Merge pull request #3684 from ollama/mxyng/scale-graph

scale graph based on gpu count

.gitattributes

@@ -0,0 +1 @@
+llm/ext_server/* linguist-vendored

.github/ISSUE_TEMPLATE/10_model_request.yml

@@ -0,0 +1,18 @@
+name: Model request
+description: Request a new model for the library
+labels: [mr]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please check if your Model request is [already available](https://ollama.com/search) or that you cannot [import it](https://github.com/ollama/ollama/blob/main/docs/import.md#import-a-model) yourself.
+        Tell us about which Model you'd like to see in the library!
+  - type: textarea
+    id: problem
+    attributes:
+      label: What model would you like?
+      description: Please provide a link to the model.
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for filing a model request!

.github/ISSUE_TEMPLATE/20_feature_request.yml

@@ -0,0 +1,41 @@
+name: Feature request
+description: Propose a new feature
+labels: [needs-triage, fr]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please check if your feature request is [already filed](https://github.com/ollama/ollama/issues).
+        Tell us about your idea!
+  - type: textarea
+    id: problem
+    attributes:
+      label: What are you trying to do?
+      description: Tell us about the problem you're trying to solve.
+    validations:
+      required: false
+  - type: textarea
+    id: solution
+    attributes:
+      label: How should we solve this?
+      description: If you have an idea of how you'd like to see this feature work, let us know.
+    validations:
+      required: false
+  - type: textarea
+    id: alternative
+    attributes:
+      label: What is the impact of not solving this?
+      description: (How) Are you currently working around the issue?
+    validations:
+      required: false
+  - type: textarea
+    id: context
+    attributes:
+      label: Anything else?
+      description: Any additional context to share, e.g., links
+    validations:
+      required: false
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for filing a feature request!

.github/ISSUE_TEMPLATE/90_bug_report.yml

@@ -0,0 +1,125 @@
+name: Bug report
+description: File a bug report. If you need help, please join our Discord server.
+labels: [needs-triage, bug]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please check if your bug is [already filed](https://github.com/ollama/ollama/issues) before filing a new one.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What is the issue?
+      description: What happened? What did you expect to happen?
+    validations:
+      required: true
+  - type: textarea
+    id: what-was-expected
+    attributes:
+      label: What did you expect to see?
+      description: What did you expect to see/happen instead?
+    validations:
+      required: false
+  - type: textarea
+    id: steps
+    attributes:
+      label: Steps to reproduce
+      description: What are the steps you took that hit this issue?
+    validations:
+      required: false
+  - type: textarea
+    id: changes
+    attributes:
+      label: Are there any recent changes that introduced the issue?
+      description: If so, what are those changes?
+    validations:
+      required: false
+  - type: dropdown
+    id: os
+    attributes:
+      label: OS
+      description: What OS are you using? You may select more than one.
+      multiple: true
+      options:
+        - Linux
+        - macOS
+        - Windows
+        - Other
+    validations:
+      required: false
+  - type: dropdown
+    id: architecture
+    attributes:
+      label: Architecture
+      description: What architecture are you using? You may select more than one.
+      multiple: true
+      options:
+        - arm64
+        - amd64
+        - x86
+        - Other
+  - type: dropdown
+    id: platform
+    attributes:
+      label: Platform
+      description: What platform are you using? You may select more than one.
+      multiple: true
+      options:
+        - Docker
+        - WSL
+        - WSL2
+    validations:
+      required: false
+  - type: input
+    id: ollama-version
+    attributes:
+      label: Ollama version
+      description: What Ollama version are you using? (`ollama --version`)
+      placeholder: e.g., 1.14.4
+    validations:
+      required: false
+  - type: dropdown
+    id: gpu
+    attributes:
+      label: GPU
+      description: What GPU, if any, are you using? You may select more than one.
+      multiple: true
+      options:
+        - Nvidia
+        - AMD
+        - Intel
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: textarea
+    id: gpu-info
+    attributes:
+      label: GPU info
+      description: What GPU info do you have? (`nvidia-smi`, `rocminfo`, `system_profiler SPDisplaysDataType`, etc.)
+    validations:
+      required: false
+  - type: dropdown
+    id: cpu
+    attributes:
+      label: CPU
+      description: What CPU are you using? You may select more than one.
+      multiple: true
+      options:
+        - Intel
+        - AMD
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: textarea
+    id: other-software
+    attributes:
+      label: Other software
+      description: What other software are you using that might be related to this issue?
+    validations:
+      required: false
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for filing a bug report!

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Help
+    url: https://discord.com/invite/ollama
+    about: Please join our Discord server for help using Ollama
+  - name: Troubleshooting
+    url: https://github.com/ollama/ollama/blob/main/docs/faq.md#faq
+    about: See the FAQ for common issues and solutions

.github/workflows/latest.yaml

@@ -0,0 +1,24 @@
+name: latest
+
+on:
+  release:
+    types: [released]
+
+jobs:
+  update-latest:
+    environment: release
+    runs-on: linux
+    steps:
+      - uses: actions/checkout@v4
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Tag images as latest
+        env:
+          PUSH: "1"
+        shell: bash
+        run: |
+          export "VERSION=${GITHUB_REF_NAME#v}"
+          ./scripts/tag_latest.sh

.github/workflows/release.yaml

@@ -0,0 +1,473 @@
+name: release
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  # Full build of the Mac assets
+  build-darwin:
+    runs-on: macos-12
+    environment: release
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: |
+          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
+      - name: key
+        env:
+          MACOS_SIGNING_KEY: ${{ secrets.MACOS_SIGNING_KEY }}
+          MACOS_SIGNING_KEY_PASSWORD: ${{ secrets.MACOS_SIGNING_KEY_PASSWORD }}
+        run: |
+          echo $MACOS_SIGNING_KEY | base64 --decode > certificate.p12
+          security create-keychain -p password build.keychain
+          security default-keychain -s build.keychain
+          security unlock-keychain -p password build.keychain
+          security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
+          security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: Build Darwin
+        env:
+          APPLE_IDENTITY: ${{ secrets.APPLE_IDENTITY }}
+          APPLE_PASSWORD: ${{ secrets.APPLE_PASSWORD }}
+          APPLE_TEAM_ID: ${{ vars.APPLE_TEAM_ID }}
+          APPLE_ID: ${{ vars.APPLE_ID }}
+          SDKROOT: /Applications/Xcode_13.4.1.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
+          DEVELOPER_DIR: /Applications/Xcode_13.4.1.app/Contents/Developer
+        run: |
+          ./scripts/build_darwin.sh
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-darwin
+          path: |
+            dist/*arwin*
+            !dist/*-cov
+
+  # Windows builds take a long time to both install the dependencies and build, so parallelize
+  # CPU generation step
+  generate-windows-cpu:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "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"
+          go generate -x ./...
+        name: go generate
+      - uses: actions/upload-artifact@v4
+        with:
+          name: generate-windows-cpu
+          path: |
+            llm/build/**/bin/*
+            llm/build/**/*.a
+
+  # ROCm generation step
+  generate-windows-rocm:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install ROCm'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading AMD HIP Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP"
+      - name: 'Verify ROCm'
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).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;$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)
+          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: llm/build/**/bin/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-rocm-deps
+          path: dist/deps/*
+
+  # CUDA generation step
+  generate-windows-cuda:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install CUDA'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading CUDA Installer"
+          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          write-host "Installing CUDA"
+          Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
+          write-host "Completed CUDA"
+          $cudaPath=((resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0].path | split-path | split-path)
+          $cudaVer=($cudaPath | split-path -leaf ) -replace 'v(\d+).(\d+)', '$1_$2' 
+          echo "$cudaPath\bin" >> $env:GITHUB_PATH
+          echo "CUDA_PATH=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" >> $env:GITHUB_ENV
+      - name: 'Verify CUDA'
+        run: nvcc -V
+      - run: go get ./...
+      - name: go generate
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $cudabin=(get-command nvcc).source | split-path
+          & "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"
+          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
+          path: llm/build/**/bin/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-cuda-deps
+          path: dist/deps/*
+
+  # Import the prior generation steps and build the final windows assets
+  build-windows:
+    environment: release
+    runs-on: windows
+    needs:
+      - generate-windows-cuda
+      - generate-windows-rocm
+      - generate-windows-cpu
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-cpu
+          path: llm/build
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-cuda
+          path: llm/build
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps
+          path: dist/deps
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-rocm-deps
+          path: dist/deps
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-rocm
+          path: llm/build
+      - run: dir llm/build
+      - run: |
+          $gopath=(get-command go).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;$env:PATH"
+          $env:OLLAMA_SKIP_GENERATE="1"
+          $env:NVIDIA_DIR=$(resolve-path ".\dist\deps")
+          $env:HIP_PATH=$(resolve-path ".\dist\deps")
+          & .\scripts\build_windows.ps1
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-windows
+          path: dist/*.exe
+
+  # Linux x86 assets built using the container based build
+  build-linux-amd64:
+    environment: release
+    runs-on: linux
+    env:
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: amd64
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: |
+          ./scripts/build_linux.sh
+          ./scripts/build_docker.sh
+          mv dist/deps/* dist/
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-linux-amd64
+          path: |
+            dist/*linux*
+            !dist/*-cov
+
+  # Linux ARM assets built using the container based build
+  # (at present, docker isn't pre-installed on arm ubunutu images)
+  build-linux-arm64:
+    environment: release
+    runs-on: linux-arm64
+    env:
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: arm64
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - name: 'Install Docker'
+        run: |
+          # Add Docker's official GPG key:
+          env
+          uname -a
+          sudo apt-get update
+          sudo apt-get install -y ca-certificates curl
+          sudo install -m 0755 -d /etc/apt/keyrings
+          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
+          sudo chmod a+r /etc/apt/keyrings/docker.asc
+
+          # Add the repository to Apt sources:
+          echo \
+            "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
+            $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
+            sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
+          sudo apt-get update
+          sudo apt-get install -y docker-ce docker-ce-cli containerd.io
+          sudo usermod -aG docker $USER
+          sudo apt-get install acl
+          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: |
+          ./scripts/build_linux.sh
+          ./scripts/build_docker.sh
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-linux-arm64
+          path: |
+            dist/*linux*
+            !dist/*-cov
+
+  # Aggregate all the assets and ship a release
+  release:
+    needs:
+      - build-darwin
+      - build-windows
+      - build-linux-amd64
+      - build-linux-arm64
+    runs-on: linux
+    environment: release
+    permissions:
+      contents: write
+    env:
+      OLLAMA_SKIP_IMAGE_BUILD: '1'
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: |
+          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: ./scripts/build_docker.sh
+      - name: Retrieve built artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: dist
+          pattern: dist-*
+          merge-multiple: true
+      - run: |
+          ls -lh dist/
+          (cd dist; sha256sum * > sha256sum.txt)
+          cat dist/sha256sum.txt
+      - uses: ncipollo/release-action@v1
+        with:
+          name: ${{ env.RELEASE_VERSION }}
+          allowUpdates: true
+          artifacts: 'dist/*'
+          draft: true
+          prerelease: true
+          omitBodyDuringUpdate: true
+          generateReleaseNotes: true
+          omitDraftDuringUpdate: true
+          omitPrereleaseDuringUpdate: true
+          replacesArtifacts: true

