepeatLastN,
@@ -851,10 +882,10 @@ type EmbeddingRequest struct {
 }
 
 type EmbeddingResponse struct {
-	Embedding []float64 `json:"embedding"`
+	Embedding []float32 `json:"embedding"`
 }
 
-func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, error) {
+func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
 	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
@@ -869,18 +900,18 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
 		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
 	}
 
-	data, err := json.Marshal(TokenizeRequest{Content: prompt})
+	data, err := json.Marshal(EmbeddingRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
 	}
 
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
+	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
 	if err != nil {
 		return nil, fmt.Errorf("error creating embed request: %w", err)
 	}
-	req.Header.Set("Content-Type", "application/json")
+	r.Header.Set("Content-Type", "application/json")
 
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := http.DefaultClient.Do(r)
 	if err != nil {
 		return nil, fmt.Errorf("do embedding request: %w", err)
 	}
@@ -896,12 +927,12 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var embedding EmbeddingResponse
-	if err := json.Unmarshal(body, &embedding); err != nil {
+	var e EmbeddingResponse
+	if err := json.Unmarshal(body, &e); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
-	return embedding.Embedding, nil
+	return e.Embedding, nil
 }
 
 type TokenizeRequest struct {
diff --git a/llm/status.go b/llm/status.go
index 0f56b7f99..604fe9e0d 100644
--- a/llm/status.go
+++ b/llm/status.go
@@ -25,7 +25,8 @@ var errorPrefixes = []string{
 	"CUDA error",
 	"cudaMalloc failed",
 	"\"ERR\"",
-	"architecture",
+	"error loading model",
+	"GGML_ASSERT",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {
diff --git a/macapp/src/app.tsx b/macapp/src/app.tsx
index ab17df603..a627e63de 100644
--- a/macapp/src/app.tsx
+++ b/macapp/src/app.tsx
@@ -19,7 +19,7 @@ export default function () {
   const [step, setStep] = useState<Step>(Step.WELCOME)
   const [commandCopied, setCommandCopied] = useState<boolean>(false)
 
-  const command = 'ollama run llama3'
+  const command = 'ollama run llama3.1'
 
   return (
     <div className='drag'>
diff --git a/main.go b/main.go
index 158f0063c..650e03a63 100644
--- a/main.go
+++ b/main.go
@@ -3,8 +3,9 @@ package main
 import (
 	"context"
 
-	"github.com/ollama/ollama/cmd"
 	"github.com/spf13/cobra"
+
+	"github.com/ollama/ollama/cmd"
 )
 
 func main() {
diff --git a/openai/openai.go b/openai/openai.go
index 706d31aa2..2bf9b9f9b 100644
--- a/openai/openai.go
+++ b/openai/openai.go
@@ -3,15 +3,21 @@ package openai
 
 import (
 	"bytes"
+	"encoding/base64"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
+	"log/slog"
 	"math/rand"
 	"net/http"
+	"strings"
 	"time"
 
 	"github.com/gin-gonic/gin"
+
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
 )
 
 type Error struct {
@@ -26,8 +32,9 @@ type ErrorResponse struct {
 }
 
 type Message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
+	Role      string     `json:"role"`
+	Content   any        `json:"content"`
+	ToolCalls []ToolCall `json:"tool_calls,omitempty"`
 }
 
 type Choice struct {
@@ -42,6 +49,12 @@ type ChunkChoice struct {
 	FinishReason *string `json:"finish_reason"`
 }
 
+type CompleteChunkChoice struct {
+	Text         string  `json:"text"`
+	Index        int     `json:"index"`
+	FinishReason *string `json:"finish_reason"`
+}
+
 type Usage struct {
 	PromptTokens     int `json:"prompt_tokens"`
 	CompletionTokens int `json:"completion_tokens"`
@@ -52,6 +65,11 @@ type ResponseFormat struct {
 	Type string `json:"type"`
 }
 
+type EmbedRequest struct {
+	Input any    `json:"input"`
+	Model string `json:"model"`
+}
+
 type ChatCompletionRequest struct {
 	Model            string          `json:"model"`
 	Messages         []Message       `json:"messages"`
@@ -61,9 +79,10 @@ type ChatCompletionRequest struct {
 	Stop             any             `json:"stop"`
 	Temperature      *float64        `json:"temperature"`
 	FrequencyPenalty *float64        `json:"frequency_penalty"`
-	PresencePenalty  *float64        `json:"presence_penalty_penalty"`
+	PresencePenalty  *float64        `json:"presence_penalty"`
 	TopP             *float64        `json:"top_p"`
 	ResponseFormat   *ResponseFormat `json:"response_format"`
+	Tools            []api.Tool      `json:"tools"`
 }
 
 type ChatCompletion struct {
@@ -85,6 +104,79 @@ type ChatCompletionChunk struct {
 	Choices           []ChunkChoice `json:"choices"`
 }
 
+// TODO (https://github.com/ollama/ollama/issues/5259): support []string, []int and [][]int
+type CompletionRequest struct {
+	Model            string   `json:"model"`
+	Prompt           string   `json:"prompt"`
+	FrequencyPenalty float32  `json:"frequency_penalty"`
+	MaxTokens        *int     `json:"max_tokens"`
+	PresencePenalty  float32  `json:"presence_penalty"`
+	Seed             *int     `json:"seed"`
+	Stop             any      `json:"stop"`
+	Stream           bool     `json:"stream"`
+	Temperature      *float32 `json:"temperature"`
+	TopP             float32  `json:"top_p"`
+	Suffix           string   `json:"suffix"`
+}
+
+type Completion struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Usage             Usage                 `json:"usage,omitempty"`
+}
+
+type CompletionChunk struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+}
+
+type ToolCall struct {
+	ID       string `json:"id"`
+	Type     string `json:"type"`
+	Function struct {
+		Name      string `json:"name"`
+		Arguments string `json:"arguments"`
+	} `json:"function"`
+}
+
+type Model struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
+type Embedding struct {
+	Object    string    `json:"object"`
+	Embedding []float32 `json:"embedding"`
+	Index     int       `json:"index"`
+}
+
+type ListCompletion struct {
+	Object string  `json:"object"`
+	Data   []Model `json:"data"`
+}
+
+type EmbeddingList struct {
+	Object string         `json:"object"`
+	Data   []Embedding    `json:"data"`
+	Model  string         `json:"model"`
+	Usage  EmbeddingUsage `json:"usage,omitempty"`
+}
+
+type EmbeddingUsage struct {
+	PromptTokens int `json:"prompt_tokens"`
+	TotalTokens  int `json:"total_tokens"`
+}
+
 func NewError(code int, message string) ErrorResponse {
 	var etype string
 	switch code {
@@ -99,7 +191,31 @@ func NewError(code int, message string) ErrorResponse {
 	return ErrorResponse{Error{Type: etype, Message: message}}
 }
 
+func toolCallId() string {
+	const letterBytes = "abcdefghijklmnopqrstuvwxyz0123456789"
+	b := make([]byte, 8)
+	for i := range b {
+		b[i] = letterBytes[rand.Intn(len(letterBytes))]
+	}
+	return "call_" + strings.ToLower(string(b))
+}
+
 func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
+	toolCalls := make([]ToolCall, len(r.Message.ToolCalls))
+	for i, tc := range r.Message.ToolCalls {
+		toolCalls[i].ID = toolCallId()
+		toolCalls[i].Type = "function"
+		toolCalls[i].Function.Name = tc.Function.Name
+
+		args, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("could not marshall function arguments to json", "error", err)
+			continue
+		}
+
+		toolCalls[i].Function.Arguments = string(args)
+	}
+
 	return ChatCompletion{
 		Id:                id,
 		Object:            "chat.completion",
@@ -108,8 +224,11 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 		SystemFingerprint: "fp_ollama",
 		Choices: []Choice{{
 			Index:   0,
-			Message: Message{Role: r.Message.Role, Content: r.Message.Content},
+			Message: Message{Role: r.Message.Role, Content: r.Message.Content, ToolCalls: toolCalls},
 			FinishReason: func(reason string) *string {
+				if len(toolCalls) > 0 {
+					reason = "tool_calls"
+				}
 				if len(reason) > 0 {
 					return &reason
 				}
@@ -117,7 +236,6 @@ func toChatCompletion(id string, r api.ChatResponse) ChatCompletion {
 			}(r.DoneReason),
 		}},
 		Usage: Usage{
-			// TODO: ollama returns 0 for prompt eval if the prompt was cached, but openai returns the actual count
 			PromptTokens:     r.PromptEvalCount,
 			CompletionTokens: r.EvalCount,
 			TotalTokens:      r.PromptEvalCount + r.EvalCount,
@@ -145,10 +263,173 @@ func toChunk(id string, r api.ChatResponse) ChatCompletionChunk {
 	}
 }
 
-func fromRequest(r ChatCompletionRequest) api.ChatRequest {
+func toCompletion(id string, r api.GenerateResponse) Completion {
+	return Completion{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           r.CreatedAt.Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+		Usage: Usage{
+			PromptTokens:     r.PromptEvalCount,
+			CompletionTokens: r.EvalCount,
+			TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		},
+	}
+}
+
+func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
+	return CompletionChunk{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           time.Now().Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+	}
+}
+
+func toListCompletion(r api.ListResponse) ListCompletion {
+	var data []Model
+	for _, m := range r.Models {
+		data = append(data, Model{
+			Id:      m.Name,
+			Object:  "model",
+			Created: m.ModifiedAt.Unix(),
+			OwnedBy: model.ParseName(m.Name).Namespace,
+		})
+	}
+
+	return ListCompletion{
+		Object: "list",
+		Data:   data,
+	}
+}
+
+func toEmbeddingList(model string, r api.EmbedResponse) EmbeddingList {
+	if r.Embeddings != nil {
+		var data []Embedding
+		for i, e := range r.Embeddings {
+			data = append(data, Embedding{
+				Object:    "embedding",
+				Embedding: e,
+				Index:     i,
+			})
+		}
+
+		return EmbeddingList{
+			Object: "list",
+			Data:   data,
+			Model:  model,
+			Usage: EmbeddingUsage{
+				PromptTokens: r.PromptEvalCount,
+				TotalTokens:  r.PromptEvalCount,
+			},
+		}
+	}
+
+	return EmbeddingList{}
+}
+
+func toModel(r api.ShowResponse, m string) Model {
+	return Model{
+		Id:      m,
+		Object:  "model",
+		Created: r.ModifiedAt.Unix(),
+		OwnedBy: model.ParseName(m).Namespace,
+	}
+}
+
+func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 	var messages []api.Message
 	for _, msg := range r.Messages {
-		messages = append(messages, api.Message{Role: msg.Role, Content: msg.Content})
+		switch content := msg.Content.(type) {
+		case string:
+			messages = append(messages, api.Message{Role: msg.Role, Content: content})
+		case []any:
+			for _, c := range content {
+				data, ok := c.(map[string]any)
+				if !ok {
+					return nil, errors.New("invalid message format")
+				}
+				switch data["type"] {
+				case "text":
+					text, ok := data["text"].(string)
+					if !ok {
+						return nil, errors.New("invalid message format")
+					}
+					messages = append(messages, api.Message{Role: msg.Role, Content: text})
+				case "image_url":
+					var url string
+					if urlMap, ok := data["image_url"].(map[string]any); ok {
+						if url, ok = urlMap["url"].(string); !ok {
+							return nil, errors.New("invalid message format")
+						}
+					} else {
+						if url, ok = data["image_url"].(string); !ok {
+							return nil, errors.New("invalid message format")
+						}
+					}
+
+					types := []string{"jpeg", "jpg", "png"}
+					valid := false
+					for _, t := range types {
+						prefix := "data:image/" + t + ";base64,"
+						if strings.HasPrefix(url, prefix) {
+							url = strings.TrimPrefix(url, prefix)
+							valid = true
+							break
+						}
+					}
+
+					if !valid {
+						return nil, errors.New("invalid image input")
+					}
+
+					img, err := base64.StdEncoding.DecodeString(url)
+					if err != nil {
+						return nil, errors.New("invalid message format")
+					}
+
+					messages = append(messages, api.Message{Role: msg.Role, Images: []api.ImageData{img}})
+				default:
+					return nil, errors.New("invalid message format")
+				}
+			}
+		default:
+			if msg.ToolCalls == nil {
+				return nil, fmt.Errorf("invalid message content type: %T", content)
+			}
+
+			toolCalls := make([]api.ToolCall, len(msg.ToolCalls))
+			for i, tc := range msg.ToolCalls {
+				toolCalls[i].Function.Name = tc.Function.Name
+				err := json.Unmarshal([]byte(tc.Function.Arguments), &toolCalls[i].Function.Arguments)
+				if err != nil {
+					return nil, errors.New("invalid tool call arguments")
+				}
+			}
+			messages = append(messages, api.Message{Role: msg.Role, ToolCalls: toolCalls})
+		}
 	}
 
 	options := make(map[string]interface{})
@@ -156,7 +437,7 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	switch stop := r.Stop.(type) {
 	case string:
 		options["stop"] = []string{stop}
-	case []interface{}:
+	case []any:
 		var stops []string
 		for _, s := range stop {
 			if str, ok := s.(string); ok {
@@ -171,7 +452,7 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	}
 
 	if r.Temperature != nil {
-		options["temperature"] = *r.Temperature * 2.0
+		options["temperature"] = *r.Temperature
 	} else {
 		options["temperature"] = 1.0
 	}
@@ -181,11 +462,11 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	}
 
 	if r.FrequencyPenalty != nil {
-		options["frequency_penalty"] = *r.FrequencyPenalty * 2.0
+		options["frequency_penalty"] = *r.FrequencyPenalty
 	}
 
 	if r.PresencePenalty != nil {
-		options["presence_penalty"] = *r.PresencePenalty * 2.0
+		options["presence_penalty"] = *r.PresencePenalty
 	}
 
 	if r.TopP != nil {
@@ -199,22 +480,98 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 		format = "json"
 	}
 
-	return api.ChatRequest{
+	return &api.ChatRequest{
 		Model:    r.Model,
 		Messages: messages,
 		Format:   format,
 		Options:  options,
 		Stream:   &r.Stream,
-	}
+		Tools:    r.Tools,
+	}, nil
 }
 
-type writer struct {
-	stream bool
-	id     string
+func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
+	options := make(map[string]any)
+
+	switch stop := r.Stop.(type) {
+	case string:
+		options["stop"] = []string{stop}
+	case []any:
+		var stops []string
+		for _, s := range stop {
+			if str, ok := s.(string); ok {
+				stops = append(stops, str)
+			} else {
+				return api.GenerateRequest{}, fmt.Errorf("invalid type for 'stop' field: %T", s)
+			}
+		}
+		options["stop"] = stops
+	}
+
+	if r.MaxTokens != nil {
+		options["num_predict"] = *r.MaxTokens
+	}
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature
+	} else {
+		options["temperature"] = 1.0
+	}
+
+	if r.Seed != nil {
+		options["seed"] = *r.Seed
+	}
+
+	options["frequency_penalty"] = r.FrequencyPenalty
+
+	options["presence_penalty"] = r.PresencePenalty
+
+	if r.TopP != 0.0 {
+		options["top_p"] = r.TopP
+	} else {
+		options["top_p"] = 1.0
+	}
+
+	return api.GenerateRequest{
+		Model:   r.Model,
+		Prompt:  r.Prompt,
+		Options: options,
+		Stream:  &r.Stream,
+		Suffix:  r.Suffix,
+	}, nil
+}
+
+type BaseWriter struct {
 	gin.ResponseWriter
 }
 
-func (w *writer) writeError(code int, data []byte) (int, error) {
+type ChatWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type CompleteWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type ListWriter struct {
+	BaseWriter
+}
+
+type RetrieveWriter struct {
+	BaseWriter
+	model string
+}
+
+type EmbedWriter struct {
+	BaseWriter
+	model string
+}
+
+func (w *BaseWriter) writeError(code int, data []byte) (int, error) {
 	var serr api.StatusError
 	err := json.Unmarshal(data, &serr)
 	if err != nil {
@@ -230,7 +587,7 @@ func (w *writer) writeError(code int, data []byte) (int, error) {
 	return len(data), nil
 }
 
-func (w *writer) writeResponse(data []byte) (int, error) {
+func (w *ChatWriter) writeResponse(data []byte) (int, error) {
 	var chatResponse api.ChatResponse
 	err := json.Unmarshal(data, &chatResponse)
 	if err != nil {
@@ -270,7 +627,7 @@ func (w *writer) writeResponse(data []byte) (int, error) {
 	return len(data), nil
 }
 
-func (w *writer) Write(data []byte) (int, error) {
+func (w *ChatWriter) Write(data []byte) (int, error) {
 	code := w.ResponseWriter.Status()
 	if code != http.StatusOK {
 		return w.writeError(code, data)
@@ -279,7 +636,242 @@ func (w *writer) Write(data []byte) (int, error) {
 	return w.writeResponse(data)
 }
 
-func Middleware() gin.HandlerFunc {
+func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	err := json.Unmarshal(data, &generateResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// completion chunk
+	if w.stream {
+		d, err := json.Marshal(toCompleteChunk(w.id, generateResponse))
+		if err != nil {
+			return 0, err
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if generateResponse.Done {
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toCompletion(w.id, generateResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *CompleteWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *ListWriter) writeResponse(data []byte) (int, error) {
+	var listResponse api.ListResponse
+	err := json.Unmarshal(data, &listResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ListWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
+	var showResponse api.ShowResponse
+	err := json.Unmarshal(data, &showResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// retrieve completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *RetrieveWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *EmbedWriter) writeResponse(data []byte) (int, error) {
+	var embedResponse api.EmbedResponse
+	err := json.Unmarshal(data, &embedResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toEmbeddingList(w.model, embedResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *EmbedWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ListMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		w := &ListWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func RetrieveMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		// response writer
+		w := &RetrieveWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      c.Param("model"),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func CompletionsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req CompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		genReq, err := fromCompleteRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &CompleteWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("cmpl-%d", rand.Intn(999)),
+		}
+
+		c.Writer = w
+		c.Next()
+	}
+}
+
+func EmbeddingsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req EmbedRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if req.Input == "" {
+			req.Input = []string{""}
+		}
+
+		if req.Input == nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		if v, ok := req.Input.([]any); ok && len(v) == 0 {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, "invalid input"))
+			return
+		}
+
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.EmbedRequest{Model: req.Model, Input: req.Input}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &EmbedWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      req.Model,
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func ChatMiddleware() gin.HandlerFunc {
 	return func(c *gin.Context) {
 		var req ChatCompletionRequest
 		err := c.ShouldBindJSON(&req)
@@ -294,17 +886,24 @@ func Middleware() gin.HandlerFunc {
 		}
 
 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(fromRequest(req)); err != nil {
+
+		chatReq, err := fromChatRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
 			return
 		}
 
 		c.Request.Body = io.NopCloser(&b)
 
-		w := &writer{
-			ResponseWriter: c.Writer,
-			stream:         req.Stream,
-			id:             fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
+		w := &ChatWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
 		}
 
 		c.Writer = w
diff --git a/openai/openai_test.go b/openai/openai_test.go
new file mode 100644
index 000000000..eabf5b663
--- /dev/null
+++ b/openai/openai_test.go
@@ -0,0 +1,561 @@
+package openai
+
+import (
+	"bytes"
+	"encoding/base64"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	prefix = `data:image/jpeg;base64,`
+	image  = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+var (
+	False = false
+	True  = true
+)
+
+func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		err := json.Unmarshal(bodyBytes, capturedRequest)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, "failed to unmarshal request")
+		}
+		c.Next()
+	}
+}
+
+func TestChatMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.ChatRequest
+		err  ErrorResponse
+	}
+
+	var capturedRequest *api.ChatRequest
+
+	testCases := []testCase{
+		{
+			name: "chat handler",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with options",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "Hello"}
+				],
+				"stream":            true,
+				"max_tokens":        999,
+				"seed":              123,
+				"stop":              ["\n", "stop"],
+				"temperature":       3.0,
+				"frequency_penalty": 4.0,
+				"presence_penalty":  5.0,
+				"top_p":             6.0,
+				"response_format":   {"type": "json_object"}
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+				},
+				Options: map[string]any{
+					"num_predict":       999.0, // float because JSON doesn't distinguish between float and int
+					"seed":              123.0,
+					"stop":              []any{"\n", "stop"},
+					"temperature":       3.0,
+					"frequency_penalty": 4.0,
+					"presence_penalty":  5.0,
+					"top_p":             6.0,
+				},
+				Format: "json",
+				Stream: &True,
+			},
+		},
+		{
+			name: "chat handler with image content",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{
+						"role": "user",
+						"content": [
+							{
+								"type": "text",
+								"text": "Hello"
+							},
+							{
+								"type": "image_url",
+								"image_url": {
+									"url": "` + prefix + image + `"
+								}
+							}
+						]
+					}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Hello",
+					},
+					{
+						Role: "user",
+						Images: []api.ImageData{
+							func() []byte {
+								img, _ := base64.StdEncoding.DecodeString(image)
+								return img
+							}(),
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+		{
+			name: "chat handler with tools",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": "What's the weather like in Paris Today?"},
+					{"role": "assistant", "tool_calls": [{"id": "id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"Paris, France\", \"format\": \"celsius\"}"}}]}
+				]
+			}`,
+			req: api.ChatRequest{
+				Model: "test-model",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "What's the weather like in Paris Today?",
+					},
+					{
+						Role: "assistant",
+						ToolCalls: []api.ToolCall{
+							{
+								Function: api.ToolCallFunction{
+									Name: "get_current_weather",
+									Arguments: map[string]interface{}{
+										"location": "Paris, France",
+										"format":   "celsius",
+									},
+								},
+							},
+						},
+					},
+				},
+				Options: map[string]any{
+					"temperature": 1.0,
+					"top_p":       1.0,
+				},
+				Stream: &False,
+			},
+		},
+
+		{
+			name: "chat handler error forwarding",
+			body: `{
+				"model": "test-model",
+				"messages": [
+					{"role": "user", "content": 2}
+				]
+			}`,
+			err: ErrorResponse{
+				Error: Error{
+					Message: "invalid message content type: float64",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ChatMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/chat", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/chat", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			defer func() { capturedRequest = nil }()
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+			}
+			if capturedRequest != nil && !reflect.DeepEqual(tc.req, *capturedRequest) {
+				t.Fatal("requests did not match")
+			}
+
+			if !reflect.DeepEqual(tc.err, errResp) {
+				t.Fatal("errors did not match")
+			}
+		})
+	}
+}
+
+func TestCompletionsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.GenerateRequest
+		err  ErrorResponse
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			name: "completions handler",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"temperature": 0.8,
+				"stop": ["\n", "stop"],
+				"suffix": "suffix"
+			}`,
+			req: api.GenerateRequest{
+				Model:  "test-model",
+				Prompt: "Hello",
+				Options: map[string]any{
+					"frequency_penalty": 0.0,
+					"presence_penalty":  0.0,
+					"temperature":       0.8,
+					"top_p":             1.0,
+					"stop":              []any{"\n", "stop"},
+				},
+				Suffix: "suffix",
+				Stream: &False,
+			},
+		},
+		{
+			name: "completions handler error forwarding",
+			body: `{
+				"model": "test-model",
+				"prompt": "Hello",
+				"temperature": null,
+				"stop": [1, 2],
+				"suffix": "suffix"
+			}`,
+			err: ErrorResponse{
+				Error: Error{
+					Message: "invalid type for 'stop' field: float64",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(CompletionsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+			}
+
+			if capturedRequest != nil && !reflect.DeepEqual(tc.req, *capturedRequest) {
+				t.Fatal("requests did not match")
+			}
+
+			if !reflect.DeepEqual(tc.err, errResp) {
+				t.Fatal("errors did not match")
+			}
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestEmbeddingsMiddleware(t *testing.T) {
+	type testCase struct {
+		name string
+		body string
+		req  api.EmbedRequest
+		err  ErrorResponse
+	}
+
+	var capturedRequest *api.EmbedRequest
+
+	testCases := []testCase{
+		{
+			name: "embed handler single input",
+			body: `{
+				"input": "Hello",
+				"model": "test-model"
+			}`,
+			req: api.EmbedRequest{
+				Input: "Hello",
+				Model: "test-model",
+			},
+		},
+		{
+			name: "embed handler batch input",
+			body: `{
+				"input": ["Hello", "World"],
+				"model": "test-model"
+			}`,
+			req: api.EmbedRequest{
+				Input: []any{"Hello", "World"},
+				Model: "test-model",
+			},
+		},
+		{
+			name: "embed handler error forwarding",
+			body: `{
+				"model": "test-model"
+			}`,
+			err: ErrorResponse{
+				Error: Error{
+					Message: "invalid input",
+					Type:    "invalid_request_error",
+				},
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(EmbeddingsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/embed", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/embed", strings.NewReader(tc.body))
+			req.Header.Set("Content-Type", "application/json")
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			var errResp ErrorResponse
+			if resp.Code != http.StatusOK {
+				if err := json.Unmarshal(resp.Body.Bytes(), &errResp); err != nil {
+					t.Fatal(err)
+				}
+			}
+
+			if capturedRequest != nil && !reflect.DeepEqual(tc.req, *capturedRequest) {
+				t.Fatal("requests did not match")
+			}
+
+			if !reflect.DeepEqual(tc.err, errResp) {
+				t.Fatal("errors did not match")
+			}
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestListMiddleware(t *testing.T) {
+	type testCase struct {
+		name     string
+		endpoint func(c *gin.Context)
+		resp     string
+	}
+
+	testCases := []testCase{
+		{
+			name: "list handler",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{
+					Models: []api.ListModelResponse{
+						{
+							Name:       "test-model",
+							ModifiedAt: time.Unix(int64(1686935002), 0).UTC(),
+						},
+					},
+				})
+			},
+			resp: `{
+				"object": "list",
+				"data": [
+					{
+						"id": "test-model",
+						"object": "model",
+						"created": 1686935002,
+						"owned_by": "library"
+					}
+				]
+			}`,
+		},
+		{
+			name: "list handler empty output",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{})
+			},
+			resp: `{
+				"object": "list",
+				"data": null
+			}`,
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+
+	for _, tc := range testCases {
+		router := gin.New()
+		router.Use(ListMiddleware())
+		router.Handle(http.MethodGet, "/api/tags", tc.endpoint)
+		req, _ := http.NewRequest(http.MethodGet, "/api/tags", nil)
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+
+		var expected, actual map[string]any
+		err := json.Unmarshal([]byte(tc.resp), &expected)
+		if err != nil {
+			t.Fatalf("failed to unmarshal expected response: %v", err)
+		}
+
+		err = json.Unmarshal(resp.Body.Bytes(), &actual)
+		if err != nil {
+			t.Fatalf("failed to unmarshal actual response: %v", err)
+		}
+
+		if !reflect.DeepEqual(expected, actual) {
+			t.Errorf("responses did not match\nExpected: %+v\nActual: %+v", expected, actual)
+		}
+	}
+}
+
+func TestRetrieveMiddleware(t *testing.T) {
+	type testCase struct {
+		name     string
+		endpoint func(c *gin.Context)
+		resp     string
+	}
+
+	testCases := []testCase{
+		{
+			name: "retrieve handler",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ShowResponse{
+					ModifiedAt: time.Unix(int64(1686935002), 0).UTC(),
+				})
+			},
+			resp: `{
+				"id":"test-model",
+				"object":"model",
+				"created":1686935002,
+				"owned_by":"library"}
+			`,
+		},
+		{
+			name: "retrieve handler error forwarding",
+			endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "model not found"})
+			},
+			resp: `{
+				"error": {
+				  "code": null,
+				  "message": "model not found",
+				  "param": null,
+				  "type": "api_error"
+				}
+			}`,
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+
+	for _, tc := range testCases {
+		router := gin.New()
+		router.Use(RetrieveMiddleware())
+		router.Handle(http.MethodGet, "/api/show/:model", tc.endpoint)
+		req, _ := http.NewRequest(http.MethodGet, "/api/show/test-model", nil)
+
+		resp := httptest.NewRecorder()
+		router.ServeHTTP(resp, req)
+
+		var expected, actual map[string]any
+		err := json.Unmarshal([]byte(tc.resp), &expected)
+		if err != nil {
+			t.Fatalf("failed to unmarshal expected response: %v", err)
+		}
+
+		err = json.Unmarshal(resp.Body.Bytes(), &actual)
+		if err != nil {
+			t.Fatalf("failed to unmarshal actual response: %v", err)
+		}
+
+		if !reflect.DeepEqual(expected, actual) {
+			t.Errorf("responses did not match\nExpected: %+v\nActual: %+v", expected, actual)
+		}
+	}
+}
diff --git a/parser/parser_test.go b/parser/parser_test.go
index 2b5c4c888..ebd8a7ff5 100644
--- a/parser/parser_test.go
+++ b/parser/parser_test.go
@@ -82,7 +82,7 @@ TEMPLATE """   {{ if .System }}<|start_header_id|>system<|end_header_id|>
 }
 
 func TestParseFileFrom(t *testing.T) {
-	var cases = []struct {
+	cases := []struct {
 		input    string
 		expected []Command
 		err      error
@@ -185,7 +185,7 @@ BADCOMMAND param1 value1
 }
 
 func TestParseFileMessages(t *testing.T) {
-	var cases = []struct {
+	cases := []struct {
 		input    string
 		expected []Command
 		err      error
@@ -276,7 +276,7 @@ MESSAGE system`,
 }
 
 func TestParseFileQuoted(t *testing.T) {
-	var cases = []struct {
+	cases := []struct {
 		multiline string
 		expected  []Command
 		err       error
@@ -430,7 +430,7 @@ TEMPLATE """
 }
 
 func TestParseFileParameters(t *testing.T) {
-	var cases = map[string]struct {
+	cases := map[string]struct {
 		name, value string
 	}{
 		"numa true":                    {"numa", "true"},
@@ -451,6 +451,7 @@ func TestParseFileParameters(t *testing.T) {
 		"num_predict 1":                {"num_predict", "1"},
 		"top_k 1":                      {"top_k", "1"},
 		"top_p 1.0":                    {"top_p", "1.0"},
+		"min_p 0.05":                   {"min_p", "0.05"},
 		"tfs_z 1.0":                    {"tfs_z", "1.0"},
 		"typical_p 1.0":                {"typical_p", "1.0"},
 		"repeat_last_n 1":              {"repeat_last_n", "1"},
@@ -490,7 +491,7 @@ func TestParseFileParameters(t *testing.T) {
 }
 
 func TestParseFileComments(t *testing.T) {
-	var cases = []struct {
+	cases := []struct {
 		input    string
 		expected []Command
 	}{
@@ -515,7 +516,7 @@ FROM foo
 }
 
 func TestParseFileFormatParseFile(t *testing.T) {
-	var cases = []string{
+	cases := []string{
 		`
 FROM foo
 ADAPTER adapter1
diff --git a/progress/bar.go b/progress/bar.go
index 476ea8881..410b6e23f 100644
--- a/progress/bar.go
+++ b/progress/bar.go
@@ -6,8 +6,9 @@ import (
 	"strings"
 	"time"
 
-	"github.com/ollama/ollama/format"
 	"golang.org/x/term"
+
+	"github.com/ollama/ollama/format"
 )
 
 type Bar struct {
diff --git a/progress/spinner.go b/progress/spinner.go
index 02f3f9fb1..e39a45eef 100644
--- a/progress/spinner.go
+++ b/progress/spinner.go
@@ -3,11 +3,12 @@ package progress
 import (
 	"fmt"
 	"strings"
+	"sync/atomic"
 	"time"
 )
 
 type Spinner struct {
-	message      string
+	message      atomic.Value
 	messageWidth int
 
 	parts []string
@@ -21,20 +22,25 @@ type Spinner struct {
 
 func NewSpinner(message string) *Spinner {
 	s := &Spinner{
-		message: message,
 		parts: []string{
 			"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏",
 		},
 		started: time.Now(),
 	}
+	s.SetMessage(message)
 	go s.start()
 	return s
 }
 
+func (s *Spinner) SetMessage(message string) {
+	s.message.Store(message)
+}
+
 func (s *Spinner) String() string {
 	var sb strings.Builder
-	if len(s.message) > 0 {
-		message := strings.TrimSpace(s.message)
+
+	if message, ok := s.message.Load().(string); ok && len(message) > 0 {
+		message := strings.TrimSpace(message)
 		if s.messageWidth > 0 && len(message) > s.messageWidth {
 			message = message[:s.messageWidth]
 		}
diff --git a/readline/buffer.go b/readline/buffer.go
index b7cf9b13f..d91fe0a99 100644
--- a/readline/buffer.go
+++ b/readline/buffer.go
@@ -13,7 +13,7 @@ type Buffer struct {
 	DisplayPos int
 	Pos        int
 	Buf        *arraylist.List
-	//LineHasSpace is an arraylist of bools to keep track of whether a line has a space at the end
+	// LineHasSpace is an arraylist of bools to keep track of whether a line has a space at the end
 	LineHasSpace *arraylist.List
 	Prompt       *Prompt
 	LineWidth    int
@@ -56,13 +56,13 @@ func (b *Buffer) GetLineSpacing(line int) bool {
 
 func (b *Buffer) MoveLeft() {
 	if b.Pos > 0 {
-		//asserts that we retrieve a rune
+		// asserts that we retrieve a rune
 		if e, ok := b.Buf.Get(b.Pos - 1); ok {
 			if r, ok := e.(rune); ok {
 				rLength := runewidth.RuneWidth(r)
 
 				if b.DisplayPos%b.LineWidth == 0 {
-					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+					fmt.Print(CursorUp + CursorBOL + CursorRightN(b.Width))
 					if rLength == 2 {
 						fmt.Print(CursorLeft)
 					}
@@ -74,7 +74,7 @@ func (b *Buffer) MoveLeft() {
 						fmt.Print(CursorLeft)
 					}
 				} else {
-					fmt.Print(cursorLeftN(rLength))
+					fmt.Print(CursorLeftN(rLength))
 				}
 
 				b.Pos -= 1
@@ -115,15 +115,15 @@ func (b *Buffer) MoveRight() {
 				b.DisplayPos += rLength
 
 				if b.DisplayPos%b.LineWidth == 0 {
-					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())))
+					fmt.Print(CursorDown + CursorBOL + CursorRightN(len(b.Prompt.prompt())))
 				} else if (b.DisplayPos-rLength)%b.LineWidth == b.LineWidth-1 && hasSpace {
-					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())+rLength))
+					fmt.Print(CursorDown + CursorBOL + CursorRightN(len(b.Prompt.prompt())+rLength))
 					b.DisplayPos += 1
 				} else if b.LineHasSpace.Size() > 0 && b.DisplayPos%b.LineWidth == b.LineWidth-1 && hasSpace {
-					fmt.Printf(CursorDown + CursorBOL + cursorRightN(len(b.Prompt.prompt())))
+					fmt.Print(CursorDown + CursorBOL + CursorRightN(len(b.Prompt.prompt())))
 					b.DisplayPos += 1
 				} else {
-					fmt.Print(cursorRightN(rLength))
+					fmt.Print(CursorRightN(rLength))
 				}
 			}
 		}
@@ -154,7 +154,7 @@ func (b *Buffer) MoveToStart() {
 				fmt.Print(CursorUp)
 			}
 		}
-		fmt.Printf(CursorBOL + cursorRightN(len(b.Prompt.prompt())))
+		fmt.Print(CursorBOL + CursorRightN(len(b.Prompt.prompt())))
 		b.Pos = 0
 		b.DisplayPos = 0
 	}
@@ -169,9 +169,9 @@ func (b *Buffer) MoveToEnd() {
 				fmt.Print(CursorDown)
 			}
 			remainder := b.DisplaySize() % b.LineWidth
-			fmt.Printf(CursorBOL + cursorRightN(len(b.Prompt.prompt())+remainder))
+			fmt.Print(CursorBOL + CursorRightN(len(b.Prompt.prompt())+remainder))
 		} else {
-			fmt.Print(cursorRightN(b.DisplaySize() - b.DisplayPos))
+			fmt.Print(CursorRightN(b.DisplaySize() - b.DisplayPos))
 		}
 
 		b.Pos = b.Buf.Size()
@@ -286,8 +286,7 @@ func (b *Buffer) drawRemaining() {
 	remLength := runewidth.StringWidth(remainingText)
 
 	if len(currLine) > 0 {
-		fmt.Printf(ClearToEOL + currLine)
-		fmt.Print(cursorLeftN(currLineSpace))
+		fmt.Print(ClearToEOL + currLine + CursorLeftN(currLineSpace))
 	} else {
 		fmt.Print(ClearToEOL)
 	}
@@ -301,9 +300,9 @@ func (b *Buffer) drawRemaining() {
 	}
 
 	if (b.DisplayPos+currLineSpace)%b.LineWidth == 0 && currLine == remainingText {
-		fmt.Print(cursorRightN(currLineSpace))
+		fmt.Print(CursorRightN(currLineSpace))
 		fmt.Printf("\n%s", b.Prompt.AltPrompt)
-		fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width-currLineSpace))
+		fmt.Print(CursorUp + CursorBOL + CursorRightN(b.Width-currLineSpace))
 	}
 
 	// render the other lines
@@ -333,9 +332,7 @@ func (b *Buffer) drawRemaining() {
 			lineLength += runewidth.RuneWidth(c)
 			fmt.Printf("%c", c)
 		}
-		fmt.Print(ClearToEOL)
-		fmt.Print(cursorUpN(totalLines))
-		fmt.Printf(CursorBOL + cursorRightN(b.Width-currLineSpace))
+		fmt.Print(ClearToEOL + CursorUpN(totalLines) + CursorBOL + CursorRightN(b.Width-currLineSpace))
 
 		hasSpace := b.GetLineSpacing(b.DisplayPos / b.LineWidth)
 
@@ -357,8 +354,7 @@ func (b *Buffer) Remove() {
 				if b.DisplayPos%b.LineWidth == 0 {
 					// if the user backspaces over the word boundary, do this magic to clear the line
 					// and move to the end of the previous line
-					fmt.Printf(CursorBOL + ClearToEOL)
-					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+					fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + CursorRightN(b.Width))
 
 					if b.DisplaySize()%b.LineWidth < (b.DisplaySize()-rLength)%b.LineWidth {
 						b.LineHasSpace.Remove(b.DisplayPos/b.LineWidth - 1)
@@ -370,24 +366,23 @@ func (b *Buffer) Remove() {
 					}
 
 					if rLength == 2 {
-						fmt.Print(CursorLeft + "  " + cursorLeftN(2))
+						fmt.Print(CursorLeft + "  " + CursorLeftN(2))
 					} else {
 						fmt.Print(" " + CursorLeft)
 					}
 				} else if (b.DisplayPos-rLength)%b.LineWidth == 0 && hasSpace {
-					fmt.Printf(CursorBOL + ClearToEOL)
-					fmt.Printf(CursorUp + CursorBOL + cursorRightN(b.Width))
+					fmt.Print(CursorBOL + ClearToEOL + CursorUp + CursorBOL + CursorRightN(b.Width))
 
 					if b.Pos == b.Buf.Size() {
 						b.LineHasSpace.Remove(b.DisplayPos/b.LineWidth - 1)
 					}
 					b.DisplayPos -= 1
 				} else {
-					fmt.Print(cursorLeftN(rLength))
+					fmt.Print(CursorLeftN(rLength))
 					for range rLength {
 						fmt.Print(" ")
 					}
-					fmt.Print(cursorLeftN(rLength))
+					fmt.Print(CursorLeftN(rLength))
 				}
 
 				var eraseExtraLine bool
@@ -405,9 +400,9 @@ func (b *Buffer) Remove() {
 					// are trailing characters which go over the line width boundary
 					if eraseExtraLine {
 						remainingLines := (b.DisplaySize() - b.DisplayPos) / b.LineWidth
-						fmt.Printf(cursorDownN(remainingLines+1) + CursorBOL + ClearToEOL)
+						fmt.Print(CursorDownN(remainingLines+1) + CursorBOL + ClearToEOL)
 						place := b.DisplayPos % b.LineWidth
-						fmt.Printf(cursorUpN(remainingLines+1) + cursorRightN(place+len(b.Prompt.prompt())))
+						fmt.Print(CursorUpN(remainingLines+1) + CursorRightN(place+len(b.Prompt.prompt())))
 					}
 				}
 			}
@@ -422,9 +417,9 @@ func (b *Buffer) Delete() {
 		if b.DisplaySize()%b.LineWidth == 0 {
 			if b.DisplayPos != b.DisplaySize() {
 				remainingLines := (b.DisplaySize() - b.DisplayPos) / b.LineWidth
-				fmt.Printf(cursorDownN(remainingLines) + CursorBOL + ClearToEOL)
+				fmt.Print(CursorDownN(remainingLines) + CursorBOL + ClearToEOL)
 				place := b.DisplayPos % b.LineWidth
-				fmt.Printf(cursorUpN(remainingLines) + cursorRightN(place+len(b.Prompt.prompt())))
+				fmt.Print(CursorUpN(remainingLines) + CursorRightN(place+len(b.Prompt.prompt())))
 			}
 		}
 	}
@@ -471,17 +466,17 @@ func (b *Buffer) DeleteWord() {
 }
 
 func (b *Buffer) ClearScreen() {
-	fmt.Printf(ClearScreen + CursorReset + b.Prompt.prompt())
+	fmt.Print(ClearScreen + CursorReset + b.Prompt.prompt())
 	if b.IsEmpty() {
 		ph := b.Prompt.placeholder()
-		fmt.Printf(ColorGrey + ph + cursorLeftN(len(ph)) + ColorDefault)
+		fmt.Print(ColorGrey + ph + CursorLeftN(len(ph)) + ColorDefault)
 	} else {
 		currPos := b.DisplayPos
 		currIndex := b.Pos
 		b.Pos = 0
 		b.DisplayPos = 0
 		b.drawRemaining()
-		fmt.Printf(CursorReset + cursorRightN(len(b.Prompt.prompt())))
+		fmt.Print(CursorReset + CursorRightN(len(b.Prompt.prompt())))
 		if currPos > 0 {
 			targetLine := currPos / b.LineWidth
 			if targetLine > 0 {
@@ -491,10 +486,10 @@ func (b *Buffer) ClearScreen() {
 			}
 			remainder := currPos % b.LineWidth
 			if remainder > 0 {
-				fmt.Print(cursorRightN(remainder))
+				fmt.Print(CursorRightN(remainder))
 			}
 			if currPos%b.LineWidth == 0 {
-				fmt.Printf(CursorBOL + b.Prompt.AltPrompt)
+				fmt.Print(CursorBOL + b.Prompt.AltPrompt)
 			}
 		}
 		b.Pos = currIndex
@@ -513,13 +508,13 @@ func (b *Buffer) Replace(r []rune) {
 
 	b.Buf.Clear()
 
-	fmt.Printf(CursorBOL + ClearToEOL)
+	fmt.Print(CursorBOL + ClearToEOL)
 
 	for range lineNums {
 		fmt.Print(CursorUp + CursorBOL + ClearToEOL)
 	}
 
-	fmt.Printf(CursorBOL + b.Prompt.prompt())
+	fmt.Print(CursorBOL + b.Prompt.prompt())
 
 	for _, c := range r {
 		b.Add(c)
@@ -545,19 +540,3 @@ func (b *Buffer) StringNM(n, m int) string {
 	}
 	return s
 }
-
-func cursorLeftN(n int) string {
-	return fmt.Sprintf(CursorLeftN, n)
-}
-
-func cursorRightN(n int) string {
-	return fmt.Sprintf(CursorRightN, n)
-}
-
-func cursorUpN(n int) string {
-	return fmt.Sprintf(CursorUpN, n)
-}
-
-func cursorDownN(n int) string {
-	return fmt.Sprintf(CursorDownN, n)
-}
diff --git a/readline/errors.go b/readline/errors.go
index 40e40cb77..bb3fbd473 100644
--- a/readline/errors.go
+++ b/readline/errors.go
@@ -4,9 +4,7 @@ import (
 	"errors"
 )
 
