some more menu options...

.dockerignore

@@ -7,5 +7,3 @@ llm/llama.cpp
 .env
 .cache
 test_data
-llm/build
-llama/build

.gitattributes

@@ -1,3 +1 @@
 llm/ext_server/* linguist-vendored
-* text=auto
-*.go text eol=lf

.github/workflows/release.yaml

@@ -28,7 +28,6 @@ jobs:
           security unlock-keychain -p password build.keychain
           security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
           security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
-          security set-keychain-settings -lut 3600 build.keychain
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
@@ -102,8 +101,7 @@ jobs:
         with:
           name: generate-windows-cpu
           path: |
-            build/**/*
+            llm/build/**/bin/*
-            build/**/*.a
             llm/build/**/*.a
             dist/windows-amd64/**
 
@@ -148,7 +146,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-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"
@@ -177,7 +175,7 @@ jobs:
         with:
           name: generate-windows-rocm
           path: |
-            build/**/*
+            llm/build/**/bin/*
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
@@ -188,13 +186,6 @@ 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:
@@ -228,11 +219,11 @@ jobs:
         with:
           go-version-file: go.mod
           cache: true
-      - name: 'Install CUDA ${{ matrix.cuda.version }}'
+      - name: 'Install CUDA'
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading CUDA Installer"
-          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
           write-host "Installing CUDA"
           Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
           write-host "Completed CUDA"
@@ -264,142 +255,16 @@ jobs:
           cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
       - uses: actions/upload-artifact@v4
         with:
-          name: generate-windows-cuda-${{ matrix.cuda.version }}
+          name: generate-windows-cuda
           path: |
-            build/**/*
+            llm/build/**/bin/*
             dist/windows-amd64/**
       - uses: actions/upload-artifact@v4
         with:
-          name: windows-cuda-deps-${{ matrix.cuda.version }}
+          name: windows-cuda-deps
           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
-      - 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
-          .\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
@@ -407,7 +272,6 @@ jobs:
       - generate-windows-cuda
       - generate-windows-rocm
       - generate-windows-cpu
-      - windows-arm64
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -449,26 +313,17 @@ jobs:
           name: generate-windows-cpu
       - uses: actions/download-artifact@v4
         with:
-          name: generate-windows-cuda-11
+          name: generate-windows-cuda
       - uses: actions/download-artifact@v4
         with:
-          name: generate-windows-cuda-12
+          name: windows-cuda-deps
-      - 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
-      - uses: actions/download-artifact@v4
+      - run: dir llm/build
-        with:
-          name: windows-arm64
-      - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent
           & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
@@ -489,7 +344,9 @@ jobs:
     environment: release
     runs-on: linux
     env:
-      PLATFORM: linux/amd64
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: amd64
+      PUSH: '1'
     steps:
       - uses: actions/checkout@v4
         with:
@@ -497,8 +354,15 @@ jobs:
       - name: Set Version
         shell: bash
         run: echo "VERSION=${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
@@ -512,7 +376,9 @@ jobs:
     environment: release
     runs-on: linux-arm64
     env:
-      PLATFORM: linux/arm64
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: arm64
+      PUSH: '1'
     steps:
       - uses: actions/checkout@v4
         with:
@@ -541,8 +407,14 @@ 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
@@ -550,178 +422,6 @@ 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:
@@ -734,7 +434,8 @@ jobs:
     permissions:
       contents: write
     env:
-      GH_TOKEN: ${{ github.token }}
+      OLLAMA_SKIP_IMAGE_BUILD: '1'
+      PUSH: '1'
     steps:
       - uses: actions/checkout@v4
       - name: Set Version
@@ -742,6 +443,12 @@ jobs:
         run: |
           echo "VERSION=${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:
@@ -750,23 +457,17 @@ jobs:
           merge-multiple: true
       - run: |
           ls -lh dist/
-          (cd dist; find . -type f | xargs sha256sum > ../sha256sum.txt)
+          (cd dist; sha256sum * > sha256sum.txt)
-          mv sha256sum.txt dist/
           cat dist/sha256sum.txt
-      - name: Create or update Release
+      - uses: ncipollo/release-action@v1
-        run: |
+        with:
-          echo "Looking for existing release for ${{ env.RELEASE_VERSION }}"
+          name: ${{ env.RELEASE_VERSION }}
-          OLD_TAG=$(gh release ls --json name,tagName | jq -r ".[] | select(.name == \"${{ env.RELEASE_VERSION }}\") | .tagName")
+          allowUpdates: true
-          if [ -n "$OLD_TAG" ]; then
+          artifacts: 'dist/*'
-            echo "Updating release ${{ env.RELEASE_VERSION }} to point to new tag ${GITHUB_REF_NAME}"
+          draft: true
-            gh release edit ${OLD_TAG} --tag ${GITHUB_REF_NAME}
+          prerelease: true
-          else
+          omitBodyDuringUpdate: true
-            echo "Creating new release ${{ env.RELEASE_VERSION }} pointing to tag ${GITHUB_REF_NAME}"
+          generateReleaseNotes: true
-            gh release create ${GITHUB_REF_NAME} \
+          omitDraftDuringUpdate: true
-              --title ${{ env.RELEASE_VERSION }} \
+          omitPrereleaseDuringUpdate: true
-              --draft \
+          replacesArtifacts: true
-              --generate-notes \
-              --prerelease
-          fi
-          echo "Uploading artifacts for tag ${GITHUB_REF_NAME}"
-          gh release upload ${GITHUB_REF_NAME} dist/* --clobber

.github/workflows/test.yaml

@@ -34,13 +34,13 @@ jobs:
             git diff-tree -r --no-commit-id --name-only \
               $(git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}) \
               ${{ github.event.pull_request.head.sha }} \
-              | xargs python3 -c "import sys; from pathlib import Path; print(any(Path(x).match(glob) for x in sys.argv[1:] for glob in '$*'.split(' ')))"
+              | xargs python3 -c "import sys; print(any([x.startswith('$1') for x in sys.argv[1:]]))"
           }
 
           {
-            echo GENERATE=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE=$(changed llm/)
-            echo GENERATE_CUDA=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE_CUDA=$(changed llm/)
-            echo GENERATE_ROCM=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE_ROCM=$(changed llm/)
           } >>$GITHUB_OUTPUT
 
   generate:
@@ -58,7 +58,6 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       GOARCH: ${{ matrix.arch }}
-      CGO_ENABLED: '1'
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
@@ -80,7 +79,12 @@ jobs:
       - run: go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
         name: 'Unix Go Generate'
-      - run: go build .
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
+          path: |
+            llm/build/**/bin/*
+            llm/build/**/*.a
   generate-cuda:
     needs: [changes]
     if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
@@ -108,13 +112,19 @@ jobs:
           go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
+      - uses: actions/upload-artifact@v4
+        with:
+          name: cuda-${{ matrix.cuda-version }}-libraries
+          path: |
+            llm/build/**/bin/*
+            dist/windows-amd64/**
   generate-rocm:
     needs: [changes]
     if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
     strategy:
       matrix:
         rocm-version:
-          - '6.1.2'
+          - '6.0.2'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
     steps:
@@ -135,6 +145,12 @@ 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:
@@ -151,7 +167,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-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"
@@ -171,6 +187,7 @@ jobs:
         name: go generate
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
 
   # CUDA generation step
   generate-windows-cuda:
@@ -212,6 +229,7 @@ jobs:
           go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
 
   lint:
     strategy:
@@ -243,7 +261,15 @@ jobs:
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
-      - uses: golangci/golangci-lint-action@v6
+      - run: |
+          mkdir -p llm/build/linux/$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@v4
         with:
           args: --timeout 8m0s -v
   test:
@@ -261,8 +287,6 @@ jobs:
       GOARCH: ${{ matrix.arch }}
       CGO_ENABLED: '1'
       OLLAMA_CPU_TARGET: 'static'
-      OLLAMA_SKIP_CPU_GENERATE: '1'
-      OLLAMA_SKIP_METAL_GENERATE: '1'
     steps:
       - uses: actions/checkout@v4
         with:
@@ -273,10 +297,23 @@ jobs:
           cache: true
       - run: |
           case ${{ matrix.arch }} in
-            amd64) echo ARCH=amd64 ;;
+            amd64) echo ARCH=x86_64 ;;
             arm64) echo ARCH=arm64 ;;
           esac >>$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

.gitignore

@@ -12,7 +12,3 @@ ggml-metal.metal
 test_data
 *.crt
 llm/build
-build/*/*/*
-!build/**/placeholder
-llama/build
-__debug_bin*

.golangci.yaml

@@ -7,41 +7,11 @@ linters:
     - bodyclose
     - containedctx
     - contextcheck
-    - errcheck
     - exportloopref
-    - gci
     - gocheckcompilerdirectives
-    - gofmt
+    # FIXME: for some reason this errors on windows
-    - gofumpt
+    # - gofmt
-    - gosimple
+    # - goimports
-    - govet
-    - ineffassign
-    - intrange
-    - makezero
     - misspell
     - nilerr
-    - nolintlint
-    - nosprintfhostport
-    - staticcheck
-    - tenv
-    - unconvert
     - unused
-    - usestdlibvars
-    - wastedassign
-    - whitespace
-linters-settings:
-  gci:
-    sections: [standard, default, localmodule]
-  staticcheck:
-    checks:
-      - all
-      - -SA1019 # omit Deprecated check
-severity:
-  default-severity: error
-  rules:
-    - linters:
-        - gofmt
-        - goimports
-        - intrange
-        - usestdlibvars
-      severity: info

CONTRIBUTING.md

@@ -1,37 +0,0 @@
-# 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).

Dockerfile

@@ -1,10 +1,8 @@
-ARG GOLANG_VERSION=1.22.5
+ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
-ARG CUDA_VERSION_11=11.3.1
+# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
-ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
+ARG CUDA_VERSION=11.3.1
-ARG CUDA_VERSION_12=12.4.0
+ARG ROCM_VERSION=6.0.2
-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
@@ -12,243 +10,131 @@ COPY .git .git
 COPY .gitmodules .gitmodules
 COPY llm llm
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION-devel-centos7 AS cuda-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:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
 
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
 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/gcc-toolset-10/root/usr/bin:$PATH
 COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${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:$PATH
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH=/opt/amdgpu/lib64
+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
-ENV GOARCH=amd64
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
-RUN --mount=type=cache,target=/root/.ccache \
+RUN mkdir /tmp/scratch && \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
+    for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
-RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
+        cp ${dep} /tmp/scratch/ || exit 1 ; \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
+    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 . )
+
 
 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 --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
-    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 --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
-    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 --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
 
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
+FROM --platform=linux/arm64 centos:7 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/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/
 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 --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
 FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN --mount=type=cache,target=/root/.ccache \
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 
 
-# Intermediate stages used for ./scripts/build_linux.sh
+# Intermediate stage 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/ llm/build/
+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/build/ build/
+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/build/ build/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
-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 --mount=type=cache,target=/root/.ccache \
+RUN go build -trimpath .
-    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/ llm/build/
+COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-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 --mount=type=cache,target=/root/.ccache \
+RUN go build -trimpath .
-    go build -trimpath -o dist/linux-arm64/bin/ollama .
-RUN cd dist/linux-$GOARCH && \
-    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
 
-FROM --platform=linux/amd64 scratch AS dist-amd64
+# Runtime stages
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
-FROM --platform=linux/arm64 scratch AS dist-arm64
+RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
-FROM dist-$TARGETARCH as dist
+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
 
-
+# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
-# Optimized container images do not cary nested payloads
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
-FROM --platform=linux/amd64 static-build-amd64 AS container-build-amd64
+RUN update-pciids
-WORKDIR /go/src/github.com/ollama/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/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

README.md

@@ -1,12 +1,12 @@
 <div align="center">
- <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+  <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>
 
 # Ollama
 
 [![Discord](https://dcbadge.vercel.app/api/server/ollama?style=flat&compact=true)](https://discord.gg/ollama)
 
-Get up and running with large language models.
+Get up and running with large language models locally.
 
 ### macOS
 
@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
+To run and chat with [Llama 3](https://ollama.com/library/llama3):
 
 ```
-ollama run llama3.1
+ollama run llama3
 ```
 
 ## Model library
@@ -49,25 +49,20 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
+| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
-| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
+| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
-| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
+| Phi-3              | 3,8B       | 2.3GB | `ollama run phi3`              |
-| 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`           |
-| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
+| Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
+| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
 | Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
-> [!NOTE]
+> 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.
-> 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
 
@@ -99,16 +94,16 @@ 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 `llama3.1` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:
 
 ```
-ollama pull llama3.1
+ollama pull llama3
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3.1
+FROM llama3
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -143,7 +138,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3.1
+ollama pull llama3
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -151,13 +146,13 @@ ollama pull llama3.1
 ### Remove a model
 
 ```
-ollama rm llama3.1
+ollama rm llama3
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3.1 my-model
+ollama cp llama3 my-model
 ```
 
 ### Multiline input
@@ -174,48 +169,48 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 ### Multimodal models
 
 ```
-ollama run llava "What's in this image? /Users/jmorgan/Desktop/smile.png"
+>>> 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
+### Pass in prompt as arguments
 
 ```
-$ ollama run llama3.1 "Summarize this file: $(cat README.md)"
+$ ollama run llama3 "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 llama3.1
-```
-
 ### List models on your computer
 
 ```
 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.
 
 ## Building
 
-See the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
+Install `cmake` and `go`:
+
+```
+brew install cmake go
+```
+
+Then generate dependencies:
+
+```
+go generate ./...
+```
+
+Then build the binary:
+
+```
+go build .
+```
+
+More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
 
