update patch

.dockerignore

@@ -3,7 +3,9 @@ ollama
 app
 macapp
 dist
+llm/llama.cpp
 .env
 .cache
 test_data
+llm/build
 llama/build

.gitattributes

@@ -1,11 +1,3 @@
-llama/**/*.cpp linguist-vendored
-llama/**/*.hpp linguist-vendored
-llama/**/*.h linguist-vendored
-llama/**/*.c linguist-vendored
-llama/**/*.cu linguist-vendored
-llama/**/*.cuh linguist-vendored
-llama/**/*.m linguist-vendored
-llama/**/*.metal linguist-vendored
-
+llm/ext_server/* linguist-vendored
 * text=auto
 *.go text eol=lf

.github/workflows/release.yaml

@@ -48,8 +48,8 @@ jobs:
         with:
           name: dist-darwin
           path: |
-            dist/Ollama-darwin.zip
-            dist/ollama-darwin
+            dist/*arwin*
+            !dist/*-cov
 
   # Windows builds take a long time to both install the dependencies and build, so parallelize
   # CPU generation step
@@ -92,19 +92,19 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
-        name: make
+          go generate -x ./...
+        name: go generate
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cpu
           path: |
             build/**/*
             build/**/*.a
+            llm/build/**/*.a
             dist/windows-amd64/**
 
   # ROCm generation step
@@ -158,21 +158,31 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
-        name: make
+          go generate -x ./...
+        name: go generate
+      - name: 'gather rocm dependencies'
+        run: |
+          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          md "dist\deps\bin\rocblas\library"
+          cp "${HIP_PATH}\bin\hipblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas\library\*" "dist\deps\bin\rocblas\library\"
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
           path: |
             build/**/*
             dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-rocm-deps
+          path: dist/deps/*
 
   # CUDA generation step
   generate-windows-cuda:
@@ -235,23 +245,34 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: make
+      - name: go generate
         run: |
           $gopath=(get-command go).source | split-path -parent
           $cudabin=(get-command nvcc).source | split-path
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$cudabin;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
+          go generate -x ./...
+      - name: 'gather cuda dependencies'
+        run: |
+          $NVIDIA_DIR=(resolve-path 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*\bin\')[0]
+          md "dist\deps"
+          cp "${NVIDIA_DIR}\cudart64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublas64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
             build/**/*
             dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-cuda-deps-${{ matrix.cuda.version }}
+          path: dist/deps/*
+
 
   # windows arm64 generate, go build, and zip file (no installer)
   # Output of this build is aggregated into the final x86 build
@@ -271,30 +292,6 @@ jobs:
           choco install -y --no-progress git gzip
           echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-      # pacman is buggy on win arm64, so we avoid using it, but rely on the binary artifacts
-      # we download the sfx (7zip bundle) which isn't fully set up, but the binaries we need to build work
-      - name: Install msys2 x64
-        run: |
-          $url="https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-base-x86_64-20240727.sfx.exe"
-          write-host "Downloading MSYS2"
-          Invoke-WebRequest -Uri "$url" -outfile "${env:RUNNER_TEMP}\msys2.exe"
-          write-host "Installing msys2"
-          Start-Process "${env:RUNNER_TEMP}\msys2.exe" -ArgumentList @(
-              '-y', '-oC:\'
-              ) -NoNewWindow -Wait
-          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-      # since pacman isn't reliable, we just download the tar file and extract directly
-      - name: Downloading and extracting msys2 make tar file
-        run: |
-          $url="https://mirror.msys2.org/msys/x86_64/make-4.4.1-2-x86_64.pkg.tar.zst"
-          write-host "Downloading make"
-          Invoke-WebRequest -Uri "$url" -outfile c:\msys64\make.tar.zst
-          cd c:\msys64; tar -xf make.tar.zst
-          rm c:\msys64\make.tar.zst
-      - name: Verify Make works properly
-        run: |
-          echo $env:PATH
-          make --version
       - name: Install Visual Studio 2022
         run: |
           $components = @(
@@ -388,9 +385,10 @@ jobs:
       - run: |
           $gopath=(get-command go).source | split-path -parent
           $gccpath=(get-command gcc).source | split-path -parent
-          import-module 'C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\Program Files\Microsoft Visual Studio\2022\Community' -skipautomaticlocation
-          $env:PATH="$gopath;$gccpath;$env:PATH"
+          & "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH;C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin"
           echo $env:PATH
           $env:ARCH="arm64"
           .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
@@ -457,6 +455,15 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-11
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps-12
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-rocm-deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
@@ -467,12 +474,11 @@ jobs:
       - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_GENERATE="1"
-          $env:ARCH="amd64"
           & .\scripts\build_windows.ps1
       - uses: actions/upload-artifact@v4
         with:

.github/workflows/test.yaml

@@ -21,7 +21,9 @@ jobs:
   changes:
     runs-on: ubuntu-latest
     outputs:
-      RUNNERS: ${{ steps.changes.outputs.RUNNERS }}
+      GENERATE: ${{ steps.changes.outputs.GENERATE }}
+      GENERATE_CUDA: ${{ steps.changes.outputs.GENERATE_CUDA }}
+      GENERATE_ROCM: ${{ steps.changes.outputs.GENERATE_ROCM }}
     steps:
       - uses: actions/checkout@v4
         with:
@@ -36,12 +38,52 @@ jobs:
           }
 
           {
-            echo RUNNERS=$(changed 'llama/**')
+            echo GENERATE=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE_CUDA=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
+            echo GENERATE_ROCM=$(changed 'llm/llama.cpp' 'llm/patches/**' 'llm/ext_server/**' 'llm/generate/**')
           } >>$GITHUB_OUTPUT
 
-  runners-linux-cuda:
+  generate:
     needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    if: ${{ needs.changes.outputs.GENERATE == 'True' }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        arch: [amd64, arm64]
+        exclude:
+          - os: ubuntu-latest
+            arch: arm64
+          - os: windows-2019
+            arch: arm64
+    runs-on: ${{ matrix.os }}
+    env:
+      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
+    steps:
+      - uses: actions/checkout@v4
+      - 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 (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
+          go generate -x ./...
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        name: 'Windows Go Generate'
+      - run: go generate -x ./...
+        if: ${{ ! startsWith(matrix.os, 'windows-') }}
+        name: 'Unix Go Generate'
+      - run: go build .
+  generate-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
     strategy:
       matrix:
         cuda-version:
@@ -51,6 +93,8 @@ jobs:
     steps:
       - run: |
           apt-get update && apt-get install -y git build-essential curl
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
+            | tar -zx -C /usr --strip-components 1
         env:
           DEBIAN_FRONTEND: noninteractive
       - uses: actions/checkout@v4
@@ -61,11 +105,12 @@ jobs:
       - run: go get ./...
       - run: |
           git config --global --add safe.directory /__w/ollama/ollama
-          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
-          make -j $cores cuda_v11
-  runners-linux-rocm:
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+  generate-rocm:
     needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
     strategy:
       matrix:
         rocm-version:
@@ -75,6 +120,8 @@ jobs:
     steps:
       - run: |
           apt-get update && apt-get install -y git build-essential curl rocm-libs
+          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
+            | tar -zx -C /usr --strip-components 1
         env:
           DEBIAN_FRONTEND: noninteractive
       - uses: actions/checkout@v4
@@ -85,13 +132,14 @@ jobs:
       - run: go get ./...
       - run: |
           git config --global --add safe.directory /__w/ollama/ollama
-          cores=$(grep '^core id' /proc/cpuinfo |sort -u|wc -l)
-          make -j $cores rocm
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
 
   # ROCm generation step
-  runners-windows-rocm:
+  generate-windows-rocm:
     needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
     runs-on: windows
     steps:
       - uses: actions/checkout@v4
@@ -113,21 +161,21 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          write-host $env:HIP_PATH
-          make -C llama print-HIP_PATH print-HIP_LIB_DIR
-          make -j $cores rocm
-        name: make
+          go generate -x ./...
+        name: go generate
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
 
   # CUDA generation step
-  runners-windows-cuda:
+  generate-windows-cuda:
     needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
     runs-on: windows
     steps:
       - uses: actions/checkout@v4
@@ -152,60 +200,19 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: make
+      - name: go generate
         run: |
           $gopath=(get-command go).source | split-path -parent
           $cudabin=(get-command nvcc).source | split-path
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$cudabin;$env:PATH"
           $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores cuda_v11
+          go generate -x ./...
         env:
           OLLAMA_SKIP_CPU_GENERATE: '1'
 
-  runners-cpu:
-    needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest, windows-2019]
-        arch: [amd64, arm64]
-        exclude:
-          - os: ubuntu-latest
-            arch: arm64
-          - os: windows-2019
-            arch: arm64
-    runs-on: ${{ matrix.os }}
-    env:
-      GOARCH: ${{ matrix.arch }}
-      ARCH: ${{ matrix.arch }}
-      CGO_ENABLED: '1'
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-go@v5
-        with:
-          go-version-file: go.mod
-          cache: true
-      - run: go get ./...
-      - name: 'Build Windows Go Runners'
-        if: ${{ startsWith(matrix.os, 'windows-') }}
-        run: |
-          $gopath=(get-command go).source | split-path -parent
-          $gccpath=(get-command gcc).source | split-path -parent
-          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
-          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$gccpath;$env:PATH"
-          echo $env:PATH
-          make -j 4      
-      - name: 'Build Unix Go Runners'
-        if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        run: make -j 4
-      - run: go build .
-
   lint:
     strategy:
       matrix:
@@ -253,6 +260,9 @@ jobs:
     env:
       GOARCH: ${{ matrix.arch }}
       CGO_ENABLED: '1'
+      OLLAMA_CPU_TARGET: 'static'
+      OLLAMA_SKIP_CPU_GENERATE: '1'
+      OLLAMA_SKIP_METAL_GENERATE: '1'
     steps:
       - uses: actions/checkout@v4
         with:
@@ -267,17 +277,6 @@ jobs:
             arm64) echo ARCH=arm64 ;;
           esac >>$GITHUB_ENV
         shell: bash
+      - run: go generate ./...
       - run: go build
       - run: go test -v ./...
-
-  patches:
-    needs: [changes]
-    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-      - name: Verify patches carry all the changes
-        run: |
-          make apply-patches sync && git diff --compact-summary --exit-code llama

.gitignore

@@ -5,6 +5,7 @@
 .swp
 dist
 ollama
+ggml-metal.metal
 .cache
 *.exe
 .idea
@@ -14,5 +15,4 @@ llm/build
 build/*/*/*
 !build/**/placeholder
 llama/build
-__debug_bin*
-llama/vendor
+__debug_bin*

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "llama.cpp"]
+	path = llm/llama.cpp
+	url = https://github.com/ggerganov/llama.cpp.git
+	shallow = true

Dockerfile

@@ -1,4 +1,3 @@
-# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
 ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION_11=11.3.1
@@ -7,134 +6,176 @@ ARG CUDA_VERSION_12=12.4.0
 ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
 
-### To create a local image for building linux binaries on mac or windows with efficient incremental builds
-#
-# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
-# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
-#
-### Then incremental builds will be much faster in this container
-#
-# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
-#
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
+# Copy the minimal context we need to run the generate scripts
+FROM scratch AS llm-code
+COPY .git .git
+COPY .gitmodules .gitmodules
+COPY llm llm
+
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
+
+FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-runner-arm64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V11_ARCHITECTURES
+ENV GOARCH=arm64
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
+    CUDA_VARIANT="_v11" \
+    bash gen_linux.sh
+
+FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-runner-arm64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG CUDA_V12_ARCHITECTURES
+ENV GOARCH=arm64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 \
+    OLLAMA_SKIP_CPU_GENERATE=1 \
+    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
+    CUDA_VARIANT="_v12" \
+    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
+    bash gen_linux.sh
+
+
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
+ARG CMAKE_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV LIBRARY_PATH=/opt/amdgpu/lib64
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+ARG CGO_CFLAGS
+ARG AMDGPU_TARGETS
+ENV GOARCH=amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
+RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
+    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
+
+FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
-ARG CUDA_VERSION_11
-ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
-ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
-ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
-    dnf clean all && \
-    dnf install -y \
-    zsh \
-    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
-    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
-# TODO intel oneapi goes here...
-ENV GOARCH amd64
-ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/ollama/ollama/
-ENTRYPOINT [ "zsh" ]
+ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$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
 
-### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
-# Note: this does not contain jetson variants
-#
-# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
-# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
-#
-FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
+FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
+FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
+RUN --mount=type=cache,target=/root/.ccache \
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
+
+FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
-ARG CUDA_VERSION_11
-ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
-    dnf config-manager --set-enabled appstream && \
-    dnf clean all && \
-    dnf install -y \
-    zsh \
-    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
-    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
-ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
-ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
-ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
-ENV GOARCH amd64
-ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/ollama/ollama/
-ENTRYPOINT [ "zsh" ]
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+ARG OLLAMA_CUSTOM_CPU_DEFS
+ARG CGO_CFLAGS
+ENV GOARCH=arm64
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
-FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
-COPY . .
-ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
-ARG OLLAMA_SKIP_ROCM_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
-ARG OLLAMA_FAST_BUILD
+FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
 RUN --mount=type=cache,target=/root/.ccache \
-    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
-        make -C llama -j $(expr $(nproc) / 2 ) ; \
-    else \
-        make -C llama -j 5 ; \
-    fi
-
-FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
-COPY . .
-ARG OLLAMA_SKIP_CUDA_GENERATE
-ARG OLLAMA_SKIP_CUDA_11_GENERATE
-ARG OLLAMA_SKIP_CUDA_12_GENERATE
-ARG CUDA_V11_ARCHITECTURES
-ARG CUDA_V12_ARCHITECTURES
-ARG OLLAMA_FAST_BUILD
+    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 \
-    make -C llama -j 8
+    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
 
 
 # Intermediate stages used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 centos:7 AS 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:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH amd64
+FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
+ENV CGO_ENABLED=1
 WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/amd64 builder-amd64 AS build-amd64
 COPY . .
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-ARG OLLAMA_SKIP_ROCM_GENERATE
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 RUN cd dist/linux-$GOARCH && \
     tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
-    cd dist/linux-$GOARCH-rocm && \
-    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
-    fi
+RUN cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz
 
-FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
-ARG CMAKE_VERSION
+FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
+ENV CGO_ENABLED=1
 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:$PATH
-ENV CGO_ENABLED 1
-ENV GOARCH arm64
 WORKDIR /go/src/github.com/ollama/ollama
-
-FROM --platform=linux/arm64 builder-arm64 AS build-arm64
 COPY . .
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
+COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
@@ -146,11 +187,11 @@ FROM --platform=linux/amd64 scratch AS dist-amd64
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
 FROM --platform=linux/arm64 scratch AS dist-arm64
 COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
-FROM dist-$TARGETARCH AS dist
+FROM dist-$TARGETARCH as dist
 
 
 # Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
+FROM --platform=linux/amd64 static-build-amd64 AS container-build-amd64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -158,7 +199,7 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 
-FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
+FROM --platform=linux/arm64 static-build-arm64 AS container-build-arm64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -166,52 +207,48 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-arm64/bin/ollama .
 
-# For amd64 container images, filter out cuda/rocm to minimize size
-FROM runners-amd64 AS runners-cuda-amd64
-RUN rm -rf \
-    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
-    ./dist/linux-amd64/lib/ollama/runners/rocm*
-
-FROM runners-amd64 AS runners-rocm-amd64
-RUN rm -rf \
-    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
-    ./dist/linux-amd64/lib/ollama/libcu*.so* \
-    ./dist/linux-amd64/lib/ollama/runners/cuda*
-
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
+    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=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+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 && \
-    rm -rf /var/lib/apt/lists/*
+    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=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+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=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+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 && \
-    rm -rf /var/lib/apt/lists/*
+    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=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-
+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

Makefile

@@ -1,4 +0,0 @@
-GOALS := $(or $(MAKECMDGOALS),all)
-.PHONY: $(GOALS)
-$(GOALS):
-	$(MAKE) -C llama $@

README.md

@@ -328,9 +328,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
 - [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
-- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG on Mac/Windows/Linux)
-- [G1](https://github.com/bklieger-groq/g1) (Prototype of using prompting strategies to improve the LLM's reasoning through o1-like reasoning chains.)
-- [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 
 ### Terminal
 
@@ -377,7 +374,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Libraries
 
-- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/integrations/chat/ollama/) with [example](https://js.langchain.com/docs/tutorials/local_rag/)
+- [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)
@@ -412,13 +409,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
-- [Ollama for Swift](https://github.com/mattt/ollama-swift)
 
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
-- [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 
 ### Extensions & Plugins

cmd/cmd.go

@@ -21,6 +21,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"runtime"
+	"slices"
 	"strconv"
 	"strings"
 	"sync/atomic"
@@ -46,58 +47,28 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-var (
-	errModelNotFound     = errors.New("no Modelfile or safetensors files found")
-	errModelfileNotFound = errors.New("specified Modelfile wasn't found")
-)
-
-func getModelfileName(cmd *cobra.Command) (string, error) {
-	fn, _ := cmd.Flags().GetString("file")
-
-	filename := fn
-	if filename == "" {
-		filename = "Modelfile"
-	}
-
-	absName, err := filepath.Abs(filename)
-	if err != nil {
-		return "", err
-	}
-
-	_, err = os.Stat(absName)
-	if err != nil {
-		return fn, err
-	}
-
-	return absName, nil
-}
-
 func CreateHandler(cmd *cobra.Command, args []string) error {
+	filename, _ := cmd.Flags().GetString("file")
+	filename, err := filepath.Abs(filename)
+	if err != nil {
+		return err
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	var reader io.Reader
-
-	filename, err := getModelfileName(cmd)
-	if os.IsNotExist(err) {
-		if filename == "" {
-			reader = strings.NewReader("FROM .\n")
-		} else {
-			return errModelfileNotFound
-		}
-	} else if err != nil {
+	f, err := os.Open(filename)
+	if err != nil {
 		return err
-	} else {
-		f, err := os.Open(filename)
-		if err != nil {
-			return err
-		}
-
-		reader = f
-		defer f.Close()
 	}
+	defer f.Close()
 
-	modelfile, err := parser.ParseFile(reader)
+	modelfile, err := parser.ParseFile(f)
 	if err != nil {
 		return err
 	}
@@ -112,11 +83,6 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p.Add(status, spinner)
 	defer p.Stop()
 
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
 	for i := range modelfile.Commands {
 		switch modelfile.Commands[i].Name {
 		case "model", "adapter":
@@ -255,7 +221,7 @@ func tempZipFiles(path string) (string, error) {
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
 	} else {
-		return "", errModelNotFound
+		return "", errors.New("no safetensors or torch files found")
 	}
 
 	// add configuration files, json files are detected as text/plain
@@ -487,7 +453,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	opts.MultiModal = len(info.ProjectorInfo) != 0
+	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
 	opts.ParentModel = info.Details.ParentModel
 
 	if interactive {
@@ -714,17 +680,6 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	// Unload the model if it's running before deletion
-	opts := &runOptions{
-		Model:     args[0],
-		KeepAlive: &api.Duration{Duration: 0},
-	}
-	if err := loadOrUnloadModel(cmd, opts); err != nil {
-		if !strings.Contains(err.Error(), "not found") {
-			return fmt.Errorf("unable to stop existing running model \"%s\": %s", args[0], err)
-		}
-	}
-
 	for _, name := range args {
 		req := api.DeleteRequest{Name: name}
 		if err := client.Delete(cmd.Context(), &req); err != nil {
@@ -1350,7 +1305,7 @@ func NewCLI() *cobra.Command {
 		RunE:    CreateHandler,
 	}
 
-	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
+	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")
 
 	showCmd := &cobra.Command{

cmd/cmd_test.go

@@ -2,17 +2,11 @@ package cmd
 
 import (
 	"bytes"
-	"context"
-	"encoding/json"
-	"net/http"
-	"net/http/httptest"
 	"os"
 	"path/filepath"
-	"strings"
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
-	"github.com/spf13/cobra"
 