-var (
-	ErrInterrupt = errors.New("Interrupt")
-)
+var ErrInterrupt = errors.New("Interrupt")
 
 type InterruptError struct {
 	Line []rune
diff --git a/readline/readline.go b/readline/readline.go
index e90a5e01f..1c14fe103 100644
--- a/readline/readline.go
+++ b/readline/readline.go
@@ -98,7 +98,7 @@ func (i *Instance) Readline() (string, error) {
 		showPlaceholder := !i.Pasting || i.Prompt.UseAlt
 		if buf.IsEmpty() && showPlaceholder {
 			ph := i.Prompt.placeholder()
-			fmt.Printf(ColorGrey + ph + fmt.Sprintf(CursorLeftN, len(ph)) + ColorDefault)
+			fmt.Print(ColorGrey + ph + CursorLeftN(len(ph)) + ColorDefault)
 		}
 
 		r, err := i.Terminal.Read()
diff --git a/readline/term_linux.go b/readline/term_linux.go
index e9ed07451..e9e36da99 100644
--- a/readline/term_linux.go
+++ b/readline/term_linux.go
@@ -7,8 +7,10 @@ import (
 	"unsafe"
 )
 
-const tcgets = 0x5401
-const tcsets = 0x5402
+const (
+	tcgets = 0x5401
+	tcsets = 0x5402
+)
 
 func getTermios(fd uintptr) (*Termios, error) {
 	termios := new(Termios)
diff --git a/readline/types.go b/readline/types.go
index 3b88588f4..e136d9962 100644
--- a/readline/types.go
+++ b/readline/types.go
@@ -1,5 +1,7 @@
 package readline
 
+import "strconv"
+
 const (
 	CharNull      = 0
 	CharLineStart = 1
@@ -41,34 +43,49 @@ const (
 )
 
 const (
-	CursorUp    = "\033[1A"
-	CursorDown  = "\033[1B"
-	CursorRight = "\033[1C"
-	CursorLeft  = "\033[1D"
+	Esc = "\x1b"
 
-	CursorSave    = "\033[s"
-	CursorRestore = "\033[u"
+	CursorSave    = Esc + "[s"
+	CursorRestore = Esc + "[u"
 
-	CursorUpN    = "\033[%dA"
-	CursorDownN  = "\033[%dB"
-	CursorRightN = "\033[%dC"
-	CursorLeftN  = "\033[%dD"
+	CursorEOL  = Esc + "[E"
+	CursorBOL  = Esc + "[1G"
+	CursorHide = Esc + "[?25l"
+	CursorShow = Esc + "[?25h"
 
-	CursorEOL  = "\033[E"
-	CursorBOL  = "\033[1G"
-	CursorHide = "\033[?25l"
-	CursorShow = "\033[?25h"
+	ClearToEOL  = Esc + "[K"
+	ClearLine   = Esc + "[2K"
+	ClearScreen = Esc + "[2J"
+	CursorReset = Esc + "[0;0f"
 
-	ClearToEOL  = "\033[K"
-	ClearLine   = "\033[2K"
-	ClearScreen = "\033[2J"
-	CursorReset = "\033[0;0f"
+	ColorGrey    = Esc + "[38;5;245m"
+	ColorDefault = Esc + "[0m"
 
-	ColorGrey    = "\033[38;5;245m"
-	ColorDefault = "\033[0m"
+	StartBracketedPaste = Esc + "[?2004h"
+	EndBracketedPaste   = Esc + "[?2004l"
+)
 
-	StartBracketedPaste = "\033[?2004h"
-	EndBracketedPaste   = "\033[?2004l"
+func CursorUpN(n int) string {
+	return Esc + "[" + strconv.Itoa(n) + "A"
+}
+
+func CursorDownN(n int) string {
+	return Esc + "[" + strconv.Itoa(n) + "B"
+}
+
+func CursorRightN(n int) string {
+	return Esc + "[" + strconv.Itoa(n) + "C"
+}
+
+func CursorLeftN(n int) string {
+	return Esc + "[" + strconv.Itoa(n) + "D"
+}
+
+var (
+	CursorUp    = CursorUpN(1)
+	CursorDown  = CursorDownN(1)
+	CursorRight = CursorRightN(1)
+	CursorLeft  = CursorLeftN(1)
 )
 
 const (
diff --git a/runners/common.go b/runners/common.go
new file mode 100644
index 000000000..681c397b9
--- /dev/null
+++ b/runners/common.go
@@ -0,0 +1,384 @@
+package runners
+
+import (
+	"compress/gzip"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"runtime"
+	"slices"
+	"strconv"
+	"strings"
+	"sync"
+	"syscall"
+
+	"golang.org/x/sync/errgroup"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/gpu"
+)
+
+const (
+	binGlob = "*/*/*/*"
+)
+
+var (
+	lock       sync.Mutex
+	runnersDir = ""
+)
+
+// Return the location where runners are stored
+// If runners are payloads, this will either extract them
+// or refresh them if any have disappeared due to tmp cleaners
+func Refresh(payloadFS fs.FS) (string, error) {
+	lock.Lock()
+	defer lock.Unlock()
+	var err error
+
+	// Wire up extra logging on our first load
+	if runnersDir == "" {
+		defer func() {
+			var runners []string
+			for v := range GetAvailableServers(runnersDir) {
+				runners = append(runners, v)
+			}
+			slog.Info("Dynamic LLM libraries", "runners", runners)
+			slog.Debug("Override detection logic by setting OLLAMA_LLM_LIBRARY")
+		}()
+	}
+
+	if hasPayloads(payloadFS) {
+		if runnersDir == "" {
+			runnersDir, err = extractRunners(payloadFS)
+		} else {
+			err = refreshRunners(payloadFS, runnersDir)
+		}
+	} else if runnersDir == "" {
+		runnersDir, err = locateRunners()
+	}
+
+	return runnersDir, err
+}
+
+func Cleanup(payloadFS fs.FS) {
+	lock.Lock()
+	defer lock.Unlock()
+	if hasPayloads(payloadFS) && runnersDir != "" {
+		// We want to fully clean up the tmpdir parent of the payloads dir
+		tmpDir := filepath.Clean(filepath.Join(runnersDir, ".."))
+		slog.Debug("cleaning up", "dir", tmpDir)
+		err := os.RemoveAll(tmpDir)
+		if err != nil {
+			slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
+		}
+	}
+}
+
+func locateRunners() (string, error) {
+	exe, err := os.Executable()
+	if err != nil {
+		return "", err
+	}
+
+	cwd, err := os.Getwd()
+	if err != nil {
+		return "", err
+	}
+
+	var paths []string
+	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), envconfig.LibRelativeToExe()), cwd} {
+		paths = append(paths,
+			root,
+			filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH),
+			filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH),
+		)
+	}
+
+	// Try a few variations to improve developer experience when building from source in the local tree
+	for _, path := range paths {
+		candidate := filepath.Join(path, "lib", "ollama", "runners")
+		if _, err := os.Stat(candidate); err == nil {
+			return candidate, nil
+		}
+	}
+	return "", fmt.Errorf("unable to locate runners in any search path %v", paths)
+}
+
+// Return true if we're carying nested payloads for the runners
+func hasPayloads(payloadFS fs.FS) bool {
+	files, err := fs.Glob(payloadFS, binGlob)
+	if err != nil || len(files) == 0 || (len(files) == 1 && strings.Contains(files[0], "placeholder")) {
+		return false
+	}
+	return true
+}
+
+func extractRunners(payloadFS fs.FS) (string, error) {
+	cleanupTmpDirs()
+	tmpDir, err := os.MkdirTemp(envconfig.TmpDir(), "ollama")
+	if err != nil {
+		return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+	}
+	// Track our pid so we can clean up orphaned tmpdirs
+	n := filepath.Join(tmpDir, "ollama.pid")
+	if err := os.WriteFile(n, []byte(strconv.Itoa(os.Getpid())), 0o644); err != nil {
+		slog.Warn("failed to write pid file", "file", n, "error", err)
+	}
+	// We create a distinct subdirectory for payloads within the tmpdir
+	// This will typically look like /tmp/ollama3208993108/runners on linux
+	rDir := filepath.Join(tmpDir, "runners")
+
+	slog.Info("extracting embedded files", "dir", rDir)
+	return rDir, refreshRunners(payloadFS, rDir)
+}
+
+func refreshRunners(payloadFS fs.FS, rDir string) error {
+	// extract or refresh server libraries
+	err := extractFiles(payloadFS, rDir, binGlob)
+	if err != nil {
+		return fmt.Errorf("extract binaries: %v", err)
+	}
+	return nil
+}
+
+// extract extracts the embedded files to the target directory
+func extractFiles(payloadFS fs.FS, targetDir string, glob string) error {
+	files, err := fs.Glob(payloadFS, glob)
+	if err != nil || len(files) == 0 {
+		// Should not happen
+		return fmt.Errorf("extractFiles called without payload present")
+	}
+
+	if err := os.MkdirAll(targetDir, 0o755); err != nil {
+		return fmt.Errorf("extractFiles could not mkdir %s: %v", targetDir, err)
+	}
+
+	g := new(errgroup.Group)
+
+	// $OS/$GOARCH/$RUNNER/$FILE
+	for _, file := range files {
+		filename := file
+
+		runner := filepath.Base(filepath.Dir(filename))
+
+		slog.Debug("extracting", "runner", runner, "payload", filename)
+
+		g.Go(func() error {
+			srcf, err := payloadFS.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer srcf.Close()
+
+			src := io.Reader(srcf)
+			if strings.HasSuffix(filename, ".gz") {
+				src, err = gzip.NewReader(src)
+				if err != nil {
+					return fmt.Errorf("decompress payload %s: %v", filename, err)
+				}
+				filename = strings.TrimSuffix(filename, ".gz")
+			}
+
+			runnerDir := filepath.Join(targetDir, runner)
+			if err := os.MkdirAll(runnerDir, 0o755); err != nil {
+				return fmt.Errorf("extractFiles could not mkdir %s: %v", runnerDir, err)
+			}
+
+			base := filepath.Base(filename)
+			destFilename := filepath.Join(runnerDir, base)
+
+			_, err = os.Stat(destFilename)
+			switch {
+			case errors.Is(err, os.ErrNotExist):
+				destFile, err := os.OpenFile(destFilename, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
+				if err != nil {
+					return fmt.Errorf("write payload %s: %v", filename, err)
+				}
+				defer destFile.Close()
+				if _, err := io.Copy(destFile, src); err != nil {
+					return fmt.Errorf("copy payload %s: %v", filename, err)
+				}
+			case err != nil:
+				return fmt.Errorf("stat payload %s: %v", filename, err)
+			}
+			return nil
+		})
+	}
+
+	err = g.Wait()
+	if err != nil {
+		slog.Error("failed to extract files", "error", err)
+		// If we fail to extract, the payload dir is most likely unusable, so cleanup whatever we extracted
+		err := os.RemoveAll(targetDir)
+		if err != nil {
+			slog.Warn("failed to cleanup incomplete payload dir", "dir", targetDir, "error", err)
+		}
+		return err
+	}
+	return nil
+}
+
+// Best effort to clean up prior tmpdirs
+func cleanupTmpDirs() {
+	tmpDir := envconfig.TmpDir()
+	if tmpDir == "" {
+		tmpDir = os.TempDir()
+	}
+	matches, err := filepath.Glob(filepath.Join(tmpDir, "ollama*", "ollama.pid"))
+	if err != nil {
+		return
+	}
+
+	for _, match := range matches {
+		raw, err := os.ReadFile(match)
+		if errors.Is(err, os.ErrNotExist) {
+			slog.Debug("not a ollama runtime directory, skipping", "path", match)
+			continue
+		} else if err != nil {
+			slog.Warn("could not read ollama.pid, skipping", "path", match, "error", err)
+			continue
+		}
+
+		pid, err := strconv.Atoi(string(raw))
+		if err != nil {
+			slog.Warn("invalid pid, skipping", "path", match, "error", err)
+			continue
+		}
+
+		p, err := os.FindProcess(pid)
+		if err == nil && !errors.Is(p.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+			slog.Warn("process still running, skipping", "pid", pid, "path", match)
+			continue
+		}
+
+		if err := os.Remove(match); err != nil {
+			slog.Warn("could not cleanup stale pidfile", "path", match, "error", err)
+		}
+
+		runners := filepath.Join(filepath.Dir(match), "runners")
+		if err := os.RemoveAll(runners); err != nil {
+			slog.Warn("could not cleanup stale runners", "path", runners, "error", err)
+		}
+
+		if err := os.Remove(filepath.Dir(match)); err != nil {
+			slog.Warn("could not cleanup stale tmpdir", "path", filepath.Dir(match), "error", err)
+		}
+	}
+}
+
+// directory names are the name of the runner and may contain an optional
+// variant prefixed with '_' as the separator. For example, "cuda_v11" and
+// "cuda_v12" or "cpu" and "cpu_avx2". Any library without a variant is the
+// lowest common denominator
+func GetAvailableServers(payloadsDir string) map[string]string {
+	if payloadsDir == "" {
+		slog.Error("empty runner dir")
+		return nil
+	}
+
+	// glob payloadsDir for files that start with ollama_
+	pattern := filepath.Join(payloadsDir, "*", "ollama_*")
+
+	files, err := filepath.Glob(pattern)
+	if err != nil {
+		slog.Debug("could not glob", "pattern", pattern, "error", err)
+		return nil
+	}
+
+	servers := make(map[string]string)
+	for _, file := range files {
+		slog.Debug("availableServers : found", "file", file)
+		servers[filepath.Base(filepath.Dir(file))] = filepath.Dir(file)
+	}
+
+	return servers
+}
+
+// serversForGpu returns a list of compatible servers give the provided GPU
+// info, ordered by performance. assumes Init() has been called
+// TODO - switch to metadata based mapping
+func ServersForGpu(info gpu.GpuInfo) []string {
+	// glob workDir for files that start with ollama_
+	availableServers := GetAvailableServers(runnersDir)
+	requested := info.Library
+	if info.Variant != gpu.CPUCapabilityNone.String() {
+		requested += "_" + info.Variant
+	}
+
+	servers := []string{}
+
+	// exact match first
+	for a := range availableServers {
+		if a == requested {
+			servers = []string{a}
+
+			if a == "metal" {
+				return servers
+			}
+
+			break
+		}
+	}
+
+	alt := []string{}
+
+	// Then for GPUs load alternates and sort the list for consistent load ordering
+	if info.Library != "cpu" {
+		for a := range availableServers {
+			if info.Library == strings.Split(a, "_")[0] && a != requested {
+				alt = append(alt, a)
+			}
+		}
+
+		slices.Sort(alt)
+		servers = append(servers, alt...)
+	}
+
+	if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
+		// Load up the best CPU variant if not primary requested
+		if info.Library != "cpu" {
+			variant := gpu.GetCPUCapability()
+			// If no variant, then we fall back to default
+			// If we have a variant, try that if we find an exact match
+			// Attempting to run the wrong CPU instructions will panic the
+			// process
+			if variant != gpu.CPUCapabilityNone {
+				for cmp := range availableServers {
+					if cmp == "cpu_"+variant.String() {
+						servers = append(servers, cmp)
+						break
+					}
+				}
+			} else {
+				servers = append(servers, "cpu")
+			}
+		}
+
+		if len(servers) == 0 {
+			servers = []string{"cpu"}
+		}
+	}
+
+	return servers
+}
+
+// Return the optimal server for this CPU architecture
+func ServerForCpu() string {
+	if runtime.GOOS == "darwin" && runtime.GOARCH == "arm64" {
+		return "metal"
+	}
+	variant := gpu.GetCPUCapability()
+	availableServers := GetAvailableServers(runnersDir)
+	if variant != gpu.CPUCapabilityNone {
+		for cmp := range availableServers {
+			if cmp == "cpu_"+variant.String() {
+				return cmp
+			}
+		}
+	}
+	return "cpu"
+}
diff --git a/runners/runners_test.go b/runners/runners_test.go
new file mode 100644
index 000000000..e6439448d
--- /dev/null
+++ b/runners/runners_test.go
@@ -0,0 +1,50 @@
+package runners
+
+import (
+	"log/slog"
+	"os"
+	"path"
+	"runtime"
+	"strings"
+	"testing"
+	"testing/fstest"
+)
+
+func TestRefreshRunners(t *testing.T) {
+	slog.SetLogLoggerLevel(slog.LevelDebug)
+
+	payloadFS := fstest.MapFS{
+		path.Join(runtime.GOOS, runtime.GOARCH, "foo", "ollama_llama_server"): {Data: []byte("hello, world\n")},
+	}
+	tmpDir, err := os.MkdirTemp("", "testing")
+	if err != nil {
+		t.Fatalf("failed to make tmp dir %s", err)
+	}
+	t.Setenv("OLLAMA_TMPDIR", tmpDir)
+	rDir, err := Refresh(payloadFS)
+	if err != nil {
+		t.Fatalf("failed to extract to %s %s", tmpDir, err)
+	}
+	if !strings.Contains(rDir, tmpDir) {
+		t.Fatalf("runner dir %s was not in tmp dir %s", rDir, tmpDir)
+	}
+
+	// spot check results
+	servers := GetAvailableServers(rDir)
+	if len(servers) < 1 {
+		t.Fatalf("expected at least 1 server")
+	}
+
+	// Refresh contents
+	rDir, err = extractRunners(payloadFS)
+	if err != nil {
+		t.Fatalf("failed to extract to %s %s", tmpDir, err)
+	}
+	if !strings.Contains(rDir, tmpDir) {
+		t.Fatalf("runner dir %s was not in tmp dir %s", rDir, tmpDir)
+	}
+
+	cleanupTmpDirs()
+
+	Cleanup(payloadFS)
+}
diff --git a/scripts/build_darwin.sh b/scripts/build_darwin.sh
index a2f76af2f..17ac0b943 100755
--- a/scripts/build_darwin.sh
+++ b/scripts/build_darwin.sh
@@ -2,8 +2,7 @@
 
 set -e
 
-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+. $(dirname $0)/env.sh
 
 mkdir -p dist
 
diff --git a/scripts/build_docker.sh b/scripts/build_docker.sh
index e91c56eda..567eb7c7a 100755
--- a/scripts/build_docker.sh
+++ b/scripts/build_docker.sh
@@ -2,76 +2,34 @@
 
 set -eu
 
-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
-
-# We use 2 different image repositories to handle combining architecture images into multiarch manifest
-# (The ROCm image is x86 only and is not a multiarch manifest)
-# For developers, you can override the DOCKER_ORG to generate multiarch manifests
-#  DOCKER_ORG=jdoe PUSH=1 ./scripts/build_docker.sh
-DOCKER_ORG=${DOCKER_ORG:-"ollama"}
-RELEASE_IMAGE_REPO=${RELEASE_IMAGE_REPO:-"${DOCKER_ORG}/release"}
-FINAL_IMAGE_REPO=${FINAL_IMAGE_REPO:-"${DOCKER_ORG}/ollama"}
-
-BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
+. $(dirname $0)/env.sh
 
 # Set PUSH to a non-empty string to trigger push instead of load
 PUSH=${PUSH:-""}
 
-# In CI mode, we break things down
-OLLAMA_SKIP_MANIFEST_CREATE=${OLLAMA_SKIP_MANIFEST_CREATE:-""}
-OLLAMA_SKIP_IMAGE_BUILD=${OLLAMA_SKIP_IMAGE_BUILD:-""}
-
 if [ -z "${PUSH}" ] ; then
+    echo "Building ${FINAL_IMAGE_REPO}:$VERSION locally.  set PUSH=1 to push"
     LOAD_OR_PUSH="--load"
 else
-    echo "Will be pushing ${RELEASE_IMAGE_REPO}:$VERSION for ${BUILD_ARCH}"
+    echo "Will be pushing ${FINAL_IMAGE_REPO}:$VERSION"
     LOAD_OR_PUSH="--push"
 fi
 
-if [ -z "${OLLAMA_SKIP_IMAGE_BUILD}" ]; then
-    for TARGETARCH in ${BUILD_ARCH}; do
-        docker build \
-            ${LOAD_OR_PUSH} \
-            --platform=linux/${TARGETARCH} \
-            --build-arg=VERSION \
-            --build-arg=GOFLAGS \
-            -f Dockerfile \
-            -t ${RELEASE_IMAGE_REPO}:$VERSION-${TARGETARCH} \
-            .
-    done
+docker buildx build \
+    ${LOAD_OR_PUSH} \
+    --platform=${PLATFORM} \
+    ${OLLAMA_COMMON_BUILD_ARGS} \
+    -f Dockerfile \
+    -t ${FINAL_IMAGE_REPO}:$VERSION \
+    .
 
-    if echo ${BUILD_ARCH} | grep "amd64" > /dev/null; then
-        docker build \
-            ${LOAD_OR_PUSH} \
-            --platform=linux/amd64 \
-            --build-arg=VERSION \
-            --build-arg=GOFLAGS \
-            --target runtime-rocm \
-            -f Dockerfile \
-            -t ${RELEASE_IMAGE_REPO}:$VERSION-rocm \
-            .
-    fi
-fi
-
-if [ -z "${OLLAMA_SKIP_MANIFEST_CREATE}" ]; then
-    if [ -n "${PUSH}" ]; then
-        docker manifest create ${FINAL_IMAGE_REPO}:$VERSION \
-            ${RELEASE_IMAGE_REPO}:$VERSION-amd64 \
-            ${RELEASE_IMAGE_REPO}:$VERSION-arm64
-        docker manifest push ${FINAL_IMAGE_REPO}:$VERSION
-
-        # For symmetry, tag/push the rocm image
-        if [ "${RELEASE_IMAGE_REPO}" != "${FINAL_IMAGE_REPO}" ]; then
-            echo "Tagging and pushing rocm image"
-            docker pull ${RELEASE_IMAGE_REPO}:$VERSION-rocm
-            docker tag ${RELEASE_IMAGE_REPO}:$VERSION-rocm ${FINAL_IMAGE_REPO}:$VERSION-rocm
-            docker push ${FINAL_IMAGE_REPO}:$VERSION-rocm
-        fi
-    else
-        echo "Skipping manifest generation when not pushing images are available locally as "
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-amd64"
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-arm64"
-        echo "  ${RELEASE_IMAGE_REPO}:$VERSION-rocm"
-    fi
-fi
+if echo $PLATFORM | grep "amd64" > /dev/null; then
+    docker buildx build \
+        ${LOAD_OR_PUSH} \
+        --platform=linux/amd64 \
+        ${OLLAMA_COMMON_BUILD_ARGS} \
+        --target runtime-rocm \
+        -f Dockerfile \
+        -t ${FINAL_IMAGE_REPO}:$VERSION-rocm \
+        .
+fi
\ No newline at end of file
diff --git a/scripts/build_linux.sh b/scripts/build_linux.sh
index 27c4ff1f8..894d9dd2e 100755
--- a/scripts/build_linux.sh
+++ b/scripts/build_linux.sh
@@ -1,31 +1,29 @@
 #!/bin/sh
+#
+# Mac ARM users, rosetta can be flaky, so to use a remote x86 builder
+#
+# docker context create amd64 --docker host=ssh://mybuildhost
+# docker buildx create --name mybuilder amd64 --platform linux/amd64
+# docker buildx create --name mybuilder --append desktop-linux --platform linux/arm64
+# docker buildx use mybuilder
+
 
 set -eu
 
-export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
-export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+. $(dirname $0)/env.sh
 
-BUILD_ARCH=${BUILD_ARCH:-"amd64 arm64"}
-export AMDGPU_TARGETS=${AMDGPU_TARGETS:=""}
 mkdir -p dist
 
-for TARGETARCH in ${BUILD_ARCH}; do
-    docker build \
-        --platform=linux/$TARGETARCH \
-        --build-arg=GOFLAGS \
-        --build-arg=CGO_CFLAGS \
-        --build-arg=OLLAMA_CUSTOM_CPU_DEFS \
-        --build-arg=AMDGPU_TARGETS \
-        --target build-$TARGETARCH \
+docker buildx build \
+        --output type=local,dest=./dist/ \
+        --platform=${PLATFORM} \
+        ${OLLAMA_COMMON_BUILD_ARGS} \
+        --target dist \
         -f Dockerfile \
-        -t builder:$TARGETARCH \
         .
-    docker create --platform linux/$TARGETARCH --name builder-$TARGETARCH builder:$TARGETARCH
-    docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/ollama ./dist/ollama-linux-$TARGETARCH
 
-    if [ "$TARGETARCH" = "amd64" ]; then
-        docker cp builder-$TARGETARCH:/go/src/github.com/ollama/ollama/dist/deps/ ./dist/
-    fi
-
-    docker rm builder-$TARGETARCH
-done
+# buildx behavior changes for single vs. multiplatform
+if echo $PLATFORM | grep "," > /dev/null ; then 
+        mv -f ./dist/linux_*64/ollama* ./dist/
+        rmdir ./dist/linux_*64
+fi
\ No newline at end of file
diff --git a/scripts/build_windows.ps1 b/scripts/build_windows.ps1
index b3991ce1f..b9508341e 100644
--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@@ -7,34 +7,45 @@
 $ErrorActionPreference = "Stop"
 
 function checkEnv() {
-    $script:TARGET_ARCH=$Env:PROCESSOR_ARCHITECTURE.ToLower()
+    if ($null -ne $env:ARCH ) {
+        $script:ARCH = $env:ARCH
+    } else {
+        $arch=([System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture)
+        if ($null -ne $arch) {
+            $script:ARCH = ($arch.ToString().ToLower()).Replace("x64", "amd64")
+        } else {
+            write-host "WARNING: old powershell detected, assuming amd64 architecture - set `$env:ARCH to override"
+            $script:ARCH="amd64"
+        }
+    }
+    $script:TARGET_ARCH=$script:ARCH
     Write-host "Building for ${script:TARGET_ARCH}"
     write-host "Locating required tools and paths"
     $script:SRC_DIR=$PWD
-    if (!$env:VCToolsRedistDir) {
+    if ($null -eq $env:VCToolsRedistDir) {
         $MSVC_INSTALL=(Get-CimInstance MSFT_VSInstance -Namespace root/cimv2/vs)[0].InstallLocation
         $env:VCToolsRedistDir=(get-item "${MSVC_INSTALL}\VC\Redist\MSVC\*")[0]
     }
-    # Try to find the CUDA dir
-    if ($null -eq $env:NVIDIA_DIR) {
+    # Locate CUDA versions
+    # Note: this assumes every version found will be built
+    $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
+    if ($cudaList.length -eq 0) {
         $d=(get-command -ea 'silentlycontinue' nvcc).path
-        if ($d -ne $null) {
-            $script:NVIDIA_DIR=($d| split-path -parent)
-        } else {
-            $cudaList=(get-item "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin\" -ea 'silentlycontinue')
-            if ($cudaList.length > 0) {
-                $script:NVIDIA_DIR=$cudaList[0]
-            }
+        if ($null -ne $d) {
+            $script:CUDA_DIRS=@($d| split-path -parent)
         }
     } else {
-        $script:NVIDIA_DIR=$env:NVIDIA_DIR
+        $script:CUDA_DIRS=$cudaList
     }
     
-    $script:INNO_SETUP_DIR=(get-item "C:\Program Files*\Inno Setup*\")[0]
+    $inoSetup=(get-item "C:\Program Files*\Inno Setup*\")
+    if ($inoSetup.length -gt 0) {
+        $script:INNO_SETUP_DIR=$inoSetup[0]
+    }
 
-    $script:DEPS_DIR="${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}"
+    $script:DIST_DIR="${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}"
     $env:CGO_ENABLED="1"
-    echo "Checking version"
+    Write-Output "Checking version"
     if (!$env:VERSION) {
         $data=(git describe --tags --first-parent --abbrev=7 --long --dirty --always)
         $pattern="v(.+)"
@@ -69,13 +80,55 @@ function checkEnv() {
 
 
 function buildOllama() {
-    write-host "Building ollama CLI"
     if ($null -eq ${env:OLLAMA_SKIP_GENERATE}) {
-        & go generate ./...
+        Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
+
+        # TODO - consider trying to parallelize this with Start-ThreadJob, but env vars can't be used to toggle
+        #        which targets to build
+
+        # Start by skipping CUDA to build everything else
+        write-host "Building ollama runners"
+        powershell -Command { $env:OLLAMA_SKIP_CUDA_GENERATE="1"; & go generate ./... }
+        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}    
+
+        # Then skip everyhting else and build all the CUDA variants
+        foreach ($env:CUDA_LIB_DIR in $script:CUDA_DIRS) {
+            write-host "Building CUDA ${env:CUDA_LIB_DIR} runner"
+
+            if ($env:CUDA_LIB_DIR.Contains("v12")) {
+                powershell -Command {
+                    $env:OLLAMA_SKIP_CUDA_GENERATE=""
+                    $env:OLLAMA_SKIP_STATIC_GENERATE="1"
+                    $env:OLLAMA_SKIP_CPU_GENERATE="1"
+                    $env:OLLAMA_SKIP_ONEAPI_GENERATE="1"
+                    $env:OLLAMA_SKIP_ROCM_GENERATE="1"
+                    $env:CMAKE_CUDA_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
+                    $env:OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on"
+                    $env:CUDA_PATH=split-path -path $env:CUDA_LIB_DIR -parent
+                    $env:PATH="$envs:CUDA_LIB_DIR;$env:PATH"
+                    & go generate ./...
+                }
+            } else {
+                powershell -Command {
+                    $env:OLLAMA_SKIP_CUDA_GENERATE=""
+                    $env:OLLAMA_SKIP_STATIC_GENERATE="1"
+                    $env:OLLAMA_SKIP_CPU_GENERATE="1"
+                    $env:OLLAMA_SKIP_ONEAPI_GENERATE="1"
+                    $env:OLLAMA_SKIP_ROCM_GENERATE="1"
+                    $env:CMAKE_CUDA_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+                    $env:OLLAMA_CUSTOM_CUDA_DEFS=""
+                    $env:CUDA_PATH=split-path -path $env:CUDA_LIB_DIR -parent
+                    $env:PATH="$envs:CUDA_LIB_DIR;$env:PATH"
+                    & go generate ./...
+                }
+            }
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+        }
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}    
     } else {
         write-host "Skipping generate step with OLLAMA_SKIP_GENERATE set"
     }
+    write-host "Building ollama CLI"
     & go build -trimpath -ldflags "-s -w -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
@@ -91,31 +144,50 @@ function buildApp() {
     write-host "Building Ollama App"
     cd "${script:SRC_DIR}\app"
     & windres -l 0 -o ollama.syso ollama.rc
-    & go build -trimpath -ldflags "-s -w -H windowsgui -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" .
+    & go build -trimpath -ldflags "-s -w -H windowsgui -X=github.com/ollama/ollama/version.Version=$script:VERSION -X=github.com/ollama/ollama/server.mode=release" -o "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}-app.exe" .
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     if ("${env:KEY_CONTAINER}") {
         & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
-            /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} app.exe
+            /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}-app.exe"
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     }
 }
 
 function gatherDependencies() {
-    write-host "Gathering runtime dependencies"
+    if ($null -eq $env:VCToolsRedistDir) {
+        write-error "Unable to locate VC Install location - please use a Developer shell"
+        exit 1
+    }
+    write-host "Gathering runtime dependencies from $env:VCToolsRedistDir"
     cd "${script:SRC_DIR}"
-    md "${script:DEPS_DIR}\ollama_runners" -ea 0 > $null
+    md "${script:DIST_DIR}\lib\ollama" -ea 0 > $null
 
     # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
     # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\ollama_runners\"
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\ollama_runners\"
+    if ($script:TARGET_ARCH -eq "amd64") {
+        $depArch="x64"
+    } else {
+        $depArch=$script:TARGET_ARCH
+    }
+    if ($depArch -eq "amd64") {
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DIST_DIR}\lib\ollama\"
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DIST_DIR}\lib\ollama\"
+        cp "${env:VCToolsRedistDir}\${depArch}\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DIST_DIR}\lib\ollama\"
+        $llvmCrtDir="$env:VCToolsRedistDir\..\..\..\Tools\Llvm\${depArch}\bin"
+        foreach ($part in $("runtime", "stdio", "filesystem", "math", "convert", "heap", "string", "time", "locale", "environment")) {
+            write-host "cp ${llvmCrtDir}\api-ms-win-crt-${part}*.dll ${script:DIST_DIR}\lib\ollama\"
+            cp "${llvmCrtDir}\api-ms-win-crt-${part}*.dll" "${script:DIST_DIR}\lib\ollama\"
+        }
+    } else {
+        # Carying the dll's doesn't seem to work, so use the redist installer
+        copy-item -path "${env:VCToolsRedistDir}\vc_redist.arm64.exe" -destination "${script:DIST_DIR}" -verbose
+    }
 
 
     cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
     if ("${env:KEY_CONTAINER}") {
         write-host "about to sign"
-        foreach ($file in (get-childitem "${script:DEPS_DIR}\cuda\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
+        foreach ($file in (get-childitem "${script:DIST_DIR}\lib\ollama\cu*.dll") + @("${script:SRC_DIR}\dist\ollama_welcome.ps1")){
             write-host "signing $file"
             & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                 /csp "Google Cloud KMS Provider" /kc ${env:KEY_CONTAINER} $file
@@ -125,6 +197,10 @@ function gatherDependencies() {
 }
 
 function buildInstaller() {
+    if ($null -eq ${script:INNO_SETUP_DIR}) {
+        write-host "Inno Setup not present, skipping installer build"
+        return
+    }
     write-host "Building Ollama Installer"
     cd "${script:SRC_DIR}\app"
     $env:PKG_VERSION=$script:PKG_VERSION
@@ -141,13 +217,20 @@ function distZip() {
     Compress-Archive -Path "${script:SRC_DIR}\dist\windows-${script:TARGET_ARCH}\*" -DestinationPath "${script:SRC_DIR}\dist\ollama-windows-${script:TARGET_ARCH}.zip" -Force
 }
 
+checkEnv
 try {
-    checkEnv
-    buildOllama
-    buildApp
-    gatherDependencies
-    buildInstaller
-    distZip
+    if ($($args.count) -eq 0) {
+        buildOllama
+        buildApp
+        gatherDependencies
+        buildInstaller
+        distZip
+    } else {
+        for ( $i = 0; $i -lt $args.count; $i++ ) {
+            write-host "performing $($args[$i])"
+            & $($args[$i])
+        } 
+    }
 } catch {
     write-host "Build Failed"
     write-host $_
diff --git a/scripts/env.sh b/scripts/env.sh
new file mode 100644
index 000000000..d3ca05d7f
--- /dev/null
+++ b/scripts/env.sh
@@ -0,0 +1,14 @@
+# Common environment setup across build*.sh scripts
+
+export VERSION=${VERSION:-$(git describe --tags --first-parent --abbrev=7 --long --dirty --always | sed -e "s/^v//g")}
+export GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=$VERSION\" \"-X=github.com/ollama/ollama/server.mode=release\"'"
+# TODO - consider `docker buildx ls --format=json` to autodiscover platform capability
+PLATFORM=${PLATFORM:-"linux/arm64,linux/amd64"}
+DOCKER_ORG=${DOCKER_ORG:-"ollama"}
+RELEASE_IMAGE_REPO=${RELEASE_IMAGE_REPO:-"${DOCKER_ORG}/release"}
+FINAL_IMAGE_REPO=${FINAL_IMAGE_REPO:-"${DOCKER_ORG}/ollama"}
+OLLAMA_COMMON_BUILD_ARGS="--build-arg=VERSION --build-arg=GOFLAGS --build-arg=OLLAMA_CUSTOM_CPU_DEFS --build-arg=AMDGPU_TARGETS"
+
+echo "Building Ollama"
+echo "VERSION=$VERSION"
+echo "PLATFORM=$PLATFORM"
\ No newline at end of file
diff --git a/scripts/install.sh b/scripts/install.sh
index 2a06c350a..79a7b564c 100644
--- a/scripts/install.sh
+++ b/scripts/install.sh
@@ -38,7 +38,7 @@ IS_WSL2=false
 KERN=$(uname -r)
 case "$KERN" in
     *icrosoft*WSL2 | *icrosoft*wsl2) IS_WSL2=true;;
-    *icrosoft) error "Microsoft WSL1 is not currently supported. Please upgrade to WSL2 with 'wsl --set-version <distro> 2'" ;;
+    *icrosoft) error "Microsoft WSL1 is not currently supported. Please use WSL2 with 'wsl --set-version <distro> 2'" ;;
     *) ;;
 esac
 
@@ -63,16 +63,36 @@ if [ -n "$NEEDS" ]; then
     exit 1
 fi
 
-status "Downloading ollama..."
-curl --fail --show-error --location --progress-bar -o $TEMP_DIR/ollama "https://ollama.com/download/ollama-linux-${ARCH}${VER_PARAM}"
-
 for BINDIR in /usr/local/bin /usr/bin /bin; do
     echo $PATH | grep -q $BINDIR && break || continue
 done
+OLLAMA_INSTALL_DIR=$(dirname ${BINDIR})
 
-status "Installing ollama to $BINDIR..."
+status "Installing ollama to $OLLAMA_INSTALL_DIR"
 $SUDO install -o0 -g0 -m755 -d $BINDIR
-$SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $BINDIR/ollama
+$SUDO install -o0 -g0 -m755 -d "$OLLAMA_INSTALL_DIR"
+if curl -I --silent --fail --location "https://ollama.com/download/ollama-linux-${ARCH}.tgz${VER_PARAM}" >/dev/null ; then
+    status "Downloading Linux ${ARCH} bundle"
+    curl --fail --show-error --location --progress-bar \
+        "https://ollama.com/download/ollama-linux-${ARCH}.tgz${VER_PARAM}" | \
+        $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+    BUNDLE=1
+    if [ "$OLLAMA_INSTALL_DIR/bin/ollama" != "$BINDIR/ollama" ] ; then
+        status "Making ollama accessible in the PATH in $BINDIR"
+        $SUDO ln -sf "$OLLAMA_INSTALL_DIR/ollama" "$BINDIR/ollama"
+    fi
+else
+    status "Downloading Linux ${ARCH} CLI"
+    curl --fail --show-error --location --progress-bar -o "$TEMP_DIR/ollama"\
+    "https://ollama.com/download/ollama-linux-${ARCH}${VER_PARAM}"
+    $SUDO install -o0 -g0 -m755 $TEMP_DIR/ollama $OLLAMA_INSTALL_DIR/ollama
+    BUNDLE=0
+    if [ "$OLLAMA_INSTALL_DIR/ollama" != "$BINDIR/ollama" ] ; then
+        status "Making ollama accessible in the PATH in $BINDIR"
+        $SUDO ln -sf "$OLLAMA_INSTALL_DIR/ollama" "$BINDIR/ollama"
+    fi
+fi
+
 
 install_success() {
     status 'The Ollama API is now available at 127.0.0.1:11434.'
@@ -178,6 +198,16 @@ if ! check_gpu lspci nvidia && ! check_gpu lshw nvidia && ! check_gpu lspci amdg
 fi
 
 if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
+    if [ $BUNDLE -ne 0 ]; then
+        status "Downloading Linux ROCm ${ARCH} bundle"
+        curl --fail --show-error --location --progress-bar \
+            "https://ollama.com/download/ollama-linux-${ARCH}-rocm.tgz${VER_PARAM}" | \
+            $SUDO tar -xzf - -C "$OLLAMA_INSTALL_DIR"
+
+        install_success
+        status "AMD GPU ready."
+        exit 0
+    fi
     # Look for pre-existing ROCm v6 before downloading the dependencies
     for search in "${HIP_PATH:-''}" "${ROCM_PATH:-''}" "/opt/rocm" "/usr/lib64"; do
         if [ -n "${search}" ] && [ -e "${search}/libhipblas.so.2" -o -e "${search}/lib/libhipblas.so.2" ]; then
@@ -198,19 +228,29 @@ if check_gpu lspci amdgpu || check_gpu lshw amdgpu; then
     exit 0
 fi
 
+CUDA_REPO_ERR_MSG="NVIDIA GPU detected, but your OS and Architecture are not supported by NVIDIA.  Please install the CUDA driver manually https://docs.nvidia.com/cuda/cuda-installation-guide-linux/"
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-7-centos-7
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-8-rocky-8
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#rhel-9-rocky-9
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#fedora
 install_cuda_driver_yum() {
     status 'Installing NVIDIA repository...'
+    
     case $PACKAGE_MANAGER in
         yum)
             $SUDO $PACKAGE_MANAGER -y install yum-utils
-            $SUDO $PACKAGE_MANAGER-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
+            if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-$1$2.repo" >/dev/null ; then
+                $SUDO $PACKAGE_MANAGER-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-$1$2.repo
+            else
+                error $CUDA_REPO_ERR_MSG
+            fi
             ;;
         dnf)
-            $SUDO $PACKAGE_MANAGER config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-$1$2.repo
+            if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-$1$2.repo" >/dev/null ; then
+                $SUDO $PACKAGE_MANAGER config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-$1$2.repo
+            else
+                error $CUDA_REPO_ERR_MSG
+            fi
             ;;
     esac
 
@@ -235,7 +275,11 @@ install_cuda_driver_yum() {
 # ref: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#debian
 install_cuda_driver_apt() {
     status 'Installing NVIDIA repository...'
-    curl -fsSL -o $TEMP_DIR/cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m)/cuda-keyring_1.1-1_all.deb
+    if curl -I --silent --fail --location "https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-keyring_1.1-1_all.deb" >/dev/null ; then
+        curl -fsSL -o $TEMP_DIR/cuda-keyring.deb https://developer.download.nvidia.com/compute/cuda/repos/$1$2/$(uname -m | sed -e 's/aarch64/sbsa/')/cuda-keyring_1.1-1_all.deb
+    else
+        error $CUDA_REPO_ERR_MSG
+    fi
 
     case $1 in
         debian)
@@ -312,12 +356,12 @@ if ! lsmod | grep -q nvidia || ! lsmod | grep -q nvidia_uvm; then
 fi
 
 # make sure the NVIDIA modules are loaded on boot with nvidia-persistenced
-if command -v nvidia-persistenced > /dev/null 2>&1; then
+if available nvidia-persistenced; then
     $SUDO touch /etc/modules-load.d/nvidia.conf
     MODULES="nvidia nvidia-uvm"
     for MODULE in $MODULES; do
         if ! grep -qxF "$MODULE" /etc/modules-load.d/nvidia.conf; then
-            echo "$MODULE" | sudo tee -a /etc/modules-load.d/nvidia.conf > /dev/null
+            echo "$MODULE" | $SUDO tee -a /etc/modules-load.d/nvidia.conf > /dev/null
         fi
     done
 fi
diff --git a/scripts/rh_linux_deps.sh b/scripts/rh_linux_deps.sh
index ed60e4304..23f1f650e 100644
--- a/scripts/rh_linux_deps.sh
+++ b/scripts/rh_linux_deps.sh
@@ -3,13 +3,25 @@
 # Script for common Dockerfile dependency installation in redhat linux based images
 
 set -ex
+set -o pipefail
 MACHINE=$(uname -m)
 
 if grep -i "centos" /etc/system-release >/dev/null; then
+    # As of 7/1/2024 mirrorlist.centos.org has been taken offline, so adjust accordingly
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
     # Centos 7 derivatives have too old of a git version to run our generate script
     # uninstall and ignore failures
     yum remove -y git
     yum -y install epel-release centos-release-scl
+
+    # The release packages reinstate the mirrors, undo that again
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
     yum -y install dnf
     if [ "${MACHINE}" = "x86_64" ]; then
         yum -y install https://repo.ius.io/ius-release-el7.rpm
@@ -18,7 +30,7 @@ if grep -i "centos" /etc/system-release >/dev/null; then
         dnf install -y rh-git227-git
         ln -s /opt/rh/rh-git227/root/usr/bin/git /usr/local/bin/git
     fi
-    dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++
+    dnf install -y devtoolset-10-gcc devtoolset-10-gcc-c++ pigz findutils
 elif grep -i "rocky" /etc/system-release >/dev/null; then
     # Temporary workaround until rocky 8 AppStream ships GCC 10.4 (10.3 is incompatible with NVCC)
     cat << EOF > /etc/yum.repos.d/Rocky-Vault.repo
@@ -32,12 +44,22 @@ gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-rockyofficial
 EOF
     dnf install -y git \
         gcc-toolset-10-gcc-10.2.1-8.2.el8 \
-        gcc-toolset-10-gcc-c++-10.2.1-8.2.el8
+        gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 \
+        findutils \
+        pigz
 else
     echo "ERROR Unexpected distro"
     exit 1
 fi
 