 ### Running local builds
 
@@ -228,7 +223,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3.1
+./ollama run llama3
 ```
 
 ## REST API
@@ -239,7 +234,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -248,7 +243,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -263,7 +258,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
-- [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
 - [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
@@ -290,41 +284,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
 - [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
-- [LLM-X](https://github.com/mrdjohnson/llm-x) (Progressive Web App)
+- [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
 - [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
 - [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
 - [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
-- [QA-Pilot](https://github.com/reid41/QA-Pilot) (Chat with Code Repository)
+- [QA-Pilot: Chat with Code Repository](https://github.com/reid41/QA-Pilot)
-- [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
+- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
-- [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
+- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
-- [RAGFlow](https://github.com/infiniflow/ragflow) (Open-source Retrieval-Augmented Generation engine based on deep document understanding)
+- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
-- [StreamDeploy](https://github.com/StreamDeploy-DevRel/streamdeploy-llm-app-scaffold) (LLM Application Scaffold)
+- [chat: chat web app for teams](https://github.com/swuecho/chat)
-- [chat](https://github.com/swuecho/chat) (chat web app for teams)
 - [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
-- [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
+- [Ollama RAG Chatbot: Local Chat with multiples PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)
-- [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)
-- [macai](https://github.com/Renset/macai) (macOS client for Ollama, ChatGPT, and other compatible API back-ends)
-- [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
 
@@ -347,13 +317,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
 - [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
 
@@ -363,28 +326,19 @@ 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)
@@ -394,24 +348,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Haystack](https://github.com/deepset-ai/haystack-integrations/blob/main/integrations/ollama.md)
 - [Elixir LangChain](https://github.com/brainlid/langchain)
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
-- [Ollama for R - ollama-r](https://github.com/hauselin/ollama-r)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
 - [Testcontainers](https://testcontainers.com/modules/ollama/)
-- [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
 
@@ -426,28 +370,18 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama Telegram Bot](https://github.com/ruecat/ollama-telegram)
 - [Hass Ollama Conversation](https://github.com/ej52/hass-ollama-conversation)
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
+- [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
 - [Cliobot](https://github.com/herval/cliobot) (Telegram bot with Ollama support)
 - [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
 - [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
-- [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 Hugging Face)
+- [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [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
-
-- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
 
+### Supported backends 
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 

SECURITY.md

@@ -1,25 +0,0 @@
-# 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

api/client.go

@@ -1,16 +1,9 @@
 // Package api implements the client-side API for code wishing to interact
 // with the ollama service. The methods of the [Client] type correspond to
-// the ollama REST API as described in [the API documentation].
+// the ollama REST API as described in https://github.com/ollama/ollama/blob/main/docs/api.md
+//
 // The ollama command-line client itself uses this package to interact with
 // the backend service.
-//
-// # Examples
-//
-// Several examples of using this package are available [in the GitHub
-// repository].
-//
-// [the API documentation]: https://github.com/ollama/ollama/blob/main/docs/api.md
-// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/examples
 package api
 
 import (
@@ -18,14 +11,15 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
+	"net"
 	"net/http"
 	"net/url"
+	"os"
 	"runtime"
+	"strings"
 
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/version"
 )
@@ -63,8 +57,36 @@ 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) {
+	defaultPort := "11434"
+
+	scheme, hostport, ok := strings.Cut(os.Getenv("OLLAMA_HOST"), "://")
+	switch {
+	case !ok:
+		scheme, hostport = "http", os.Getenv("OLLAMA_HOST")
+	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
+		}
+	}
+
 	return &Client{
-		base: envconfig.Host(),
+		base: &url.URL{
+			Scheme: scheme,
+			Host:   net.JoinHostPort(host, port),
+		},
 		http: http.DefaultClient,
 	}, nil
 }
@@ -173,7 +195,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 		}
 
 		if errorResponse.Error != "" {
-			return errors.New(errorResponse.Error)
+			return fmt.Errorf(errorResponse.Error)
 		}
 
 		if response.StatusCode >= http.StatusBadRequest {
@@ -250,14 +272,8 @@ func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc
 	})
 }
 
-// PushProgressFunc is a function that [Client.Push] invokes when progress is
-// made.
-// It's similar to other progress function types like [PullProgressFunc].
 type PushProgressFunc func(ProgressResponse) error
 
-// Push uploads a model to the model library; requires registering for ollama.ai
-// and adding a public key first. fn is called each time progress is made on
-// the request and can be used to display a progress bar, etc.
 func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/push", req, func(bts []byte) error {
 		var resp ProgressResponse
@@ -269,15 +285,8 @@ func (c *Client) Push(ctx context.Context, req *PushRequest, fn PushProgressFunc
 	})
 }
 
-// CreateProgressFunc is a function that [Client.Create] invokes when progress
-// is made.
-// It's similar to other progress function types like [PullProgressFunc].
 type CreateProgressFunc func(ProgressResponse) error
 
-// Create creates a model from a [Modelfile]. fn is a progress function that
-// behaves similarly to other methods (see [Client.Pull]).
-//
-// [Modelfile]: https://github.com/ollama/ollama/blob/main/docs/modelfile.md
 func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/create", req, func(bts []byte) error {
 		var resp ProgressResponse
@@ -289,7 +298,6 @@ func (c *Client) Create(ctx context.Context, req *CreateRequest, fn CreateProgre
 	})
 }
 
-// List lists models that are available locally.
 func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	var lr ListResponse
 	if err := c.do(ctx, http.MethodGet, "/api/tags", nil, &lr); err != nil {
@@ -298,17 +306,6 @@ func (c *Client) List(ctx context.Context) (*ListResponse, error) {
 	return &lr, nil
 }
 
-// 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 {
-		return nil, err
-	}
-	return &lr, nil
-}
-
-// Copy copies a model - creating a model with another name from an existing
-// model.
 func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {
 	if err := c.do(ctx, http.MethodPost, "/api/copy", req, nil); err != nil {
 		return err
@@ -316,7 +313,6 @@ func (c *Client) Copy(ctx context.Context, req *CopyRequest) error {
 	return nil
 }
 
-// Delete deletes a model and its data.
 func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	if err := c.do(ctx, http.MethodDelete, "/api/delete", req, nil); err != nil {
 		return err
@@ -324,7 +320,6 @@ func (c *Client) Delete(ctx context.Context, req *DeleteRequest) error {
 	return nil
 }
 
-// Show obtains model information, including details, modelfile, license etc.
 func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, error) {
 	var resp ShowResponse
 	if err := c.do(ctx, http.MethodPost, "/api/show", req, &resp); err != nil {
@@ -333,25 +328,12 @@ func (c *Client) Show(ctx context.Context, req *ShowRequest) (*ShowResponse, err
 	return &resp, nil
 }
 
-// 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 {
 		return err
 	}
 	return nil
 }
-
-// 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 {
@@ -360,13 +342,10 @@ func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*Embedd
 	return &resp, nil
 }
 
-// CreateBlob creates a blob from a file on the server. digest is the
-// expected SHA256 digest of the file, and r represents the file.
 func (c *Client) CreateBlob(ctx context.Context, digest string, r io.Reader) error {
 	return c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil)
 }
 
-// Version returns the Ollama server version as a string.
 func (c *Client) Version(ctx context.Context) (string, error) {
 	var version struct {
 		Version string `json:"version"`

api/client_test.go

@@ -1,8 +1,6 @@
 package api
 
-import (
+import "testing"
-	"testing"
-)
 
 func TestClientFromEnvironment(t *testing.T) {
 	type testCase struct {

api/types.go

@@ -2,8 +2,8 @@ package api
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
-	"log/slog"
 	"math"
 	"os"
 	"reflect"
@@ -12,7 +12,6 @@ import (
 	"time"
 )
 
-// StatusError is an error with and HTTP status code.
 type StatusError struct {
 	StatusCode   int
 	Status       string
@@ -33,7 +32,6 @@ func (e StatusError) Error() string {
 	}
 }
 
-// ImageData represents the raw binary data of an image file.
 type ImageData []byte
 
 // GenerateRequest describes a request sent by [Client.Generate]. While you
@@ -47,9 +45,6 @@ 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"`
 
@@ -82,112 +77,26 @@ type GenerateRequest struct {
 	Options map[string]interface{} `json:"options"`
 }
 
-// ChatRequest describes a request sent by [Client.Chat].
 type ChatRequest struct {
-	// Model is the model name, as in [GenerateRequest].
+	Model     string    `json:"model"`
-	Model string `json:"model"`
+	Messages  []Message `json:"messages"`
-
+	Stream    *bool     `json:"stream,omitempty"`
-	// Messages is the messages of the chat - can be used to keep a chat memory.
+	Format    string    `json:"format"`
-	Messages []Message `json:"messages"`
-
-	// Stream enable streaming of returned response; true by default.
-	Stream *bool `json:"stream,omitempty"`
-
-	// Format is the format to return the response in (e.g. "json").
-	Format string `json:"format"`
-
-	// KeepAlive controls how long the model will stay loaded into memory
-	// 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"`
+	Role    string      `json:"role"` // one of ["system", "user", "assistant"]
-	Content   string      `json:"content"`
+	Content string      `json:"content"`
-	Images    []ImageData `json:"images,omitempty"`
+	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
-// similar to [GenerateResponse].
 type ChatResponse struct {
-	Model      string    `json:"model"`
+	Model     string    `json:"model"`
-	CreatedAt  time.Time `json:"created_at"`
+	CreatedAt time.Time `json:"created_at"`
-	Message    Message   `json:"message"`
+	Message   Message   `json:"message"`
-	DoneReason string    `json:"done_reason,omitempty"`
 
 	Done bool `json:"done"`
 
@@ -203,8 +112,7 @@ type Metrics struct {
 	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
 }
 
-// Options specified in [GenerateRequest], if you add a new option here add it
+// Options specified in GenerateRequest, if you add a new option here add it to the API docs also
-// to the API docs also.
 type Options struct {
 	Runner
 
@@ -214,7 +122,6 @@ 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"`
@@ -231,129 +138,82 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	NumCtx    int   `json:"num_ctx,omitempty"`
+	UseNUMA   bool `json:"numa,omitempty"`
-	NumBatch  int   `json:"num_batch,omitempty"`
+	NumCtx    int  `json:"num_ctx,omitempty"`
-	NumGPU    int   `json:"num_gpu,omitempty"`
+	NumBatch  int  `json:"num_batch,omitempty"`
-	MainGPU   int   `json:"main_gpu,omitempty"`
+	NumGQA    int  `json:"num_gqa,omitempty"`
-	LowVRAM   bool  `json:"low_vram,omitempty"`
+	NumGPU    int  `json:"num_gpu,omitempty"`
-	F16KV     bool  `json:"f16_kv,omitempty"`
+	MainGPU   int  `json:"main_gpu,omitempty"`
-	LogitsAll bool  `json:"logits_all,omitempty"`
+	LowVRAM   bool `json:"low_vram,omitempty"`
-	VocabOnly bool  `json:"vocab_only,omitempty"`
+	F16KV     bool `json:"f16_kv,omitempty"`
-	UseMMap   *bool `json:"use_mmap,omitempty"`
+	LogitsAll bool `json:"logits_all,omitempty"`
-	UseMLock  bool  `json:"use_mlock,omitempty"`
+	VocabOnly bool `json:"vocab_only,omitempty"`
-	NumThread int   `json:"num_thread,omitempty"`
+	UseMMap   bool `json:"use_mmap,omitempty"`
+	UseMLock  bool `json:"use_mlock,omitempty"`
+	NumThread int  `json:"num_thread,omitempty"`
+
+	// Unused: RopeFrequencyBase is ignored. Instead the value in the model will be used
+	RopeFrequencyBase float32 `json:"rope_frequency_base,omitempty"`
+	// Unused: RopeFrequencyScale is ignored. Instead the value in the model will be used
+	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
 }
 
-// EmbedRequest is the request passed to [Client.Embed].
-type EmbedRequest struct {
-	// Model is the model name.
-	Model string `json:"model"`
-
-	// Input is the input to embed.
-	Input any `json:"input"`
-
-	// 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"`
-}
-
-// 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].
 type EmbeddingRequest struct {
-	// Model is the model name.
+	Model     string    `json:"model"`
-	Model string `json:"model"`
+	Prompt    string    `json:"prompt"`
-
-	// Prompt is the textual prompt to embed.
-	Prompt string `json:"prompt"`
-
-	// KeepAlive controls how long the model will stay loaded in memory following
-	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
-	// Options lists model-specific options.
 	Options map[string]interface{} `json:"options"`
 }
 
-// EmbeddingResponse is the response from [Client.Embeddings].
 type EmbeddingResponse struct {
 	Embedding []float64 `json:"embedding"`
 }
 
-// CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
-	Model     string `json:"model"`
+	Model        string `json:"model"`
-	Modelfile string `json:"modelfile"`
+	Path         string `json:"path"`
-	Stream    *bool  `json:"stream,omitempty"`
+	Modelfile    string `json:"modelfile"`
-	Quantize  string `json:"quantize,omitempty"`
+	Stream       *bool  `json:"stream,omitempty"`
-
-	// Deprecated: set the model name with Model instead
-	Name string `json:"name"`
-
-	// Deprecated: set the file content with Modelfile instead
-	Path string `json:"path"`
-
-	// Deprecated: use Quantize instead
 	Quantization string `json:"quantization,omitempty"`
+
+	// Name is deprecated, see Model
+	Name string `json:"name"`
 }
 
-// DeleteRequest is the request passed to [Client.Delete].
 type DeleteRequest struct {
 	Model string `json:"model"`
 
-	// Deprecated: set the model name with Model instead
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
-// ShowRequest is the request passed to [Client.Show].
 type ShowRequest struct {
-	Model  string `json:"model"`
+	Model    string `json:"model"`
-	System string `json:"system"`
+	System   string `json:"system"`
-
-	// Template is deprecated
 	Template string `json:"template"`
-	Verbose  bool   `json:"verbose"`
 
 	Options map[string]interface{} `json:"options"`
 
-	// Deprecated: set the model name with Model instead
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
-// ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License       string         `json:"license,omitempty"`
+	License    string       `json:"license,omitempty"`
-	Modelfile     string         `json:"modelfile,omitempty"`
+	Modelfile  string       `json:"modelfile,omitempty"`
-	Parameters    string         `json:"parameters,omitempty"`
+	Parameters string       `json:"parameters,omitempty"`
-	Template      string         `json:"template,omitempty"`
+	Template   string       `json:"template,omitempty"`
-	System        string         `json:"system,omitempty"`
+	System     string       `json:"system,omitempty"`
-	Details       ModelDetails   `json:"details,omitempty"`
+	Details    ModelDetails `json:"details,omitempty"`
-	Messages      []Message      `json:"messages,omitempty"`
+	Messages   []Message    `json:"messages,omitempty"`
-	ModelInfo     map[string]any `json:"model_info,omitempty"`
-	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
-	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
 }
 
-// CopyRequest is the request passed to [Client.Copy].
 type CopyRequest struct {
 	Source      string `json:"source"`
 	Destination string `json:"destination"`
 }
 
-// PullRequest is the request passed to [Client.Pull].
 type PullRequest struct {
 	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
@@ -361,12 +221,10 @@ type PullRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
 
-	// Deprecated: set the model name with Model instead
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
-// ProgressResponse is the response passed to progress functions like
-// [PullProgressFunc] and [PushProgressFunc].
 type ProgressResponse struct {
 	Status    string `json:"status"`
 	Digest    string `json:"digest,omitempty"`
@@ -374,7 +232,6 @@ type ProgressResponse struct {
 	Completed int64  `json:"completed,omitempty"`
 }
 
-// PushRequest is the request passed to [Client.Push].
 type PushRequest struct {
 	Model    string `json:"model"`
 	Insecure bool   `json:"insecure,omitempty"`
@@ -382,22 +239,15 @@ type PushRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`
 
-	// Deprecated: set the model name with Model instead
+	// Name is deprecated, see Model
 	Name string `json:"name"`
 }
 
-// ListResponse is the response from [Client.List].
 type ListResponse struct {
-	Models []ListModelResponse `json:"models"`
+	Models []ModelResponse `json:"models"`
 }
 
-// ProcessResponse is the response from [Client.Process].
+type ModelResponse struct {
-type ProcessResponse struct {
-	Models []ProcessModelResponse `json:"models"`
-}
-
-// ListModelResponse is a single model description in [ListResponse].
-type ListModelResponse struct {
 	Name       string       `json:"name"`
 	Model      string       `json:"model"`
 	ModifiedAt time.Time    `json:"modified_at"`
@@ -406,53 +256,21 @@ type ListModelResponse struct {
 	Details    ModelDetails `json:"details,omitempty"`
 }
 
-// ProcessModelResponse is a single model description in [ProcessResponse].