.github/workflows/test.yaml

@@ -5,11 +5,35 @@ on:
     paths:
       - '**/*'
       - '!docs/**'
-      - '!examples/**'
       - '!README.md'
 
 jobs:
+  changes:
+    runs-on: ubuntu-latest
+    outputs:
+      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:
+          fetch-depth: 0
+      - id: changes
+        run: |
+          changed() {
+            git diff-tree -r --no-commit-id --name-only ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }} \
+              | xargs python3 -c "import sys; print(any([x.startswith('$1') for x in sys.argv[1:]]))"
+          }
+
+          {
+            echo GENERATE=$(changed llm/)
+            echo GENERATE_CUDA=$(changed llm/)
+            echo GENERATE_ROCM=$(changed llm/)
+          } >>$GITHUB_OUTPUT
+
   generate:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE == 'True' }}
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-2019]
@@ -26,26 +50,32 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5
         with:
-          go-version: '1.22'
+          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;$env:PATH"
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
           go generate -x ./...
         if: ${{ startsWith(matrix.os, 'windows-') }}
-        name: "Windows Go Generate"
+        name: 'Windows Go Generate'
       - run: go generate -x ./...
         if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        name: "Unix Go Generate"
+        name: 'Unix Go Generate'
       - uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: |
+            llm/build/**/bin/*
+            llm/build/**/*.a
   generate-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
     strategy:
       matrix:
         cuda-version:
@@ -62,7 +92,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v4
         with:
-          go-version: '1.22'
+          go-version-file: go.mod
           cache: true
       - run: go get ./...
       - run: |
@@ -73,12 +103,14 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: cuda-${{ matrix.cuda-version }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: llm/build/**/bin/*
   generate-rocm:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
     strategy:
       matrix:
         rocm-version:
-          - '6.0'
+          - '6.0.2'
     runs-on: linux
     container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
     steps:
@@ -91,7 +123,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v4
         with:
-          go-version: '1.22'
+          go-version-file: go.mod
           cache: true
       - run: go get ./...
       - run: |
@@ -102,7 +134,87 @@ jobs:
       - uses: actions/upload-artifact@v4
         with:
           name: rocm-${{ matrix.rocm-version }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: llm/build/**/bin/*
+
+  # ROCm generation step
+  generate-windows-rocm:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
+    runs-on: windows
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install ROCm'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading AMD HIP Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP"
+      - name: 'Verify ROCm'
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).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;$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)
+          go generate -x ./...
+        name: go generate
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
+
+  # CUDA generation step
+  generate-windows-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
+    runs-on: windows
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install CUDA'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading CUDA Installer"
+          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          write-host "Installing CUDA"
+          Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
+          write-host "Completed CUDA"
+          $cudaPath=((resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0].path | split-path | split-path)
+          $cudaVer=($cudaPath | split-path -leaf ) -replace 'v(\d+).(\d+)', '$1_$2' 
+          echo "$cudaPath\bin" >> $env:GITHUB_PATH
+          echo "CUDA_PATH=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" >> $env:GITHUB_ENV
+      - name: 'Verify CUDA'
+        run: nvcc -V
+      - run: go get ./...
+      - name: go generate
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $cudabin=(get-command nvcc).source | split-path
+          & "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"
+          $env:OLLAMA_SKIP_CPU_GENERATE="1"
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
+
   lint:
     strategy:
       matrix:
@@ -125,24 +237,31 @@ jobs:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version: '1.22'
+          go-version-file: go.mod
           cache: false
       - run: |
-          mkdir -p llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/stub.so
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=x86_64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
+      - run: |
+          mkdir -p llm/build/linux/$ARCH/stub/bin
+          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'ubuntu-') }}
       - run: |
-          mkdir -p llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/stub.dylib
-          touch llm/llama.cpp/ggml-metal.metal
+          mkdir -p llm/build/darwin/$ARCH/stub/bin
+          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'macos-') }}
       - run: |
-          mkdir -p llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/stub.dll
+          mkdir -p llm/build/windows/$ARCH/stub/bin
+          touch llm/build/windows/$ARCH/stub/bin/ollama_llama_server
         if: ${{ startsWith(matrix.os, 'windows-') }}
-      - uses: golangci/golangci-lint-action@v3
+        shell: bash
+      - uses: golangci/golangci-lint-action@v4
+        with:
+          args: --timeout 8m0s
   test:
-    needs: generate
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-2019]
@@ -156,19 +275,36 @@ jobs:
     env:
       GOARCH: ${{ matrix.arch }}
       CGO_ENABLED: '1'
+      OLLAMA_CPU_TARGET: 'static'
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: recursive
       - uses: actions/setup-go@v5
         with:
-          go-version: '1.22'
+          go-version-file: go.mod
           cache: true
       - run: go get
-      - uses: actions/download-artifact@v4
-        with:
-          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: llm/llama.cpp/build
+      - run: |
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=x86_64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
+      - run: |
+          mkdir -p llm/build/linux/$ARCH/stub/bin
+          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
+      - run: |
+          mkdir -p llm/build/darwin/$ARCH/stub/bin
+          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'macos-') }}
+      - run: |
+          mkdir -p llm/build/windows/$ARCH/stub/bin
+          touch llm/build/windows/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        shell: bash
+      - run: go generate ./...
       - run: go build
       - run: go test -v ./...
       - uses: actions/upload-artifact@v4

.gitignore

@@ -10,4 +10,5 @@ ggml-metal.metal
 *.exe
 .idea
 test_data
-*.crt
+*.crt
+llm/build

.golangci.yaml

@@ -15,13 +15,3 @@ linters:
     - misspell
     - nilerr
     - unused
-linters-settings:
-  errcheck:
-    # exclude the following functions since we don't generally
-    # need to be concerned with the returned errors
-    exclude-functions:
-      - encoding/binary.Read
-      - (*os.File).Seek
-      - (*bufio.Writer).WriteString
-      - (*github.com/spf13/pflag.FlagSet).Set
-      - (*github.com/jmorganca/ollama/llm.readSeekOffset).Seek

Dockerfile

@@ -1,7 +1,8 @@
 ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
+# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
 ARG CUDA_VERSION=11.3.1
-ARG ROCM_VERSION=6.0
+ARG ROCM_VERSION=6.0.2
 
 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@@ -14,8 +15,8 @@ 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/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
@@ -24,8 +25,8 @@ 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/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 
@@ -35,18 +36,18 @@ 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/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
 RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 RUN mkdir /tmp/scratch && \
-    for dep in $(cat /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/x86_64/rocm*/lib/deps.txt) ; do \
+    for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
         cp ${dep} /tmp/scratch/ || exit 1 ; \
     done && \
     (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
-    mkdir -p /go/src/github.com/jmorganca/ollama/dist/deps/ && \
-    (cd /tmp/scratch/ && tar czvf /go/src/github.com/jmorganca/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
+    mkdir -p /go/src/github.com/ollama/ollama/dist/deps/ && \
+    (cd /tmp/scratch/ && tar czvf /go/src/github.com/ollama/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
 
 
 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
@@ -55,11 +56,13 @@ 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
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 
+FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
 RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
@@ -67,56 +70,62 @@ RUN OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
 RUN OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
 
-FROM --platform=linux/arm64 centos:7 AS cpu-build-arm64
+FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
-# Note, we only build the "base" CPU variant on arm since avx/avx2 are x86 features
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+
+FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
 RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
 
+
 # Intermediate stage used for ./scripts/build_linux.sh
 FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
 ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/jmorganca/ollama
+WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=cuda-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/dist/deps/ ./dist/deps/
+COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build .
+RUN go build -trimpath .
 