 	"github.com/ollama/ollama/api"
 )
@@ -210,162 +204,3 @@ Weigh anchor!
 		}
 	})
 }
-
-func TestDeleteHandler(t *testing.T) {
-	stopped := false
-	mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path == "/api/delete" && r.Method == http.MethodDelete {
-			var req api.DeleteRequest
-			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-				http.Error(w, err.Error(), http.StatusBadRequest)
-				return
-			}
-			if req.Name == "test-model" {
-				w.WriteHeader(http.StatusOK)
-			} else {
-				w.WriteHeader(http.StatusNotFound)
-			}
-			return
-		}
-		if r.URL.Path == "/api/generate" && r.Method == http.MethodPost {
-			var req api.GenerateRequest
-			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-				http.Error(w, err.Error(), http.StatusBadRequest)
-				return
-			}
-			if req.Model == "test-model" {
-				w.WriteHeader(http.StatusOK)
-				if err := json.NewEncoder(w).Encode(api.GenerateResponse{
-					Done: true,
-				}); err != nil {
-					http.Error(w, err.Error(), http.StatusInternalServerError)
-				}
-				stopped = true
-				return
-			} else {
-				w.WriteHeader(http.StatusNotFound)
-				if err := json.NewEncoder(w).Encode(api.GenerateResponse{
-					Done: false,
-				}); err != nil {
-					http.Error(w, err.Error(), http.StatusInternalServerError)
-				}
-			}
-		}
-	}))
-
-	t.Setenv("OLLAMA_HOST", mockServer.URL)
-	t.Cleanup(mockServer.Close)
-
-	cmd := &cobra.Command{}
-	cmd.SetContext(context.TODO())
-	if err := DeleteHandler(cmd, []string{"test-model"}); err != nil {
-		t.Fatalf("DeleteHandler failed: %v", err)
-	}
-	if !stopped {
-		t.Fatal("Model was not stopped before deletion")
-	}
-
-	err := DeleteHandler(cmd, []string{"test-model-not-found"})
-	if err == nil || !strings.Contains(err.Error(), "unable to stop existing running model \"test-model-not-found\"") {
-		t.Fatalf("DeleteHandler failed: expected error about stopping non-existent model, got %v", err)
-	}
-}
-
-func TestGetModelfileName(t *testing.T) {
-	tests := []struct {
-		name          string
-		modelfileName string
-		fileExists    bool
-		expectedName  string
-		expectedErr   error
-	}{
-		{
-			name:          "no modelfile specified, no modelfile exists",
-			modelfileName: "",
-			fileExists:    false,
-			expectedName:  "",
-			expectedErr:   os.ErrNotExist,
-		},
-		{
-			name:          "no modelfile specified, modelfile exists",
-			modelfileName: "",
-			fileExists:    true,
-			expectedName:  "Modelfile",
-			expectedErr:   nil,
-		},
-		{
-			name:          "modelfile specified, no modelfile exists",
-			modelfileName: "crazyfile",
-			fileExists:    false,
-			expectedName:  "crazyfile",
-			expectedErr:   os.ErrNotExist,
-		},
-		{
-			name:          "modelfile specified, modelfile exists",
-			modelfileName: "anotherfile",
-			fileExists:    true,
-			expectedName:  "anotherfile",
-			expectedErr:   nil,
-		},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			cmd := &cobra.Command{
-				Use: "fakecmd",
-			}
-			cmd.Flags().String("file", "", "path to modelfile")
-
-			var expectedFilename string
-
-			if tt.fileExists {
-				tempDir, err := os.MkdirTemp("", "modelfiledir")
-				defer os.RemoveAll(tempDir)
-				if err != nil {
-					t.Fatalf("temp modelfile dir creation failed: %v", err)
-				}
-				var fn string
-				if tt.modelfileName != "" {
-					fn = tt.modelfileName
-				} else {
-					fn = "Modelfile"
-				}
-
-				tempFile, err := os.CreateTemp(tempDir, fn)
-				if err != nil {
-					t.Fatalf("temp modelfile creation failed: %v", err)
-				}
-
-				expectedFilename = tempFile.Name()
-				err = cmd.Flags().Set("file", expectedFilename)
-				if err != nil {
-					t.Fatalf("couldn't set file flag: %v", err)
-				}
-			} else {
-				if tt.modelfileName != "" {
-					expectedFilename = tt.modelfileName
-					err := cmd.Flags().Set("file", tt.modelfileName)
-					if err != nil {
-						t.Fatalf("couldn't set file flag: %v", err)
-					}
-				}
-			}
-
-			actualFilename, actualErr := getModelfileName(cmd)
-
-			if actualFilename != expectedFilename {
-				t.Errorf("expected filename: '%s' actual filename: '%s'", expectedFilename, actualFilename)
-			}
-
-			if tt.expectedErr != os.ErrNotExist {
-				if actualErr != tt.expectedErr {
-					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
-				}
-			} else {
-				if !os.IsNotExist(actualErr) {
-					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
-				}
-			}
-		})
-	}
-}

cmd/interactive.go

@@ -442,6 +442,13 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					return err
 				}
 
+				// clear all previous images for better responses
+				if len(images) > 0 {
+					for i := range opts.Messages {
+						opts.Messages[i].Images = nil
+					}
+				}
+
 				newMessage.Content = msg
 				newMessage.Images = images
 			}
@@ -494,22 +501,28 @@ func buildModelfile(opts runOptions) string {
 }
 
 func normalizeFilePath(fp string) string {
-	return strings.NewReplacer(
-		"\\ ", " ", // Escaped space
-		"\\(", "(", // Escaped left parenthesis
-		"\\)", ")", // Escaped right parenthesis
-		"\\[", "[", // Escaped left square bracket
-		"\\]", "]", // Escaped right square bracket
-		"\\{", "{", // Escaped left curly brace
-		"\\}", "}", // Escaped right curly brace
-		"\\$", "$", // Escaped dollar sign
-		"\\&", "&", // Escaped ampersand
-		"\\;", ";", // Escaped semicolon
-		"\\'", "'", // Escaped single quote
-		"\\\\", "\\", // Escaped backslash
-		"\\*", "*", // Escaped asterisk
-		"\\?", "?", // Escaped question mark
-	).Replace(fp)
+	// Define a map of escaped characters and their replacements
+	replacements := map[string]string{
+		"\\ ":  " ",  // Escaped space
+		"\\(":  "(",  // Escaped left parenthesis
+		"\\)":  ")",  // Escaped right parenthesis
+		"\\[":  "[",  // Escaped left square bracket
+		"\\]":  "]",  // Escaped right square bracket
+		"\\{":  "{",  // Escaped left curly brace
+		"\\}":  "}",  // Escaped right curly brace
+		"\\$":  "$",  // Escaped dollar sign
+		"\\&":  "&",  // Escaped ampersand
+		"\\;":  ";",  // Escaped semicolon
+		"\\'":  "'",  // Escaped single quote
+		"\\\\": "\\", // Escaped backslash
+		"\\*":  "*",  // Escaped asterisk
+		"\\?":  "?",  // Escaped question mark
+	}
+
+	for escaped, actual := range replacements {
+		fp = strings.ReplaceAll(fp, escaped, actual)
+	}
+	return fp
 }
 
 func extractFileNames(input string) []string {
@@ -529,9 +542,10 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 	for _, fp := range filePaths {
 		nfp := normalizeFilePath(fp)
 		data, err := getImageData(nfp)
-		if errors.Is(err, os.ErrNotExist) {
-			continue
-		} else if err != nil {
+		if err != nil {
+			if os.IsNotExist(err) {
+				continue
+			}
 			fmt.Fprintf(os.Stderr, "Couldn't process image: %q\n", err)
 			return "", imgs, err
 		}
@@ -539,7 +553,7 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 		input = strings.ReplaceAll(input, fp, "")
 		imgs = append(imgs, data)
 	}
-	return strings.TrimSpace(input), imgs, nil
+	return input, imgs, nil
 }
 
 func getImageData(filePath string) ([]byte, error) {

convert/convert.go

@@ -9,7 +9,7 @@ import (
 	"log/slog"
 	"strings"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type ModelParameters struct {
@@ -27,8 +27,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) fileutils.KV {
-	kv := fileutils.KV{
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
+	kv := llm.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -54,7 +54,7 @@ func (ModelParameters) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() fileutils.KV {
+func (p AdapterParameters) KV() llm.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -62,7 +62,7 @@ func (p AdapterParameters) KV() fileutils.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := fileutils.KV{
+	kv := llm.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -79,19 +79,19 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
-	return fileutils.WriteGGUF(ws, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }
 
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
-	return fileutils.WriteGGUF(ws, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }
 
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) fileutils.KV
+	KV(*Tokenizer) llm.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []fileutils.Tensor
+	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
@@ -99,7 +99,7 @@ type ModelConverter interface {
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
 	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, fileutils.KV, []fileutils.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
 type moreParser interface {
@@ -108,17 +108,17 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(fileutils.KV) fileutils.KV
+	KV(llm.KV) llm.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []fileutils.Tensor
+	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, fileutils.KV, []fileutils.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }
 
-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV fileutils.KV) 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

convert/convert_bert.go

@@ -8,7 +8,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type bertModel struct {
@@ -85,7 +85,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) fileutils.KV {
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p *bertModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []fileutils.Tensor {
 			continue
 		}
 
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type gemmaModel struct {
@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) fileutils.KV {
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p *gemmaModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+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, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_gemma2.go

@@ -1,7 +1,7 @@
 package convert
 
 import (
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type gemma2Model struct {
@@ -11,7 +11,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) fileutils.KV {
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV fileutils.KV) fileutils.KV {
+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) []fileutils.Tensor {
-	var out []fileutils.Tensor
+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]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama.go

@@ -9,7 +9,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type llamaModel struct {
@@ -46,7 +46,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) fileutils.KV {
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
@@ -120,11 +120,11 @@ func (p *llamaModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p *llamaModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 
 	if p.RopeScaling.factors != nil {
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -138,7 +138,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama_adapter.go

@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
+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"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+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]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mixtral.go

@@ -6,7 +6,7 @@ import (
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type mixtralModel struct {
@@ -15,7 +15,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) fileutils.KV {
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {
@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p *mixtralModel) Tensors(ts []Tensor) []fileutils.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []fileutils.Tensor {
 		return true
 	})
 
-	var out []fileutils.Tensor
+	var out []llm.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),

convert/convert_phi3.go

@@ -8,7 +8,7 @@ import (
 	"strings"
 	"sync"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type phi3Model struct {
@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) fileutils.KV {
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }
 
-func (p *phi3Model) Tensors(ts []Tensor) []fileutils.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 	var addRopeFactors sync.Once
 
-	out := make([]fileutils.Tensor, 0, len(ts)+2)
+	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, fileutils.Tensor{
+				out = append(out, llm.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, fileutils.Tensor{
+				}, llm.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []fileutils.Tensor {
 			})
 		}
 
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_test.go

@@ -20,7 +20,7 @@ import (
 
 	"golang.org/x/exp/maps"
 
-	"github.com/ollama/ollama/fileutils"
+	"github.com/ollama/ollama/llm"
 )
 
 type tensorData struct {
@@ -29,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, *fileutils.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -48,7 +48,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, *fileutils.T
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
+	m, _, err := llm.DecodeGGML(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -60,7 +60,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, *fileutils.T
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv fileutils.KV, tensors *fileutils.Tensors) map[string]string {
+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 {
@@ -330,7 +330,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
 			if err != nil {
 				t.Fatal(err)
 			}

discover/README.md

@@ -1,3 +0,0 @@
-# `discover`
-
-This package is responsible for discovering information about the system and the capabilities to run LLM.  This includes GPU and CPU discovery so the optimal runner can be chosen for a given model.  The ollama scheduler relies on up-to-date available memory information, so this package provides the ability to refresh free memory as efficiently as possible.

discover/gpu_linux.go

@@ -1,186 +0,0 @@
-package discover
-
-import (
-	"bufio"
-	"fmt"
-	"os"
-	"reflect"
-	"regexp"
-	"strings"
-
-	"github.com/ollama/ollama/format"
-)
-
-var CudartGlobs = []string{
-	"/usr/local/cuda/lib64/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
-	"/usr/lib/wsl/lib/libcudart.so*",
-	"/usr/lib/wsl/drivers/*/libcudart.so*",
-	"/opt/cuda/lib64/libcudart.so*",
-	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
-	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
-	"/usr/local/cuda/lib*/libcudart.so*",
-	"/usr/lib*/libcudart.so*",
-	"/usr/local/lib*/libcudart.so*",
-}
-
-var NvmlGlobs = []string{}
-
-var NvcudaGlobs = []string{
-	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
-	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
-	"/usr/lib/*-linux-gnu/libcuda.so*",
-	"/usr/lib/wsl/lib/libcuda.so*",
-	"/usr/lib/wsl/drivers/*/libcuda.so*",
-	"/opt/cuda/lib*/libcuda.so*",
-	"/usr/local/cuda/lib*/libcuda.so*",
-	"/usr/lib*/libcuda.so*",
-	"/usr/local/lib*/libcuda.so*",
-}
-
-var OneapiGlobs = []string{
-	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
-	"/usr/lib*/libze_intel_gpu.so*",
-}
-
-var (
-	CudartMgmtName = "libcudart.so*"
-	NvcudaMgmtName = "libcuda.so*"
-	NvmlMgmtName   = "" // not currently wired on linux
-	OneapiMgmtName = "libze_intel_gpu.so*"
-)
-
-func GetCPUMem() (memInfo, error) {
-	var mem memInfo
-	var total, available, free, buffers, cached, freeSwap uint64
-	f, err := os.Open("/proc/meminfo")
-	if err != nil {
-		return mem, err
-	}
-	defer f.Close()
-	s := bufio.NewScanner(f)
-	for s.Scan() {
-		line := s.Text()
-		switch {
-		case strings.HasPrefix(line, "MemTotal:"):
-			_, err = fmt.Sscanf(line, "MemTotal:%d", &total)
-		case strings.HasPrefix(line, "MemAvailable:"):
-			_, err = fmt.Sscanf(line, "MemAvailable:%d", &available)
-		case strings.HasPrefix(line, "MemFree:"):
-			_, err = fmt.Sscanf(line, "MemFree:%d", &free)
-		case strings.HasPrefix(line, "Buffers:"):
-			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
-		case strings.HasPrefix(line, "Cached:"):
-			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
-		case strings.HasPrefix(line, "SwapFree:"):
-			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
-		default:
-			continue
-		}
-		if err != nil {
-			return mem, err
-		}
-	}
-	mem.TotalMemory = total * format.KibiByte
-	mem.FreeSwap = freeSwap * format.KibiByte
-	if available > 0 {
-		mem.FreeMemory = available * format.KibiByte
-	} else {
-		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
-	}
-	return mem, nil
-}
-
-const CpuInfoFilename = "/proc/cpuinfo"
-
-type linuxCpuInfo struct {
-	ID         string `cpuinfo:"processor"`
-	VendorID   string `cpuinfo:"vendor_id"`
-	ModelName  string `cpuinfo:"model name"`
-	PhysicalID string `cpuinfo:"physical id"`
-	Siblings   string `cpuinfo:"siblings"`
-	CoreID     string `cpuinfo:"core id"`
-}
-
-func GetCPUDetails() ([]CPU, error) {
-	file, err := os.Open(CpuInfoFilename)
-	if err != nil {
-		return nil, err
-	}
-	reColumns := regexp.MustCompile("\t+: ")
-	scanner := bufio.NewScanner(file)
-	cpuInfos := []linuxCpuInfo{}
-	cpu := &linuxCpuInfo{}
-	for scanner.Scan() {
-		line := scanner.Text()
-		if sl := reColumns.Split(line, 2); len(sl) > 1 {
-			t := reflect.TypeOf(cpu).Elem()
-			s := reflect.ValueOf(cpu).Elem()
-			for i := range t.NumField() {
-				field := t.Field(i)
-				tag := field.Tag.Get("cpuinfo")
-				if tag == sl[0] {
-					s.FieldByName(field.Name).SetString(sl[1])
-					break
-				}
-			}
-		} else if strings.TrimSpace(line) == "" && cpu.ID != "" {
-			cpuInfos = append(cpuInfos, *cpu)
-			cpu = &linuxCpuInfo{}
-		}
-	}
-
-	// Process the sockets/cores/threads
-	socketByID := map[string]*CPU{}
-	coreBySocket := map[string]map[string]struct{}{}
-	threadsByCoreBySocket := map[string]map[string]int{}
-	for _, c := range cpuInfos {
-		if _, found := socketByID[c.PhysicalID]; !found {
-			socketByID[c.PhysicalID] = &CPU{
-				ID:        c.PhysicalID,
-				VendorID:  c.VendorID,
-				ModelName: c.ModelName,
-			}
-			coreBySocket[c.PhysicalID] = map[string]struct{}{}
-			threadsByCoreBySocket[c.PhysicalID] = map[string]int{}
-		}
-		if c.CoreID != "" {
-			coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID] = struct{}{}
-			threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.CoreID]++
-		} else {
-			coreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID] = struct{}{}
-			threadsByCoreBySocket[c.PhysicalID][c.PhysicalID+":"+c.ID]++
-		}
-	}
-
-	// Tally up the values from the tracking maps
-	for id, s := range socketByID {
-		s.CoreCount = len(coreBySocket[id])
-		s.ThreadCount = 0
-		for _, tc := range threadsByCoreBySocket[id] {
-			s.ThreadCount += tc
-		}
-
-		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
-		efficiencyCoreCount := 0
-		for _, threads := range threadsByCoreBySocket[id] {
-			if threads == 1 {
-				efficiencyCoreCount++
-			}
-		}
-		if efficiencyCoreCount == s.CoreCount {
-			// 1:1 mapping means they're not actually efficiency cores, but regular cores
-			s.EfficiencyCoreCount = 0
-		} else {
-			s.EfficiencyCoreCount = efficiencyCoreCount
-		}
-	}
-
-	result := []CPU{}
-	for _, c := range socketByID {
-		result = append(result, *c)
-	}
-	return result, nil
-}