-type ProcessModelResponse struct {
-	Name      string       `json:"name"`
-	Model     string       `json:"model"`
-	Size      int64        `json:"size"`
-	Digest    string       `json:"digest"`
-	Details   ModelDetails `json:"details,omitempty"`
-	ExpiresAt time.Time    `json:"expires_at"`
-	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"`
 }
 
-// GenerateResponse is the response passed into [GenerateResponseFunc].
 type GenerateResponse struct {
-	// Model is the model name that generated the response.
+	Model     string    `json:"model"`
-	Model string `json:"model"`
-
-	// CreatedAt is the timestamp of the response.
 	CreatedAt time.Time `json:"created_at"`
+	Response  string    `json:"response"`
 
-	// Response is the textual response itself.
+	Done    bool  `json:"done"`
-	Response string `json:"response"`
-
-	// Done specifies if the response is complete.
-	Done bool `json:"done"`
-
-	// DoneReason is the reason the model stopped generating text.
-	DoneReason string `json:"done_reason,omitempty"`
-
-	// Context is an encoding of the conversation used in this response; this
-	// can be sent in the next request to keep a conversational memory.
 	Context []int `json:"context,omitempty"`
 
 	Metrics
 }
 
-// ModelDetails provides details about a model.
 type ModelDetails struct {
 	ParentModel       string   `json:"parent_model"`
 	Format            string   `json:"format"`
@@ -490,6 +308,8 @@ func (m *Metrics) Summary() {
 	}
 }
 
+var ErrInvalidOpts = errors.New("invalid options")
+
 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
@@ -503,87 +323,76 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 		}
 	}
 
+	invalidOpts := []string{}
 	for key, val := range m {
-		opt, ok := jsonOpts[key]
+		if opt, ok := jsonOpts[key]; ok {
-		if !ok {
+			field := valueOpts.FieldByName(opt.Name)
-			slog.Warn("invalid option provided", "option", key)
+			if field.IsValid() && field.CanSet() {
-			continue
+				if val == nil {
-		}
+					continue
+				}
 
-		field := valueOpts.FieldByName(opt.Name)
+				switch field.Kind() {
-		if field.IsValid() && field.CanSet() {
+				case reflect.Int:
-			if val == nil {
+					switch t := val.(type) {
-				continue
+					case int64:
-			}
+						field.SetInt(t)
-
+					case float64:
-			switch field.Kind() {
+						// when JSON unmarshals numbers, it uses float64, not int
-			case reflect.Int:
+						field.SetInt(int64(t))
-				switch t := val.(type) {
+					default:
-				case int64:
+						return fmt.Errorf("option %q must be of type integer", key)
-					field.SetInt(t)
-				case float64:
-					// when JSON unmarshals numbers, it uses float64, not int
-					field.SetInt(int64(t))
-				default:
-					return fmt.Errorf("option %q must be of type integer", key)
-				}
-			case reflect.Bool:
-				val, ok := val.(bool)
-				if !ok {
-					return fmt.Errorf("option %q must be of type boolean", key)
-				}
-				field.SetBool(val)
-			case reflect.Float32:
-				// JSON unmarshals to float64
-				val, ok := val.(float64)
-				if !ok {
-					return fmt.Errorf("option %q must be of type float32", key)
-				}
-				field.SetFloat(val)
-			case reflect.String:
-				val, ok := val.(string)
-				if !ok {
-					return fmt.Errorf("option %q must be of type string", key)
-				}
-				field.SetString(val)
-			case reflect.Slice:
-				// JSON unmarshals to []interface{}, not []string
-				val, ok := val.([]interface{})
-				if !ok {
-					return fmt.Errorf("option %q must be of type array", key)
-				}
-				// convert []interface{} to []string
-				slice := make([]string, len(val))
-				for i, item := range val {
-					str, ok := item.(string)
-					if !ok {
-						return fmt.Errorf("option %q must be of an array of strings", key)
 					}
-					slice[i] = str
+				case reflect.Bool:
-				}
-				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))
+					field.SetBool(val)
-				} else {
+				case reflect.Float32:
-					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
+					// JSON unmarshals to float64
+					val, ok := val.(float64)
+					if !ok {
+						return fmt.Errorf("option %q must be of type float32", key)
+					}
+					field.SetFloat(val)
+				case reflect.String:
+					val, ok := val.(string)
+					if !ok {
+						return fmt.Errorf("option %q must be of type string", key)
+					}
+					field.SetString(val)
+				case reflect.Slice:
+					// JSON unmarshals to []interface{}, not []string
+					val, ok := val.([]interface{})
+					if !ok {
+						return fmt.Errorf("option %q must be of type array", key)
+					}
+					// convert []interface{} to []string
+					slice := make([]string, len(val))
+					for i, item := range val {
+						str, ok := item.(string)
+						if !ok {
+							return fmt.Errorf("option %q must be of an array of strings", key)
+						}
+						slice[i] = str
+					}
+					field.Set(reflect.ValueOf(slice))
+				default:
+					return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 				}
-			default:
-				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
+		} else {
+			invalidOpts = append(invalidOpts, key)
 		}
 	}
 
+	if len(invalidOpts) > 0 {
+		return fmt.Errorf("%w: %v", ErrInvalidOpts, strings.Join(invalidOpts, ", "))
+	}
 	return nil
 }
 
-// DefaultOptions is the default set of options for [GenerateRequest]; these
-// values are used unless the user specifies other values explicitly.
 func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
@@ -611,11 +420,13 @@ func DefaultOptions() Options {
 			NumCtx:    2048,
 			NumBatch:  512,
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
-			NumThread: 0,  // let the runtime decide
+			NumGQA:    1,
+			NumThread: 0, // let the runtime decide
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   nil,
+			UseMMap:   true,
+			UseNUMA:   false,
 		},
 	}
 }
@@ -624,13 +435,6 @@ type Duration struct {
 	time.Duration
 }
 
-func (d Duration) MarshalJSON() ([]byte, error) {
-	if d.Duration < 0 {
-		return []byte("-1"), nil
-	}
-	return []byte("\"" + d.Duration.String() + "\""), nil
-}
-
 func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 	var v any
 	if err := json.Unmarshal(b, &v); err != nil {
@@ -644,7 +448,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if t < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		} else {
-			d.Duration = time.Duration(int(t) * int(time.Second))
+			d.Duration = time.Duration(t * float64(time.Second))
 		}
 	case string:
 		d.Duration, err = time.ParseDuration(t)
@@ -654,8 +458,6 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if d.Duration < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		}
-	default:
-		return fmt.Errorf("Unsupported type: '%s'", reflect.TypeOf(v))
 	}
 
 	return nil
@@ -711,17 +513,6 @@ 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)
 				}

api/types_test.go

@@ -2,7 +2,6 @@ package api
 
 import (
 	"encoding/json"
-	"errors"
 	"math"
 	"testing"
 	"time"
@@ -22,11 +21,6 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 			req:  `{ "keep_alive": 42 }`,
 			exp:  &Duration{42 * time.Second},
 		},
-		{
-			name: "Positive Float",
-			req:  `{ "keep_alive": 42.5 }`,
-			exp:  &Duration{42 * time.Second},
-		},
 		{
 			name: "Positive Integer String",
 			req:  `{ "keep_alive": "42m" }`,
@@ -37,11 +31,6 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 			req:  `{ "keep_alive": -1 }`,
 			exp:  &Duration{math.MaxInt64},
 		},
-		{
-			name: "Negative Float",
-			req:  `{ "keep_alive": -3.14 }`,
-			exp:  &Duration{math.MaxInt64},
-		},
 		{
 			name: "Negative Integer String",
 			req:  `{ "keep_alive": "-1m" }`,
@@ -59,175 +48,3 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		})
 	}
 }
-
-func TestDurationMarshalUnmarshal(t *testing.T) {
-	tests := []struct {
-		name     string
-		input    time.Duration
-		expected time.Duration
-	}{
-		{
-			"negative duration",
-			time.Duration(-1),
-			time.Duration(math.MaxInt64),
-		},
-		{
-			"positive duration",
-			42 * time.Second,
-			42 * time.Second,
-		},
-		{
-			"another positive duration",
-			42 * time.Minute,
-			42 * time.Minute,
-		},
-		{
-			"zero duration",
-			time.Duration(0),
-			time.Duration(0),
-		},
-		{
-			"max duration",
-			time.Duration(math.MaxInt64),
-			time.Duration(math.MaxInt64),
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			b, err := json.Marshal(Duration{test.input})
-			require.NoError(t, err)
-
-			var d Duration
-			err = json.Unmarshal(b, &d)
-			require.NoError(t, err)
-
-			assert.Equal(t, test.expected, d.Duration, "input %v, marshalled %v, got %v", test.input, string(b), d.Duration)
-		})
-	}
-}
-
-func TestUseMmapParsingFromJSON(t *testing.T) {
-	tr := true
-	fa := false
-	tests := []struct {
-		name string
-		req  string
-		exp  *bool
-	}{
-		{
-			name: "Undefined",
-			req:  `{ }`,
-			exp:  nil,
-		},
-		{
-			name: "True",
-			req:  `{ "use_mmap": true }`,
-			exp:  &tr,
-		},
-		{
-			name: "False",
-			req:  `{ "use_mmap": false }`,
-			exp:  &fa,
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			var oMap map[string]interface{}
-			err := json.Unmarshal([]byte(test.req), &oMap)
-			require.NoError(t, err)
-			opts := DefaultOptions()
-			err = opts.FromMap(oMap)
-			require.NoError(t, err)
-			assert.Equal(t, test.exp, opts.UseMMap)
-		})
-	}
-}
-
-func TestUseMmapFormatParams(t *testing.T) {
-	tr := true
-	fa := false
-	tests := []struct {
-		name string
-		req  map[string][]string
-		exp  *bool
-		err  error
-	}{
-		{
-			name: "True",
-			req: map[string][]string{
-				"use_mmap": {"true"},
-			},
-			exp: &tr,
-			err: nil,
-		},
-		{
-			name: "False",
-			req: map[string][]string{
-				"use_mmap": {"false"},
-			},
-			exp: &fa,
-			err: nil,
-		},
-		{
-			name: "Numeric True",
-			req: map[string][]string{
-				"use_mmap": {"1"},
-			},
-			exp: &tr,
-			err: nil,
-		},
-		{
-			name: "Numeric False",
-			req: map[string][]string{
-				"use_mmap": {"0"},
-			},
-			exp: &fa,
-			err: nil,
-		},
-		{
-			name: "invalid string",
-			req: map[string][]string{
-				"use_mmap": {"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, test.err, err)
-			respVal, ok := resp["use_mmap"]
-			if test.exp != nil {
-				assert.True(t, ok, "resp: %v", resp)
-				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)
-		}
-	}
-}

app/.gitignore

@@ -1 +1,2 @@
 ollama.syso
+app

app/AppDelegate.h

@@ -0,0 +1,7 @@
+#import <Cocoa/Cocoa.h>
+
+@interface AppDelegate : NSObject <NSApplicationDelegate>
+
+- (void)applicationDidFinishLaunching:(NSNotification *)aNotification;
+
+@end

app/README.md

@@ -1,10 +1,6 @@
 # Ollama App
 
-## Linux
+## macOS
-
-TODO
-
-## MacOS
 
 TODO
 

app/app_darwin.go

@@ -0,0 +1,76 @@
+package main
+
+// #cgo CFLAGS: -x objective-c
+// #cgo LDFLAGS: -framework Cocoa -framework LocalAuthentication -framework ServiceManagement
+// #include "app_darwin.h"
+import "C"
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"syscall"
+)
+
+func init() {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		panic(err)
+	}
+
+	ServerLogFile = filepath.Join(home, ".ollama", "logs", "server.log")
+}
+
+func run() {
+	initLogging()
+	slog.Info("ollama macOS app started")
+
+	// Ask to move to applications directory
+	moving := C.askToMoveToApplications()
+	if moving {
+		return
+	}
+
+	C.killOtherInstances()
+
+	code := C.installSymlink()
+	if code != 0 {
+		slog.Error("Failed to install symlink")
+	}
+
+	exe, err := os.Executable()
+	if err != nil {
+		panic(err)
+	}
+
+	var options ServerOptions
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var done chan int
+
+	done, err = SpawnServer(ctx, filepath.Join(filepath.Dir(exe), "..", "Resources", "ollama"), options)
+	if err != nil {
+		slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
+		done = make(chan int, 1)
+		done <- 1
+	}
+
+	// Run the native macOS app
+	// Note: this will block until the app is closed
+	C.run()
+
+	slog.Info("ollama macOS app closed")
+
+	cancel()
+	slog.Info("Waiting for ollama server to shutdown...")
+	if done != nil {
+		<-done
+	}
+	slog.Info("Ollama app exiting")
+}
+
+//export Quit
+func Quit() {
+	syscall.Kill(os.Getpid(), syscall.SIGTERM)
+}

app/app_darwin.h

@@ -0,0 +1,13 @@
+#import <Cocoa/Cocoa.h>
+
+@interface AppDelegate : NSObject <NSApplicationDelegate>
+- (void)applicationDidFinishLaunching:(NSNotification *)aNotification;
+@end
+
+void run();
+void killOtherInstances();
+bool askToMoveToApplications();
+int createSymlinkWithAuthorization();
+int installSymlink();
+extern void Restart();
+extern void Quit();

app/app_darwin.m

@@ -0,0 +1,282 @@
+#import <AppKit/AppKit.h>
+#import <Cocoa/Cocoa.h>
+#import <CoreServices/CoreServices.h>
+#import <Security/Security.h>
+#import <ServiceManagement/ServiceManagement.h>
+#import "app_darwin.h"
+
+@interface AppDelegate ()
+
+@property (strong, nonatomic) NSStatusItem *statusItem;
+
+@end
+
+@implementation AppDelegate
+
+- (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
+    // show status menu
+    NSMenu *menu = [[NSMenu alloc] init];
+
+    NSMenuItem *aboutMenuItem = [[NSMenuItem alloc] initWithTitle:@"About Ollama" action:@selector(aboutOllama) keyEquivalent:@""];
+    [aboutMenuItem setTarget:self];
+    [menu addItem:aboutMenuItem];
+
+    // Settings submenu
+    NSMenu *settingsMenu = [[NSMenu alloc] initWithTitle:@"Settings"];
+
+    // Submenu items
+    NSMenuItem *chooseModelDirectoryItem = [[NSMenuItem alloc] initWithTitle:@"Choose model directory..." action:@selector(chooseModelDirectory) keyEquivalent:@""];
+    [chooseModelDirectoryItem setTarget:self];
+    [chooseModelDirectoryItem setEnabled:YES];
+    [settingsMenu addItem:chooseModelDirectoryItem];
+
+    NSMenuItem *exposeExternallyItem = [[NSMenuItem alloc] initWithTitle:@"Allow external connections" action:@selector(toggleExposeExternally:) keyEquivalent:@""];
+    [exposeExternallyItem setTarget:self];
+    [exposeExternallyItem setState:NSOffState]; // Set initial state to off
+    [exposeExternallyItem setEnabled:YES];
+    [settingsMenu addItem:exposeExternallyItem];
+
+    NSMenuItem *allowCrossOriginItem = [[NSMenuItem alloc] initWithTitle:@"Allow browser requests" action:@selector(toggleCrossOrigin:) keyEquivalent:@""];
+    [allowCrossOriginItem setTarget:self];
+    [allowCrossOriginItem setState:NSOffState]; // Set initial state to off
+    [allowCrossOriginItem setEnabled:YES];
+    [settingsMenu addItem:allowCrossOriginItem];
+
+    NSMenuItem *settingsMenuItem = [[NSMenuItem alloc] initWithTitle:@"Settings" action:nil keyEquivalent:@""];
+    [settingsMenuItem setSubmenu:settingsMenu];
+    [menu addItem:settingsMenuItem];
+
+    [menu addItemWithTitle:@"Quit Ollama" action:@selector(quit) keyEquivalent:@"q"];
+
+    self.statusItem = [[NSStatusBar systemStatusBar] statusItemWithLength:NSVariableStatusItemLength];
+    [self.statusItem addObserver:self forKeyPath:@"button.effectiveAppearance" options:NSKeyValueObservingOptionNew|NSKeyValueObservingOptionInitial context:nil];
+
+    self.statusItem.menu = menu;
+    [self showIcon];
+}
+
+- (void)aboutOllama {
+    [[NSApplication sharedApplication] orderFrontStandardAboutPanel:nil];
+}
+
+- (void)toggleCrossOrigin:(id)sender {
+    NSMenuItem *item = (NSMenuItem *)sender;
+    if ([item state] == NSOffState) {
+        // Do something when cross-origin requests are allowed
+        [item setState:NSOnState];
+    } else {
+        // Do something when cross-origin requests are disallowed
+        [item setState:NSOffState];
+    }
+}
+
+- (void)toggleExposeExternally:(id)sender {
+    NSMenuItem *item = (NSMenuItem *)sender;
+    if ([item state] == NSOffState) {
+        // Do something when Ollama is exposed externally
+        [item setState:NSOnState];
+    } else {
+        // Do something when Ollama is not exposed externally
+        [item setState:NSOffState];
+    }
+}
+
+- (void)chooseModelDirectory {
+    NSOpenPanel *openPanel = [NSOpenPanel openPanel];
+    [openPanel setCanChooseFiles:NO];
+    [openPanel setCanChooseDirectories:YES];
+    [openPanel setAllowsMultipleSelection:NO];
+
+    NSInteger result = [openPanel runModal];
+    if (result == NSModalResponseOK) {
+        NSURL *selectedDirectoryURL = [openPanel URLs].firstObject;
+        // Do something with the selected directory URL
+    }
+}
+
+-(void) showIcon {
+    NSAppearance* appearance = self.statusItem.button.effectiveAppearance;
+    NSString* appearanceName = (NSString*)(appearance.name);
+    NSString* iconName = [[appearanceName lowercaseString] containsString:@"dark"] ? @"iconDark" : @"icon";
+    NSImage* statusImage = [NSImage imageNamed:iconName];
+    [statusImage setTemplate:YES];
+    self.statusItem.button.image = statusImage;
+}
+
+-(void)observeValueForKeyPath:(NSString *)keyPath ofObject:(id)object change:(NSDictionary<NSKeyValueChangeKey,id> *)change context:(void *)context {
+    [self showIcon];
+}
+
+- (void)quit {
+    [NSApp stop:nil];
+}
+
+@end
+
+void run() {
+    @autoreleasepool {
+        [NSApplication sharedApplication];
+        AppDelegate *appDelegate = [[AppDelegate alloc] init];
+        [NSApp setDelegate:appDelegate];
+        [NSApp run];
+    }
+}
+
+// killOtherInstances kills all other instances of the app currently
+// running. This way we can ensure that only the most recently started
+// instance of Ollama is running
+void killOtherInstances() {
+    pid_t pid = getpid();
+    NSArray *all = [[NSWorkspace sharedWorkspace] runningApplications];
+    NSMutableArray *apps = [NSMutableArray array];
+
+    for (NSRunningApplication *app in all) {
+        if ([app.bundleIdentifier isEqualToString:[[NSBundle mainBundle] bundleIdentifier]] ||
+            [app.bundleIdentifier isEqualToString:@"ai.ollama.ollama"] ||
+            [app.bundleIdentifier isEqualToString:@"com.electron.ollama"]) {
+            if (app.processIdentifier != pid) {
+                [apps addObject:app];
+            }
+        }
+    }
+
+    for (NSRunningApplication *app in apps) {
+        kill(app.processIdentifier, SIGTERM);
+    }
+
+    NSDate *startTime = [NSDate date];
+    for (NSRunningApplication *app in apps) {
+        while (!app.terminated) {
+            if (-[startTime timeIntervalSinceNow] >= 5) {
+                kill(app.processIdentifier, SIGKILL);
+                break;
+            }
+
+            [[NSRunLoop currentRunLoop] runUntilDate:[NSDate dateWithTimeIntervalSinceNow:0.1]];
+        }
+    }
+}
+
+bool askToMoveToApplications() {
+    NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
+    if ([bundlePath hasPrefix:@"/Applications"]) {
+        return false;
+    }
+
+    NSAlert *alert = [[NSAlert alloc] init];
+    [alert setMessageText:@"Move to Applications?"];
+    [alert setInformativeText:@"Ollama works best when run from the Applications directory."];