 # Intermediate stage used for ./scripts/build_linux.sh
 FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
 ENV CGO_ENABLED 1
 ARG GOLANG_VERSION
-WORKDIR /go/src/github.com/jmorganca/ollama
+WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=cuda-build-arm64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 ARG GOFLAGS
 ARG CGO_CFLAGS
-RUN go build .
+RUN go build -trimpath .
 
 # Runtime stages
 FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
 FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
 
 # Radeon images are much larger so we keep it distinct from the CPU/CUDA image
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
 RUN update-pciids
-COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
 EXPOSE 11434
 ENV OLLAMA_HOST 0.0.0.0
 

README.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <img alt="ollama" height="200px" src="https://github.com/jmorganca/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+  <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>
 
 # Ollama
@@ -22,7 +22,7 @@ Get up and running with large language models locally.
 curl -fsSL https://ollama.com/install.sh | sh
 ```
 
-[Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
+[Manual install instructions](https://github.com/ollama/ollama/blob/main/docs/linux.md)
 
 ### Docker
 
@@ -64,6 +64,7 @@ Here are some example models that can be downloaded:
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
 | Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
+| Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
 
 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
 
@@ -213,7 +214,7 @@ Then build the binary:
 go build .
 ```
 
-More detailed instructions can be found in the [developer guide](https://github.com/jmorganca/ollama/blob/main/docs/development.md)
+More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)
 
 ### Running local builds
 
@@ -259,9 +260,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
+- [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
+- [LibreChat](https://github.com/danny-avila/LibreChat)
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
+- [Saddle](https://github.com/jikkuatwork/saddle)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
@@ -272,14 +276,24 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
 - [Ollama-SwiftUI](https://github.com/kghandour/Ollama-SwiftUI)
+- [Dify.AI](https://github.com/langgenius/dify)
 - [MindMac](https://mindmac.app)
 - [NextJS Web Interface for Ollama](https://github.com/jakobhoeg/nextjs-ollama-llm-ui)
 - [Msty](https://msty.app)
 - [Chatbox](https://github.com/Bin-Huang/Chatbox)
 - [WinForm Ollama Copilot](https://github.com/tgraupmann/WinForm_Ollama_Copilot)
 - [NextChat](https://github.com/ChatGPTNextWeb/ChatGPT-Next-Web) with [Get Started Doc](https://docs.nextchat.dev/models/ollama)
+- [Alpaca WebUI](https://github.com/mmo80/alpaca-webui)
+- [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
+- [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
 - [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
+- [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
+- [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
+- [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
+- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
+- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
+- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
 
 ### Terminal
 
@@ -288,18 +302,23 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Emacs client](https://github.com/zweifisch/ollama)
 - [gen.nvim](https://github.com/David-Kunz/gen.nvim)
 - [ollama.nvim](https://github.com/nomnivore/ollama.nvim)
+- [ollero.nvim](https://github.com/marco-souza/ollero.nvim)
 - [ollama-chat.nvim](https://github.com/gerazov/ollama-chat.nvim)
 - [ogpt.nvim](https://github.com/huynle/ogpt.nvim)
 - [gptel Emacs client](https://github.com/karthink/gptel)
 - [Oatmeal](https://github.com/dustinblackman/oatmeal)
 - [cmdh](https://github.com/pgibler/cmdh)
+- [ooo](https://github.com/npahlfer/ooo)
 - [tenere](https://github.com/pythops/tenere)
 - [llm-ollama](https://github.com/taketwo/llm-ollama) for [Datasette's LLM CLI](https://llm.datasette.io/en/stable/).
+- [typechat-cli](https://github.com/anaisbetts/typechat-cli)
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
+- [tlm](https://github.com/yusufcanb/tlm)
 
 ### Database
 
-- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md)
+- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
+- [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)
 
 ### Package managers
 
@@ -312,7 +331,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
-- [LangChain4j](https://github.com/langchain4j/langchain4j/tree/main/langchain4j-ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
@@ -329,6 +347,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
+- [Testcontainers](https://testcontainers.com/modules/ollama/)
 
 ### Mobile
 
@@ -350,9 +369,12 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
+- [Cliobot](https://github.com/herval/cliobot) (Telegram bot with Ollama support)
 - [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
 - [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
+- [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
+- [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)

api/client.go

@@ -1,3 +1,9 @@
+// Package api implements the client-side API for code wishing to interact
+// with the ollama service. The methods of the [Client] type correspond to
+// the ollama REST API as described in https://github.com/ollama/ollama/blob/main/docs/api.md
+//
+// The ollama command-line client itself uses this package to interact with
+// the backend service.
 package api
 
 import (
@@ -5,7 +11,6 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"net"
@@ -15,10 +20,12 @@ import (
 	"runtime"
 	"strings"
 
-	"github.com/jmorganca/ollama/format"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/version"
 )
 
+// Client encapsulates client state for interacting with the ollama
+// service. Use [ClientFromEnvironment] to create new Clients.
 type Client struct {
 	base *url.URL
 	http *http.Client
@@ -40,6 +47,15 @@ func checkError(resp *http.Response, body []byte) error {
 	return apiError
 }
 
+// ClientFromEnvironment creates a new [Client] using configuration from the
+// environment variable OLLAMA_HOST, which points to the network host and
+// port on which the ollama service is listenting. The format of this variable
+// is:
+//
+//	<scheme>://<host>:<port>
+//
+// If the variable is not specified, a default ollama host and port will be
+// used.
 func ClientFromEnvironment() (*Client, error) {
 	defaultPort := "11434"
 
@@ -191,8 +207,14 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	return nil
 }
 
+// GenerateResponseFunc is a function that [Client.Generate] invokes every time
+// a response is received from the service. If this function returns an error,
+// [Client.Generate] will stop generating and return this error.
 type GenerateResponseFunc func(GenerateResponse) error
 
+// Generate generates a response for a given prompt. The req parameter should
+// be populated with prompt details. fn is called for each response (there may
+// be multiple responses, e.g. in case streaming is enabled).
 func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn GenerateResponseFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/generate", req, func(bts []byte) error {
 		var resp GenerateResponse
@@ -204,8 +226,15 @@ func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn Generate
 	})
 }
 
+// ChatResponseFunc is a function that [Client.Chat] invokes every time
+// a response is received from the service. If this function returns an error,
+// [Client.Chat] will stop generating and return this error.
 type ChatResponseFunc func(ChatResponse) error
 
+// Chat generates the next message in a chat. [ChatRequest] may contain a
+// sequence of messages which can be used to maintain chat history with a model.
+// fn is called for each response (there may be multiple responses, e.g. if case
+// streaming is enabled).
 func (c *Client) Chat(ctx context.Context, req *ChatRequest, fn ChatResponseFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/chat", req, func(bts []byte) error {
 		var resp ChatResponse
@@ -217,8 +246,14 @@ func (c *Client) Chat(ctx context.Context, req *ChatRequest, fn ChatResponseFunc
 	})
 }
 
+// PullProgressFunc is a function that [Client.Pull] invokes every time there
+// is progress with a "pull" request sent to the service. If this function
+// returns an error, [Client.Pull] will stop the process and return this error.
 type PullProgressFunc func(ProgressResponse) error
 
+// Pull downloads a model from the ollama library. fn is called each time
+// progress is made on the request and can be used to display a progress bar,
+// etc.
 func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/pull", req, func(bts []byte) error {
 		var resp ProgressResponse
@@ -301,18 +336,7 @@ func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*Embedd
 }
 
 func (c *Client) CreateBlob(ctx context.Context, digest string, r io.Reader) error {
-	if err := c.do(ctx, http.MethodHead, fmt.Sprintf("/api/blobs/%s", digest), nil, nil); err != nil {
-		var statusError StatusError
-		if !errors.As(err, &statusError) || statusError.StatusCode != http.StatusNotFound {
-			return err
-		}
-
-		if err := c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil); err != nil {
-			return err
-		}
-	}
-
-	return nil
+	return c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil)
 }
 
 func (c *Client) Version(ctx context.Context) (string, error) {

api/types.go

@@ -33,18 +33,46 @@ func (e StatusError) Error() string {
 
 type ImageData []byte
 
+// GenerateRequest describes a request sent by [Client.Generate]. While you
+// have to specify the Model and Prompt fields, all the other fields have
+// reasonable defaults for basic uses.
 type GenerateRequest struct {
-	Model     string      `json:"model"`
-	Prompt    string      `json:"prompt"`
-	System    string      `json:"system"`
-	Template  string      `json:"template"`
-	Context   []int       `json:"context,omitempty"`
-	Stream    *bool       `json:"stream,omitempty"`
-	Raw       bool        `json:"raw,omitempty"`
-	Format    string      `json:"format"`
-	KeepAlive *Duration   `json:"keep_alive,omitempty"`
-	Images    []ImageData `json:"images,omitempty"`
+	// Model is the model name; it should be a name familiar to Ollama from
+	// the library at https://ollama.com/library
+	Model string `json:"model"`
 