discover/gpu_windows.go

@@ -1,234 +0,0 @@
-package discover
-
-import (
-	"fmt"
-	"log/slog"
-	"syscall"
-	"unsafe"
-)
-
-type MEMORYSTATUSEX struct {
-	length               uint32
-	MemoryLoad           uint32
-	TotalPhys            uint64
-	AvailPhys            uint64
-	TotalPageFile        uint64
-	AvailPageFile        uint64
-	TotalVirtual         uint64
-	AvailVirtual         uint64
-	AvailExtendedVirtual uint64
-}
-
-var (
-	k32                              = syscall.NewLazyDLL("kernel32.dll")
-	globalMemoryStatusExProc         = k32.NewProc("GlobalMemoryStatusEx")
-	sizeofMemoryStatusEx             = uint32(unsafe.Sizeof(MEMORYSTATUSEX{}))
-	GetLogicalProcessorInformationEx = k32.NewProc("GetLogicalProcessorInformationEx")
-)
-
-var CudartGlobs = []string{
-	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
-}
-
-var NvmlGlobs = []string{
-	"c:\\Windows\\System32\\nvml.dll",
-}
-
-var NvcudaGlobs = []string{
-	"c:\\windows\\system*\\nvcuda.dll",
-}
-
-var OneapiGlobs = []string{
-	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
-}
-
-var (
-	CudartMgmtName = "cudart64_*.dll"
-	NvcudaMgmtName = "nvcuda.dll"
-	NvmlMgmtName   = "nvml.dll"
-	OneapiMgmtName = "ze_intel_gpu64.dll"
-)
-
-func GetCPUMem() (memInfo, error) {
-	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
-	r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
-	if r1 == 0 {
-		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
-	}
-	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
-}
-
-type LOGICAL_PROCESSOR_RELATIONSHIP uint32
-
-const (
-	RelationProcessorCore LOGICAL_PROCESSOR_RELATIONSHIP = iota
-	RelationNumaNode
-	RelationCache
-	RelationProcessorPackage
-	RelationGroup
-	RelationProcessorDie
-	RelationNumaNodeEx
-	RelationProcessorModule
-)
-const RelationAll LOGICAL_PROCESSOR_RELATIONSHIP = 0xffff
-
-type GROUP_AFFINITY struct {
-	Mask     uintptr // KAFFINITY
-	Group    uint16
-	Reserved [3]uint16
-}
-
-type PROCESSOR_RELATIONSHIP struct {
-	Flags           byte
-	EfficiencyClass byte
-	Reserved        [20]byte
-	GroupCount      uint16
-	GroupMask       [1]GROUP_AFFINITY // len GroupCount
-}
-
-// Omitted unused structs: NUMA_NODE_RELATIONSHIP CACHE_RELATIONSHIP GROUP_RELATIONSHIP
-
-type SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX struct {
-	Relationship LOGICAL_PROCESSOR_RELATIONSHIP
-	Size         uint32
-	U            [1]byte // Union len Size
-	// PROCESSOR_RELATIONSHIP
-	// NUMA_NODE_RELATIONSHIP
-	// CACHE_RELATIONSHIP
-	// GROUP_RELATIONSHIP
-}
-
-func (group *GROUP_AFFINITY) IsMember(target *GROUP_AFFINITY) bool {
-	if group == nil || target == nil {
-		return false
-	}
-	return group.Mask&target.Mask != 0
-}
-
-type winPackage struct {
-	groups              []*GROUP_AFFINITY
-	coreCount           int // performance cores = coreCount - efficiencyCoreCount
-	efficiencyCoreCount int
-	threadCount         int
-}
-
-func (pkg *winPackage) IsMember(target *GROUP_AFFINITY) bool {
-	for _, group := range pkg.groups {
-		if group.IsMember(target) {
-			return true
-		}
-	}
-	return false
-}
-
-func getLogicalProcessorInformationEx() ([]byte, error) {
-	buf := make([]byte, 1)
-	bufSize := len(buf)
-	ret, _, err := GetLogicalProcessorInformationEx.Call(
-		uintptr(RelationAll),
-		uintptr(unsafe.Pointer(&buf[0])),
-		uintptr(unsafe.Pointer(&bufSize)),
-	)
-	if ret != 0 {
-		return nil, fmt.Errorf("failed to determine size info ret:%d %w", ret, err)
-	}
-
-	buf = make([]byte, bufSize)
-	ret, _, err = GetLogicalProcessorInformationEx.Call(
-		uintptr(RelationAll),
-		uintptr(unsafe.Pointer(&buf[0])),
-		uintptr(unsafe.Pointer(&bufSize)),
-	)
-	if ret == 0 {
-		return nil, fmt.Errorf("failed to gather processor information ret:%d buflen:%d %w", ret, bufSize, err)
-	}
-	return buf, nil
-}
-
-func processSystemLogicalProcessorInforationList(buf []byte) []*winPackage {
-	var slpi *SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX
-	// Find all the packages first
-	packages := []*winPackage{}
-	for bufOffset := 0; bufOffset < len(buf); bufOffset += int(slpi.Size) {
-		slpi = (*SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX)(unsafe.Pointer(&buf[bufOffset]))
-		if slpi.Relationship != RelationProcessorPackage {
-			continue
-		}
-		pr := (*PROCESSOR_RELATIONSHIP)(unsafe.Pointer(&slpi.U[0]))
-		pkg := &winPackage{}
-		ga0 := unsafe.Pointer(&pr.GroupMask[0])
-		for j := range pr.GroupCount {
-			gm := (*GROUP_AFFINITY)(unsafe.Pointer(uintptr(ga0) + uintptr(j)*unsafe.Sizeof(GROUP_AFFINITY{})))
-			pkg.groups = append(pkg.groups, gm)
-		}
-		packages = append(packages, pkg)
-	}
-
-	slog.Info("packages", "count", len(packages))
-
-	// To identify efficiency cores we have to compare the relative values
-	// Larger values are "less efficient" (aka, more performant)
-	var maxEfficiencyClass byte
-	for bufOffset := 0; bufOffset < len(buf); bufOffset += int(slpi.Size) {
-		slpi = (*SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX)(unsafe.Pointer(&buf[bufOffset]))
-		if slpi.Relationship != RelationProcessorCore {
-			continue
-		}
-		pr := (*PROCESSOR_RELATIONSHIP)(unsafe.Pointer(&slpi.U[0]))
-		if pr.EfficiencyClass > maxEfficiencyClass {
-			maxEfficiencyClass = pr.EfficiencyClass
-		}
-	}
-	if maxEfficiencyClass > 0 {
-		slog.Info("efficiency cores detected", "maxEfficiencyClass", maxEfficiencyClass)
-	}
-
-	// then match up the Cores to the Packages, count up cores, threads and efficiency cores
-	for bufOffset := 0; bufOffset < len(buf); bufOffset += int(slpi.Size) {
-		slpi = (*SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX)(unsafe.Pointer(&buf[bufOffset]))
-		if slpi.Relationship != RelationProcessorCore {
-			continue
-		}
-		pr := (*PROCESSOR_RELATIONSHIP)(unsafe.Pointer(&slpi.U[0]))
-		ga0 := unsafe.Pointer(&pr.GroupMask[0])
-		for j := range pr.GroupCount {
-			gm := (*GROUP_AFFINITY)(unsafe.Pointer(uintptr(ga0) + uintptr(j)*unsafe.Sizeof(GROUP_AFFINITY{})))
-			for _, pkg := range packages {
-				if pkg.IsMember(gm) {
-					pkg.coreCount++
-					if pr.Flags == 0 {
-						pkg.threadCount++
-					} else {
-						pkg.threadCount += 2
-					}
-					if pr.EfficiencyClass < maxEfficiencyClass {
-						pkg.efficiencyCoreCount++
-					}
-				}
-			}
-		}
-	}
-
-	// Sumarize the results
-	for i, pkg := range packages {
-		slog.Info("", "package", i, "cores", pkg.coreCount, "efficiency", pkg.efficiencyCoreCount, "threads", pkg.threadCount)
-	}
-
-	return packages
-}
-
-func GetCPUDetails() ([]CPU, error) {
-	buf, err := getLogicalProcessorInformationEx()
-	if err != nil {
-		return nil, err
-	}
-	packages := processSystemLogicalProcessorInforationList(buf)
-	cpus := make([]CPU, len(packages))
-
-	for i, pkg := range packages {
-		cpus[i].CoreCount = pkg.coreCount
-		cpus[i].EfficiencyCoreCount = pkg.efficiencyCoreCount
-		cpus[i].ThreadCount = pkg.threadCount
-	}
-	return cpus, nil
-}

docs/development.md

@@ -2,13 +2,15 @@
 
 Install required tools:
 
+- cmake version 3.24 or higher
 - go version 1.22 or higher
 - gcc version 11.4.0 or higher
 
-
 ### MacOS
 
-[Download Go](https://go.dev/dl/)
+```bash
+brew install go cmake gcc
+```
 
 Optionally enable debugging and more verbose logging:
 
@@ -20,10 +22,10 @@ export CGO_CFLAGS="-g"
 export OLLAMA_DEBUG=1
 ```
 
-Get the required libraries and build the native LLM code:  (Adjust the job count based on your number of processors for a faster build)
+Get the required libraries and build the native LLM code:
 
 ```bash
-make -j 5
+go generate ./...
 ```
 
 Then build ollama:
@@ -38,17 +40,13 @@ Now you can run `ollama`:
 ./ollama
 ```
 
-#### Xcode 15 warnings
-
-If you are using Xcode newer than version 14, you may see a warning during `go build` about `ld: warning: ignoring duplicate libraries: '-lobjc'` due to Golang issue https://github.com/golang/go/issues/67799 which can be safely ignored.  You can suppress the warning with `export CGO_LDFLAGS="-Wl,-no_warn_duplicate_libraries"`
-
 ### Linux
 
 #### Linux CUDA (NVIDIA)
 
 _Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
 
-Install `make`, `gcc` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
 development and runtime packages.
 
 Typically the build scripts will auto-detect CUDA, however, if your Linux distro
@@ -57,10 +55,10 @@ specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
 libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
 a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
 
-Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
+Then generate dependencies:
 
 ```
-make -j 5
+go generate ./...
 ```
 
 Then build the binary:
@@ -73,7 +71,7 @@ go build .
 
 _Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
 
-Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `make`, `gcc`, and `golang`.
+Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `cmake` and `golang`.
 
 Typically the build scripts will auto-detect ROCm, however, if your Linux distro
 or installation approach uses unusual paths, you can specify the location by
@@ -82,10 +80,8 @@ install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
 CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
 the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
 
-Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
-
 ```
-make -j 5
+go generate ./...
 ```
 
 Then build the binary:
@@ -98,59 +94,60 @@ ROCm requires elevated privileges to access the GPU at runtime. On most distros
 
 #### Advanced CPU Settings
 
-By default, running `make` will compile a few different variations
+By default, running `go generate ./...` will compile a few different variations
 of the LLM library based on common CPU families and vector math capabilities,
 including a lowest-common-denominator which should run on almost any 64 bit CPU
 somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
-load. 
+load. If you would like to build a CPU-based build customized for your
+processor, you can set `OLLAMA_CUSTOM_CPU_DEFS` to the llama.cpp flags you would
+like to use. For example, to compile an optimized binary for an Intel i9-9880H,
+you might use:
 
-Custom CPU settings are not currently supported in the new Go server build but will be added back after we complete the transition.
+```
+OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
+go build .
+```
 
 #### Containerized Linux Build
 
-If you have Docker available, you can build linux binaries with `OLLAMA_NEW_RUNNERS=1 ./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
+If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
 
 ### Windows
 
-The following tools are required as a minimal development environment to build CPU inference support.
+Note: The Windows build for Ollama is still under development.
 
+First, install required tools:
+
+- MSVC toolchain - C/C++ and cmake as minimal requirements
 - Go version 1.22 or higher
-  - https://go.dev/dl/
-- Git
-  - https://git-scm.com/download/win
-- GCC and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
+- MinGW (pick one variant) with GCC.
+  - [MinGW-w64](https://www.mingw-w64.org/)
   - [MSYS2](https://www.msys2.org/)
-    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-ucrt-x86_64-gcc make` to install the required tools
-  - Assuming you used the default install prefix for msys2 above, add `c:\msys64\ucrt64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
 
 Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"
-make -j 8
+go generate ./...
 go build .
 ```
 
-#### GPU Support
-
-The GPU tools require the Microsoft native build tools.  To build either CUDA or ROCm, you must first install MSVC via Visual Studio:
-
-- Make sure to select `Desktop development with C++` as a Workload during the Visual Studio install
-- You must complete the Visual Studio install and run it once **BEFORE** installing CUDA or ROCm for the tools to properly register
-- Add the location of the **64 bit (x64)** compiler (`cl.exe`) to your `PATH`
-- Note: the default Developer Shell may configure the 32 bit (x86) compiler which will lead to build failures.  Ollama requires a 64 bit toolchain.
-
 #### Windows CUDA (NVIDIA)
 
-In addition to the common Windows development tools and MSVC described above:
+In addition to the common Windows development tools described above, install CUDA after installing MSVC.
 
 - [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)
 
+
 #### Windows ROCm (AMD Radeon)
 
-In addition to the common Windows development tools and MSVC described above:
+In addition to the common Windows development tools described above, install AMDs HIP package after installing MSVC.
 
 - [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+- [Strawberry Perl](https://strawberryperl.com/)
+
+Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
 
 #### Windows arm64
 
@@ -169,4 +166,4 @@ Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64
 pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
 ```
 
-You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)

docs/openai.md

@@ -37,7 +37,7 @@ response = client.chat.completions.create(
                 {"type": "text", "text": "What's in this image?"},
                 {
                     "type": "image_url",
-                    "image_url": "data:image/png;base64,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",
+                    "image_url": "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",
                 },
             ],
         }
@@ -86,7 +86,7 @@ const response = await openai.chat.completions.create({
             { type: "text", text: "What's in this image?" },
             {
             type: "image_url",
-            image_url: "data:image/png;base64,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",
+            image_url: "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",
             },
         ],
         },
@@ -142,7 +142,7 @@ curl http://localhost:11434/v1/chat/completions \
           {
             "type": "image_url",
             "image_url": {
-               "url": "data:image/png;base64,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"
+               "url": "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"
             }
           }
         ]

envconfig/config.go

@@ -72,7 +72,6 @@ func Origins() (origins []string) {
 		"app://*",
 		"file://*",
 		"tauri://*",
-		"vscode-webview://*",
 	)
 
 	return origins
@@ -161,8 +160,6 @@ var (
 	SchedSpread = Bool("OLLAMA_SCHED_SPREAD")
 	// IntelGPU enables experimental Intel GPU detection.
 	IntelGPU = Bool("OLLAMA_INTEL_GPU")
-	// MultiUserCache optimizes prompt caching for multi-user scenarios
-	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
 )
 
 func String(s string) func() string {
@@ -248,7 +245,6 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
-		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
 
 		// Informational
 		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},

envconfig/config_test.go

@@ -68,7 +68,6 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
-			"vscode-webview://*",
 		}},
 		{"http://10.0.0.1", []string{
 			"http://10.0.0.1",
@@ -87,7 +86,6 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
-			"vscode-webview://*",
 		}},
 		{"http://172.16.0.1,https://192.168.0.1", []string{
 			"http://172.16.0.1",
@@ -107,7 +105,6 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
-			"vscode-webview://*",
 		}},
 		{"http://totally.safe,http://definitely.legit", []string{
 			"http://totally.safe",
@@ -127,7 +124,6 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
-			"vscode-webview://*",
 		}},
 	}
 	for _, tt := range cases {

fileutils/README.md

@@ -1,3 +0,0 @@
-# `modelfile`
-
-This package provides utilities for loading and inspecting model files

fileutils/ggml_test.go

@@ -1 +0,0 @@
-package fileutils

go.mod

@@ -22,7 +22,6 @@ require (
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
-	golang.org/x/image v0.14.0
 )
 
 require (

go.sum

@@ -230,8 +230,6 @@ golang.org/x/image v0.0.0-20200430140353-33d19683fad8/go.mod h1:FeLwcggjj3mMvU+o
 golang.org/x/image v0.0.0-20200618115811-c13761719519/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20201208152932-35266b937fa6/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20210216034530-4410531fe030/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
-golang.org/x/image v0.14.0 h1:tNgSxAFe3jC4uYqvZdTr84SZoM1KfwdC9SKIFrLjFn4=
-golang.org/x/image v0.14.0/go.mod h1:HUYqC05R2ZcZ3ejNQsIHQDQiwWM4JBqmm6MKANTp4LE=
 golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
 golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
 golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=

gpu/amd_common.go

@@ -1,6 +1,6 @@
 //go:build linux || windows
 
-package discover
+package gpu
 
 import (
 	"errors"

gpu/amd_hip_windows.go

@@ -1,4 +1,4 @@
-package discover
+package gpu
 
 import (
 	"errors"

gpu/amd_linux.go

@@ -1,4 +1,4 @@
-package discover
+package gpu
 
 import (
 	"bufio"
@@ -47,11 +47,10 @@ var (
 )
 
 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
-// Only called once during bootstrap
-func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
+func AMDGetGPUInfo() []RocmGPUInfo {
 	resp := []RocmGPUInfo{}
 	if !AMDDetected() {
-		return resp, fmt.Errorf("AMD GPUs not detected")
+		return resp
 	}
 
 	// Opportunistic logging of driver version to aid in troubleshooting
@@ -195,9 +194,13 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 
 		// Shouldn't happen, but just in case...
 		if gpuID < 0 {
-			err := fmt.Errorf("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
-			slog.Error(err.Error())
-			return nil, err
+			slog.Error("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
+			return nil
+		}
+
+		if int(major) < RocmComputeMin {
+			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch), "gpu", gpuID)
+			continue
 		}
 
 		// Look up the memory for the current node
@@ -267,12 +270,19 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			break
 		}
 
+		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
+		if totalMemory < IGPUMemLimit {
+			slog.Info("unsupported Radeon iGPU detected skipping", "id", gpuID, "total", format.HumanBytes2(totalMemory))
+			continue
+		}
 		var name string
 		// TODO - PCI ID lookup
 		if vendor > 0 && device > 0 {
 			name = fmt.Sprintf("%04x:%04x", vendor, device)
 		}
 
+		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
 		gpuInfo := RocmGPUInfo{
 			GpuInfo: GpuInfo{
 				Library: "rocm",
@@ -290,31 +300,6 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			usedFilepath: usedFile,
 		}
 
-		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
-		if totalMemory < IGPUMemLimit {
-			reason := "unsupported Radeon iGPU detected skipping"
-			slog.Info(reason, "id", gpuID, "total", format.HumanBytes2(totalMemory))
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			continue
-		}
-
-		if int(major) < RocmComputeMin {
-			reason := fmt.Sprintf("amdgpu too old gfx%d%x%x", major, minor, patch)
-			slog.Warn(reason, "gpu", gpuID)
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-
-			continue
-		}
-
-		slog.Debug("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
-		slog.Debug("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
-
 		// If the user wants to filter to a subset of devices, filter out if we aren't a match
 		if len(visibleDevices) > 0 {
 			include := false
@@ -325,13 +310,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				}
 			}
 			if !include {
-				reason := "filtering out device per user request"
-				slog.Info(reason, "id", gpuInfo.ID, "visible_devices", visibleDevices)
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  reason,
-				})
-
+				slog.Info("filtering out device per user request", "id", gpuInfo.ID, "visible_devices", visibleDevices)
 				continue
 			}
 		}
@@ -341,13 +320,8 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		if libDir == "" {
 			libDir, err = AMDValidateLibDir()
 			if err != nil {
-				err = fmt.Errorf("unable to verify rocm library: %w", err)
-				slog.Warn(err.Error())
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  err.Error(),
-				})
-				return nil, err
+				slog.Warn("unable to verify rocm library, will use cpu", "error", err)
+				return nil
 			}
 		}
 		gpuInfo.DependencyPath = libDir
@@ -357,25 +331,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			if len(supported) == 0 {
 				supported, err = GetSupportedGFX(libDir)
 				if err != nil {
-					err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
-					slog.Warn(err.Error())
-					unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-						GpuInfo: gpuInfo.GpuInfo,
-						Reason:  err.Error(),
-					})
-					return nil, err
+					slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
+					return nil
 				}
 				slog.Debug("rocm supported GPUs", "types", supported)
 			}
 			gfx := gpuInfo.Compute
 			if !slices.Contains[[]string, string](supported, gfx) {
-				reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
-				slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
-				unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-					GpuInfo: gpuInfo.GpuInfo,
-					Reason:  reason,
-				})
-
+				slog.Warn("amdgpu is not supported", "gpu", gpuInfo.ID, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
 				continue
@@ -395,16 +358,13 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		resp = append(resp, gpuInfo)
 	}
 	if len(resp) == 0 {
-		err := fmt.Errorf("no compatible amdgpu devices detected")
-		slog.Info(err.Error())
-		return nil, err
+		slog.Info("no compatible amdgpu devices detected")
 	}
 	if err := verifyKFDDriverAccess(); err != nil {
-		err = fmt.Errorf("amdgpu devices detected but permission problems block access: %w", err)
-		slog.Error(err.Error())
-		return nil, err
+		slog.Error("amdgpu devices detected but permission problems block access", "error", err)
+		return nil
 	}
-	return resp, nil
+	return resp
 }
 
 // Quick check for AMD driver so we can skip amdgpu discovery if not present

gpu/amd_windows.go

@@ -1,9 +1,8 @@
-package discover
+package gpu
 
 import (
 	"bytes"
 	"errors"
-	"fmt"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -27,13 +26,12 @@ var (
 	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\6.1\\bin"} // TODO glob?
 )
 
-// Only called once during bootstrap
-func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
+func AMDGetGPUInfo() []RocmGPUInfo {
 	resp := []RocmGPUInfo{}
 	hl, err := NewHipLib()
 	if err != nil {
 		slog.Debug(err.Error())
-		return nil, err
+		return nil
 	}
 	defer hl.Release()
 
@@ -46,15 +44,12 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
 	if count == 0 {
-		err := fmt.Errorf("no compatible amdgpu devices detected")
-		slog.Info(err.Error())
-		return nil, err
+		return nil
 	}
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
-		err = fmt.Errorf("unable to verify rocm library: %w", err)
-		slog.Warn(err.Error())
-		return nil, err
+		slog.Warn("unable to verify rocm library, will use cpu", "error", err)
+		return nil
 	}
 
 	var supported []string
@@ -62,9 +57,8 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
-			err = fmt.Errorf("failed to lookup supported GFX types: %w", err)
-			slog.Warn(err.Error())
-			return nil, err
+			slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
+			return nil
 		}
 	} else {
 		slog.Info("skipping rocm gfx compatibility check", "HSA_OVERRIDE_GFX_VERSION", gfxOverride)
@@ -93,6 +87,21 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
 		// slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
+		if strings.EqualFold(name, iGPUName) {
+			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
+			continue
+		}
+		if gfxOverride == "" {
+			// Strip off Target Features when comparing
+			if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
+				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
+				// TODO - consider discrete markdown just for ROCM troubleshooting?
+				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
+				continue
+			} else {
+				slog.Debug("amdgpu is supported", "gpu", i, "gpu_type", gfx)
+			}
+		}
 
 		freeMemory, totalMemory, err := hl.HipMemGetInfo()
 		if err != nil {
@@ -100,6 +109,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			continue
 		}
 
+		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
+		if totalMemory < IGPUMemLimit {
+			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
+			continue
+		}
+
+		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
+		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
 		gpuInfo := RocmGPUInfo{
 			GpuInfo: GpuInfo{
 				Library: "rocm",
@@ -121,38 +138,10 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			index: i,
 		}
 