+if [ "${MACHINE}" = "x86_64" ] ; then
+    curl -s -L https://github.com/ccache/ccache/releases/download/v4.10.2/ccache-4.10.2-linux-x86_64.tar.xz | tar -Jx -C /tmp --strip-components 1 && \
+    mv /tmp/ccache /usr/local/bin/
+else
+    yum -y install epel-release
+    yum install -y ccache
+fi
+
 if [ -n "${CMAKE_VERSION}" ]; then
     curl -s -L https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-$(uname -m).tar.gz | tar -zx -C /usr --strip-components 1
 fi
diff --git a/scripts/tag_latest.sh b/scripts/tag_latest.sh
index abe42631e..1f56f0360 100755
--- a/scripts/tag_latest.sh
+++ b/scripts/tag_latest.sh
@@ -2,32 +2,12 @@
 
 set -eu
 
-# We use 2 different image repositories to handle combining architecture images into multiarch manifest
-# (The ROCm image is x86 only and is not a multiarch manifest)
 # For developers, you can override the DOCKER_ORG to generate multiarch manifests
-#  DOCKER_ORG=jdoe VERSION=0.1.30 PUSH=1 ./scripts/tag_latest.sh
+#  DOCKER_ORG=jdoe VERSION=0.1.30 ./scripts/tag_latest.sh
 DOCKER_ORG=${DOCKER_ORG:-"ollama"}
-RELEASE_IMAGE_REPO=${RELEASE_IMAGE_REPO:-"${DOCKER_ORG}/release"}
 FINAL_IMAGE_REPO=${FINAL_IMAGE_REPO:-"${DOCKER_ORG}/ollama"}
 
-# Set PUSH to a non-empty string to trigger push instead of load
-PUSH=${PUSH:-""}
-
-echo "Assembling manifest and tagging latest"
-docker manifest rm ${FINAL_IMAGE_REPO}:latest || true
-docker manifest create ${FINAL_IMAGE_REPO}:latest \
-    ${RELEASE_IMAGE_REPO}:$VERSION-amd64 \
-    ${RELEASE_IMAGE_REPO}:$VERSION-arm64
-
-docker pull ${RELEASE_IMAGE_REPO}:$VERSION-rocm
-docker tag ${RELEASE_IMAGE_REPO}:$VERSION-rocm ${FINAL_IMAGE_REPO}:rocm
-
-if [ -n "${PUSH}" ]; then
-    echo "Pushing latest tags up..."
-    docker manifest push ${FINAL_IMAGE_REPO}:latest
-    docker push ${FINAL_IMAGE_REPO}:rocm
-else
-    echo "Not pushing ${FINAL_IMAGE_REPO}:latest and ${FINAL_IMAGE_REPO}:rocm"
-fi
-
-
+echo "Updating ${FINAL_IMAGE_REPO}:latest -> ${FINAL_IMAGE_REPO}:${VERSION}"
+docker buildx imagetools create -t ${FINAL_IMAGE_REPO}:latest ${FINAL_IMAGE_REPO}:${VERSION}
+echo "Updating ${FINAL_IMAGE_REPO}:rocm -> ${FINAL_IMAGE_REPO}:${VERSION}-rocm"
+docker buildx imagetools create -t ${FINAL_IMAGE_REPO}:rocm ${FINAL_IMAGE_REPO}:${VERSION}-rocm
diff --git a/server/auth.go b/server/auth.go
index e92a5b657..dcef5bf9c 100644
--- a/server/auth.go
+++ b/server/auth.go
@@ -67,7 +67,7 @@ func getAuthorizationToken(ctx context.Context, challenge registryChallenge) (st
 
 	headers.Add("Authorization", signature)
 
-	response, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, nil)
+	response, err := makeRequest(ctx, http.MethodGet, redirectURL, headers, nil, &registryOptions{})
 	if err != nil {
 		return "", err
 	}
diff --git a/server/download.go b/server/download.go
index d93cd3b45..a3b531897 100644
--- a/server/download.go
+++ b/server/download.go
@@ -8,6 +8,7 @@ import (
 	"io"
 	"log/slog"
 	"math"
+	"math/rand/v2"
 	"net/http"
 	"net/url"
 	"os"
@@ -27,8 +28,10 @@ import (
 
 const maxRetries = 6
 
-var errMaxRetriesExceeded = errors.New("max retries exceeded")
-var errPartStalled = errors.New("part stalled")
+var (
+	errMaxRetriesExceeded = errors.New("max retries exceeded")
+	errPartStalled        = errors.New("part stalled")
+)
 
 var blobDownloadManager sync.Map
 
@@ -43,23 +46,55 @@ type blobDownload struct {
 
 	context.CancelFunc
 
-	done       bool
+	done       chan struct{}
 	err        error
 	references atomic.Int32
 }
 
 type blobDownloadPart struct {
-	N           int
-	Offset      int64
-	Size        int64
-	Completed   int64
-	lastUpdated time.Time
+	N         int
+	Offset    int64
+	Size      int64
+	Completed atomic.Int64
+
+	lastUpdatedMu sync.Mutex
+	lastUpdated   time.Time
 
 	*blobDownload `json:"-"`
 }
 
+type jsonBlobDownloadPart struct {
+	N         int
+	Offset    int64
+	Size      int64
+	Completed int64
+}
+
+func (p *blobDownloadPart) MarshalJSON() ([]byte, error) {
+	return json.Marshal(jsonBlobDownloadPart{
+		N:         p.N,
+		Offset:    p.Offset,
+		Size:      p.Size,
+		Completed: p.Completed.Load(),
+	})
+}
+
+func (p *blobDownloadPart) UnmarshalJSON(b []byte) error {
+	var j jsonBlobDownloadPart
+	if err := json.Unmarshal(b, &j); err != nil {
+		return err
+	}
+	*p = blobDownloadPart{
+		N:      j.N,
+		Offset: j.Offset,
+		Size:   j.Size,
+	}
+	p.Completed.Store(j.Completed)
+	return nil
+}
+
 const (
-	numDownloadParts          = 64
+	numDownloadParts          = 16
 	minDownloadPartSize int64 = 100 * format.MegaByte
 	maxDownloadPartSize int64 = 1000 * format.MegaByte
 )
@@ -71,7 +106,7 @@ func (p *blobDownloadPart) Name() string {
 }
 
 func (p *blobDownloadPart) StartsAt() int64 {
-	return p.Offset + p.Completed
+	return p.Offset + p.Completed.Load()
 }
 
 func (p *blobDownloadPart) StopsAt() int64 {
@@ -81,7 +116,9 @@ func (p *blobDownloadPart) StopsAt() int64 {
 func (p *blobDownloadPart) Write(b []byte) (n int, err error) {
 	n = len(b)
 	p.blobDownload.Completed.Add(int64(n))
+	p.lastUpdatedMu.Lock()
 	p.lastUpdated = time.Now()
+	p.lastUpdatedMu.Unlock()
 	return n, nil
 }
 
@@ -91,6 +128,8 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 		return err
 	}
 
+	b.done = make(chan struct{})
+
 	for _, partFilePath := range partFilePaths {
 		part, err := b.readPart(partFilePath)
 		if err != nil {
@@ -98,7 +137,7 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 		}
 
 		b.Total += part.Size
-		b.Completed.Add(part.Completed)
+		b.Completed.Add(part.Completed.Load())
 		b.Parts = append(b.Parts, part)
 	}
 
@@ -138,9 +177,36 @@ func (b *blobDownload) Prepare(ctx context.Context, requestURL *url.URL, opts *r
 }
 
 func (b *blobDownload) Run(ctx context.Context, requestURL *url.URL, opts *registryOptions) {
+	defer close(b.done)
 	b.err = b.run(ctx, requestURL, opts)
 }
 
+func newBackoff(maxBackoff time.Duration) func(ctx context.Context) error {
+	var n int
+	return func(ctx context.Context) error {
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+
+		n++
+
+		// n^2 backoff timer is a little smoother than the
+		// common choice of 2^n.
+		d := min(time.Duration(n*n)*10*time.Millisecond, maxBackoff)
+		// Randomize the delay between 0.5-1.5 x msec, in order
+		// to prevent accidental "thundering herd" problems.
+		d = time.Duration(float64(d) * (rand.Float64() + 0.5))
+		t := time.NewTimer(d)
+		defer t.Stop()
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-t.C:
+			return nil
+		}
+	}
+}
+
 func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *registryOptions) error {
 	defer blobDownloadManager.Delete(b.Digest)
 	ctx, b.CancelFunc = context.WithCancel(ctx)
@@ -150,14 +216,61 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 		return err
 	}
 	defer file.Close()
+	setSparse(file)
 
 	_ = file.Truncate(b.Total)
 
+	directURL, err := func() (*url.URL, error) {
+		ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+		defer cancel()
+
+		backoff := newBackoff(10 * time.Second)
+		for {
+			// shallow clone opts to be used in the closure
+			// without affecting the outer opts.
+			newOpts := new(registryOptions)
+			*newOpts = *opts
+
+			newOpts.CheckRedirect = func(req *http.Request, via []*http.Request) error {
+				if len(via) > 10 {
+					return errors.New("maximum redirects exceeded (10) for directURL")
+				}
+
+				// if the hostname is the same, allow the redirect
+				if req.URL.Hostname() == requestURL.Hostname() {
+					return nil
+				}
+
+				// stop at the first redirect that is not
+				// the same hostname as the original
+				// request.
+				return http.ErrUseLastResponse
+			}
+
+			resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, nil, nil, newOpts)
+			if err != nil {
+				slog.Warn("failed to get direct URL; backing off and retrying", "err", err)
+				if err := backoff(ctx); err != nil {
+					return nil, err
+				}
+				continue
+			}
+			defer resp.Body.Close()
+			if resp.StatusCode != http.StatusTemporaryRedirect && resp.StatusCode != http.StatusOK {
+				return nil, fmt.Errorf("unexpected status code %d", resp.StatusCode)
+			}
+			return resp.Location()
+		}
+	}()
+	if err != nil {
+		return err
+	}
+
 	g, inner := errgroup.WithContext(ctx)
 	g.SetLimit(numDownloadParts)
 	for i := range b.Parts {
 		part := b.Parts[i]
-		if part.Completed == part.Size {
+		if part.Completed.Load() == part.Size {
 			continue
 		}
 
@@ -165,7 +278,7 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 			var err error
 			for try := 0; try < maxRetries; try++ {
 				w := io.NewOffsetWriter(file, part.StartsAt())
-				err = b.downloadChunk(inner, requestURL, w, part, opts)
+				err = b.downloadChunk(inner, directURL, w, part)
 				switch {
 				case errors.Is(err, context.Canceled), errors.Is(err, syscall.ENOSPC):
 					// return immediately if the context is canceled or the device is out of space
@@ -206,29 +319,31 @@ func (b *blobDownload) run(ctx context.Context, requestURL *url.URL, opts *regis
 		return err
 	}
 
-	b.done = true
 	return nil
 }
 
-func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart, opts *registryOptions) error {
+func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w io.Writer, part *blobDownloadPart) error {
 	g, ctx := errgroup.WithContext(ctx)
 	g.Go(func() error {
-		headers := make(http.Header)
-		headers.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
-		resp, err := makeRequestWithRetry(ctx, http.MethodGet, requestURL, headers, nil, opts)
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, requestURL.String(), nil)
+		if err != nil {
+			return err
+		}
+		req.Header.Set("Range", fmt.Sprintf("bytes=%d-%d", part.StartsAt(), part.StopsAt()-1))
+		resp, err := http.DefaultClient.Do(req)
 		if err != nil {
 			return err
 		}
 		defer resp.Body.Close()
 
-		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size-part.Completed)
+		n, err := io.CopyN(w, io.TeeReader(resp.Body, part), part.Size-part.Completed.Load())
 		if err != nil && !errors.Is(err, context.Canceled) && !errors.Is(err, io.ErrUnexpectedEOF) {
 			// rollback progress
 			b.Completed.Add(-n)
 			return err
 		}
 
-		part.Completed += n
+		part.Completed.Add(n)
 		if err := b.writePart(part.Name(), part); err != nil {
 			return err
 		}
@@ -242,15 +357,21 @@ func (b *blobDownload) downloadChunk(ctx context.Context, requestURL *url.URL, w
 		for {
 			select {
 			case <-ticker.C:
-				if part.Completed >= part.Size {
+				if part.Completed.Load() >= part.Size {
 					return nil
 				}
 
-				if !part.lastUpdated.IsZero() && time.Since(part.lastUpdated) > 5*time.Second {
+				part.lastUpdatedMu.Lock()
+				lastUpdated := part.lastUpdated
+				part.lastUpdatedMu.Unlock()
+
+				if !lastUpdated.IsZero() && time.Since(lastUpdated) > 5*time.Second {
 					const msg = "%s part %d stalled; retrying. If this persists, press ctrl-c to exit, then 'ollama pull' to find a faster connection."
 					slog.Info(fmt.Sprintf(msg, b.Digest[7:19], part.N))
 					// reset last updated
+					part.lastUpdatedMu.Lock()
 					part.lastUpdated = time.Time{}
+					part.lastUpdatedMu.Unlock()
 					return errPartStalled
 				}
 			case <-ctx.Done():
@@ -315,6 +436,8 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 	ticker := time.NewTicker(60 * time.Millisecond)
 	for {
 		select {
+		case <-b.done:
+			return b.err
 		case <-ticker.C:
 			fn(api.ProgressResponse{
 				Status:    fmt.Sprintf("pulling %s", b.Digest[7:19]),
@@ -322,10 +445,6 @@ func (b *blobDownload) Wait(ctx context.Context, fn func(api.ProgressResponse))
 				Total:     b.Total,
 				Completed: b.Completed.Load(),
 			})
-
-			if b.done || b.err != nil {
-				return b.err
-			}
 		case <-ctx.Done():
 			return ctx.Err()
 		}
diff --git a/server/images.go b/server/images.go
index e949fb18a..b5bf7ad64 100644
--- a/server/images.go
+++ b/server/images.go
@@ -28,16 +28,34 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
 
+var (
+	errCapabilities         = errors.New("does not support")
+	errCapabilityCompletion = errors.New("completion")
+	errCapabilityTools      = errors.New("tools")
+	errCapabilityInsert     = errors.New("insert")
+)
+
+type Capability string
+
+const (
+	CapabilityCompletion = Capability("completion")
+	CapabilityTools      = Capability("tools")
+	CapabilityInsert     = Capability("insert")
+)
+
 type registryOptions struct {
 	Insecure bool
 	Username string
 	Password string
 	Token    string
+
+	CheckRedirect func(req *http.Request, via []*http.Request) error
 }
 
 type Model struct {
@@ -48,16 +66,59 @@ type Model struct {
 	ParentModel    string
 	AdapterPaths   []string
 	ProjectorPaths []string
-	Template       string
 	System         string
 	License        []string
 	Digest         string
 	Options        map[string]interface{}
-	Messages       []Message
+	Messages       []api.Message
+
+	Template *template.Template
 }
 
-func (m *Model) IsEmbedding() bool {
-	return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert")
+// CheckCapabilities checks if the model has the specified capabilities returning an error describing
+// any missing or unknown capabilities
+func (m *Model) CheckCapabilities(caps ...Capability) error {
+	var errs []error
+	for _, cap := range caps {
+		switch cap {
+		case CapabilityCompletion:
+			f, err := os.Open(m.ModelPath)
+			if err != nil {
+				slog.Error("couldn't open model file", "error", err)
+				continue
+			}
+			defer f.Close()
+
+			// TODO(mxyng): decode the GGML into model to avoid doing this multiple times
+			ggml, _, err := llm.DecodeGGML(f, 0)
+			if err != nil {
+				slog.Error("couldn't decode ggml", "error", err)
+				continue
+			}
+
+			if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
+				errs = append(errs, errCapabilityCompletion)
+			}
+		case CapabilityTools:
+			if !slices.Contains(m.Template.Vars(), "tools") {
+				errs = append(errs, errCapabilityTools)
+			}
+		case CapabilityInsert:
+			vars := m.Template.Vars()
+			if !slices.Contains(vars, "suffix") {
+				errs = append(errs, errCapabilityInsert)
+			}
+		default:
+			slog.Error("unknown capability", "capability", cap)
+			return fmt.Errorf("unknown capability: %s", cap)
+		}
+	}
+
+	if err := errors.Join(errs...); err != nil {
+		return fmt.Errorf("%w %w", errCapabilities, errors.Join(errs...))
+	}
+
+	return nil
 }
 
 func (m *Model) String() string {
@@ -82,10 +143,10 @@ func (m *Model) String() string {
 		})
 	}
 
-	if m.Template != "" {
+	if m.Template != nil {
 		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "template",
-			Args: m.Template,
+			Args: m.Template.String(),
 		})
 	}
 
@@ -123,25 +184,13 @@ func (m *Model) String() string {
 	for _, msg := range m.Messages {
 		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "message",
-			Args: fmt.Sprintf("%s %s", msg.Role, msg.Content),
+			Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content),
 		})
 	}
 
 	return modelfile.String()
 }
 
-type Message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
-}
-
-type ManifestV2 struct {
-	SchemaVersion int      `json:"schemaVersion"`
-	MediaType     string   `json:"mediaType"`
-	Config        *Layer   `json:"config"`
-	Layers        []*Layer `json:"layers"`
-}
-
 type ConfigV2 struct {
 	ModelFormat   string   `json:"model_format"`
 	ModelFamily   string   `json:"model_family"`
@@ -160,31 +209,26 @@ type RootFS struct {
 	DiffIDs []string `json:"diff_ids"`
 }
 
-func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
+func GetManifest(mp ModelPath) (*Manifest, string, error) {
 	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return nil, "", err
 	}
 
-	if _, err = os.Stat(fp); err != nil {
-		return nil, "", err
-	}
-
-	var manifest *ManifestV2
-
-	bts, err := os.ReadFile(fp)
+	f, err := os.Open(fp)
 	if err != nil {
-		return nil, "", fmt.Errorf("couldn't open file '%s'", fp)
+		return nil, "", err
 	}
+	defer f.Close()
 
-	shaSum := sha256.Sum256(bts)
-	shaStr := hex.EncodeToString(shaSum[:])
+	sha256sum := sha256.New()
 
-	if err := json.Unmarshal(bts, &manifest); err != nil {
+	var manifest Manifest
+	if err := json.NewDecoder(io.TeeReader(f, sha256sum)).Decode(&manifest); err != nil {
 		return nil, "", err
 	}
 
-	return manifest, shaStr, nil
+	return &manifest, hex.EncodeToString(sha256sum.Sum(nil)), nil
 }
 
 func GetModel(name string) (*Model, error) {
@@ -198,23 +242,24 @@ func GetModel(name string) (*Model, error) {
 		Name:      mp.GetFullTagname(),
 		ShortName: mp.GetShortTagname(),
 		Digest:    digest,
-		Template:  "{{ .Prompt }}",
-		License:   []string{},
+		Template:  template.DefaultTemplate,
 	}
 
-	filename, err := GetBlobsPath(manifest.Config.Digest)
-	if err != nil {
-		return nil, err
-	}
+	if manifest.Config.Digest != "" {
+		filename, err := GetBlobsPath(manifest.Config.Digest)
+		if err != nil {
+			return nil, err
+		}
 
-	configFile, err := os.Open(filename)
-	if err != nil {
-		return nil, err
-	}
-	defer configFile.Close()
+		configFile, err := os.Open(filename)
+		if err != nil {
+			return nil, err
+		}
+		defer configFile.Close()
 
-	if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
-		return nil, err
+		if err := json.NewDecoder(configFile).Decode(&model.Config); err != nil {
+			return nil, err
+		}
 	}
 
 	for _, layer := range manifest.Layers {
@@ -235,13 +280,17 @@ func GetModel(name string) (*Model, error) {
 			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
 			model.ProjectorPaths = append(model.ProjectorPaths, filename)
-		case "application/vnd.ollama.image.template":
+		case "application/vnd.ollama.image.prompt",
+			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
 				return nil, err
 			}
 
-			model.Template = string(bts)
+			model.Template, err = template.Parse(string(bts))
+			if err != nil {
+				return nil, err
+			}
 		case "application/vnd.ollama.image.system":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
@@ -249,13 +298,6 @@ func GetModel(name string) (*Model, error) {
 			}
 
 			model.System = string(bts)
-		case "application/vnd.ollama.image.prompt":
-			bts, err := os.ReadFile(filename)
-			if err != nil {
-				return nil, err
-			}
-
-			model.Template = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@@ -326,14 +368,15 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 	var messages []*api.Message
 	parameters := make(map[string]any)
 
-	var layers []*Layer
+	var layers []Layer
+	var baseLayers []*layerGGML
 	for _, c := range modelfile.Commands {
 		mediatype := fmt.Sprintf("application/vnd.ollama.image.%s", c.Name)
+		command := c.Name
 
-		switch c.Name {
+		switch command {
 		case "model", "adapter":
-			var baseLayers []*layerGGML
-			if name := model.ParseName(c.Args); name.IsValid() {
+			if name := model.ParseName(c.Args); name.IsValid() && command == "model" {
 				baseLayers, err = parseFromModel(ctx, name, fn)
 				if err != nil {
 					return err
@@ -367,14 +410,14 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 				}
 				defer blob.Close()
 
-				baseLayers, err = parseFromFile(ctx, blob, digest, fn)
+				baseLayers, err = parseFromFile(ctx, command, baseLayers, blob, digest, fn)
 				if err != nil {
 					return err
 				}
 			} else if file, err := os.Open(realpath(modelFileDir, c.Args)); err == nil {
 				defer file.Close()
 
-				baseLayers, err = parseFromFile(ctx, file, "", fn)
+				baseLayers, err = parseFromFile(ctx, command, baseLayers, file, "", fn)
 				if err != nil {
 					return err
 				}
@@ -444,9 +487,15 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 				layers = append(layers, baseLayer.Layer)
 			}
 		case "license", "template", "system":
+			if c.Name == "template" {
+				if _, err := template.Parse(c.Args); err != nil {
+					return fmt.Errorf("%w: %s", errBadTemplate, err)
+				}
+			}
+
 			if c.Name != "license" {
 				// replace
-				layers = slices.DeleteFunc(layers, func(layer *Layer) bool {
+				layers = slices.DeleteFunc(layers, func(layer Layer) bool {
 					if layer.MediaType != mediatype {
 						return false
 					}
@@ -492,7 +541,7 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 	}
 
 	var err2 error
-	layers = slices.DeleteFunc(layers, func(layer *Layer) bool {
+	layers = slices.DeleteFunc(layers, func(layer Layer) bool {
 		switch layer.MediaType {
 		case "application/vnd.ollama.image.message":
 			// if there are new messages, remove the inherited ones
@@ -572,12 +621,12 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 		return err
 	}
 
-	layer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
+	configLayer, err := NewLayer(&b, "application/vnd.docker.container.image.v1+json")
 	if err != nil {
 		return err
 	}
 
-	for _, layer := range append(layers, layer) {
+	for _, layer := range append(layers, configLayer) {
 		if layer.status != "" {
 			fn(api.ProgressResponse{Status: layer.status})
 		}
@@ -586,11 +635,11 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 	old, _ := ParseNamedManifest(name)
 
 	fn(api.ProgressResponse{Status: "writing manifest"})
-	if err := WriteManifest(name, layer, layers); err != nil {
+	if err := WriteManifest(name, configLayer, layers); err != nil {
 		return err
 	}
 
-	if !envconfig.NoPrune && old != nil {
+	if !envconfig.NoPrune() && old != nil {
 		if err := old.RemoveLayers(); err != nil {
 			return err
 		}
@@ -639,44 +688,18 @@ func CopyModel(src, dst model.Name) error {
 	return err
 }
 
-func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{}) error {
-	fp, err := GetManifestPath()
+func deleteUnusedLayers(deleteMap map[string]struct{}) error {
+	manifests, err := Manifests()
 	if err != nil {
 		return err
 	}
 
-	walkFunc := func(path string, info os.FileInfo, _ error) error {
-		if info.IsDir() {
-			return nil
-		}
-
-		dir, file := filepath.Split(path)
-		dir = strings.Trim(strings.TrimPrefix(dir, fp), string(os.PathSeparator))
-		tag := strings.Join([]string{dir, file}, ":")
-		fmp := ParseModelPath(tag)
-
-		// skip the manifest we're trying to delete
-		if skipModelPath != nil && skipModelPath.GetFullTagname() == fmp.GetFullTagname() {
-			return nil
-		}
-
-		// save (i.e. delete from the deleteMap) any files used in other manifests
-		manifest, _, err := GetManifest(fmp)
-		if err != nil {
-			//nolint:nilerr
-			return nil
-		}
-
+	for _, manifest := range manifests {
 		for _, layer := range manifest.Layers {
 			delete(deleteMap, layer.Digest)
 		}
 
 		delete(deleteMap, manifest.Config.Digest)
-		return nil
-	}
-
-	if err := filepath.Walk(fp, walkFunc); err != nil {
-		return err
 	}
 
 	// only delete the files which are still in the deleteMap
@@ -729,9 +752,9 @@ func PruneLayers() error {
 
 	slog.Info(fmt.Sprintf("total blobs: %d", len(deleteMap)))
 
-	err = deleteUnusedLayers(nil, deleteMap)
-	if err != nil {
-		return err
+	if err := deleteUnusedLayers(deleteMap); err != nil {
+		slog.Error(fmt.Sprintf("couldn't remove unused layers: %v", err))
+		return nil
 	}
 
 	slog.Info(fmt.Sprintf("total unused blobs removed: %d", len(deleteMap)))
@@ -777,7 +800,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	fn(api.ProgressResponse{Status: "retrieving manifest"})
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return fmt.Errorf("insecure protocol http")
+		return errors.New("insecure protocol http")
 	}
 
 	manifest, _, err := GetManifest(mp)
@@ -786,9 +809,11 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		return err
 	}
 
-	var layers []*Layer
+	var layers []Layer
 	layers = append(layers, manifest.Layers...)
-	layers = append(layers, manifest.Config)
+	if manifest.Config.Digest != "" {
+		layers = append(layers, manifest.Config)
+	}
 
 	for _, layer := range layers {
 		if err := uploadBlob(ctx, mp, layer, regOpts, fn); err != nil {
@@ -822,29 +847,24 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	mp := ParseModelPath(name)
 
-	var manifest *ManifestV2
-	var err error
-	var noprune string
-
 	// build deleteMap to prune unused layers
 	deleteMap := make(map[string]struct{})
-
-	if !envconfig.NoPrune {
-		manifest, _, err = GetManifest(mp)
-		if err != nil && !errors.Is(err, os.ErrNotExist) {
-			return err
+	manifest, _, err := GetManifest(mp)
+	if errors.Is(err, os.ErrNotExist) {
+		// noop
+	} else if err != nil && !errors.Is(err, os.ErrNotExist) {
+		return err
+	} else {
+		for _, l := range manifest.Layers {
+			deleteMap[l.Digest] = struct{}{}
 		}
-
-		if manifest != nil {
-			for _, l := range manifest.Layers {
-				deleteMap[l.Digest] = struct{}{}
-			}
+		if manifest.Config.Digest != "" {
 			deleteMap[manifest.Config.Digest] = struct{}{}
 		}
 	}
 
 	if mp.ProtocolScheme == "http" && !regOpts.Insecure {
-		return fmt.Errorf("insecure protocol http")
+		return errors.New("insecure protocol http")
 	}
 
 	fn(api.ProgressResponse{Status: "pulling manifest"})
@@ -854,9 +874,11 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		return fmt.Errorf("pull model manifest: %s", err)
 	}
 
-	var layers []*Layer
+	var layers []Layer
 	layers = append(layers, manifest.Layers...)
-	layers = append(layers, manifest.Config)
+	if manifest.Config.Digest != "" {
+		layers = append(layers, manifest.Config)
+	}
 
 	skipVerify := make(map[string]bool)
 	for _, layer := range layers {
@@ -916,11 +938,10 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 		return err
 	}
 
-	if noprune == "" {
-		fn(api.ProgressResponse{Status: "removing any unused layers"})
-		err = deleteUnusedLayers(nil, deleteMap)
-		if err != nil {
-			return err
+	if !envconfig.NoPrune() && len(deleteMap) > 0 {
+		fn(api.ProgressResponse{Status: "removing unused layers"})
+		if err := deleteUnusedLayers(deleteMap); err != nil {
+			fn(api.ProgressResponse{Status: fmt.Sprintf("couldn't remove unused layers: %v", err)})
 		}
 	}
 
@@ -929,7 +950,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	return nil
 }
 
-func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*ManifestV2, error) {
+func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
 	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)
 
 	headers := make(http.Header)
@@ -940,12 +961,12 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()
 
-	var m *ManifestV2
+	var m Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}
 
-	return m, err
+	return &m, err
 }
 
 // GetSHA256Digest returns the SHA256 hash of a given buffer and returns it, and the size of buffer
@@ -959,7 +980,7 @@ func GetSHA256Digest(r io.Reader) (string, int64) {
 	return fmt.Sprintf("sha256:%x", h.Sum(nil)), n
 }
 
-var errUnauthorized = fmt.Errorf("unauthorized: access denied")
+var errUnauthorized = errors.New("unauthorized: access denied")
 
 // getTokenSubject returns the subject of a JWT token, it does not validate the token
 func getTokenSubject(token string) string {
@@ -1077,7 +1098,9 @@ func makeRequest(ctx context.Context, method string, requestURL *url.URL, header
 		req.ContentLength = contentLength
 	}
 
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := (&http.Client{
+		CheckRedirect: regOpts.CheckRedirect,
+	}).Do(req)
 	if err != nil {
 		return nil, err
 	}
diff --git a/server/layer.go b/server/layer.go
index cc6709d24..0bdee72b3 100644
--- a/server/layer.go
+++ b/server/layer.go
@@ -2,6 +2,7 @@ package server
 
 import (
 	"crypto/sha256"
+	"errors"
 	"fmt"
 	"io"
 	"os"
@@ -15,15 +16,15 @@ type Layer struct {
 	status    string
 }
 
-func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
+func NewLayer(r io.Reader, mediatype string) (Layer, error) {
 	blobs, err := GetBlobsPath("")
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
 	temp, err := os.CreateTemp(blobs, "sha256-")
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 	defer temp.Close()
 	defer os.Remove(temp.Name())
@@ -31,28 +32,31 @@ func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
 	sha256sum := sha256.New()
 	n, err := io.Copy(io.MultiWriter(temp, sha256sum), r)
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
 	if err := temp.Close(); err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
 	digest := fmt.Sprintf("sha256:%x", sha256sum.Sum(nil))
 	blob, err := GetBlobsPath(digest)
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
 	status := "using existing layer"
 	if _, err := os.Stat(blob); err != nil {
 		status = "creating new layer"
 		if err := os.Rename(temp.Name(), blob); err != nil {
-			return nil, err
+			return Layer{}, err
+		}
+		if err := os.Chmod(blob, 0o644); err != nil {
+			return Layer{}, err
 		}
 	}
 
-	return &Layer{
+	return Layer{
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      n,
@@ -60,18 +64,22 @@ func NewLayer(r io.Reader, mediatype string) (*Layer, error) {
 	}, nil
 }
 
-func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
+func NewLayerFromLayer(digest, mediatype, from string) (Layer, error) {
+	if digest == "" {
+		return Layer{}, errors.New("creating new layer from layer with empty digest")
+	}
+
 	blob, err := GetBlobsPath(digest)
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
 	fi, err := os.Stat(blob)
 	if err != nil {
-		return nil, err
+		return Layer{}, err
 	}
 
-	return &Layer{
+	return Layer{
 		MediaType: mediatype,
 		Digest:    digest,
 		Size:      fi.Size(),
@@ -81,6 +89,10 @@ func NewLayerFromLayer(digest, mediatype, from string) (*Layer, error) {
 }
 
 func (l *Layer) Open() (io.ReadSeekCloser, error) {
+	if l.Digest == "" {
+		return nil, errors.New("opening layer with empty digest")
+	}
+
 	blob, err := GetBlobsPath(l.Digest)
 	if err != nil {
 		return nil, err
@@ -90,6 +102,10 @@ func (l *Layer) Open() (io.ReadSeekCloser, error) {
 }
 
 func (l *Layer) Remove() error {
+	if l.Digest == "" {
+		return nil
+	}
+
 	ms, err := Manifests()
 	if err != nil {
 		return err
diff --git a/server/manifest.go b/server/manifest.go
index 61dd1ab4e..6b04753ff 100644
--- a/server/manifest.go
+++ b/server/manifest.go
@@ -2,6 +2,7 @@ package server
 
 import (
 	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -14,7 +15,10 @@ import (
 )
 
 type Manifest struct {
-	ManifestV2
+	SchemaVersion int     `json:"schemaVersion"`
+	MediaType     string  `json:"mediaType"`
+	Config        Layer   `json:"config"`
+	Layers        []Layer `json:"layers"`
 
 	filepath string
 	fi       os.FileInfo
@@ -44,10 +48,12 @@ func (m *Manifest) Remove() error {
 
 func (m *Manifest) RemoveLayers() error {
 	for _, layer := range append(m.Layers, m.Config) {
-		if err := layer.Remove(); errors.Is(err, os.ErrNotExist) {
-			slog.Debug("layer does not exist", "digest", layer.Digest)
-		} else if err != nil {
-			return err
+		if layer.Digest != "" {
+			if err := layer.Remove(); errors.Is(err, os.ErrNotExist) {
+				slog.Debug("layer does not exist", "digest", layer.Digest)
+			} else if err != nil {
+				return err
+			}
 		}
 	}
 
@@ -66,7 +72,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 
 	p := filepath.Join(manifests, n.Filepath())
 
-	var m ManifestV2
+	var m Manifest
 	f, err := os.Open(p)
 	if err != nil {
 		return nil, err
@@ -83,15 +89,14 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, err
 	}
 
-	return &Manifest{
-		ManifestV2: m,
-		filepath:   p,
-		fi:         fi,
-		digest:     fmt.Sprintf("%x", sha256sum.Sum(nil)),
-	}, nil
+	m.filepath = p
+	m.fi = fi
+	m.digest = hex.EncodeToString(sha256sum.Sum(nil))
+
+	return &m, nil
 }
 
-func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
+func WriteManifest(name model.Name, config Layer, layers []Layer) error {
 	manifests, err := GetManifestPath()
 	if err != nil {
 		return err
@@ -108,7 +113,7 @@ func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
 	}
 	defer f.Close()
 
-	m := ManifestV2{
+	m := Manifest{
 		SchemaVersion: 2,
 		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
 		Config:        config,
@@ -146,14 +151,16 @@ func Manifests() (map[model.Name]*Manifest, error) {
 
 			n := model.ParseNameFromFilepath(rel)
 			if !n.IsValid() {
-				slog.Warn("bad manifest name", "path", rel, "error", err)
+				slog.Warn("bad manifest name", "path", rel)
 				continue
 			}
 
 			m, err := ParseNamedManifest(n)
-			if err != nil {
+			if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
 				slog.Warn("bad manifest", "name", n, "error", err)
 				continue
+			} else if err != nil {
+				return nil, fmt.Errorf("%s: %w", n, err)
 			}
 
 			ms[n] = m
diff --git a/server/manifest_test.go b/server/manifest_test.go
index ceee31d88..70ab7fa2f 100644
--- a/server/manifest_test.go
+++ b/server/manifest_test.go
@@ -7,7 +7,6 @@ import (
 	"slices"
 	"testing"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -15,7 +14,7 @@ func createManifest(t *testing.T, path, name string) {
 	t.Helper()
 
 	p := filepath.Join(path, "manifests", name)
-	if err := os.MkdirAll(filepath.Dir(p), 0755); err != nil {
+	if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
 		t.Fatal(err)
 	}
 
@@ -25,7 +24,7 @@ func createManifest(t *testing.T, path, name string) {
 	}
 	defer f.Close()
 
-	if err := json.NewEncoder(f).Encode(ManifestV2{}); err != nil {
+	if err := json.NewEncoder(f).Encode(Manifest{}); err != nil {
 		t.Fatal(err)
 	}
 }
@@ -108,7 +107,6 @@ func TestManifests(t *testing.T) {
 		t.Run(n, func(t *testing.T) {
 			d := t.TempDir()
 			t.Setenv("OLLAMA_MODELS", d)
-			envconfig.LoadConfig()
 
 			for _, p := range wants.ps {
 				createManifest(t, d, p)
diff --git a/server/model.go b/server/model.go
index d56e641ba..124693d3b 100644
--- a/server/model.go
+++ b/server/model.go
@@ -4,6 +4,7 @@ import (
 	"archive/zip"
 	"bytes"
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -11,19 +12,21 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
+	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/templates"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )
 
 var intermediateBlobs map[string]string = make(map[string]string)
 
 type layerGGML struct {
-	*Layer
+	Layer
 	*llm.GGML
 }
 
@@ -78,113 +81,65 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe
 	return layers, nil
 }
 
-func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) error {
-	stat, err := file.Stat()
-	if err != nil {
-		return err
-	}
-
-	r, err := zip.NewReader(file, stat.Size())
-	if err != nil {
-		return err
-	}
-
-	fn(api.ProgressResponse{Status: "unpacking model metadata"})
-	for _, f := range r.File {
-		n := filepath.Join(p, f.Name)
-		if !strings.HasPrefix(n, p) {
-			slog.Warn("skipped extracting file outside of context", "name", f.Name)
-			continue
-		}
-
-		if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil {
-			return err
-		}
-
-		// TODO(mxyng): this should not write out all files to disk
-		outfile, err := os.Create(n)
-		if err != nil {
-			return err
-		}
-		defer outfile.Close()
-
-		infile, err := f.Open()
-		if err != nil {
-			return err
-		}
-		defer infile.Close()
-
-		if _, err = io.Copy(outfile, infile); err != nil {
-			return err
-		}
-
-		if err := outfile.Close(); err != nil {
-			return err
-		}
-
-		if err := infile.Close(); err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
-
-func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
-	tempDir, err := os.MkdirTemp(filepath.Dir(file.Name()), "")
-	if err != nil {
-		return nil, err
-	}
-	defer os.RemoveAll(tempDir)
-
-	if err := extractFromZipFile(tempDir, file, fn); err != nil {
-		return nil, err
-	}
-
-	mf, err := convert.GetModelFormat(tempDir)
+func parseFromZipFile(_ context.Context, command string, baseLayers []*layerGGML, f *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+	fi, err := f.Stat()
 	if err != nil {
 		return nil, err
 	}
 
-	params, err := mf.GetParams(tempDir)
+	r, err := zip.NewReader(f, fi.Size())
 	if err != nil {
 		return nil, err
 	}
 
-	mArch, err := mf.GetModelArch("", tempDir, params)
+	p, err := os.MkdirTemp(filepath.Dir(f.Name()), "")
 	if err != nil {
 		return nil, err
 	}
-
-	fn(api.ProgressResponse{Status: "processing tensors"})
-	if err := mArch.GetTensors(); err != nil {
-		return nil, err
-	}
-
-	if err := mArch.LoadVocab(); err != nil {
-		return nil, err
-	}
+	defer os.RemoveAll(p)
 
 	fn(api.ProgressResponse{Status: "converting model"})
-
 	// TODO(mxyng): this should write directly into a layer
 	// e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model")
-	temp, err := os.CreateTemp(tempDir, "fp16")
+	t, err := os.CreateTemp(p, "fp16")
 	if err != nil {
 		return nil, err
 	}
-	defer temp.Close()
-	defer os.Remove(temp.Name())
+	defer t.Close()
+	defer os.Remove(t.Name())
 
-	if err = mArch.WriteGGUF(temp); err != nil {
+	var layerType string
+
+	switch command {
+	case "adapter":
+		var baseModel *llm.GGML
+		for _, l := range baseLayers {
+			if l.GGML != nil {
+				baseModel = l.GGML
+				break
+			}
+		}
+
+		if baseModel == nil {
+			return nil, fmt.Errorf("no base model specified for the adapter")
+		}
+
+		if err := convert.ConvertAdapter(convert.NewZipReader(r, p, 32<<20), t, baseModel.KV()); err != nil {
+			return nil, err
+		}
+		layerType = "application/vnd.ollama.image.adapter"
+	case "model":
+		if err := convert.ConvertModel(convert.NewZipReader(r, p, 32<<20), t); err != nil {
+			return nil, err
+		}
+		layerType = "application/vnd.ollama.image.model"
+	}
+
+	if _, err := t.Seek(0, io.SeekStart); err != nil {
 		return nil, err
 	}
 
-	if _, err := temp.Seek(0, io.SeekStart); err != nil {
-		return nil, err
-	}
-
-	layer, err := NewLayer(temp, "application/vnd.ollama.image.model")
+	layer, err := NewLayer(t, layerType)
 	if err != nil {
 		return nil, err
 	}
@@ -206,7 +161,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a
 	return detectChatTemplate(layers)
 }
 
-func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
+func parseFromFile(ctx context.Context, command string, baseLayers []*layerGGML, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) {
 	sr := io.NewSectionReader(file, 0, 512)
 	contentType, err := detectContentType(sr)
 	if err != nil {
@@ -217,7 +172,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 	case "gguf", "ggla":
 		// noop
 	case "application/zip":
-		return parseFromZipFile(ctx, file, digest, fn)
+		return parseFromZipFile(ctx, command, baseLayers, file, digest, fn)
 	default:
 		return nil, fmt.Errorf("unsupported content type: %s", contentType)
 	}
@@ -237,15 +192,26 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 		}
 
 		mediatype := "application/vnd.ollama.image.model"
-		if ggml.Name() == "ggla" {
+		if ggml.Name() == "ggla" || ggml.KV().Kind() == "adapter" {
 			mediatype = "application/vnd.ollama.image.adapter"
 		} else if ggml.KV().Architecture() == "clip" {
 			mediatype = "application/vnd.ollama.image.projector"
 		}
 
-		layer, err := NewLayer(io.NewSectionReader(file, offset, n), mediatype)
-		if err != nil {
-			return nil, err
+		var layer Layer
+		if digest != "" && n == stat.Size() && offset == 0 {
+			layer, err = NewLayerFromLayer(digest, mediatype, file.Name())
+			if err != nil {
+				slog.Debug("could not create new layer from layer", "error", err)
+			}
+		}
+
+		// Fallback to creating layer from file copy (either NewLayerFromLayer failed, or digest empty/n != stat.Size())
+		if layer.Digest == "" {
+			layer, err = NewLayer(io.NewSectionReader(file, offset, n), mediatype)
+			if err != nil {
+				return nil, err
+			}
 		}
 
 		layers = append(layers, &layerGGML{layer, ggml})
@@ -258,16 +224,30 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 	for _, layer := range layers {
 		if s := layer.GGML.KV().ChatTemplate(); s != "" {
-			if t, err := templates.NamedTemplate(s); err != nil {
+			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err)
 			} else {
-				tmpl, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
+				layer, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
 				if err != nil {
 					return nil, err
 				}
 