+    [alert addButtonWithTitle:@"Move to Applications"];
+    [alert addButtonWithTitle:@"Don't move"];
+
+    [NSApp activateIgnoringOtherApps:YES];
+
+    if ([alert runModal] != NSAlertFirstButtonReturn) {
+        return false;
+    }
+
+    // move to applications
+    NSString *applicationsPath = @"/Applications";
+    NSString *newPath = [applicationsPath stringByAppendingPathComponent:@"Ollama.app"];
+    NSFileManager *fileManager = [NSFileManager defaultManager];
+
+    // Check if the newPath already exists
+    if ([fileManager fileExistsAtPath:newPath]) {
+        NSError *removeError = nil;
+        [fileManager removeItemAtPath:newPath error:&removeError];
+        if (removeError) {
+            NSLog(@"Error removing file at %@: %@", newPath, removeError);
+            return false; // or handle the error
+        }
+    }
+
+    NSError *moveError = nil;
+    [fileManager moveItemAtPath:bundlePath toPath:newPath error:&moveError];
+    if (moveError) {
+        NSLog(@"Error moving file from %@ to %@: %@", bundlePath, newPath, moveError);
+        return false;
+    }
+
+    NSLog(@"Opening %@", newPath);
+    NSError *error = nil;
+    NSWorkspace *workspace = [NSWorkspace sharedWorkspace];
+#pragma clang diagnostic ignored "-Wdeprecated-declarations"
+    [workspace launchApplicationAtURL:[NSURL fileURLWithPath:newPath]
+               options:NSWorkspaceLaunchNewInstance | NSWorkspaceLaunchDefault
+               configuration:@{}
+               error:&error];
+
+    return true;
+}
+
+int installSymlink() {
+    NSString *linkPath = @"/usr/local/bin/ollama";
+    NSError *error = nil;
+
+    NSFileManager *fileManager = [NSFileManager defaultManager];
+    NSString *symlinkPath = [fileManager destinationOfSymbolicLinkAtPath:linkPath error:&error];
+    NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
+    NSString *execPath = [[NSBundle mainBundle] executablePath];
+    NSString *resPath = [[NSBundle mainBundle] pathForResource:@"ollama" ofType:nil];
+
+    // if the symlink already exists and points to the right place, don't prompt
+    if ([symlinkPath isEqualToString:resPath]) {
+        NSLog(@"symbolic link already exists and points to the right place");
+        return 0;
+    }
+
+    NSString *authorizationPrompt = @"Ollama is trying to install its command line interface (CLI) tool.";
+
+    AuthorizationRef auth = NULL;
+    OSStatus createStatus = AuthorizationCreate(NULL, kAuthorizationEmptyEnvironment, kAuthorizationFlagDefaults, &auth);
+    if (createStatus != errAuthorizationSuccess) {
+        NSLog(@"Error creating authorization");
+        return -1;
+    }
+
+    NSString * bundleIdentifier = [[NSBundle mainBundle] bundleIdentifier];
+    NSString *rightNameString = [NSString stringWithFormat:@"%@.%@", bundleIdentifier, @"auth3"];
+    const char *rightName = rightNameString.UTF8String;
+
+    OSStatus getRightResult = AuthorizationRightGet(rightName, NULL);
+    if (getRightResult == errAuthorizationDenied) {
+        if (AuthorizationRightSet(auth, rightName, (__bridge CFTypeRef _Nonnull)(@(kAuthorizationRuleAuthenticateAsAdmin)), (__bridge CFStringRef _Nullable)(authorizationPrompt), NULL, NULL) != errAuthorizationSuccess) {
+            NSLog(@"Failed to set right");
+            return -1;
+        }
+    }
+
+    AuthorizationItem right = { .name = rightName, .valueLength = 0, .value = NULL, .flags = 0 };
+    AuthorizationRights rights = { .count = 1, .items = &right };
+    AuthorizationFlags flags = (AuthorizationFlags)(kAuthorizationFlagExtendRights | kAuthorizationFlagInteractionAllowed);
+    AuthorizationItem iconAuthorizationItem = {.name = kAuthorizationEnvironmentIcon, .valueLength = 0, .value = NULL, .flags = 0};
+    AuthorizationEnvironment authorizationEnvironment = {.count = 0, .items = NULL};
+
+    BOOL failedToUseSystemDomain = NO;
+    OSStatus copyStatus = AuthorizationCopyRights(auth, &rights, &authorizationEnvironment, flags, NULL);
+    if (copyStatus != errAuthorizationSuccess) {
+        failedToUseSystemDomain = YES;
+
+        if (copyStatus == errAuthorizationCanceled) {
+            NSLog(@"User cancelled authorization");
+            return -1;
+        } else {
+            NSLog(@"Failed copying system domain rights: %d", copyStatus);
+            return -1;
+        }
+    }
+
+    const char *toolPath = "/bin/ln";
+    const char *args[] = {"-s", "-F", [resPath UTF8String], "/usr/local/bin/ollama", NULL};
+    FILE *pipe = NULL;
+
+#pragma clang diagnostic ignored "-Wdeprecated-declarations"
+    OSStatus status = AuthorizationExecuteWithPrivileges(auth, toolPath, kAuthorizationFlagDefaults, (char *const *)args, &pipe);
+    if (status != errAuthorizationSuccess) {
+        NSLog(@"Failed to create symlink");
+        return -1;
+    }
+
+    AuthorizationFree(auth, kAuthorizationFlagDestroyRights);
+    return 0;
+}

app/app_windows.go

@@ -0,0 +1,166 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log"
+	"log/slog"
+	"os"
+	"os/exec"
+	"os/signal"
+	"path/filepath"
+	"strings"
+	"syscall"
+
+	"github.com/ollama/ollama/app/lifecycle"
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/tray"
+	"github.com/ollama/ollama/app/updater"
+)
+
+func init() {
+	AppName += ".exe"
+	CLIName += ".exe"
+	// Logs, configs, downloads go to LOCALAPPDATA
+	localAppData := os.Getenv("LOCALAPPDATA")
+	AppDataDir = filepath.Join(localAppData, "Ollama")
+	AppLogFile = filepath.Join(AppDataDir, "app.log")
+	ServerLogFile = filepath.Join(AppDataDir, "server.log")
+
+	// Executables are stored in APPDATA
+	AppDir = filepath.Join(localAppData, "Programs", "Ollama")
+
+	// Make sure we have PATH set correctly for any spawned children
+	paths := strings.Split(os.Getenv("PATH"), ";")
+	// Start with whatever we find in the PATH/LD_LIBRARY_PATH
+	found := false
+	for _, path := range paths {
+		d, err := filepath.Abs(path)
+		if err != nil {
+			continue
+		}
+		if strings.EqualFold(AppDir, d) {
+			found = true
+		}
+	}
+	if !found {
+		paths = append(paths, AppDir)
+
+		pathVal := strings.Join(paths, ";")
+		slog.Debug("setting PATH=" + pathVal)
+		err := os.Setenv("PATH", pathVal)
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to update PATH: %s", err))
+		}
+	}
+
+	// Make sure our logging dir exists
+	_, err := os.Stat(AppDataDir)
+	if errors.Is(err, os.ErrNotExist) {
+		if err := os.MkdirAll(AppDataDir, 0o755); err != nil {
+			slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
+		}
+	}
+}
+
+func ShowLogs() {
+	cmd_path := "c:\\Windows\\system32\\cmd.exe"
+	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
+	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
+	err := cmd.Start()
+	if err != nil {
+		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
+	}
+}
+
+func Start() {
+	cmd_path := "c:\\Windows\\system32\\cmd.exe"
+	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
+	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
+	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
+	err := cmd.Start()
+	if err != nil {
+		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
+	}
+}
+
+func run() {
+	initLogging()
+
+	slog.Info("ollama windows app started")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var done chan int
+
+	t, err := tray.NewTray()
+	if err != nil {
+		log.Fatalf("Failed to start: %s", err)
+	}
+	callbacks := t.GetCallbacks()
+
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
+
+	go func() {
+		slog.Debug("starting callback loop")
+		for {
+			select {
+			case <-callbacks.Quit:
+				slog.Debug("quit called")
+				t.Quit()
+			case <-signals:
+				slog.Debug("shutting down due to signal")
+				t.Quit()
+			case <-callbacks.Update:
+				err := updater.DoUpgrade(cancel, done)
+				if err != nil {
+					slog.Warn(fmt.Sprintf("upgrade attempt failed: %s", err))
+				}
+			case <-callbacks.ShowLogs:
+				ShowLogs()
+			case <-callbacks.DoFirstUse:
+				err := lifecycle.GetStarted()
+				if err != nil {
+					slog.Warn(fmt.Sprintf("Failed to launch getting started shell: %s", err))
+				}
+			}
+		}
+	}()
+
+	if !store.GetFirstTimeRun() {
+		slog.Debug("First time run")
+		err = t.DisplayFirstUseNotification()
+		if err != nil {
+			slog.Debug(fmt.Sprintf("XXX failed to display first use notification %v", err))
+		}
+		store.SetFirstTimeRun(true)
+	} else {
+		slog.Debug("Not first time, skipping first run notification")
+	}
+
+	if isServerRunning(ctx) {
+		slog.Info("Detected another instance of ollama running, exiting")
+		os.Exit(1)
+	}
+
+	done, err = SpawnServer(ctx, CLIName)
+	if err != nil {
+		// TODO - should we retry in a backoff loop?
+		// TODO - should we pop up a warning and maybe add a menu item to view application logs?
+		slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
+		done = make(chan int, 1)
+		done <- 1
+	}
+
+	updater.StartBackgroundUpdaterChecker(ctx, t.UpdateAvailable)
+
+	t.Run()
+	cancel()
+	slog.Info("Waiting for ollama server to shutdown...")
+	if done != nil {
+		<-done
+	}
+	slog.Info("Ollama app exiting")
+}

app/darwin/Ollama.app/Contents/Info.plist

@@ -0,0 +1,40 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+  <dict>
+    <key>CFBundleDisplayName</key>
+    <string>Ollama</string>
+    <key>CFBundleExecutable</key>
+    <string>Ollama</string>
+    <key>CFBundleIconFile</key>
+    <string>icon.icns</string>
+    <key>CFBundleIdentifier</key>
+    <string>com.ollama.ollama</string>
+    <key>CFBundleInfoDictionaryVersion</key>
+    <string>6.0</string>
+    <key>CFBundleName</key>
+    <string>Ollama</string>
+    <key>CFBundlePackageType</key>
+    <string>APPL</string>
+    <key>CFBundleShortVersionString</key>
+    <string>0.0.0</string>
+    <key>CFBundleVersion</key>
+    <string>0.0.0</string>
+    <key>DTCompiler</key>
+    <string>com.apple.compilers.llvm.clang.1_0</string>
+    <key>DTSDKBuild</key>
+    <string>22E245</string>
+    <key>DTSDKName</key>
+    <string>macosx13.3</string>
+    <key>DTXcode</key>
+    <string>1431</string>
+    <key>DTXcodeBuild</key>
+    <string>14E300c</string>
+    <key>LSApplicationCategoryType</key>
+    <string>public.app-category.developer-tools</string>
+    <key>LSMinimumSystemVersion</key>
+    <string>11.0</string>
+    <key>LSUIElement</key>
+    <true/>
+  </dict>
+</plist>

app/lifecycle/getstarted_darwin.go

@@ -0,0 +1,7 @@
+package lifecycle
+
+import "fmt"
+
+func GetStarted() error {
+	return fmt.Errorf("GetStarted not implemented")
+}

app/lifecycle/getstarted_nonwindows.go

@@ -1,9 +0,0 @@
-//go:build !windows
-
-package lifecycle
-
-import "errors"
-
-func GetStarted() error {
-	return errors.New("not implemented")
-}

app/lifecycle/getstarted_windows.go

@@ -34,6 +34,7 @@ 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)
 	}

app/lifecycle/lifecycle.go

@@ -1,92 +0,0 @@
-package lifecycle
-
-import (
-	"context"
-	"fmt"
-	"log"
-	"log/slog"
-	"os"
-	"os/signal"
-	"syscall"
-
-	"github.com/ollama/ollama/app/store"
-	"github.com/ollama/ollama/app/tray"
-)
-
-func Run() {
-	InitLogging()
-
-	ctx, cancel := context.WithCancel(context.Background())
-	var done chan int
-
-	t, err := tray.NewTray()
-	if err != nil {
-		log.Fatalf("Failed to start: %s", err)
-	}
-	callbacks := t.GetCallbacks()
-
-	signals := make(chan os.Signal, 1)
-	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM)
-
-	go func() {
-		slog.Debug("starting callback loop")
-		for {
-			select {
-			case <-callbacks.Quit:
-				slog.Debug("quit called")
-				t.Quit()
-			case <-signals:
-				slog.Debug("shutting down due to signal")
-				t.Quit()
-			case <-callbacks.Update:
-				err := DoUpgrade(cancel, done)
-				if err != nil {
-					slog.Warn(fmt.Sprintf("upgrade attempt failed: %s", err))
-				}
-			case <-callbacks.ShowLogs:
-				ShowLogs()
-			case <-callbacks.DoFirstUse:
-				err := GetStarted()
-				if err != nil {
-					slog.Warn(fmt.Sprintf("Failed to launch getting started shell: %s", err))
-				}
-			}
-		}
-	}()
-
-	// Are we first use?
-	if !store.GetFirstTimeRun() {
-		slog.Debug("First time run")
-		err = t.DisplayFirstUseNotification()
-		if err != nil {
-			slog.Debug(fmt.Sprintf("XXX failed to display first use notification %v", err))
-		}
-		store.SetFirstTimeRun(true)
-	} else {
-		slog.Debug("Not first time, skipping first run notification")
-	}
-
-	if IsServerRunning(ctx) {
-		slog.Info("Detected another instance of ollama running, exiting")
-		os.Exit(1)
-	} else {
-		done, err = SpawnServer(ctx, CLIName)
-		if err != nil {
-			// TODO - should we retry in a backoff loop?
-			// TODO - should we pop up a warning and maybe add a menu item to view application logs?
-			slog.Error(fmt.Sprintf("Failed to spawn ollama server %s", err))
-			done = make(chan int, 1)
-			done <- 1
-		}
-	}
-
-	StartBackgroundUpdaterChecker(ctx, t.UpdateAvailable)
-
-	t.Run()
-	cancel()
-	slog.Info("Waiting for ollama server to shutdown...")
-	if done != nil {
-		<-done
-	}
-	slog.Info("Ollama app exiting")
-}

app/lifecycle/logging.go

@@ -1,80 +0,0 @@
-package lifecycle
-
-import (
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"strconv"
-	"strings"
-
-	"github.com/ollama/ollama/envconfig"
-)
-
-func InitLogging() {
-	level := slog.LevelInfo
-
-	if envconfig.Debug() {
-		level = slog.LevelDebug
-	}
-
-	var logFile *os.File
-	var err error
-	// Detect if we're a GUI app on windows, and if not, send logs to console
-	if os.Stderr.Fd() != 0 {
-		// Console app detected
-		logFile = os.Stderr
-		// 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, 0o755)
-		if err != nil {
-			slog.Error(fmt.Sprintf("failed to create server log %v", err))
-			return
-		}
-	}
-	handler := slog.NewTextHandler(logFile, &slog.HandlerOptions{
-		Level:     level,
-		AddSource: true,
-		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
-			if attr.Key == slog.SourceKey {
-				source := attr.Value.Any().(*slog.Source)
-				source.File = filepath.Base(source.File)
-			}
-			return attr
-		},
-	})
-
-	slog.SetDefault(slog.New(handler))
-
-	slog.Info("ollama app started")
-}
-
-func rotateLogs(logFile string) {
-	if _, err := os.Stat(logFile); os.IsNotExist(err) {
-		return
-	}
-	index := strings.LastIndex(logFile, ".")
-	pre := logFile[:index]
-	post := "." + logFile[index+1:]
-	for i := LogRotationCount; i > 0; i-- {
-		older := pre + "-" + strconv.Itoa(i) + post
-		newer := pre + "-" + strconv.Itoa(i-1) + post
-		if i == 1 {
-			newer = pre + post
-		}
-		if _, err := os.Stat(newer); err == nil {
-			if _, err := os.Stat(older); err == nil {
-				err := os.Remove(older)
-				if err != nil {
-					slog.Warn("Failed to remove older log", "older", older, "error", err)
-					continue
-				}
-			}
-			err := os.Rename(newer, older)
-			if err != nil {
-				slog.Warn("Failed to rotate log", "older", older, "newer", newer, "error", err)
-			}
-		}
-	}
-}

app/lifecycle/logging_nonwindows.go

@@ -1,9 +0,0 @@
-//go:build !windows
-
-package lifecycle
-
-import "log/slog"
-
-func ShowLogs() {
-	slog.Warn("not implemented")
-}

app/lifecycle/logging_test.go

@@ -1,44 +0,0 @@
-package lifecycle
-
-import (
-	"os"
-	"path/filepath"
-	"strconv"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestRotateLogs(t *testing.T) {
-	logDir := t.TempDir()
-	logFile := filepath.Join(logDir, "testlog.log")
-
-	// No log exists
-	rotateLogs(logFile)
-
-	require.NoError(t, os.WriteFile(logFile, []byte("1"), 0o644))
-	assert.FileExists(t, logFile)
-	// First rotation
-	rotateLogs(logFile)
-	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
-	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
-	assert.NoFileExists(t, logFile)
-
-	// Should be a no-op without a new log
-	rotateLogs(logFile)
-	assert.FileExists(t, filepath.Join(logDir, "testlog-1.log"))
-	assert.NoFileExists(t, filepath.Join(logDir, "testlog-2.log"))
-	assert.NoFileExists(t, logFile)
-
-	for i := 2; i <= LogRotationCount+1; i++ {
-		require.NoError(t, os.WriteFile(logFile, []byte(strconv.Itoa(i)), 0o644))
-		assert.FileExists(t, logFile)
-		rotateLogs(logFile)
-		assert.NoFileExists(t, logFile)
-		for j := 1; j < i; j++ {
-			assert.FileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(j)+".log"))
-		}
-		assert.NoFileExists(t, filepath.Join(logDir, "testlog-"+strconv.Itoa(i+1)+".log"))
-	}
-}

app/lifecycle/logging_windows.go

@@ -1,19 +0,0 @@
-package lifecycle
-
-import (
-	"fmt"
-	"log/slog"
-	"os/exec"
-	"syscall"
-)
-
-func ShowLogs() {
-	cmd_path := "c:\\Windows\\system32\\cmd.exe"
-	slog.Debug(fmt.Sprintf("viewing logs with start %s", AppDataDir))
-	cmd := exec.Command(cmd_path, "/c", "start", AppDataDir)
-	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: false, CreationFlags: 0x08000000}
-	err := cmd.Start()
-	if err != nil {
-		slog.Error(fmt.Sprintf("Failed to open log dir: %s", err))
-	}
-}

app/lifecycle/paths.go

@@ -16,12 +16,11 @@ var (
 	AppDir     = "/opt/Ollama"
 	AppDataDir = "/opt/Ollama"
 	// TODO - should there be a distinct log dir?
-	UpdateStageDir   = "/tmp"
+	UpdateStageDir = "/tmp"
-	AppLogFile       = "/tmp/ollama_app.log"
+	AppLogFile     = "/tmp/ollama_app.log"
-	ServerLogFile    = "/tmp/ollama.log"
+	ServerLogFile  = "/tmp/ollama.log"
-	UpgradeLogFile   = "/tmp/ollama_update.log"
+	UpgradeLogFile = "/tmp/ollama_update.log"
-	Installer        = "OllamaSetup.exe"
+	Installer      = "OllamaSetup.exe"
-	LogRotationCount = 5
 )
 
 func init() {
@@ -70,10 +69,6 @@ func init() {
 				slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
 			}
 		}
-	} else if runtime.GOOS == "darwin" {
+
-		// TODO
-		AppName += ".app"
-		// } else if runtime.GOOS == "linux" {
-		// TODO
 	}
 }

app/log.go

@@ -0,0 +1,44 @@
+package main
+
+import (
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+)
+
+func initLogging() {
+	level := slog.LevelInfo
+
+	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+		level = slog.LevelDebug
+	}
+
+	var logFile *os.File
+	var err error
+	// Detect if we're a GUI app on windows, and if not, send logs to console
+	if os.Stderr.Fd() != 0 {
+		// Console app detected
+		logFile = os.Stderr
+		// TODO - write one-line to the app.log file saying we're running in console mode to help avoid confusion
+	} else {
+		logFile, err = os.OpenFile(AppLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
+		if err != nil {
+			slog.Error(fmt.Sprintf("failed to create server log %v", err))
+			return
+		}
+	}
+	handler := slog.NewTextHandler(logFile, &slog.HandlerOptions{
+		Level:     level,
+		AddSource: true,
+		ReplaceAttr: func(_ []string, attr slog.Attr) slog.Attr {
+			if attr.Key == slog.SourceKey {
+				source := attr.Value.Any().(*slog.Source)
+				source.File = filepath.Base(source.File)
+			}
+			return attr
+		},
+	})
+
+	slog.SetDefault(slog.New(handler))
+}

app/main.go

@@ -2,11 +2,15 @@ package main
 