-		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
-		if strings.EqualFold(name, iGPUName) || totalMemory < IGPUMemLimit {
-			reason := "unsupported Radeon iGPU detected skipping"
-			slog.Info(reason, "id", gpuInfo.ID, "total", format.HumanBytes2(totalMemory))
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			continue
-		}
-
-		// Strip off Target Features when comparing
-		if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
-			reason := fmt.Sprintf("amdgpu is not supported (supported types:%s)", supported)
-			slog.Warn(reason, "gpu_type", gfx, "gpu", gpuInfo.ID, "library", libDir)
-			unsupportedGPUs = append(unsupportedGPUs, UnsupportedGPUInfo{
-				GpuInfo: gpuInfo.GpuInfo,
-				Reason:  reason,
-			})
-			// HSA_OVERRIDE_GFX_VERSION not supported on windows
-			continue
-		} else {
-			slog.Debug("amdgpu is supported", "gpu", i, "gpu_type", gfx)
-		}
-
-		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
-		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
-
 		resp = append(resp, gpuInfo)
 	}
 
-	return resp, nil
+	return resp
 }
 
 func AMDValidateLibDir() (string, error) {

gpu/cpu_common.go

@@ -1,4 +1,4 @@
-package discover
+package gpu
 
 import (
 	"os"

gpu/cuda_common.go

@@ -1,6 +1,6 @@
 //go:build linux || windows
 
-package discover
+package gpu
 
 import (
 	"log/slog"

gpu/gpu.go

@@ -1,6 +1,6 @@
 //go:build linux || windows
 
-package discover
+package gpu
 
 /*
 #cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
@@ -54,13 +54,6 @@ var (
 	nvmlLibPath   string
 	rocmGPUs      []RocmGPUInfo
 	oneapiGPUs    []OneapiGPUInfo
-
-	// If any discovered GPUs are incompatible, report why
-	unsupportedGPUs []UnsupportedGPUInfo
-
-	// Keep track of errors during bootstrapping so that if GPUs are missing
-	// they expected to be present this may explain why
-	bootstrapErrors []error
 )
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
@@ -77,17 +70,16 @@ func initCudaHandles() *cudaHandles {
 
 	cHandles := &cudaHandles{}
 	// Short Circuit if we already know which library to use
-	// ignore bootstrap errors in this case since we already recorded them
 	if nvmlLibPath != "" {
-		cHandles.nvml, _, _ = loadNVMLMgmt([]string{nvmlLibPath})
+		cHandles.nvml, _ = LoadNVMLMgmt([]string{nvmlLibPath})
 		return cHandles
 	}
 	if nvcudaLibPath != "" {
-		cHandles.deviceCount, cHandles.nvcuda, _, _ = loadNVCUDAMgmt([]string{nvcudaLibPath})
+		cHandles.deviceCount, cHandles.nvcuda, _ = LoadNVCUDAMgmt([]string{nvcudaLibPath})
 		return cHandles
 	}
 	if cudartLibPath != "" {
-		cHandles.deviceCount, cHandles.cudart, _, _ = loadCUDARTMgmt([]string{cudartLibPath})
+		cHandles.deviceCount, cHandles.cudart, _ = LoadCUDARTMgmt([]string{cudartLibPath})
 		return cHandles
 	}
 
@@ -110,21 +102,18 @@ func initCudaHandles() *cudaHandles {
 	if len(NvmlGlobs) > 0 {
 		nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
 		if len(nvmlLibPaths) > 0 {
-			nvml, libPath, err := loadNVMLMgmt(nvmlLibPaths)
+			nvml, libPath := LoadNVMLMgmt(nvmlLibPaths)
 			if nvml != nil {
 				slog.Debug("nvidia-ml loaded", "library", libPath)
 				cHandles.nvml = nvml
 				nvmlLibPath = libPath
 			}
-			if err != nil {
-				bootstrapErrors = append(bootstrapErrors, err)
-			}
 		}
 	}
 
 	nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
 	if len(nvcudaLibPaths) > 0 {
-		deviceCount, nvcuda, libPath, err := loadNVCUDAMgmt(nvcudaLibPaths)
+		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
 		if nvcuda != nil {
 			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
 			cHandles.nvcuda = nvcuda
@@ -132,14 +121,11 @@ func initCudaHandles() *cudaHandles {
 			nvcudaLibPath = libPath
 			return cHandles
 		}
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
 	}
 
 	cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
-		deviceCount, cudart, libPath, err := loadCUDARTMgmt(cudartLibPaths)
+		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
 			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
 			cHandles.cudart = cudart
@@ -147,9 +133,6 @@ func initCudaHandles() *cudaHandles {
 			cudartLibPath = libPath
 			return cHandles
 		}
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
 	}
 
 	return cHandles
@@ -160,19 +143,14 @@ func initOneAPIHandles() *oneapiHandles {
 	oHandles := &oneapiHandles{}
 
 	// Short Circuit if we already know which library to use
-	// ignore bootstrap errors in this case since we already recorded them
 	if oneapiLibPath != "" {
-		oHandles.deviceCount, oHandles.oneapi, _, _ = loadOneapiMgmt([]string{oneapiLibPath})
+		oHandles.deviceCount, oHandles.oneapi, _ = LoadOneapiMgmt([]string{oneapiLibPath})
 		return oHandles
 	}
 
 	oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
 	if len(oneapiLibPaths) > 0 {
-		var err error
-		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath, err = loadOneapiMgmt(oneapiLibPaths)
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
+		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath = LoadOneapiMgmt(oneapiLibPaths)
 	}
 
 	return oHandles
@@ -219,7 +197,6 @@ func GetGPUInfo() GpuInfoList {
 
 	if !bootstrapped {
 		slog.Info("looking for compatible GPUs")
-		bootstrapErrors = []error{}
 		needRefresh = false
 		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
@@ -229,10 +206,7 @@ func GetGPUInfo() GpuInfoList {
 			slog.Warn("error looking up system memory", "error", err)
 		}
 		depPath := LibraryDir()
-		details, err := GetCPUDetails()
-		if err != nil {
-			slog.Warn("failed to lookup CPU details", "error", err)
-		}
+
 		cpus = []CPUInfo{
 			{
 				GpuInfo: GpuInfo{
@@ -242,15 +216,12 @@ func GetGPUInfo() GpuInfoList {
 					ID:             "0",
 					DependencyPath: depPath,
 				},
-				CPUs: details,
 			},
 		}
 
 		// Fallback to CPU mode if we're lacking required vector extensions on x86
 		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
-			err := fmt.Errorf("CPU does not have minimum vector extensions, GPU inference disabled.  Required:%s  Detected:%s", GPURunnerCPUCapability, cpuCapability)
-			slog.Warn(err.Error())
-			bootstrapErrors = append(bootstrapErrors, err)
+			slog.Warn("CPU does not have minimum vector extensions, GPU inference disabled", "required", GPURunnerCPUCapability, "detected", cpuCapability)
 			bootstrapped = true
 			// No need to do any GPU discovery, since we can't run on them
 			return GpuInfoList{cpus[0].GpuInfo}
@@ -282,6 +253,10 @@ func GetGPUInfo() GpuInfoList {
 					C.free(unsafe.Pointer(memInfo.err))
 					continue
 				}
+				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+					continue
+				}
 				gpuInfo.TotalMemory = uint64(memInfo.total)
 				gpuInfo.FreeMemory = uint64(memInfo.free)
 				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
@@ -304,15 +279,6 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.Variant = variant
 
-				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
-					unsupportedGPUs = append(unsupportedGPUs,
-						UnsupportedGPUInfo{
-							GpuInfo: gpuInfo.GpuInfo,
-						})
-					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
-					continue
-				}
-
 				// query the management library as well so we can record any skew between the two
 				// which represents overhead on the GPU we must set aside on subsequent updates
 				if cHandles.nvml != nil {
@@ -375,10 +341,7 @@ func GetGPUInfo() GpuInfoList {
 			}
 		}
 
-		rocmGPUs, err = AMDGetGPUInfo()
-		if err != nil {
-			bootstrapErrors = append(bootstrapErrors, err)
-		}
+		rocmGPUs = AMDGetGPUInfo()
 		bootstrapped = true
 		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
 			slog.Info("no compatible GPUs were discovered")
@@ -563,114 +526,92 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	return gpuLibPaths
 }
 
-// Bootstrap the runtime library
-// Returns: num devices, handle, libPath, error
-func loadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string, error) {
+func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
 	var resp C.cudart_init_resp_t
 	resp.ch.verbose = getVerboseState()
-	var err error
 	for _, libPath := range cudartLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.cudart_init(lib, &resp)
 		if resp.err != nil {
-			err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
-			slog.Debug(err.Error())
+			slog.Debug("Unable to load cudart", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			err = nil
-			return int(resp.num_devices), &resp.ch, libPath, err
+			return int(resp.num_devices), &resp.ch, libPath
 		}
 	}
-	return 0, nil, "", err
+	return 0, nil, ""
 }
 
-// Bootstrap the driver library
-// Returns: num devices, handle, libPath, error
-func loadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string, error) {
+func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 	var resp C.nvcuda_init_resp_t
 	resp.ch.verbose = getVerboseState()
-	var err error
 	for _, libPath := range nvcudaLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
 			// Decide what log level based on the type of error message to help users understand why
+			msg := C.GoString(resp.err)
 			switch resp.cudaErr {
 			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
-				err = fmt.Errorf("version mismatch between driver and cuda driver library - reboot or upgrade may be required: library %s", libPath)
-				slog.Warn(err.Error())
+				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
 			case C.CUDA_ERROR_NO_DEVICE:
-				err = fmt.Errorf("no nvidia devices detected by library %s", libPath)
-				slog.Info(err.Error())
+				slog.Info("no nvidia devices detected", "library", libPath)
 			case C.CUDA_ERROR_UNKNOWN:
-				err = fmt.Errorf("unknown error initializing cuda driver library %s: %s. see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information", libPath, C.GoString(resp.err))
-				slog.Warn(err.Error())
+				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
+				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
 			default:
-				msg := C.GoString(resp.err)
 				if strings.Contains(msg, "wrong ELF class") {
 					slog.Debug("skipping 32bit library", "library", libPath)
 				} else {
-					err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
-					slog.Info(err.Error())
+					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
 				}
 			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			err = nil
-			return int(resp.num_devices), &resp.ch, libPath, err
+			return int(resp.num_devices), &resp.ch, libPath
 		}
 	}
-	return 0, nil, "", err
+	return 0, nil, ""
 }
 
-// Bootstrap the management library
-// Returns: handle, libPath, error
-func loadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string, error) {
+func LoadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string) {
 	var resp C.nvml_init_resp_t
 	resp.ch.verbose = getVerboseState()
-	var err error
 	for _, libPath := range nvmlLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.nvml_init(lib, &resp)
 		if resp.err != nil {
-			err = fmt.Errorf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err))
-			slog.Info(err.Error())
+			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			err = nil
-			return &resp.ch, libPath, err
+			return &resp.ch, libPath
 		}
 	}
-	return nil, "", err
+	return nil, ""
 }
 
-// bootstrap the Intel GPU library
-// Returns: num devices, handle, libPath, error
-func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, error) {
+func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
 	var resp C.oneapi_init_resp_t
 	num_devices := 0
 	resp.oh.verbose = getVerboseState()
-	var err error
 	for _, libPath := range oneapiLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.oneapi_init(lib, &resp)
 		if resp.err != nil {
-			err = fmt.Errorf("Unable to load oneAPI management library %s: %s", libPath, C.GoString(resp.err))
-			slog.Debug(err.Error())
+			slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			err = nil
 			for i := range resp.oh.num_drivers {
 				num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
 			}
-			return num_devices, &resp.oh, libPath, err
+			return num_devices, &resp.oh, libPath
 		}
 	}
-	return 0, nil, "", err
+	return 0, nil, ""
 }
 
 func getVerboseState() C.uint16_t {
@@ -728,23 +669,3 @@ func LibraryDir() string {
 	slog.Warn("unable to locate gpu dependency libraries")
 	return ""
 }
-
-func GetSystemInfo() SystemInfo {
-	gpus := GetGPUInfo()
-	gpuMutex.Lock()
-	defer gpuMutex.Unlock()
-	discoveryErrors := []string{}
-	for _, err := range bootstrapErrors {
-		discoveryErrors = append(discoveryErrors, err.Error())
-	}
-	if len(gpus) == 1 && gpus[0].Library == "cpu" {
-		gpus = []GpuInfo{}
-	}
-
-	return SystemInfo{
-		System:          cpus[0],
-		GPUs:            gpus,
-		UnsupportedGPUs: unsupportedGPUs,
-		DiscoveryErrors: discoveryErrors,
-	}
-}

gpu/gpu_darwin.go

@@ -1,6 +1,6 @@
 //go:build darwin
 
-package discover
+package gpu
 
 /*
 #cgo CFLAGS: -x objective-c
@@ -10,9 +10,7 @@ package discover
 import "C"
 
 import (
-	"log/slog"
 	"runtime"
-	"syscall"
 
 	"github.com/ollama/ollama/format"
 )
@@ -68,34 +66,3 @@ func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	// No-op on darwin
 	return "", ""
 }
-
-func GetSystemInfo() SystemInfo {
-	mem, _ := GetCPUMem()
-	query := "hw.perflevel0.physicalcpu"
-	perfCores, err := syscall.SysctlUint32(query)
-	if err != nil {
-		slog.Warn("failed to discover physical CPU details", "query", query, "error", err)
-	}
-	query = "hw.perflevel1.physicalcpu"
-	efficiencyCores, _ := syscall.SysctlUint32(query) // On x86 xeon this wont return data
-
-	// Determine thread count
-	query = "hw.logicalcpu"
-	logicalCores, _ := syscall.SysctlUint32(query)
-
-	return SystemInfo{
-		System: CPUInfo{
-			GpuInfo: GpuInfo{
-				memInfo: mem,
-			},
-			CPUs: []CPU{
-				{
-					CoreCount:           int(perfCores + efficiencyCores),
-					EfficiencyCoreCount: int(efficiencyCores),
-					ThreadCount:         int(logicalCores),
-				},
-			},
-		},
-		GPUs: GetGPUInfo(),
-	}
-}

gpu/gpu_linux.go

@@ -0,0 +1,92 @@
+package gpu
+
+import (
+	"bufio"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/ollama/ollama/format"
+)
+
+var CudartGlobs = []string{
+	"/usr/local/cuda/lib64/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
+	"/usr/lib/wsl/lib/libcudart.so*",
+	"/usr/lib/wsl/drivers/*/libcudart.so*",
+	"/opt/cuda/lib64/libcudart.so*",
+	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
+	"/usr/local/cuda/lib*/libcudart.so*",
+	"/usr/lib*/libcudart.so*",
+	"/usr/local/lib*/libcudart.so*",
+}
+
+var NvmlGlobs = []string{}
+
+var NvcudaGlobs = []string{
+	"/usr/local/cuda*/targets/*/lib/libcuda.so*",
+	"/usr/lib/*-linux-gnu/nvidia/current/libcuda.so*",
+	"/usr/lib/*-linux-gnu/libcuda.so*",
+	"/usr/lib/wsl/lib/libcuda.so*",
+	"/usr/lib/wsl/drivers/*/libcuda.so*",
+	"/opt/cuda/lib*/libcuda.so*",
+	"/usr/local/cuda/lib*/libcuda.so*",
+	"/usr/lib*/libcuda.so*",
+	"/usr/local/lib*/libcuda.so*",
+}
+
+var OneapiGlobs = []string{
+	"/usr/lib/x86_64-linux-gnu/libze_intel_gpu.so*",
+	"/usr/lib*/libze_intel_gpu.so*",
+}
+
+var (
+	CudartMgmtName = "libcudart.so*"
+	NvcudaMgmtName = "libcuda.so*"
+	NvmlMgmtName   = "" // not currently wired on linux
+	OneapiMgmtName = "libze_intel_gpu.so*"
+)
+
+func GetCPUMem() (memInfo, error) {
+	var mem memInfo
+	var total, available, free, buffers, cached, freeSwap uint64
+	f, err := os.Open("/proc/meminfo")
+	if err != nil {
+		return mem, err
+	}
+	defer f.Close()
+	s := bufio.NewScanner(f)
+	for s.Scan() {
+		line := s.Text()
+		switch {
+		case strings.HasPrefix(line, "MemTotal:"):
+			_, err = fmt.Sscanf(line, "MemTotal:%d", &total)
+		case strings.HasPrefix(line, "MemAvailable:"):
+			_, err = fmt.Sscanf(line, "MemAvailable:%d", &available)
+		case strings.HasPrefix(line, "MemFree:"):
+			_, err = fmt.Sscanf(line, "MemFree:%d", &free)
+		case strings.HasPrefix(line, "Buffers:"):
+			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
+		case strings.HasPrefix(line, "Cached:"):
+			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
+		case strings.HasPrefix(line, "SwapFree:"):
+			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
+		default:
+			continue
+		}
+		if err != nil {
+			return mem, err
+		}
+	}
+	mem.TotalMemory = total * format.KibiByte
+	mem.FreeSwap = freeSwap * format.KibiByte
+	if available > 0 {
+		mem.FreeMemory = available * format.KibiByte
+	} else {
+		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	}
+	return mem, nil
+}

gpu/gpu_oneapi.go

@@ -1,6 +1,6 @@
 //go:build linux || windows
 
-package discover
+package gpu
 
 import (
 	"log/slog"

gpu/gpu_test.go

@@ -1,4 +1,4 @@
-package discover
+package gpu
 
 import (
 	"runtime"

gpu/gpu_windows.go

@@ -0,0 +1,57 @@
+package gpu
+
+import (
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+type MEMORYSTATUSEX struct {
+	length               uint32
+	MemoryLoad           uint32
+	TotalPhys            uint64
+	AvailPhys            uint64
+	TotalPageFile        uint64
+	AvailPageFile        uint64
+	TotalVirtual         uint64
+	AvailVirtual         uint64
+	AvailExtendedVirtual uint64
+}
+
+var (
+	k32                      = syscall.NewLazyDLL("kernel32.dll")
+	globalMemoryStatusExProc = k32.NewProc("GlobalMemoryStatusEx")
+	sizeofMemoryStatusEx     = uint32(unsafe.Sizeof(MEMORYSTATUSEX{}))
+)
+
+var CudartGlobs = []string{
+	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
+}
+
+var NvmlGlobs = []string{
+	"c:\\Windows\\System32\\nvml.dll",
+}
+
+var NvcudaGlobs = []string{
+	"c:\\windows\\system*\\nvcuda.dll",
+}
+
+var OneapiGlobs = []string{
+	"c:\\Windows\\System32\\DriverStore\\FileRepository\\*\\ze_intel_gpu64.dll",
+}
+
+var (
+	CudartMgmtName = "cudart64_*.dll"
+	NvcudaMgmtName = "nvcuda.dll"
+	NvmlMgmtName   = "nvml.dll"
+	OneapiMgmtName = "ze_intel_gpu64.dll"
+)
+
+func GetCPUMem() (memInfo, error) {
+	memStatus := MEMORYSTATUSEX{length: sizeofMemoryStatusEx}
+	r1, _, err := globalMemoryStatusExProc.Call(uintptr(unsafe.Pointer(&memStatus)))
+	if r1 == 0 {
+		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
+	}
+	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
+}

gpu/types.go

@@ -1,4 +1,4 @@
-package discover
+package gpu
 
 import (
 	"fmt"
@@ -10,11 +10,11 @@ import (
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
-	FreeSwap    uint64 `json:"free_swap,omitempty"` // TODO split this out for system only
+	FreeSwap    uint64 `json:"free_swap,omitempty"`
 }
 
 // Beginning of an `ollama info` command
-type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
+type GpuInfo struct {
 	memInfo
 	Library string `json:"library,omitempty"`
 