+	// Prompt is the textual prompt to send to the model.
+	Prompt string `json:"prompt"`
+
+	// System overrides the model's default system message/prompt.
+	System string `json:"system"`
+
+	// Template overrides the model's default prompt template.
+	Template string `json:"template"`
+
+	// Context is the context parameter returned from a previous call to
+	// Generate call. It can be used to keep a short conversational memory.
+	Context []int `json:"context,omitempty"`
+
+	// Stream specifies whether the response is streaming; it is true by default.
+	Stream *bool `json:"stream,omitempty"`
+
+	// Raw set to true means that no formatting will be applied to the prompt.
+	Raw bool `json:"raw,omitempty"`
+
+	// Format specifies the format to return a response in.
+	Format string `json:"format"`
+
+	// KeepAlive controls how long the model will stay loaded in memory following
+	// this request.
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
+
+	// Images is an optional list of base64-encoded images accompanying this
+	// request, for multimodal models.
+	Images []ImageData `json:"images,omitempty"`
+
+	// Options lists model-specific options. For example, temperature can be
+	// set through this field, if the model supports it.
 	Options map[string]interface{} `json:"options"`
 }
 
@@ -109,21 +137,24 @@ type Options struct {
 
 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA            bool    `json:"numa,omitempty"`
-	NumCtx             int     `json:"num_ctx,omitempty"`
-	NumBatch           int     `json:"num_batch,omitempty"`
-	NumGQA             int     `json:"num_gqa,omitempty"`
-	NumGPU             int     `json:"num_gpu,omitempty"`
-	MainGPU            int     `json:"main_gpu,omitempty"`
-	LowVRAM            bool    `json:"low_vram,omitempty"`
-	F16KV              bool    `json:"f16_kv,omitempty"`
-	LogitsAll          bool    `json:"logits_all,omitempty"`
-	VocabOnly          bool    `json:"vocab_only,omitempty"`
-	UseMMap            bool    `json:"use_mmap,omitempty"`
-	UseMLock           bool    `json:"use_mlock,omitempty"`
-	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
+	UseNUMA   bool `json:"numa,omitempty"`
+	NumCtx    int  `json:"num_ctx,omitempty"`
+	NumBatch  int  `json:"num_batch,omitempty"`
+	NumGQA    int  `json:"num_gqa,omitempty"`
+	NumGPU    int  `json:"num_gpu,omitempty"`
+	MainGPU   int  `json:"main_gpu,omitempty"`
+	LowVRAM   bool `json:"low_vram,omitempty"`
+	F16KV     bool `json:"f16_kv,omitempty"`
+	LogitsAll bool `json:"logits_all,omitempty"`
+	VocabOnly bool `json:"vocab_only,omitempty"`
+	UseMMap   bool `json:"use_mmap,omitempty"`
+	UseMLock  bool `json:"use_mlock,omitempty"`
+	NumThread int  `json:"num_thread,omitempty"`
+
+	// Unused: RopeFrequencyBase is ignored. Instead the value in the model will be used
+	RopeFrequencyBase float32 `json:"rope_frequency_base,omitempty"`
+	// Unused: RopeFrequencyScale is ignored. Instead the value in the model will be used
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
-	NumThread          int     `json:"num_thread,omitempty"`
 }
 
 type EmbeddingRequest struct {
@@ -139,10 +170,11 @@ type EmbeddingResponse struct {
 }
 
 type CreateRequest struct {
-	Model     string `json:"model"`
-	Path      string `json:"path"`
-	Modelfile string `json:"modelfile"`
-	Stream    *bool  `json:"stream,omitempty"`
+	Model        string `json:"model"`
+	Path         string `json:"path"`
+	Modelfile    string `json:"modelfile"`
+	Stream       *bool  `json:"stream,omitempty"`
+	Quantization string `json:"quantization,omitempty"`
 
 	// Name is deprecated, see Model
 	Name string `json:"name"`
@@ -382,18 +414,16 @@ func DefaultOptions() Options {
 
 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:             2048,
-			RopeFrequencyBase:  10000.0,
-			RopeFrequencyScale: 1.0,
-			NumBatch:           512,
-			NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
-			NumGQA:             1,
-			NumThread:          0, // let the runtime decide
-			LowVRAM:            false,
-			F16KV:              true,
-			UseMLock:           false,
-			UseMMap:            true,
-			UseNUMA:            false,
+			NumCtx:    2048,
+			NumBatch:  512,
+			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
+			NumGQA:    1,
+			NumThread: 0, // let the runtime decide
+			LowVRAM:   false,
+			F16KV:     true,
+			UseMLock:  false,
+			UseMMap:   true,
+			UseNUMA:   false,
 		},
 	}
 }

api/types_test.go

@@ -0,0 +1,50 @@
+package api
+
+import (
+	"encoding/json"
+	"math"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestKeepAliveParsingFromJSON(t *testing.T) {
+	tests := []struct {
+		name string
+		req  string
+		exp  *Duration
+	}{
+		{
+			name: "Positive Integer",
+			req:  `{ "keep_alive": 42 }`,
+			exp:  &Duration{42 * time.Second},
+		},
+		{
+			name: "Positive Integer String",
+			req:  `{ "keep_alive": "42m" }`,
+			exp:  &Duration{42 * time.Minute},
+		},
+		{
+			name: "Negative Integer",
+			req:  `{ "keep_alive": -1 }`,
+			exp:  &Duration{math.MaxInt64},
+		},
+		{
+			name: "Negative Integer String",
+			req:  `{ "keep_alive": "-1m" }`,
+			exp:  &Duration{math.MaxInt64},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var dec ChatRequest
+			err := json.Unmarshal([]byte(test.req), &dec)
+			require.NoError(t, err)
+
+			assert.Equal(t, test.exp, dec.KeepAlive)
+		})
+	}
+}

app/lifecycle/lifecycle.go

@@ -9,8 +9,8 @@ import (
 	"os/signal"
 	"syscall"
 
-	"github.com/jmorganca/ollama/app/store"
-	"github.com/jmorganca/ollama/app/tray"
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/tray"
 )
 
 func Run() {

app/lifecycle/server.go

@@ -11,7 +11,7 @@ import (
 	"path/filepath"
 	"time"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )
 
 func getCLIFullPath(command string) string {
@@ -83,6 +83,38 @@ func SpawnServer(ctx context.Context, command string) (chan int, error) {
 		io.Copy(logFile, stderr) //nolint:errcheck
 	}()
 
+	// Re-wire context done behavior to attempt a graceful shutdown of the server
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			err := terminate(cmd)
+			if err != nil {
+				slog.Warn("error trying to gracefully terminate server", "err", err)
+				return cmd.Process.Kill()
+			}
+
+			tick := time.NewTicker(10 * time.Millisecond)
+			defer tick.Stop()
+
+			for {
+				select {
+				case <-tick.C:
+					exited, err := isProcessExited(cmd.Process.Pid)
+					if err != nil {
+						return err
+					}
+
+					if exited {
+						return nil
+					}
+				case <-time.After(5 * time.Second):
+					slog.Warn("graceful server shutdown timeout, killing", "pid", cmd.Process.Pid)
+					return cmd.Process.Kill()
+				}
+			}
+		}
+		return nil
+	}
+
 	// run the command and wait for it to finish
 	if err := cmd.Start(); err != nil {
 		return done, fmt.Errorf("failed to start server %w", err)
@@ -105,7 +137,7 @@ func SpawnServer(ctx context.Context, command string) (chan int, error) {
 
 			select {
 			case <-ctx.Done():
-				slog.Debug(fmt.Sprintf("server shutdown with exit code %d", code))
+				slog.Info(fmt.Sprintf("server shutdown with exit code %d", code))
 				done <- code
 				return
 			default:

app/lifecycle/server_unix.go

@@ -4,9 +4,35 @@ package lifecycle
 
 import (
 	"context"
+	"errors"
+	"fmt"
+	"os"
 	"os/exec"
+	"syscall"
 )
 
 func getCmd(ctx context.Context, cmd string) *exec.Cmd {
 	return exec.CommandContext(ctx, cmd, "serve")
 }
+
+func terminate(cmd *exec.Cmd) error {
+	return cmd.Process.Signal(os.Interrupt)
+}
+
+func isProcessExited(pid int) (bool, error) {
+	proc, err := os.FindProcess(pid)
+	if err != nil {
+		return false, fmt.Errorf("failed to find process: %v", err)
+	}
+
+	err = proc.Signal(syscall.Signal(0))
+	if err != nil {
+		if errors.Is(err, os.ErrProcessDone) || errors.Is(err, syscall.ESRCH) {
+			return true, nil
+		}
+
+		return false, fmt.Errorf("error signaling process: %v", err)
+	}
+
+	return false, nil
+}

app/lifecycle/server_windows.go

@@ -2,12 +2,88 @@ package lifecycle
 
 import (
 	"context"
+	"fmt"
 	"os/exec"
 	"syscall"
+
+	"golang.org/x/sys/windows"
 )
 