 // Compile with the following to get rid of the cmd pop up on windows
 // go build -ldflags="-H windowsgui" .
-
+var (
-import (
+	AppName       string
-	"github.com/ollama/ollama/app/lifecycle"
+	CLIName       string
+	AppDir        string
+	AppDataDir    string
+	AppLogFile    string
+	ServerLogFile string
 )
 
 func main() {
-	lifecycle.Run()
+	run()
 }

app/server.go

@@ -1,4 +1,4 @@
-package lifecycle
+package main
 
 import (
 	"context"
@@ -14,37 +14,28 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
-func getCLIFullPath(command string) string {
+type ServerOptions struct {
-	var cmdPath string
+	Cors       bool
-	appExe, err := os.Executable()
+	Expose     bool
-	if err == nil {
+	ModelsPath string
-		cmdPath = filepath.Join(filepath.Dir(appExe), command)
-		_, err := os.Stat(cmdPath)
-		if err == nil {
-			return cmdPath
-		}
-	}
-	cmdPath, err = exec.LookPath(command)
-	if err == nil {
-		_, err := os.Stat(cmdPath)
-		if err == nil {
-			return cmdPath
-		}
-	}
-	pwd, err := os.Getwd()
-	if err == nil {
-		cmdPath = filepath.Join(pwd, command)
-		_, err = os.Stat(cmdPath)
-		if err == nil {
-			return cmdPath
-		}
-	}
-
-	return command
 }
 
-func start(ctx context.Context, command string) (*exec.Cmd, error) {
+func start(ctx context.Context, command string, options ServerOptions) (*exec.Cmd, error) {
-	cmd := getCmd(ctx, getCLIFullPath(command))
+	cmd := getCmd(ctx, command)
+
+	// set environment variables
+	if options.ModelsPath != "" {
+		cmd.Env = append(cmd.Env, fmt.Sprintf("OLLAMA_MODELS=%s", options.ModelsPath))
+	}
+
+	if options.Cors {
+		cmd.Env = append(cmd.Env, "OLLAMA_ORIGINS=*")
+	}
+
+	if options.Expose {
+		cmd.Env = append(cmd.Env, "OLLAMA_HOST=0.0.0.0")
+	}
+
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
 		return nil, fmt.Errorf("failed to spawn server stdout pipe: %w", err)
@@ -54,24 +45,11 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 		return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
 	}
 
-	rotateLogs(ServerLogFile)
+	// TODO - rotation
-	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0o755)
+	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 	if err != nil {
 		return nil, fmt.Errorf("failed to create server log: %w", err)
 	}
-
-	logDir := filepath.Dir(ServerLogFile)
-	_, err = os.Stat(logDir)
-	if err != nil {
-		if !errors.Is(err, os.ErrNotExist) {
-			return nil, fmt.Errorf("stat ollama server log dir %s: %v", logDir, err)
-		}
-
-		if err := os.MkdirAll(logDir, 0o755); err != nil {
-			return nil, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
-		}
-	}
-
 	go func() {
 		defer logFile.Close()
 		io.Copy(logFile, stdout) //nolint:errcheck
@@ -125,20 +103,25 @@ func start(ctx context.Context, command string) (*exec.Cmd, error) {
 	return cmd, nil
 }
 
-func SpawnServer(ctx context.Context, command string) (chan int, error) {
+func SpawnServer(ctx context.Context, command string, options ServerOptions) (chan int, error) {
+	logDir := filepath.Dir(ServerLogFile)
+	_, err := os.Stat(logDir)
+	if errors.Is(err, os.ErrNotExist) {
+		if err := os.MkdirAll(logDir, 0o755); err != nil {
+			return nil, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
+		}
+	}
+
 	done := make(chan int)
 
 	go func() {
 		// Keep the server running unless we're shuttind down the app
 		crashCount := 0
 		for {
-			slog.Info("starting server...")
+			slog.Info(fmt.Sprintf("starting server..."))
-			cmd, err := start(ctx, command)
+			cmd, err := start(ctx, command, options)
 			if err != nil {
-				crashCount++
 				slog.Error(fmt.Sprintf("failed to start server %s", err))
-				time.Sleep(500 * time.Millisecond * time.Duration(crashCount))
-				continue
 			}
 
 			cmd.Wait() //nolint:errcheck
@@ -164,7 +147,7 @@ func SpawnServer(ctx context.Context, command string) (chan int, error) {
 	return done, nil
 }
 
-func IsServerRunning(ctx context.Context) bool {
+func isServerRunning(ctx context.Context) bool {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
 		slog.Info("unable to connect to server")

app/server_darwin.go

@@ -1,6 +1,4 @@
-//go:build !windows
+package main
-
-package lifecycle
 
 import (
 	"context"

app/server_windows.go

@@ -1,4 +1,4 @@
-package lifecycle
+package main
 
 import (
 	"context"
@@ -24,8 +24,7 @@ func terminate(cmd *exec.Cmd) error {
 	if err != nil {
 		return err
 	}
-	//nolint:errcheck
+	defer dll.Release() // nolint: errcheck
-	defer dll.Release()
 
 	pid := cmd.Process.Pid
 
@@ -74,8 +73,7 @@ func isProcessExited(pid int) (bool, error) {
 	if err != nil {
 		return false, fmt.Errorf("failed to open process: %v", err)
 	}
-	//nolint:errcheck
+	defer windows.CloseHandle(hProcess) // nolint: errcheck
-	defer windows.CloseHandle(hProcess)
 
 	var exitCode uint32
 	err = windows.GetExitCodeProcess(hProcess, &exitCode)

app/store/store.go

@@ -29,6 +29,7 @@ func GetID() string {
 		initStore()
 	}
 	return store.ID
+
 }
 
 func GetFirstTimeRun() bool {

app/tray/tray_darwin.go

@@ -1,13 +1,11 @@
-//go:build !windows
-
 package tray
 
 import (
-	"errors"
+	"fmt"
 
 	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
-	return nil, errors.New("not implemented")
+	return nil, fmt.Errorf("NOT IMPLEMENTED YET")
 }

app/tray/wintray/eventloop.go

@@ -11,7 +11,9 @@ import (
 	"golang.org/x/sys/windows"
 )
 
-var quitOnce sync.Once
+var (
+	quitOnce sync.Once
+)
 
 func (t *winTray) Run() {
 	nativeLoop()
@@ -45,6 +47,7 @@ func nativeLoop() {
 		default:
 			pTranslateMessage.Call(uintptr(unsafe.Pointer(m))) //nolint:errcheck
 			pDispatchMessage.Call(uintptr(unsafe.Pointer(m)))  //nolint:errcheck
+
 		}
 	}
 }
@@ -157,8 +160,8 @@ func (t *winTray) wndProc(hWnd windows.Handle, message uint32, wParam, lParam ui
 		lResult, _, _ = pDefWindowProc.Call(
 			uintptr(hWnd),
 			uintptr(message),
-			wParam,
+			uintptr(wParam),
-			lParam,
+			uintptr(lParam),
 		)
 	}
 	return

app/tray/wintray/menus.go

@@ -1,71 +1,71 @@
-//go:build windows
+//go:build windows
-
+
-package wintray
+package wintray
-
+
-import (
+import (
-	"fmt"
+	"fmt"
-	"log/slog"
+	"log/slog"
-	"unsafe"
+	"unsafe"
-
+
-	"golang.org/x/sys/windows"
+	"golang.org/x/sys/windows"
-)
+)
-
+
-const (
+const (
-	updateAvailableMenuID = 1
+	updatAvailableMenuID = 1
-	updateMenuID          = updateAvailableMenuID + 1
+	updateMenuID         = updatAvailableMenuID + 1
-	separatorMenuID       = updateMenuID + 1
+	separatorMenuID      = updateMenuID + 1
-	diagLogsMenuID        = separatorMenuID + 1
+	diagLogsMenuID       = separatorMenuID + 1
-	diagSeparatorMenuID   = diagLogsMenuID + 1
+	diagSeparatorMenuID  = diagLogsMenuID + 1
-	quitMenuID            = diagSeparatorMenuID + 1
+	quitMenuID           = diagSeparatorMenuID + 1
-)
+)
-
+
-func (t *winTray) initMenus() error {
+func (t *winTray) initMenus() error {
-	if err := t.addOrUpdateMenuItem(diagLogsMenuID, 0, diagLogsMenuTitle, false); err != nil {
+	if err := t.addOrUpdateMenuItem(diagLogsMenuID, 0, diagLogsMenuTitle, false); err != nil {
-		return fmt.Errorf("unable to create menu entries %w\n", err)
+		return fmt.Errorf("unable to create menu entries %w\n", err)
-	}
+	}
-	if err := t.addSeparatorMenuItem(diagSeparatorMenuID, 0); err != nil {
+	if err := t.addSeparatorMenuItem(diagSeparatorMenuID, 0); err != nil {
-		return fmt.Errorf("unable to create menu entries %w", err)
+		return fmt.Errorf("unable to create menu entries %w", err)
-	}
+	}
-	if err := t.addOrUpdateMenuItem(quitMenuID, 0, quitMenuTitle, false); err != nil {
+	if err := t.addOrUpdateMenuItem(quitMenuID, 0, quitMenuTitle, false); err != nil {
-		return fmt.Errorf("unable to create menu entries %w\n", err)
+		return fmt.Errorf("unable to create menu entries %w\n", err)
-	}
+	}
-	return nil
+	return nil
-}
+}
-
+
-func (t *winTray) UpdateAvailable(ver string) error {
+func (t *winTray) UpdateAvailable(ver string) error {
-	if !t.updateNotified {
+	if !t.updateNotified {
-		slog.Debug("updating menu and sending notification for new update")
+		slog.Debug("updating menu and sending notification for new update")
-		if err := t.addOrUpdateMenuItem(updateAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
+		if err := t.addOrUpdateMenuItem(updatAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
-			return fmt.Errorf("unable to create menu entries %w", err)
+			return fmt.Errorf("unable to create menu entries %w", err)
-		}
+		}
-		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenutTitle, false); err != nil {
+		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenutTitle, false); err != nil {
-			return fmt.Errorf("unable to create menu entries %w", err)
+			return fmt.Errorf("unable to create menu entries %w", err)
-		}
+		}
-		if err := t.addSeparatorMenuItem(separatorMenuID, 0); err != nil {
+		if err := t.addSeparatorMenuItem(separatorMenuID, 0); err != nil {
-			return fmt.Errorf("unable to create menu entries %w", err)
+			return fmt.Errorf("unable to create menu entries %w", err)
-		}
+		}
-		iconFilePath, err := iconBytesToFilePath(wt.updateIcon)
+		iconFilePath, err := iconBytesToFilePath(wt.updateIcon)
-		if err != nil {
+		if err != nil {
-			return fmt.Errorf("unable to write icon data to temp file: %w", err)
+			return fmt.Errorf("unable to write icon data to temp file: %w", err)
-		}
+		}
-		if err := wt.setIcon(iconFilePath); err != nil {
+		if err := wt.setIcon(iconFilePath); err != nil {
-			return fmt.Errorf("unable to set icon: %w", err)
+			return fmt.Errorf("unable to set icon: %w", err)
-		}
+		}
-		t.updateNotified = true
+		t.updateNotified = true
-
+
-		t.pendingUpdate = true
+		t.pendingUpdate = true
-		// Now pop up the notification
+		// Now pop up the notification
-		t.muNID.Lock()
+		t.muNID.Lock()
-		defer t.muNID.Unlock()
+		defer t.muNID.Unlock()
-		copy(t.nid.InfoTitle[:], windows.StringToUTF16(updateTitle))
+		copy(t.nid.InfoTitle[:], windows.StringToUTF16(updateTitle))
-		copy(t.nid.Info[:], windows.StringToUTF16(fmt.Sprintf(updateMessage, ver)))
+		copy(t.nid.Info[:], windows.StringToUTF16(fmt.Sprintf(updateMessage, ver)))
-		t.nid.Flags |= NIF_INFO
+		t.nid.Flags |= NIF_INFO
-		t.nid.Timeout = 10
+		t.nid.Timeout = 10
-		t.nid.Size = uint32(unsafe.Sizeof(*wt.nid))
+		t.nid.Size = uint32(unsafe.Sizeof(*wt.nid))
-		err = t.nid.modify()
+		err = t.nid.modify()
-		if err != nil {
+		if err != nil {
-			return err
+			return err
-		}
+		}
-	}
+	}
-	return nil
+	return nil
-}
+}

app/tray/wintray/tray.go

@@ -11,12 +11,10 @@ import (
 	"path/filepath"
 	"sort"
 	"sync"
-	"syscall"
 	"unsafe"
 
-	"golang.org/x/sys/windows"
-
 	"github.com/ollama/ollama/app/tray/commontray"
+	"golang.org/x/sys/windows"
 )
 
 // Helpful sources: https://github.com/golang/exp/blob/master/shiny/driver/internal/win32
@@ -188,7 +186,7 @@ func (t *winTray) initInstance() error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid = &notifyIconData{
-		Wnd:             t.window,
+		Wnd:             windows.Handle(t.window),
 		ID:              100,
 		Flags:           NIF_MESSAGE,
 		CallbackMessage: t.wmSystrayMessage,
@@ -199,6 +197,7 @@ func (t *winTray) initInstance() error {
 }
 
 func (t *winTray) createMenu() error {
+
 	menuHandle, _, err := pCreatePopupMenu.Call()
 	if menuHandle == 0 {
 		return err
@@ -247,7 +246,7 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title
 	mi := menuItemInfo{
 		Mask:     MIIM_FTYPE | MIIM_STRING | MIIM_ID | MIIM_STATE,
 		Type:     MFT_STRING,
-		ID:       menuItemId,
+		ID:       uint32(menuItemId),
 		TypeData: titlePtr,
 		Cch:      uint32(len(title)),
 	}
@@ -303,10 +302,11 @@ func (t *winTray) addOrUpdateMenuItem(menuItemId uint32, parentId uint32, title
 }
 
 func (t *winTray) addSeparatorMenuItem(menuItemId, parentId uint32) error {
+
 	mi := menuItemInfo{
 		Mask: MIIM_FTYPE | MIIM_ID | MIIM_STATE,
 		Type: MFT_SEPARATOR,
-		ID:   menuItemId,
+		ID:   uint32(menuItemId),
 	}
 
 	mi.Size = uint32(unsafe.Sizeof(mi))
@@ -416,7 +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, 0o644); err != nil {
+		if err := os.WriteFile(iconFilePath, iconBytes, 0644); err != nil {
 			return "", err
 		}
 	}
@@ -426,6 +426,7 @@ func iconBytesToFilePath(iconBytes []byte) (string, error) {
 // Loads an image from file and shows it in tray.
 // Shell_NotifyIcon: https://msdn.microsoft.com/en-us/library/windows/desktop/bb762159(v=vs.85).aspx
 func (t *winTray) setIcon(src string) error {
+
 	h, err := t.loadIconFrom(src)
 	if err != nil {
 		return err
@@ -434,12 +435,7 @@ func (t *winTray) setIcon(src string) error {
 	t.muNID.Lock()
 	defer t.muNID.Unlock()
 	t.nid.Icon = h
-	t.nid.Flags |= NIF_ICON | NIF_TIP
+	t.nid.Flags |= NIF_ICON
-	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()
@@ -448,6 +444,7 @@ func (t *winTray) setIcon(src string) error {
 // Loads an image from file to be shown in tray or menu item.
 // LoadImage: https://msdn.microsoft.com/en-us/library/windows/desktop/ms648045(v=vs.85).aspx
 func (t *winTray) loadIconFrom(src string) (windows.Handle, error) {
+
 	// Save and reuse handles of loaded images
 	t.muLoadedImages.RLock()
 	h, ok := t.loadedImages[src]

app/tray/wintray/w32api.go

@@ -61,7 +61,6 @@ const (
 	MIIM_SUBMENU        = 0x00000004
 	MIM_APPLYTOSUBMENUS = 0x80000000
 	NIF_ICON            = 0x00000002
-	NIF_TIP             = 0x00000004
 	NIF_INFO            = 0x00000010
 	NIF_MESSAGE         = 0x00000001
 	SW_HIDE             = 0

app/updater/updater.go

@@ -1,4 +1,4 @@
-package lifecycle
+package updater
 
 import (
 	"context"
@@ -15,7 +15,6 @@ import (
 	"path"
 	"path/filepath"
 	"runtime"
-	"strconv"
 	"strings"
 	"time"
 
@@ -23,6 +22,10 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
+var (
+	UpdateStageDir string
+)
+
 var (
 	UpdateCheckURLBase  = "https://ollama.com/api/update"
 	UpdateDownloaded    = false
@@ -47,7 +50,7 @@ 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", strconv.FormatInt(time.Now().Unix(), 10))
+	query.Add("ts", fmt.Sprintf("%d", time.Now().Unix()))
 
 	nonce, err := auth.NewNonce(rand.Reader, 16)
 	if err != nil {
@@ -79,7 +82,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 	}
 	defer resp.Body.Close()
 
-	if resp.StatusCode == http.StatusNoContent {
+	if resp.StatusCode == 204 {
 		slog.Debug("check update response 204 (current version is up to date)")
 		return false, updateResp
 	}
@@ -88,7 +91,7 @@ func IsNewReleaseAvailable(ctx context.Context) (bool, UpdateResponse) {
 		slog.Warn(fmt.Sprintf("failed to read body response: %s", err))
 	}
 
-	if resp.StatusCode != http.StatusOK {
+	if resp.StatusCode != 200 {
 		slog.Info(fmt.Sprintf("check update error %d - %.96s", resp.StatusCode, string(body)))
 		return false, updateResp
 	}
@@ -115,7 +118,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 	if err != nil {
 		return fmt.Errorf("error checking update: %w", err)
 	}
-	if resp.StatusCode != http.StatusOK {
+	if resp.StatusCode != 200 {
 		return fmt.Errorf("unexpected status attempting to download update %d", resp.StatusCode)
 	}
 	resp.Body.Close()
@@ -124,7 +127,7 @@ func DownloadNewRelease(ctx context.Context, updateResp UpdateResponse) error {
 		slog.Debug("no etag detected, falling back to filename based dedup")
 		etag = "_"
 	}
-	filename := Installer
+	filename := "OllamaSetup.exe"
 	_, params, err := mime.ParseMediaType(resp.Header.Get("content-disposition"))
 	if err == nil {
 		filename = params["filename"]

app/updater/updater_darwin.go

@@ -1,12 +1,10 @@
-//go:build !windows
+package updater
-
-package lifecycle
 
 import (
 	"context"
-	"errors"
+	"fmt"
 )
 
 func DoUpgrade(cancel context.CancelFunc, done chan int) error {
-	return errors.New("not implemented")
+	return fmt.Errorf("DoUpgrade not yet implemented")
 }

app/updater/updater_windows.go

@@ -1,8 +1,7 @@
-package lifecycle
+package updater
 
 import (
 	"context"
-	"errors"
 	"fmt"
 	"log/slog"
 	"os"
@@ -10,13 +9,19 @@ import (
 	"path/filepath"
 )
 
+func init() {
+	UpdateStageDir = filepath.Join(os.Getenv("LOCALAPPDATA"), "Ollama", "updates")
+}
+
 func DoUpgrade(cancel context.CancelFunc, done chan int) error {
+	logFile := filepath.Join(os.Getenv("LOCALAPPDATA"), "Ollama", "upgrade.log")
+
 	files, err := filepath.Glob(filepath.Join(UpdateStageDir, "*", "*.exe")) // TODO generalize for multiplatform
 	if err != nil {
 		return fmt.Errorf("failed to lookup downloads: %s", err)
 	}
 	if len(files) == 0 {
-		return errors.New("no update downloads found")
+		return fmt.Errorf("no update downloads found")
 	} else if len(files) > 1 {
 		// Shouldn't happen
 		slog.Warn(fmt.Sprintf("multiple downloads found, using first one %v", files))
@@ -24,21 +29,24 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 	installerExe := files[0]
 
 	slog.Info("starting upgrade with " + installerExe)
-	slog.Info("upgrade log file " + UpgradeLogFile)
+	slog.Info("upgrade log file " + logFile)
 
 	// When running in debug mode, we'll be "verbose" and let the installer pop up and prompt
 	installArgs := []string{
-		"/CLOSEAPPLICATIONS",                    // Quit the tray app if it's still running
+		"/CLOSEAPPLICATIONS",             // Quit the tray app if it's still running
-		"/LOG=" + filepath.Base(UpgradeLogFile), // Only relative seems reliable, so set pwd
+		"/LOG=" + filepath.Base(logFile), // Only relative seems reliable, so set pwd
-		"/FORCECLOSEAPPLICATIONS",               // Force close the tray app - might be needed
+		"/FORCECLOSEAPPLICATIONS",        // Force close the tray app - might be needed
 	}
-	// make the upgrade as quiet as possible (no GUI, no prompts)
+	// When we're not in debug mode, make the upgrade as quiet as possible (no GUI, no prompts)
+	// TODO - temporarily disable since we're pinning in debug mode for the preview
+	// if debug := os.Getenv("OLLAMA_DEBUG"); debug == "" {
 	installArgs = append(installArgs,
 		"/SP", // Skip the "This will install... Do you wish to continue" prompt
 		"/SUPPRESSMSGBOXES",
 		"/SILENT",
 		"/VERYSILENT",
 	)
+	// }
 
 	// Safeguard in case we have requests in flight that need to drain...
 	slog.Info("Waiting for server to shutdown")
@@ -51,7 +59,7 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 	}
 
 	slog.Debug(fmt.Sprintf("starting installer: %s %v", installerExe, installArgs))
-	os.Chdir(filepath.Dir(UpgradeLogFile)) //nolint:errcheck
+	os.Chdir(filepath.Dir(logFile)) //nolint:errcheck
 	cmd := exec.Command(installerExe, installArgs...)
 
 	if err := cmd.Start(); err != nil {
@@ -65,7 +73,7 @@ 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 errors.New("installer process did not start")
+		return fmt.Errorf("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?

app/windows/ollama.iss

@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64compatible arm64
+ArchitecturesAllowed=x64 arm64
-ArchitecturesInstallIn64BitMode=x64compatible arm64