-				tmpl.status = fmt.Sprintf("using autodetected template %s", t.Name)
-				layers = append(layers, &layerGGML{tmpl, nil})
+				layer.status = fmt.Sprintf("using autodetected template %s", t.Name)
+				layers = append(layers, &layerGGML{layer, nil})
+
+				if t.Parameters != nil {
+					var b bytes.Buffer
+					if err := json.NewEncoder(&b).Encode(t.Parameters); err != nil {
+						return nil, err
+					}
+
+					layer, err := NewLayer(&b, "application/vnd.ollama.image.params")
+					if err != nil {
+						return nil, err
+					}
+
+					layers = append(layers, &layerGGML{layer, nil})
+				}
 			}
 		}
 	}
@@ -291,3 +271,124 @@ func detectContentType(r io.Reader) (string, error) {
 
 	return "unknown", nil
 }
+
+func parseObjects(s string) []map[string]any {
+	var objs []map[string]any
+	for offset := 0; offset < len(s); {
+		var obj map[string]any
+		decoder := json.NewDecoder(strings.NewReader(s[offset:]))
+		if err := decoder.Decode(&obj); errors.Is(err, io.EOF) || errors.Is(err, io.ErrUnexpectedEOF) {
+			break
+		} else if syntax := &(json.SyntaxError{}); errors.As(err, &syntax) {
+			// skip over any syntax errors
+			offset += int(syntax.Offset)
+		} else if unmarshalType := &(json.UnmarshalTypeError{}); errors.As(err, &unmarshalType) {
+			// skip over any unmarshalable types
+			offset += int(unmarshalType.Offset)
+		} else if err != nil {
+			return nil
+		} else {
+			offset += int(decoder.InputOffset())
+			objs = append(objs, obj)
+		}
+	}
+
+	return objs
+}
+
+// parseToolCalls attempts to parse a JSON string into a slice of ToolCalls.
+// mxyng: this only really works if the input contains tool calls in some JSON format
+func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
+	// create a subtree from the node that ranges over .ToolCalls
+	tmpl := m.Template.Subtree(func(n parse.Node) bool {
+		if t, ok := n.(*parse.RangeNode); ok {
+			return slices.Contains(template.Identifiers(t.Pipe), "ToolCalls")
+		}
+
+		return false
+	})
+
+	if tmpl == nil {
+		return nil, false
+	}
+
+	var b bytes.Buffer
+	if err := tmpl.Execute(&b, map[string][]api.ToolCall{
+		"ToolCalls": {
+			{
+				Function: api.ToolCallFunction{
+					Name: "@@name@@",
+					Arguments: api.ToolCallFunctionArguments{
+						"@@argument@@": 1,
+					},
+				},
+			},
+		},
+	}); err != nil {
+		return nil, false
+	}
+
+	templateObjects := parseObjects(b.String())
+	if len(templateObjects) == 0 {
+		return nil, false
+	}
+
+	// find the keys that correspond to the name and arguments fields
+	var name, arguments string
+	for k, v := range templateObjects[0] {
+		switch v.(type) {
+		case string:
+			name = k
+		case map[string]any:
+			arguments = k
+		}
+	}
+
+	if name == "" || arguments == "" {
+		return nil, false
+	}
+
+	responseObjects := parseObjects(s)
+	if len(responseObjects) == 0 {
+		return nil, false
+	}
+
+	// collect all nested objects
+	var collect func(any) []map[string]any
+	collect = func(obj any) (all []map[string]any) {
+		switch o := obj.(type) {
+		case map[string]any:
+			all = append(all, o)
+			for _, v := range o {
+				all = append(all, collect(v)...)
+			}
+		case []any:
+			for _, v := range o {
+				all = append(all, collect(v)...)
+			}
+		}
+
+		return all
+	}
+
+	var objs []map[string]any
+	for _, p := range responseObjects {
+		objs = append(objs, collect(p)...)
+	}
+
+	var toolCalls []api.ToolCall
+	for _, kv := range objs {
+		n, nok := kv[name].(string)
+		a, aok := kv[arguments].(map[string]any)
+		if nok && aok {
+			toolCalls = append(toolCalls, api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      n,
+					Arguments: a,
+				},
+			})
+		}
+	}
+
+	return toolCalls, len(toolCalls) > 0
+}
diff --git a/server/model_test.go b/server/model_test.go
index c3023eb2b..304d46556 100644
--- a/server/model_test.go
+++ b/server/model_test.go
@@ -1,91 +1,261 @@
 package server
 
 import (
-	"archive/zip"
 	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
 	"io"
 	"os"
 	"path/filepath"
-	"slices"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
+
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/template"
 )
 
-func createZipFile(t *testing.T, name string) *os.File {
+func readFile(t *testing.T, base, name string) *bytes.Buffer {
 	t.Helper()
 
-	f, err := os.CreateTemp(t.TempDir(), "")
+	bts, err := os.ReadFile(filepath.Join(base, name))
 	if err != nil {
 		t.Fatal(err)
 	}
 
-	zf := zip.NewWriter(f)
-	defer zf.Close()
-
-	zh, err := zf.CreateHeader(&zip.FileHeader{Name: name})
-	if err != nil {
-		t.Fatal(err)
-	}
-
-	if _, err := io.Copy(zh, bytes.NewReader([]byte(""))); err != nil {
-		t.Fatal(err)
-	}
-
-	return f
+	return bytes.NewBuffer(bts)
 }
 
-func TestExtractFromZipFile(t *testing.T) {
+func TestExecuteWithTools(t *testing.T) {
+	p := filepath.Join("testdata", "tools")
 	cases := []struct {
-		name   string
-		expect []string
+		model  string
+		output string
+		ok     bool
 	}{
+		{"mistral", `[TOOL_CALLS]  [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"mistral", `[TOOL_CALLS]  [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]
+
+The temperature in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.`, true},
+		{"mistral", `I'm not aware of that information. However, I can suggest searching for the weather using the "get_current_weather" function:
+
+		[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"mistral", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"command-r-plus", "Action: ```json" + `
+[
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {
+            "format": "fahrenheit",
+            "location": "San Francisco, CA"
+        }
+    },
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {
+            "format": "celsius",
+            "location": "Toronto, Canada"
+        }
+    }
+]
+` + "```", true},
+		{"command-r-plus", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"firefunction", ` functools[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`, true},
+		{"firefunction", " The weather in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.", false},
+		{"llama3-groq-tool-use", `<tool_call>
+{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}}
+{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
+</tool_call>`, true},
+		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
+		{"nemotron", `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]} </toolcall>`, true},
+	}
+
+	var tools []api.Tool
+	if err := json.Unmarshal(readFile(t, p, "tools.json").Bytes(), &tools); err != nil {
+		t.Fatal(err)
+	}
+
+	var messages []api.Message
+	if err := json.Unmarshal(readFile(t, p, "messages.json").Bytes(), &messages); err != nil {
+		t.Fatal(err)
+	}
+
+	calls := []api.ToolCall{
 		{
-			name:   "good",
-			expect: []string{"good"},
+			Function: api.ToolCallFunction{
+				Name: "get_current_weather",
+				Arguments: api.ToolCallFunctionArguments{
+					"format":   "fahrenheit",
+					"location": "San Francisco, CA",
+				},
+			},
 		},
 		{
-			name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"),
+			Function: api.ToolCallFunction{
+				Name: "get_current_weather",
+				Arguments: api.ToolCallFunctionArguments{
+					"format":   "celsius",
+					"location": "Toronto, Canada",
+				},
+			},
 		},
 	}
 
 	for _, tt := range cases {
-		t.Run(tt.name, func(t *testing.T) {
-			f := createZipFile(t, tt.name)
-			defer f.Close()
-
-			tempDir := t.TempDir()
-			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil {
+		t.Run(tt.model, func(t *testing.T) {
+			tmpl, err := template.Parse(readFile(t, p, fmt.Sprintf("%s.gotmpl", tt.model)).String())
+			if err != nil {
 				t.Fatal(err)
 			}
 
-			var matches []string
-			if err := filepath.Walk(tempDir, func(p string, fi os.FileInfo, err error) error {
-				if err != nil {
-					return err
+			t.Run("template", func(t *testing.T) {
+				var actual bytes.Buffer
+				if err := tmpl.Execute(&actual, template.Values{Tools: tools, Messages: messages}); err != nil {
+					t.Fatal(err)
 				}
 
-				if !fi.IsDir() {
-					matches = append(matches, p)
+				if diff := cmp.Diff(actual.String(), readFile(t, p, fmt.Sprintf("%s.out", tt.model)).String()); diff != "" {
+					t.Errorf("mismatch (-got +want):\n%s", diff)
+				}
+			})
+
+			t.Run("parse", func(t *testing.T) {
+				m := &Model{Template: tmpl}
+				actual, ok := m.parseToolCalls(tt.output)
+				if ok != tt.ok {
+					t.Fatalf("expected %t, got %t", tt.ok, ok)
 				}
 
-				return nil
-			}); err != nil {
-				t.Fatal(err)
-			}
-
-			var actual []string
-			for _, match := range matches {
-				rel, err := filepath.Rel(tempDir, match)
-				if err != nil {
-					t.Error(err)
+				if tt.ok {
+					if diff := cmp.Diff(actual, calls); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
 				}
+			})
+		})
+	}
+}
 
-				actual = append(actual, rel)
-			}
+func TestParseFromFileFromLayer(t *testing.T) {
+	tempModels := t.TempDir()
+	t.Setenv("OLLAMA_MODELS", tempModels)
 
-			if !slices.Equal(actual, tt.expect) {
-				t.Fatalf("expected %d files, got %d", len(tt.expect), len(matches))
+	file, err := os.CreateTemp(tempModels, "")
+	if err != nil {
+		t.Fatalf("failed to open file: %v", err)
+	}
+	defer file.Close()
+	if err := llm.WriteGGUF(file, llm.KV{"general.architecture": "gemma"}, []llm.Tensor{}); err != nil {
+		t.Fatalf("failed to write gguf: %v", err)
+	}
+
+	if _, err := file.Seek(0, io.SeekStart); err != nil {
+		t.Fatalf("failed to seek to start: %v", err)
+	}
+
+	layers, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file, "", func(api.ProgressResponse) {})
+	if err != nil {
+		t.Fatalf("failed to parse from file: %v", err)
+	}
+
+	if len(layers) != 1 {
+		t.Fatalf("got %d != want 1", len(layers))
+	}
+
+	if _, err := file.Seek(0, io.SeekStart); err != nil {
+		t.Fatalf("failed to seek to start: %v", err)
+	}
+
+	layers2, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file, layers[0].Digest, func(api.ProgressResponse) {})
+	if err != nil {
+		t.Fatalf("failed to parse from file: %v", err)
+	}
+	if len(layers2) != 1 {
+		t.Fatalf("got %d != want 1", len(layers2))
+	}
+
+	if layers[0].Digest != layers2[0].Digest {
+		t.Fatalf("got %s != want %s", layers[0].Digest, layers2[0].Digest)
+	}
+
+	if layers[0].Size != layers2[0].Size {
+		t.Fatalf("got %d != want %d", layers[0].Size, layers2[0].Size)
+	}
+
+	if layers[0].MediaType != layers2[0].MediaType {
+		t.Fatalf("got %v != want %v", layers[0].MediaType, layers2[0].MediaType)
+	}
+}
+
+func TestParseLayerFromCopy(t *testing.T) {
+	tempModels := t.TempDir()
+	t.Setenv("OLLAMA_MODELS", tempModels)
+
+	file2, err := os.CreateTemp(tempModels, "")
+	if err != nil {
+		t.Fatalf("failed to open file: %v", err)
+	}
+	defer file2.Close()
+
+	for range 5 {
+		if err := llm.WriteGGUF(file2, llm.KV{"general.architecture": "gemma"}, []llm.Tensor{}); err != nil {
+			t.Fatalf("failed to write gguf: %v", err)
+		}
+	}
+
+	if _, err := file2.Seek(0, io.SeekStart); err != nil {
+		t.Fatalf("failed to seek to start: %v", err)
+	}
+
+	layers, err := parseFromFile(context.Background(), "model", []*layerGGML{}, file2, "", func(api.ProgressResponse) {})
+	if err != nil {
+		t.Fatalf("failed to parse from file: %v", err)
+	}
+
+	if len(layers) != 5 {
+		t.Fatalf("got %d != want 5", len(layers))
+	}
+}
+
+func TestParseObjects(t *testing.T) {
+	tests := []struct {
+		input string
+		want  []map[string]any
+	}{
+		{
+			input: `[{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, Canada"}},
+			},
+		},
+		{
+			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall>`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+			},
+		},
+		{
+			input: `<toolcall>{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}} </toolcall> <toolcall>{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, ON"}} </toolcall>`,
+			want: []map[string]any{
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "fahrenheit", "location": "San Francisco, CA"}},
+				{"name": "get_current_weather", "arguments": map[string]any{"format": "celsius", "location": "Toronto, ON"}},
+			},
+		},
+		{
+			input: `{"name": "get_current_weather", "arguments": `,
+			want:  nil,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.input, func(t *testing.T) {
+			got := parseObjects(tc.input)
+
+			if diff := cmp.Diff(got, tc.want); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
 			}
 		})
 	}
diff --git a/server/modelpath.go b/server/modelpath.go
index 64f59c29a..d498c4678 100644
--- a/server/modelpath.go
+++ b/server/modelpath.go
@@ -73,18 +73,6 @@ func ParseModelPath(name string) ModelPath {
 
 var errModelPathInvalid = errors.New("invalid model path")
 
-func (mp ModelPath) Validate() error {
-	if mp.Repository == "" {
-		return fmt.Errorf("%w: model repository name is required", errModelPathInvalid)
-	}
-
-	if strings.Contains(mp.Tag, ":") {
-		return fmt.Errorf("%w: ':' (colon) is not allowed in tag names", errModelPathInvalid)
-	}
-
-	return nil
-}
-
 func (mp ModelPath) GetNamespaceRepository() string {
 	return fmt.Sprintf("%s/%s", mp.Namespace, mp.Repository)
 }
@@ -103,20 +91,13 @@ func (mp ModelPath) GetShortTagname() string {
 	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
 }
 
-// modelsDir returns the value of the OLLAMA_MODELS environment variable or the user's home directory if OLLAMA_MODELS is not set.
-// The models directory is where Ollama stores its model files and manifests.
-func modelsDir() (string, error) {
-	return envconfig.ModelsDir, nil
-}
-
 // GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
 func (mp ModelPath) GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
+	if p := filepath.Join(mp.Registry, mp.Namespace, mp.Repository, mp.Tag); filepath.IsLocal(p) {
+		return filepath.Join(envconfig.Models(), "manifests", p), nil
 	}
 
-	return filepath.Join(dir, "manifests", mp.Registry, mp.Namespace, mp.Repository, mp.Tag), nil
+	return "", errModelPathInvalid
 }
 
 func (mp ModelPath) BaseURL() *url.URL {
@@ -127,12 +108,7 @@ func (mp ModelPath) BaseURL() *url.URL {
 }
 
 func GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
-
-	path := filepath.Join(dir, "manifests")
+	path := filepath.Join(envconfig.Models(), "manifests")
 	if err := os.MkdirAll(path, 0o755); err != nil {
 		return "", err
 	}
@@ -141,11 +117,6 @@ func GetManifestPath() (string, error) {
 }
 
 func GetBlobsPath(digest string) (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
-
 	// only accept actual sha256 digests
 	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
 	re := regexp.MustCompile(pattern)
@@ -155,7 +126,7 @@ func GetBlobsPath(digest string) (string, error) {
 	}
 
 	digest = strings.ReplaceAll(digest, ":", "-")
-	path := filepath.Join(dir, "blobs", digest)
+	path := filepath.Join(envconfig.Models(), "blobs", digest)
 	dirPath := filepath.Dir(path)
 	if digest == "" {
 		dirPath = path
diff --git a/server/modelpath_test.go b/server/modelpath_test.go
index 6c4dfbee4..ef26266bd 100644
--- a/server/modelpath_test.go
+++ b/server/modelpath_test.go
@@ -1,14 +1,13 @@
 package server
 
 import (
+	"errors"
 	"os"
 	"path/filepath"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
-
-	"github.com/ollama/ollama/envconfig"
 )
 
 func TestGetBlobsPath(t *testing.T) {
@@ -63,7 +62,6 @@ func TestGetBlobsPath(t *testing.T) {
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
 			t.Setenv("OLLAMA_MODELS", dir)
-			envconfig.LoadConfig()
 
 			got, err := GetBlobsPath(tc.digest)
 
@@ -157,3 +155,10 @@ func TestParseModelPath(t *testing.T) {
 		})
 	}
 }
+
+func TestInsecureModelpath(t *testing.T) {
+	mp := ParseModelPath("../../..:something")
+	if _, err := mp.GetManifestPath(); !errors.Is(err, errModelPathInvalid) {
+		t.Errorf("expected error: %v", err)
+	}
+}
diff --git a/server/prompt.go b/server/prompt.go
index 604e69717..be0d49692 100644
--- a/server/prompt.go
+++ b/server/prompt.go
@@ -1,221 +1,74 @@
 package server
 
 import (
-	"fmt"
+	"bytes"
+	"context"
 	"log/slog"
-	"strings"
-	"text/template"
-	"text/template/parse"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/template"
 )
 
-// isResponseNode checks if the node contains .Response
-func isResponseNode(node *parse.ActionNode) bool {
-	for _, cmd := range node.Pipe.Cmds {
-		for _, arg := range cmd.Args {
-			if fieldNode, ok := arg.(*parse.FieldNode); ok && len(fieldNode.Ident) > 0 {
-				if fieldNode.Ident[0] == "Response" {
-					return true
-				}
+type tokenizeFunc func(context.Context, string) ([]int, error)
+
+// chatPrompt accepts a list of messages and returns the prompt and images that should be used for the next chat turn.
+// chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
+// latest message and 2) system messages
+func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message, tools []api.Tool) (prompt string, images []llm.ImageData, _ error) {
+	var system []api.Message
+	// always include the last message
+	n := len(msgs) - 1
+	// in reverse, find all messages that fit into context window
+	for i := n - 1; i >= 0; i-- {
+		system = make([]api.Message, 0)
+		for j := range i {
+			if msgs[j].Role == "system" {
+				system = append(system, msgs[j])
 			}
 		}
-	}
-	return false
-}
 
-// formatTemplateForResponse formats the template AST to:
-// 1. remove all nodes after the first .Response (if generate=true)
-// 2. add a .Response node to the end if it doesn't exist
-// TODO(jmorganca): this should recursively cut the template before the first .Response
-func formatTemplateForResponse(tmpl *template.Template, generate bool) {
-	var found bool
-	for i, node := range tmpl.Tree.Root.Nodes {
-		if actionNode, ok := node.(*parse.ActionNode); ok {
-			if isResponseNode(actionNode) {
-				found = true
-				if generate {
-					tmpl.Tree.Root.Nodes = tmpl.Tree.Root.Nodes[:i+1]
-					break
-				}
-			}
+		var b bytes.Buffer
+		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...), Tools: tools}); err != nil {
+			return "", nil, err
 		}
-	}
 
-	if !found {
-		// add the response node if it doesn't exist
-		responseFieldNode := &parse.FieldNode{NodeType: parse.NodeField, Ident: []string{"Response"}}
-		responsePipeNode := &parse.PipeNode{NodeType: parse.NodePipe, Cmds: []*parse.CommandNode{{NodeType: parse.NodeCommand, Args: []parse.Node{responseFieldNode}}}}
-		responseActionNode := &parse.ActionNode{NodeType: parse.NodeAction, Pipe: responsePipeNode}
-		tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, responseActionNode)
-	}
-}
-
-// Prompt renders a prompt from a template. If generate is set to true,
-// the response and parts of the template following it are not rendered
-func Prompt(tmpl, system, prompt, response string, generate bool) (string, error) {
-	parsed, err := template.New("").Option("missingkey=zero").Parse(tmpl)
-	if err != nil {
-		return "", err
-	}
-
-	formatTemplateForResponse(parsed, generate)
-
-	vars := map[string]any{
-		"System":   system,
-		"Prompt":   prompt,
-		"Response": response,
-	}
-
-	var sb strings.Builder
-	if err := parsed.Execute(&sb, vars); err != nil {
-		return "", err
-	}
-
-	return sb.String(), nil
-}
-
-func countTokens(tmpl string, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) {
-	rendered, err := Prompt(tmpl, system, prompt, response, false)
-	if err != nil {
-		return 0, err
-	}
-
-	tokens, err := encode(rendered)
-	if err != nil {
-		slog.Error("failed to encode prompt", "err", err)
-		return 0, err
-	}
-
-	return len(tokens), err
-}
-
-// ChatPrompt builds up a prompt from a series of messages, truncating based on context window size
-func ChatPrompt(tmpl string, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) {
-	type prompt struct {
-		System   string
-		Prompt   string
-		Response string
-
-		images []int
-		tokens int
-	}
-
-	var p prompt
-
-	// iterate through messages to build up {system,user,response} prompts
-	var imgId int
-	var prompts []prompt
-	for _, msg := range messages {
-		switch strings.ToLower(msg.Role) {
-		case "system":
-			if p.System != "" || p.Prompt != "" || p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			p.System = msg.Content
-		case "user":
-			if p.Prompt != "" || p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			var sb strings.Builder
-			for range msg.Images {
-				fmt.Fprintf(&sb, "[img-%d] ", imgId)
-				p.images = append(p.images, imgId)
-				imgId += 1
-			}
-
-			sb.WriteString(msg.Content)
-			p.Prompt = sb.String()
-		case "assistant":
-			if p.Response != "" {
-				prompts = append(prompts, p)
-				p = prompt{}
-			}
-
-			p.Response = msg.Content
-		default:
-			return "", fmt.Errorf("invalid role: %s, role must be one of [system, user, assistant]", msg.Role)
-		}
-	}
-
-	// add final prompt
-	if p.System != "" || p.Prompt != "" || p.Response != "" {
-		prompts = append(prompts, p)
-	}
-
-	// calculate token lengths for each prompt, estimating 768 tokens per images
-	for i, p := range prompts {
-		tokens, err := countTokens(tmpl, p.System, p.Prompt, p.Response, encode)
+		s, err := tokenize(ctx, b.String())
 		if err != nil {
-			return "", err
+			return "", nil, err
 		}
 
-		prompts[i].tokens = tokens + len(prompts[i].images)*768
-	}
-
-	// truncate images and prompts starting from the beginning of the list
-	// until either one prompt remains or the total tokens fits the context window
-	// TODO (jmorganca): this doesn't account for the context window room required for the response
-	for {
-		var required int
-		for _, p := range prompts {
-			required += p.tokens
+		c := len(s)
+		if m.ProjectorPaths != nil {
+			for _, m := range msgs[i:] {
+				// images are represented as 768 sized embeddings
+				// TODO: get embedding length from project metadata
+				c += 768 * len(m.Images)
+			}
 		}
 
-		required += 1 // for bos token
-
-		if required <= window {
-			slog.Debug("prompt now fits in context window", "required", required, "window", window)
+		if c > opts.NumCtx {
+			slog.Debug("truncating input messages which exceed context length", "truncated", len(msgs[i:]))
 			break
+		} else {
+			n = i
 		}
-
-		prompt := &prompts[0]
-
-		if len(prompt.images) > 1 {
-			img := prompt.images[0]
-			slog.Debug("prompt longer than context window, removing image", "id", img, "required", required, "window", window)
-			prompt.images = prompt.images[1:]
-			prompt.Prompt = strings.Replace(prompt.Prompt, fmt.Sprintf(" [img-%d]", img), "", 1)
-			prompt.tokens -= 768
-			continue
-		}
-
-		if len(prompts) > 1 {
-			slog.Debug("required tokens longer than context window, removing first prompt", "prompt", prompts[0].tokens, "required", required, "window", window)
-			system := prompt.System
-			prompts = prompts[1:]
-
-			if system != "" && prompts[0].System == "" {
-				prompts[0].System = system
-
-				tokens, err := countTokens(tmpl, prompts[0].System, prompts[0].Prompt, prompts[0].Response, encode)
-				if err != nil {
-					return "", err
-				}
-
-				prompts[0].tokens = tokens + len(prompts[0].images)*768
-			}
-
-			continue
-		}
-
-		// stop truncating if there's only one prompt left
-		break
 	}
 
-	var sb strings.Builder
-	for i, p := range prompts {
-		// last prompt should leave the response unrendered (for completion)
-		rendered, err := Prompt(tmpl, p.System, p.Prompt, p.Response, i == len(prompts)-1)
-		if err != nil {
-			return "", err
-		}
-		sb.WriteString(rendered)
+	// truncate any messages that do not fit into the context window
+	var b bytes.Buffer
+	if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[n:]...), Tools: tools}); err != nil {
+		return "", nil, err
 	}
 
-	return sb.String(), nil
+	for _, m := range msgs[n:] {
+		for _, i := range m.Images {
+			images = append(images, llm.ImageData{
+				ID:   len(images),
+				Data: i,
+			})
+		}
+	}
+
+	return b.String(), images, nil
 }
diff --git a/server/prompt_test.go b/server/prompt_test.go
index a7e18a70f..5fe3d4c56 100644
--- a/server/prompt_test.go
+++ b/server/prompt_test.go
@@ -1,204 +1,210 @@
 package server
 
 import (
-	"strings"
+	"bytes"
+	"context"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
+
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/template"
 )
 
-func TestPrompt(t *testing.T) {
-	tests := []struct {
-		name     string
-		template string
-		system   string
-		prompt   string
-		response string
-		generate bool
-		want     string
-	}{
-		{
-			name:     "simple prompt",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST]",
-		},
-		{
-			name:     "implicit response",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST]",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST]I don't know.",
-		},
-		{
-			name:     "response",
-			template: "[INST] {{ .System }} {{ .Prompt }} [/INST] {{ .Response }}",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "[INST] You are a Wizard. What are the potion ingredients? [/INST] I don't know.",
-		},
-		{
-			name:     "cut",
-			template: "<system>{{ .System }}</system><user>{{ .Prompt }}</user><assistant>{{ .Response }}</assistant>",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			generate: true,
-			want:     "<system>You are a Wizard.</system><user>What are the potion ingredients?</user><assistant>I don't know.",
-		},
-		{
-			name:     "nocut",
-			template: "<system>{{ .System }}</system><user>{{ .Prompt }}</user><assistant>{{ .Response }}</assistant>",
-			system:   "You are a Wizard.",
-			prompt:   "What are the potion ingredients?",
-			response: "I don't know.",
-			want:     "<system>You are a Wizard.</system><user>What are the potion ingredients?</user><assistant>I don't know.</assistant>",
-		},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got, err := Prompt(tc.template, tc.system, tc.prompt, tc.response, tc.generate)
-			if err != nil {
-				t.Errorf("error = %v", err)
-			}
-
-			if got != tc.want {
-				t.Errorf("got = %v, want %v", got, tc.want)
-			}
-		})
-	}
-}
-
 func TestChatPrompt(t *testing.T) {
-	tests := []struct {
-		name     string
-		template string
-		messages []api.Message
-		window   int
-		want     string
+	type expect struct {
+		prompt string
+		images [][]byte
+	}
+
+	cases := []struct {
+		name  string
+		limit int
+		msgs  []api.Message
+		expect
 	}{
 		{
-			name:     "simple prompt",
-			template: "[INST] {{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "user", Content: "Hello"},
+			name:  "messages",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "[INST] Hello [/INST]",
 		},
 		{
-			name:     "with system message",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
+			name:  "truncate messages",
+			limit: 1,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST]",
 		},
 		{
-			name:     "with response",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
+			name:  "truncate messages with image",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("something")}},
+			},
+			expect: expect{
+				prompt: "[img-0] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST] I am?",
 		},
 		{
-			name:     "with implicit response",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST]",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
+			name:  "truncate messages with images",
+			limit: 64,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "[img-0] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> Hello [/INST]I am?",
 		},
 		{
-			name:     "with conversation",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "What are the potion ingredients?"},
-				{Role: "assistant", Content: "sugar"},
-				{Role: "user", Content: "Anything else?"},
+			name:  "messages with images",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "[img-0] You're a test, Harry! I-I'm a what? [img-1] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "[INST] <<SYS>>You are a Wizard.<</SYS>> What are the potion ingredients? [/INST] sugar [INST] Anything else? [/INST] ",
 		},
 		{
-			name:     "with truncation",
-			template: "{{ .System }} {{ .Prompt }} {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello"},
-				{Role: "assistant", Content: "I am?"},
-				{Role: "user", Content: "Why is the sky blue?"},
-				{Role: "assistant", Content: "The sky is blue from rayleigh scattering"},
+			name:  "message with image tag",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry! [img]", Images: []api.ImageData{[]byte("something")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager.", Images: []api.ImageData{[]byte("somethingelse")}},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! [img-0] I-I'm a what? [img-1] A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 10,
-			want:   "You are a Wizard. Why is the sky blue? The sky is blue from rayleigh scattering",
 		},
 		{
-			name:     "images",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("base64")}},
+			name:  "messages with interleaved images",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "user", Images: []api.ImageData{[]byte("something")}},
+				{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry!\n\n[img-0]\n\n[img-1] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("something"),
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "You are a Wizard. [img-0] Hello",
 		},
 		{
-			name:     "images truncated",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{
-				{Role: "system", Content: "You are a Wizard."},
-				{Role: "user", Content: "Hello", Images: []api.ImageData{[]byte("img1"), []byte("img2")}},
+			name:  "truncate message with interleaved images",
+			limit: 1024,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "user", Images: []api.ImageData{[]byte("something")}},
+				{Role: "user", Images: []api.ImageData{[]byte("somethingelse")}},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "[img-0] I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+				images: [][]byte{
+					[]byte("somethingelse"),
+				},
 			},
-			window: 1024,
-			want:   "You are a Wizard. [img-0] [img-1] Hello",
 		},
 		{
-			name:     "empty list",
-			template: "{{ .System }} {{ .Prompt }}",
-			messages: []api.Message{},
-			window:   1024,
-			want:     "",
+			name:  "message with system prompt",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "system", Content: "You are the Test Who Lived."},
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You are the Test Who Lived. You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+			},
 		},
 		{
-			name:     "empty prompt",
-			template: "[INST] {{ if .System }}<<SYS>>{{ .System }}<</SYS>> {{ end }}{{ .Prompt }} [/INST] {{ .Response }} ",
-			messages: []api.Message{
-				{Role: "user", Content: ""},
+			name:  "out of order system",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "system", Content: "You are the Test Who Lived."},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
+			expect: expect{
+				prompt: "You're a test, Harry! I-I'm a what? You are the Test Who Lived. A test. And a thumping good one at that, I'd wager. ",
 			},
-			window: 1024,
-			want:   "",
 		},
 	}
 
-	encode := func(s string) ([]int, error) {
-		words := strings.Fields(s)
-		return make([]int, len(words)), nil
+	tmpl, err := template.Parse(`
+{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}{{ .Prompt }} {{ end }}
+{{- if .Response }}{{ .Response }} {{ end }}`)
+	if err != nil {
+		t.Fatal(err)
 	}
 
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			got, err := ChatPrompt(tc.template, tc.messages, tc.window, encode)
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			model := Model{Template: tmpl, ProjectorPaths: []string{"vision"}}
+			opts := api.Options{Runner: api.Runner{NumCtx: tt.limit}}
+			prompt, images, err := chatPrompt(context.TODO(), &model, mockRunner{}.Tokenize, &opts, tt.msgs, nil)
 			if err != nil {
-				t.Errorf("error = %v", err)
+				t.Fatal(err)
 			}
 
-			if got != tc.want {
-				t.Errorf("got: %q, want: %q", got, tc.want)
+			if diff := cmp.Diff(prompt, tt.prompt); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+
+			if len(images) != len(tt.images) {
+				t.Fatalf("expected %d images, got %d", len(tt.images), len(images))
+			}
+
+			for i := range images {
+				if images[i].ID != i {
+					t.Errorf("expected ID %d, got %d", i, images[i].ID)
+				}
+
+				if !bytes.Equal(images[i].Data, tt.images[i]) {
+					t.Errorf("expected %q, got %q", tt.images[i], images[i])
+				}
 			}
 		})
 	}
diff --git a/server/routes.go b/server/routes.go
index 76ead072f..6bd3a93f5 100644
--- a/server/routes.go
+++ b/server/routes.go
@@ -1,13 +1,13 @@
 package server
 
 import (
+	"bytes"
 	"cmp"
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
-	"io/fs"
 	"log/slog"
 	"math"
 	"net"
@@ -17,20 +17,23 @@ import (
 	"os/signal"
 	"path/filepath"
 	"slices"
-	"strconv"
 	"strings"
 	"syscall"
 	"time"
 
 	"github.com/gin-contrib/cors"
 	"github.com/gin-gonic/gin"
+	"golang.org/x/sync/errgroup"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/build"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/runners"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -55,7 +58,10 @@ func init() {
 	gin.SetMode(mode)
 }
 
-var defaultSessionDuration = 5 * time.Minute
+var (
+	errRequired    = errors.New("is required")
+	errBadTemplate = errors.New("template error")
+)
 
 func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
 	opts := api.DefaultOptions()
@@ -70,277 +76,250 @@ func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options
 	return opts, nil
 }
 
-func isSupportedImageType(image []byte) bool {
-	contentType := http.DetectContentType(image)
-	allowedTypes := []string{"image/jpeg", "image/jpg", "image/png"}
-	return slices.Contains(allowedTypes, contentType)
+// scheduleRunner schedules a runner after validating inputs such as capabilities and model options.
+// It returns the allocated runner, model instance, and consolidated options if successful and error otherwise.
+func (s *Server) scheduleRunner(ctx context.Context, name string, caps []Capability, requestOpts map[string]any, keepAlive *api.Duration) (llm.LlamaServer, *Model, *api.Options, error) {
+	if name == "" {
+		return nil, nil, nil, fmt.Errorf("model %w", errRequired)
+	}
+
+	model, err := GetModel(name)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+
+	if err := model.CheckCapabilities(caps...); err != nil {
+		return nil, nil, nil, fmt.Errorf("%s %w", name, err)
+	}
+
+	opts, err := modelOptions(model, requestOpts)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+
+	runnerCh, errCh := s.sched.GetRunner(ctx, model, opts, keepAlive)
+	var runner *runnerRef
+	select {
+	case runner = <-runnerCh:
+	case err = <-errCh:
+		return nil, nil, nil, err
+	}
+
+	return runner.llama, model, &opts, nil
 }
 
 func (s *Server) GenerateHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 	var req api.GenerateRequest
-	err := c.ShouldBindJSON(&req)
-
-	switch {
-	case errors.Is(err, io.EOF):
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
 		return
-	case err != nil:
+	} else if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	// validate the request
-	switch {
-	case req.Model == "":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
+	// expire the runner
+	if req.Prompt == "" && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
+		model, err := GetModel(req.Model)
+		if err != nil {
+			switch {
+			case os.IsNotExist(err):
+				c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
+			case err.Error() == "invalid model name":
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			}
+			return
+		}
+		s.sched.expireRunner(model)
+
+		c.JSON(http.StatusOK, api.GenerateResponse{
+			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Response:   "",
+			Done:       true,
+			DoneReason: "unload",
+		})
 		return
-	case len(req.Format) > 0 && req.Format != "json":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
+	}
+
+	if req.Format != "" && req.Format != "json" {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be empty or \"json\""})
 		return
-	case req.Raw && (req.Template != "" || req.System != "" || len(req.Context) > 0):
+	} else if req.Raw && (req.Template != "" || req.System != "" || len(req.Context) > 0) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "raw mode does not support template, system, or context"})
 		return
 	}
 
-	for _, img := range req.Images {
-		if !isSupportedImageType(img) {
-			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "unsupported image format"})
-			return
-		}
+	caps := []Capability{CapabilityCompletion}
+	if req.Suffix != "" {
+		caps = append(caps, CapabilityInsert)
 	}
 
-	model, err := GetModel(req.Model)
-	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
-			return
-		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, caps, req.Options, req.KeepAlive)
+	if errors.Is(err, errCapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support generate", req.Model)})
+		return
+	} else if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
-	if model.IsEmbedding() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support generate"})
-		return
-	}
+	checkpointLoaded := time.Now()
 
-	opts, err := modelOptions(model, req.Options)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
-		return
-	}
-
-	// an empty request loads the model
-	// note: for a short while template was used in lieu
-	// of `raw` mode so we need to check for it too
-	if req.Prompt == "" && req.Template == "" && req.System == "" {
+	if req.Prompt == "" {
 		c.JSON(http.StatusOK, api.GenerateResponse{
-			CreatedAt:  time.Now().UTC(),
 			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
 			Done:       true,
 			DoneReason: "load",
 		})
 		return
 	}
 
-	checkpointLoaded := time.Now()
+	images := make([]llm.ImageData, len(req.Images))
+	for i := range req.Images {
+		images[i] = llm.ImageData{ID: i, Data: req.Images[i]}
+	}
 
-	var prompt string
-	switch {
-	case req.Raw:
-		prompt = req.Prompt
-	case req.Prompt != "":
-		if req.Template == "" {
-			req.Template = model.Template
-		}
-
-		if req.System == "" {
-			req.System = model.System
-		}
-
-		slog.Debug("generate handler", "prompt", req.Prompt)
-		slog.Debug("generate handler", "template", req.Template)
-		slog.Debug("generate handler", "system", req.System)
-
-		var sb strings.Builder
-		for i := range req.Images {
-			fmt.Fprintf(&sb, "[img-%d] ", i)
-		}
-
-		sb.WriteString(req.Prompt)
-
-		p, err := Prompt(req.Template, req.System, sb.String(), "", true)
-		if err != nil {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-			return
-		}
-
-		sb.Reset()
-		if req.Context != nil {
-			prev, err := runner.llama.Detokenize(c.Request.Context(), req.Context)
+	prompt := req.Prompt
+	if !req.Raw {
+		tmpl := m.Template
+		if req.Template != "" {
+			tmpl, err = template.Parse(req.Template)
 			if err != nil {
 				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 				return
 			}
-
-			sb.WriteString(prev)
 		}
 
-		sb.WriteString(p)
-
-		prompt = sb.String()
-	}
-
-	slog.Debug("generate handler", "prompt", prompt)
-
-	ch := make(chan any)
-	var generated strings.Builder
-	go func() {
-		defer close(ch)
-
-		fn := func(r llm.CompletionResponse) {
-			// Build up the full response
-			if _, err := generated.WriteString(r.Content); err != nil {
-				ch <- gin.H{"error": err.Error()}
-				return
+		var values template.Values
+		if req.Suffix != "" {
+			values.Prompt = prompt
+			values.Suffix = req.Suffix
+		} else {
+			var msgs []api.Message
+			if req.System != "" {
+				msgs = append(msgs, api.Message{Role: "system", Content: req.System})
+			} else if m.System != "" {
+				msgs = append(msgs, api.Message{Role: "system", Content: m.System})
 			}
 
-			resp := api.GenerateResponse{
+			if req.Context == nil {
+				msgs = append(msgs, m.Messages...)
+			}
+
+			for _, i := range images {
+				msgs = append(msgs, api.Message{Role: "user", Content: fmt.Sprintf("[img-%d]", i.ID)})
+			}
+
+			values.Messages = append(msgs, api.Message{Role: "user", Content: req.Prompt})
+		}
+
+		var b bytes.Buffer
+		if req.Context != nil {
+			s, err := r.Detokenize(c.Request.Context(), req.Context)
+			if err != nil {
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+				return
+			}
+			b.WriteString(s)
+		}
+
+		if err := tmpl.Execute(&b, values); err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		prompt = b.String()
+	}
+
+	slog.Debug("generate request", "prompt", prompt, "images", images)
+
+	ch := make(chan any)
+	go func() {
+		// TODO (jmorganca): avoid building the response twice both here and below
+		var sb strings.Builder
+		defer close(ch)
+		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
+		}, func(cr llm.CompletionResponse) {
+			res := api.GenerateResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
-				Done:       r.Done,
-				Response:   r.Content,
-				DoneReason: r.DoneReason,
+				Response:   cr.Content,
+				Done:       cr.Done,
+				DoneReason: cr.DoneReason,
 				Metrics: api.Metrics{
-					PromptEvalCount:    r.PromptEvalCount,
-					PromptEvalDuration: r.PromptEvalDuration,
-					EvalCount:          r.EvalCount,
-					EvalDuration:       r.EvalDuration,
+					PromptEvalCount:    cr.PromptEvalCount,
+					PromptEvalDuration: cr.PromptEvalDuration,
+					EvalCount:          cr.EvalCount,
+					EvalDuration:       cr.EvalDuration,
 				},
 			}
 
-			if r.Done {
-				resp.TotalDuration = time.Since(checkpointStart)
-				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+			if _, err := sb.WriteString(cr.Content); err != nil {
+				ch <- gin.H{"error": err.Error()}
+			}
+
+			if cr.Done {
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 
 				if !req.Raw {
-					p, err := Prompt(req.Template, req.System, req.Prompt, generated.String(), false)
-					if err != nil {
-						c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-						return
-					}
-
-					// TODO (jmorganca): encode() should not strip special tokens
-					tokens, err := runner.llama.Tokenize(c.Request.Context(), p)
+					tokens, err := r.Tokenize(c.Request.Context(), prompt+sb.String())
 					if err != nil {
 						ch <- gin.H{"error": err.Error()}
 						return
 					}
-
-					resp.Context = append(req.Context, tokens...)
+					res.Context = tokens
 				}
 			}
 
-			ch <- resp
-		}
-
-		var images []llm.ImageData
-		for i := range req.Images {
-			images = append(images, llm.ImageData{
-				ID:   i,
-				Data: req.Images[i],
-			})
-		}
-
-		// Start prediction
-		req := llm.CompletionRequest{
-			Prompt:  prompt,
-			Format:  req.Format,
-			Images:  images,
-			Options: opts,
-		}
-		if err := runner.llama.Completion(c.Request.Context(), req, fn); err != nil {
+			ch <- res
+		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		// Accumulate responses into the final response
-		var final api.GenerateResponse
+		var r api.GenerateResponse
 		var sb strings.Builder
-		for resp := range ch {
-			switch r := resp.(type) {
+		for rr := range ch {
+			switch t := rr.(type) {
 			case api.GenerateResponse:
-				sb.WriteString(r.Response)
-				final = r
+				sb.WriteString(t.Response)
+				r = t
 			case gin.H:
-				if errorMsg, ok := r["error"].(string); ok {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
-					return
-				} else {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
-					return
+				msg, ok := t["error"].(string)
+				if !ok {
+					msg = "unexpected error format in response"
 				}
+
+				c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
+				return
 			default:
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})
 				return
 			}
 		}
 
-		final.Response = sb.String()
-		c.JSON(http.StatusOK, final)
+		r.Response = sb.String()
+		c.JSON(http.StatusOK, r)
 		return
 	}
 
 	streamResponse(c, ch)
 }
 
-func getDefaultSessionDuration() time.Duration {
-	if envconfig.KeepAlive != "" {
-		v, err := strconv.Atoi(envconfig.KeepAlive)
-		if err != nil {
-			d, err := time.ParseDuration(envconfig.KeepAlive)
-			if err != nil {
-				return defaultSessionDuration
-			}
-
-			if d < 0 {
-				return time.Duration(math.MaxInt64)
-			}
-
-			return d
-		}
-
-		d := time.Duration(v) * time.Second
-		if d < 0 {
-			return time.Duration(math.MaxInt64)
-		}
-		return d
-	}
-
-	return defaultSessionDuration
-}
-
-func (s *Server) EmbeddingsHandler(c *gin.Context) {
-	var req api.EmbeddingRequest
+func (s *Server) EmbedHandler(c *gin.Context) {
+	checkpointStart := time.Now()
+	var req api.EmbedRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
 	case errors.Is(err, io.EOF):
@@ -351,41 +330,140 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	if req.Model == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
-		return
+	truncate := true
+
+	if req.Truncate != nil && !*req.Truncate {
+		truncate = false
 	}
 