 func getCmd(ctx context.Context, exePath string) *exec.Cmd {
 	cmd := exec.CommandContext(ctx, exePath, "serve")
-	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true, CreationFlags: 0x08000000}
+	cmd.SysProcAttr = &syscall.SysProcAttr{
+		HideWindow:    true,
+		CreationFlags: windows.CREATE_NEW_PROCESS_GROUP,
+	}
+
 	return cmd
 }
+
+func terminate(cmd *exec.Cmd) error {
+	dll, err := windows.LoadDLL("kernel32.dll")
+	if err != nil {
+		return err
+	}
+	defer dll.Release() // nolint: errcheck
+
+	pid := cmd.Process.Pid
+
+	f, err := dll.FindProc("AttachConsole")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err := f.Call(uintptr(pid))
+	if r1 == 0 && err != syscall.ERROR_ACCESS_DENIED {
+		return err
+	}
+
+	f, err = dll.FindProc("SetConsoleCtrlHandler")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err = f.Call(0, 1)
+	if r1 == 0 {
+		return err
+	}
+
+	f, err = dll.FindProc("GenerateConsoleCtrlEvent")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err = f.Call(windows.CTRL_BREAK_EVENT, uintptr(pid))
+	if r1 == 0 {
+		return err
+	}
+
+	r1, _, err = f.Call(windows.CTRL_C_EVENT, uintptr(pid))
+	if r1 == 0 {
+		return err
+	}
+
+	return nil
+}
+
+const STILL_ACTIVE = 259
+
+func isProcessExited(pid int) (bool, error) {
+	hProcess, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION, false, uint32(pid))
+	if err != nil {
+		return false, fmt.Errorf("failed to open process: %v", err)
+	}
+	defer windows.CloseHandle(hProcess) // nolint: errcheck
+
+	var exitCode uint32
+	err = windows.GetExitCodeProcess(hProcess, &exitCode)
+	if err != nil {
+		return false, fmt.Errorf("failed to get exit code: %v", err)
+	}
+
+	if exitCode == STILL_ACTIVE {
+		return false, nil
+	}
+
+	return true, nil
+}

app/lifecycle/updater.go

@@ -18,8 +18,8 @@ import (
 	"strings"
 	"time"
 
-	"github.com/jmorganca/ollama/auth"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/auth"
+	"github.com/ollama/ollama/version"
 )
 
 var (

app/main.go

@@ -4,7 +4,7 @@ package main
 // go build -ldflags="-H windowsgui" .
 
 import (
-	"github.com/jmorganca/ollama/app/lifecycle"
+	"github.com/ollama/ollama/app/lifecycle"
 )
 
 func main() {

app/ollama.iss

@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64
-ArchitecturesInstallIn64BitMode=x64
+ArchitecturesAllowed=x64 arm64
+ArchitecturesInstallIn64BitMode=x64 arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes

app/tray/tray.go

@@ -4,8 +4,8 @@ import (
 	"fmt"
 	"runtime"
 
-	"github.com/jmorganca/ollama/app/assets"
-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/assets"
+	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 func NewTray() (commontray.OllamaTray, error) {
@@ -24,10 +24,5 @@ func NewTray() (commontray.OllamaTray, error) {
 		return nil, fmt.Errorf("failed to load icon %s: %w", iconName, err)
 	}
 
-	tray, err := InitPlatformTray(icon, updateIcon)
-	if err != nil {
-		return nil, err
-	}
-
-	return tray, nil
+	return InitPlatformTray(icon, updateIcon)
 }

app/tray/tray_nonwindows.go

@@ -5,7 +5,7 @@ package tray
 import (
 	"fmt"
 
-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/commontray"
 )
 
 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {

app/tray/tray_windows.go

@@ -1,8 +1,8 @@
 package tray
 
 import (
-	"github.com/jmorganca/ollama/app/tray/commontray"
-	"github.com/jmorganca/ollama/app/tray/wintray"
+	"github.com/ollama/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/wintray"
 )
 
 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {

app/tray/wintray/tray.go

@@ -13,7 +13,7 @@ import (
 	"sync"
 	"unsafe"
 
-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/commontray"
 	"golang.org/x/sys/windows"
 )
 

cmd/cmd.go

@@ -30,12 +30,12 @@ import (
 	"golang.org/x/exp/slices"
 	"golang.org/x/term"
 
-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/format"
-	"github.com/jmorganca/ollama/parser"
-	"github.com/jmorganca/ollama/progress"
-	"github.com/jmorganca/ollama/server"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/server"
+	"github.com/ollama/ollama/version"
 )
 
 func CreateHandler(cmd *cobra.Command, args []string) error {
@@ -105,24 +105,48 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 
 				zf := zip.NewWriter(tf)
 
-				files, err := filepath.Glob(filepath.Join(path, "model-*.safetensors"))
+				files := []string{}
+
+				tfiles, err := filepath.Glob(filepath.Join(path, "pytorch_model-*.bin"))
 				if err != nil {
 					return err
+				} else if len(tfiles) == 0 {
+					tfiles, err = filepath.Glob(filepath.Join(path, "model-*.safetensors"))
+					if err != nil {
+						return err
+					}
 				}
 
+				files = append(files, tfiles...)
+
 				if len(files) == 0 {
-					return fmt.Errorf("no safetensors files were found in '%s'", path)
+					return fmt.Errorf("no models were found in '%s'", path)
 				}
 
-				// add the safetensor config file + tokenizer
+				// add the safetensor/torch config file + tokenizer
 				files = append(files, filepath.Join(path, "config.json"))
+				files = append(files, filepath.Join(path, "params.json"))
 				files = append(files, filepath.Join(path, "added_tokens.json"))
 				files = append(files, filepath.Join(path, "tokenizer.model"))
 
 				for _, fn := range files {
 					f, err := os.Open(fn)
-					if os.IsNotExist(err) && strings.HasSuffix(fn, "added_tokens.json") {
-						continue
+
+					// just skip whatever files aren't there
+					if os.IsNotExist(err) {
+						if strings.HasSuffix(fn, "tokenizer.model") {
+							// try the parent dir before giving up
+							parentDir := filepath.Dir(path)
+							newFn := filepath.Join(parentDir, "tokenizer.model")
+							f, err = os.Open(newFn)
+							if os.IsNotExist(err) {
+								continue
+							} else if err != nil {
+								return err
+							}
+						} else {
+							continue
+						}
 					} else if err != nil {
 						return err
 					}
@@ -194,7 +218,9 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	request := api.CreateRequest{Name: args[0], Modelfile: string(modelfile)}
+	quantization, _ := cmd.Flags().GetString("quantization")
+
+	request := api.CreateRequest{Name: args[0], Modelfile: string(modelfile), Quantization: quantization}
 	if err := client.Create(cmd.Context(), &request, fn); err != nil {
 		return err
 	}
@@ -213,7 +239,10 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 	if _, err := io.Copy(hash, bin); err != nil {
 		return "", err
 	}
-	bin.Seek(0, io.SeekStart)
+
+	if _, err := bin.Seek(0, io.SeekStart); err != nil {
+		return "", err
+	}
 
 	digest := fmt.Sprintf("sha256:%x", hash.Sum(nil))
 	if err = client.CreateBlob(cmd.Context(), digest, bin); err != nil {
@@ -900,8 +929,7 @@ func NewCLI() *cobra.Command {
 	cobra.EnableCommandSorting = false
 
 	if runtime.GOOS == "windows" {
-		// Enable colorful ANSI escape code in Windows terminal (disabled by default)
-		console.ConsoleFromFile(os.Stdout) //nolint:errcheck
+		console.ConsoleFromFile(os.Stdin) //nolint:errcheck
 	}
 
 	rootCmd := &cobra.Command{
@@ -933,6 +961,7 @@ func NewCLI() *cobra.Command {
 	}
 
 	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
+	createCmd.Flags().StringP("quantization", "q", "", "Quantization level.")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -970,9 +999,11 @@ func NewCLI() *cobra.Command {
 	serveCmd.SetUsageTemplate(serveCmd.UsageTemplate() + `
 Environment Variables:
 
-    OLLAMA_HOST       The host:port to bind to (default "127.0.0.1:11434")
-    OLLAMA_ORIGINS    A comma separated list of allowed origins.
-    OLLAMA_MODELS     The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_HOST         The host:port to bind to (default "127.0.0.1:11434")
+    OLLAMA_ORIGINS      A comma separated list of allowed origins.
+    OLLAMA_MODELS       The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default is "5m")
+    OLLAMA_DEBUG        Set to 1 to enable additional debug logging
 `)
 
 	pullCmd := &cobra.Command{

cmd/interactive.go

@@ -14,9 +14,9 @@ import (
 	"github.com/spf13/cobra"
 	"golang.org/x/exp/slices"
 
-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/progress"
-	"github.com/jmorganca/ollama/readline"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
 )
 
 type MultilineState int
@@ -295,10 +295,14 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					opts.WordWrap = false
 					fmt.Println("Set 'nowordwrap' mode.")
 				case "verbose":
-					cmd.Flags().Set("verbose", "true")
+					if err := cmd.Flags().Set("verbose", "true"); err != nil {
+						return err
+					}
 					fmt.Println("Set 'verbose' mode.")
 				case "quiet":
-					cmd.Flags().Set("verbose", "false")
+					if err := cmd.Flags().Set("verbose", "false"); err != nil {
+						return err
+					}
 					fmt.Println("Set 'quiet' mode.")
 				case "format":
 					if len(args) < 3 || args[2] != "json" {

cmd/interactive_test.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/stretchr/testify/assert"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )
 
 func TestExtractFilenames(t *testing.T) {

cmd/start_darwin.go

@@ -7,7 +7,7 @@ import (
 	"os/exec"
 	"strings"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {

cmd/start_default.go

@@ -6,7 +6,7 @@ import (
 	"context"
 	"fmt"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {

cmd/start_windows.go

@@ -10,7 +10,7 @@ import (
 	"strings"
 	"syscall"
 