@@ -49,17 +49,6 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 
 type CPUInfo struct {
 	GpuInfo
-	CPUs []CPU
-}
-
-// CPU type represents a CPU Package occupying a socket
-type CPU struct {
-	ID                  string `cpuinfo:"processor"`
-	VendorID            string `cpuinfo:"vendor_id"`
-	ModelName           string `cpuinfo:"model name"`
-	CoreCount           int
-	EfficiencyCoreCount int // Performance = CoreCount - Efficiency
-	ThreadCount         int
 }
 
 type CudaGPUInfo struct {
@@ -87,11 +76,6 @@ type OneapiGPUInfoList []OneapiGPUInfo
 
 type GpuInfoList []GpuInfo
 
-type UnsupportedGPUInfo struct {
-	GpuInfo
-	Reason string `json:"reason"`
-}
-
 // Split up the set of gpu info's by Library and variant
 func (l GpuInfoList) ByLibrary() []GpuInfoList {
 	resp := []GpuInfoList{}
@@ -162,19 +146,3 @@ func (c CPUCapability) String() string {
 		return "no vector extensions"
 	}
 }
-
-type SystemInfo struct {
-	System          CPUInfo              `json:"system"`
-	GPUs            []GpuInfo            `json:"gpus"`
-	UnsupportedGPUs []UnsupportedGPUInfo `json:"unsupported_gpus"`
-	DiscoveryErrors []string             `json:"discovery_errors"`
-}
-
-// Return the optimal number of threads to use for inference
-func (si SystemInfo) GetOptimalThreadCount() int {
-	if len(si.System.CPUs) == 0 {
-		return 0
-	}
-	// Allocate thread count matching the performance cores on a single socket
-	return si.System.CPUs[0].CoreCount - si.System.CPUs[0].EfficiencyCoreCount
-}

integration/basic_test.go

@@ -30,22 +30,6 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }
 
-func TestUnicodeOutput(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
-	defer cancel()
-	// Set up the test data
-	req := api.GenerateRequest{
-		Model:  "gemma2:2b",
-		Prompt: "Output some smily face emoji",
-		Stream: &stream,
-		Options: map[string]interface{}{
-			"temperature": 0,
-			"seed":        123,
-		},
-	}
-	GenerateTestHelper(ctx, t, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"})
-}
-
 func TestUnicodeModelDir(t *testing.T) {
 	// This is only useful for Windows with utf-16 characters, so skip this test for other platforms
 	if runtime.GOOS != "windows" {

integration/concurrency_test.go

@@ -42,7 +42,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 		}
 		resp = [2][]string{
 			{"sunlight"},
-			{"england", "english", "massachusetts", "pilgrims", "british", "festival"},
+			{"england", "english", "massachusetts", "pilgrims", "british"},
 		}
 	)
 	var wg sync.WaitGroup

integration/utils_test.go

@@ -275,7 +275,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 				break
 			}
 		}
-		require.True(t, atLeastOne, "%s: none of %v found in %s", genReq.Model, anyResp, response)
+		require.True(t, atLeastOne, "none of %v found in %s", anyResp, response)
 		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
 	case <-ctx.Done():
 		t.Error("outer test context done while waiting for generate")

llama/.gitignore

@@ -1,3 +0,0 @@
-*.bin
-*.gguf
-build/

llama/Makefile

@@ -1,57 +0,0 @@
-# top level makefile for Go server
-include make/common-defs.make
-
-RUNNER_TARGETS := default
-
-# Determine which if any GPU runners we should build
-ifeq ($(OS),windows)
-	CUDA_PATH?=$(shell cygpath -m -s "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\" 2>/dev/null)unknown
-	CUDA_BASE_DIR := $(dir $(shell cygpath -m -s "$(CUDA_PATH)\\.." 2>/dev/null))
-	CUDA_11:=$(shell ls -d $(CUDA_BASE_DIR)/v11.? 2>/dev/null)
-	CUDA_12:=$(shell ls -d $(CUDA_BASE_DIR)/v12.? 2>/dev/null)
-	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
-else ifeq ($(OS),linux)
-	HIP_PATH?=/opt/rocm
-	HIP_LIB_DIR := $(shell ls -d $(HIP_PATH)/lib 2>/dev/null)
-	CUDA_PATH?=/usr/local/cuda
-	CUDA_11:=$(shell ls -d $(CUDA_PATH)-11 2>/dev/null)
-	CUDA_12:=$(shell ls -d $(CUDA_PATH)-12 2>/dev/null)
-endif
-
-ifeq ($(OLLAMA_SKIP_CUDA_GENERATE),)
-ifneq ($(CUDA_11),)
-	RUNNER_TARGETS += cuda_v11
-endif
-ifneq ($(CUDA_12),)
-	RUNNER_TARGETS += cuda_v12
-endif
-endif
-ifeq ($(OLLAMA_SKIP_ROCM_GENERATE),)
-ifneq ($(HIP_LIB_DIR),)
-	RUNNER_TARGETS += rocm
-endif
-endif
-
-
-all: clean-payload .WAIT runners
-
-runners: $(RUNNER_TARGETS)
-
-$(RUNNER_TARGETS):
-	$(MAKE) -f make/Makefile.$@
-
-help-sync apply-patches create-patches sync:
-	$(MAKE) -f make/Makefile.sync $@
-
-clean:
-	rm -rf $(BUILD_DIR) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
-	go clean -cache
-
-clean-payload:
-	rm -rf $(addprefix $(RUNNERS_PAYLOAD_DIR)/, $(RUNNER_TARGETS) metal cpu cpu_avx cpu_avx2)
-
-.PHONY: all runners clean clean-payload $(RUNNER_TARGETS) .WAIT
-
-# Handy debugging for make variables
-print-%:
-	@echo '$*=$($*)'

llama/README.md

@@ -1,160 +0,0 @@
-# `llama`
-
-This package integrates the [llama.cpp](https://github.com/ggerganov/llama.cpp) library as a Go package and makes it easy to build it with tags for different CPU and GPU processors.
-
-Supported:
-
-- [x] CPU
-- [x] avx, avx2
-- [x] macOS Metal
-- [x] Windows CUDA
-- [x] Windows ROCm
-- [x] Linux CUDA
-- [x] Linux ROCm
-- [x] Llava
-
-Extra build steps are required for CUDA and ROCm on Windows since `nvcc` and `hipcc` both require using msvc as the host compiler. For these shared libraries are created:
-
-- `ggml_cuda.dll` on Windows or `ggml_cuda.so` on Linux
-- `ggml_hipblas.dll` on Windows or `ggml_hipblas.so` on Linux
-
-> Note: it's important that memory is allocated and freed by the same compiler (e.g. entirely by code compiled with msvc or mingw). Issues from this should be rare, but there are some places where pointers are returned by the CUDA or HIP runtimes and freed elsewhere, causing a a crash. In a future change the same runtime should be used in both cases to avoid crashes.
-
-## Building
-
-```
-go build .
-```
-
-### AVX
-
-```shell
-go build -tags avx .
-```
-
-### AVX2
-
-```shell
-# go doesn't recognize `-mfma` as a valid compiler flag
-# see https://github.com/golang/go/issues/17895
-go env -w "CGO_CFLAGS_ALLOW=-mfma|-mf16c"
-go env -w "CGO_CXXFLAGS_ALLOW=-mfma|-mf16c"
-go build -tags=avx,avx2 .
-```
-
-## Linux
-
-### CUDA
-
-Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
-
-```shell
-make ggml_cuda.so
-go build -tags avx,cuda .
-```
-
-### ROCm
-
-Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
-
-```shell
-make ggml_hipblas.so
-go build -tags avx,rocm .
-```
-
-## Windows
-
-Download [w64devkit](https://github.com/skeeto/w64devkit/releases/latest) for a simple MinGW development environment.
-
-### CUDA
-
-Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive) then build the cuda code:
-
-```shell
-make ggml_cuda.dll
-go build -tags avx,cuda .
-```
-
-### ROCm
-
-Install [ROCm 5.7.1](https://rocm.docs.amd.com/en/docs-5.7.1/).
-
-```shell
-make ggml_hipblas.dll
-go build -tags avx,rocm .
-```
-
-## Building runners
-
-```shell
-# build all runners for this platform
-make -j
-```
-
-## Vendoring
-
-Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model.  While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit.  A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.
-
-If you update the vendoring code, start by running the following command to establish the tracking llama.cpp repo in the `./vendor/` directory.
-
-```
-make apply-patches
-```
-
-### Updating Base Commit
-
-**Pin to new base commit**
-
-To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring.env`
-
-#### Applying patches
-
-When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.
-
-Start by applying the patches.  If any of the patches have conflicts, the `git am` will stop at the first failure.
-
-```
-make apply-patches
-```
-
-If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed.  Save the file(s) and continue the patch series with `git am --continue` .  If any additional patches fail, follow the same pattern until the full patch series is applied.  Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.
-
-```
-make create-patches sync
-```
-
-Build and test Ollama, and make any necessary changes to the Go code based on the new base commit.  Submit your PR to the Ollama repo.
-
-### Generating Patches
-
-When working on new fixes or features that impact vendored code, use the following model.  First get a clean tracking repo with all current patches applied:
-
-```
-make apply-patches
-```
-
-Now edit the upstream native code in the `./vendor/` directory.  You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing.  Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:
-
-```
-make sync
-make -j 8
-go build .
-```
-
-> [!IMPORTANT]
-> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo.  It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).
-
-Iterate until you're ready to submit PRs.  Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with
-
-```
-make create-patches
-```
-
-> [!IMPORTANT]
-> Once you have completed this step, it is safe to run `apply-patches` since your change is preserved in the patches.
-
-In your `./vendor/` directory, create a branch, and cherry-pick the new commit to that branch, then submit a PR upstream to llama.cpp.
-
-Commit the changes in the ollama repo and submit a PR to Ollama, which will include the vendored code update with your change, along with the patches.
-
-After your PR upstream is merged, follow the **Updating Base Commit** instructions above, however first remove your patch before running `apply-patches` since the new base commit contains your change already.

llama/base64.hpp

@@ -1,392 +0,0 @@
-/*
-This is free and unencumbered software released into the public domain.
-
-Anyone is free to copy, modify, publish, use, compile, sell, or
-distribute this software, either in source code form or as a compiled
-binary, for any purpose, commercial or non-commercial, and by any
-means.
-
-In jurisdictions that recognize copyright laws, the author or authors
-of this software dedicate any and all copyright interest in the
-software to the public domain. We make this dedication for the benefit
-of the public at large and to the detriment of our heirs and
-successors. We intend this dedication to be an overt act of
-relinquishment in perpetuity of all present and future rights to this
-software under copyright law.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
-EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
-IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
-OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
-ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
-OTHER DEALINGS IN THE SOFTWARE.
-
-For more information, please refer to <http://unlicense.org>
-*/
-
-#ifndef PUBLIC_DOMAIN_BASE64_HPP_
-#define PUBLIC_DOMAIN_BASE64_HPP_
-
-#include <cstdint>
-#include <iterator>
-#include <stdexcept>
-#include <string>
-
-class base64_error : public std::runtime_error
-{
-public:
-    using std::runtime_error::runtime_error;
-};
-
-class base64
-{
-public:
-    enum class alphabet
-    {
-        /** the alphabet is detected automatically */
-        auto_,
-        /** the standard base64 alphabet is used */
-        standard,
-        /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/
-        url_filename_safe
-    };
-
-    enum class decoding_behavior
-    {
-        /** if the input is not padded, the remaining bits are ignored */
-        moderate,
-        /** if a padding character is encounter decoding is finished */
-        loose
-    };
-
-    /**
-     Encodes all the elements from `in_begin` to `in_end` to `out`.
-
-     @warning The source and destination cannot overlap. The destination must be able to hold at least
-     `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator.
-
-     @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than
-     8 bits
-     @tparam Output_iterator the destination; the elements written to it are from the type `char`
-     @param in_begin the beginning of the source
-     @param in_end the ending of the source
-     @param out the destination iterator
-     @param alphabet which alphabet should be used
-     @returns the iterator to the next element past the last element copied
-     @throws see `Input_iterator` and `Output_iterator`
-    */
-    template<typename Input_iterator, typename Output_iterator>
-    static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
-                                  alphabet alphabet = alphabet::standard)
-    {
-        constexpr auto pad = '=';
-        const char* alpha  = alphabet == alphabet::url_filename_safe
-                                ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
-                                : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
-
-        while (in_begin != in_end) {
-            std::uint8_t i0 = 0, i1 = 0, i2 = 0;
-
-            // first character
-            i0 = static_cast<std::uint8_t>(*in_begin);
-            ++in_begin;
-
-            *out = alpha[i0 >> 2 & 0x3f];
-            ++out;
-
-            // part of first character and second
-            if (in_begin != in_end) {
-                i1 = static_cast<std::uint8_t>(*in_begin);
-                ++in_begin;
-
-                *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)];
-                ++out;
-            } else {
-                *out = alpha[(i0 & 0x3) << 4];
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                break;
-            }
-
-            // part of second character and third
-            if (in_begin != in_end) {
-                i2 = static_cast<std::uint8_t>(*in_begin);
-                ++in_begin;
-
-                *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)];
-                ++out;
-            } else {
-                *out = alpha[(i1 & 0xf) << 2];
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                break;
-            }
-
-            // rest of third
-            *out = alpha[i2 & 0x3f];
-            ++out;
-        }
-
-        return out;
-    }
-    /**
-     Encodes a string.
-
-     @param str the string that should be encoded
-     @param alphabet which alphabet should be used
-     @returns the encoded base64 string
-     @throws see base64::encode()
-    */
-    static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard)
-    {
-        std::string result;
-
-        result.reserve(required_encode_size(str.length()) + 1);
-
-        encode(str.begin(), str.end(), std::back_inserter(result), alphabet);
-
-        return result;
-    }
-    /**
-     Encodes a char array.
-
-     @param buffer the char array
-     @param size the size of the array
-     @param alphabet which alphabet should be used
-     @returns the encoded string
-    */
-    static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard)
-    {
-        std::string result;
-
-        result.reserve(required_encode_size(size) + 1);
-
-        encode(buffer, buffer + size, std::back_inserter(result), alphabet);
-
-        return result;
-    }
-    /**
-     Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`,
-     in other words: inplace decoding is possible.
-
-     @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`,
-     otherwise the behavior depends on the output iterator.
-
-     @tparam Input_iterator the source; the returned elements are cast to `char`
-     @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t`
-     @param in_begin the beginning of the source
-     @param in_end the ending of the source
-     @param out the destination iterator
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the iterator to the next element past the last element copied
-     @throws base64_error depending on the set behavior
-     @throws see `Input_iterator` and `Output_iterator`
-    */
-    template<typename Input_iterator, typename Output_iterator>
-    static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
-                                  alphabet alphabet          = alphabet::auto_,
-                                  decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        //constexpr auto pad = '=';
-        std::uint8_t last  = 0;
-        auto bits          = 0;
-
-        while (in_begin != in_end) {
-            auto c = *in_begin;
-            ++in_begin;
-
-            if (c == '=') {
-                break;
-            }
-
-            auto part = _base64_value(alphabet, c);
-
-            // enough bits for one byte
-            if (bits + 6 >= 8) {
-                *out = (last << (8 - bits)) | (part >> (bits - 2));
-                ++out;
-
-                bits -= 2;
-            } else {
-                bits += 6;
-            }
-
-            last = part;
-        }
-
-        // check padding
-        if (behavior != decoding_behavior::loose) {
-            while (in_begin != in_end) {
-                auto c = *in_begin;
-                ++in_begin;
-
-                if (c != '=') {
-                    throw base64_error("invalid base64 character.");
-                }
-            }
-        }
-
-        return out;
-    }
-    /**
-     Decodes a string.
-
-     @param str the base64 encoded string
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the decoded string
-     @throws see base64::decode()
-    */
-    static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_,
-                              decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        std::string result;
-
-        result.reserve(max_decode_size(str.length()));
-
-        decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior);
-
-        return result;
-    }
-    /**
-     Decodes a string.
-
-     @param buffer the base64 encoded buffer
-     @param size the size of the buffer
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the decoded string
-     @throws see base64::decode()
-    */
-    static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_,
-                              decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        std::string result;
-
-        result.reserve(max_decode_size(size));
-
-        decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior);
-
-        return result;
-    }
-    /**
-     Decodes a string inplace.
-
-     @param[in,out] str the base64 encoded string
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @throws base64::decode_inplace()
-    */
-    static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_,
-                               decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin());
-    }
-    /**
-     Decodes a char array inplace.
-
-     @param[in,out] str the string array
-     @param size the length of the array
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the pointer to the next element past the last element decoded
-     @throws base64::decode_inplace()
-    */
-    static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_,
-                                decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        return decode(str, str + size, str, alphabet, behavior);
-    }
-    /**
-     Returns the required decoding size for a given size. The value is calculated with the following formula:
-
-     $$
-     \lceil \frac{size}{4} \rceil \cdot 3
-     $$
-
-     @param size the size of the encoded input
-     @returns the size of the resulting decoded buffer; this the absolute maximum
-    */
-    static std::size_t max_decode_size(std::size_t size) noexcept
-    {
-        return (size / 4 + (size % 4 ? 1 : 0)) * 3;
-    }
-    /**
-     Returns the required encoding size for a given size. The value is calculated with the following formula:
-
-     $$
-     \lceil \frac{size}{3} \rceil \cdot 4
-     $$
-
-     @param size the size of the decoded input
-     @returns the size of the resulting encoded buffer
-    */
-    static std::size_t required_encode_size(std::size_t size) noexcept
-    {
-        return (size / 3 + (size % 3 ? 1 : 0)) * 4;
-    }
-
-private:
-    static std::uint8_t _base64_value(alphabet& alphabet, char c)
-    {
-        if (c >= 'A' && c <= 'Z') {
-            return c - 'A';
-        } else if (c >= 'a' && c <= 'z') {
-            return c - 'a' + 26;
-        } else if (c >= '0' && c <= '9') {
-            return c - '0' + 52;
-        }
-
-        // comes down to alphabet
-        if (alphabet == alphabet::standard) {
-            if (c == '+') {
-                return 62;
-            } else if (c == '/') {
-                return 63;
-            }
-        } else if (alphabet == alphabet::url_filename_safe) {
-            if (c == '-') {
-                return 62;
-            } else if (c == '_') {
-                return 63;
-            }
-        } // auto detect
-        else {
-            if (c == '+') {
-                alphabet = alphabet::standard;
-
-                return 62;
-            } else if (c == '/') {
-                alphabet = alphabet::standard;
-
-                return 63;
-            } else if (c == '-') {
-                alphabet = alphabet::url_filename_safe;
-
-                return 62;
-            } else if (c == '_') {
-                alphabet = alphabet::url_filename_safe;
-
-                return 63;
-            }
-        }
-
-        throw base64_error("invalid base64 character.");
-    }
-};
-
-#endif // !PUBLIC_DOMAIN_BASE64_HPP_

llama/build-info.cpp

@@ -1,4 +0,0 @@
-int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "3f1ae2e32cde00c39b96be6d01c2997c29bae555";
-char const *LLAMA_COMPILER = "";
-char const *LLAMA_BUILD_TARGET = "";

llama/clip.h

@@ -1,120 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#ifndef CLIP_H
-#define CLIP_H
-
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define CLIP_API __declspec(dllexport)
-#        else
-#            define CLIP_API __declspec(dllimport)
-#        endif
-#    else
-#        define CLIP_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define CLIP_API
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct clip_ctx;
-
-struct clip_image_size {
-    int width;
-    int height;
-};
-
-struct clip_image_u8_batch {
-    struct clip_image_u8 * data;
-    size_t size;
-};
-
-struct clip_image_f32_batch {
-    struct clip_image_f32 * data;
-    size_t size;
-};
-
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
-
-CLIP_API void clip_free(struct clip_ctx * ctx);
-
-CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-
-CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
-
-// TODO: should be enum, not string
-CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
-
-CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
-
-CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
-CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
-
-CLIP_API struct clip_image_size * clip_image_size_init();
-CLIP_API struct clip_image_u8  * clip_image_u8_init ();
-CLIP_API struct clip_image_f32 * clip_image_f32_init();
-
-CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
-CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
-CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
-CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
-
-CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
-
-/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
-CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
-
-/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
-
-CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
-
-CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
-CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
-
-CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
-
-CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // CLIP_H