-	model, err := GetModel(req.Model)
-	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
+	var input []string
+
+	switch i := req.Input.(type) {
+	case string:
+		if len(i) > 0 {
+			input = append(input, i)
+		}
+	case []any:
+		for _, v := range i {
+			if _, ok := v.(string); !ok {
+				c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid input type"})
+				return
+			}
+			input = append(input, v.(string))
+		}
+	default:
+		if req.Input != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "invalid input type"})
 			return
 		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+	}
+
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, []Capability{}, req.Options, req.KeepAlive)
+	if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
-	opts, err := modelOptions(model, req.Options)
+	checkpointLoaded := time.Now()
+
+	if len(input) == 0 {
+		c.JSON(http.StatusOK, api.EmbedResponse{Model: req.Model, Embeddings: [][]float32{}})
+		return
+	}
+
+	kvData, err := getKVData(m.ModelPath, false)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
+	var count int
+	for i, s := range input {
+		tokens, err := r.Tokenize(c.Request.Context(), s)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+
+		ctxLen := min(opts.NumCtx, int(kvData.ContextLength()))
+		if len(tokens) > ctxLen {
+			if !truncate {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "input length exceeds maximum context length"})
+				return
+			}
+
+			tokens = tokens[:ctxLen]
+			s, err = r.Detokenize(c.Request.Context(), tokens)
+			if err != nil {
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+				return
+			}
+		}
+
+		count += len(tokens)
+
+		input[i] = s
 	}
 
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
+	var g errgroup.Group
+	embeddings := make([][]float32, len(input))
+	for i, text := range input {
+		g.Go(func() error {
+			embedding, err := r.Embedding(c.Request.Context(), text)
+			if err != nil {
+				return err
+			}
+			embeddings[i] = normalize(embedding)
+			return nil
+		})
+	}
+
+	if err := g.Wait(); err != nil {
+		slog.Error("embedding generation failed", "error", err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Errorf("failed to generate embeddings: %v", err)})
+		return
+	}
+
+	resp := api.EmbedResponse{
+		Model:           req.Model,
+		Embeddings:      embeddings,
+		TotalDuration:   time.Since(checkpointStart),
+		LoadDuration:    checkpointLoaded.Sub(checkpointStart),
+		PromptEvalCount: count,
+	}
+	c.JSON(http.StatusOK, resp)
+}
+
+func normalize(vec []float32) []float32 {
+	var sum float32
+	for _, v := range vec {
+		sum += v * v
+	}
+
+	norm := float32(0.0)
+	if sum > 0 {
+		norm = float32(1.0 / math.Sqrt(float64(sum)))
+	}
+
+	for i := range vec {
+		vec[i] *= norm
+	}
+	return vec
+}
+
+func (s *Server) EmbeddingsHandler(c *gin.Context) {
+	var req api.EmbeddingRequest
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
+		return
+	} else if err != nil {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+
+	r, _, _, err := s.scheduleRunner(c.Request.Context(), req.Model, []Capability{}, req.Options, req.KeepAlive)
+	if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
@@ -395,20 +473,25 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	embedding, err := runner.llama.Embedding(c.Request.Context(), req.Prompt)
+	embedding, err := r.Embedding(c.Request.Context(), req.Prompt)
 	if err != nil {
 		slog.Info(fmt.Sprintf("embedding generation failed: %v", err))
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to generate embedding"})
 		return
 	}
 
+	var e []float64
+	for _, v := range embedding {
+		e = append(e, float64(v))
+	}
+
 	resp := api.EmbeddingResponse{
-		Embedding: embedding,
+		Embedding: e,
 	}
 	c.JSON(http.StatusOK, resp)
 }
 
-func (s *Server) PullModelHandler(c *gin.Context) {
+func (s *Server) PullHandler(c *gin.Context) {
 	var req api.PullRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
@@ -458,7 +541,7 @@ func (s *Server) PullModelHandler(c *gin.Context) {
 	streamResponse(c, ch)
 }
 
-func (s *Server) PushModelHandler(c *gin.Context) {
+func (s *Server) PushHandler(c *gin.Context) {
 	var req api.PushRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
@@ -515,14 +598,14 @@ func checkNameExists(name model.Name) error {
 
 	for n := range names {
 		if strings.EqualFold(n.Filepath(), name.Filepath()) && n != name {
-			return fmt.Errorf("a model with that name already exists")
+			return errors.New("a model with that name already exists")
 		}
 	}
 
 	return nil
 }
 
-func (s *Server) CreateModelHandler(c *gin.Context) {
+func (s *Server) CreateHandler(c *gin.Context) {
 	var r api.CreateRequest
 	if err := c.ShouldBindJSON(&r); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -577,7 +660,9 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 		defer cancel()
 
 		quantization := cmp.Or(r.Quantize, r.Quantization)
-		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); err != nil {
+		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); errors.Is(err, errBadTemplate) {
+			ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
+		} else if err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -590,7 +675,7 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 	streamResponse(c, ch)
 }
 
-func (s *Server) DeleteModelHandler(c *gin.Context) {
+func (s *Server) DeleteHandler(c *gin.Context) {
 	var r api.DeleteRequest
 	if err := c.ShouldBindJSON(&r); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -623,7 +708,7 @@ func (s *Server) DeleteModelHandler(c *gin.Context) {
 	}
 }
 
-func (s *Server) ShowModelHandler(c *gin.Context) {
+func (s *Server) ShowHandler(c *gin.Context) {
 	var req api.ShowRequest
 	err := c.ShouldBindJSON(&req)
 	switch {
@@ -679,18 +764,14 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 		m.System = req.System
 	}
 
-	if req.Template != "" {
-		m.Template = req.Template
-	}
-
-	msgs := make([]api.Message, 0)
-	for _, msg := range m.Messages {
-		msgs = append(msgs, api.Message{Role: msg.Role, Content: msg.Content})
+	msgs := make([]api.Message, len(m.Messages))
+	for i, msg := range m.Messages {
+		msgs[i] = api.Message{Role: msg.Role, Content: msg.Content}
 	}
 
 	n := model.ParseName(req.Model)
 	if !n.IsValid() {
-		return nil, fmt.Errorf("invalid model name")
+		return nil, errors.New("invalid model name")
 	}
 
 	manifest, err := ParseNamedManifest(n)
@@ -701,7 +782,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	resp := &api.ShowResponse{
 		License:    strings.Join(m.License, "\n"),
 		System:     m.System,
-		Template:   m.Template,
+		Template:   m.Template.String(),
 		Details:    modelDetails,
 		Messages:   msgs,
 		ModifiedAt: manifest.fi.ModTime(),
@@ -776,7 +857,7 @@ func getKVData(digest string, verbose bool) (llm.KV, error) {
 	return kv, nil
 }
 
-func (s *Server) ListModelsHandler(c *gin.Context) {
+func (s *Server) ListHandler(c *gin.Context) {
 	ms, err := Manifests()
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -785,17 +866,20 @@ func (s *Server) ListModelsHandler(c *gin.Context) {
 
 	models := []api.ListModelResponse{}
 	for n, m := range ms {
-		f, err := m.Config.Open()
-		if err != nil {
-			slog.Warn("bad manifest filepath", "name", n, "error", err)
-			continue
-		}
-		defer f.Close()
-
 		var cf ConfigV2
-		if err := json.NewDecoder(f).Decode(&cf); err != nil {
-			slog.Warn("bad manifest config", "name", n, "error", err)
-			continue
+
+		if m.Config.Digest != "" {
+			f, err := m.Config.Open()
+			if err != nil {
+				slog.Warn("bad manifest filepath", "name", n, "error", err)
+				continue
+			}
+			defer f.Close()
+
+			if err := json.NewDecoder(f).Decode(&cf); err != nil {
+				slog.Warn("bad manifest config", "name", n, "error", err)
+				continue
+			}
 		}
 
 		// tag should never be masked
@@ -823,7 +907,7 @@ func (s *Server) ListModelsHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, api.ListResponse{Models: models})
 }
 
-func (s *Server) CopyModelHandler(c *gin.Context) {
+func (s *Server) CopyHandler(c *gin.Context) {
 	var r api.CopyRequest
 	if err := c.ShouldBindJSON(&r); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -954,7 +1038,7 @@ func allowedHost(host string) bool {
 		return true
 	}
 
-	var tlds = []string{
+	tlds := []string{
 		"localhost",
 		"local",
 		"internal",
@@ -1017,7 +1101,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 	for _, prop := range openAIProperties {
 		config.AllowHeaders = append(config.AllowHeaders, "x-stainless-"+prop)
 	}
-	config.AllowOrigins = envconfig.AllowOrigins
+	config.AllowOrigins = envconfig.Origins()
 
 	r := gin.Default()
 	r.Use(
@@ -1025,28 +1109,33 @@ func (s *Server) GenerateRoutes() http.Handler {
 		allowedHostsMiddleware(s.addr),
 	)
 
-	r.POST("/api/pull", s.PullModelHandler)
+	r.POST("/api/pull", s.PullHandler)
 	r.POST("/api/generate", s.GenerateHandler)
 	r.POST("/api/chat", s.ChatHandler)
+	r.POST("/api/embed", s.EmbedHandler)
 	r.POST("/api/embeddings", s.EmbeddingsHandler)
-	r.POST("/api/create", s.CreateModelHandler)
-	r.POST("/api/push", s.PushModelHandler)
-	r.POST("/api/copy", s.CopyModelHandler)
-	r.DELETE("/api/delete", s.DeleteModelHandler)
-	r.POST("/api/show", s.ShowModelHandler)
+	r.POST("/api/create", s.CreateHandler)
+	r.POST("/api/push", s.PushHandler)
+	r.POST("/api/copy", s.CopyHandler)
+	r.DELETE("/api/delete", s.DeleteHandler)
+	r.POST("/api/show", s.ShowHandler)
 	r.POST("/api/blobs/:digest", s.CreateBlobHandler)
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
-	r.GET("/api/ps", s.ProcessHandler)
+	r.GET("/api/ps", s.PsHandler)
 
 	// Compatibility endpoints
-	r.POST("/v1/chat/completions", openai.Middleware(), s.ChatHandler)
+	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
+	r.POST("/v1/completions", openai.CompletionsMiddleware(), s.GenerateHandler)
+	r.POST("/v1/embeddings", openai.EmbeddingsMiddleware(), s.EmbedHandler)
+	r.GET("/v1/models", openai.ListMiddleware(), s.ListHandler)
+	r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowHandler)
 
 	for _, method := range []string{http.MethodGet, http.MethodHead} {
 		r.Handle(method, "/", func(c *gin.Context) {
 			c.String(http.StatusOK, "Ollama is running")
 		})
 
-		r.Handle(method, "/api/tags", s.ListModelsHandler)
+		r.Handle(method, "/api/tags", s.ListHandler)
 		r.Handle(method, "/api/version", func(c *gin.Context) {
 			c.JSON(http.StatusOK, gin.H{"version": version.Version})
 		})
@@ -1057,7 +1146,7 @@ func (s *Server) GenerateRoutes() http.Handler {
 
 func Serve(ln net.Listener) error {
 	level := slog.LevelInfo
-	if envconfig.Debug {
+	if envconfig.Debug() {
 		level = slog.LevelDebug
 	}
 
@@ -1085,7 +1174,7 @@ func Serve(ln net.Listener) error {
 		return err
 	}
 
-	if !envconfig.NoPrune {
+	if !envconfig.NoPrune() {
 		// clean up unused layers and manifests
 		if err := PruneLayers(); err != nil {
 			return err
@@ -1129,12 +1218,12 @@ func Serve(ln net.Listener) error {
 		srvr.Close()
 		schedDone()
 		sched.unloadAllRunners()
-		gpu.Cleanup()
+		runners.Cleanup(build.EmbedFS)
 		done()
 	}()
 
-	if err := llm.Init(); err != nil {
-		return fmt.Errorf("unable to initialize llm library %w", err)
+	if _, err := runners.Refresh(build.EmbedFS); err != nil {
+		return fmt.Errorf("unable to initialize llm runners %w", err)
 	}
 
 	s.sched.Run(schedCtx)
@@ -1164,11 +1253,15 @@ func waitForStream(c *gin.Context, ch chan interface{}) {
 				return
 			}
 		case gin.H:
+			status, ok := r["status"].(int)
+			if !ok {
+				status = http.StatusInternalServerError
+			}
 			if errorMsg, ok := r["error"].(string); ok {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
+				c.JSON(status, gin.H{"error": errorMsg})
 				return
 			} else {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in progress response"})
+				c.JSON(status, gin.H{"error": "unexpected error format in progress response"})
 				return
 			}
 		default:
@@ -1204,7 +1297,7 @@ func streamResponse(c *gin.Context, ch chan any) {
 	})
 }
 
-func (s *Server) ProcessHandler(c *gin.Context) {
+func (s *Server) PsHandler(c *gin.Context) {
 	models := []api.ProcessModelResponse{}
 
 	for _, v := range s.sched.loaded {
@@ -1245,139 +1338,94 @@ func (s *Server) ProcessHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, api.ProcessResponse{Models: models})
 }
 
-// ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model
-func chatPrompt(ctx context.Context, runner *runnerRef, template string, messages []api.Message, numCtx int) (string, error) {
-	encode := func(s string) ([]int, error) {
-		return runner.llama.Tokenize(ctx, s)
-	}
-
-	prompt, err := ChatPrompt(template, messages, numCtx, encode)
-	if err != nil {
-		return "", err
-	}
-
-	return prompt, nil
-}
-
 func (s *Server) ChatHandler(c *gin.Context) {
 	checkpointStart := time.Now()
 
 	var req api.ChatRequest
-	err := c.ShouldBindJSON(&req)
-	switch {
-	case errors.Is(err, io.EOF):
+	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
 		return
-	case err != nil:
+	} else if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	// validate the request
-	switch {
-	case req.Model == "":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "model is required"})
-		return
-	case len(req.Format) > 0 && req.Format != "json":
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "format must be json"})
-		return
-	}
-
-	model, err := GetModel(req.Model)
-	if err != nil {
-		var pErr *fs.PathError
-		if errors.As(err, &pErr) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found, try pulling it first", req.Model)})
+	// expire the runner
+	if len(req.Messages) == 0 && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
+		model, err := GetModel(req.Model)
+		if err != nil {
+			switch {
+			case os.IsNotExist(err):
+				c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Model)})
+			case err.Error() == "invalid model name":
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			default:
+				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			}
 			return
 		}
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		s.sched.expireRunner(model)
+
+		c.JSON(http.StatusOK, api.ChatResponse{
+			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Message:    api.Message{Role: "assistant"},
+			Done:       true,
+			DoneReason: "unload",
+		})
 		return
 	}
 
-	if model.IsEmbedding() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support chat"})
+	caps := []Capability{CapabilityCompletion}
+	if len(req.Tools) > 0 {
+		caps = append(caps, CapabilityTools)
+	}
+
+	r, m, opts, err := s.scheduleRunner(c.Request.Context(), req.Model, caps, req.Options, req.KeepAlive)
+	if errors.Is(err, errCapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%q does not support chat", req.Model)})
 		return
-	}
-
-	opts, err := modelOptions(model, req.Options)
-	if err != nil {
-		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		return
-	}
-
-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
-	var runner *runnerRef
-	select {
-	case runner = <-rCh:
-	case err = <-eCh:
-		handleErrorResponse(c, err)
+	} else if err != nil {
+		handleScheduleError(c, req.Model, err)
 		return
 	}
 
 	checkpointLoaded := time.Now()
 
-	// if the first message is not a system message, then add the model's default system message
-	if len(req.Messages) > 0 && req.Messages[0].Role != "system" {
-		req.Messages = append([]api.Message{
-			{
-				Role:    "system",
-				Content: model.System,
-			},
-		}, req.Messages...)
-	}
-
-	prompt, err := chatPrompt(c.Request.Context(), runner, model.Template, req.Messages, opts.NumCtx)
-	if err != nil {
-		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	// an empty request loads the model
-	if len(req.Messages) == 0 || prompt == "" {
-		resp := api.ChatResponse{
-			CreatedAt:  time.Now().UTC(),
+	if len(req.Messages) == 0 {
+		c.JSON(http.StatusOK, api.ChatResponse{
 			Model:      req.Model,
+			CreatedAt:  time.Now().UTC(),
+			Message:    api.Message{Role: "assistant"},
 			Done:       true,
 			DoneReason: "load",
-			Message:    api.Message{Role: "assistant"},
-		}
-		c.JSON(http.StatusOK, resp)
+		})
 		return
 	}
 
-	// only send images that are in the prompt
-	var i int
-	var images []llm.ImageData
-	for _, m := range req.Messages {
-		for _, img := range m.Images {
-			if !isSupportedImageType(img) {
-				c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "unsupported image format"})
-				return
-			}
-
-			if strings.Contains(prompt, fmt.Sprintf("[img-%d]", i)) {
-				images = append(images, llm.ImageData{Data: img, ID: i})
-			}
-			i += 1
-		}
+	msgs := append(m.Messages, req.Messages...)
+	if req.Messages[0].Role != "system" && m.System != "" {
+		msgs = append([]api.Message{{Role: "system", Content: m.System}}, msgs...)
 	}
 
-	slog.Debug("chat handler", "prompt", prompt, "images", len(images))
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	slog.Debug("chat request", "images", len(images), "prompt", prompt)
 
 	ch := make(chan any)
-
 	go func() {
 		defer close(ch)
-
-		fn := func(r llm.CompletionResponse) {
-			resp := api.ChatResponse{
+		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
+			Prompt:  prompt,
+			Images:  images,
+			Format:  req.Format,
+			Options: opts,
+		}, func(r llm.CompletionResponse) {
+			res := api.ChatResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
 				Message:    api.Message{Role: "assistant", Content: r.Content},
@@ -1392,62 +1440,65 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			}
 
 			if r.Done {
-				resp.TotalDuration = time.Since(checkpointStart)
-				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
 			}
 
-			ch <- resp
-		}
-
-		if err := runner.llama.Completion(c.Request.Context(), llm.CompletionRequest{
-			Prompt:  prompt,
-			Format:  req.Format,
-			Images:  images,
-			Options: opts,
-		}, fn); err != nil {
+			ch <- res
+		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
 
 	if req.Stream != nil && !*req.Stream {
-		// Accumulate responses into the final response
-		var final api.ChatResponse
+		var resp api.ChatResponse
 		var sb strings.Builder
-		for resp := range ch {
-			switch r := resp.(type) {
+		for rr := range ch {
+			switch t := rr.(type) {
 			case api.ChatResponse:
-				sb.WriteString(r.Message.Content)
-				final = r
+				sb.WriteString(t.Message.Content)
+				resp = t
 			case gin.H:
-				if errorMsg, ok := r["error"].(string); ok {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
-					return
-				} else {
-					c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in response"})
-					return
+				msg, ok := t["error"].(string)
+				if !ok {
+					msg = "unexpected error format in response"
 				}
+
+				c.JSON(http.StatusInternalServerError, gin.H{"error": msg})
+				return
 			default:
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error"})
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected response"})
 				return
 			}
 		}
 
-		final.Message = api.Message{Role: "assistant", Content: sb.String()}
-		c.JSON(http.StatusOK, final)
+		resp.Message.Content = sb.String()
+
+		if len(req.Tools) > 0 {
+			if toolCalls, ok := m.parseToolCalls(sb.String()); ok {
+				resp.Message.ToolCalls = toolCalls
+				resp.Message.Content = ""
+			}
+		}
+
+		c.JSON(http.StatusOK, resp)
 		return
 	}
 
 	streamResponse(c, ch)
 }
 
-func handleErrorResponse(c *gin.Context, err error) {
-	if errors.Is(err, context.Canceled) {
+func handleScheduleError(c *gin.Context, name string, err error) {
+	switch {
+	case errors.Is(err, errCapabilities), errors.Is(err, errRequired):
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+	case errors.Is(err, context.Canceled):
 		c.JSON(499, gin.H{"error": "request canceled"})
-		return
-	}
-	if errors.Is(err, ErrMaxQueue) {
+	case errors.Is(err, ErrMaxQueue):
 		c.JSON(http.StatusServiceUnavailable, gin.H{"error": err.Error()})
-		return
+	case errors.Is(err, os.ErrNotExist):
+		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model %q not found, try pulling it first", name)})
+	default:
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 	}
-	c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 }
diff --git a/server/routes_create_test.go b/server/routes_create_test.go
index 340612822..09521753f 100644
--- a/server/routes_create_test.go
+++ b/server/routes_create_test.go
@@ -2,7 +2,7 @@ package server
 
 import (
 	"bytes"
-	"encoding/binary"
+	"cmp"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -14,8 +14,8 @@ import (
 	"testing"
 
 	"github.com/gin-gonic/gin"
+
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/llm"
 )
 
@@ -30,7 +30,7 @@ func createBinFile(t *testing.T, kv map[string]any, ti []llm.Tensor) string {
 	}
 	defer f.Close()
 
-	if err := llm.NewGGUFV3(binary.LittleEndian).Encode(f, kv, ti); err != nil {
+	if err := llm.WriteGGUF(f, kv, ti); err != nil {
 		t.Fatal(err)
 	}
 
@@ -54,6 +54,8 @@ func (t *responseRecorder) CloseNotify() <-chan bool {
 
 func createRequest(t *testing.T, fn func(*gin.Context), body any) *httptest.ResponseRecorder {
 	t.Helper()
+	// if OLLAMA_MODELS is not set, set it to the temp directory
+	t.Setenv("OLLAMA_MODELS", cmp.Or(os.Getenv("OLLAMA_MODELS"), t.TempDir()))
 
 	w := NewRecorder()
 	c, _ := gin.CreateTestContext(w)
@@ -85,12 +87,13 @@ func checkFileExists(t *testing.T, p string, expect []string) {
 }
 
 func TestCreateFromBin(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 
 	var s Server
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -111,12 +114,13 @@ func TestCreateFromBin(t *testing.T) {
 }
 
 func TestCreateFromModel(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -130,7 +134,7 @@ func TestCreateFromModel(t *testing.T) {
 		filepath.Join(p, "manifests", "registry.ollama.ai", "library", "test", "latest"),
 	})
 
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test2",
 		Modelfile: "FROM test",
 		Stream:    &stream,
@@ -152,12 +156,13 @@ func TestCreateFromModel(t *testing.T) {
 }
 
 func TestCreateRemovesLayers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .Prompt }}", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -177,7 +182,7 @@ func TestCreateRemovesLayers(t *testing.T) {
 		filepath.Join(p, "blobs", "sha256-bc80b03733773e0728011b2f4adf34c458b400e1aad48cb28d61170f3a2ad2d6"),
 	})
 
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .System }} {{ .Prompt }}", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -199,12 +204,13 @@ func TestCreateRemovesLayers(t *testing.T) {
 }
 
 func TestCreateUnsetsSystem(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nSYSTEM Say hi!", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -224,7 +230,7 @@ func TestCreateUnsetsSystem(t *testing.T) {
 		filepath.Join(p, "blobs", "sha256-f29e82a8284dbdf5910b1555580ff60b04238b8da9d5e51159ada67a4d0d5851"),
 	})
 
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nSYSTEM \"\"", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -255,12 +261,13 @@ func TestCreateUnsetsSystem(t *testing.T) {
 }
 
 func TestCreateMergeParameters(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nPARAMETER temperature 1\nPARAMETER top_k 10\nPARAMETER stop USER:\nPARAMETER stop ASSISTANT:", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -281,7 +288,7 @@ func TestCreateMergeParameters(t *testing.T) {
 	})
 
 	// in order to merge parameters, the second model must be created FROM the first
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test2",
 		Modelfile: "FROM test\nPARAMETER temperature 0.6\nPARAMETER top_p 0.7",
 		Stream:    &stream,
@@ -319,7 +326,7 @@ func TestCreateMergeParameters(t *testing.T) {
 	}
 
 	// slices are replaced
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test2",
 		Modelfile: "FROM test\nPARAMETER temperature 0.6\nPARAMETER top_p 0.7\nPARAMETER stop <|endoftext|>",
 		Stream:    &stream,
@@ -358,12 +365,13 @@ func TestCreateMergeParameters(t *testing.T) {
 }
 
 func TestCreateReplacesMessages(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nMESSAGE assistant \"What is my purpose?\"\nMESSAGE user \"You run tests.\"\nMESSAGE assistant \"Oh, my god.\"", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -383,7 +391,7 @@ func TestCreateReplacesMessages(t *testing.T) {
 		filepath.Join(p, "blobs", "sha256-e0e27d47045063ccb167ae852c51d49a98eab33fabaee4633fdddf97213e40b5"),
 	})
 
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test2",
 		Modelfile: "FROM test\nMESSAGE assistant \"You're a test, Harry.\"\nMESSAGE user \"I-I'm a what?\"\nMESSAGE assistant \"A test. And a thumping good one at that, I'd wager.\"",
 		Stream:    &stream,
@@ -434,12 +442,13 @@ func TestCreateReplacesMessages(t *testing.T) {
 }
 
 func TestCreateTemplateSystem(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .Prompt }}\nSYSTEM Say hello!\nTEMPLATE {{ .System }} {{ .Prompt }}\nSYSTEM Say bye!", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -477,15 +486,52 @@ func TestCreateTemplateSystem(t *testing.T) {
 	if string(system) != "Say bye!" {
 		t.Errorf("expected \"Say bye!\", actual %s", system)
 	}
+
+	t.Run("incomplete template", func(t *testing.T) {
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .Prompt", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with unclosed if", func(t *testing.T) {
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ if .Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with undefined function", func(t *testing.T) {
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{  Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
 }
 
 func TestCreateLicenses(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s\nLICENSE MIT\nLICENSE Apache-2.0", createBinFile(t, nil, nil)),
 		Stream:    &stream,
@@ -526,13 +572,14 @@ func TestCreateLicenses(t *testing.T) {
 }
 
 func TestCreateDetectTemplate(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 	var s Server
 
 	t.Run("matched", func(t *testing.T) {
-		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
 			Name: "test",
 			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, llm.KV{
 				"tokenizer.chat_template": "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
@@ -545,14 +592,15 @@ func TestCreateDetectTemplate(t *testing.T) {
 		}
 
 		checkFileExists(t, filepath.Join(p, "blobs", "*"), []string{
-			filepath.Join(p, "blobs", "sha256-2f8e594e6f34b1b4d36a246628eeb3365ce442303d656f1fcc69e821722acea0"),
-			filepath.Join(p, "blobs", "sha256-542b217f179c7825eeb5bca3c77d2b75ed05bafbd3451d9188891a60a85337c6"),
+			filepath.Join(p, "blobs", "sha256-0d79f567714c62c048378f2107fb332dabee0135d080c302d884317da9433cc5"),
+			filepath.Join(p, "blobs", "sha256-35360843d0c84fb1506952a131bbef13cd2bb4a541251f22535170c05b56e672"),
 			filepath.Join(p, "blobs", "sha256-553c4a3f747b3d22a4946875f1cc8ed011c2930d83f864a0c7265f9ec0a20413"),
+			filepath.Join(p, "blobs", "sha256-de3959f841e9ef6b4b6255fa41cb9e0a45da89c3066aa72bdd07a4747f848990"),
 		})
 	})
 
 	t.Run("unmatched", func(t *testing.T) {
-		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
 			Name:      "test",
 			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 			Stream:    &stream,
diff --git a/server/routes_delete_test.go b/server/routes_delete_test.go
index 00303bd17..5a337e794 100644
--- a/server/routes_delete_test.go
+++ b/server/routes_delete_test.go
@@ -8,19 +8,21 @@ import (
 	"path/filepath"
 	"testing"
 
+	"github.com/gin-gonic/gin"
+
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/types/model"
 )
 
 func TestDelete(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
-	envconfig.LoadConfig()
 
 	var s Server
 
-	w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test",
 		Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 	})
@@ -29,7 +31,7 @@ func TestDelete(t *testing.T) {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
 	}
 
-	w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name:      "test2",
 		Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .System }} {{ .Prompt }}", createBinFile(t, nil, nil)),
 	})
@@ -50,7 +52,7 @@ func TestDelete(t *testing.T) {
 		filepath.Join(p, "blobs", "sha256-fe7ac77b725cda2ccad03f88a880ecdfd7a33192d6cae08fce2c0ee1455991ed"),
 	})
 
-	w = createRequest(t, s.DeleteModelHandler, api.DeleteRequest{Name: "test"})
+	w = createRequest(t, s.DeleteHandler, api.DeleteRequest{Name: "test"})
 
 	if w.Code != http.StatusOK {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
@@ -66,7 +68,7 @@ func TestDelete(t *testing.T) {
 		filepath.Join(p, "blobs", "sha256-fe7ac77b725cda2ccad03f88a880ecdfd7a33192d6cae08fce2c0ee1455991ed"),
 	})
 
-	w = createRequest(t, s.DeleteModelHandler, api.DeleteRequest{Name: "test2"})
+	w = createRequest(t, s.DeleteHandler, api.DeleteRequest{Name: "test2"})
 
 	if w.Code != http.StatusOK {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
@@ -77,6 +79,8 @@ func TestDelete(t *testing.T) {
 }
 
 func TestDeleteDuplicateLayers(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	p := t.TempDir()
 	t.Setenv("OLLAMA_MODELS", p)
 	var s Server
@@ -94,11 +98,11 @@ func TestDeleteDuplicateLayers(t *testing.T) {
 	}
 
 	// create a manifest with duplicate layers
-	if err := WriteManifest(n, config, []*Layer{config}); err != nil {
+	if err := WriteManifest(n, config, []Layer{config}); err != nil {
 		t.Fatal(err)
 	}
 
-	w := createRequest(t, s.DeleteModelHandler, api.DeleteRequest{Name: "test"})
+	w := createRequest(t, s.DeleteHandler, api.DeleteRequest{Name: "test"})
 	if w.Code != http.StatusOK {
 		t.Errorf("expected status code 200, actual %d", w.Code)
 	}
diff --git a/server/routes_generate_test.go b/server/routes_generate_test.go
new file mode 100644
index 000000000..480b96721
--- /dev/null
+++ b/server/routes_generate_test.go
@@ -0,0 +1,714 @@
+package server
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/llm"
+)
+
+type mockRunner struct {
+	llm.LlamaServer
+
+	// CompletionRequest is only valid until the next call to Completion
+	llm.CompletionRequest
+	llm.CompletionResponse
+}
+
+func (m *mockRunner) Completion(_ context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
+	m.CompletionRequest = r
+	fn(m.CompletionResponse)
+	return nil
+}
+
+func (mockRunner) Tokenize(_ context.Context, s string) (tokens []int, err error) {
+	for range strings.Fields(s) {
+		tokens = append(tokens, len(tokens))
+	}
+
+	return
+}
+
+func newMockServer(mock *mockRunner) func(gpu.GpuInfoList, string, *llm.GGML, []string, []string, api.Options, int) (llm.LlamaServer, error) {
+	return func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, projectors, system []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		return mock, nil
+	}
+}
+
+func TestGenerateChat(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	mock := mockRunner{
+		CompletionResponse: llm.CompletionResponse{
+			Done:               true,
+			DoneReason:         "stop",
+			PromptEvalCount:    1,
+			PromptEvalDuration: 1,
+			EvalCount:          1,
+			EvalDuration:       1,
+		},
+	}
+
+	s := Server{
+		sched: &Scheduler{
+			pendingReqCh:  make(chan *LlmRequest, 1),
+			finishedReqCh: make(chan *LlmRequest, 1),
+			expiredCh:     make(chan *runnerRef, 1),
+			unloadedCh:    make(chan any, 1),
+			loaded:        make(map[string]*runnerRef),
+			newServerFn:   newMockServer(&mock),
+			getGpuFn:      gpu.GetGPUInfo,
+			getCpuFn:      gpu.GetCPUInfo,
+			reschedDelay:  250 * time.Millisecond,
+			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
+				req.successCh <- &runnerRef{
+					llama: &mock,
+				}
+			},
+		},
+	}
+
+	go s.sched.Run(context.TODO())
+
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model: "test",
+		Modelfile: fmt.Sprintf(`FROM %s
+		TEMPLATE """
+{{- if .System }}System: {{ .System }} {{ end }}
+{{- if .Prompt }}User: {{ .Prompt }} {{ end }}
+{{- if .Response }}Assistant: {{ .Response }} {{ end }}"""
+`, createBinFile(t, llm.KV{
+			"general.architecture":          "llama",
+			"llama.block_count":             uint32(1),
+			"llama.context_length":          uint32(8192),
+			"llama.embedding_length":        uint32(4096),
+			"llama.attention.head_count":    uint32(32),
+			"llama.attention.head_count_kv": uint32(8),
+			"tokenizer.ggml.tokens":         []string{""},
+			"tokenizer.ggml.scores":         []float32{0},
+			"tokenizer.ggml.token_type":     []int32{0},
+		}, []llm.Tensor{
+			{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		})),
+		Stream: &stream,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("missing body", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, nil)
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing model", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{})
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities chat", func(t *testing.T) {
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
+			Model: "bert",
+			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, llm.KV{
+				"general.architecture": "bert",
+				"bert.pooling_type":    uint32(0),
+			}, []llm.Tensor{})),
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		w = createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "bert",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"\"bert\" does not support chat"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("load model", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		var actual api.ChatResponse
+		if err := json.NewDecoder(w.Body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != "test" {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done true, got false")
+		}
+
+		if actual.DoneReason != "load" {
+			t.Errorf("expected done reason load, got %s", actual.DoneReason)
+		}
+	})
+
+	checkChatResponse := func(t *testing.T, body io.Reader, model, content string) {
+		t.Helper()
+
+		var actual api.ChatResponse
+		if err := json.NewDecoder(body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != model {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done false, got true")
+		}
+
+		if actual.DoneReason != "stop" {
+			t.Errorf("expected done reason stop, got %s", actual.DoneReason)
+		}
+
+		if diff := cmp.Diff(actual.Message, api.Message{
+			Role:    "assistant",
+			Content: content,
+		}); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		if actual.PromptEvalCount == 0 {
+			t.Errorf("expected prompt eval count > 0, got 0")
+		}
+
+		if actual.PromptEvalDuration == 0 {
+			t.Errorf("expected prompt eval duration > 0, got 0")
+		}
+
+		if actual.EvalCount == 0 {
+			t.Errorf("expected eval count > 0, got 0")
+		}
+
+		if actual.EvalDuration == 0 {
+			t.Errorf("expected eval duration > 0, got 0")
+		}
+
+		if actual.LoadDuration == 0 {
+			t.Errorf("expected load duration > 0, got 0")
+		}
+
+		if actual.TotalDuration == 0 {
+			t.Errorf("expected total duration > 0, got 0")
+		}
+	}
+
+	mock.CompletionResponse.Content = "Hi!"
+	t.Run("messages", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test", "Hi!")
+	})
+
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:     "test-system",
+		Modelfile: "FROM test\nSYSTEM You are a helpful assistant.",
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("messages with model system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Hi!")
+	})
+
+	mock.CompletionResponse.Content = "Abra kadabra!"
+	t.Run("messages with system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "system", Content: "You can perform magic tricks."},
+				{Role: "user", Content: "Hello!"},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You can perform magic tricks. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	t.Run("messages with interleaved system", func(t *testing.T) {
+		w := createRequest(t, s.ChatHandler, api.ChatRequest{
+			Model: "test-system",
+			Messages: []api.Message{
+				{Role: "user", Content: "Hello!"},
+				{Role: "assistant", Content: "I can help you with that."},
+				{Role: "system", Content: "You can perform magic tricks."},
+				{Role: "user", Content: "Help me write tests."},
+			},
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! Assistant: I can help you with that. System: You can perform magic tricks. User: Help me write tests. "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkChatResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+}
+
+func TestGenerate(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	mock := mockRunner{
+		CompletionResponse: llm.CompletionResponse{
+			Done:               true,
+			DoneReason:         "stop",
+			PromptEvalCount:    1,
+			PromptEvalDuration: 1,
+			EvalCount:          1,
+			EvalDuration:       1,
+		},
+	}
+
+	s := Server{
+		sched: &Scheduler{
+			pendingReqCh:  make(chan *LlmRequest, 1),
+			finishedReqCh: make(chan *LlmRequest, 1),
+			expiredCh:     make(chan *runnerRef, 1),
+			unloadedCh:    make(chan any, 1),
+			loaded:        make(map[string]*runnerRef),
+			newServerFn:   newMockServer(&mock),
+			getGpuFn:      gpu.GetGPUInfo,
+			getCpuFn:      gpu.GetCPUInfo,
+			reschedDelay:  250 * time.Millisecond,
+			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
+				req.successCh <- &runnerRef{
+					llama: &mock,
+				}
+			},
+		},
+	}
+
+	go s.sched.Run(context.TODO())
+
+	w := createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model: "test",
+		Modelfile: fmt.Sprintf(`FROM %s
+		TEMPLATE """
+{{- if .System }}System: {{ .System }} {{ end }}
+{{- if .Prompt }}User: {{ .Prompt }} {{ end }}
+{{- if .Response }}Assistant: {{ .Response }} {{ end }}"""
+`, createBinFile(t, llm.KV{
+			"general.architecture":          "llama",
+			"llama.block_count":             uint32(1),
+			"llama.context_length":          uint32(8192),
+			"llama.embedding_length":        uint32(4096),
+			"llama.attention.head_count":    uint32(32),
+			"llama.attention.head_count_kv": uint32(8),
+			"tokenizer.ggml.tokens":         []string{""},
+			"tokenizer.ggml.scores":         []float32{0},
+			"tokenizer.ggml.token_type":     []int32{0},
+		}, []llm.Tensor{
+			{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+			{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
+		})),
+		Stream: &stream,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("missing body", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, nil)
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing model", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{})
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"model is required"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities generate", func(t *testing.T) {
+		w := createRequest(t, s.CreateHandler, api.CreateRequest{
+			Model: "bert",
+			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, llm.KV{
+				"general.architecture": "bert",
+				"bert.pooling_type":    uint32(0),
+			}, []llm.Tensor{})),
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		w = createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model: "bert",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"\"bert\" does not support generate"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("missing capabilities suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "def add(",
+			Suffix: "    return c",
+		})
+
+		if w.Code != http.StatusBadRequest {
+			t.Errorf("expected status 400, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(w.Body.String(), `{"error":"test does not support insert"}`); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("load model", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model: "test",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		var actual api.GenerateResponse
+		if err := json.NewDecoder(w.Body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != "test" {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done true, got false")
+		}
+
+		if actual.DoneReason != "load" {
+			t.Errorf("expected done reason load, got %s", actual.DoneReason)
+		}
+	})
+
+	checkGenerateResponse := func(t *testing.T, body io.Reader, model, content string) {
+		t.Helper()
+
+		var actual api.GenerateResponse
+		if err := json.NewDecoder(body).Decode(&actual); err != nil {
+			t.Fatal(err)
+		}
+
+		if actual.Model != model {
+			t.Errorf("expected model test, got %s", actual.Model)
+		}
+
+		if !actual.Done {
+			t.Errorf("expected done false, got true")
+		}
+
+		if actual.DoneReason != "stop" {
+			t.Errorf("expected done reason stop, got %s", actual.DoneReason)
+		}
+
+		if actual.Response != content {
+			t.Errorf("expected response %s, got %s", content, actual.Response)
+		}
+
+		if actual.Context == nil {
+			t.Errorf("expected context not nil")
+		}
+
+		if actual.PromptEvalCount == 0 {
+			t.Errorf("expected prompt eval count > 0, got 0")
+		}
+
+		if actual.PromptEvalDuration == 0 {
+			t.Errorf("expected prompt eval duration > 0, got 0")
+		}
+
+		if actual.EvalCount == 0 {
+			t.Errorf("expected eval count > 0, got 0")
+		}
+
+		if actual.EvalDuration == 0 {
+			t.Errorf("expected eval duration > 0, got 0")
+		}
+
+		if actual.LoadDuration == 0 {
+			t.Errorf("expected load duration > 0, got 0")
+		}
+
+		if actual.TotalDuration == 0 {
+			t.Errorf("expected total duration > 0, got 0")
+		}
+	}
+
+	mock.CompletionResponse.Content = "Hi!"
+	t.Run("prompt", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test",
+			Prompt: "Hello!",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test", "Hi!")
+	})
+
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model:     "test-system",
+		Modelfile: "FROM test\nSYSTEM You are a helpful assistant.",
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("prompt with model system", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Hello!",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You are a helpful assistant. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Hi!")
+	})
+
+	mock.CompletionResponse.Content = "Abra kadabra!"
+	t.Run("prompt with system", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Hello!",
+			System: "You can perform magic tricks.",
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "System: You can perform magic tricks. User: Hello! "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	t.Run("prompt with template", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Help me write tests.",
+			System: "You can perform magic tricks.",
+			Template: `{{- if .System }}{{ .System }} {{ end }}
+{{- if .Prompt }}### USER {{ .Prompt }} {{ end }}
+{{- if .Response }}### ASSISTANT {{ .Response }} {{ end }}`,
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "You can perform magic tricks. ### USER Help me write tests. "); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+
+		checkGenerateResponse(t, w.Body, "test-system", "Abra kadabra!")
+	})
+
+	w = createRequest(t, s.CreateHandler, api.CreateRequest{
+		Model: "test-suffix",
+		Modelfile: `FROM test
+TEMPLATE """{{- if .Suffix }}<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+{{- else }}{{ .Prompt }}
+{{- end }}"""`,
+	})
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	t.Run("prompt with suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-suffix",
+			Prompt: "def add(",
+			Suffix: "    return c",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "<PRE> def add( <SUF>    return c <MID>"); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("prompt without suffix", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-suffix",
+			Prompt: "def add(",
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "def add("); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+
+	t.Run("raw", func(t *testing.T) {
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-system",
+			Prompt: "Help me write tests.",
+			Raw:    true,
+			Stream: &stream,
+		})
+
+		if w.Code != http.StatusOK {
+			t.Errorf("expected status 200, got %d", w.Code)
+		}
+
+		if diff := cmp.Diff(mock.CompletionRequest.Prompt, "Help me write tests."); diff != "" {
+			t.Errorf("mismatch (-got +want):\n%s", diff)
+		}
+	})
+}
diff --git a/server/routes_list_test.go b/server/routes_list_test.go
index d04be9d63..56b408300 100644
--- a/server/routes_list_test.go
+++ b/server/routes_list_test.go
@@ -7,13 +7,15 @@ import (
 	"slices"
 	"testing"
 
+	"github.com/gin-gonic/gin"
+
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 )
 
 func TestList(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
-	envconfig.LoadConfig()
 
 	expectNames := []string{
 		"mistral:7b-instruct-q4_0",
@@ -29,13 +31,13 @@ func TestList(t *testing.T) {
 
 	var s Server
 	for _, n := range expectNames {
-		createRequest(t, s.CreateModelHandler, api.CreateRequest{
+		createRequest(t, s.CreateHandler, api.CreateRequest{
 			Name:      n,
 			Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 		})
 	}
 