-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )
 
 func startApp(ctx context.Context, client *api.Client) error {

convert/convert.go

@@ -1,23 +1,20 @@
 package convert
 
 import (
-	"bytes"
 	"cmp"
 	"encoding/binary"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
-	"regexp"
 	"slices"
+	"strings"
 
-	"github.com/mitchellh/mapstructure"
 	"google.golang.org/protobuf/proto"
 
-	"github.com/jmorganca/ollama/convert/sentencepiece"
-	"github.com/jmorganca/ollama/llm"
+	"github.com/ollama/ollama/convert/sentencepiece"
+	"github.com/ollama/ollama/llm"
 )
 
 type Params struct {
@@ -30,137 +27,58 @@ type Params struct {
 	AttentionHeads   int      `json:"num_attention_heads"` // n_head
 	KeyValHeads      int      `json:"num_key_value_heads"`
 	NormEPS          float64  `json:"rms_norm_eps"`
-	RopeFreqBase     float64  `json:"rope_theta"`
 	BoSTokenID       int      `json:"bos_token_id"`
 	EoSTokenID       int      `json:"eos_token_id"`
+	HeadDimension    int      `json:"head_dim"`
+	PaddingTokenID   int      `json:"pad_token_id"`
+
+	ByteOrder
 }
 
-type MetaData struct {
-	Type    string `mapstructure:"dtype"`
-	Shape   []int  `mapstructure:"shape"`
-	Offsets []int  `mapstructure:"data_offsets"`
+type ByteOrder interface {
+	binary.ByteOrder
+	binary.AppendByteOrder
 }
 
-func ReadSafeTensors(fn string, offset uint64) ([]llm.Tensor, uint64, error) {
-	f, err := os.Open(fn)
-	if err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-	defer f.Close()
-
-	var jsonSize uint64
-	binary.Read(f, binary.LittleEndian, &jsonSize)
-
-	buf := make([]byte, jsonSize)
-	_, err = io.ReadFull(f, buf)
-	if err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-
-	d := json.NewDecoder(bytes.NewBuffer(buf))
-	d.UseNumber()
-	var parsed map[string]interface{}
-	if err = d.Decode(&parsed); err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-
-	var keys []string
-	for k := range parsed {
-		keys = append(keys, k)
-	}
-
-	slices.Sort(keys)
-
-	slog.Info("converting layers")
-
-	var tensors []llm.Tensor
-	for _, k := range keys {
-		vals := parsed[k].(map[string]interface{})
-		var data MetaData
-		if err = mapstructure.Decode(vals, &data); err != nil {
-			return []llm.Tensor{}, 0, err
-		}
-
-		var size uint64
-		var kind uint32
-		switch len(data.Shape) {
-		case 0:
-			// metadata
-			continue
-		case 1:
-			// convert to float32
-			kind = 0
-			size = uint64(data.Shape[0] * 4)
-		case 2:
-			// convert to float16
-			kind = 1
-			size = uint64(data.Shape[0] * data.Shape[1] * 2)
-		}
-
-		ggufName, err := GetTensorName(k)
-		if err != nil {
-			slog.Error("%v", err)
-			return []llm.Tensor{}, 0, err
-		}
-
-		shape := [4]uint64{1, 1, 1, 1}
-		for cnt, s := range data.Shape {
-			shape[cnt] = uint64(s)
-		}
-
-		t := llm.Tensor{
-			Name:          ggufName,
-			Kind:          kind,
-			Offset:        offset,
-			Shape:         shape[:],
-			FileName:      fn,
-			OffsetPadding: 8 + jsonSize,
-			FileOffsets:   []uint64{uint64(data.Offsets[0]), uint64(data.Offsets[1])},
-		}
-		slog.Debug(fmt.Sprintf("%v", t))
-		tensors = append(tensors, t)
-		offset += size
-	}
-	return tensors, offset, nil
+type ModelArch interface {
+	GetTensors() error
+	LoadVocab() error
+	WriteGGUF() (string, error)
 }
 
-func GetSafeTensors(dirpath string) ([]llm.Tensor, error) {
-	var tensors []llm.Tensor
-	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
-	if err != nil {
-		return []llm.Tensor{}, err
-	}
-
-	var offset uint64
-	for _, f := range files {
-		var t []llm.Tensor
-		var err error
-		t, offset, err = ReadSafeTensors(f, offset)
-		if err != nil {
-			slog.Error("%v", err)
-			return []llm.Tensor{}, err
-		}
-		tensors = append(tensors, t...)
-	}
-	return tensors, nil
+type ModelFormat interface {
+	GetLayerName(string) (string, error)
+	GetTensors(string, *Params) ([]llm.Tensor, error)
+	GetParams(string) (*Params, error)
+	GetModelArch(string, string, *Params) (ModelArch, error)
 }
 
-func GetParams(dirpath string) (*Params, error) {
-	f, err := os.Open(filepath.Join(dirpath, "config.json"))
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
+type ModelData struct {
+	Path    string
+	Name    string
+	Params  *Params
+	Vocab   *Vocab
+	Tensors []llm.Tensor
+	Format  ModelFormat
+}
 
-	var params Params
-
-	d := json.NewDecoder(f)
-	err = d.Decode(&params)
+func GetModelFormat(dirname string) (ModelFormat, error) {
+	files, err := filepath.Glob(filepath.Join(dirname, "*"))
 	if err != nil {
 		return nil, err
 	}
 
-	return &params, nil
+	for _, fn := range files {
+		slog.Debug(fmt.Sprintf("file = %s", fn))
+		if strings.HasSuffix(fn, ".safetensors") {
+			return &SafetensorFormat{}, nil
+		} else if strings.HasSuffix(fn, ".bin") {
+			slog.Debug("model is torch")
+			return &TorchFormat{}, nil
+		}
+	}
+
+	return nil, fmt.Errorf("couldn't determine model format")
 }
 
 // Details on gguf's tokenizer can be found at:
@@ -171,7 +89,7 @@ type Vocab struct {
 	Types  []int32
 }
 
-func LoadTokens(dirpath string) (*Vocab, error) {
+func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 	slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
 	in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
 	if err != nil {
@@ -196,6 +114,14 @@ func LoadTokens(dirpath string) (*Vocab, error) {
 		v.Tokens = append(v.Tokens, p.GetPiece())
 		v.Scores = append(v.Scores, p.GetScore())
 		t := p.GetType()
+		switch t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN:
+		case sentencepiece.ModelProto_SentencePiece_CONTROL:
+		case sentencepiece.ModelProto_SentencePiece_UNUSED:
+		case sentencepiece.ModelProto_SentencePiece_BYTE:
+		default:
+			t = sentencepiece.ModelProto_SentencePiece_NORMAL
+		}
 		v.Types = append(v.Types, int32(t))
 	}
 
@@ -243,89 +169,15 @@ func LoadTokens(dirpath string) (*Vocab, error) {
 	}
 	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
 
-	return v, nil
-}
-
-func GetTensorName(n string) (string, error) {
-	tMap := map[string]string{
-		"model.embed_tokens.weight":                           "token_embd.weight",
-		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
-		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
-		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
-		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
-		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
-		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
-		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
-		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
-		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
-		"lm_head.weight":    "output.weight",
-		"model.norm.weight": "output_norm.weight",
-	}
-
-	v, ok := tMap[n]
-	if ok {
-		return v, nil
-	}
-
-	// quick hack to rename the layers to gguf format
-	for k, v := range tMap {
-		re := regexp.MustCompile(k)
-		newName := re.ReplaceAllString(n, v)
-		if newName != n {
-			return newName, nil
+	if params.VocabSize > len(v.Tokens) {
+		missingTokens := params.VocabSize - len(v.Tokens)
+		slog.Warn(fmt.Sprintf("vocab is missing %d tokens", missingTokens))
+		for cnt := 0; cnt < missingTokens; cnt++ {
+			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
+			v.Scores = append(v.Scores, -1)
+			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
 		}
 	}
 