llama/common.h

@@ -1,581 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-// Various helper functions and utilities
-
-#pragma once
-
-#include "llama.h"
-
-#include <string>
-#include <vector>
-#include <sstream>
-
-#ifdef _WIN32
-#define DIRECTORY_SEPARATOR '\\'
-#else
-#define DIRECTORY_SEPARATOR '/'
-#endif // _WIN32
-
-#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
-#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
-
-#define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
-} while(0)
-
-#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
-
-struct llama_lora_adapter_info {
-    std::string path;
-    float scale;
-};
-
-struct llama_lora_adapter_container : llama_lora_adapter_info {
-    struct llama_lora_adapter * adapter;
-};
-
-// build info
-extern int LLAMA_BUILD_NUMBER;
-extern char const * LLAMA_COMMIT;
-extern char const * LLAMA_COMPILER;
-extern char const * LLAMA_BUILD_TARGET;
-
-struct llama_control_vector_load_info;
-
-//
-// CPU utils
-//
-
-struct cpu_params {
-    int      n_threads                   = -1;
-    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
-    bool     mask_valid                  = false;   // Default: any CPU
-    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
-    bool     strict_cpu                  = false;   // Use strict CPU placement
-    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
-};
-
-int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
-
-//
-// Common params
-//
-
-enum llama_example {
-    LLAMA_EXAMPLE_COMMON,
-    LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_MAIN,
-    LLAMA_EXAMPLE_INFILL,
-    LLAMA_EXAMPLE_EMBEDDING,
-    LLAMA_EXAMPLE_PERPLEXITY,
-    LLAMA_EXAMPLE_RETRIEVAL,
-    LLAMA_EXAMPLE_PASSKEY,
-    LLAMA_EXAMPLE_IMATRIX,
-    LLAMA_EXAMPLE_BENCH,
-    LLAMA_EXAMPLE_SERVER,
-    LLAMA_EXAMPLE_CVECTOR_GENERATOR,
-    LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
-    LLAMA_EXAMPLE_LOOKUP,
-    LLAMA_EXAMPLE_PARALLEL,
-
-    LLAMA_EXAMPLE_COUNT,
-};
-
-enum gpt_sampler_type {
-    GPT_SAMPLER_TYPE_NONE        = 0,
-    GPT_SAMPLER_TYPE_TOP_K       = 1,
-    GPT_SAMPLER_TYPE_TOP_P       = 2,
-    GPT_SAMPLER_TYPE_MIN_P       = 3,
-    GPT_SAMPLER_TYPE_TFS_Z       = 4,
-    GPT_SAMPLER_TYPE_TYPICAL_P   = 5,
-    GPT_SAMPLER_TYPE_TEMPERATURE = 6,
-};
-
-// dimensionality reduction methods, used by cvector-generator
-enum dimre_method {
-    DIMRE_METHOD_PCA,
-    DIMRE_METHOD_MEAN,
-};
-
-// sampler parameters
-struct gpt_sampler_params {
-    uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
-
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
-    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range    = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.00f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = false; // consider newlines as a repeatable token
-    bool    ignore_eos        = false;
-    bool    no_perf           = false; // disable performance metrics
-
-    std::vector<enum gpt_sampler_type> samplers = {
-        GPT_SAMPLER_TYPE_TOP_K,
-        GPT_SAMPLER_TYPE_TFS_Z,
-        GPT_SAMPLER_TYPE_TYPICAL_P,
-        GPT_SAMPLER_TYPE_TOP_P,
-        GPT_SAMPLER_TYPE_MIN_P,
-        GPT_SAMPLER_TYPE_TEMPERATURE
-    };
-
-    std::string grammar; // optional BNF-like grammar to constrain sampling
-
-    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
-
-    // print the parameters into a string
-    std::string print() const;
-};
-
-struct gpt_params {
-    int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
-    int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                =     0; // number of tokens to keep from initial prompt
-    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
-    int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel            =     1; // number of parallel sequences to decode
-    int32_t n_sequences           =     1; // number of sequences to decode
-    float   p_split               =  0.1f; // speculative decoding split probability
-    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
-    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
-    int32_t grp_attn_n            =     1; // group-attention factor
-    int32_t grp_attn_w            =   512; // group-attention width
-    int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base        =  0.0f; // RoPE base frequency
-    float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
-    float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
-    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
-    int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
-
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
-    struct cpu_params draft_cpuparams;
-    struct cpu_params draft_cpuparams_batch;
-
-    ggml_backend_sched_eval_callback cb_eval = nullptr;
-    void * cb_eval_user_data                 = nullptr;
-
-    ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
-
-    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
-    enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-    enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
-    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
-
-    struct gpt_sampler_params sparams;
-
-    std::string model                = ""; // model path                                                    // NOLINT
-    std::string model_draft          = ""; // draft model for speculative decoding                          // NOLINT
-    std::string model_alias          = "unknown"; // model alias                                            // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
-    std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
-    std::string prompt               = "";                                                                  // NOLINT
-    std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
-    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
-    std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
-    std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
-    std::string logdir               = ""; // directory in which to save YAML log files                     // NOLINT
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
-    std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
-    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // NOLINT
-
-    std::vector<std::string> in_files;   // all input files
-    std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
-    std::vector<llama_model_kv_override> kv_overrides;
-
-    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
-    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
-
-    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
-
-    int32_t verbosity                  = 0;
-    int32_t control_vector_layer_start = -1; // layer range for control vector
-    int32_t control_vector_layer_end   = -1; // layer range for control vector
-
-    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
-    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
-                                     //                                       (which is more convenient to use for plotting)
-                                     //
-    bool   hellaswag        = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
-    size_t hellaswag_tasks  = 400;   // number of tasks to use when computing the HellaSwag score
-
-    bool   winogrande       = false; // compute Winogrande score over random tasks from datafile supplied in prompt
-    size_t winogrande_tasks = 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
-
-    bool   multiple_choice  = false;  // compute TruthfulQA score over random tasks from datafile supplied in prompt
-    size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
-
-    bool   kl_divergence    = false; // compute KL divergence
-
-    bool usage             = false; // print usage
-    bool use_color         = false; // use color to distinguish generations and inputs
-    bool special           = false; // enable special token output
-    bool interactive       = false; // interactive mode
-    bool interactive_first = false; // wait for user input immediately
-    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
-    bool prompt_cache_all  = false; // save user input and generations to prompt cache
-    bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
-
-    bool escape            = true;  // escape "\n", "\r", "\t", "\'", "\"", and "\\"
-    bool multiline_input   = false; // reverse the usage of `\`
-    bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
-    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
-    bool flash_attn        = false; // flash attention
-    bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = true;  // context shift on inifinite text generation
-
-    bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool logits_all        = false; // return logits for all tokens in the batch
-    bool use_mmap          = true;  // use mmap for faster loads
-    bool use_mlock         = false; // use mlock to keep model in memory
-    bool verbose_prompt    = false; // print prompt tokens before generation
-    bool display_prompt    = true;  // print prompt before generation
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
-    bool no_kv_offload     = false; // disable KV offloading
-    bool warmup            = true;  // warmup run
-    bool check_tensors     = false; // validate tensor data
-
-    std::string cache_type_k = "f16"; // KV cache data type for the K
-    std::string cache_type_v = "f16"; // KV cache data type for the V
-
-    // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
-    std::vector<std::string> image; // path to image file(s)
-
-    // embedding
-    bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
-    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-    std::string embd_sep   = "\n";  // separator of embendings
-    bool reranking         = false; // enable reranking support on server
-
-    // server params
-    int32_t port           = 8080;         // server listens on this network port
-    int32_t timeout_read   = 600;          // http read timeout in seconds
-    int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-
-    std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";                                                                         // NOLINT
-    std::string chat_template = "";                                                                         // NOLINT
-    std::string system_prompt = "";                                                                         // NOLINT
-    bool enable_chat_template = true;
-
-    std::vector<std::string> api_keys;
-
-    std::string ssl_file_key  = "";                                                                         // NOLINT
-    std::string ssl_file_cert = "";                                                                         // NOLINT
-
-    bool endpoint_slots   = true;
-    bool endpoint_metrics = false;
-
-    bool log_json = false;
-
-    std::string slot_save_path;
-
-    float slot_prompt_similarity = 0.5f;
-
-    // batched-bench params
-    bool is_pp_shared = false;
-
-    std::vector<int32_t> n_pp;
-    std::vector<int32_t> n_tg;
-    std::vector<int32_t> n_pl;
-
-    // retrieval params
-    std::vector<std::string> context_files; // context files to embed
-
-    int32_t chunk_size = 64; // chunk size for context embedding
-
-    std::string chunk_separator = "\n"; // chunk separator for context embedding
-
-    // passkey params
-    int32_t n_junk = 250; // number of times to repeat the junk text
-    int32_t i_pos  = -1;  // position of the passkey in the junk text
-
-    // imatrix params
-    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
-
-    int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
-    int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
-    int32_t i_chunk     =  0; // start processing from this chunk
-
-    bool process_output = false; // collect data for the output tensor
-    bool compute_ppl    = true;  // whether to compute perplexity
-
-    // cvector-generator params
-    int n_pca_batch = 100;
-    int n_pca_iterations = 1000;
-    dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_outfile       = "control_vector.gguf";
-    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
-
-    bool spm_infill = false; // suffix/prefix/middle pattern for infill
-
-    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
-
-    // batched-bench params
-    bool batched_bench_output_jsonl = false;
-};
-
-// call once at the start of a program if it uses libcommon
-// initializes the logging system and prints info about the build
-void gpt_init();
-
-std::string gpt_params_get_system_info(const gpt_params & params);
-
-bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
-bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
-void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
-bool set_process_priority(enum ggml_sched_priority prio);
-
-//
-// String utils
-//
-
-std::vector<std::string> string_split(std::string input, char separator);
-
-std::string string_strip(const std::string & str);
-std::string string_get_sortable_timestamp();
-
-void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
-
-template<class T>
-static std::vector<T> string_split(const std::string & str, char delim) {
-    std::vector<T> values;
-    std::istringstream str_stream(str);
-    std::string token;
-    while (std::getline(str_stream, token, delim)) {
-        T value;
-        std::istringstream token_stream(token);
-        token_stream >> value;
-        values.push_back(value);
-    }
-    return values;
-}
-
-bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
-void string_process_escapes(std::string & input);
-
-std::string string_from(bool value);
-std::string string_from(const std::vector<int> & values);
-std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
-std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
-
-//
-// Filesystem utils
-//
-
-bool fs_validate_filename(const std::string & filename);
-bool fs_create_directory_with_parents(const std::string & path);
-
-std::string fs_get_cache_directory();
-std::string fs_get_cache_file(const std::string & filename);
-
-//
-// Model utils
-//
-
-struct llama_init_result {
-    struct llama_model   * model   = nullptr;
-    struct llama_context * context = nullptr;
-    std::vector<llama_lora_adapter_container> lora_adapters;
-};
-
-struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
-
-struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
-struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
-
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
-
-// clear LoRA adapters from context, then apply new list of adapters
-void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
-
-// Batch utils
-
-void llama_batch_clear(struct llama_batch & batch);
-
-void llama_batch_add(
-                 struct llama_batch & batch,
-                        llama_token   id,
-                          llama_pos   pos,
-    const std::vector<llama_seq_id> & seq_ids,
-                               bool   logits);
-
-//
-// Vocab utils
-//
-
-// tokenizes a string into a vector of tokens
-// should work similar to Python's `tokenizer.encode`
-std::vector<llama_token> llama_tokenize(
-  const struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_special,
-                        bool   parse_special = false);
-
-std::vector<llama_token> llama_tokenize(
-    const struct llama_model * model,
-           const std::string & text,
-                        bool   add_special,
-                        bool   parse_special = false);
-
-// tokenizes a token into a piece, optionally renders special/control tokens
-// should work similar to Python's `tokenizer.id_to_piece`
-std::string llama_token_to_piece(
-        const struct llama_context * ctx,
-                       llama_token   token,
-                       bool          special = true);
-
-// detokenizes a vector of tokens into a string
-// should work similar to Python's `tokenizer.decode`
-// optionally renders special/control tokens
-std::string llama_detokenize(
-                         llama_context * ctx,
-        const std::vector<llama_token> & tokens,
-                                  bool   special = true);
-
-//
-// Chat template utils
-//
-
-// same with llama_chat_message, but uses std::string
-struct llama_chat_msg {
-    std::string role;
-    std::string content;
-};
-
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool llama_chat_verify_template(const std::string & tmpl);
-
-// CPP wrapper for llama_chat_apply_template
-// If the built-in template is not supported, we default to chatml
-// If the custom "tmpl" is not supported, we throw an error
-std::string llama_chat_apply_template(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<llama_chat_msg> & chat,
-        bool add_ass);
-
-// Format single message, while taking into account the position of that message in chat history
-std::string llama_chat_format_single(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<llama_chat_msg> & past_msg,
-        const llama_chat_msg & new_msg,
-        bool add_ass);
-
-// Returns an example of formatted chat
-std::string llama_chat_format_example(const struct llama_model * model,
-        const std::string & tmpl);
-
-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// Dump the KV cache view showing individual sequences in each cell (long output).
-void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
-//
-// Embedding utils
-//
-
-void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
-
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
-
-//
-// Control vector utils
-//
-
-struct llama_control_vector_data {
-    int n_embd;
-
-    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
-    std::vector<float> data;
-};
-
-struct llama_control_vector_load_info {
-    float strength;
-
-    std::string fname;
-};
-
-// Load control vectors, scale each by strength, and add them together.
-// On error, returns {-1, empty}
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
-
-//
-// Split utils
-//
-
-static const char * const LLM_KV_SPLIT_NO            = "split.no";
-static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
-static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
-
-//
-// YAML utils
-//
-
-void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
-void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
-void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
-
-void yaml_dump_non_result_info(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

llama/ggml-aarch64.h

@@ -1,65 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-#ifdef __cplusplus
-}
-#endif
-

llama/ggml-alloc.h

@@ -1,102 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-typedef struct      ggml_backend_buffer * ggml_backend_buffer_t;
-typedef struct             ggml_backend * ggml_backend_t;
-
-// Tensor allocator
-struct ggml_tallocr {
-    ggml_backend_buffer_t buffer;
-    void * base;
-    size_t alignment;
-    size_t offset;
-};
-
-GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
-GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
-
-// Graph allocator
-/*
-  Example usage:
-    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
-
-    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
-    ggml_gallocr_reserve(galloc, build_graph(max_batch));
-
-    // allocate the graph
-    struct ggml_cgraph * graph = build_graph(batch);
-    ggml_gallocr_alloc_graph(galloc, graph);
-
-    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
-
-    // evaluate the graph
-    ggml_backend_graph_compute(backend, graph);
-*/
-
-// special tensor flags for use with the graph allocator:
-//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
-//   ggml_set_output(): output tensors are never freed and never overwritten
-
-typedef struct ggml_gallocr * ggml_gallocr_t;
-
-GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
-GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
-GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
-
-// pre-allocate buffers from a measure graph - does not allocate or modify the graph
-// call with a worst-case graph to avoid buffer reallocations
-// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
-// returns false if the buffer allocation failed
-GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-GGML_API bool ggml_gallocr_reserve_n(
-    ggml_gallocr_t galloc,
-    struct ggml_cgraph * graph,
-    const int * node_buffer_ids,
-    const int * leaf_buffer_ids);
-
-// automatic reallocation if the topology changes when using a single buffer
-// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
-GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-
-GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
-
-// Utils
-// Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-backend-impl.h

@@ -1,180 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        // allocate a buffer of this type
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        // tensor alignment
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft);
-        // max buffer size that can be allocated
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft);
-        // data size needed to allocate the tensor, including padding
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // check if tensor data is in host memory
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)      (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)   (ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)      (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL memset_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
-        void         (*GGML_CALL set_tensor)    (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor)    (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor)    (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)         (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)         (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-    };
-
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
-        ggml_backend_buffer_context_t context;
-        size_t size;
-        enum ggml_backend_buffer_usage usage;
-    };
-
-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
-
-        void (*GGML_CALL free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
-
-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
-
-        // compute graph with a plan (not used currently)
-        // create a new plan for a graph
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
-        void                      (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
-        // compute the graph with the plan
-        enum ggml_status          (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend can compute an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend can use tensors allocated in a buffer type
-        bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        // create a new event that can record events on this backend instance
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        // record an event on the backend instance that created it
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        // wait for an event on on a different backend instance
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        // block until an event is recorded
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
-    };
-
-    struct ggml_backend {
-        ggml_guid_t guid;
-
-        struct ggml_backend_i iface;
-        ggml_backend_context_t context;
-    };
-
-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-backend.h

@@ -1,267 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
-
-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-        GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
-    };
-
-    GGML_API           const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
-    GGML_API           ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
-
-    //
-    // Backend
-    //
-
-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    // "offset" refers to the offset of the tensor data for setting/getting data
-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event);
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
-
-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name); // returns index of backend with name, or SIZE_MAX if not found
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is backend_name:params (params is optional)
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
-
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
-
-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
-
-        // manually assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
-
-        ggml_backend_sched_reserve(sched, reserve_graph);
-
-        // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
-
-    // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
-
-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
-
-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
-
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-blas.cpp

@@ -1,398 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#ifdef GGML_USE_BLAS
-
-#include "ggml-impl.h"
-#include "ggml-blas.h"
-#include "ggml-backend-impl.h"
-
-#include <future>
-#include <vector>
-
-#if defined(GGML_USE_ACCELERATE)
-#   include <Accelerate/Accelerate.h>
-#elif defined(GGML_BLAS_USE_MKL)
-#   include <mkl.h>
-#elif defined(GGML_BLAS_USE_BLIS)
-#   include <blis.h>
-#elif defined(GGML_BLAS_USE_NVPL)
-#   include <nvpl_blas.h>
-#else
-#   include <cblas.h>
-#endif
-
-struct ggml_backend_blas_context {
-    int n_threads = GGML_DEFAULT_N_THREADS;
-    std::unique_ptr<char[]> work_data;
-    size_t work_size = 0;
-#ifndef GGML_USE_OPENMP
-    std::vector<std::future<void>> tasks;
-#endif
-};
-
-// helper function to determine if it is better to use BLAS or not
-// for large matrices, BLAS is faster
-static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    const int64_t ne10 = src1->ne[0];
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-
-    // TODO: find the optimal values for these
-    if (ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) &&
-        src1->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
-
-        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
-        return true;
-    }
-
-    return false;
-}
-
-static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const enum ggml_type type = src0->type;
-
-    GGML_ASSERT(ne0 == ne01);
-    GGML_ASSERT(ne1 == ne11);
-    GGML_ASSERT(ne2 == ne12);
-    GGML_ASSERT(ne3 == ne13);
-
-    // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == ggml_type_size(type));
-    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-    const int64_t ne_plane      = ne01*ne00;
-    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
-
-    if (ctx->work_size < desired_wsize) {
-        ctx->work_data.reset(new char[desired_wsize]);
-        ctx->work_size = desired_wsize;
-    }
-    void * wdata = ctx->work_data.get();
-
-    // convert src0 to float
-    if (type != GGML_TYPE_F32) {
-        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
-        ggml_to_float_t const to_float = type_traits.to_float;
-
-        for (int64_t i03 = 0; i03 < ne03; i03++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
-                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
-
-                const int min_cols_per_thread = 4096;
-                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
-                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
-
-#ifdef GGML_USE_OPENMP
-                #pragma omp parallel for num_threads(n_threads)
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                }
-#else
-                for (int i = 1; i < n_threads; i++) {
-                    const int64_t start =       i*ne01/n_threads;
-                    const int64_t end   = (i + 1)*ne01/n_threads;
-                    if (start < end) {
-                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
-                            for (int64_t i01 = start; i01 < end; i01++) {
-                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                            }
-                        }));
-                    }
-                }
-                {
-                    // reuse the current thread for the first task
-                    const int64_t start = 0;
-                    const int64_t end   = ne01/n_threads;
-                    for (int64_t i01 = start; i01 < end; i01++) {
-                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
-                    }
-                }
-#endif
-            }
-        }
-
-#ifndef GGML_USE_OPENMP
-        // wait for all tasks to finish
-        for (auto & task : ctx->tasks) {
-            task.get();
-        }
-        ctx->tasks.clear();
-#endif
-    }
-
-#if defined(OPENBLAS_VERSION)
-    openblas_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_BLIS)
-    bli_thread_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_NVPL)
-    nvpl_blas_set_num_threads(ctx->n_threads);
-#endif
-
-    for (int64_t i13 = 0; i13 < ne13; i13++) {
-        for (int64_t i12 = 0; i12 < ne12; i12++) {
-            const int64_t i03 = i13/r3;
-            const int64_t i02 = i12/r2;
-
-            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
-            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
-
-            if (type != GGML_TYPE_F32) {
-                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
-            }
-
-            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne1, ne01, ne10,
-                        1.0f,   y, ne10,
-                                x, ne00,
-                        0.0f,   d, ne01);
-        }
-    }
-}
-
-static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne0  == ne00);
-    GGML_ASSERT(ne1  == ne10);
-    GGML_ASSERT(ne2  == ne02);
-    GGML_ASSERT(ne02 == ne12);
-    GGML_ASSERT(ne3  == ne13);
-    GGML_ASSERT(ne03 == ne13);
-
-    // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == sizeof(float));
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    // GGML_ASSERT(nb0 <= nb1);
-    // GGML_ASSERT(nb1 <= nb2);
-    // GGML_ASSERT(nb2 <= nb3);
-
-    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
-    // src0: (k,n)
-    // src1: (k,m)
-    // dst:  (m,n)
-    //
-    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
-    // Also expressed as (major,minor)
-    // a: (m,k): so src1 transposed
-    // b: (k,n): so src0
-    // c: (m,n)
-    //
-    // However, if ggml_is_transposed(src1) is true, then
-    // src1->data already contains a transposed version, so sgemm mustn't
-    // transpose it further.
-
-    int n = src0->ne[0];
-    int k = src0->ne[1];
-    int m = src1->ne[0];
-
-    CBLAS_TRANSPOSE transposeA;
-    int lda;
-
-    if (!ggml_is_transposed(src1)) {
-        transposeA = CblasTrans;
-        lda = m;
-    } else {
-        transposeA = CblasNoTrans;
-        lda = k;
-    }
-
-    float * a = (float *) ((char *) src1->data);
-    float * b = (float *) ((char *) src0->data);
-    float * c = (float *) ((char *) dst->data);
-
-    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
-
-    GGML_UNUSED(ctx);
-}
-
-// backend interface
-
-GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
-    return "BLAS";
-
-    GGML_UNUSED(backend);
-}
-
-GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
-    delete ctx;
-    delete backend;
-}
-
-GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(backend);
-}
-
-GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
-
-        switch (node->op) {
-            case GGML_OP_MUL_MAT:
-                ggml_backend_blas_mul_mat(ctx, node);
-                break;
-
-            case GGML_OP_OUT_PROD:
-                ggml_backend_blas_out_prod(ctx, node);
-                break;
-
-            case GGML_OP_NONE:
-            case GGML_OP_RESHAPE:
-            case GGML_OP_VIEW:
-            case GGML_OP_PERMUTE:
-            case GGML_OP_TRANSPOSE:
-                break;
-
-            default:
-                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
-        }
-    }
-
-    return GGML_STATUS_SUCCESS;
-
-    GGML_UNUSED(backend);
-}
-
-GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    const struct ggml_tensor * src0 = op->src[0];
-    const struct ggml_tensor * src1 = op->src[1];
-
-    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
-           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
-                                          op->src[1]->type == GGML_TYPE_F32 &&
-                                          ggml_is_matrix(src0) &&
-                                          ggml_is_matrix(src1) &&
-                                          ggml_is_contiguous(src0) &&
-                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
-
-    GGML_UNUSED(backend);
-}
-
-GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i blas_backend_i = {
-    /* .get_name                = */ ggml_backend_blas_name,
-    /* .free                    = */ ggml_backend_blas_free,
-    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
-    /* .supports_op             = */ ggml_backend_blas_supports_op,
-    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
-    /* .offload_op              = */ NULL,
-    /* .event_new               = */ NULL,
-    /* .event_free              = */ NULL,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-    /* .event_synchronize       = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_blas_guid(void) {
-    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_blas_init(void) {
-    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
-
-    ggml_backend_t backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_blas_guid(),
-        /* .interface = */ blas_backend_i,
-        /* .context   = */ ctx,
-    };
-
-#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
-    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
-        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
-    }
-#endif
-
-#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
-    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
-#endif
-
-    return backend;
-}
-
-GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
-}
-
-void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
-
-    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
-    ctx->n_threads = n_threads;
-}
-
-#endif

llama/ggml-blas.h

@@ -1,49 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void);
-
-GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend);
-
-// number of threads used for conversion to float
-// for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
-
-
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-cpu-impl.h