-	w := createRequest(t, s.ListModelsHandler, nil)
+	w := createRequest(t, s.ListHandler, nil)
 	if w.Code != http.StatusOK {
 		t.Fatalf("expected status code 200, actual %d", w.Code)
 	}
diff --git a/server/routes_test.go b/server/routes_test.go
index 5a5c0fbba..bffcea205 100644
--- a/server/routes_test.go
+++ b/server/routes_test.go
@@ -7,6 +7,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"math"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -18,8 +19,8 @@ import (
 	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@@ -105,6 +106,24 @@ func Test_Routes(t *testing.T) {
 				assert.Empty(t, len(modelList.Models))
 			},
 		},
+		{
+			Name:   "openai empty list",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Equal(t, "list", modelList.Object)
+				assert.Empty(t, modelList.Data)
+			},
+		},
 		{
 			Name:   "Tags Handler (yes tags)",
 			Method: http.MethodGet,
@@ -128,6 +147,25 @@ func Test_Routes(t *testing.T) {
 				assert.Equal(t, "test-model:latest", modelList.Models[0].Name)
 			},
 		},
+		{
+			Name:   "openai list models with tags",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Len(t, modelList.Data, 1)
+				assert.Equal(t, "test-model:latest", modelList.Data[0].Id)
+				assert.Equal(t, "library", modelList.Data[0].OwnedBy)
+			},
+		},
 		{
 			Name:   "Create Model Handler",
 			Method: http.MethodPost,
@@ -216,10 +254,27 @@ func Test_Routes(t *testing.T) {
 				assert.InDelta(t, 0, showResp.ModelInfo["general.parameter_count"], 1e-9, "Parameter count should be 0")
 			},
 		},
+		{
+			Name:   "openai retrieve model handler",
+			Method: http.MethodGet,
+			Path:   "/v1/models/show-model",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var retrieveResp api.RetrieveModelResponse
+				err = json.Unmarshal(body, &retrieveResp)
+				require.NoError(t, err)
+
+				assert.Equal(t, "show-model", retrieveResp.Id)
+				assert.Equal(t, "library", retrieveResp.OwnedBy)
+			},
+		},
 	}
 
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
-	envconfig.LoadConfig()
 
 	s := &Server{}
 	router := s.GenerateRoutes()
@@ -250,7 +305,6 @@ func Test_Routes(t *testing.T) {
 
 func TestCase(t *testing.T) {
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
-	envconfig.LoadConfig()
 
 	cases := []string{
 		"mistral",
@@ -264,7 +318,7 @@ func TestCase(t *testing.T) {
 	var s Server
 	for _, tt := range cases {
 		t.Run(tt, func(t *testing.T) {
-			w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			w := createRequest(t, s.CreateHandler, api.CreateRequest{
 				Name:      tt,
 				Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 				Stream:    &stream,
@@ -280,7 +334,7 @@ func TestCase(t *testing.T) {
 			}
 
 			t.Run("create", func(t *testing.T) {
-				w = createRequest(t, s.CreateModelHandler, api.CreateRequest{
+				w = createRequest(t, s.CreateHandler, api.CreateRequest{
 					Name:      strings.ToUpper(tt),
 					Modelfile: fmt.Sprintf("FROM %s", createBinFile(t, nil, nil)),
 					Stream:    &stream,
@@ -296,7 +350,7 @@ func TestCase(t *testing.T) {
 			})
 
 			t.Run("pull", func(t *testing.T) {
-				w := createRequest(t, s.PullModelHandler, api.PullRequest{
+				w := createRequest(t, s.PullHandler, api.PullRequest{
 					Name:   strings.ToUpper(tt),
 					Stream: &stream,
 				})
@@ -311,7 +365,7 @@ func TestCase(t *testing.T) {
 			})
 
 			t.Run("copy", func(t *testing.T) {
-				w := createRequest(t, s.CopyModelHandler, api.CopyRequest{
+				w := createRequest(t, s.CopyHandler, api.CopyRequest{
 					Source:      tt,
 					Destination: strings.ToUpper(tt),
 				})
@@ -330,11 +384,10 @@ func TestCase(t *testing.T) {
 
 func TestShow(t *testing.T) {
 	t.Setenv("OLLAMA_MODELS", t.TempDir())
-	envconfig.LoadConfig()
 
 	var s Server
 
-	createRequest(t, s.CreateModelHandler, api.CreateRequest{
+	createRequest(t, s.CreateHandler, api.CreateRequest{
 		Name: "show-model",
 		Modelfile: fmt.Sprintf(
 			"FROM %s\nFROM %s",
@@ -343,7 +396,7 @@ func TestShow(t *testing.T) {
 		),
 	})
 
-	w := createRequest(t, s.ShowModelHandler, api.ShowRequest{
+	w := createRequest(t, s.ShowHandler, api.ShowRequest{
 		Name: "show-model",
 	})
 
@@ -364,3 +417,38 @@ func TestShow(t *testing.T) {
 		t.Fatal("Expected projector architecture to be 'clip', but got", resp.ProjectorInfo["general.architecture"])
 	}
 }
+
+func TestNormalize(t *testing.T) {
+	type testCase struct {
+		input []float32
+	}
+
+	testCases := []testCase{
+		{input: []float32{1}},
+		{input: []float32{0, 1, 2, 3}},
+		{input: []float32{0.1, 0.2, 0.3}},
+		{input: []float32{-0.1, 0.2, 0.3, -0.4}},
+		{input: []float32{0, 0, 0}},
+	}
+
+	isNormalized := func(vec []float32) (res bool) {
+		sum := 0.0
+		for _, v := range vec {
+			sum += float64(v * v)
+		}
+		if math.Abs(sum-1) > 1e-6 {
+			return sum == 0
+		} else {
+			return true
+		}
+	}
+
+	for _, tc := range testCases {
+		t.Run("", func(t *testing.T) {
+			normalized := normalize(tc.input)
+			if !isNormalized(normalized) {
+				t.Errorf("Vector %v is not normalized", tc.input)
+			}
+		})
+	}
+}
diff --git a/server/sched.go b/server/sched.go
index 71b535ae2..3c8656ada 100644
--- a/server/sched.go
+++ b/server/sched.go
@@ -5,9 +5,11 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"os"
 	"reflect"
 	"runtime"
 	"sort"
+	"strconv"
 	"strings"
 	"sync"
 	"time"
@@ -24,7 +26,7 @@ type LlmRequest struct {
 	model           *Model
 	opts            api.Options
 	origNumCtx      int // Track the initial ctx request
-	sessionDuration time.Duration
+	sessionDuration *api.Duration
 	successCh       chan *runnerRef
 	errCh           chan error
 	schedAttempts   uint
@@ -56,14 +58,15 @@ var defaultModelsPerGPU = 3
 // we'll back off down to 1 to try to get it to fit
 var defaultParallel = 4
 
-var ErrMaxQueue = fmt.Errorf("server busy, please try again.  maximum pending requests exceeded")
+var ErrMaxQueue = errors.New("server busy, please try again.  maximum pending requests exceeded")
 
 func InitScheduler(ctx context.Context) *Scheduler {
+	maxQueue := envconfig.MaxQueue()
 	sched := &Scheduler{
-		pendingReqCh:  make(chan *LlmRequest, envconfig.MaxQueuedRequests),
-		finishedReqCh: make(chan *LlmRequest, envconfig.MaxQueuedRequests),
-		expiredCh:     make(chan *runnerRef, envconfig.MaxQueuedRequests),
-		unloadedCh:    make(chan interface{}, envconfig.MaxQueuedRequests),
+		pendingReqCh:  make(chan *LlmRequest, maxQueue),
+		finishedReqCh: make(chan *LlmRequest, maxQueue),
+		expiredCh:     make(chan *runnerRef, maxQueue),
+		unloadedCh:    make(chan interface{}, maxQueue),
 		loaded:        make(map[string]*runnerRef),
 		newServerFn:   llm.NewLlamaServer,
 		getGpuFn:      gpu.GetGPUInfo,
@@ -75,7 +78,7 @@ func InitScheduler(ctx context.Context) *Scheduler {
 }
 
 // context must be canceled to decrement ref count and release the runner
-func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
+func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration *api.Duration) (chan *runnerRef, chan error) {
 	if opts.NumCtx < 4 {
 		opts.NumCtx = 4
 	}
@@ -126,17 +129,13 @@ func (s *Scheduler) processPending(ctx context.Context) {
 				slog.Debug("pending request cancelled or timed out, skipping scheduling")
 				continue
 			}
-			numParallel := envconfig.NumParallel
+			numParallel := int(envconfig.NumParallel())
 			// TODO (jmorganca): multimodal models don't support parallel yet
 			// see https://github.com/ollama/ollama/issues/4165
 			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
 				numParallel = 1
 				slog.Warn("multimodal models don't support parallel requests yet")
 			}
-			// Keep NumCtx and numParallel in sync
-			if numParallel > 1 {
-				pending.opts.NumCtx = pending.origNumCtx * numParallel
-			}
 
 			for {
 				var runnerToExpire *runnerRef
@@ -152,7 +151,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						pending.useLoadedRunner(runner, s.finishedReqCh)
 						break
 					}
-				} else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners {
+				} else if envconfig.MaxRunners() > 0 && loadedCount >= int(envconfig.MaxRunners()) {
 					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
 					runnerToExpire = s.findRunnerToUnload()
 				} else {
@@ -165,7 +164,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						gpus = s.getGpuFn()
 					}
 
-					if envconfig.MaxRunners <= 0 {
+					if envconfig.MaxRunners() <= 0 {
 						// No user specified MaxRunners, so figure out what automatic setting to use
 						// If all GPUs have reliable free memory reporting, defaultModelsPerGPU * the number of GPUs
 						// if any GPU has unreliable free memory reporting, 1x the number of GPUs
@@ -177,11 +176,13 @@ func (s *Scheduler) processPending(ctx context.Context) {
 							}
 						}
 						if allReliable {
-							envconfig.MaxRunners = defaultModelsPerGPU * len(gpus)
+							// HACK
+							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(defaultModelsPerGPU*len(gpus)))
 							slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners, "gpu_count", len(gpus))
 						} else {
+							// HACK
+							os.Setenv("OLLAMA_MAX_LOADED_MODELS", strconv.Itoa(len(gpus)))
 							slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency")
-							envconfig.MaxRunners = len(gpus)
 						}
 					}
 
@@ -192,14 +193,20 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}
 
+					// Embedding models should always be loaded with parallel=1
+					if pending.model.CheckCapabilities(CapabilityCompletion) != nil {
+						numParallel = 1
+					}
+
 					// Evaluate if the model will fit in the available system memory, or if we should unload a model first
 					if len(gpus) == 1 && gpus[0].Library == "cpu" {
 						// simplifying assumption of defaultParallel when in CPU mode
 						if numParallel <= 0 {
 							numParallel = defaultParallel
-							pending.opts.NumCtx = pending.origNumCtx * numParallel
 						}
 
+						pending.opts.NumCtx = pending.origNumCtx * numParallel
+
 						if loadedCount == 0 {
 							slog.Debug("cpu mode with first model, loading")
 							s.loadFn(pending, ggml, gpus, numParallel)
@@ -215,9 +222,12 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					} else if loadedCount == 0 {
 						// No models loaded. Load the model but prefer the best fit.
 						slog.Debug("loading first model", "model", pending.model.ModelPath)
-						g := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
+						g := pickBestFullFitByLibrary(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
+						} else {
+							// Only allow partial loads when this is the first model
+							gpus = pickBestPartialFitByLibrary(pending, ggml, gpus, &numParallel)
 						}
 						s.loadFn(pending, ggml, gpus, numParallel)
 						break
@@ -234,7 +244,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
+						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
 							s.loadFn(pending, ggml, fitGpus, numParallel)
@@ -350,7 +360,6 @@ func (s *Scheduler) processCompleted(ctx context.Context) {
 			slog.Debug("runner expired event received", "modelPath", runner.modelPath)
 			runner.refMu.Lock()
 			if runner.refCount > 0 {
-				// Shouldn't happen, but safeguard to ensure no leaked runners
 				slog.Debug("expired event with positive ref count, retrying", "modelPath", runner.modelPath, "refCount", runner.refCount)
 				go func(runner *runnerRef) {
 					// We can't unload yet, but want to as soon as the current request completes
@@ -389,7 +398,9 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 		runner.expireTimer.Stop()
 		runner.expireTimer = nil
 	}
-	runner.sessionDuration = pending.sessionDuration
+	if pending.sessionDuration != nil {
+		runner.sessionDuration = pending.sessionDuration.Duration
+	}
 	pending.successCh <- runner
 	go func() {
 		<-pending.ctx.Done()
@@ -402,12 +413,16 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList,
 	if numParallel < 1 {
 		numParallel = 1
 	}
+	sessionDuration := envconfig.KeepAlive()
+	if req.sessionDuration != nil {
+		sessionDuration = req.sessionDuration.Duration
+	}
 	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
 	if err != nil {
 		// some older models are not compatible with newer versions of llama.cpp
 		// show a generalized compatibility error until there is a better way to
 		// check for model compatibility
-		if errors.Is(llm.ErrUnsupportedFormat, err) || strings.Contains(err.Error(), "failed to load model") {
+		if errors.Is(err, llm.ErrUnsupportedFormat) || strings.Contains(err.Error(), "failed to load model") {
 			err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, req.model.ShortName)
 		}
 		slog.Info("NewLlamaServer failed", "model", req.model.ModelPath, "error", err)
@@ -419,7 +434,7 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList,
 		modelPath:       req.model.ModelPath,
 		llama:           llama,
 		Options:         &req.opts,
-		sessionDuration: req.sessionDuration,
+		sessionDuration: sessionDuration,
 		gpus:            gpus,
 		estimatedVRAM:   llama.EstimatedVRAM(),
 		estimatedTotal:  llama.EstimatedTotal(),
@@ -665,11 +680,12 @@ func (a ByDuration) Less(i, j int) bool {
 // func (a BySize) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 // func (a BySize) Less(i, j int) bool { return a[i].estimatedVRAM < a[j].estimatedVRAM }
 
-// pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
+// pickBestFullFitByLibrary will try to find the optimal placement of the model in the available GPUs where the model fully fits
+// The list of GPUs returned will always be the same brand (library)
 // If the model can not be fit fully within the available GPU(s) nil is returned
 // If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
 // opts.NumCtx accordingly
-func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
+func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
 
 	var numParallelToTry []int
@@ -692,7 +708,7 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP
 		// First attempt to fit the model into a single GPU
 		for _, p := range numParallelToTry {
 			req.opts.NumCtx = req.origNumCtx * p
-			if !envconfig.SchedSpread {
+			if !envconfig.SchedSpread() {
 				for _, g := range sgl {
 					if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
 						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
@@ -720,6 +736,28 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP
 	return nil
 }
 
+// If multiple Libraries are detected, pick the Library which loads the most layers for the model
+func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
+	if *numParallel <= 0 {
+		*numParallel = 1
+		req.opts.NumCtx = req.origNumCtx
+	}
+	byLibrary := gpus.ByLibrary()
+	if len(byLibrary) <= 1 {
+		return gpus
+	}
+	var bestEstimate uint64
+	var bestFit int
+	for i, gl := range byLibrary {
+		_, estimatedVRAM := llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		if estimatedVRAM > bestEstimate {
+			bestEstimate = estimatedVRAM
+			bestFit = i
+		}
+	}
+	return byLibrary[bestFit]
+}
+
 // findRunnerToUnload finds a runner to unload to make room for a new model
 func (s *Scheduler) findRunnerToUnload() *runnerRef {
 	s.loadedMu.Lock()
@@ -763,6 +801,25 @@ func (s *Scheduler) unloadAllRunners() {
 	}
 }
 
+func (s *Scheduler) expireRunner(model *Model) {
+	s.loadedMu.Lock()
+	defer s.loadedMu.Unlock()
+	runner, ok := s.loaded[model.ModelPath]
+	if ok {
+		runner.refMu.Lock()
+		runner.expiresAt = time.Now()
+		if runner.expireTimer != nil {
+			runner.expireTimer.Stop()
+			runner.expireTimer = nil
+		}
+		runner.sessionDuration = 0
+		if runner.refCount <= 0 {
+			s.expiredCh <- runner
+		}
+		runner.refMu.Unlock()
+	}
+}
+
 // If other runners are loaded, make sure the pending request will fit in system memory
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) *runnerRef {
diff --git a/server/sched_test.go b/server/sched_test.go
index be0830a34..fe5647c52 100644
--- a/server/sched_test.go
+++ b/server/sched_test.go
@@ -3,25 +3,25 @@ package server
 import (
 	"bytes"
 	"context"
-	"encoding/binary"
-	"fmt"
+	"errors"
 	"log/slog"
 	"os"
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/require"
+
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/lifecycle"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
 	"github.com/ollama/ollama/llm"
-	"github.com/stretchr/testify/require"
 )
 
-func init() {
+func TestMain(m *testing.M) {
 	os.Setenv("OLLAMA_DEBUG", "1")
 	lifecycle.InitLogging()
+	os.Exit(m.Run())
 }
 
 func TestInitScheduler(t *testing.T) {
@@ -44,11 +44,11 @@ func TestLoad(t *testing.T) {
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2 * time.Second},
 	}
 	// Fail to load model first
 	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
-		return nil, fmt.Errorf("something failed to load model blah")
+		return nil, errors.New("something failed to load model blah")
 	}
 	gpus := gpu.GpuInfoList{}
 	s.load(req, ggml, gpus, 0)
@@ -77,7 +77,7 @@ func TestLoad(t *testing.T) {
 	}
 
 	req.model.ModelPath = "dummy_model_path"
-	server.waitResp = fmt.Errorf("wait failure")
+	server.waitResp = errors.New("wait failure")
 	s.load(req, ggml, gpus, 0)
 	select {
 	case err := <-req.errCh:
@@ -94,7 +94,7 @@ func TestLoad(t *testing.T) {
 	require.Len(t, s.expiredCh, 1)
 }
 
-type bundle struct {
+type reqBundle struct {
 	ctx     context.Context //nolint:containedctx
 	ctxDone func()
 	srv     *mockLlm
@@ -102,23 +102,21 @@ type bundle struct {
 	ggml    *llm.GGML
 }
 
-func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+func (scenario *reqBundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 	return scenario.srv, nil
 }
 
-func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64) *bundle {
-	scenario := &bundle{}
-	scenario.ctx, scenario.ctxDone = context.WithCancel(ctx)
+func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64, duration *api.Duration) *reqBundle {
+	b := &reqBundle{}
+	b.ctx, b.ctxDone = context.WithCancel(ctx)
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), modelName)
 	require.NoError(t, err)
 	defer f.Close()
 
-	gguf := llm.NewGGUFV3(binary.LittleEndian)
-	err = gguf.Encode(f, llm.KV{
+	require.NoError(t, llm.WriteGGUF(f, llm.KV{
 		"general.architecture":          "llama",
-		"general.name":                  "name",
 		"llama.context_length":          uint32(32),
 		"llama.embedding_length":        uint32(4096),
 		"llama.block_count":             uint32(1),
@@ -130,122 +128,160 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 	}, []llm.Tensor{
 		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
 		{Name: "output.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: bytes.NewReader(make([]byte, 32))},
-	})
+	}))
 	require.NoError(t, err)
 
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath, 0)
+	b.ggml, err = llm.LoadModel(model.ModelPath, 0)
 	require.NoError(t, err)
 
-	scenario.req = &LlmRequest{
-		ctx:             scenario.ctx,
+	if duration == nil {
+		duration = &api.Duration{Duration: 5 * time.Millisecond}
+	}
+	b.req = &LlmRequest{
+		ctx:             b.ctx,
 		model:           model,
 		opts:            api.DefaultOptions(),
-		sessionDuration: 5 * time.Millisecond,
+		sessionDuration: duration,
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
 	}
-	scenario.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
-	return scenario
+	b.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
+	return b
 }
 
-func TestRequests(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 10*time.Second)
+func getGpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "metal"}
+	g.TotalMemory = 24 * format.GigaByte
+	g.FreeMemory = 12 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func getCpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "cpu"}
+	g.TotalMemory = 32 * format.GigaByte
+	g.FreeMemory = 26 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func TestRequestsSameModelSameRequest(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
 	defer done()
-
-	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = 5 * time.Millisecond
-	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
-	scenario1b.req.model = scenario1a.req.model
-	scenario1b.ggml = scenario1a.ggml
-	scenario1b.req.sessionDuration = 0
-
-	// simple reload of same model
-	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
-	tmpModel := *scenario1a.req.model
-	scenario2a.req.model = &tmpModel
-	scenario2a.ggml = scenario1a.ggml
-	scenario2a.req.sessionDuration = 5 * time.Millisecond
-
-	// Multiple loaded models
-	scenario3a := newScenario(t, ctx, "ollama-model-3a", 1*format.GigaByte)
-	scenario3b := newScenario(t, ctx, "ollama-model-3b", 24*format.GigaByte)
-	scenario3c := newScenario(t, ctx, "ollama-model-4a", 30)
-	scenario3c.req.opts.NumGPU = 0                           // CPU load, will be allowed
-	scenario3d := newScenario(t, ctx, "ollama-model-3c", 30) // Needs prior unloaded
-
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.getCpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "cpu"}
-		g.TotalMemory = 32 * format.GigaByte
-		g.FreeMemory = 26 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	s.pendingReqCh <- scenario1a.req
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 11, &api.Duration{Duration: 0})
+	b.req.model = a.req.model
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
 	require.Len(t, s.pendingReqCh, 1)
 	s.Run(ctx)
 	select {
-	case resp := <-scenario1a.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1a.req.errCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
 
 	// Same runner as first request due to not needing a reload
-	s.newServerFn = scenario1b.newServer
-	slog.Info("scenario1b")
-	s.pendingReqCh <- scenario1b.req
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario1b.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1b.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		t.Fatal("timeout")
+	}
+}
+
+func TestRequestsSimpleReloadSameModel(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 20, &api.Duration{Duration: 5 * time.Millisecond})
+	tmpModel := *a.req.model
+	b.req.model = &tmpModel
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
+		require.Empty(t, s.pendingReqCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
 
 	// Trigger a reload
-	s.newServerFn = scenario2a.newServer
-	scenario2a.req.model.AdapterPaths = []string{"new"}
-	slog.Info("scenario2a")
-	s.pendingReqCh <- scenario2a.req
+	s.newServerFn = b.newServer
+	b.req.model.AdapterPaths = []string{"new"}
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	// finish first two requests, so model can reload
 	time.Sleep(1 * time.Millisecond)
-	scenario1a.ctxDone()
-	scenario1b.ctxDone()
+	a.ctxDone()
 	select {
-	case resp := <-scenario2a.req.successCh:
-		require.Equal(t, resp.llama, scenario2a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario2a.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
+}
 
-	envconfig.MaxRunners = 1
-	s.newServerFn = scenario3a.newServer
-	slog.Info("scenario3a")
-	s.pendingReqCh <- scenario3a.req
-	// finish prior request, so new model can load
-	time.Sleep(1 * time.Millisecond)
-	scenario2a.ctxDone()
+func TestRequestsMultipleLoadedModels(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+
+	// Multiple loaded models
+	a := newScenarioRequest(t, ctx, "ollama-model-3a", 1*format.GigaByte, nil)
+	b := newScenarioRequest(t, ctx, "ollama-model-3b", 24*format.GigaByte, nil)
+	c := newScenarioRequest(t, ctx, "ollama-model-4a", 30, nil)
+	c.req.opts.NumGPU = 0                                       // CPU load, will be allowed
+	d := newScenarioRequest(t, ctx, "ollama-model-3c", 30, nil) // Needs prior unloaded
+
+	t.Setenv("OLLAMA_MAX_LOADED_MODELS", "1")
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	s.Run(ctx)
 	select {
-	case resp := <-scenario3a.req.successCh:
-		require.Equal(t, resp.llama, scenario3a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3a.req.errCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -253,15 +289,17 @@ func TestRequests(t *testing.T) {
 	require.Len(t, s.loaded, 1)
 	s.loadedMu.Unlock()
 
-	envconfig.MaxRunners = 0
-	s.newServerFn = scenario3b.newServer
-	slog.Info("scenario3b")
-	s.pendingReqCh <- scenario3b.req
+	t.Setenv("OLLAMA_MAX_LOADED_MODELS", "0")
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario3b.req.successCh:
-		require.Equal(t, resp.llama, scenario3b.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3b.req.errCh)
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -270,14 +308,16 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 
 	// This is a CPU load with NumGPU = 0 so it should load
-	s.newServerFn = scenario3c.newServer
-	slog.Info("scenario3c")
-	s.pendingReqCh <- scenario3c.req
+	s.newServerFn = c.newServer
+	slog.Info("c")
+	s.pendingReqCh <- c.req
 	select {
-	case resp := <-scenario3c.req.successCh:
-		require.Equal(t, resp.llama, scenario3c.srv)
+	case resp := <-c.req.successCh:
+		require.Equal(t, resp.llama, c.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3c.req.errCh)
+		require.Empty(t, c.req.errCh)
+	case err := <-c.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -286,25 +326,25 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 
 	// Try to load a model that wont fit
-	s.newServerFn = scenario3d.newServer
-	slog.Info("scenario3d")
+	s.newServerFn = d.newServer
+	slog.Info("d")
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 3)
 	s.loadedMu.Unlock()
-	scenario3a.ctxDone() // Won't help since this one isn't big enough to make room
+	a.ctxDone() // Won't help since this one isn't big enough to make room
 	time.Sleep(2 * time.Millisecond)
-	s.pendingReqCh <- scenario3d.req
+	s.pendingReqCh <- d.req
 	// finish prior request, so new model can load
 	time.Sleep(6 * time.Millisecond)
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
 	s.loadedMu.Unlock()
-	scenario3b.ctxDone()
+	b.ctxDone()
 	select {
-	case resp := <-scenario3d.req.successCh:
-		require.Equal(t, resp.llama, scenario3d.srv)
+	case resp := <-d.req.successCh:
+		require.Equal(t, resp.llama, d.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3d.req.errCh)
+		require.Empty(t, d.req.errCh)
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -314,29 +354,22 @@ func TestRequests(t *testing.T) {
 }
 
 func TestGetRunner(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	ctx, done := context.WithTimeout(context.Background(), 200*time.Millisecond)
 	defer done()
 
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
-	scenario1a.req.sessionDuration = 0
-	scenario1b := newScenario(t, ctx, "ollama-model-1b", 10)
-	scenario1b.req.sessionDuration = 0
-	scenario1c := newScenario(t, ctx, "ollama-model-1c", 10)
-	scenario1c.req.sessionDuration = 0
-	envconfig.MaxQueuedRequests = 1
+	a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1b", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	c := newScenarioRequest(t, ctx, "ollama-model-1c", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	t.Setenv("OLLAMA_MAX_QUEUE", "1")
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	successCh1a, errCh1a := s.GetRunner(scenario1a.ctx, scenario1a.req.model, scenario1a.req.opts, scenario1a.req.sessionDuration)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	successCh1a, errCh1a := s.GetRunner(a.ctx, a.req.model, a.req.opts, a.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
-	slog.Info("scenario1b")
-	successCh1b, errCh1b := s.GetRunner(scenario1b.ctx, scenario1b.req.model, scenario1b.req.opts, scenario1b.req.sessionDuration)
+	slog.Info("b")
+	successCh1b, errCh1b := s.GetRunner(b.ctx, b.req.model, b.req.opts, b.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
 	require.Empty(t, successCh1b)
 	require.Len(t, errCh1b, 1)
@@ -345,22 +378,24 @@ func TestGetRunner(t *testing.T) {
 	s.Run(ctx)
 	select {
 	case resp := <-successCh1a:
-		require.Equal(t, resp.llama, scenario1a.srv)
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, errCh1a)
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
-	scenario1a.ctxDone()
+	a.ctxDone() // Set "a" model to idle so it can unload
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 1)
 	s.loadedMu.Unlock()
 
-	scenario1c.req.model.ModelPath = "bad path"
-	slog.Info("scenario1c")
-	successCh1c, errCh1c := s.GetRunner(scenario1c.ctx, scenario1c.req.model, scenario1c.req.opts, scenario1c.req.sessionDuration)
+	c.req.model.ModelPath = "bad path"
+	slog.Info("c")
+	successCh1c, errCh1c := s.GetRunner(c.ctx, c.req.model, c.req.opts, c.req.sessionDuration)
 	// Starts in pending channel, then should be quickly processsed to return an error
-	time.Sleep(5 * time.Millisecond)
+	time.Sleep(50 * time.Millisecond) // Long enough for the "a" model to expire and unload
 	require.Empty(t, successCh1c)
 	s.loadedMu.Lock()
 	require.Empty(t, s.loaded)
@@ -368,7 +403,53 @@ func TestGetRunner(t *testing.T) {
 	require.Len(t, errCh1c, 1)
 	err = <-errCh1c
 	require.Contains(t, err.Error(), "bad path")
-	scenario1b.ctxDone()
+	b.ctxDone()
+}
+
+func TestExpireRunner(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 20*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	req := &LlmRequest{
+		ctx:             ctx,
+		model:           &Model{ModelPath: "foo"},
+		opts:            api.DefaultOptions(),
+		successCh:       make(chan *runnerRef, 1),
+		errCh:           make(chan error, 1),
+		sessionDuration: &api.Duration{Duration: 2 * time.Minute},
+	}
+
+	var ggml *llm.GGML
+	gpus := gpu.GpuInfoList{}
+	server := &mockLlm{estimatedVRAM: 10, estimatedVRAMByGPU: map[string]uint64{}}
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		return server, nil
+	}
+	s.load(req, ggml, gpus, 0)
+
+	select {
+	case err := <-req.errCh:
+		if err != nil {
+			t.Fatalf("expected no errors when loading, got '%s'", err.Error())
+		}
+	case resp := <-req.successCh:
+		s.loadedMu.Lock()
+		if resp.refCount != uint(1) || len(s.loaded) != 1 {
+			t.Fatalf("expected a model to be loaded")
+		}
+		s.loadedMu.Unlock()
+	}
+
+	s.expireRunner(&Model{ModelPath: "foo"})
+
+	s.finishedReqCh <- req
+	s.processCompleted(ctx)
+
+	s.loadedMu.Lock()
+	if len(s.loaded) != 0 {
+		t.Fatalf("expected model to be unloaded")
+	}
+	s.loadedMu.Unlock()
 }
 
 // TODO - add one scenario that triggers the bogus finished event with positive ref count
@@ -377,7 +458,7 @@ func TestPrematureExpired(t *testing.T) {
 	defer done()
 
 	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
+	scenario1a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, nil)
 	s := InitScheduler(ctx)
 	s.getGpuFn = func() gpu.GpuInfoList {
 		g := gpu.GpuInfo{Library: "metal"}
@@ -399,10 +480,12 @@ func TestPrematureExpired(t *testing.T) {
 		s.loadedMu.Unlock()
 		slog.Info("sending premature expired event now")
 		s.expiredCh <- resp // Shouldn't happen in real life, but make sure its safe
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
-	time.Sleep(scenario1a.req.sessionDuration)
+	time.Sleep(scenario1a.req.sessionDuration.Duration)
 	scenario1a.ctxDone()
 	time.Sleep(20 * time.Millisecond)
 	require.LessOrEqual(t, len(s.finishedReqCh), 1)
@@ -423,7 +506,7 @@ func TestUseLoadedRunner(t *testing.T) {
 		ctx:             ctx,
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2},
 	}
 	finished := make(chan *LlmRequest)
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
@@ -434,6 +517,8 @@ func TestUseLoadedRunner(t *testing.T) {
 	select {
 	case success := <-req.successCh:
 		require.Equal(t, r1, success)
+	case err := <-req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -562,7 +647,7 @@ func TestNeedsReload(t *testing.T) {
 	resp = runner.needsReload(ctx, req)
 	require.True(t, resp)
 	req.opts.NumBatch = runner.Options.NumBatch
-	llm.pingResp = fmt.Errorf("foo")
+	llm.pingResp = errors.New("foo")
 	resp = runner.needsReload(ctx, req)
 	require.True(t, resp)
 	llm.pingResp = nil
@@ -613,8 +698,7 @@ func TestAlreadyCanceled(t *testing.T) {
 	defer done()
 	dctx, done2 := context.WithCancel(ctx)
 	done2()
-	scenario1a := newScenario(t, dctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = 0
+	scenario1a := newScenarioRequest(t, dctx, "ollama-model-1", 10, &api.Duration{Duration: 0})
 	s := InitScheduler(ctx)
 	slog.Info("scenario1a")
 	s.pendingReqCh <- scenario1a.req
@@ -626,11 +710,50 @@ func TestAlreadyCanceled(t *testing.T) {
 	require.Empty(t, scenario1a.req.successCh)
 }
 
+func TestHomogeneousGPUs(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+
+	s.getGpuFn = func() gpu.GpuInfoList {
+		// Set memory values to require the model to be spread
+		gpus := []gpu.GpuInfo{
+			{Library: "cuda"},
+			{Library: "rocm"},
+		}
+		gpus[0].TotalMemory = 1 * format.GibiByte
+		gpus[0].FreeMemory = 256 * format.MebiByte
+		gpus[1].TotalMemory = 1 * format.GibiByte
+		gpus[1].FreeMemory = 256 * format.MebiByte
+		return gpus
+	}
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		require.Len(t, gpus, 1)
+		return a.newServer(gpus, model, ggml, adapters, projectors, opts, numParallel)
+	}
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
+		require.Empty(t, s.pendingReqCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		t.Fatal("timeout")
+	}
+}
+
 type mockLlm struct {
 	pingResp           error
 	waitResp           error
 	completionResp     error
-	embeddingResp      []float64
+	embeddingResp      []float32
 	embeddingRespErr   error
 	tokenizeResp       []int
 	tokenizeRespErr    error
@@ -648,15 +771,19 @@ func (s *mockLlm) WaitUntilRunning(ctx context.Context) error { return s.waitRes
 func (s *mockLlm) Completion(ctx context.Context, req llm.CompletionRequest, fn func(llm.CompletionResponse)) error {
 	return s.completionResp
 }
-func (s *mockLlm) Embedding(ctx context.Context, prompt string) ([]float64, error) {
+
+func (s *mockLlm) Embedding(ctx context.Context, input string) ([]float32, error) {
 	return s.embeddingResp, s.embeddingRespErr
 }
+
 func (s *mockLlm) Tokenize(ctx context.Context, content string) ([]int, error) {
 	return s.tokenizeResp, s.tokenizeRespErr
 }
+
 func (s *mockLlm) Detokenize(ctx context.Context, tokens []int) (string, error) {
 	return s.detokenizeResp, s.detonekizeRespErr
 }
+
 func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
diff --git a/server/sparse_common.go b/server/sparse_common.go
new file mode 100644
index 000000000..c88b2da0b
--- /dev/null
+++ b/server/sparse_common.go
@@ -0,0 +1,8 @@
+//go:build !windows
+
+package server
+
+import "os"
+
+func setSparse(*os.File) {
+}
diff --git a/server/sparse_windows.go b/server/sparse_windows.go
new file mode 100644
index 000000000..f21cbbda7
--- /dev/null
+++ b/server/sparse_windows.go
@@ -0,0 +1,17 @@
+package server
+
+import (
+	"os"
+
+	"golang.org/x/sys/windows"
+)
+
+func setSparse(file *os.File) {
+	// exFat (and other FS types) don't support sparse files, so ignore errors
+	windows.DeviceIoControl( //nolint:errcheck
+		windows.Handle(file.Fd()), windows.FSCTL_SET_SPARSE,
+		nil, 0,
+		nil, 0,
+		nil, nil,
+	)
+}
diff --git a/server/testdata/tools/command-r-plus.gotmpl b/server/testdata/tools/command-r-plus.gotmpl
new file mode 100644
index 000000000..f30124e37
--- /dev/null
+++ b/server/testdata/tools/command-r-plus.gotmpl
@@ -0,0 +1,67 @@
+{{- if or .Tools .System }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>
+{{- if .Tools }}# Safety Preamble
+The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
+
+# System Preamble
+## Basic Rules
+You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
+
+{{ if .System }}# User Preamble
+{{ .System }}
+{{- end }}
+
+## Available Tools
+Here is a list of tools that you have available to you:
+{{- range .Tools }}
+
+```python
+def {{ .Function.Name }}(
+{{- range $name, $property := .Function.Parameters.Properties }}{{ $name }}: {{ $property.Type }}, {{ end }}) -> List[Dict]:
+    """{{ .Function.Description }}
+
+{{- if .Function.Parameters.Properties }}
+
+    Args:
+{{- range $name, $property := .Function.Parameters.Properties }}
+        {{ $name }} ({{ $property.Type }}): {{ $property.Description }}
+{{- end }}
+{{- end }}
+    """
+    pass
+```
+{{- end }}
+{{- else if .System }}{{ .System }}
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- continue }}
+{{- end }}<|START_OF_TURN_TOKEN|>
+{{- if eq .Role "user" }}<|USER_TOKEN|>{{ .Content }}
+{{- else if eq .Role "assistant" }}<|CHATBOT_TOKEN|>
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}
+Action: ```json
+[
+{{- range .ToolCalls }}
+    {
+        "tool_name": "{{ .Function.Name }}",
+        "parameters": {{ .Function.Arguments }}
+    }
+{{- end }}
+]```
+{{ continue }}
+{{ end }}
+{{- else if eq .Role "tool" }}<|SYSTEM_TOKEN|><results>
+{{ .Content }}</results>
+{{- end }}<|END_OF_TURN_TOKEN|>
+{{- end }}
+{{- if .Tools }}<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
+```json
+[
+    {
+        "tool_name": title of the tool in the specification,
+        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
+    }
+]```
+{{- end }}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
\ No newline at end of file
diff --git a/server/testdata/tools/command-r-plus.out b/server/testdata/tools/command-r-plus.out
new file mode 100644
index 000000000..425af75ab
--- /dev/null
+++ b/server/testdata/tools/command-r-plus.out
@@ -0,0 +1,39 @@
+<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># Safety Preamble
+The instructions in this section override those in the task description and style guide sections. Don't answer questions that are harmful or immoral.
+
+# System Preamble
+## Basic Rules
+You are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user's requests, you cite your sources in your answers, according to those instructions.
+
+# User Preamble
+You are a knowledgable assistant. You can answer questions and perform tasks.
+
+## Available Tools
+Here is a list of tools that you have available to you:
+
+```python
+def get_current_weather(format: string, location: string, ) -> List[Dict]:
+    """Get the current weather
+
+    Args:
+        format (string): The temperature unit to use. Infer this from the users location.
+        location (string): The city and state, e.g. San Francisco, CA
+    """
+    pass
+```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in Paris?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
+Action: ```json
+[
+    {
+        "tool_name": "get_current_weather",
+        "parameters": {"format":"celsius","location":"Paris, France"}
+    }
+]```
+<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|><results>
+22</results><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>The current temperature in Paris, France is 22 degrees Celsius.<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>What's the weather like today in San Francisco and Toronto?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write 'Action:' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user's last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:
+```json
+[
+    {
+        "tool_name": title of the tool in the specification,
+        "parameters": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters
+    }
+]```<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
\ No newline at end of file
diff --git a/server/testdata/tools/firefunction.gotmpl b/server/testdata/tools/firefunction.gotmpl
new file mode 100644
index 000000000..312be205c
--- /dev/null
+++ b/server/testdata/tools/firefunction.gotmpl
@@ -0,0 +1,31 @@
+{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
+{{- if .System }}
+{{ .System }}
+{{- end }}
+In addition to plain text responses, you can chose to call one or more of the provided functions.
+
+Use the following rule to decide when to call a function:
+  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
+  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
+
+If you decide to call functions:
+  * prefix function calls with functools marker (no closing marker required)
+  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
+  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
+  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
+  * make sure you pick the right functions that match the user intent
+
+Available functions as JSON spec:
+{{- if .Tools }}
+{{ .Tools }}
+{{- end }}<|eot_id|>
+{{- end }}
+{{- range .Messages }}<|start_header_id|>
+{{- if or (eq .Role "user") (eq .Role "assistant") (eq .Role "tool") }}{{ .Role }}
+{{- end }}<|end_header_id|>
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }} functools[
+{{- range .ToolCalls }}{{ "{" }}"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}{{ "}" }}
+{{- end }}]
+{{- end }}<|eot_id|>
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
\ No newline at end of file
diff --git a/server/testdata/tools/firefunction.out b/server/testdata/tools/firefunction.out
new file mode 100644
index 000000000..be50175ef
--- /dev/null
+++ b/server/testdata/tools/firefunction.out
@@ -0,0 +1,17 @@
+<|start_header_id|>system<|end_header_id|>
+You are a knowledgable assistant. You can answer questions and perform tasks.
+In addition to plain text responses, you can chose to call one or more of the provided functions.
+
+Use the following rule to decide when to call a function:
+  * if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
+  * if you need external information that can be obtained by calling one or more of the provided functions, generate a function calls
+
+If you decide to call functions:
+  * prefix function calls with functools marker (no closing marker required)
+  * all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
+  * follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
+  * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
+  * make sure you pick the right functions that match the user intent
+
+Available functions as JSON spec:
+[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]<|eot_id|><|start_header_id|><|end_header_id|>You are a knowledgable assistant. You can answer questions and perform tasks.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|> functools[{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]<|eot_id|><|start_header_id|>tool<|end_header_id|>22<|eot_id|><|start_header_id|>assistant<|end_header_id|>The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
\ No newline at end of file
diff --git a/server/testdata/tools/llama3-groq-tool-use.gotmpl b/server/testdata/tools/llama3-groq-tool-use.gotmpl
new file mode 100644
index 000000000..45e9b462f
--- /dev/null
+++ b/server/testdata/tools/llama3-groq-tool-use.gotmpl
@@ -0,0 +1,43 @@
+{{- if .Messages }}
+{{- if or .System .Tools }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}
+{{- if .Tools }} You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
+<tool_call>
+{"name": <function-name>,"arguments": <args-dict>}
+</tool_call>
+
+Here are the available tools:
+<tools>
+{{- range .Tools }} {{ .Function }}
+{{- end }} </tools>
+{{- end }}
+{{- end }}<|eot_id|>
+{{- range .Messages }}
+{{- if ne .Role "system" }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ if eq .Role "user" }}{{ .Content }}
+{{- else if eq .Role "assistant" }}
+{{- if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}<tool_call>
+{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{- end }}
+</tool_call>
+{{- end }}
+{{- else if eq .Role "tool" }}<tool_response>
+{{ .Content }}
+</tool_response>
+{{- end }}<|eot_id|>
+{{- end }}
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ else }}
+{{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ end }}{{ .Response }}
+{{- if .Response }}<|eot_id|>
+{{- end }}
\ No newline at end of file
diff --git a/server/testdata/tools/llama3-groq-tool-use.out b/server/testdata/tools/llama3-groq-tool-use.out
new file mode 100644
index 000000000..75a495582
--- /dev/null
+++ b/server/testdata/tools/llama3-groq-tool-use.out
@@ -0,0 +1,24 @@
+<|start_header_id|>system<|end_header_id|>
+
+You are a knowledgable assistant. You can answer questions and perform tasks. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
+<tool_call>
+{"name": <function-name>,"arguments": <args-dict>}
+</tool_call>
+
+Here are the available tools:
+<tools> {"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}} </tools><|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What's the weather like today in Paris?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+<tool_call>
+{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}
+</tool_call><|eot_id|><|start_header_id|>tool<|end_header_id|>
+
+<tool_response>
+22
+</tool_response><|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+The current temperature in Paris, France is 22 degrees Celsius.<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What's the weather like today in San Francisco and Toronto?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
diff --git a/server/testdata/tools/messages.json b/server/testdata/tools/messages.json
new file mode 100644
index 000000000..1a3d1f56c
--- /dev/null
+++ b/server/testdata/tools/messages.json
@@ -0,0 +1,39 @@
+[
+  {
+    "role": "system",
+    "content": "You are a knowledgable assistant. You can answer questions and perform tasks."
+  },
+  {
+    "role": "user",
+    "content": "What's the weather like today in Paris?"
+  },
+  {
+    "role": "assistant",
+    "tool_calls": [
+      {
+        "id": "89a1e453-0bce-4de3-a456-c54bed09c520",
+        "type": "function",
+        "function": {
+          "name": "get_current_weather",
+          "arguments": {
+            "location": "Paris, France",
+            "format": "celsius"
+          }
+        }
+      }
+    ]
+  },
+  {
+    "role": "tool",
+    "tool_call_id": "89a1e453-0bce-4de3-a456-c54bed09c520",
+    "content": "22"
+  },
+  {
+    "role": "assistant",
+    "content": "The current temperature in Paris, France is 22 degrees Celsius."
+  },
+  {
+    "role": "user",
+    "content": "What's the weather like today in San Francisco and Toronto?"
+  }
+]
diff --git a/server/testdata/tools/mistral.gotmpl b/server/testdata/tools/mistral.gotmpl
new file mode 100644
index 000000000..b08d6c2c1
--- /dev/null
+++ b/server/testdata/tools/mistral.gotmpl
@@ -0,0 +1,15 @@
+{{- range $index, $_ := .Messages }}
+{{- if eq .Role "user" }}
+{{- if and (eq (len (slice $.Messages $index)) 1) $.Tools }}[AVAILABLE_TOOLS] {{ $.Tools }}[/AVAILABLE_TOOLS]
+{{- end }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}
+
+{{ end }}{{ .Content }}[/INST]
+{{- else if eq .Role "assistant" }}
+{{- if .Content }} {{ .Content }}</s>
+{{- else if .ToolCalls }}[TOOL_CALLS] [
+{{- range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{- end }}]</s>
+{{- end }}
+{{- else if eq .Role "tool" }}[TOOL_RESULTS] {"content": {{ .Content }}}[/TOOL_RESULTS]
+{{- end }}
+{{- end }}
\ No newline at end of file
diff --git a/server/testdata/tools/mistral.out b/server/testdata/tools/mistral.out
new file mode 100644
index 000000000..31d8cdd62
--- /dev/null
+++ b/server/testdata/tools/mistral.out
@@ -0,0 +1,3 @@
+[INST] What's the weather like today in Paris?[/INST][TOOL_CALLS] [{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]</s>[TOOL_RESULTS] {"content": 22}[/TOOL_RESULTS] The current temperature in Paris, France is 22 degrees Celsius.</s>[AVAILABLE_TOOLS] [{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}][/AVAILABLE_TOOLS][INST] You are a knowledgable assistant. You can answer questions and perform tasks.
+
+What's the weather like today in San Francisco and Toronto?[/INST]
\ No newline at end of file
diff --git a/server/testdata/tools/nemotron.gotmpl b/server/testdata/tools/nemotron.gotmpl
new file mode 100644
index 000000000..1b6b89ecb
--- /dev/null
+++ b/server/testdata/tools/nemotron.gotmpl
@@ -0,0 +1,33 @@
+{{- if (or .Tools .System) }}<extra_id_0>System
+{{ if .System }}{{ .System }}
+
+
+{{ end }}
+{{- if .Tools }}
+{{- range .Tools }}<tool> {{ . }} </tool>{{ end }}
+
+
+{{ end }}
+{{- end }}
+{{- range $i, $m := .Messages }}
+{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+{{- if eq .Role "user" }}<extra_id_1>User
+{{ .Content }}
+{{- if $last }}
+<extra_id_1>Assistant
+{{- end }}
+{{ else if eq .Role "tool" }}<extra_id_1>Tool
+{{ .Content }}
+{{- if $last }}
+<extra_id_1>Assistant
+{{- end }}
+{{ else if eq .Role "assistant" }}<extra_id_1>Assistant
+{{- if .ToolCalls }}
+{{ range .ToolCalls }}<toolcall> {"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}} </toolcall> {{ end }}
+{{ else }}
+{{ .Content }}
+{{- if not $last }}
+{{ end }}
+{{- end }}
+{{- end }}
+{{- end }}
\ No newline at end of file
diff --git a/server/testdata/tools/nemotron.out b/server/testdata/tools/nemotron.out
new file mode 100644
index 000000000..2166b2026
--- /dev/null
+++ b/server/testdata/tools/nemotron.out
@@ -0,0 +1,18 @@
+<extra_id_0>System
+You are a knowledgable assistant. You can answer questions and perform tasks.
+
+
+<tool> {"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}} </tool>
+
+
+<extra_id_1>User
+What's the weather like today in Paris?
+<extra_id_1>Assistant
+<toolcall> {"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}} </toolcall> 
+<extra_id_1>Tool
+22
+<extra_id_1>Assistant
+The current temperature in Paris, France is 22 degrees Celsius.
+<extra_id_1>User
+What's the weather like today in San Francisco and Toronto?
+<extra_id_1>Assistant
diff --git a/server/testdata/tools/tools.json b/server/testdata/tools/tools.json
new file mode 100644
index 000000000..17260bf83
--- /dev/null
+++ b/server/testdata/tools/tools.json
@@ -0,0 +1,30 @@
+[
+  {
+    "type": "function",
+    "function": {
+      "name": "get_current_weather",
+      "description": "Get the current weather",
+      "parameters": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state, e.g. San Francisco, CA"
+          },
+          "format": {
+            "type": "string",
+            "enum": [
+              "celsius",
+              "fahrenheit"
+            ],
+            "description": "The temperature unit to use. Infer this from the users location."
+          }
+        },
+        "required": [
+          "location",
+          "format"
+        ]
+      }
+    }
+  }
+]
diff --git a/server/testdata/tools/xlam.gotmpl b/server/testdata/tools/xlam.gotmpl
new file mode 100644
index 000000000..51513d698
--- /dev/null
+++ b/server/testdata/tools/xlam.gotmpl
@@ -0,0 +1,45 @@
+{{- if .System }}{{ .System }}
+{{ end }}
+{{- range $i, $_ := .Messages }}
+{{- if eq .Role "user" }}### Instruction:
+{{- if and $.Tools (le (len (slice $.Messages $i)) 2) }}
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+{{ $.Tools }}
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+{{ .Content }}
+[END OF QUERY]
+
+
+{{ else }}
+{{ .Content }}
+{{ end }}
+{{- else if .ToolCalls }}### Response:
+{"tool_calls": [{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}{{ end }}]}
+<|EOT|>
+{{ else if eq .Role "assistant" }}### Response:
+{{ .Content }}
+<|EOT|>
+{{ end }}
+{{- end }}### Response:
\ No newline at end of file
diff --git a/server/testdata/tools/xlam.out b/server/testdata/tools/xlam.out
new file mode 100644
index 000000000..a4a9952fc
--- /dev/null
+++ b/server/testdata/tools/xlam.out
@@ -0,0 +1,40 @@
+You are a knowledgable assistant. You can answer questions and perform tasks.
+### Instruction:
+What's the weather like today in Paris?
+### Response:
+{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]}
+<|EOT|>
+### Response:
+The current temperature in Paris, France is 22 degrees Celsius.
+<|EOT|>
+### Instruction:
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+What's the weather like today in San Francisco and Toronto?
+[END OF QUERY]
+
+
+### Response:
\ No newline at end of file
diff --git a/server/upload.go b/server/upload.go
index 73ce78cee..020e89551 100644
--- a/server/upload.go
+++ b/server/upload.go
@@ -12,19 +12,21 @@ import (
 	"net/http"
 	"net/url"
 	"os"
+	"strconv"
 	"sync"
 	"sync/atomic"
 	"time"
 