-	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
-}
-
-func WriteGGUF(name string, tensors []llm.Tensor, params *Params, vocab *Vocab) (string, error) {
-	c := llm.ContainerGGUF{
-		ByteOrder: binary.LittleEndian,
-	}
-
-	m := llm.NewGGUFModel(&c)
-	m.Tensors = tensors
-	m.KV["general.architecture"] = "llama"
-	m.KV["general.name"] = name
-	m.KV["llama.context_length"] = uint32(params.ContextSize)
-	m.KV["llama.embedding_length"] = uint32(params.HiddenSize)
-	m.KV["llama.block_count"] = uint32(params.HiddenLayers)
-	m.KV["llama.feed_forward_length"] = uint32(params.IntermediateSize)
-	m.KV["llama.rope.dimension_count"] = uint32(128)
-	m.KV["llama.attention.head_count"] = uint32(params.AttentionHeads)
-	m.KV["llama.attention.head_count_kv"] = uint32(params.KeyValHeads)
-	m.KV["llama.attention.layer_norm_rms_epsilon"] = float32(params.NormEPS)
-	m.KV["llama.rope.freq_base"] = float32(params.RopeFreqBase)
-	m.KV["general.file_type"] = uint32(1)
-	m.KV["tokenizer.ggml.model"] = "llama"
-
-	m.KV["tokenizer.ggml.tokens"] = vocab.Tokens
-	m.KV["tokenizer.ggml.scores"] = vocab.Scores
-	m.KV["tokenizer.ggml.token_type"] = vocab.Types
-
-	m.KV["tokenizer.ggml.bos_token_id"] = uint32(params.BoSTokenID)
-	m.KV["tokenizer.ggml.eos_token_id"] = uint32(params.EoSTokenID)
-	m.KV["tokenizer.ggml.unknown_token_id"] = uint32(0)
-	m.KV["tokenizer.ggml.add_bos_token"] = true
-	m.KV["tokenizer.ggml.add_eos_token"] = false
-
-	// llamacpp sets the chat template, however we don't need to set it since we pass it in through a layer
-	// m.KV["tokenizer.chat_template"] = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" // XXX removeme
-
-	c.V3.NumTensor = uint64(len(tensors))
-	c.V3.NumKV = uint64(len(m.KV))
-
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	err = m.Encode(f)
-	if err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
+	return v, nil
 }

convert/gemma.go

@@ -0,0 +1,137 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type GemmaModel struct {
+	ModelData
+}
+
+func gemmaLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	slog.Debug(fmt.Sprintf("converting '%s'", r.t.Name))
+
+	data := make([]byte, r.end-r.start)
+	if err := binary.Read(f, r.bo, data); err != nil {
+		return err
+	}
+
+	tDataF32 := bfloat16.DecodeFloat32(data)
+
+	var err error
+	tDataF32, err = addOnes(tDataF32, int(r.t.Shape[0]))
+	if err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, r.bo, tDataF32); err != nil {
+		return err
+	}
+	return nil
+}
+
+func addOnes(data []float32, vectorSize int) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, vectorSize)
+
+	var err error
+	n, err = n.Add(ones)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	newN, err := native.SelectF32(n, 0)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	var fullTensor []float32
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+
+	return fullTensor, nil
+}
+
+func (m *GemmaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
+
+	m.Tensors = []llm.Tensor{}
+	for _, l := range t {
+		if strings.HasSuffix(l.Name, "norm.weight") {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = gemmaLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *GemmaModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *GemmaModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "gemma",
+		"general.name":                           m.Name,
+		"gemma.context_length":                   uint32(m.Params.ContextSize),
+		"gemma.embedding_length":                 uint32(m.Params.HiddenSize),
+		"gemma.block_count":                      uint32(m.Params.HiddenLayers),
+		"gemma.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"gemma.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"gemma.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"gemma.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"gemma.attention.key_length":             uint32(m.Params.HeadDimension),
+		"gemma.attention.value_length":           uint32(m.Params.HeadDimension),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.padding_token_id": uint32(m.Params.PaddingTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(3),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}

convert/llama.go

@@ -0,0 +1,176 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type LlamaModel struct {
+	ModelData
+}
+
+func llamaLayerHandler(w io.Writer, r torchWriterTo) error {
+	slog.Debug(fmt.Sprintf("repacking layer '%s'", r.t.Name))
+
+	data := r.storage.(*pytorch.HalfStorage).Data
+	tData := make([]uint16, len(data))
+	for cnt, v := range data {
+		tData[cnt] = uint16(float16.Fromfloat32(v))
+	}
+
+	var err error
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	slog.Debug(fmt.Sprintf("heads = %d", heads))
+
+	tData, err = llamaRepack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	if err = binary.Write(w, r.bo, tData); err != nil {
+		return err
+	}
+	return nil
+}
+
+func llamaRepack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *LlamaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			slog.Debug(fmt.Sprintf("setting handler for: %s", l.Name))
+			wt := l.WriterTo.(torchWriterTo)
+			wt.handler = llamaLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *LlamaModel) LoadVocab() error {
+	var v *Vocab
+	var err error
+
+	slog.Debug("loading vocab")
+	v, err = LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug("vocab loaded")
+
+	m.Vocab = v
+	return nil
+}
+
+func (m *LlamaModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.vocab_size":                       uint32(len(m.Vocab.Tokens)),
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	slog.Debug(fmt.Sprintf("gguf file = %s", f.Name()))
+
+	return f.Name(), nil
+}

convert/mistral.go

@@ -0,0 +1,173 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MistralModel struct {
+	ModelData
+}
+
+func mistralLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	layerSize := r.end - r.start
+
+	var err error
+	tData := make([]uint16, layerSize/2)
+	if err = binary.Read(f, r.bo, tData); err != nil {
+		return err
+	}
+
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	tData, err = repack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	var buf []byte
+	for _, n := range tData {
+		buf = r.bo.AppendUint16(buf, n)
+	}
+
+	tempBuf := make([]uint16, len(tData))
+	tDataF32 := bfloat16.DecodeFloat32(buf)
+	for cnt, v := range tDataF32 {
+		tDataF16 := float16.Fromfloat32(v)
+		tempBuf[cnt] = uint16(tDataF16)
+	}
+
+	if err = binary.Write(w, r.bo, tempBuf); err != nil {
+		return err
+	}
+	return nil
+}
+
+func repack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *MistralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = mistralLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *MistralModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *MistralModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}

convert/safetensors.go

@@ -0,0 +1,304 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"slices"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/mitchellh/mapstructure"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type safetensorWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	filename string
+
+	start, end, padding uint64
+	handler             func(w io.Writer, r safetensorWriterTo, f *os.File) error
+}
+
+type tensorMetaData struct {
+	Type    string `mapstructure:"dtype"`
+	Shape   []int  `mapstructure:"shape"`
+	Offsets []int  `mapstructure:"data_offsets"`
+}
+
+type SafetensorFormat struct{}
+
+func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting tensor data")
+	var tensors []llm.Tensor
+	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
+	if err != nil {
+		return nil, err
+	}
+
+	var offset uint64
+	for _, f := range files {
+		var t []llm.Tensor
+		var err error
+		t, offset, err = m.readTensors(f, offset, params)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, err
+		}
+		tensors = append(tensors, t...)
+	}
+	slog.Debug(fmt.Sprintf("all tensors = %d", len(tensors)))
+	return tensors, nil
+}
+
+func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params) ([]llm.Tensor, uint64, error) {
+	f, err := os.Open(fn)
+	if err != nil {
+		return nil, 0, err
+	}
+	defer f.Close()
+
+	var jsonSize uint64
+	if err := binary.Read(f, binary.LittleEndian, &jsonSize); err != nil {
+		return nil, 0, err
+	}
+
+	buf := make([]byte, jsonSize)
+	_, err = io.ReadFull(f, buf)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	d := json.NewDecoder(bytes.NewBuffer(buf))
+	d.UseNumber()
+	var parsed map[string]interface{}
+	if err = d.Decode(&parsed); err != nil {
+		return nil, 0, err
+	}
+
+	var keys []string
+	for k := range parsed {
+		keys = append(keys, k)
+	}
+
+	slices.Sort(keys)
+
+	slog.Info("converting layers")
+
+	var tensors []llm.Tensor
+	for _, k := range keys {
+		vals := parsed[k].(map[string]interface{})
+		var data tensorMetaData
+		if err = mapstructure.Decode(vals, &data); err != nil {
+			slog.Error("couldn't decode properly")
+			return nil, 0, err
+		}
+
+		slog.Debug(fmt.Sprintf("metadata = %#v", data))
+		var size uint64
+		var kind uint32
+		switch len(data.Shape) {
+		case 0:
+			// metadata
+			continue
+		case 1:
+			// convert to float32
+			kind = 0
+			size = uint64(data.Shape[0] * 4)
+		case 2:
+			// convert to float16
+			kind = 1
+			size = uint64(data.Shape[0] * data.Shape[1] * 2)
+		}
+
+		ggufName, err := m.GetLayerName(k)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, 0, err
+		}
+
+		shape := []uint64{0, 0, 0, 0}
+		for i := range data.Shape {
+			shape[i] = uint64(data.Shape[i])
+		}
+
+		t := llm.Tensor{
+			Name:   ggufName,
+			Kind:   kind,
+			Offset: offset,
+			Shape:  shape[:],
+		}
+
+		t.WriterTo = safetensorWriterTo{
+			t:        &t,
+			params:   params,
+			bo:       params.ByteOrder,
+			filename: fn,
+			start:    uint64(data.Offsets[0]),
+			end:      uint64(data.Offsets[1]),
+			padding:  8 + jsonSize,
+		}
+
+		tensors = append(tensors, t)
+		offset += size
+	}
+	slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
+	slog.Debug(fmt.Sprintf("offset = %d", offset))
+	return tensors, offset, nil
+}
+
+func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var params Params
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	tMap := map[string]string{
+		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range tMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	f, err := os.Open(r.filename)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+
+	if _, err = f.Seek(int64(r.padding+r.start), 0); err != nil {
+		return 0, err
+	}
+
+	// use the handler if one is present
+	if r.handler != nil {
+		return 0, r.handler(w, r, f)
+	}
+
+	remaining := r.end - r.start
+
+	bufSize := uint64(10240)
+	var finished bool
+	for {
+		data := make([]byte, min(bufSize, remaining))
+
+		b, err := io.ReadFull(f, data)
+		remaining -= uint64(b)
+
+		if err == io.EOF || remaining <= 0 {
+			finished = true
+		} else if err != nil {
+			return 0, err
+		}
+
+		// convert bfloat16 -> ieee float32
+		tDataF32 := bfloat16.DecodeFloat32(data)
+
+		switch r.t.Kind {
+		case 0:
+			if err := binary.Write(w, r.bo, tDataF32); err != nil {
+				return 0, err
+			}
+		case 1:
+			// convert float32 -> float16
+			tempBuf := make([]uint16, len(data)/2)
+			for cnt, v := range tDataF32 {
+				tDataF16 := float16.Fromfloat32(v)
+				tempBuf[cnt] = uint16(tDataF16)
+			}
+			if err := binary.Write(w, r.bo, tempBuf); err != nil {
+				return 0, err
+			}
+		}
+		if finished {
+			break
+		}
+	}
+	return 0, nil
+}
+
+func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "MistralForCausalLM":
+			return &MistralModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "GemmaForCausalLM":
+			return &GemmaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}