+ArchitecturesInstallIn64BitMode=x64 arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@@ -48,7 +48,6 @@ OutputDir=..\dist\
 SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
-RestartIfNeededByRun=no
 
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
@@ -87,21 +86,16 @@ Name: "english"; MessagesFile: "compiler:Default.isl"
 DialogFontSize=12
 
 [Files]
-#if DirExists("..\dist\windows-amd64")
+Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\dist\windows-amd64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit
+Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+Source: "..\dist\windows-amd64\ollama_runners\*"; DestDir: "{app}\ollama_runners"; 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
+#if DirExists("..\dist\windows-amd64\rocm")
+  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
+#endif
+
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@@ -109,9 +103,6 @@ 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]
@@ -131,10 +122,6 @@ 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.
@@ -142,7 +129,7 @@ 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 llama3.1
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3
 ;ClickFinish=%n
 
 [Registry]
@@ -167,39 +154,3 @@ 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;

app/windows/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 llama3.1"
+write-host "`tollama run llama2"
 write-host ""

auth/auth.go

@@ -5,50 +5,17 @@ import (
 	"context"
 	"crypto/rand"
 	"encoding/base64"
-	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
-	"strings"
 
 	"golang.org/x/crypto/ssh"
 )
 
 const defaultPrivateKey = "id_ed25519"
 
-func keyPath() (string, error) {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return "", err
-	}
-
-	return filepath.Join(home, ".ollama", defaultPrivateKey), nil
-}
-
-func GetPublicKey() (string, error) {
-	keyPath, err := keyPath()
-	if err != nil {
-		return "", err
-	}
-
-	privateKeyFile, err := os.ReadFile(keyPath)
-	if err != nil {
-		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
-		return "", err
-	}
-
-	privateKey, err := ssh.ParsePrivateKey(privateKeyFile)
-	if err != nil {
-		return "", err
-	}
-
-	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
-
-	return strings.TrimSpace(string(publicKey)), nil
-}
-
 func NewNonce(r io.Reader, length int) (string, error) {
 	nonce := make([]byte, length)
 	if _, err := io.ReadFull(r, nonce); err != nil {
@@ -59,11 +26,13 @@ func NewNonce(r io.Reader, length int) (string, error) {
 }
 
 func Sign(ctx context.Context, bts []byte) (string, error) {
-	keyPath, err := keyPath()
+	home, err := os.UserHomeDir()
 	if err != nil {
 		return "", err
 	}
 
+	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
+
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
@@ -79,7 +48,7 @@ func Sign(ctx context.Context, bts []byte) (string, error) {
 	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
 	parts := bytes.Split(publicKey, []byte(" "))
 	if len(parts) < 2 {
-		return "", errors.New("malformed public key")
+		return "", fmt.Errorf("malformed public key")
 	}
 
 	signedData, err := privateKey.Sign(rand.Reader, bts)

build/darwin/amd64/placeholder

@@ -1 +0,0 @@
-This is here to make sure the build/ directory exists for the go:embed command

build/darwin/arm64/placeholder

@@ -1 +0,0 @@
-This is here to make sure the build/ directory exists for the go:embed command

build/embed_darwin_amd64.go

@@ -1,8 +0,0 @@
-package build
-
-import "embed"
-
-// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
-
-//go:embed darwin/amd64/*
-var EmbedFS embed.FS

build/embed_darwin_arm64.go

@@ -1,8 +0,0 @@
-package build
-
-import "embed"
-
-// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
-
-//go:embed darwin/arm64/*
-var EmbedFS embed.FS

build/embed_linux.go

@@ -1,6 +0,0 @@
-package build
-
-import "embed"
-
-//go:embed linux/*
-var EmbedFS embed.FS

build/embed_unused.go

@@ -1,8 +0,0 @@
-//go:build !linux && !darwin
-
-package build
-
-import "embed"
-
-// unused on windows
-var EmbedFS embed.FS

build/linux/amd64/placeholder

@@ -1 +0,0 @@
-This is here to make sure the build/ directory exists for the go:embed command

build/linux/arm64/placeholder

@@ -1 +0,0 @@
-This is here to make sure the build/ directory exists for the go:embed command

cmd/cmd.go

@@ -2,7 +2,6 @@ package cmd
 
 import (
 	"archive/zip"
-	"bufio"
 	"bytes"
 	"context"
 	"crypto/ed25519"
@@ -13,7 +12,6 @@ import (
 	"fmt"
 	"io"
 	"log"
-	"math"
 	"net"
 	"net/http"
 	"os"
@@ -21,29 +19,23 @@ import (
 	"path/filepath"
 	"regexp"
 	"runtime"
-	"slices"
-	"strconv"
 	"strings"
-	"sync/atomic"
 	"syscall"
 	"time"
 
 	"github.com/containerd/console"
-	"github.com/mattn/go-runewidth"
+
 	"github.com/olekukonko/tablewriter"
 	"github.com/spf13/cobra"
 	"golang.org/x/crypto/ssh"
+	"golang.org/x/exp/slices"
 	"golang.org/x/term"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/auth"
-	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/server"
-	"github.com/ollama/ollama/types/errtypes"
-	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
 
@@ -62,13 +54,12 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	f, err := os.Open(filename)
+	modelfile, err := os.ReadFile(filename)
 	if err != nil {
 		return err
 	}
-	defer f.Close()
 
-	modelfile, err := parser.ParseFile(f)
+	commands, err := parser.Parse(bytes.NewReader(modelfile))
 	if err != nil {
 		return err
 	}
@@ -81,12 +72,11 @@ 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 {
+	for _, c := range commands {
-		switch modelfile.Commands[i].Name {
+		switch c.Name {
 		case "model", "adapter":
-			path := modelfile.Commands[i].Args
+			path := c.Args
 			if path == "~" {
 				path = home
 			} else if strings.HasPrefix(path, "~/") {
@@ -98,7 +88,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 			}
 
 			fi, err := os.Stat(path)
-			if errors.Is(err, os.ErrNotExist) && modelfile.Commands[i].Name == "model" {
+			if errors.Is(err, os.ErrNotExist) && c.Name == "model" {
 				continue
 			} else if err != nil {
 				return err
@@ -116,12 +106,18 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				path = tempfile
 			}
 
-			digest, err := createBlob(cmd, client, path, spinner)
+			digest, err := createBlob(cmd, client, path)
 			if err != nil {
 				return err
 			}
 
-			modelfile.Commands[i].Args = "@" + digest
+			name := c.Name
+			if c.Name == "model" {
+				name = "from"
+			}
+
+			re := regexp.MustCompile(fmt.Sprintf(`(?im)^(%s)\s+%s\s*$`, name, c.Args))
+			modelfile = re.ReplaceAll(modelfile, []byte("$1 @"+digest))
 		}
 	}
 
@@ -149,9 +145,9 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	quantize, _ := cmd.Flags().GetString("quantize")
+	quantization, _ := cmd.Flags().GetString("quantization")
 
-	request := api.CreateRequest{Name: args[0], Modelfile: modelfile.String(), Quantize: quantize}
+	request := api.CreateRequest{Name: args[0], Modelfile: string(modelfile), Quantization: quantization}
 	if err := client.Create(cmd.Context(), &request, fn); err != nil {
 		return err
 	}
@@ -166,6 +162,9 @@ func tempZipFiles(path string) (string, error) {
 	}
 	defer tempfile.Close()
 
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
 	detectContentType := func(path string) (string, error) {
 		f, err := os.Open(path)
 		if err != nil {
@@ -206,17 +205,11 @@ 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
 		files = append(files, pt...)
-	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/zip"); len(pt) > 0 {
+	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/octet-stream"); len(pt) > 0 {
 		// pytorch files might also be unresolved git lfs references; skip if they are
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
@@ -231,14 +224,6 @@ 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
@@ -248,9 +233,6 @@ func tempZipFiles(path string) (string, error) {
 		files = append(files, tks...)
 	}
 
-	zipfile := zip.NewWriter(tempfile)
-	defer zipfile.Close()
-
 	for _, file := range files {
 		f, err := os.Open(file)
 		if err != nil {
@@ -268,11 +250,6 @@ 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
@@ -286,20 +263,13 @@ func tempZipFiles(path string) (string, error) {
 	return tempfile.Name(), nil
 }
 
-func createBlob(cmd *cobra.Command, client *api.Client, path string, spinner *progress.Spinner) (string, error) {
+func createBlob(cmd *cobra.Command, client *api.Client, path string) (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
@@ -309,83 +279,46 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string, spinner *pr
 		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, io.TeeReader(bin, &pw)); err != nil {
+	if err = client.CreateBlob(cmd.Context(), digest, bin); err != nil {
 		return "", err
 	}
 	return digest, nil
 }
 
-type progressWriter struct {
+func RunHandler(cmd *cobra.Command, args []string) error {
-	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{
+	name := args[0]
-		Model:     opts.Model,
-		KeepAlive: opts.KeepAlive,
-	}
 
-	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+	// check if the model exists on the server
-}
+	show, err := client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-
+	var statusError api.StatusError
-func StopHandler(cmd *cobra.Command, args []string) error {
+	switch {
-	opts := &runOptions{
+	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
-		Model:     args[0],
+		if err := PullHandler(cmd, []string{name}); err != nil {
-		KeepAlive: &api.Duration{Duration: 0},
+			return err
-	}
-	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 {
+		show, err = client.Show(cmd.Context(), &api.ShowRequest{Name: name})
+		if err != nil {
+			return err
+		}
+	case err != nil:
+		return err
+	}
+
 	interactive := true
 
 	opts := runOptions{
-		Model:    args[0],
+		Model:       args[0],
-		WordWrap: os.Getenv("TERM") == "xterm-256color",
+		WordWrap:    os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]interface{}{},
+		Options:     map[string]interface{}{},
+		MultiModal:  slices.Contains(show.Details.Families, "clip"),
+		ParentModel: show.Details.ParentModel,
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -394,18 +327,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.Format = format
 
-	keepAlive, err := cmd.Flags().GetString("keepalive")
-	if err != nil {
-		return err
-	}
-	if keepAlive != "" {
-		d, err := time.ParseDuration(keepAlive)
-		if err != nil {
-			return err
-		}
-		opts.KeepAlive = &api.Duration{Duration: d}
-	}
-
 	prompts := args[1:]
 	// prepend stdin to the prompt if provided
 	if !term.IsTerminal(int(os.Stdin.Fd())) {
@@ -429,94 +350,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 	opts.WordWrap = !nowrap
 
-	// Fill out the rest of the options based on information about the
+	if !interactive {
-	// model.
+		return generate(cmd, opts)
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
 	}
 
-	name := args[0]
+	return generateInteractive(cmd, opts)
-	info, err := func() (*api.ShowResponse, error) {
-		showReq := &api.ShowRequest{Name: name}
-		info, err := client.Show(cmd.Context(), showReq)
-		var se api.StatusError
-		if errors.As(err, &se) && se.StatusCode == http.StatusNotFound {
-			if err := PullHandler(cmd, []string{name}); err != nil {
-				return nil, err
-			}
-			return client.Show(cmd.Context(), &api.ShowRequest{Name: name})
-		}
-		return info, err
-	}()
-	if err != nil {
-		return err
-	}
-
-	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
-	opts.ParentModel = info.Details.ParentModel
-
-	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)
-}
-
-func errFromUnknownKey(unknownKeyErr error) error {
-	// find SSH public key in the error message
-	sshKeyPattern := `ssh-\w+ [^\s"]+`
-	re := regexp.MustCompile(sshKeyPattern)
-	matches := re.FindStringSubmatch(unknownKeyErr.Error())
-
-	if len(matches) > 0 {
-		serverPubKey := matches[0]
-
-		localPubKey, err := auth.GetPublicKey()
-		if err != nil {
-			return unknownKeyErr
-		}
-
-		if runtime.GOOS == "linux" && serverPubKey != localPubKey {
-			// try the ollama service public key
-			svcPubKey, err := os.ReadFile("/usr/share/ollama/.ollama/id_ed25519.pub")
-			if err != nil {
-				return unknownKeyErr
-			}
-			localPubKey = strings.TrimSpace(string(svcPubKey))
-		}
-
-		// check if the returned public key matches the local public key, this prevents adding a remote key to the user's account
-		if serverPubKey != localPubKey {
-			return unknownKeyErr
-		}
-
-		var msg strings.Builder
-		msg.WriteString(unknownKeyErr.Error())
-		msg.WriteString("\n\nYour ollama key is:\n")
-		msg.WriteString(localPubKey)
-		msg.WriteString("\nAdd your key at:\n")
-		msg.WriteString("https://ollama.com/settings/keys")
-
-		return errors.New(msg.String())
-	}
-
-	return unknownKeyErr
 }
 
 func PushHandler(cmd *cobra.Command, args []string) error {
@@ -566,20 +404,6 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 	request := api.PushRequest{Name: args[0], Insecure: insecure}
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
-		if spinner != nil {
-			spinner.Stop()
-		}
-		if strings.Contains(err.Error(), "access denied") {
-			return errors.New("you are not authorized to push to this namespace, create the model under a namespace you own")
-		}
-		host := model.ParseName(args[0]).Host
-		isOllamaHost := strings.HasSuffix(host, ".ollama.ai") || strings.HasSuffix(host, ".ollama.com")
-		if strings.Contains(err.Error(), errtypes.UnknownOllamaKeyErrMsg) && isOllamaHost {
-			// the user has not added their ollama key to ollama.com
-			// re-throw an error with a more user-friendly message
-			return errFromUnknownKey(err)
-		}
-
 		return err
 	}
 
@@ -613,61 +437,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("    ")
+	table.SetTablePadding("\t")
-	table.AppendBulk(data)
-	table.Render()
-
-	return nil
-}
-
-func ListRunningHandler(cmd *cobra.Command, args []string) error {
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	models, err := client.ListRunning(cmd.Context())
-	if err != nil {
-		return err
-	}
-
-	var data [][]string
-
-	for _, m := range models.Models {
-		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
-			var procStr string
-			switch {
-			case m.SizeVRAM == 0:
-				procStr = "100% CPU"
-			case m.SizeVRAM == m.Size:
-				procStr = "100% GPU"
-			case m.SizeVRAM > m.Size || m.Size == 0:
-				procStr = "Unknown"
-			default:
-				sizeCPU := m.Size - m.SizeVRAM
-				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
-				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
-			}
-
-			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})
-		}
-	}
-
-	table := tablewriter.NewWriter(os.Stdout)
-	table.SetHeader([]string{"NAME", "ID", "SIZE", "PROCESSOR", "UNTIL"})
-	table.SetHeaderAlignment(tablewriter.ALIGN_LEFT)
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-	table.SetHeaderLine(false)
-	table.SetBorder(false)
-	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()
 
@@ -696,6 +466,10 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
+	if len(args) != 1 {
+		return errors.New("missing model name")
+	}
+
 	license, errLicense := cmd.Flags().GetBool("license")
 	modelfile, errModelfile := cmd.Flags().GetBool("modelfile")
 	parameters, errParams := cmd.Flags().GetBool("parameters")
@@ -738,6 +512,8 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 
 	if flagsSet > 1 {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
+	} else if flagsSet == 0 {
+		return errors.New("one of '--license', '--modelfile', '--parameters', '--system', or '--template' must be specified")
 	}
 
 	req := api.ShowRequest{Name: args[0]}
@@ -746,103 +522,17 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	if flagsSet == 1 {
+	switch showType {
-		switch showType {
+	case "license":
-		case "license":
+		fmt.Println(resp.License)
-			fmt.Println(resp.License)
+	case "modelfile":
-		case "modelfile":
+		fmt.Println(resp.Modelfile)
-			fmt.Println(resp.Modelfile)
+	case "parameters":
-		case "parameters":
+		fmt.Println(resp.Parameters)
-			fmt.Println(resp.Parameters)
+	case "system":
-		case "system":
+		fmt.Println(resp.System)
-			fmt.Println(resp.System)
+	case "template":
-		case "template":
+		fmt.Println(resp.Template)
-			fmt.Println(resp.Template)
-		}
-
-		return nil
-	}
-
-	return showInfo(resp, os.Stdout)
-}
-
-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("    ")
-
-		switch header {
-		case "Template", "System", "License":
-			table.SetColWidth(100)
-		}
-
-		table.AppendBulk(rows())
-		table.Render()
-		fmt.Fprintln(w)
-	}
-
-	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 {
-		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 != "" {
-		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 != "" {
-		tableRender("System", func() [][]string {
-			return head(resp.System, 2)
-		})
-	}
-
-	if resp.License != "" {
-		tableRender("License", func() [][]string {
-			return head(resp.License, 2)
-		})
 	}
 
 	return nil
@@ -926,10 +616,10 @@ type runOptions struct {
 	WordWrap    bool
 	Format      string
 	System      string
+	Template    string
 	Images      []api.ImageData
 	Options     map[string]interface{}
 	MultiModal  bool
-	KeepAlive   *api.Duration
 }
 
 type displayResponseState struct {
@@ -942,7 +632,7 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 	if wordWrap && termWidth >= 10 {
 		for _, ch := range content {
 			if state.lineLength+1 > termWidth-5 {
-				if runewidth.StringWidth(state.wordBuffer) > termWidth-10 {
+				if len(state.wordBuffer) > termWidth-10 {
 					fmt.Printf("%s%c", state.wordBuffer, ch)
 					state.wordBuffer = ""
 					state.lineLength = 0
@@ -950,22 +640,12 @@ func displayResponse(content string, wordWrap bool, state *displayResponseState)
 				}
 
 				// backtrack the length of the last word and clear to the end of the line
-				a := runewidth.StringWidth(state.wordBuffer)
+				fmt.Printf("\x1b[%dD\x1b[K\n", len(state.wordBuffer))
-				if a > 0 {
-					fmt.Printf("\x1b[%dD", a)
-				}
-				fmt.Printf("\x1b[K\n")
 				fmt.Printf("%s%c", state.wordBuffer, ch)
-				chWidth := runewidth.RuneWidth(ch)
+				state.lineLength = len(state.wordBuffer) + 1
-
-				state.lineLength = runewidth.StringWidth(state.wordBuffer) + chWidth
 			} else {
 				fmt.Print(string(ch))
-				state.lineLength += runewidth.RuneWidth(ch)
+				state.lineLength += 1
-				if runewidth.RuneWidth(ch) >= 2 {
-					state.wordBuffer = ""
-					continue
-				}
 
 				switch ch {
 				case ' ':
@@ -1034,10 +714,6 @@ func chat(cmd *cobra.Command, opts runOptions) (*api.Message, error) {
 		Options:  opts.Options,
 	}
 
-	if opts.KeepAlive != nil {
-		req.KeepAlive = opts.KeepAlive
-	}
-
 	if err := client.Chat(cancelCtx, req, fn); err != nil {
 		if errors.Is(err, context.Canceled) {
 			return nil, nil