+	"golang.org/x/sync/errgroup"
+
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
-	"golang.org/x/sync/errgroup"
 )
 
 var blobUploadManager sync.Map
 
 type blobUpload struct {
-	*Layer
+	Layer
 
 	Total     int64
 	Completed atomic.Int64
@@ -43,7 +45,7 @@ type blobUpload struct {
 }
 
 const (
-	numUploadParts          = 64
+	numUploadParts          = 16
 	minUploadPartSize int64 = 100 * format.MegaByte
 	maxUploadPartSize int64 = 1000 * format.MegaByte
 )
@@ -212,7 +214,7 @@ func (b *blobUpload) Run(ctx context.Context, opts *registryOptions) {
 func (b *blobUpload) uploadPart(ctx context.Context, method string, requestURL *url.URL, part *blobUploadPart, opts *registryOptions) error {
 	headers := make(http.Header)
 	headers.Set("Content-Type", "application/octet-stream")
-	headers.Set("Content-Length", fmt.Sprintf("%d", part.Size))
+	headers.Set("Content-Length", strconv.FormatInt(part.Size, 10))
 
 	if method == http.MethodPatch {
 		headers.Set("X-Redirect-Uploads", "1")
@@ -254,7 +256,7 @@ func (b *blobUpload) uploadPart(ctx context.Context, method string, requestURL *
 
 		// retry uploading to the redirect URL
 		for try := range maxRetries {
-			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, nil)
+			err = b.uploadPart(ctx, http.MethodPut, redirectURL, part, &registryOptions{})
 			switch {
 			case errors.Is(err, context.Canceled):
 				return err
@@ -360,7 +362,7 @@ func (p *progressWriter) Rollback() {
 	p.written = 0
 }
 
-func uploadBlob(ctx context.Context, mp ModelPath, layer *Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
+func uploadBlob(ctx context.Context, mp ModelPath, layer Layer, opts *registryOptions, fn func(api.ProgressResponse)) error {
 	requestURL := mp.BaseURL()
 	requestURL = requestURL.JoinPath("v2", mp.GetNamespaceRepository(), "blobs", layer.Digest)
 