@@ -1,640 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-// GGML CPU internal header
-
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-//#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-#ifdef __cplusplus
-}
-#endif

llama/ggml-cuda.h

@@ -1,75 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#elif defined(GGML_USE_MUSA)
-#define GGML_CUDA_NAME "MUSA"
-#define GGML_CUBLAS_NAME "muBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int ggml_backend_cuda_reg_devices();
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
-GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
-#ifdef  __cplusplus
-}
-#endif

llama/ggml-cuda/acc.cu

@@ -1,73 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "acc.cuh"
-
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
-    if (i >= ne) {
-        return;
-    }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
-    }
-}
-
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
-}
-
-void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
-
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
-
-    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
-}

llama/ggml-cuda/acc.cuh

@@ -1,31 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-#define CUDA_ACC_BLOCK_SIZE 256
-
-void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/arange.cu

@@ -1,60 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "arange.cuh"
-
-static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
-    // blockIDx.x: idx of ne0 / BLOCK_SIZE
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    dst[nidx] = start + step * nidx;
-}
-
-static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
-    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
-}
-
-void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    float start;
-    float stop;
-    float step;
-    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
-    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
-    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
-
-    int64_t steps = (int64_t)ceil((stop - start) / step);
-    GGML_ASSERT(ggml_nelements(dst) == steps);
-
-    arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
-}

llama/ggml-cuda/arange.cuh

@@ -1,31 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-#define CUDA_ARANGE_BLOCK_SIZE 256
-
-void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/argsort.cu

@@ -1,130 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "argsort.cuh"
-
-template<typename T>
-static inline __device__ void ggml_cuda_swap(T & a, T & b) {
-    T tmp = a;
-    a = b;
-    b = tmp;
-}
-
-template<ggml_sort_order order>
-static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
-    // bitonic sort
-    int col = threadIdx.x;
-    int row = blockIdx.y;
-
-    if (col >= ncols_pad) {
-        return;
-    }
-
-    const float * x_row = x + row * ncols;
-    extern __shared__ int dst_row[];
-
-    // initialize indices
-    dst_row[col] = col;
-
-    __syncthreads();
-
-    for (int k = 2; k <= ncols_pad; k *= 2) {
-        for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= ncols ||
-                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (dst_row[ixj] >= ncols ||
-                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
-                    }
-                }
-            }
-            __syncthreads();
-        }
-    }
-
-    // copy the result to dst without the padding
-    if (col < ncols) {
-        dst[row * ncols + col] = dst_row[col];
-    }
-}
-
-static int next_power_of_2(int x) {
-    int n = 1;
-    while (n < x) {
-        n *= 2;
-    }
-    return n;
-}
-
-static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
-    // bitonic sort requires ncols to be power of 2
-    const int ncols_pad = next_power_of_2(ncols);
-
-    const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
-    const size_t shared_mem = ncols_pad * sizeof(int);
-
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
-
-    if (order == GGML_SORT_ORDER_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-    } else if (order == GGML_SORT_ORDER_DESC) {
-        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
-    } else {
-        GGML_ABORT("fatal error");
-    }
-}
-
-void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_I32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-
-    const int64_t ncols = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
-
-    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
-}

llama/ggml-cuda/argsort.cuh

@@ -1,29 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/binbcast.cu

@@ -1,381 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "binbcast.cuh"
-#include <cstdint>
-
-static __device__ __forceinline__ float op_repeat(const float a, const float b) {
-    return b;
-    GGML_UNUSED(a);
-}
-
-static __device__ __forceinline__ float op_add(const float a, const float b) {
-    return a + b;
-}
-
-static __device__ __forceinline__ float op_sub(const float a, const float b) {
-    return a - b;
-}
-
-static __device__ __forceinline__ float op_mul(const float a, const float b) {
-    return a * b;
-}
-
-static __device__ __forceinline__ float op_div(const float a, const float b) {
-    return a / b;
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13) {
-    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
-    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
-    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-    }
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13) {
-
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-}
-
-template <typename T>
-static __global__ void k_repeat_back(
-    const T * __restrict__ src, T * __restrict__ dst, const int64_t ne00, const int64_t ne01, const int64_t ne02,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2) {
-
-    const int64_t tid0 = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
-    const int64_t tid1 = (int64_t) blockIdx.y*blockDim.y + threadIdx.y;
-    const int64_t tid2 = (int64_t) blockIdx.z*blockDim.z + threadIdx.z;
-
-    if (tid0 >= ne0) {
-        return;
-    }
-
-    T sum = 0;
-    for (int64_t i2 = tid2; i2 < ne02; i2 += ne2) {
-        for (int64_t i1 = tid1; i1 < ne01; i1 += ne1) {
-            for (int64_t i0 = tid0; i0 < ne00; i0 += ne0) {
-                sum += src[i2*ne01*ne00 + i1*ne00 + i0];
-            }
-        }
-    }
-    dst[tid2*ne1*ne0 + tid1*ne0 + tid0] = sum;
-}
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_cuda {
-    template<typename src0_t, typename src1_t, typename dst_t>
-    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
-            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
-            cudaStream_t stream) {
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        int nr0 = ne10/ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
-                }
-            }
-        }
-
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            //int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
-            //int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
-            //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
-            //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s00 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            dim3 block_dims;
-            block_dims.x = std::min<unsigned int>(hne0, block_size);
-            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
-            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
-
-            dim3 block_nums(
-                (hne0 + block_dims.x - 1) / block_dims.x,
-                (ne1 + block_dims.y - 1) / block_dims.y,
-                (ne2*ne3 + block_dims.z - 1) / block_dims.z
-            );
-
-            if (block_nums.z > 65535) {
-                // this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
-                    /* s10, */ s11, s12, s13);
-            } else {
-                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
-                    /* s10, */ s11, s12, s13);
-            }
-        }
-    }
-};
-
-template <typename T>
-static void repeat_back_cuda(
-    const T * src, T * dst, const int64_t ne00, const int64_t ne01, const int64_t ne02,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, cudaStream_t stream) {
-
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    const dim3 block_nums((ne0 + WARP_SIZE - 1) / WARP_SIZE, ne1, ne2);
-    k_repeat_back<T><<<block_nums, block_dims, 0, stream>>>(src, dst, ne00, ne01, ne02, ne0, ne1, ne2);
-}
-
-template<class op>
-static void ggml_cuda_op_bin_bcast(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ABORT("fatal error");
-    }
-}
-
-void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_sub>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
-}
-
-void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-
-    GGML_ASSERT(src0->type == dst->type);
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(dst));
-    GGML_ASSERT(ggml_can_repeat(dst, src0));
-
-    cudaStream_t stream = ctx.stream();
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    GGML_ASSERT(src0->ne[3] == 1);
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-    const int64_t ne2 = dst->ne[2];
-    GGML_ASSERT(dst->ne[3] == 1);
-
-    switch (dst->type) {
-        case GGML_TYPE_F32: {
-            const float * src0_d = (const float *) src0->data;
-            float       * dst_d  = (float       *) dst->data;
-            repeat_back_cuda<float>(src0_d, dst_d, ne00, ne01, ne02, ne0, ne1, ne2, stream);
-        } break;
-        default: {
-            GGML_ASSERT(false);
-        } break;
-    }
-}

llama/ggml-cuda/binbcast.cuh

@@ -1,35 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-
-void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/clamp.cu

@@ -1,60 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "clamp.cuh"
-
-static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
-}
-
-static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
-    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
-}
-
-
-void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    float min;
-    float max;
-    memcpy(&min, dst->op_params, sizeof(float));
-    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
-
-    clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
-}

llama/ggml-cuda/clamp.cuh

@@ -1,31 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-#define CUDA_CLAMP_BLOCK_SIZE 256
-
-void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/common.cuh

@@ -1,702 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-cuda.h"
-
-#include <cstdint>
-#include <memory>
-
-#if defined(GGML_USE_HIPBLAS)
-#define GGML_COMMON_DECL_HIP
-#define GGML_COMMON_IMPL_HIP
-#else
-#define GGML_COMMON_DECL_CUDA
-#define GGML_COMMON_IMPL_CUDA
-#if defined(GGML_USE_MUSA)
-#define GGML_COMMON_DECL_MUSA
-#define GGML_COMMON_IMPL_MUSA
-#endif
-#endif
-#include "ggml-common.h"
-
-#include <cstdio>
-#include <array>
-#include <cassert>
-#include <cfloat>
-#include <string>
-#include <vector>
-
-#if defined(GGML_USE_HIPBLAS)
-#include "vendors/hip.h"
-#elif defined(GGML_USE_MUSA)
-#include "vendors/musa.h"
-#else
-#include "vendors/cuda.h"
-#endif // defined(GGML_USE_HIPBLAS)
-
-#define STRINGIZE_IMPL(...) #__VA_ARGS__
-#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
-
-#define WARP_SIZE 32
-#define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
-#define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
-
-#define CC_PASCAL     600
-#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define CC_VOLTA      700
-#define CC_TURING     750
-#define CC_AMPERE     800
-#define CC_OFFSET_AMD 1000000
-#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
-#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
-#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
-#define CC_QY1        210
-#define CC_QY2        220
-
-#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-#define GGML_CUDA_MAX_STREAMS 8
-
-[[noreturn]]
-void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
-
-#define CUDA_CHECK_GEN(err, success, error_fn)                                      \
-     do {                                                                           \
-        auto err_ = (err);                                                          \
-        if (err_ != (success)) {                                                    \
-            ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_));    \
-        }                                                                           \
-    } while (0)
-
-#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
-
-#if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
-    static const char * cublas_get_error_str(const cublasStatus_t err) {
-        return cublasGetStatusString(err);
-    }
-#else
-    static const char * cublas_get_error_str(const cublasStatus_t err) {
-        switch (err) {
-            case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
-            case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
-            case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
-            case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
-            case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
-            case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
-            case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
-            case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
-            case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
-            default: return "unknown error";
-        }
-    }
-#endif // CUDART_VERSION >= 12000
-
-#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
-
-#if !defined(GGML_USE_HIPBLAS)
-static const char * cu_get_error_str(CUresult err) {
-    const char * err_str;
-    cuGetErrorString(err, &err_str);
-    return err_str;
-}
-#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
-#endif
-
-#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
-#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
-#else
-#define GGML_CUDA_ASSUME(x)
-#endif // CUDART_VERSION >= 11100
-
-#ifdef GGML_CUDA_F16
-typedef half dfloat; // dequantize float
-typedef half2 dfloat2;
-#else
-typedef float dfloat; // dequantize float
-typedef float2 dfloat2;
-#endif // GGML_CUDA_F16
-
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
-#define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
-
-#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
-#define FAST_FP16_AVAILABLE
-#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
-
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
-#define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
-
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
-#define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
-
-#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
-#define FLASH_ATTN_AVAILABLE
-#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
-
-static constexpr bool fast_fp16_available(const int cc) {
-    return cc >= CC_PASCAL && cc != 610;
-}
-
-static constexpr bool fp16_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
-}
-
-static constexpr bool int8_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
-}
-
-[[noreturn]]
-static __device__ void no_device_code(
-    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
-           file_name, line, function_name, arch);
-    GGML_UNUSED(arch_list);
-#else
-    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
-           file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    __trap();
-
-    GGML_UNUSED(no_device_code); // suppress unused function warning
-}
-
-#ifdef __CUDA_ARCH__
-#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
-#else
-#define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
-#endif // __CUDA_ARCH__
-
-static __device__ __forceinline__ float warp_reduce_sum(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
-    }
-    return x;
-}
-
-static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
-    }
-    return a;
-}
-
-static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-#ifdef FP16_AVAILABLE
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
-        reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
-        reinterpret_cast<half&>(a.y) += __high2half(a_other);
-    }
-    return a;
-#else
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-    }
-    return a;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
-#else
-    NO_DEVICE_CODE;
-    return a;
-#endif // FP16_AVAILABLE
-}
-
-static __device__ __forceinline__ float warp_reduce_max(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-    }
-    return x;
-}
-
-static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
-#ifdef FP16_AVAILABLE
-
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-    return __float2half(fmaxf(__half2float(a), __half2float(b)));
-#else
-    return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-
-#else
-   NO_DEVICE_CODE;
-   GGML_UNUSED(b);
-   return a;
-#endif // FP16_AVAILABLE
-}
-
-static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-
-#if CUDART_VERSION >= CUDART_HMAX
-    return __hmax2(a, b);
-#else
-    half2 ret;
-    reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
-    reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
-    return ret;
-#endif // CUDART_VERSION >= CUDART_HMAX
-
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-}
-
-static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-   }
-   return x;
-#else
-   GGML_UNUSED(x);
-   NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
-
-#if CUDART_VERSION < CUDART_HMASK
-static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
-    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
-    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
-    return mask_low | mask_high;
-}
-#endif // CUDART_VERSION < CUDART_HMASK
-
-static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
-    c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
-    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
-#elif defined(__gfx1010__) || defined(__gfx900__)
-    int tmp1;
-    int tmp2;
-    asm("\n \
-        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
-        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
-        v_add3_u32 %0, %1, %2, %0 \n \
-        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
-        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
-        v_add3_u32 %0, %1, %2, %0 \n \
-        "
-        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
-        : "v"(a), "v"(b)
-    );
-#else
-    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
-    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
-#endif
-    return c;
-
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A
-    return __dp4a(a, b, c);
-#else // __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int8_t * a8 = (const int8_t *) &a;
-    const int8_t * b8 = (const int8_t *) &b;
-    return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-}
-
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
-typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
-
-static __device__ __forceinline__ float get_alibi_slope(
-    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
-) {
-    if (max_bias <= 0.0f) {
-        return 1.0f;
-    }
-    const float base = h < n_head_log2 ? m0 : m1;
-    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-    return powf(base, exph);
-}
-
-template <ggml_type type>
-struct ggml_cuda_type_traits;
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_F16> {
-    static constexpr int qk = 1;
-    static constexpr int qr = 1;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
-    static constexpr int qk = QK4_0;
-    static constexpr int qr = QR4_0;
-    static constexpr int qi = QI4_0;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q4_1> {
-    static constexpr int qk = QK4_1;
-    static constexpr int qr = QR4_1;
-    static constexpr int qi = QI4_1;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q5_0> {
-    static constexpr int qk = QK5_0;
-    static constexpr int qr = QR5_0;
-    static constexpr int qi = QI5_0;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q5_1> {
-    static constexpr int qk = QK5_1;
-    static constexpr int qr = QR5_1;
-    static constexpr int qi = QI5_1;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
-    static constexpr int qk = QK8_0;
-    static constexpr int qr = QR8_0;
-    static constexpr int qi = QI8_0;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR2_K;
-    static constexpr int qi = QI2_K;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q3_K> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR3_K;
-    static constexpr int qi = QI3_K;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q4_K> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR4_K;
-    static constexpr int qi = QI4_K;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q5_K> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR5_K;
-    static constexpr int qi = QI5_K;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_Q6_K> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR6_K;
-    static constexpr int qi = QI6_K;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XXS> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR2_XXS;
-    static constexpr int qi = QI2_XXS;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XS> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR2_XS;
-    static constexpr int qi = QI2_XS;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ2_S> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR2_S;
-    static constexpr int qi = QI2_S;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ3_XXS> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR3_XXS;
-    static constexpr int qi = QI3_XXS;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR1_S;
-    static constexpr int qi = QI1_S;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR1_M;
-    static constexpr int qi = QI1_M;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ4_NL> {
-    static constexpr int qk = QK4_NL;
-    static constexpr int qr = QR4_NL;
-    static constexpr int qi = QI4_NL;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ4_XS> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR4_XS;
-    static constexpr int qi = QI4_XS;
-};
-
-template<>
-struct ggml_cuda_type_traits<GGML_TYPE_IQ3_S> {
-    static constexpr int qk = QK_K;
-    static constexpr int qr = QR3_S;
-    static constexpr int qi = QI3_S;
-};
-
-//////////////////////
-
-struct ggml_cuda_device_info {
-    int device_count;
-
-    struct cuda_device_info {
-        int     cc;                 // compute capability
-        int     nsm;                // number of streaming multiprocessors
-        size_t  smpb;               // max. shared memory per block
-        size_t  smpbo;              // max. shared memory per block (with opt-in)
-        bool    vmm;                // virtual memory support
-        size_t  vmm_granularity;    // granularity of virtual memory
-        size_t  total_vram;
-    };
-
-    cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
-
-    std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
-};
-
-const ggml_cuda_device_info & ggml_cuda_info();
-
-void ggml_cuda_set_device(int device);
-int ggml_cuda_get_device();
-
-struct ggml_cuda_pool {
-    virtual ~ggml_cuda_pool() = default;
-
-    virtual void * alloc(size_t size, size_t * actual_size) = 0;
-    virtual void free(void * ptr, size_t size) = 0;
-};
-
-template<typename T>
-struct ggml_cuda_pool_alloc {
-    ggml_cuda_pool * pool = nullptr;
-    T * ptr = nullptr;
-    size_t actual_size = 0;
-
-    ggml_cuda_pool_alloc() = default;
-
-    explicit ggml_cuda_pool_alloc(ggml_cuda_pool & pool) : pool(&pool) {
-    }
-
-    ggml_cuda_pool_alloc(ggml_cuda_pool & pool, size_t size) : pool(&pool) {
-        alloc(size);
-    }
-
-    ~ggml_cuda_pool_alloc() {
-        if (ptr != nullptr) {
-            pool->free(ptr, actual_size);
-        }
-    }
-
-    // size is in number of elements
-    T * alloc(size_t size) {
-        GGML_ASSERT(pool != nullptr);
-        GGML_ASSERT(ptr == nullptr);
-        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
-        return ptr;
-    }
-
-    T * alloc(ggml_cuda_pool & pool, size_t size) {
-        this->pool = &pool;
-        return alloc(size);
-    }
-
-    T * get() {
-        return ptr;
-    }
-
-    ggml_cuda_pool_alloc(const ggml_cuda_pool_alloc &) = delete;
-    ggml_cuda_pool_alloc(ggml_cuda_pool_alloc &&) = delete;
-    ggml_cuda_pool_alloc& operator=(const ggml_cuda_pool_alloc &) = delete;
-    ggml_cuda_pool_alloc& operator=(ggml_cuda_pool_alloc &&) = delete;
-};
-
-
-// backend interface
-
-struct ggml_tensor_extra_gpu {
-    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
-    cudaEvent_t events[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS]; // events for synchronizing multiple GPUs
-};
-
-
-#if (CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)
-#define USE_CUDA_GRAPH
-#endif
-
-struct ggml_graph_node_properties {
-    void * node_address;
-    ggml_op node_op;
-    int64_t ne[GGML_MAX_DIMS];
-    size_t nb[GGML_MAX_DIMS];
-    void * src_address[GGML_MAX_SRC];
-    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
-};
-
-struct ggml_cuda_graph {
-#ifdef USE_CUDA_GRAPH
-    ~ggml_cuda_graph() {
-        if (instance != nullptr) {
-            CUDA_CHECK(cudaGraphExecDestroy(instance));
-        }
-        if (graph != nullptr) {
-            CUDA_CHECK(cudaGraphDestroy(graph));
-        }
-    }
-    cudaGraph_t graph = nullptr;
-    cudaGraphExec_t instance = nullptr;
-    size_t num_nodes = 0;
-    std::vector<cudaGraphNode_t> nodes;
-    std::vector<cudaKernelNodeParams> params;
-    bool disable_due_to_gpu_arch = false;
-    bool disable_due_to_too_many_updates = false;
-    bool disable_due_to_failed_graph_capture = false;
-    int number_consecutive_updates = 0;
-    std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    std::vector<char **> updated_kernel_arg;
-#endif
-};
-
-struct ggml_backend_cuda_context {
-    int device;
-    std::string name;
-    cudaEvent_t copy_event = nullptr;
-
-    cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
-    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
-
-    std::unique_ptr<ggml_cuda_graph> cuda_graph;
-
-    explicit ggml_backend_cuda_context(int device) :
-        device(device),
-        name(GGML_CUDA_NAME + std::to_string(device)) {
-    }
-
-    ~ggml_backend_cuda_context() {
-        if (copy_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(copy_event));
-        }
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-                if (streams[i][j] != nullptr) {
-                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-                }
-            }
-            if (cublas_handles[i] != nullptr) {
-                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-            }
-        }
-    }
-
-    cudaStream_t stream(int device, int stream) {
-        if (streams[device][stream] == nullptr) {
-            ggml_cuda_set_device(device);
-            CUDA_CHECK(cudaStreamCreateWithFlags(&streams[device][stream], cudaStreamNonBlocking));
-        }
-        return streams[device][stream];
-    }
-
-    cudaStream_t stream() {
-        return stream(device, 0);
-    }
-
-    cublasHandle_t cublas_handle(int device) {
-        if (cublas_handles[device] == nullptr) {
-            ggml_cuda_set_device(device);
-            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
-            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
-        }
-        return cublas_handles[device];
-    }
-
-    cublasHandle_t cublas_handle() {
-        return cublas_handle(device);
-    }
-
-    // pool
-    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
-
-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
-
-    ggml_cuda_pool & pool(int device) {
-        if (pools[device] == nullptr) {
-            pools[device] = new_pool_for_device(device);
-        }
-        return *pools[device];
-    }
-
-    ggml_cuda_pool & pool() {
-        return pool(device);
-    }
-};