@@ -1113,14 +789,14 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	}
 
 	request := api.GenerateRequest{
-		Model:     opts.Model,
+		Model:    opts.Model,
-		Prompt:    opts.Prompt,
+		Prompt:   opts.Prompt,
-		Context:   generateContext,
+		Context:  generateContext,
-		Images:    opts.Images,
+		Images:   opts.Images,
-		Format:    opts.Format,
+		Format:   opts.Format,
-		System:    opts.System,
+		System:   opts.System,
-		Options:   opts.Options,
+		Template: opts.Template,
-		KeepAlive: opts.KeepAlive,
+		Options:  opts.Options,
 	}
 
 	if err := client.Generate(ctx, &request, fn); err != nil {
@@ -1154,22 +830,25 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 	return nil
 }
 
-func RunServer(_ *cobra.Command, _ []string) error {
+func RunServer(cmd *cobra.Command, _ []string) error {
+	host, port, err := net.SplitHostPort(strings.Trim(os.Getenv("OLLAMA_HOST"), "\"'"))
+	if err != nil {
+		host, port = "127.0.0.1", "11434"
+		if ip := net.ParseIP(strings.Trim(os.Getenv("OLLAMA_HOST"), "[]")); ip != nil {
+			host = ip.String()
+		}
+	}
+
 	if err := initializeKeypair(); err != nil {
 		return err
 	}
 
-	ln, err := net.Listen("tcp", envconfig.Host().Host)
+	ln, err := net.Listen("tcp", net.JoinHostPort(host, port))
 	if err != nil {
 		return err
 	}
 
-	err = server.Serve(ln)
+	return server.Serve(ln)
-	if errors.Is(err, http.ErrServerClosed) {
-		return nil
-	}
-
-	return err
 }
 
 func initializeKeypair() error {
@@ -1218,6 +897,24 @@ func initializeKeypair() error {
 	return nil
 }
 
+//nolint:unused
+func waitForServer(ctx context.Context, client *api.Client) error {
+	// wait for the server to start
+	timeout := time.After(5 * time.Second)
+	tick := time.Tick(500 * time.Millisecond)
+	for {
+		select {
+		case <-timeout:
+			return errors.New("timed out waiting for server to start")
+		case <-tick:
+			if err := client.Heartbeat(ctx); err == nil {
+				return nil // server has started
+			}
+		}
+	}
+
+}
+
 func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -1228,7 +925,7 @@ func checkServerHeartbeat(cmd *cobra.Command, _ []string) error {
 			return err
 		}
 		if err := startApp(cmd.Context(), client); err != nil {
-			return errors.New("could not connect to ollama app, is it running?")
+			return fmt.Errorf("could not connect to ollama app, is it running?")
 		}
 	}
 	return nil
@@ -1254,19 +951,12 @@ func versionHandler(cmd *cobra.Command, _ []string) {
 	}
 }
 
-func appendEnvDocs(cmd *cobra.Command, envs []envconfig.EnvVar) {
+func appendHostEnvDocs(cmd *cobra.Command) {
-	if len(envs) == 0 {
+	const hostEnvDocs = `
-		return
-	}
-
-	envUsage := `
 Environment Variables:
+      OLLAMA_HOST        The host:port or base URL of the Ollama server (e.g. http://localhost:11434)
 `
-	for _, e := range envs {
+	cmd.SetUsageTemplate(cmd.UsageTemplate() + hostEnvDocs)
-		envUsage += fmt.Sprintf("      %-24s   %s\n", e.Name, e.Description)
-	}
-
-	cmd.SetUsageTemplate(cmd.UsageTemplate() + envUsage)
 }
 
 func NewCLI() *cobra.Command {
@@ -1305,8 +995,8 @@ func NewCLI() *cobra.Command {
 		RunE:    CreateHandler,
 	}
 
-	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile")
+	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
-	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
+	createCmd.Flags().StringP("quantization", "q", "", "Quantization level.")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1330,20 +1020,10 @@ func NewCLI() *cobra.Command {
 		RunE:    RunHandler,
 	}
 
-	runCmd.Flags().String("keepalive", "", "Duration to keep a model loaded (e.g. 5m)")
 	runCmd.Flags().Bool("verbose", false, "Show timings for response")
 	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"},
@@ -1351,6 +1031,15 @@ func NewCLI() *cobra.Command {
 		Args:    cobra.ExactArgs(0),
 		RunE:    RunServer,
 	}
+	serveCmd.SetUsageTemplate(serveCmd.UsageTemplate() + `
+Environment Variables:
+
+    OLLAMA_HOST         The host:port to bind to (default "127.0.0.1:11434")
+    OLLAMA_ORIGINS      A comma separated list of allowed origins.
+    OLLAMA_MODELS       The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default is "5m")
+    OLLAMA_DEBUG        Set to 1 to enable additional debug logging
+`)
 
 	pullCmd := &cobra.Command{
 		Use:     "pull MODEL",
@@ -1379,16 +1068,8 @@ func NewCLI() *cobra.Command {
 		PreRunE: checkServerHeartbeat,
 		RunE:    ListHandler,
 	}
-
-	psCmd := &cobra.Command{
-		Use:     "ps",
-		Short:   "List running models",
-		PreRunE: checkServerHeartbeat,
-		RunE:    ListRunningHandler,
-	}
-
 	copyCmd := &cobra.Command{
-		Use:     "cp SOURCE DESTINATION",
+		Use:     "cp SOURCE TARGET",
 		Short:   "Copy a model",
 		Args:    cobra.ExactArgs(2),
 		PreRunE: checkServerHeartbeat,
@@ -1403,47 +1084,17 @@ func NewCLI() *cobra.Command {
 		RunE:    DeleteHandler,
 	}
 
-	envVars := envconfig.AsMap()
-
-	envs := []envconfig.EnvVar{envVars["OLLAMA_HOST"]}
-
 	for _, cmd := range []*cobra.Command{
 		createCmd,
 		showCmd,
 		runCmd,
-		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
-		psCmd,
 		copyCmd,
 		deleteCmd,
-		serveCmd,
 	} {
-		switch cmd {
+		appendHostEnvDocs(cmd)
-		case runCmd:
-			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
-		case serveCmd:
-			appendEnvDocs(cmd, []envconfig.EnvVar{
-				envVars["OLLAMA_DEBUG"],
-				envVars["OLLAMA_HOST"],
-				envVars["OLLAMA_KEEP_ALIVE"],
-				envVars["OLLAMA_MAX_LOADED_MODELS"],
-				envVars["OLLAMA_MAX_QUEUE"],
-				envVars["OLLAMA_MODELS"],
-				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_GPU_OVERHEAD"],
-				envVars["OLLAMA_LOAD_TIMEOUT"],
-			})
-		default:
-			appendEnvDocs(cmd, envs)
-		}
 	}
 
 	rootCmd.AddCommand(
@@ -1451,11 +1102,9 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
-		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
-		psCmd,
 		copyCmd,
 		deleteCmd,
 	)

cmd/cmd_test.go

@@ -1,206 +0,0 @@
-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)
-		}
-	})
-}

cmd/interactive.go

@@ -1,7 +1,6 @@
 package cmd
 
 import (
-	"cmp"
 	"errors"
 	"fmt"
 	"io"
@@ -9,17 +8,15 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
-	"slices"
+	"sort"
 	"strings"
 
 	"github.com/spf13/cobra"
-	"golang.org/x/exp/maps"
+	"golang.org/x/exp/slices"
 
 	"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"
 )
 
 type MultilineState int
@@ -28,16 +25,75 @@ const (
 	MultilineNone MultilineState = iota
 	MultilinePrompt
 	MultilineSystem
+	MultilineTemplate
 )
 
+func loadModel(cmd *cobra.Command, opts *runOptions) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	showReq := api.ShowRequest{Name: opts.Model}
+	showResp, err := client.Show(cmd.Context(), &showReq)
+	if err != nil {
+		return err
+	}
+	opts.MultiModal = slices.Contains(showResp.Details.Families, "clip")
+	opts.ParentModel = showResp.Details.ParentModel
+
+	if len(showResp.Messages) > 0 {
+		opts.Messages = append(opts.Messages, showResp.Messages...)
+	}
+
+	chatReq := &api.ChatRequest{
+		Model:    opts.Model,
+		Messages: []api.Message{},
+	}
+	err = client.Chat(cmd.Context(), chatReq, func(resp api.ChatResponse) error {
+		p.StopAndClear()
+		if len(opts.Messages) > 0 {
+			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
+	})
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
+	opts.Messages = make([]api.Message, 0)
+
+	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")
 		fmt.Fprintln(os.Stderr, "  /show           Show model information")
 		fmt.Fprintln(os.Stderr, "  /load <model>   Load a session or model")
 		fmt.Fprintln(os.Stderr, "  /save <model>   Save your current session")
-		fmt.Fprintln(os.Stderr, "  /clear          Clear session context")
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
@@ -55,6 +111,7 @@ 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")
@@ -74,7 +131,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "   Alt + f            Move forward (right) one word")
 		fmt.Fprintln(os.Stderr, "  Ctrl + k            Delete the sentence after the cursor")
 		fmt.Fprintln(os.Stderr, "  Ctrl + u            Delete the sentence before the cursor")
-		fmt.Fprintln(os.Stderr, "  Ctrl + w            Delete the word before the cursor")
 		fmt.Fprintln(os.Stderr, "")
 		fmt.Fprintln(os.Stderr, "  Ctrl + l            Clear the screen")
 		fmt.Fprintln(os.Stderr, "  Ctrl + c            Stop the model from responding")
@@ -100,13 +156,12 @@ 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")
 		fmt.Fprintln(os.Stderr, "  /set parameter repeat_last_n <int>    Set how far back to look for repetitions")
 		fmt.Fprintln(os.Stderr, "  /set parameter num_gpu <int>          The number of layers to send to the GPU")
-		fmt.Fprintln(os.Stderr, "  /set parameter stop <string> <string> ...   Set the stop parameters")
+		fmt.Fprintln(os.Stderr, "  /set parameter stop \"<string>\", ...   Set the stop parameters")
 		fmt.Fprintln(os.Stderr, "")
 	}
 
@@ -120,10 +175,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		return err
 	}
 
-	if envconfig.NoHistory() {
-		scanner.HistoryDisable()
-	}
-
 	fmt.Print(readline.StartBracketedPaste)
 	defer fmt.Printf(readline.EndBracketedPaste)
 
@@ -165,6 +216,10 @@ 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
@@ -196,7 +251,7 @@ 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 := loadOrUnloadModel(cmd, &opts); err != nil {
+			if err := loadModel(cmd, &opts); err != nil {
 				return err
 			}
 			continue
@@ -220,22 +275,11 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			fn := func(resp api.ProgressResponse) error { return nil }
 			err = client.Create(cmd.Context(), req, fn)
 			if err != nil {
-				if strings.Contains(err.Error(), errtypes.InvalidModelNameErrMsg) {
+				fmt.Println("error: couldn't save model")
-					fmt.Printf("error: The model name '%s' is invalid\n", args[1])
-					continue
-				}
 				return err
 			}
 			fmt.Printf("Created new model '%s'\n", args[1])
 			continue
-		case strings.HasPrefix(line, "/clear"):
-			opts.Messages = []api.Message{}
-			if opts.System != "" {
-				newMessage := api.Message{Role: "system", Content: opts.System}
-				opts.Messages = append(opts.Messages, newMessage)
-			}
-			fmt.Println("Cleared session context")
-			continue
 		case strings.HasPrefix(line, "/set"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
@@ -283,13 +327,17 @@ 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":
+				case "system", "template":
 					if len(args) < 3 {
 						usageSet()
 						continue
 					}
 
-					multiline = MultilineSystem
+					if args[1] == "system" {
+						multiline = MultilineSystem
+					} else if args[1] == "template" {
+						multiline = MultilineTemplate
+					}
 
 					line := strings.Join(args[2:], " ")
 					line, ok := strings.CutPrefix(line, `"""`)
@@ -309,17 +357,23 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						continue
 					}
 
-					opts.System = sb.String() // for display in modelfile
+					if args[1] == "system" {
-					newMessage := api.Message{Role: "system", Content: sb.String()}
+						opts.System = sb.String() // for display in modelfile
-					// Check if the slice is not empty and the last message is from 'system'
+						newMessage := api.Message{Role: "system", Content: sb.String()}
-					if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
+						// Check if the slice is not empty and the last message is from 'system'
-						// Replace the last message
+						if len(opts.Messages) > 0 && opts.Messages[len(opts.Messages)-1].Role == "system" {
-						opts.Messages[len(opts.Messages)-1] = newMessage
+							// Replace the last message
-					} else {
+							opts.Messages[len(opts.Messages)-1] = newMessage
-						opts.Messages = append(opts.Messages, 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()
 					}
-					fmt.Println("Set system message.")
-					sb.Reset()
 
 					sb.Reset()
 					continue
@@ -338,9 +392,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					return err
 				}
 				req := &api.ShowRequest{
-					Name:    opts.Model,
+					Name:     opts.Model,
-					System:  opts.System,
+					System:   opts.System,
-					Options: opts.Options,
+					Template: opts.Template,
+					Options:  opts.Options,
 				}
 				resp, err := client.Show(cmd.Context(), req)
 				if err != nil {
@@ -350,7 +405,15 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					_ = showInfo(resp, os.Stderr)
+					fmt.Println("Model details:")
+					if len(resp.Details.Families) > 0 {
+						fmt.Printf("Family              %s\n", strings.Join(resp.Details.Families, ", "))
+					} else if resp.Details.Family != "" {
+						fmt.Printf("Family              %s\n", resp.Details.Family)
+					}
+					fmt.Printf("Parameter Size      %s\n", resp.Details.ParameterSize)
+					fmt.Printf("Quantization Level  %s\n", resp.Details.QuantizationLevel)
+					fmt.Println("")
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -383,9 +446,12 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						fmt.Println("No system message was specified for this model.")
 					}
 				case "template":
-					if resp.Template != "" {
+					switch {
+					case opts.Template != "":
+						fmt.Println(opts.Template + "\n")
+					case resp.Template != "":
 						fmt.Println(resp.Template)
-					} else {
+					default:
 						fmt.Println("No prompt template was specified for this model.")
 					}
 				default:
@@ -469,35 +535,35 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func buildModelfile(opts runOptions) string {
-	var f parser.File
+	var mf strings.Builder
-	f.Commands = append(f.Commands, parser.Command{Name: "model", Args: cmp.Or(opts.ParentModel, opts.Model)})
+	model := opts.ParentModel
-
+	if model == "" {
+		model = opts.Model
+	}
+	fmt.Fprintf(&mf, "FROM %s\n", model)
 	if opts.System != "" {
-		f.Commands = append(f.Commands, parser.Command{Name: "system", Args: opts.System})
+		fmt.Fprintf(&mf, "SYSTEM \"\"\"%s\"\"\"\n", opts.System)
 	}
 
-	keys := maps.Keys(opts.Options)
+	if opts.Template != "" {
-	slices.Sort(keys)
+		fmt.Fprintf(&mf, "TEMPLATE \"\"\"%s\"\"\"\n", opts.Template)
+	}
+
+	keys := make([]string, 0)
+	for k := range opts.Options {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
 	for _, k := range keys {
-		v := opts.Options[k]
+		fmt.Fprintf(&mf, "PARAMETER %s %v\n", k, 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 {
-		f.Commands = append(f.Commands, parser.Command{Name: "message", Args: fmt.Sprintf("%s: %s", msg.Role, msg.Content)})
+		fmt.Fprintf(&mf, "MESSAGE %s \"\"\"%s\"\"\"\n", msg.Role, msg.Content)
 	}
 
-	return f.String()
+	return mf.String()
 }
 
 func normalizeFilePath(fp string) string {
@@ -583,7 +649,7 @@ 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, errors.New("file size exceeds maximum limit (100MB)")
+		return nil, fmt.Errorf("file size exceeds maximum limit (100MB)")
 	}
 
 	buf = make([]byte, info.Size())

cmd/interactive_test.go

@@ -1,9 +1,10 @@
 package cmd
 
 import (
+	"bytes"
 	"testing"
+	"text/template"
 
-	"github.com/google/go-cmp/cmp"
 	"github.com/stretchr/testify/assert"
 
 	"github.com/ollama/ollama/api"
@@ -55,53 +56,61 @@ d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8
 
 func TestModelfileBuilder(t *testing.T) {
 	opts := runOptions{
-		Model:  "hork",
+		Model:    "hork",
-		System: "You are part horse and part shark, but all hork. Do horklike things",
+		System:   "You are part horse and part shark, but all hork. Do horklike things",
+		Template: "This is a template.",
 		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]any{
+		Options: map[string]interface{}{},
-			"temperature":      0.9,
-			"seed":             42,
-			"penalize_newline": false,
-			"stop":             []string{"hi", "there"},
-		},
 	}
 
-	t.Run("model", func(t *testing.T) {
+	opts.Options["temperature"] = 0.9
-		expect := `FROM hork
+	opts.Options["seed"] = 42
-SYSTEM You are part horse and part shark, but all hork. Do horklike things
+	opts.Options["penalize_newline"] = false
+	opts.Options["stop"] = []string{"hi", "there"}
+
+	mf := buildModelfile(opts)
+	expectedModelfile := `FROM {{.Model}}
+SYSTEM """{{.System}}"""
+TEMPLATE """{{.Template}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop hi
+PARAMETER stop [hi there]
-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."""
 `
 
-		actual := buildModelfile(opts)
+	tmpl, err := template.New("").Parse(expectedModelfile)
-		if diff := cmp.Diff(expect, actual); diff != "" {
+	assert.Nil(t, err)
-			t.Errorf("mismatch (-want +got):\n%s", diff)
-		}
-	})
 
-	t.Run("parent model", func(t *testing.T) {
+	var buf bytes.Buffer
-		opts.ParentModel = "horseshark"
+	err = tmpl.Execute(&buf, opts)
-		expect := `FROM horseshark
+	assert.Nil(t, err)
-SYSTEM You are part horse and part shark, but all hork. Do horklike things
+	assert.Equal(t, buf.String(), mf)
+
+	opts.ParentModel = "horseshark"
+	mf = buildModelfile(opts)
+	expectedModelfile = `FROM {{.ParentModel}}
+SYSTEM """{{.System}}"""
+TEMPLATE """{{.Template}}"""
 PARAMETER penalize_newline false
 PARAMETER seed 42
-PARAMETER stop hi
+PARAMETER stop [hi there]
-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."""
 `
-		actual := buildModelfile(opts)
+
-		if diff := cmp.Diff(expect, actual); diff != "" {
+	tmpl, err = template.New("").Parse(expectedModelfile)
-			t.Errorf("mismatch (-want +got):\n%s", diff)
+	assert.Nil(t, err)
-		}
+
-	})
+	var parentBuf bytes.Buffer
+	err = tmpl.Execute(&parentBuf, opts)
+	assert.Nil(t, err)
+	assert.Equal(t, parentBuf.String(), mf)
 }

cmd/start.go

@@ -1,27 +0,0 @@
-//go:build darwin || windows
-
-package cmd
-
-import (
-	"context"
-	"errors"
-	"time"
-
-	"github.com/ollama/ollama/api"
-)
-
-func waitForServer(ctx context.Context, client *api.Client) error {
-	// wait for the server to start
-	timeout := time.After(5 * time.Second)
-	tick := time.Tick(500 * time.Millisecond)
-	for {
-		select {
-		case <-timeout:
-			return errors.New("timed out waiting for server to start")
-		case <-tick:
-			if err := client.Heartbeat(ctx); err == nil {
-				return nil // server has started
-			}
-		}
-	}
-}

cmd/start_darwin.go

@@ -2,7 +2,7 @@ package cmd
 
 import (
 	"context"
-	"errors"
+	"fmt"
 	"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 errors.New("could not find ollama app")
+		return fmt.Errorf("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 {

cmd/start_default.go

@@ -4,11 +4,11 @@ package cmd
 
 import (
 	"context"
-	"errors"
+	"fmt"
 
 	"github.com/ollama/ollama/api"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {
-	return errors.New("could not connect to ollama server, run 'ollama serve' to start it")
+	return fmt.Errorf("could not connect to ollama server, run 'ollama serve' to start it")
 }

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 errors.New("could not locate ollama app")
+				return fmt.Errorf("could not locate ollama app")