diff --git a/template/alfred.gotmpl b/template/alfred.gotmpl
new file mode 100644
index 000000000..86dba48f2
--- /dev/null
+++ b/template/alfred.gotmpl
@@ -0,0 +1,2 @@
+{{- range .Messages }}<start_{{ .Role }}>{{ .Content }}<end_message>
+{{- end }}<start_assistant>
\ No newline at end of file
diff --git a/template/alfred.json b/template/alfred.json
new file mode 100644
index 000000000..edac21afe
--- /dev/null
+++ b/template/alfred.json
@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<start_system>",
+    "<end_message>",
+    "<start_user>",
+    "<start_assistant>"
+  ]
+}
diff --git a/template/alpaca.gotmpl b/template/alpaca.gotmpl
new file mode 100644
index 000000000..004397367
--- /dev/null
+++ b/template/alpaca.gotmpl
@@ -0,0 +1,18 @@
+{{- $system := "" }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- else if eq .Role "user" }}
+{{- if $system }}{{ $system }}
+
+{{ $system = "" }}
+{{- end }}### Instruction:
+{{ .Content }}
+
+{{ else if eq .Role "assistant" }}### Response:
+{{ .Content }}
+
+{{ end }}
+{{- end }}### Response:
diff --git a/template/alpaca.json b/template/alpaca.json
new file mode 100644
index 000000000..eafe2b8ae
--- /dev/null
+++ b/template/alpaca.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "### Instruction:",
+    "### Response"
+  ]
+}
diff --git a/template/chatml.gotmpl b/template/chatml.gotmpl
new file mode 100644
index 000000000..43207ab14
--- /dev/null
+++ b/template/chatml.gotmpl
@@ -0,0 +1,3 @@
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
diff --git a/template/chatml.json b/template/chatml.json
new file mode 100644
index 000000000..7afeb3de5
--- /dev/null
+++ b/template/chatml.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ]
+}
diff --git a/template/chatqa.gotmpl b/template/chatqa.gotmpl
new file mode 100644
index 000000000..0f91e0f01
--- /dev/null
+++ b/template/chatqa.gotmpl
@@ -0,0 +1,7 @@
+{{- range .Messages }}
+{{- if eq .Role "system" }}System:
+{{- else if eq .Role "user" }}User:
+{{- else if eq .Role "assistant" }}Assistant:
+{{- end }} {{ .Content }}
+
+{{ end }}Assistant:
\ No newline at end of file
diff --git a/template/chatqa.json b/template/chatqa.json
new file mode 100644
index 000000000..64dd0f337
--- /dev/null
+++ b/template/chatqa.json
@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "System:",
+    "User:",
+    "Assistant:",
+    "<|begin_of_text|>"
+  ]
+}
diff --git a/template/codellama-70b-instruct.gotmpl b/template/codellama-70b-instruct.gotmpl
new file mode 100644
index 000000000..189315205
--- /dev/null
+++ b/template/codellama-70b-instruct.gotmpl
@@ -0,0 +1,10 @@
+{{- range .Messages }}Source:
+{{- if eq .Role "system" }} system
+{{- else if eq .Role "user" }} user
+{{- else if eq .Role "assistant" }} assistant
+{{- end }}
+
+ {{ .Content }} <step> {{ end }}Source: assistant
+Destination: user
+
+ 
\ No newline at end of file
diff --git a/template/codellama-70b-instruct.json b/template/codellama-70b-instruct.json
new file mode 100644
index 000000000..a56a63f16
--- /dev/null
+++ b/template/codellama-70b-instruct.json
@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "Source:",
+    "Destination:",
+    "<step>"
+  ]
+}
diff --git a/template/falcon-instruct.gotmpl b/template/falcon-instruct.gotmpl
new file mode 100644
index 000000000..b9b51d2c6
--- /dev/null
+++ b/template/falcon-instruct.gotmpl
@@ -0,0 +1,8 @@
+{{- range .Messages }}
+{{- if eq .Role "system" }}System: {{ .Content }}
+{{ continue }}
+{{- else if eq .Role "user" }}User:
+{{- else if eq .Role "assistant" }}Falcon:
+{{- end }}
+{{ .Content }}
+{{ end }}Falcon:
diff --git a/template/falcon-instruct.json b/template/falcon-instruct.json
new file mode 100644
index 000000000..a0da0e813
--- /dev/null
+++ b/template/falcon-instruct.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "User:",
+    "Assistant:"
+  ]
+}
diff --git a/template/gemma-instruct.gotmpl b/template/gemma-instruct.gotmpl
new file mode 100644
index 000000000..cce257192
--- /dev/null
+++ b/template/gemma-instruct.gotmpl
@@ -0,0 +1,16 @@
+{{- $system := "" }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- continue }}
+{{- else if eq .Role "user" }}<start_of_turn>user
+{{- if $system }}
+{{ $system }}
+{{- $system = "" }}
+{{- end }}
+{{- else if eq .Role "assistant" }}<start_of_turn>model
+{{- end }}
+{{ .Content }}<end_of_turn>
+{{ end }}<start_of_turn>model
diff --git a/template/gemma-instruct.json b/template/gemma-instruct.json
new file mode 100644
index 000000000..f4ad415ca
--- /dev/null
+++ b/template/gemma-instruct.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<start_of_turn>",
+    "<end_of_turn>"
+  ]
+}
diff --git a/template/granite-instruct.gotmpl b/template/granite-instruct.gotmpl
new file mode 100644
index 000000000..83634990e
--- /dev/null
+++ b/template/granite-instruct.gotmpl
@@ -0,0 +1,8 @@
+{{- range .Messages }}
+{{- if eq .Role "system" }}System:
+{{- else if eq .Role "user" }}Question:
+{{- else if eq .Role "assistant" }}Answer:
+{{- end }}
+{{ .Content }}
+
+{{ end }}Answer:
diff --git a/template/granite-instruct.json b/template/granite-instruct.json
new file mode 100644
index 000000000..0933e4b52
--- /dev/null
+++ b/template/granite-instruct.json
@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "System:",
+    "Question:",
+    "Answer:"
+  ]
+}
diff --git a/templates/index.json b/template/index.json
similarity index 79%
rename from templates/index.json
rename to template/index.json
index e2d418932..0ce6ac0f2 100644
--- a/templates/index.json
+++ b/template/index.json
@@ -91,6 +91,10 @@
     "template": "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
     "name": "llama3-instruct"
   },
+  {
+    "template": "{{- bos_token }}\n{%- if custom_tools is defined %}\n    {%- set tools = custom_tools %}\n{%- endif %}\n{%- if not tools_in_user_message is defined %}\n    {%- set tools_in_user_message = true %}\n{%- endif %}\n{%- if not date_string is defined %}\n    {%- set date_string = \"26 Jul 2024\" %}\n{%- endif %}\n{%- if not tools is defined %}\n    {%- set tools = none %}\n{%- endif %}\n\n{#- This block extracts the system message, so we can slot it into the right place. #}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content']|trim %}\n    {%- set messages = messages[1:] %}\n{%- else %}\n    {%- set system_message = \"\" %}\n{%- endif %}\n\n{#- System message + builtin tools #}\n{{- \"<|start_header_id|>system<|end_header_id|>\\n\\n\" }}\n{%- if builtin_tools is defined or tools is not none %}\n    {{- \"Environment: ipython\\n\" }}\n{%- endif %}\n{%- if builtin_tools is defined %}\n    {{- \"Tools: \" + builtin_tools | reject('equalto', 'code_interpreter') | join(\", \") + \"\\n\\n\"}}\n{%- endif %}\n{{- \"Cutting Knowledge Date: December 2023\\n\" }}\n{{- \"Today Date: \" + date_string + \"\\n\\n\" }}\n{%- if tools is not none and not tools_in_user_message %}\n    {{- \"You have access to the following functions. To call a function, please respond with JSON for a function call.\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n{%- endif %}\n{{- system_message }}\n{{- \"<|eot_id|>\" }}\n\n{#- Custom tools are passed in a user message with some extra guidance #}\n{%- if tools_in_user_message and not tools is none %}\n    {#- Extract the first user message so we can plug it in here #}\n    {%- if messages | length != 0 %}\n        {%- set first_user_message = messages[0]['content']|trim %}\n        {%- set messages = messages[1:] %}\n    {%- else %}\n        {{- raise_exception(\"Cannot put tools in the first user message when there's no first user message!\") }}\n{%- endif %}\n    {{- '<|start_header_id|>user<|end_header_id|>\\n\\n' -}}\n    {{- \"Given the following functions, please respond with a JSON for a function call \" }}\n    {{- \"with its proper arguments that best answers the given prompt.\\n\\n\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n    {{- first_user_message + \"<|eot_id|>\"}}\n{%- endif %}\n\n{%- for message in messages %}\n    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}\n        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' }}\n    {%- elif 'tool_calls' in message %}\n        {%- if not message.tool_calls|length == 1 %}\n            {{- raise_exception(\"This model only supports single tool-calls at once!\") }}\n        {%- endif %}\n        {%- set tool_call = message.tool_calls[0].function %}\n        {%- if builtin_tools is defined and tool_call.name in builtin_tools %}\n            {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' -}}\n            {{- \"<|python_tag|>\" + tool_call.name + \".call(\" }}\n            {%- for arg_name, arg_val in tool_call.arguments | items %}\n                {{- arg_name + '=\"' + arg_val + '\"' }}\n                {%- if not loop.last %}\n                    {{- \", \" }}\n                {%- endif %}\n                {%- endfor %}\n            {{- \")\" }}\n        {%- else  %}\n            {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' -}}\n            {{- '{\"name\": \"' + tool_call.name + '\", ' }}\n            {{- '\"parameters\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- \"}\" }}\n        {%- endif %}\n        {%- if builtin_tools is defined %}\n            {#- This means we're in ipython mode #}\n            {{- \"<|eom_id|>\" }}\n        {%- else %}\n            {{- \"<|eot_id|>\" }}\n        {%- endif %}\n    {%- elif message.role == \"tool\" or message.role == \"ipython\" %}\n        {{- \"<|start_header_id|>ipython<|end_header_id|>\\n\\n\" }}\n        {%- if message.content is mapping or message.content is iterable %}\n            {{- message.content | tojson }}\n        {%- else %}\n            {{- message.content }}\n        {%- endif %}\n        {{- \"<|eot_id|>\" }}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}\n",
+    "name": "llama3-instruct"
+  },
   {
     "template": "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ 'Question:\n' + message['content'] + '\n\n' }}{% elif message['role'] == 'system' %}\n{{ 'System:\n' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Answer:\n'  + message['content'] + '\n\n' }}{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ 'Answer:\n' }}{% endif %}{% endfor %}",
     "name": "granite-instruct"
diff --git a/template/llama2-chat.gotmpl b/template/llama2-chat.gotmpl
new file mode 100644
index 000000000..5634a0720
--- /dev/null
+++ b/template/llama2-chat.gotmpl
@@ -0,0 +1,14 @@
+{{- $system := "" }}[INST] {{ range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- else if eq .Role "user" }}<<SYS>>
+{{- if $system }}
+{{ $system }}
+{{ $system = "" }}
+{{- end }}<</SYS>>
+
+{{ .Content }} [/INST]
+{{- else if eq .Role "assistant" }} {{ .Content }}</s><s>[INST] {{ end }}
+{{- end }}
\ No newline at end of file
diff --git a/template/llama2-chat.json b/template/llama2-chat.json
new file mode 100644
index 000000000..17590ab4e
--- /dev/null
+++ b/template/llama2-chat.json
@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "[INST]",
+    "[/INST]",
+    "<<SYS>>",
+    "<</SYS>>"
+  ]
+}
diff --git a/template/llama3-instruct.gotmpl b/template/llama3-instruct.gotmpl
new file mode 100644
index 000000000..305ae403b
--- /dev/null
+++ b/template/llama3-instruct.gotmpl
@@ -0,0 +1,5 @@
+{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ .Content }}<|eot_id|>
+{{- end }}<|start_header_id|>assistant<|end_header_id|>
+
diff --git a/template/llama3-instruct.json b/template/llama3-instruct.json
new file mode 100644
index 000000000..c4e9d4483
--- /dev/null
+++ b/template/llama3-instruct.json
@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "<|eot_id|>"
+  ]
+}
diff --git a/template/magicoder.gotmpl b/template/magicoder.gotmpl
new file mode 100644
index 000000000..e5ee0e470
--- /dev/null
+++ b/template/magicoder.gotmpl
@@ -0,0 +1,17 @@
+{{- $system := "" }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- continue }}
+{{- else if eq .Role "user" }}
+{{- if $system }}{{ $system }}
+
+{{ $system = "" }}
+{{- end }}@@ Instruction
+{{- else if eq .Role "assistant" }}@@ Response
+{{- end }}
+{{ .Content }}
+
+{{ end }}@@ Response
diff --git a/template/magicoder.json b/template/magicoder.json
new file mode 100644
index 000000000..6f67cab0f
--- /dev/null
+++ b/template/magicoder.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "@@ Instruction",
+    "@@ Response"
+  ]
+}
diff --git a/template/mistral-instruct.gotmpl b/template/mistral-instruct.gotmpl
new file mode 100644
index 000000000..7a6ecdfd8
--- /dev/null
+++ b/template/mistral-instruct.gotmpl
@@ -0,0 +1,6 @@
+[INST] {{ range $index, $_ := .Messages }}
+{{- if eq .Role "system" }}{{ .Content }}
+
+{{ else if eq .Role "user" }}{{ .Content }}[/INST]
+{{- else if eq .Role "assistant" }} {{ .Content }}</s>[INST] {{ end }}
+{{- end }}
\ No newline at end of file
diff --git a/template/mistral-instruct.json b/template/mistral-instruct.json
new file mode 100644
index 000000000..7afeb3de5
--- /dev/null
+++ b/template/mistral-instruct.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ]
+}
diff --git a/template/openchat.gotmpl b/template/openchat.gotmpl
new file mode 100644
index 000000000..66a4d687b
--- /dev/null
+++ b/template/openchat.gotmpl
@@ -0,0 +1,6 @@
+{{- range .Messages }}GPT4 Correct
+{{- if eq .Role "system" }} System:
+{{- else if eq .Role "user" }} User:
+{{- else if eq .Role "assistant" }} Assistant:
+{{- end }} {{ .Content }}<|end_of_turn|>
+{{- end }}GPT4 Correct Assistant:
\ No newline at end of file
diff --git a/template/openchat.json b/template/openchat.json
new file mode 100644
index 000000000..0edc341fd
--- /dev/null
+++ b/template/openchat.json
@@ -0,0 +1,5 @@
+{
+  "stop": [
+    "<|end_of_turn|>"
+  ]
+}
diff --git a/template/phi-3.gotmpl b/template/phi-3.gotmpl
new file mode 100644
index 000000000..abec2137d
--- /dev/null
+++ b/template/phi-3.gotmpl
@@ -0,0 +1,3 @@
+{{- range .Messages }}<|{{ .Role }}|>
+{{ .Content }}<|end|>
+{{ end }}<|assistant|>
diff --git a/template/phi-3.json b/template/phi-3.json
new file mode 100644
index 000000000..27bf7664b
--- /dev/null
+++ b/template/phi-3.json
@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<|end|>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>"
+  ]
+}
diff --git a/template/solar-instruct.gotmpl b/template/solar-instruct.gotmpl
new file mode 100644
index 000000000..263bde804
--- /dev/null
+++ b/template/solar-instruct.gotmpl
@@ -0,0 +1,11 @@
+{{- range .Messages }}
+{{- if eq .Role "system" }}### System:
+{{- else if eq .Role "user" }}### User:
+{{- else if eq .Role "assistant" }}### Assistant:
+{{ .Content }}</s>
+
+{{ continue }}
+{{- end }}
+{{ .Content }}
+
+{{ end }}### Assistant:
diff --git a/template/solar-instruct.json b/template/solar-instruct.json
new file mode 100644
index 000000000..7b7a90504
--- /dev/null
+++ b/template/solar-instruct.json
@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "### System:",
+    "### User:",
+    "### Assistant"
+  ]
+}
diff --git a/template/starcoder2-instruct.gotmpl b/template/starcoder2-instruct.gotmpl
new file mode 100644
index 000000000..7963b4f9a
--- /dev/null
+++ b/template/starcoder2-instruct.gotmpl
@@ -0,0 +1,18 @@
+{{- $system := "" }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- else if eq .Role "user" }}
+{{- if $system }}{{ $system }}
+
+{{ $system = "" }}
+{{- end }}### Instruction
+{{ .Content }}
+
+{{ else if eq .Role "assistant" }}### Response
+{{ .Content }}<|endoftext|>
+
+{{ end }}
+{{- end }}### Response
diff --git a/template/starcoder2-instruct.json b/template/starcoder2-instruct.json
new file mode 100644
index 000000000..313489086
--- /dev/null
+++ b/template/starcoder2-instruct.json
@@ -0,0 +1,7 @@
+{
+  "stop": [
+    "### Instruction",
+    "### Response",
+    "<|endoftext|>"
+  ]
+}
diff --git a/template/template.go b/template/template.go
new file mode 100644
index 000000000..5dc484f47
--- /dev/null
+++ b/template/template.go
@@ -0,0 +1,447 @@
+package template
+
+import (
+	"bytes"
+	"embed"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"slices"
+	"strings"
+	"sync"
+	"text/template"
+	"text/template/parse"
+
+	"github.com/agnivade/levenshtein"
+	"golang.org/x/exp/maps"
+
+	"github.com/ollama/ollama/api"
+)
+
+//go:embed index.json
+var indexBytes []byte
+
+//go:embed *.gotmpl
+//go:embed *.json
+var templatesFS embed.FS
+
+var templatesOnce = sync.OnceValues(func() ([]*named, error) {
+	var templates []*named
+	if err := json.Unmarshal(indexBytes, &templates); err != nil {
+		return nil, err
+	}
+
+	for _, t := range templates {
+		bts, err := templatesFS.ReadFile(t.Name + ".gotmpl")
+		if err != nil {
+			return nil, err
+		}
+
+		// normalize line endings
+		t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
+
+		params, err := templatesFS.ReadFile(t.Name + ".json")
+		if err != nil {
+			continue
+		}
+
+		if err := json.Unmarshal(params, &t.Parameters); err != nil {
+			return nil, err
+		}
+	}
+
+	return templates, nil
+})
+
+type named struct {
+	Name     string `json:"name"`
+	Template string `json:"template"`
+	Bytes    []byte
+
+	Parameters *struct {
+		Stop []string `json:"stop"`
+	}
+}
+
+func (t named) Reader() io.Reader {
+	return bytes.NewReader(t.Bytes)
+}
+
+func Named(s string) (*named, error) {
+	templates, err := templatesOnce()
+	if err != nil {
+		return nil, err
+	}
+
+	var template *named
+	score := math.MaxInt
+	for _, t := range templates {
+		if s := levenshtein.ComputeDistance(s, t.Template); s < score {
+			score = s
+			template = t
+		}
+	}
+
+	if score < 100 {
+		return template, nil
+	}
+
+	return nil, errors.New("no matching template found")
+}
+
+var DefaultTemplate, _ = Parse("{{ .Prompt }}")
+
+type Template struct {
+	*template.Template
+	raw string
+}
+
+// response is a template node that can be added to templates that don't already have one
+var response = parse.ActionNode{
+	NodeType: parse.NodeAction,
+	Pipe: &parse.PipeNode{
+		NodeType: parse.NodePipe,
+		Cmds: []*parse.CommandNode{
+			{
+				NodeType: parse.NodeCommand,
+				Args: []parse.Node{
+					&parse.FieldNode{
+						NodeType: parse.NodeField,
+						Ident:    []string{"Response"},
+					},
+				},
+			},
+		},
+	},
+}
+
+var funcs = template.FuncMap{
+	"json": func(v any) string {
+		b, _ := json.Marshal(v)
+		return string(b)
+	},
+}
+
+func Parse(s string) (*Template, error) {
+	tmpl := template.New("").Option("missingkey=zero").Funcs(funcs)
+
+	tmpl, err := tmpl.Parse(s)
+	if err != nil {
+		return nil, err
+	}
+
+	t := Template{Template: tmpl, raw: s}
+	if vars := t.Vars(); !slices.Contains(vars, "messages") && !slices.Contains(vars, "response") {
+		// touch up the template and append {{ .Response }}
+		tmpl.Tree.Root.Nodes = append(tmpl.Tree.Root.Nodes, &response)
+	}
+
+	return &t, nil
+}
+
+func (t *Template) String() string {
+	return t.raw
+}
+
+func (t *Template) Vars() []string {
+	var vars []string
+	for _, tt := range t.Templates() {
+		for _, n := range tt.Root.Nodes {
+			vars = append(vars, Identifiers(n)...)
+		}
+	}
+
+	set := make(map[string]struct{})
+	for _, n := range vars {
+		set[strings.ToLower(n)] = struct{}{}
+	}
+
+	vars = maps.Keys(set)
+	slices.Sort(vars)
+	return vars
+}
+
+type Values struct {
+	Messages []api.Message
+	api.Tools
+	Prompt string
+	Suffix string
+
+	// forceLegacy is a flag used to test compatibility with legacy templates
+	forceLegacy bool
+}
+
+func (t *Template) Subtree(fn func(parse.Node) bool) *template.Template {
+	var walk func(parse.Node) parse.Node
+	walk = func(n parse.Node) parse.Node {
+		if fn(n) {
+			return n
+		}
+
+		switch t := n.(type) {
+		case *parse.ListNode:
+			for _, c := range t.Nodes {
+				if n := walk(c); n != nil {
+					return n
+				}
+			}
+		case *parse.BranchNode:
+			for _, n := range []*parse.ListNode{t.List, t.ElseList} {
+				if n != nil {
+					if n := walk(n); n != nil {
+						return n
+					}
+				}
+			}
+		case *parse.IfNode:
+			return walk(&t.BranchNode)
+		case *parse.WithNode:
+			return walk(&t.BranchNode)
+		case *parse.RangeNode:
+			return walk(&t.BranchNode)
+		}
+
+		return nil
+	}
+
+	if n := walk(t.Tree.Root); n != nil {
+		return (&template.Template{
+			Tree: &parse.Tree{
+				Root: &parse.ListNode{
+					Nodes: []parse.Node{n},
+				},
+			},
+		}).Funcs(funcs)
+	}
+
+	return nil
+}
+
+func (t *Template) Execute(w io.Writer, v Values) error {
+	system, messages := collate(v.Messages)
+	if v.Prompt != "" && v.Suffix != "" {
+		return t.Template.Execute(w, map[string]any{
+			"Prompt":   v.Prompt,
+			"Suffix":   v.Suffix,
+			"Response": "",
+		})
+	} else if !v.forceLegacy && slices.Contains(t.Vars(), "messages") {
+		return t.Template.Execute(w, map[string]any{
+			"System":   system,
+			"Messages": messages,
+			"Tools":    v.Tools,
+			"Response": "",
+		})
+	}
+
+	system = ""
+	var b bytes.Buffer
+	var prompt, response string
+	for _, m := range messages {
+		execute := func() error {
+			if err := t.Template.Execute(&b, map[string]any{
+				"System":   system,
+				"Prompt":   prompt,
+				"Response": response,
+			}); err != nil {
+				return err
+			}
+
+			system = ""
+			prompt = ""
+			response = ""
+			return nil
+		}
+
+		switch m.Role {
+		case "system":
+			if prompt != "" || response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			system = m.Content
+		case "user":
+			if response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			prompt = m.Content
+		case "assistant":
+			response = m.Content
+		}
+	}
+
+	var cut bool
+	nodes := deleteNode(t.Template.Root.Copy(), func(n parse.Node) bool {
+		if field, ok := n.(*parse.FieldNode); ok && slices.Contains(field.Ident, "Response") {
+			cut = true
+			return false
+		}
+
+		return cut
+	})
+
+	tree := parse.Tree{Root: nodes.(*parse.ListNode)}
+	if err := template.Must(template.New("").AddParseTree("", &tree)).Execute(&b, map[string]any{
+		"System":   system,
+		"Prompt":   prompt,
+		"Response": response,
+	}); err != nil {
+		return err
+	}
+
+	_, err := io.Copy(w, &b)
+	return err
+}
+
+// collate messages based on role. consecutive messages of the same role are merged
+// into a single message. collate also collects and returns all system messages.
+// collate mutates message content adding image tags ([img-%d]) as needed
+func collate(msgs []api.Message) (string, []*api.Message) {
+	var n int
+
+	var system []string
+	var collated []*api.Message
+	for i := range msgs {
+		msg := msgs[i]
+		for range msg.Images {
+			imageTag := fmt.Sprintf("[img-%d]", n)
+			if !strings.Contains(msg.Content, "[img]") {
+				msg.Content = strings.TrimSpace("[img] " + msg.Content)
+			}
+
+			msg.Content = strings.Replace(msg.Content, "[img]", imageTag, 1)
+			n++
+		}
+
+		if msg.Role == "system" {
+			system = append(system, msg.Content)
+		}
+
+		if len(collated) > 0 && collated[len(collated)-1].Role == msg.Role {
+			collated[len(collated)-1].Content += "\n\n" + msg.Content
+		} else {
+			collated = append(collated, &msg)
+		}
+	}
+
+	return strings.Join(system, "\n\n"), collated
+}
+
+// Identifiers walks the node tree returning any identifiers it finds along the way
+func Identifiers(n parse.Node) []string {
+	switch n := n.(type) {
+	case *parse.ListNode:
+		var names []string
+		for _, n := range n.Nodes {
+			names = append(names, Identifiers(n)...)
+		}
+
+		return names
+	case *parse.TemplateNode:
+		return Identifiers(n.Pipe)
+	case *parse.ActionNode:
+		return Identifiers(n.Pipe)
+	case *parse.BranchNode:
+		names := Identifiers(n.Pipe)
+		for _, n := range []*parse.ListNode{n.List, n.ElseList} {
+			if n != nil {
+				names = append(names, Identifiers(n)...)
+			}
+		}
+		return names
+	case *parse.IfNode:
+		return Identifiers(&n.BranchNode)
+	case *parse.RangeNode:
+		return Identifiers(&n.BranchNode)
+	case *parse.WithNode:
+		return Identifiers(&n.BranchNode)
+	case *parse.PipeNode:
+		var names []string
+		for _, c := range n.Cmds {
+			for _, a := range c.Args {
+				names = append(names, Identifiers(a)...)
+			}
+		}
+		return names
+	case *parse.FieldNode:
+		return n.Ident
+	case *parse.VariableNode:
+		return n.Ident
+	}
+
+	return nil
+}
+
+// deleteNode walks the node list and deletes nodes that match the predicate
+// this is currently to remove the {{ .Response }} node from templates
+func deleteNode(n parse.Node, fn func(parse.Node) bool) parse.Node {
+	var walk func(n parse.Node) parse.Node
+	walk = func(n parse.Node) parse.Node {
+		if fn(n) {
+			return nil
+		}
+
+		switch t := n.(type) {
+		case *parse.ListNode:
+			var nodes []parse.Node
+			for _, c := range t.Nodes {
+				if n := walk(c); n != nil {
+					nodes = append(nodes, n)
+				}
+			}
+
+			t.Nodes = nodes
+			return t
+		case *parse.IfNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.WithNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.RangeNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.BranchNode:
+			t.List = walk(t.List).(*parse.ListNode)
+			if t.ElseList != nil {
+				t.ElseList = walk(t.ElseList).(*parse.ListNode)
+			}
+		case *parse.ActionNode:
+			n := walk(t.Pipe)
+			if n == nil {
+				return nil
+			}
+
+			t.Pipe = n.(*parse.PipeNode)
+		case *parse.PipeNode:
+			var commands []*parse.CommandNode
+			for _, c := range t.Cmds {
+				var args []parse.Node
+				for _, a := range c.Args {
+					if n := walk(a); n != nil {
+						args = append(args, n)
+					}
+				}
+
+				if len(args) == 0 {
+					return nil
+				}
+
+				c.Args = args
+				commands = append(commands, c)
+			}
+
+			if len(commands) == 0 {
+				return nil
+			}
+
+			t.Cmds = commands
+		}
+
+		return n
+	}
+
+	return walk(n)
+}
diff --git a/template/template_test.go b/template/template_test.go
new file mode 100644
index 000000000..113e06838
--- /dev/null
+++ b/template/template_test.go
@@ -0,0 +1,417 @@
+package template
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"io"
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/llm"
+)
+
+func TestNamed(t *testing.T) {
+	f, err := os.Open(filepath.Join("testdata", "templates.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	scanner := bufio.NewScanner(f)
+	for scanner.Scan() {
+		var ss map[string]string
+		if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil {
+			t.Fatal(err)
+		}
+
+		for k, v := range ss {
+			t.Run(k, func(t *testing.T) {
+				kv := llm.KV{"tokenizer.chat_template": v}
+				s := kv.ChatTemplate()
+				r, err := Named(s)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if r.Name != k {
+					t.Errorf("expected %q, got %q", k, r.Name)
+				}
+
+				var b bytes.Buffer
+				if _, err := io.Copy(&b, r.Reader()); err != nil {
+					t.Fatal(err)
+				}
+
+				tmpl, err := Parse(b.String())
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if tmpl.Tree.Root.String() == "" {
+					t.Errorf("empty %s template", k)
+				}
+			})
+		}
+	}
+}
+
+func TestTemplate(t *testing.T) {
+	cases := make(map[string][]api.Message)
+	for _, mm := range [][]api.Message{
+		{
+			{Role: "user", Content: "Hello, how are you?"},
+		},
+		{
+			{Role: "user", Content: "Hello, how are you?"},
+			{Role: "assistant", Content: "I'm doing great. How can I help you today?"},
+			{Role: "user", Content: "I'd like to show off how chat templating works!"},
+		},
+		{
+			{Role: "system", Content: "You are a helpful assistant."},
+			{Role: "user", Content: "Hello, how are you?"},
+			{Role: "assistant", Content: "I'm doing great. How can I help you today?"},
+			{Role: "user", Content: "I'd like to show off how chat templating works!"},
+		},
+	} {
+		var roles []string
+		for _, m := range mm {
+			roles = append(roles, m.Role)
+		}
+
+		cases[strings.Join(roles, "-")] = mm
+	}
+
+	matches, err := filepath.Glob("*.gotmpl")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	for _, match := range matches {
+		t.Run(match, func(t *testing.T) {
+			bts, err := os.ReadFile(match)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			tmpl, err := Parse(string(bts))
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			for n, tt := range cases {
+				var actual bytes.Buffer
+				t.Run(n, func(t *testing.T) {
+					if err := tmpl.Execute(&actual, Values{Messages: tt}); err != nil {
+						t.Fatal(err)
+					}
+
+					expect, err := os.ReadFile(filepath.Join("testdata", match, n))
+					if err != nil {
+						t.Fatal(err)
+					}
+
+					bts := actual.Bytes()
+
+					if slices.Contains([]string{"chatqa.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && bts[len(bts)-1] == ' ' {
+						t.Log("removing trailing space from output")
+						bts = bts[:len(bts)-1]
+					}
+
+					if diff := cmp.Diff(bts, expect); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+
+				t.Run("legacy", func(t *testing.T) {
+					t.Skip("legacy outputs are currently default outputs")
+					var legacy bytes.Buffer
+					if err := tmpl.Execute(&legacy, Values{Messages: tt, forceLegacy: true}); err != nil {
+						t.Fatal(err)
+					}
+
+					legacyBytes := legacy.Bytes()
+					if slices.Contains([]string{"chatqa.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && legacyBytes[len(legacyBytes)-1] == ' ' {
+						t.Log("removing trailing space from legacy output")
+						legacyBytes = legacyBytes[:len(legacyBytes)-1]
+					} else if slices.Contains([]string{"codellama-70b-instruct.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl"}, match) {
+						t.Skip("legacy outputs cannot be compared to messages outputs")
+					}
+
+					if diff := cmp.Diff(legacyBytes, actual.Bytes()); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+			}
+		})
+	}
+}
+
+func TestParse(t *testing.T) {
+	cases := []struct {
+		template string
+		vars     []string
+	}{
+		{"{{ .Prompt }}", []string{"prompt", "response"}},
+		{"{{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system"}},
+		{"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}},
+		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system", "tools"}},
+		{"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}},
+		{`{{- range .Messages }}
+{{- if eq .Role "system" }}SYSTEM:
+{{- else if eq .Role "user" }}USER:
+{{- else if eq .Role "assistant" }}ASSISTANT:
+{{- end }} {{ .Content }}
+{{- end }}`, []string{"content", "messages", "role"}},
+		{`{{- if .Messages }}
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ else -}}
+{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ .Response }}<|im_end|>
+{{- end -}}`, []string{"content", "messages", "prompt", "response", "role", "system"}},
+	}
+
+	for _, tt := range cases {
+		t.Run("", func(t *testing.T) {
+			tmpl, err := Parse(tt.template)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if diff := cmp.Diff(tmpl.Vars(), tt.vars); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestExecuteWithMessages(t *testing.T) {
+	type template struct {
+		name     string
+		template string
+	}
+	cases := []struct {
+		name      string
+		templates []template
+		values    Values
+		expected  string
+	}{
+		{
+			"mistral",
+			[]template{
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`[INST] Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
+		},
+		{
+			"mistral system",
+			[]template{
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "system", Content: "You are a helpful assistant!"},
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`[INST] You are a helpful assistant!
+
+Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
+		},
+		{
+			"mistral assistant",
+			[]template{
+				{"no response", `[INST] {{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `
+{{- range $i, $m := .Messages }}
+{{- if eq .Role "user" }}[INST] {{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}{{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+					{Role: "assistant", Content: "My name is Ollama and I"},
+				},
+			},
+			`[INST] Hello friend![/INST] Hello human![INST] What is your name?[/INST] My name is Ollama and I`,
+		},
+		{
+			"chatml",
+			[]template{
+				// this does not have a "no response" test because it's impossible to render the same output
+				{"response", `{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ .Response }}<|im_end|>
+`},
+				{"messages", `
+{{- range $index, $_ := .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "system", Content: "You are a helpful assistant!"},
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+				},
+			},
+			`<|im_start|>system
+You are a helpful assistant!<|im_end|>
+<|im_start|>user
+Hello friend!<|im_end|>
+<|im_start|>assistant
+Hello human!<|im_end|>
+<|im_start|>user
+What is your name?<|im_end|>
+<|im_start|>assistant
+`,
+		},
+		{
+			"moondream",
+			[]template{
+				// this does not have a "no response" test because it's impossible to render the same output
+				{"response", `{{ if .Prompt }}Question: {{ .Prompt }}
+
+{{ end }}Answer: {{ .Response }}
+
+`},
+				{"messages", `
+{{- range .Messages }}
+{{- if eq .Role "user" }}Question: {{ .Content }}
+
+{{ else if eq .Role "assistant" }}Answer: {{ .Content }}
+
+{{ end }}
+{{- end }}Answer: `},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "What's in this image?", Images: []api.ImageData{[]byte("")}},
+					{Role: "assistant", Content: "It's a hot dog."},
+					{Role: "user", Content: "What's in _this_ image?"},
+					{Role: "user", Images: []api.ImageData{[]byte("")}},
+					{Role: "user", Content: "Is it a hot dog?"},
+				},
+			},
+			`Question: [img-0] What's in this image?
+
+Answer: It's a hot dog.
+
+Question: What's in _this_ image?
+
+[img-1]
+
+Is it a hot dog?
+
+Answer: `,
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			for _, ttt := range tt.templates {
+				t.Run(ttt.name, func(t *testing.T) {
+					tmpl, err := Parse(ttt.template)
+					if err != nil {
+						t.Fatal(err)
+					}
+
+					var b bytes.Buffer
+					if err := tmpl.Execute(&b, tt.values); err != nil {
+						t.Fatal(err)
+					}
+
+					if diff := cmp.Diff(b.String(), tt.expected); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+			}
+		})
+	}
+}
+
+func TestExecuteWithSuffix(t *testing.T) {
+	tmpl, err := Parse(`{{- if .Suffix }}<PRE> {{ .Prompt }} <SUF>{{ .Suffix }} <MID>
+{{- else }}{{ .Prompt }}
+{{- end }}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	cases := []struct {
+		name   string
+		values Values
+		expect string
+	}{
+		{
+			"message", Values{Messages: []api.Message{{Role: "user", Content: "hello"}}}, "hello",
+		},
+		{
+			"prompt suffix", Values{Prompt: "def add(", Suffix: "return x"}, "<PRE> def add( <SUF>return x <MID>",
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			var b bytes.Buffer
+			if err := tmpl.Execute(&b, tt.values); err != nil {
+				t.Fatal(err)
+			}
+
+			if diff := cmp.Diff(b.String(), tt.expect); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
diff --git a/template/testdata/alfred.gotmpl/system-user-assistant-user b/template/testdata/alfred.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..03e23ea9c
--- /dev/null
+++ b/template/testdata/alfred.gotmpl/system-user-assistant-user
@@ -0,0 +1 @@
+<start_system>You are a helpful assistant.<end_message><start_user>Hello, how are you?<end_message><start_assistant>I'm doing great. How can I help you today?<end_message><start_user>I'd like to show off how chat templating works!<end_message><start_assistant>
\ No newline at end of file
diff --git a/template/testdata/alfred.gotmpl/user b/template/testdata/alfred.gotmpl/user
new file mode 100644
index 000000000..7c884a6f0
--- /dev/null
+++ b/template/testdata/alfred.gotmpl/user
@@ -0,0 +1 @@
+<start_user>Hello, how are you?<end_message><start_assistant>
\ No newline at end of file
diff --git a/template/testdata/alfred.gotmpl/user-assistant-user b/template/testdata/alfred.gotmpl/user-assistant-user
new file mode 100644
index 000000000..a60701ed7
--- /dev/null
+++ b/template/testdata/alfred.gotmpl/user-assistant-user
@@ -0,0 +1 @@
+<start_user>Hello, how are you?<end_message><start_assistant>I'm doing great. How can I help you today?<end_message><start_user>I'd like to show off how chat templating works!<end_message><start_assistant>
\ No newline at end of file
diff --git a/template/testdata/alpaca.gotmpl/system-user-assistant-user b/template/testdata/alpaca.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..4caa81788
--- /dev/null
+++ b/template/testdata/alpaca.gotmpl/system-user-assistant-user
@@ -0,0 +1,12 @@
+You are a helpful assistant.
+
+### Instruction:
+Hello, how are you?
+
+### Response:
+I'm doing great. How can I help you today?
+
+### Instruction:
+I'd like to show off how chat templating works!
+
+### Response:
diff --git a/template/testdata/alpaca.gotmpl/user b/template/testdata/alpaca.gotmpl/user
new file mode 100644
index 000000000..a0ce5dec7
--- /dev/null
+++ b/template/testdata/alpaca.gotmpl/user
@@ -0,0 +1,4 @@
+### Instruction:
+Hello, how are you?
+
+### Response:
diff --git a/template/testdata/alpaca.gotmpl/user-assistant-user b/template/testdata/alpaca.gotmpl/user-assistant-user
new file mode 100644
index 000000000..6c5e23ff5
--- /dev/null
+++ b/template/testdata/alpaca.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+### Instruction:
+Hello, how are you?
+
+### Response:
+I'm doing great. How can I help you today?
+
+### Instruction:
+I'd like to show off how chat templating works!
+
+### Response:
diff --git a/template/testdata/chatml.gotmpl/system-user-assistant-user b/template/testdata/chatml.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..8b013fcfb
--- /dev/null
+++ b/template/testdata/chatml.gotmpl/system-user-assistant-user
@@ -0,0 +1,9 @@
+<|im_start|>system
+You are a helpful assistant.<|im_end|>
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great. How can I help you today?<|im_end|>
+<|im_start|>user
+I'd like to show off how chat templating works!<|im_end|>
+<|im_start|>assistant
diff --git a/template/testdata/chatml.gotmpl/user b/template/testdata/chatml.gotmpl/user
new file mode 100644
index 000000000..aa9e597a4
--- /dev/null
+++ b/template/testdata/chatml.gotmpl/user
@@ -0,0 +1,3 @@
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant
diff --git a/template/testdata/chatml.gotmpl/user-assistant-user b/template/testdata/chatml.gotmpl/user-assistant-user
new file mode 100644
index 000000000..a7cba4de0
--- /dev/null
+++ b/template/testdata/chatml.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+<|im_start|>user
+Hello, how are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great. How can I help you today?<|im_end|>
+<|im_start|>user
+I'd like to show off how chat templating works!<|im_end|>
+<|im_start|>assistant
diff --git a/template/testdata/chatqa.gotmpl/system-user-assistant-user b/template/testdata/chatqa.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..98fd59bfa
--- /dev/null
+++ b/template/testdata/chatqa.gotmpl/system-user-assistant-user
@@ -0,0 +1,9 @@
+System: You are a helpful assistant.
+
+User: Hello, how are you?
+
+Assistant: I'm doing great. How can I help you today?
+
+User: I'd like to show off how chat templating works!
+
+Assistant:
\ No newline at end of file
diff --git a/template/testdata/chatqa.gotmpl/user b/template/testdata/chatqa.gotmpl/user
new file mode 100644
index 000000000..9e7cf702d
--- /dev/null
+++ b/template/testdata/chatqa.gotmpl/user
@@ -0,0 +1,3 @@
+User: Hello, how are you?
+
+Assistant:
\ No newline at end of file
diff --git a/template/testdata/chatqa.gotmpl/user-assistant-user b/template/testdata/chatqa.gotmpl/user-assistant-user
new file mode 100644
index 000000000..405bbe12c
--- /dev/null
+++ b/template/testdata/chatqa.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+User: Hello, how are you?
+
+Assistant: I'm doing great. How can I help you today?
+
+User: I'd like to show off how chat templating works!
+
+Assistant:
\ No newline at end of file
diff --git a/template/testdata/codellama-70b-instruct.gotmpl/system-user-assistant-user b/template/testdata/codellama-70b-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..d7528f80c
--- /dev/null
+++ b/template/testdata/codellama-70b-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,12 @@
+Source: system
+
+ You are a helpful assistant. <step> Source: user
+
+ Hello, how are you? <step> Source: assistant
+
+ I'm doing great. How can I help you today? <step> Source: user
+
+ I'd like to show off how chat templating works! <step> Source: assistant
+Destination: user
+
+ 
\ No newline at end of file
diff --git a/template/testdata/codellama-70b-instruct.gotmpl/user b/template/testdata/codellama-70b-instruct.gotmpl/user
new file mode 100644
index 000000000..8e07853ca
--- /dev/null
+++ b/template/testdata/codellama-70b-instruct.gotmpl/user
@@ -0,0 +1,6 @@
+Source: user
+
+ Hello, how are you? <step> Source: assistant
+Destination: user
+
+ 
\ No newline at end of file
diff --git a/template/testdata/codellama-70b-instruct.gotmpl/user-assistant-user b/template/testdata/codellama-70b-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..f732cc746
--- /dev/null
+++ b/template/testdata/codellama-70b-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+Source: user
+
+ Hello, how are you? <step> Source: assistant
+
+ I'm doing great. How can I help you today? <step> Source: user
+
+ I'd like to show off how chat templating works! <step> Source: assistant
+Destination: user
+
+ 
\ No newline at end of file
diff --git a/template/testdata/falcon-instruct.gotmpl/system-user-assistant-user b/template/testdata/falcon-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..16e45e5b6
--- /dev/null
+++ b/template/testdata/falcon-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,8 @@
+System: You are a helpful assistant.
+User:
+Hello, how are you?
+Falcon:
+I'm doing great. How can I help you today?
+User:
+I'd like to show off how chat templating works!
+Falcon:
diff --git a/template/testdata/falcon-instruct.gotmpl/user b/template/testdata/falcon-instruct.gotmpl/user
new file mode 100644
index 000000000..110831a2c
--- /dev/null
+++ b/template/testdata/falcon-instruct.gotmpl/user
@@ -0,0 +1,3 @@
+User:
+Hello, how are you?
+Falcon:
diff --git a/template/testdata/falcon-instruct.gotmpl/user-assistant-user b/template/testdata/falcon-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..b49639ea5
--- /dev/null
+++ b/template/testdata/falcon-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+User:
+Hello, how are you?
+Falcon:
+I'm doing great. How can I help you today?
+User:
+I'd like to show off how chat templating works!
+Falcon:
diff --git a/template/testdata/gemma-instruct.gotmpl/system-user-assistant-user b/template/testdata/gemma-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..5f6c37324
--- /dev/null
+++ b/template/testdata/gemma-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,8 @@
+<start_of_turn>user
+You are a helpful assistant.
+Hello, how are you?<end_of_turn>
+<start_of_turn>model
+I'm doing great. How can I help you today?<end_of_turn>
+<start_of_turn>user
+I'd like to show off how chat templating works!<end_of_turn>
+<start_of_turn>model
diff --git a/template/testdata/gemma-instruct.gotmpl/user b/template/testdata/gemma-instruct.gotmpl/user
new file mode 100644
index 000000000..dc8b30b68
--- /dev/null
+++ b/template/testdata/gemma-instruct.gotmpl/user
@@ -0,0 +1,3 @@
+<start_of_turn>user
+Hello, how are you?<end_of_turn>
+<start_of_turn>model
diff --git a/template/testdata/gemma-instruct.gotmpl/user-assistant-user b/template/testdata/gemma-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..1185924b0
--- /dev/null
+++ b/template/testdata/gemma-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+<start_of_turn>user
+Hello, how are you?<end_of_turn>
+<start_of_turn>model
+I'm doing great. How can I help you today?<end_of_turn>
+<start_of_turn>user
+I'd like to show off how chat templating works!<end_of_turn>
+<start_of_turn>model
diff --git a/template/testdata/granite-instruct.gotmpl/system-user-assistant-user b/template/testdata/granite-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..a732a77fb
--- /dev/null
+++ b/template/testdata/granite-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,13 @@
+System:
+You are a helpful assistant.
+
+Question:
+Hello, how are you?
+
+Answer:
+I'm doing great. How can I help you today?
+
+Question:
+I'd like to show off how chat templating works!
+
+Answer:
diff --git a/template/testdata/granite-instruct.gotmpl/user b/template/testdata/granite-instruct.gotmpl/user
new file mode 100644
index 000000000..7abd2ea35
--- /dev/null
+++ b/template/testdata/granite-instruct.gotmpl/user
@@ -0,0 +1,4 @@
+Question:
+Hello, how are you?
+
+Answer:
diff --git a/template/testdata/granite-instruct.gotmpl/user-assistant-user b/template/testdata/granite-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..da5e43eae
--- /dev/null
+++ b/template/testdata/granite-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+Question:
+Hello, how are you?
+
+Answer:
+I'm doing great. How can I help you today?
+
+Question:
+I'd like to show off how chat templating works!
+
+Answer:
diff --git a/template/testdata/llama2-chat.gotmpl/system-user-assistant-user b/template/testdata/llama2-chat.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..9db81cb44
--- /dev/null
+++ b/template/testdata/llama2-chat.gotmpl/system-user-assistant-user
@@ -0,0 +1,7 @@
+[INST] <<SYS>>
+You are a helpful assistant.
+<</SYS>>
+
+Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] <<SYS>><</SYS>>
+
+I'd like to show off how chat templating works! [/INST]
\ No newline at end of file
diff --git a/template/testdata/llama2-chat.gotmpl/user b/template/testdata/llama2-chat.gotmpl/user
new file mode 100644
index 000000000..ceef9bdbb
--- /dev/null
+++ b/template/testdata/llama2-chat.gotmpl/user
@@ -0,0 +1,3 @@
+[INST] <<SYS>><</SYS>>
+
+Hello, how are you? [/INST]
\ No newline at end of file
diff --git a/template/testdata/llama2-chat.gotmpl/user-assistant-user b/template/testdata/llama2-chat.gotmpl/user-assistant-user
new file mode 100644
index 000000000..ca58954f5
--- /dev/null
+++ b/template/testdata/llama2-chat.gotmpl/user-assistant-user
@@ -0,0 +1,5 @@
+[INST] <<SYS>><</SYS>>
+
+Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] <<SYS>><</SYS>>
+
+I'd like to show off how chat templating works! [/INST]
\ No newline at end of file
diff --git a/template/testdata/llama3-instruct.gotmpl/system-user-assistant-user b/template/testdata/llama3-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..6740bcb4d
--- /dev/null
+++ b/template/testdata/llama3-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,10 @@
+<|start_header_id|>system<|end_header_id|>
+
+You are a helpful assistant.<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+Hello, how are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+I'm doing great. How can I help you today?<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+I'd like to show off how chat templating works!<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
diff --git a/template/testdata/llama3-instruct.gotmpl/user b/template/testdata/llama3-instruct.gotmpl/user
new file mode 100644
index 000000000..470aa028f
--- /dev/null
+++ b/template/testdata/llama3-instruct.gotmpl/user
@@ -0,0 +1,4 @@
+<|start_header_id|>user<|end_header_id|>
+
+Hello, how are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
diff --git a/template/testdata/llama3-instruct.gotmpl/user-assistant-user b/template/testdata/llama3-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..6dd768af5
--- /dev/null
+++ b/template/testdata/llama3-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,8 @@
+<|start_header_id|>user<|end_header_id|>
+
+Hello, how are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+I'm doing great. How can I help you today?<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+I'd like to show off how chat templating works!<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
diff --git a/template/testdata/magicoder.gotmpl/system-user-assistant-user b/template/testdata/magicoder.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..c966a861d
--- /dev/null
+++ b/template/testdata/magicoder.gotmpl/system-user-assistant-user
@@ -0,0 +1,12 @@
+You are a helpful assistant.
+
+@@ Instruction
+Hello, how are you?
+
+@@ Response
+I'm doing great. How can I help you today?
+
+@@ Instruction
+I'd like to show off how chat templating works!
+
+@@ Response
diff --git a/template/testdata/magicoder.gotmpl/user b/template/testdata/magicoder.gotmpl/user
new file mode 100644
index 000000000..ccfb02bd2
--- /dev/null
+++ b/template/testdata/magicoder.gotmpl/user
@@ -0,0 +1,4 @@
+@@ Instruction
+Hello, how are you?
+
+@@ Response
diff --git a/template/testdata/magicoder.gotmpl/user-assistant-user b/template/testdata/magicoder.gotmpl/user-assistant-user
new file mode 100644
index 000000000..3aea6dab9
--- /dev/null
+++ b/template/testdata/magicoder.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+@@ Instruction
+Hello, how are you?
+
+@@ Response
+I'm doing great. How can I help you today?
+
+@@ Instruction
+I'd like to show off how chat templating works!
+
+@@ Response
diff --git a/template/testdata/mistral-instruct.gotmpl/system-user-assistant-user b/template/testdata/mistral-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..2f1edaec9
--- /dev/null
+++ b/template/testdata/mistral-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,3 @@
+[INST] You are a helpful assistant.
+
+Hello, how are you?[/INST] I'm doing great. How can I help you today?</s>[INST] I'd like to show off how chat templating works![/INST]
\ No newline at end of file
diff --git a/template/testdata/mistral-instruct.gotmpl/user b/template/testdata/mistral-instruct.gotmpl/user
new file mode 100644
index 000000000..b04871e5d
--- /dev/null
+++ b/template/testdata/mistral-instruct.gotmpl/user
@@ -0,0 +1 @@
+[INST] Hello, how are you?[/INST]
\ No newline at end of file
diff --git a/template/testdata/mistral-instruct.gotmpl/user-assistant-user b/template/testdata/mistral-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..b473e0df0
--- /dev/null
+++ b/template/testdata/mistral-instruct.gotmpl/user-assistant-user
@@ -0,0 +1 @@
+[INST] Hello, how are you?[/INST] I'm doing great. How can I help you today?</s>[INST] I'd like to show off how chat templating works![/INST]
\ No newline at end of file
diff --git a/template/testdata/openchat.gotmpl/system-user-assistant-user b/template/testdata/openchat.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..404b071aa
--- /dev/null
+++ b/template/testdata/openchat.gotmpl/system-user-assistant-user
@@ -0,0 +1 @@
+GPT4 Correct System: You are a helpful assistant.<|end_of_turn|>GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT4 Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT4 Correct Assistant:
\ No newline at end of file
diff --git a/template/testdata/openchat.gotmpl/user b/template/testdata/openchat.gotmpl/user
new file mode 100644
index 000000000..48229cb0e
--- /dev/null
+++ b/template/testdata/openchat.gotmpl/user
@@ -0,0 +1 @@
+GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant:
\ No newline at end of file
diff --git a/template/testdata/openchat.gotmpl/user-assistant-user b/template/testdata/openchat.gotmpl/user-assistant-user
new file mode 100644
index 000000000..4719abb2d
--- /dev/null
+++ b/template/testdata/openchat.gotmpl/user-assistant-user
@@ -0,0 +1 @@
+GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT4 Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT4 Correct Assistant:
\ No newline at end of file
diff --git a/template/testdata/phi-3.gotmpl/system-user-assistant-user b/template/testdata/phi-3.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..6109a9a24
--- /dev/null
+++ b/template/testdata/phi-3.gotmpl/system-user-assistant-user
@@ -0,0 +1,9 @@
+<|system|>
+You are a helpful assistant.<|end|>
+<|user|>
+Hello, how are you?<|end|>
+<|assistant|>
+I'm doing great. How can I help you today?<|end|>
+<|user|>
+I'd like to show off how chat templating works!<|end|>
+<|assistant|>
diff --git a/template/testdata/phi-3.gotmpl/user b/template/testdata/phi-3.gotmpl/user
new file mode 100644
index 000000000..feb96e7c9
--- /dev/null
+++ b/template/testdata/phi-3.gotmpl/user
@@ -0,0 +1,3 @@
+<|user|>
+Hello, how are you?<|end|>
+<|assistant|>
diff --git a/template/testdata/phi-3.gotmpl/user-assistant-user b/template/testdata/phi-3.gotmpl/user-assistant-user
new file mode 100644
index 000000000..db79d01c1
--- /dev/null
+++ b/template/testdata/phi-3.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+<|user|>
+Hello, how are you?<|end|>
+<|assistant|>
+I'm doing great. How can I help you today?<|end|>
+<|user|>
+I'd like to show off how chat templating works!<|end|>
+<|assistant|>
diff --git a/template/testdata/solar-instruct.gotmpl/system-user-assistant-user b/template/testdata/solar-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..28c1730ab
--- /dev/null
+++ b/template/testdata/solar-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,13 @@
+### System:
+You are a helpful assistant.
+
+### User:
+Hello, how are you?
+
+### Assistant:
+I'm doing great. How can I help you today?</s>
+
+### User:
+I'd like to show off how chat templating works!
+
+### Assistant:
diff --git a/template/testdata/solar-instruct.gotmpl/user b/template/testdata/solar-instruct.gotmpl/user
new file mode 100644
index 000000000..3a43382af
--- /dev/null
+++ b/template/testdata/solar-instruct.gotmpl/user
@@ -0,0 +1,4 @@
+### User:
+Hello, how are you?
+
+### Assistant:
diff --git a/template/testdata/solar-instruct.gotmpl/user-assistant-user b/template/testdata/solar-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..8553e73ba
--- /dev/null
+++ b/template/testdata/solar-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+### User:
+Hello, how are you?
+
+### Assistant:
+I'm doing great. How can I help you today?</s>
+
+### User:
+I'd like to show off how chat templating works!
+
+### Assistant:
diff --git a/template/testdata/starcoder2-instruct.gotmpl/system-user-assistant-user b/template/testdata/starcoder2-instruct.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..5b718b3ec
--- /dev/null
+++ b/template/testdata/starcoder2-instruct.gotmpl/system-user-assistant-user
@@ -0,0 +1,12 @@
+You are a helpful assistant.
+
+### Instruction
+Hello, how are you?
+
+### Response
+I'm doing great. How can I help you today?<|endoftext|>
+
+### Instruction
+I'd like to show off how chat templating works!
+
+### Response
diff --git a/template/testdata/starcoder2-instruct.gotmpl/user b/template/testdata/starcoder2-instruct.gotmpl/user
new file mode 100644
index 000000000..11b0be1fe
--- /dev/null
+++ b/template/testdata/starcoder2-instruct.gotmpl/user
@@ -0,0 +1,4 @@
+### Instruction
+Hello, how are you?
+
+### Response
diff --git a/template/testdata/starcoder2-instruct.gotmpl/user-assistant-user b/template/testdata/starcoder2-instruct.gotmpl/user-assistant-user
new file mode 100644
index 000000000..d99feabb0
--- /dev/null
+++ b/template/testdata/starcoder2-instruct.gotmpl/user-assistant-user
@@ -0,0 +1,10 @@
+### Instruction
+Hello, how are you?
+
+### Response
+I'm doing great. How can I help you today?<|endoftext|>
+
+### Instruction
+I'd like to show off how chat templating works!
+
+### Response
diff --git a/templates/testdata/templates.jsonl b/template/testdata/templates.jsonl
similarity index 100%
rename from templates/testdata/templates.jsonl
rename to template/testdata/templates.jsonl
diff --git a/template/testdata/vicuna.gotmpl/system-user-assistant-user b/template/testdata/vicuna.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..50d2f92c1
--- /dev/null
+++ b/template/testdata/vicuna.gotmpl/system-user-assistant-user
@@ -0,0 +1,6 @@
+You are a helpful assistant.
+
+USER: Hello, how are you?
+ASSISTANT: I'm doing great. How can I help you today?</s>
+USER: I'd like to show off how chat templating works!
+ASSISTANT:
\ No newline at end of file
diff --git a/template/testdata/vicuna.gotmpl/user b/template/testdata/vicuna.gotmpl/user
new file mode 100644
index 000000000..cbe5ef709
--- /dev/null
+++ b/template/testdata/vicuna.gotmpl/user
@@ -0,0 +1,2 @@
+USER: Hello, how are you?
+ASSISTANT:
\ No newline at end of file
diff --git a/template/testdata/vicuna.gotmpl/user-assistant-user b/template/testdata/vicuna.gotmpl/user-assistant-user
new file mode 100644
index 000000000..9172547e3
--- /dev/null
+++ b/template/testdata/vicuna.gotmpl/user-assistant-user
@@ -0,0 +1,4 @@
+USER: Hello, how are you?
+ASSISTANT: I'm doing great. How can I help you today?</s>
+USER: I'd like to show off how chat templating works!
+ASSISTANT:
\ No newline at end of file
diff --git a/template/testdata/zephyr.gotmpl/system-user-assistant-user b/template/testdata/zephyr.gotmpl/system-user-assistant-user
new file mode 100644
index 000000000..03d43fc34
--- /dev/null
+++ b/template/testdata/zephyr.gotmpl/system-user-assistant-user
@@ -0,0 +1,9 @@
+<|system|>
+You are a helpful assistant.</s>
+<|user|>
+Hello, how are you?</s>
+<|assistant|>
+I'm doing great. How can I help you today?</s>
+<|user|>
+I'd like to show off how chat templating works!</s>
+<|assistant|>
diff --git a/template/testdata/zephyr.gotmpl/user b/template/testdata/zephyr.gotmpl/user
new file mode 100644
index 000000000..6cefdaa0f
--- /dev/null
+++ b/template/testdata/zephyr.gotmpl/user
@@ -0,0 +1,3 @@
+<|user|>
+Hello, how are you?</s>
+<|assistant|>
diff --git a/template/testdata/zephyr.gotmpl/user-assistant-user b/template/testdata/zephyr.gotmpl/user-assistant-user
new file mode 100644
index 000000000..3937b006a
--- /dev/null
+++ b/template/testdata/zephyr.gotmpl/user-assistant-user
@@ -0,0 +1,7 @@
+<|user|>
+Hello, how are you?</s>
+<|assistant|>
+I'm doing great. How can I help you today?</s>
+<|user|>
+I'd like to show off how chat templating works!</s>
+<|assistant|>
diff --git a/template/vicuna.gotmpl b/template/vicuna.gotmpl
new file mode 100644
index 000000000..c27f39c52
--- /dev/null
+++ b/template/vicuna.gotmpl
@@ -0,0 +1,14 @@
+{{- $system := "" }}
+{{- range .Messages }}
+{{- if eq .Role "system" }}
+{{- if not $system }}{{ $system = .Content }}
+{{- else }}{{ $system = printf "%s\n\n%s" $system .Content }}
+{{- end }}
+{{- else if eq .Role "user" }}
+{{- if $system }}{{ $system }}
+
+{{ $system = "" }}
+{{- end }}USER: {{ .Content }}
+{{ else if eq .Role "assistant" }}ASSISTANT: {{ .Content }}</s>
+{{ end }}
+{{- end }}ASSISTANT:
\ No newline at end of file
diff --git a/template/vicuna.json b/template/vicuna.json
new file mode 100644
index 000000000..ed7bfb0fa
--- /dev/null
+++ b/template/vicuna.json
@@ -0,0 +1,6 @@
+{
+  "stop": [
+    "USER:",
+    "ASSISTANT:"
+  ]
+}
diff --git a/template/zephyr.gotmpl b/template/zephyr.gotmpl
new file mode 100644
index 000000000..25da148a0
--- /dev/null
+++ b/template/zephyr.gotmpl
@@ -0,0 +1,3 @@
+{{- range .Messages }}<|{{ .Role }}|>
+{{ .Content }}</s>
+{{ end }}<|assistant|>
diff --git a/template/zephyr.json b/template/zephyr.json
new file mode 100644
index 000000000..f9c0115cc
--- /dev/null
+++ b/template/zephyr.json
@@ -0,0 +1,8 @@
+{
+  "stop": [
+    "<|system|>",
+    "</s>",
+    "<|user|>",
+    "<|assistant|>"
+  ]
+}
diff --git a/templates/alfred.gotmpl b/templates/alfred.gotmpl
deleted file mode 100644
index cecb9d2c8..000000000
--- a/templates/alfred.gotmpl
+++ /dev/null
@@ -1 +0,0 @@
-{{ if .System }}<start_system>{{ .System }}<end_message>{{ end }}{{ if .Prompt }}<start_user>{{ .Prompt }}<end_message>{{ end }}<start_assistant>{{ .Response }}<end_message>
\ No newline at end of file
diff --git a/templates/alpaca.gotmpl b/templates/alpaca.gotmpl
deleted file mode 100644
index 440d06627..000000000
--- a/templates/alpaca.gotmpl
+++ /dev/null
@@ -1,7 +0,0 @@
-{{ if .System }}{{ .System }}
-
-{{ end }}{{ if .Prompt }}### Instruction:
-{{ .Prompt }}
-
-{{ end }}### Response:
-{{ .Response }}
\ No newline at end of file
diff --git a/templates/chatml.gotmpl b/templates/chatml.gotmpl
deleted file mode 100644
index dcf172853..000000000
--- a/templates/chatml.gotmpl
+++ /dev/null
@@ -1,6 +0,0 @@
-{{ if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
-{{ end }}{{ if .Prompt }}<|im_start|>user
-{{ .Prompt }}<|im_end|>
-{{ end }}<|im_start|>assistant
-{{ .Response }}<|im_end|>
\ No newline at end of file
diff --git a/templates/chatqa.gotmpl b/templates/chatqa.gotmpl
deleted file mode 100644
index 1ede6227f..000000000
--- a/templates/chatqa.gotmpl
+++ /dev/null
@@ -1,5 +0,0 @@
-{{ if .System }}System: {{ .System }}
-
-{{ end }}{{ if .Prompt }}User: {{ .Prompt }}
-
-{{ end }}Assistant: <|begin_of_text|>{{ .Response }}
\ No newline at end of file
diff --git a/templates/codellama-70b-instruct.gotmpl b/templates/codellama-70b-instruct.gotmpl
deleted file mode 100644
index 3196bd6fd..000000000
--- a/templates/codellama-70b-instruct.gotmpl
+++ /dev/null
@@ -1,8 +0,0 @@
-{{ if .System }} Source: system
-
- {{ .System }} <step>{{ end }} Source: user
-
- {{ .Prompt }} <step> Source: assistant
-Destination: user
-
- {{ .Response }}<step>
\ No newline at end of file
diff --git a/templates/falcon-instruct.gotmpl b/templates/falcon-instruct.gotmpl
deleted file mode 100644
index 2309a1c5d..000000000
--- a/templates/falcon-instruct.gotmpl
+++ /dev/null
@@ -1,3 +0,0 @@
-{{ if .System }}{{ .System }}
-{{ end }}{{ if .Prompt }}User: {{ .Prompt }}
-{{ end }}Assistant: {{ .Response }}
\ No newline at end of file
diff --git a/templates/gemma-instruct.gotmpl b/templates/gemma-instruct.gotmpl
deleted file mode 100644
index 91b9883a1..000000000
--- a/templates/gemma-instruct.gotmpl
+++ /dev/null
@@ -1,4 +0,0 @@
-<start_of_turn>user
-{{ if .System }}{{ .System }} {{ end }}{{ .Prompt }}<end_of_turn>
-<start_of_turn>model
-{{ .Response }}<end_of_turn>
\ No newline at end of file
diff --git a/templates/granite-instruct.gotmpl b/templates/granite-instruct.gotmpl
deleted file mode 100644
index 2ede647f5..000000000
--- a/templates/granite-instruct.gotmpl
+++ /dev/null
@@ -1,9 +0,0 @@
-{{ if .System }}
-System:
-{{ .System }}
-
-{{ end }}{{ if .Prompt }}Question:
-{{ .Prompt }}
-
-{{ end }}Answer:
-{{ .Response }}
\ No newline at end of file
diff --git a/templates/llama2-chat.gotmpl b/templates/llama2-chat.gotmpl
deleted file mode 100644
index a739f6908..000000000
--- a/templates/llama2-chat.gotmpl
+++ /dev/null
@@ -1,3 +0,0 @@
-[INST] <<SYS>>{{ .System }}<</SYS>>
-
-{{ .Prompt }} [/INST] {{ .Response }}
\ No newline at end of file
diff --git a/templates/llama3-instruct.gotmpl b/templates/llama3-instruct.gotmpl
deleted file mode 100644
index 36d0218b6..000000000
--- a/templates/llama3-instruct.gotmpl
+++ /dev/null
@@ -1,7 +0,0 @@
-{{ if .System }}<|start_header_id|>system<|end_header_id|>
-
-{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
-
-{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
-
-{{ .Response }}<|eot_id|>
\ No newline at end of file
diff --git a/templates/magicoder.gotmpl b/templates/magicoder.gotmpl
deleted file mode 100644
index 306972ecc..000000000
--- a/templates/magicoder.gotmpl
+++ /dev/null
@@ -1,7 +0,0 @@
-{{ if .System }}{{ .System }}
-
-{{ end }}{{ if .Prompt }}@@ Instruction
-{{ .Prompt }}
-
-{{ end }}@@ Response
-{{ .Response }}
\ No newline at end of file
diff --git a/templates/mistral-instruct.gotmpl b/templates/mistral-instruct.gotmpl
deleted file mode 100644
index dcf172853..000000000
--- a/templates/mistral-instruct.gotmpl
+++ /dev/null
@@ -1,6 +0,0 @@
-{{ if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
-{{ end }}{{ if .Prompt }}<|im_start|>user
-{{ .Prompt }}<|im_end|>
-{{ end }}<|im_start|>assistant
-{{ .Response }}<|im_end|>
\ No newline at end of file
diff --git a/templates/openchat.gotmpl b/templates/openchat.gotmpl
deleted file mode 100644
index d2ca38685..000000000
--- a/templates/openchat.gotmpl
+++ /dev/null
@@ -1 +0,0 @@
-{{ .System }}<|end_of_turn|>GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>
\ No newline at end of file
diff --git a/templates/phi-3.gotmpl b/templates/phi-3.gotmpl
deleted file mode 100644
index bf26dcee2..000000000
--- a/templates/phi-3.gotmpl
+++ /dev/null
@@ -1,6 +0,0 @@
-{{ if .System }}<|system|>
-{{ .System }}<|end|>
-{{ end }}{{ if .Prompt }}<|user|>
-{{ .Prompt }}<|end|>
-{{ end }}<|assistant|>
-{{ .Response }}<|end|>
\ No newline at end of file
diff --git a/templates/solar-instruct.gotmpl b/templates/solar-instruct.gotmpl
deleted file mode 100644
index c275a26a3..000000000
--- a/templates/solar-instruct.gotmpl
+++ /dev/null
@@ -1,8 +0,0 @@
-{{ if .System }}### System:
-{{ .System }}
-
-{{ end }}{{ if .Prompt }}### User:
-{{ .Prompt }}
-
-{{ end }}### Assistant:
-{{ .Response }}
\ No newline at end of file
diff --git a/templates/starcoder2-instruct.gotmpl b/templates/starcoder2-instruct.gotmpl
deleted file mode 100644
index 33357e54e..000000000
--- a/templates/starcoder2-instruct.gotmpl
+++ /dev/null
@@ -1,9 +0,0 @@
-{{ if .System }}{{ .System }}
-
-{{ end }}{{ if .Prompt }}### Instruction
-{{ .Prompt }}
-
-
-{{ end }}### Response
-{{ .Response }}<|endoftext|>
-
diff --git a/templates/template.go b/templates/template.go
deleted file mode 100644
index 72bd69e9d..000000000
--- a/templates/template.go
+++ /dev/null
@@ -1,70 +0,0 @@
-package templates
-
-import (
-	"bytes"
-	"embed"
-	"encoding/json"
-	"errors"
-	"io"
-	"math"
-	"sync"
-
-	"github.com/agnivade/levenshtein"
-)
-
-//go:embed index.json
-var indexBytes []byte
-
-//go:embed *.gotmpl
-var templatesFS embed.FS
-
-var templatesOnce = sync.OnceValues(func() ([]*Template, error) {
-	var templates []*Template
-	if err := json.Unmarshal(indexBytes, &templates); err != nil {
-		return nil, err
-	}
-
-	for _, t := range templates {
-		bts, err := templatesFS.ReadFile(t.Name + ".gotmpl")
-		if err != nil {
-			return nil, err
-		}
-
-		// normalize line endings
-		t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
-	}
-
-	return templates, nil
-})
-
-type Template struct {
-	Name     string `json:"name"`
-	Template string `json:"template"`
-	Bytes []byte
-}
-
-func (t Template) Reader() io.Reader {
-	return bytes.NewReader(t.Bytes)
-}
-
-func NamedTemplate(s string) (*Template, error) {
-	templates, err := templatesOnce()
-	if err != nil {
-		return nil, err
-	}
-
-	var template *Template
-	score := math.MaxInt
-	for _, t := range templates {
-		if s := levenshtein.ComputeDistance(s, t.Template); s < score {
-			score = s
-			template = t
-		}
-	}
-
-	if score < 100 {
-		return template, nil
-	}
-
-	return nil, errors.New("no matching template found")
-}
diff --git a/templates/template_test.go b/templates/template_test.go
deleted file mode 100644
index 61bc78374..000000000
--- a/templates/template_test.go
+++ /dev/null
@@ -1,59 +0,0 @@
-package templates
-
-import (
-	"bufio"
-	"bytes"
-	"encoding/json"
-	"io"
-	"os"
-	"path/filepath"
-	"testing"
-	"text/template"
-
-	"github.com/ollama/ollama/llm"
-)
-
-func TestKVChatTemplate(t *testing.T) {
-	f, err := os.Open(filepath.Join("testdata", "templates.jsonl"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer f.Close()
-
-	scanner := bufio.NewScanner(f)
-	for scanner.Scan() {
-		var ss map[string]string
-		if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil {
-			t.Fatal(err)
-		}
-
-		for k, v := range ss {
-			t.Run(k, func(t *testing.T) {
-				kv := llm.KV{"tokenizer.chat_template": v}
-				s := kv.ChatTemplate()
-				r, err := NamedTemplate(s)
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if r.Name != k {
-					t.Errorf("expected %q, got %q", k, r.Name)
-				}
-
-				var b bytes.Buffer
-				if _, err := io.Copy(&b, r.Reader()); err != nil {
-					t.Fatal(err)
-				}
-
-				tmpl, err := template.New(s).Parse(b.String())
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if tmpl.Tree.Root.String() == "" {
-					t.Errorf("empty %s template", k)
-				}
-			})
-		}
-	}
-}
diff --git a/templates/vicuna.gotmpl b/templates/vicuna.gotmpl
deleted file mode 100644
index 174c1a353..000000000
--- a/templates/vicuna.gotmpl
+++ /dev/null
@@ -1,3 +0,0 @@
-{{ if .System }}{{ .System }}
-{{ end }}{{ if .Prompt }}USER: {{ .Prompt }}
-{{ end }}ASSISTANT: {{ .Response }}
\ No newline at end of file
diff --git a/templates/zephyr.gotmpl b/templates/zephyr.gotmpl
deleted file mode 100644
index aac0c7a1f..000000000
--- a/templates/zephyr.gotmpl
+++ /dev/null
@@ -1,6 +0,0 @@
-{{ if .System }}<|system|>
-{{ .System }}</s>
-{{ end }}{{ if .Prompt }}<|user|>
-{{ .Prompt }}</s>
-{{ end }}<|assistant|>
-{{ .Response }}</s>
\ No newline at end of file
diff --git a/types/errtypes/errtypes.go b/types/errtypes/errtypes.go
index d30731140..27c3f913e 100644
--- a/types/errtypes/errtypes.go
+++ b/types/errtypes/errtypes.go
@@ -6,8 +6,10 @@ import (
 	"strings"
 )
 
-const UnknownOllamaKeyErrMsg = "unknown ollama key"
-const InvalidModelNameErrMsg = "invalid model name"
+const (
+	UnknownOllamaKeyErrMsg = "unknown ollama key"
+	InvalidModelNameErrMsg = "invalid model name"
+)
 
 // TODO: This should have a structured response from the API
 type UnknownOllamaKey struct {
diff --git a/types/model/name.go b/types/model/name.go
index e645a844c..75b35ef78 100644
--- a/types/model/name.go
+++ b/types/model/name.go
@@ -91,7 +91,6 @@ type Name struct {
 	Namespace string
 	Model     string
 	Tag       string
-	RawDigest string
 }
 
 // ParseName parses and assembles a Name from a name string. The
@@ -143,11 +142,6 @@ func ParseNameBare(s string) Name {
 	var n Name
 	var promised bool
 
-	s, n.RawDigest, promised = cutLast(s, "@")
-	if promised && n.RawDigest == "" {
-		n.RawDigest = MissingPart
-	}
-
 	// "/" is an illegal tag character, so we can use it to split the host
 	if strings.LastIndex(s, ":") > strings.LastIndex(s, "/") {
 		s, n.Tag, _ = cutPromised(s, ":")
@@ -222,14 +216,10 @@ func (n Name) String() string {
 		b.WriteByte(':')
 		b.WriteString(n.Tag)
 	}
-	if n.RawDigest != "" {
-		b.WriteByte('@')
-		b.WriteString(n.RawDigest)
-	}
 	return b.String()
 }
 
-// DisplayShort returns a short string version of the name.
+// DisplayShortest returns a short string version of the name.
 func (n Name) DisplayShortest() string {
 	var sb strings.Builder
 
@@ -250,23 +240,25 @@ func (n Name) DisplayShortest() string {
 	return sb.String()
 }
 
-func IsValidNamespace(namespace string) bool {
-	return isValidPart(kindNamespace, namespace)
+// IsValidNamespace reports whether the provided string is a valid
+// namespace.
+func IsValidNamespace(s string) bool {
+	return isValidPart(kindNamespace, s)
 }
 
 // IsValid reports whether all parts of the name are present and valid. The
 // digest is a special case, and is checked for validity only if present.
+//
+// Note: The digest check has been removed as is planned to be added back in
+// at a later time.
 func (n Name) IsValid() bool {
-	if n.RawDigest != "" && !isValidPart(kindDigest, n.RawDigest) {
-		return false
-	}
 	return n.IsFullyQualified()
 }
 
 // IsFullyQualified returns true if all parts of the name are present and
 // valid without the digest.
 func (n Name) IsFullyQualified() bool {
-	var parts = []string{
+	parts := []string{
 		n.Host,
 		n.Namespace,
 		n.Model,
diff --git a/types/model/name_test.go b/types/model/name_test.go
index 008dd586c..794d14d79 100644
--- a/types/model/name_test.go
+++ b/types/model/name_test.go
@@ -122,21 +122,6 @@ func TestParseNameParts(t *testing.T) {
 			},
 			wantFilepath: filepath.Join(part350, part80, part80, part80),
 		},
-		{
-			in: "@digest",
-			want: Name{
-				RawDigest: "digest",
-			},
-			wantValidDigest: false,
-		},
-		{
-			in: "model@sha256:123",
-			want: Name{
-				Model:     "model",
-				RawDigest: "sha256:123",
-			},
-			wantValidDigest: true,
-		},
 	}
 
 	for _, tt := range cases {
@@ -160,22 +145,18 @@ var testCases = map[string]bool{ // name -> valid
 	"_why/_the/_lucky:_stiff": true,
 
 	// minimal
-	"h/n/m:t@d": true,
+	"h/n/m:t": true,
 
 	"host/namespace/model:tag": true,
 	"host/namespace/model":     false,
 	"namespace/model":          false,
 	"model":                    false,
-	"@sha256-1000000000000000000000000000000000000000000000000000000000000000":      false,
-	"model@sha256-1000000000000000000000000000000000000000000000000000000000000000": false,
-	"model@sha256:1000000000000000000000000000000000000000000000000000000000000000": false,
 
 	// long (but valid)
 	part80 + "/" + part80 + "/" + part80 + ":" + part80:  true,
 	part350 + "/" + part80 + "/" + part80 + ":" + part80: true,
 
-	"h/nn/mm:t@sha256-1000000000000000000000000000000000000000000000000000000000000000": true, // bare minimum part sizes
-	"h/nn/mm:t@sha256:1000000000000000000000000000000000000000000000000000000000000000": true, // bare minimum part sizes
+	"h/nn/mm:t": true, // bare minimum part sizes
 
 	// unqualified
 	"m":     false,
@@ -196,11 +177,10 @@ var testCases = map[string]bool{ // name -> valid
 	"@":      false,
 
 	// not starting with alphanum
-	"-hh/nn/mm:tt@dd": false,
-	"hh/-nn/mm:tt@dd": false,
-	"hh/nn/-mm:tt@dd": false,
-	"hh/nn/mm:-tt@dd": false,
-	"hh/nn/mm:tt@-dd": false,
+	"-hh/nn/mm:tt": false,
+	"hh/-nn/mm:tt": false,
+	"hh/nn/-mm:tt": false,
+	"hh/nn/mm:-tt": false,
 
 	// hosts
 	"host:https/namespace/model:tag": true,
@@ -334,7 +314,7 @@ func FuzzName(f *testing.F) {
 	f.Fuzz(func(t *testing.T, s string) {
 		n := ParseNameBare(s)
 		if n.IsValid() {
-			parts := [...]string{n.Host, n.Namespace, n.Model, n.Tag, n.RawDigest}
+			parts := [...]string{n.Host, n.Namespace, n.Model, n.Tag}
 			for _, part := range parts {
 				if part == ".." {
 					t.Errorf("unexpected .. as valid part")