llama/ggml-cuda/concat.cu

@@ -1,222 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "concat.cuh"
-
-// contiguous kernels
-static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (nidx < ne00) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            (nidx - ne00) +
-            blockIdx.y * (ne0 - ne00) +
-            blockIdx.z * (ne0 - ne00) * gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.y < ne01) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * ne01;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            (blockIdx.y - ne01) * ne0 +
-            blockIdx.z * ne0 * (gridDim.y - ne01);
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.z < ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
-    if (dim == 0) {
-        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
-        return;
-    }
-    if (dim == 1) {
-        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
-        return;
-    }
-    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
-}
-
-// non-contiguous kernel (slow)
-static __global__ void concat_f32_non_cont(
-        const char * src0,
-        const char * src1,
-              char * dst,
-           int64_t   ne00,
-           int64_t   ne01,
-           int64_t   ne02,
-           int64_t   ne03,
-          uint64_t   nb00,
-          uint64_t   nb01,
-          uint64_t   nb02,
-          uint64_t   nb03,
-           int64_t /*ne10*/,
-           int64_t /*ne11*/,
-           int64_t /*ne12*/,
-           int64_t /*ne13*/,
-          uint64_t   nb10,
-          uint64_t   nb11,
-          uint64_t   nb12,
-          uint64_t   nb13,
-           int64_t   ne0,
-           int64_t /*ne1*/,
-           int64_t /*ne2*/,
-           int64_t /*ne3*/,
-          uint64_t   nb0,
-          uint64_t   nb1,
-          uint64_t   nb2,
-          uint64_t   nb3,
-          int32_t   dim) {
-    const int64_t i3 = blockIdx.z;
-    const int64_t i2 = blockIdx.y;
-    const int64_t i1 = blockIdx.x;
-
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
-
-    const float * x;
-
-    for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
-        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
-        } else {
-            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
-        }
-
-        float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-        *y = *x;
-    }
-}
-
-
-void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    cudaStream_t stream = ctx.stream();
-
-    const int32_t dim = ((int32_t *) dst->op_params)[0];
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
-
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        const float * src0_d = (const float *)src0->data;
-        const float * src1_d = (const float *)src1->data;
-
-        float * dst_d = (float *)dst->data;
-
-        if (dim != 3) {
-            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-                concat_f32_cuda(
-                        src0_d + i3 * (src0->nb[3] / 4),
-                        src1_d + i3 * (src1->nb[3] / 4),
-                        dst_d + i3 * ( dst->nb[3] / 4),
-                        src0->ne[0], src0->ne[1], src0->ne[2],
-                        dst->ne[0],  dst->ne[1],  dst->ne[2], dim, stream);
-            }
-        } else {
-            const size_t size0 = ggml_nbytes(src0);
-            const size_t size1 = ggml_nbytes(src1);
-
-            CUDA_CHECK(cudaMemcpyAsync(dst_d,           src0_d, size0, cudaMemcpyDeviceToDevice, stream));
-            CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
-        }
-    } else {
-        dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
-        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
-                (const char *)src0->data,
-                (const char *)src1->data,
-                (      char *)dst->data,
-                src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
-                src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
-                src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
-                src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
-                dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
-                dst->nb[0],  dst->nb[1],  dst->nb[2],  dst->nb[3], dim);
-    }
-}

llama/ggml-cuda/concat.cuh

@@ -1,31 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-#define CUDA_CONCAT_BLOCK_SIZE 256
-
-void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/conv-transpose-1d.cu

@@ -1,113 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "conv-transpose-1d.cuh"
-
-static  __global__ void conv_transpose_1d_kernel(
-        const int s0, const int p0, const int d0, const int output_size,
-        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
-        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
-        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
-        const float * src0, const float * src1,  float * dst) {
-    int global_index = threadIdx.x + blockIdx.x * blockDim.x;
-    if (global_index >= output_size) {
-        return;
-    }
-
-    int out_index = global_index / dst_ne0;
-
-    float accumulator = 0;
-
-    for (int c = 0; c < src0_ne2; c++) {
-        int idx = global_index % dst_ne0;
-
-        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
-        int input_offset = src1_ne0 * c;
-
-        for (int i = 0; i < src1_ne0; i++) {
-            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
-                continue;
-            }
-            int weight_idx = idx - i*s0;
-
-            float kernel_weight = src0[kernel_offset + weight_idx];
-            float input_value =  src1[input_offset+i];
-
-            accumulator += kernel_weight * input_value;
-        }
-    }
-    dst[global_index] = accumulator;
-}
-
-static void conv_transpose_1d_f32_f32_cuda(
-        const int s0, const int p0, const int d0, const int output_size,
-        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
-        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
-        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
-        const float * src0, const float * src1,  float * dst,
-        cudaStream_t stream) {
-
-    const int num_blocks = (output_size + CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE;
-    conv_transpose_1d_kernel<<<num_blocks,CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE, 0, stream>>>(
-        s0,p0,d0,output_size,
-        src0_ne0, src0_ne1,  src0_ne2, src0_ne3,
-        src1_ne0, src1_ne1,  src1_ne2, src1_ne3,
-        dst_ne0,  dst_ne1,   dst_ne2,  dst_ne3,
-        src0,src1, dst);
-}
-
-void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-
-    const ggml_tensor * src1 = dst->src[1];
-    const float * src1_d = (const float *)src1->data;
-
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-
-    const int32_t * opts = (const int32_t *)dst->op_params;
-
-    const int s0 = opts[0];
-    const int p0 = 0;//opts[3];
-    const int d0 = 1;//opts[4];
-
-    const int64_t kernel_size = ggml_nelements(src0);
-    const int64_t input_size = ggml_nelements(src1);
-    const int64_t output_size = ggml_nelements(dst);
-
-    conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size,
-        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
-        src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-        src0_d, src1_d, dst_d, stream);
-}

llama/ggml-cuda/conv-transpose-1d.cuh

@@ -1,31 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "common.cuh"
-
-#define CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE 256
-
-void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

llama/ggml-cuda/convert.cu

@@ -1,712 +0,0 @@
-/**
- * llama.cpp - commit 3f1ae2e32cde00c39b96be6d01c2997c29bae555 - do not edit this file
- *
- * MIT License
- *
- * Copyright (c) 2023-2024 The ggml authors
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to deal
- * in the Software without restriction, including without limitation the rights
- * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
- * copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include "convert.cuh"
-#include "dequantize.cuh"
-
-#define CUDA_Q8_0_NE_ALIGN 2048
-
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
-
-    if (i >= k) {
-        return;
-    }
-
-    const int64_t ib = i/qk; // block index
-    const int64_t iqs = (i%qk)/qr; // quant index
-    const int64_t iybs = i - i%qk; // y block start index
-    const int64_t y_offset = qr == 1 ? 1 : qk/2;
-
-    // dequantize
-    dfloat2 v;
-    dequantize_kernel(vx, ib, iqs, v);
-
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
-}
-
-template <bool need_check>
-static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
-#if __CUDA_ARCH__ >= CC_PASCAL
-    constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
-
-    const int64_t   i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
-    const int * x0 = ((int *) vx) + blockIdx.x * nint;
-    half2 * y2 = (half2 *) (y + i0);
-
-    __shared__ int vals[nint];
-
-#pragma unroll
-    for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) {
-        if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) {
-            break;
-        }
-
-        const int ix = ix0 + threadIdx.x;
-        vals[ix] = x0[ix];
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
-        if (need_check && i0 + iy + 2*threadIdx.x >= k) {
-            return;
-        }
-
-        const half * b0 = ((const half  *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0);
-        const half    d = *b0;
-        const char2  qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)];
-
-        y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
-    }
-#else
-    GGML_UNUSED(vx);
-    GGML_UNUSED(y);
-    GGML_UNUSED(k);
-    NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= CC_PASCAL
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
-
-    const int64_t i = blockIdx.x;
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/8;
-    const int64_t ir  = tid%8;
-    const int64_t ib = 8*i + ir;
-    if (ib >= nb32) {
-        return;
-    }
-
-    dst_t * y = yy + 256*i + 32*ir + 4*il;
-
-    const block_q4_0 * x = (const block_q4_0 *)vx + ib;
-    const float d = __half2float(x->d);
-    const float dm = -8*d;
-
-    const uint8_t * q = x->qs + 4*il;
-
-    for (int l = 0; l < 4; ++l) {
-        y[l+ 0] = d * (q[l] & 0xF) + dm;
-        y[l+16] = d * (q[l] >>  4) + dm;
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
-
-    const int64_t i = blockIdx.x;
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/8;
-    const int64_t ir  = tid%8;
-    const int64_t ib = 8*i + ir;
-    if (ib >= nb32) {
-        return;
-    }
-
-    dst_t * y = yy + 256*i + 32*ir + 4*il;
-
-    const block_q4_1 * x = (const block_q4_1 *)vx + ib;
-    const float2 d = __half22float2(x->dm);
-
-    const uint8_t * q = x->qs + 4*il;
-
-    for (int l = 0; l < 4; ++l) {
-        y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
-        y[l+16] = d.x * (q[l] >>  4) + d.y;
-    }
-}
-
-//================================== k-quants
-
-template<typename dst_t>
-static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_q2_K * x = (const block_q2_K *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t n   = tid/32;
-    const int64_t l   = tid - 32*n;
-    const int64_t is  = 8*n + l/16;
-
-    const uint8_t q = x[i].qs[32*n + l];
-    dst_t * y = yy + i*QK_K + 128*n;
-
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
-    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
-    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i = blockIdx.x;
-    const block_q3_K * x = (const block_q3_K *) vx;
-
-    const int64_t r = threadIdx.x/4;
-    const int64_t tid = r/2;
-    const int64_t is0 = r%2;
-    const int64_t l0 = 16*is0 + 4*(threadIdx.x%4);
-    const int64_t n = tid / 4;
-    const int64_t j = tid - 4*n;
-
-    uint8_t m = 1 << (4*n + j);
-    int64_t is = 8*n + 2*j + is0;
-    int shift = 2*j;
-
-    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
-                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
-                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
-                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
-    float d_all = x[i].d;
-    float dl = d_all * (us - 32);
-
-    dst_t * y = yy + i*QK_K + 128*n + 32*j;
-    const uint8_t * q = x[i].qs + 32*n;
-    const uint8_t * hm = x[i].hmask;
-
-    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-}
-
-static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
-    if (j < 4) {
-        d = q[j] & 63; m = q[j + 4] & 63;
-    } else {
-        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q4_K * x = (const block_q4_K *) vx;
-
-    const int64_t i = blockIdx.x;
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/8;
-    const int64_t ir  = tid%8;
-    const int64_t is  = 2*il;
-    const int64_t n   = 4;
-
-    dst_t * y = yy + i*QK_K + 64*il + n*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * q = x[i].qs + 32*il + n*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-    for (int l = 0; l < n; ++l) {
-        y[l + 0] = d1 * (q[l] & 0xF) - m1;
-        y[l +32] = d2 * (q[l] >>  4) - m2;
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q5_K * x = (const block_q5_K *) vx;
-
-    const int64_t i = blockIdx.x;
-
-    // assume 64 threads - this is very slightly better than the one below
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/16;   // il is in 0...3
-    const int64_t ir  = tid%16;   // ir is in 0...15
-    const int64_t is  = 2*il;     // is is in 0...6
-
-    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
-    const uint8_t * qh = x[i].qh + 2*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-
-    uint8_t   hm  = 1 << (2*il);
-    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
-    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
-    hm <<= 1;
-    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
-    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q6_K * x = (const block_q6_K *) vx;
-
-    const int64_t i = blockIdx.x;
-
-    // assume 64 threads - this is very slightly better than the one below
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/32;   // ip is 0 or 1
-    const int64_t il  = tid - 32*ip; // 0...32
-    const int64_t is  = 8*ip + il/16;
-
-    dst_t * y = yy + i*QK_K + 128*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t * ql = x[i].ql + 64*ip + il;
-    const uint8_t   qh = x[i].qh[32*ip + il];
-    const int8_t  * sc = x[i].scales + is;
-
-    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
-    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * q2 = x[i].qs + 4*ib;
-    const uint8_t  * aux8 = (const uint8_t *)q2;
-    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
-    const uint32_t aux32 = q2[2] | (q2[3] << 16);
-    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
-    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq2_xs * x = (const block_iq2_xs *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * q2 = x[i].qs + 4*ib;
-    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
-    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
-    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq2_s * x = (const block_iq2_s *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
-    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
-    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t  * q3 = x[i].qs + 8*ib;
-    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
-    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
-    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
-    const uint32_t aux32 = gas[0] | (gas[1] << 16);
-    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
-    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 4; ++j) {
-        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
-        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq3_s * x = (const block_iq3_s *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t * qs = x[i].qs + 8*ib;
-    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
-    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
-    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
-    const uint8_t signs = x[i].signs[4*ib + il];
-    for (int j = 0; j < 4; ++j) {
-        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
-        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq1_s * x = (const block_iq1_s  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
-    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
-    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
-    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
-    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
-    grid32[0] &= 0x0f0f0f0f;
-    for (int j = 0; j < 8; ++j) {
-        y[j] = d * (q[j] + delta);
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq1_m * x = (const block_iq1_m  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * sc = (const uint16_t *)x[i].scales;
-    iq1m_scale_t scale;
-    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
-    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
-    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
-    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
-    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
-    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
-    grid32[0] &= 0x0f0f0f0f;
-    for (int j = 0; j < 8; ++j) {
-        y[j] = d * (q[j] + delta);
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int64_t i   = blockIdx.x;
-    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[ib].qs + 4*il;
-    const float d = (float)x[ib].d;
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
-        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
-    }
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const int64_t i   = blockIdx.x;
-    const block_iq4_xs * x = (const block_iq4_xs *)vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
-    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
-        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
-    }
-}
-
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
-    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN;
-    if (k % CUDA_Q8_0_NE_ALIGN == 0) {
-        const bool need_check = false;
-        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
-    } else {
-        const bool need_check = true;
-        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
-    }
-}
-
-template<typename dst_t>
-static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb32 = k / 32;
-    const int nb = (k + 255) / 256;
-    dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
-}
-
-template<typename dst_t>
-static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb32 = k / 32;
-    const int nb = (k + 255) / 256;
-    dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
-}
-
-template<typename dst_t>
-static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = (k + QK_K - 1) / QK_K;
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
-    const int nb = (k + QK_K - 1) / QK_K;
-    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template <typename src_t, typename dst_t>
-static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    const src_t * x = (src_t *) vx;
-
-    y[i] = x[i];
-}
-
-template <typename src_t, typename dst_t>
-static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_row_q4_0_cuda;
-        case GGML_TYPE_Q4_1:
-            return dequantize_row_q4_1_cuda;
-        case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
-        case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
-        case GGML_TYPE_Q8_0:
-            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
-                return dequantize_block_q8_0_f16_cuda;
-            }
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_IQ2_XXS:
-            return dequantize_row_iq2_xxs_cuda;
-        case GGML_TYPE_IQ2_XS:
-            return dequantize_row_iq2_xs_cuda;
-        case GGML_TYPE_IQ2_S:
-            return dequantize_row_iq2_s_cuda;
-        case GGML_TYPE_IQ3_XXS:
-            return dequantize_row_iq3_xxs_cuda;
-        case GGML_TYPE_IQ1_S:
-            return dequantize_row_iq1_s_cuda;
-        case GGML_TYPE_IQ1_M:
-            return dequantize_row_iq1_m_cuda;
-        case GGML_TYPE_IQ4_NL:
-            return dequantize_row_iq4_nl_cuda;
-        case GGML_TYPE_IQ4_XS:
-            return dequantize_row_iq4_xs_cuda;
-        case GGML_TYPE_IQ3_S:
-            return dequantize_row_iq3_s_cuda;
-        case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
-        default:
-            return nullptr;
-    }
-}
-
-to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_row_q4_0_cuda;
-        case GGML_TYPE_Q4_1:
-            return dequantize_row_q4_1_cuda;
-        case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
-        case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
-        case GGML_TYPE_Q8_0:
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_IQ2_XXS:
-            return dequantize_row_iq2_xxs_cuda;
-        case GGML_TYPE_IQ2_XS:
-            return dequantize_row_iq2_xs_cuda;
-        case GGML_TYPE_IQ2_S:
-            return dequantize_row_iq2_s_cuda;
-        case GGML_TYPE_IQ3_XXS:
-            return dequantize_row_iq3_xxs_cuda;
-        case GGML_TYPE_IQ1_S:
-            return dequantize_row_iq1_s_cuda;
-        case GGML_TYPE_IQ1_M:
-            return dequantize_row_iq1_m_cuda;
-        case GGML_TYPE_IQ4_NL:
-            return dequantize_row_iq4_nl_cuda;
-        case GGML_TYPE_IQ4_XS:
-            return dequantize_row_iq4_xs_cuda;
-        case GGML_TYPE_IQ3_S:
-            return dequantize_row_iq3_s_cuda;
-        case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
-        default:
-            return nullptr;
-    }
-}