 			}
 		}
 	}

convert/convert.go

@@ -1,232 +1,187 @@
 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"
 )
 
-type ModelParameters struct {
+type Params struct {
-	Architectures []string `json:"architectures"`
+	Architectures     []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+	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"`
+
+	ByteOrder
 }
 
-type AdapterParameters struct {
+type ByteOrder interface {
-	Alpha          uint32 `json:"lora_alpha"`
+	binary.ByteOrder
-	LoraLayers     uint32 `json:"lora_layers"`
+	binary.AppendByteOrder
-	LoraParameters struct {
-		Rank  uint32  `json:"rank"`
-		Alpha float32 `json:"alpha"`
-		Scale float32 `json:"scale"`
-	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) llm.KV {
+type ModelArch interface {
-	kv := llm.KV{
+	GetTensors() error
-		"general.file_type":            uint32(1),
+	LoadVocab() error
-		"general.quantization_version": uint32(2),
+	WriteGGUF() (string, error)
-		"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
 }
 
-func (p AdapterParameters) KV() llm.KV {
+type ModelFormat interface {
-	var alpha float32
+	GetLayerName(string) (string, error)
-	if p.LoraParameters.Alpha == 0 {
+	GetTensors(string, *Params) ([]llm.Tensor, error)
-		alpha = float32(p.Alpha)
+	GetParams(string) (*Params, error)
-	} else {
+	GetModelArch(string, string, *Params) (ModelArch, error)
-		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
 }
 
-func (ModelParameters) specialTokenTypes() []string {
+type ModelData struct {
-	return []string{
+	Path    string
-		"bos", "eos", "unk", "sep", "pad", "cls", "mask",
+	Name    string
-	}
+	Params  *Params
+	Vocab   *Vocab
+	Tensors []llm.Tensor
+	Format  ModelFormat
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+func GetModelFormat(dirname string) (ModelFormat, error) {
-	return llm.WriteGGUF(ws, kv, ts)
+	files, err := filepath.Glob(filepath.Join(dirname, "*"))
-}
-
-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 err
+		return nil, err
 	}
 
-	var p AdapterParameters
+	for _, fn := range files {
-	if err := json.Unmarshal(bts, &p); err != nil {
+		slog.Debug(fmt.Sprintf("file = %s", fn))
-		return err
+		if strings.HasSuffix(fn, ".safetensors") {
-	}
+			return &SafetensorFormat{}, nil
-
+		} else if strings.HasSuffix(fn, ".bin") {
-	arch, ok := baseKV["general.architecture"]
+			slog.Debug("model is torch")
-	if !ok {
+			return &TorchFormat{}, nil
-		return errors.New("architecture not set for the base model")
-	}
-
-	var conv AdapterConverter
-	switch arch {
-	case "llama":
-		conv = &llamaAdapter{}
-	case "gemma2":
-		conv = &gemma2Adapter{}
-	default:
-		return errors.New("unsupported architecture")
-	}
-
-	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
-	if err != nil {
-		return err
-	}
-
-	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
-	}
-
-	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())
+	return nil, fmt.Errorf("couldn't determine model format")
-	if err != nil {
+}
-		return err
+
-	}
+// Details on gguf's tokenizer can be found at:
-
+// https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#tokenizer
-	vocabSize := int(p.VocabSize)
+type Vocab struct {
-	switch {
+	Tokens []string
-	case vocabSize > len(t.Vocabulary.Tokens):
+	Scores []float32
-		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
+	Types  []int32
-		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)
+func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
-			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
+	slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
-		}
+	in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
-	case vocabSize < len(t.Vocabulary.Tokens):
+	if err != nil {
-		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
+		return nil, err
-	default:
+	}
-		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
+
-	}
+	// To regenerate sentencepiece from the protobufs use:
-
+	// protoc -I=./ --go_out=./ sentencepiece_model.proto
-	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
+	modelProto := &sentencepiece.ModelProto{}
-	if err != nil {
+	if err := proto.Unmarshal(in, modelProto); err != nil {
-		return err
+		return nil, err
-	}
+	}
-
+
-	return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts))
+	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, int32(llm.GGUFTokenUserDefined))
+	}
+	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 := 0; cnt < missingTokens; cnt++ {
+			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
+			v.Scores = append(v.Scores, -1)
+			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
+		}
+	}
+
+	return v, nil
 }

convert/convert_bert.go

@@ -1,174 +0,0 @@
-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",
-	}
-}

convert/convert_gemma.go

@@ -1,100 +0,0 @@
-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
-}

convert/convert_gemma2.go

@@ -1,53 +0,0 @@
-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",
-	}
-}

convert/convert_gemma2_adapter.go

@@ -1,91 +0,0 @@
-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
-}

convert/convert_llama.go

@@ -1,213 +0,0 @@
-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
-}

convert/convert_llama_adapter.go

@@ -1,169 +0,0 @@
-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
-}

convert/convert_mixtral.go

@@ -1,94 +0,0 @@
-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
-}

convert/convert_phi3.go

@@ -1,123 +0,0 @@
-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
-}

convert/convert_test.go

@@ -1,476 +0,0 @@
-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"
-)
-
-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()
-
-	f, err := os.CreateTemp(t.TempDir(), "f16")
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer f.Close()
-
-	if err := ConvertModel(fsys, f); err != nil {
-		t.Fatal(err)
-	}
-
-	r, err := os.Open(f.Name())
-	if err != nil {
-		t.Fatal(err)
-	}
-	t.Cleanup(func() { 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)
-	}
-
-	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)
-			}
-
-			f, kv, tensors := convertFull(t, os.DirFS(p))
-			actual := generateResultsJSON(t, f, kv, tensors)
-
-			expectFile, err := os.Open(filepath.Join("testdata", fmt.Sprintf("%s.json", tt)))
-			if err != nil {
-				t.Fatal(err)
-			}
-
-			var expect map[string]string
-			if err := json.NewDecoder(expectFile).Decode(&expect); err != nil {
-				t.Fatal(err)
-			}
-
-			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)
-	}
-}

convert/fs.go

@@ -1,58 +0,0 @@
-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)
-}

convert/gemma.go

@@ -0,0 +1,137 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type GemmaModel struct {
+	ModelData
+}
+
+func gemmaLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	slog.Debug(fmt.Sprintf("converting '%s'", r.t.Name))
+
+	data := make([]byte, r.end-r.start)
+	if err := binary.Read(f, r.bo, data); err != nil {
+		return err
+	}
+
+	tDataF32 := bfloat16.DecodeFloat32(data)
+
+	var err error
+	tDataF32, err = addOnes(tDataF32, int(r.t.Shape[0]))
+	if err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, r.bo, tDataF32); err != nil {
+		return err
+	}
+	return nil
+}
+
+func addOnes(data []float32, vectorSize int) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, vectorSize)
+
+	var err error
+	n, err = n.Add(ones)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	newN, err := native.SelectF32(n, 0)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	var fullTensor []float32
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+
+	return fullTensor, 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)))
+
+	m.Tensors = []llm.Tensor{}
+	for _, l := range t {
+		if strings.HasSuffix(l.Name, "norm.weight") {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = gemmaLayerHandler
+			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) WriteGGUF() (string, 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,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}

convert/llama.go

@@ -0,0 +1,176 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type LlamaModel struct {
+	ModelData
+}
+
+func llamaLayerHandler(w io.Writer, r torchWriterTo) error {
+	slog.Debug(fmt.Sprintf("repacking layer '%s'", r.t.Name))
+
+	data := r.storage.(*pytorch.HalfStorage).Data
+	tData := make([]uint16, len(data))
+	for cnt, v := range data {
+		tData[cnt] = uint16(float16.Fromfloat32(v))
+	}
+
+	var err error
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	slog.Debug(fmt.Sprintf("heads = %d", heads))
+
+	tData, err = llamaRepack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	if err = binary.Write(w, r.bo, tData); err != nil {
+		return err
+	}
+	return nil
+}
+
+func llamaRepack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *LlamaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	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 {
+			slog.Debug(fmt.Sprintf("setting handler for: %s", l.Name))
+			wt := l.WriterTo.(torchWriterTo)
+			wt.handler = llamaLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *LlamaModel) LoadVocab() error {
+	var v *Vocab
+	var err error
+
+	slog.Debug("loading vocab")
+	v, err = LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug("vocab loaded")
+
+	m.Vocab = v
+	return nil
+}
+
+func (m *LlamaModel) WriteGGUF() (string, 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.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.unknown_token_id": uint32(0),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	slog.Debug(fmt.Sprintf("gguf file = %s", f.Name()))
+
+	return f.Name(), nil
+}

convert/mistral.go

@@ -0,0 +1,173 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MistralModel struct {
+	ModelData
+}
+
+func mistralLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	layerSize := r.end - r.start
+
+	var err error
+	tData := make([]uint16, layerSize/2)
+	if err = binary.Read(f, r.bo, tData); err != nil {
+		return err
+	}
+
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	tData, err = repack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	var buf []byte
+	for _, n := range tData {
+		buf = r.bo.AppendUint16(buf, n)
+	}
+
+	tempBuf := make([]uint16, len(tData))
+	tDataF32 := bfloat16.DecodeFloat32(buf)
+	for cnt, v := range tDataF32 {
+		tDataF16 := float16.Fromfloat32(v)
+		tempBuf[cnt] = uint16(tDataF16)
+	}
+
+	if err = binary.Write(w, r.bo, tempBuf); err != nil {
+		return err
+	}
+	return nil
+}
+
+func repack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *MistralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	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.handler = mistralLayerHandler
+			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() (string, 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),
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}

convert/mixtral.go

@@ -0,0 +1,96 @@
+package convert
+
+import (
+	"os"
+	"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
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	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.handler = mistralLayerHandler
+			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() (string, 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,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}

convert/reader.go

@@ -1,86 +0,0 @@
-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")
-}

convert/reader_safetensors.go

@@ -1,163 +0,0 @@
-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())
-	}
-}

convert/reader_torch.go

@@ -1,48 +0,0 @@
-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
-}

convert/safetensors.go

@@ -0,0 +1,317 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"slices"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/mitchellh/mapstructure"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type safetensorWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	filename string
+
+	start, end, padding uint64
+	handler             func(w io.Writer, r safetensorWriterTo, f *os.File) error
+}
+
+type tensorMetaData struct {
+	Type    string `mapstructure:"dtype"`
+	Shape   []int  `mapstructure:"shape"`
+	Offsets []int  `mapstructure:"data_offsets"`
+}
+
+type SafetensorFormat struct{}
+
+func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting tensor data")
+	var tensors []llm.Tensor
+	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
+	if err != nil {
+		return nil, err
+	}
+
+	var offset uint64
+	for _, f := range files {
+		var t []llm.Tensor
+		var err error
+		t, offset, err = m.readTensors(f, offset, params)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, err
+		}
+		tensors = append(tensors, t...)
+	}
+	slog.Debug(fmt.Sprintf("all tensors = %d", len(tensors)))
+	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 jsonSize uint64
+	if err := binary.Read(f, binary.LittleEndian, &jsonSize); err != nil {
+		return nil, 0, err
+	}
+
+	buf := make([]byte, jsonSize)
+	_, err = io.ReadFull(f, buf)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	d := json.NewDecoder(bytes.NewBuffer(buf))
+	d.UseNumber()
+	var parsed map[string]interface{}
+	if err = d.Decode(&parsed); err != nil {
+		return nil, 0, err
+	}
+
+	var keys []string
+	for k := range parsed {
+		keys = append(keys, k)
+	}
+
+	slices.Sort(keys)
+	slog.Info("converting layers")
+
+	var tensors []llm.Tensor
+	for _, k := range keys {
+		vals := parsed[k].(map[string]interface{})
+		var data tensorMetaData
+		if err = mapstructure.Decode(vals, &data); err != nil {
+			slog.Error("couldn't decode properly")
+			return nil, 0, err
+		}
+
+		var size uint64
+		var kind uint32
+		switch len(data.Shape) {
+		case 0:
+			// metadata
+			continue
+		case 1:
+			// convert to float32
+			kind = 0
+			size = uint64(data.Shape[0] * 4)
+		case 2:
+			// convert to float16
+			kind = 1
+			size = uint64(data.Shape[0] * data.Shape[1] * 2)
+		}
+
+		ggufName, err := m.GetLayerName(k)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, 0, err
+		}
+
+		shape := []uint64{0, 0, 0, 0}
+		for i := range data.Shape {
+			shape[i] = uint64(data.Shape[i])
+		}
+
+		t := llm.Tensor{
+			Name:   ggufName,
+			Kind:   kind,
+			Offset: offset,
+			Shape:  shape[:],
+		}
+
+		t.WriterTo = safetensorWriterTo{
+			t:        &t,
+			params:   params,
+			bo:       params.ByteOrder,
+			filename: fn,
+			start:    uint64(data.Offsets[0]),
+			end:      uint64(data.Offsets[1]),
+			padding:  8 + jsonSize,
+		}
+
+		offset += size
+		tensors = append(tensors, t)
+	}
+
+	slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
+	slog.Debug(fmt.Sprintf("offset = %d", offset))
+
+	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
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if 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(int64(r.padding+r.start), 0); err != nil {
+		return 0, err
+	}
+
+	// use the handler if one is present
+	if r.handler != nil {
+		return 0, r.handler(w, r, f)
+	}
+
+	remaining := r.end - r.start
+
+	bufSize := uint64(10240)
+	var finished bool
+	for {
+		data := make([]byte, min(bufSize, remaining))
+
+		b, err := io.ReadFull(f, data)
+		remaining -= uint64(b)
+
+		if err == io.EOF || remaining <= 0 {
+			finished = true
+		} else if err != nil {
+			return 0, err
+		}
+
+		// convert bfloat16 -> ieee float32
+		tDataF32 := bfloat16.DecodeFloat32(data)
+
+		switch r.t.Kind {
+		case 0:
+			if err := binary.Write(w, r.bo, tDataF32); err != nil {
+				return 0, err
+			}
+		case 1:
+			// convert float32 -> float16
+			tempBuf := make([]uint16, len(data)/2)
+			for cnt, v := range tDataF32 {
+				tDataF16 := float16.Fromfloat32(v)
+				tempBuf[cnt] = uint16(tDataF16)
+			}
+			if err := binary.Write(w, r.bo, tempBuf); err != nil {
+				return 0, err
+			}
+		}
+		if finished {
+			break
+		}
+	}
+	return 0, nil
+}
+
+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 "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")
+}

convert/testdata/Meta-Llama-3-8B-Instruct.json

@@ -1,313 +0,0 @@
-{
-  "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"
-}

convert/testdata/Meta-Llama-3.1-8B-Instruct.json

@@ -1,3 +0,0 @@
-{
-  "rope_freqs.weight": "80fd5efb2f729381785b293a091a268cfeceb0079167f6ece9b07070e662b222"
-}

convert/testdata/Mistral-7B-Instruct-v0.2.json

@@ -1,313 +0,0 @@
-{
-  "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"
-}

convert/testdata/Mixtral-8x7B-Instruct-v0.1.json

@@ -1,348 +0,0 @@
-{
-  "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"
-}

convert/testdata/Phi-3-mini-128k-instruct.json

@@ -1,225 +0,0 @@
-{
-  "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"
-}

convert/testdata/all-MiniLM-L6-v2.json

@@ -1,124 +0,0 @@
-{
-  "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"
-}

convert/testdata/gemma-2-2b-it.json

@@ -1,312 +0,0 @@
-{
-  "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"
-}

convert/testdata/gemma-2-9b-it.json

@@ -1,6 +0,0 @@
-{
-  "general.architecture": "gemma2",
-  "gemma2.attention.sliding_window": "4096",
-  "gemma2.attn_logit_softcapping": "50",
-  "gemma2.final_logit_softcapping": "30"
-}

convert/testdata/gemma-2b-it.json

@@ -1,188 +0,0 @@
-{
-  "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"
-}

convert/tokenizer.go

@@ -1,278 +0,0 @@
-package convert
-
-import (
-	"crypto/sha256"
-	"encoding/hex"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"io/fs"
-	"log/slog"
-	"os"
-	"slices"
-
-	"golang.org/x/exp/maps"
-)
-
-const (
-	_ int32 = iota
-	tokenTypeNormal
-	tokenTypeUnknown
-	tokenTypeControl
-	tokenTypeUserDefined
-	tokenTypeUnused
-	tokenTypeByte
-)
-
-type Tokenizer struct {
-	*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 {
-			Type    string `json:"type"`
-			Pattern struct {
-				Regex string `json:"Regex"`
-			} `json:"pattern"`
-		} `json:"pretokenizers"`
-	} `json:"pre_tokenizer"`
-}
-
-type token struct {
-	ID          int    `json:"id"`
-	Content     string `json:"content"`
-	Special     bool   `json:"special"`
-	UserDefined bool
-}
-
-type Vocabulary struct {
-	Model  string
-	Tokens []string
-	Scores []float32
-	Types  []int32
-}
-
-func parseVocabularyFromTokenizer(fsys fs.FS) (*Vocabulary, error) {
-	f, err := fsys.Open("tokenizer.json")
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
-
-	var t tokenizer
-	if err := json.NewDecoder(f).Decode(&t); err != nil {
-		return nil, err
-	}
-
-	tokens := make(map[int]token, len(t.Model.Vocab))
-	for k, v := range t.Model.Vocab {
-		tokens[v] = token{
-			ID:      v,
-			Content: k,
-		}
-	}
-
-	for _, token := range t.AddedTokens {
-		token.UserDefined = true
-		tokens[token.ID] = token
-	}
-
-	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")
-}

convert/tokenizer_spm.go

@@ -1,113 +0,0 @@
-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
-}