convert/torch.go

@@ -0,0 +1,286 @@
+package convert
+
+import (
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/nlpodyssey/gopickle/types"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type torchWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	storage pytorch.StorageInterface
+	handler func(w io.Writer, r torchWriterTo) error
+}
+
+type TorchFormat struct{}
+
+func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting torch tensors")
+
+	files, err := filepath.Glob(filepath.Join(dirpath, "pytorch_model-*.bin"))
+	if err != nil {
+		slog.Error("didn't find any torch files")
+		return nil, err
+	}
+
+	var offset uint64
+
+	var tensors []llm.Tensor
+	for _, fn := range files {
+		m, err := pytorch.Load(fn)
+		if err != nil {
+			slog.Error(fmt.Sprintf("error unpickling: %q", err))
+			return []llm.Tensor{}, err
+		}
+
+		for _, k := range m.(*types.Dict).Keys() {
+			if strings.HasSuffix(k.(string), "self_attn.rotary_emb.inv_freq") {
+				continue
+			}
+
+			t, _ := m.(*types.Dict).Get(k)
+			tshape := t.(*pytorch.Tensor).Size
+
+			var size uint64
+			var kind uint32
+			switch len(tshape) {
+			case 0:
+				continue
+			case 1:
+				// convert to float32
+				kind = 0
+				size = uint64(tshape[0] * 4)
+			case 2:
+				// convert to float16
+				kind = 1
+				size = uint64(tshape[0] * tshape[1] * 2)
+			}
+
+			ggufName, err := tf.GetLayerName(k.(string))
+			if err != nil {
+				slog.Error("%v", err)
+				return nil, err
+			}
+			slog.Debug(fmt.Sprintf("finding name for '%s' -> '%s'", k.(string), ggufName))
+
+			shape := []uint64{0, 0, 0, 0}
+			for i := range tshape {
+				shape[i] = uint64(tshape[i])
+			}
+
+			tensor := llm.Tensor{
+				Name:   ggufName,
+				Kind:   kind,
+				Offset: offset, // calculate the offset
+				Shape:  shape[:],
+			}
+
+			tensor.WriterTo = torchWriterTo{
+				t:       &tensor,
+				params:  params,
+				bo:      params.ByteOrder,
+				storage: t.(*pytorch.Tensor).Source,
+			}
+
+			tensors = append(tensors, tensor)
+			offset += size
+		}
+	}
+
+	return tensors, nil
+
+}
+
+func getAltParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "params.json"))
+	if err != nil {
+		slog.Error("no params.json")
+		return nil, err
+	}
+	defer f.Close()
+
+	type TorchParams struct {
+		HiddenSize     int     `json:"dim"`
+		AttentionHeads int     `json:"n_heads"`
+		KeyValHeads    int     `json:"n_kv_heads"`
+		HiddenLayers   int     `json:"n_layers"`
+		RopeTheta      int     `json:"rope_theta"`
+		NormEPS        float64 `json:"norm_eps"`
+	}
+
+	var tparams TorchParams
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&tparams)
+	if err != nil {
+		return nil, err
+	}
+
+	params := &Params{
+		HiddenSize:     tparams.HiddenSize,
+		AttentionHeads: tparams.AttentionHeads,
+		KeyValHeads:    tparams.KeyValHeads,
+		HiddenLayers:   tparams.HiddenLayers,
+		NormEPS:        tparams.NormEPS,
+	}
+
+	switch {
+	case tparams.RopeTheta == 1000000:
+		// Codellama
+		params.ContextSize = 16384
+	case tparams.NormEPS == 1e-06:
+		// llama2
+		slog.Debug("Found llama2 - setting context size to 4096")
+		params.ContextSize = 4096
+	default:
+		params.ContextSize = 2048
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return params, nil
+}
+
+func (m *TorchFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		if os.IsNotExist(err) {
+			// try params.json instead
+			return getAltParams(dirpath)
+		} else {
+			return nil, err
+		}
+	}
+
+	var params Params
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *TorchFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"tok_embeddings.weight":     "token_embd.weight",
+		"output.weight":             "output.weight",
+		"norm.weight":               "output_norm.weight",
+		"rope.freqs":                "rope_freqs.weight",
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	lMap := map[string]string{
+		"layers.(\\d+).attention_norm.weight":                 "blk.$1.attn_norm.weight",
+		"layers.(\\d+).attention_output_norm.weight":          "blk.$1.attn_norm.weight",
+		"layers.(\\d+).feed_forward.w2.weight":                "blk.$1.ffn_down.weight",
+		"layers.(\\d+).feed_forward.w1.weight":                "blk.$1.ffn_gate.weight",
+		"layers.(\\d+).feed_forward.w3.weight":                "blk.$1.ffn_up.weight",
+		"layers.(\\d+).ffn_norm.weight":                       "blk.$1.ffn_norm.weight",
+		"layers.(\\d+).attention.wk.weight":                   "blk.$1.attn_k.weight",
+		"layers.(\\d+).attention.wo.weight":                   "blk.$1.attn_output.weight",
+		"layers.(\\d+).attention.wq.weight":                   "blk.$1.attn_q.weight",
+		"layers.(\\d+).attention.wv.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range lMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r torchWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	// use the handler if one is present
+	if r.handler != nil {
+		return 0, r.handler(w, r)
+	}
+
+	switch r.storage.(type) {
+	case *pytorch.FloatStorage:
+		slog.Warn(fmt.Sprintf("unexpected storage found for layer '%s'; skipping", r.t.Name))
+		return 0, nil
+	case *pytorch.HalfStorage:
+		switch r.t.Kind {
+		case 0:
+			data := r.storage.(*pytorch.HalfStorage).Data
+			slog.Debug(fmt.Sprintf("%35s F32 (%d)", r.t.Name, len(data)))
+			if err := binary.Write(w, r.bo, data); err != nil {
+				return 0, err
+			}
+		case 1:
+			data := r.storage.(*pytorch.HalfStorage).Data
+			tData := make([]uint16, len(data))
+			for cnt, v := range data {
+				tData[cnt] = uint16(float16.Fromfloat32(v))
+			}
+			slog.Debug(fmt.Sprintf("%35s F16 (%d)", r.t.Name, len(tData)))
+			if err := binary.Write(w, r.bo, tData); err != nil {
+				return 0, err
+			}
+		}
+	}
+
+	return 0, nil
+}
+
+func (m *TorchFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "LlamaForCausalLM":
+			return &LlamaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}

docs/README.md

@@ -3,7 +3,7 @@
 ### Getting Started
 * [Quickstart](../README.md#quickstart)
 * [Examples](../examples)
-* [Importing models](./import.md) from GGUF, Pytorch and Safetensors
+* [Importing models](./import.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
 * [Docker Documentation](https://hub.docker.com/r/ollama/ollama)

docs/api.md

@@ -394,7 +394,6 @@ Advanced parameters (optional):
 
 - `format`: the format to return a response in. Currently the only accepted value is `json`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
-- `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 

docs/development.md

@@ -69,7 +69,7 @@ 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/deploy/linux/quick_start.html) development packages first, as well as `cmake` 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
@@ -116,29 +116,30 @@ Note: The windows build for Ollama is still under development.
 
 Install required tools:
 
-- MSVC toolchain - C/C++ and cmake as minimal requirements - You must build from a "Developer Shell" with the environment variables set
-- go version 1.22 or higher
+- MSVC toolchain - C/C++ and cmake as minimal requirements
+- Go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
-  - <https://www.mingw-w64.org/>
-  - <https://www.msys2.org/>
+  - [MinGW-w64](https://www.mingw-w64.org/)
+  - [MSYS2](https://www.msys2.org/)
 
 ```powershell
 $env:CGO_ENABLED="1"
-
 go generate ./...
-
 go build .
 ```
 
 #### Windows CUDA (NVIDIA)
 
-In addition to the common Windows development tools described above, install CUDA **AFTER** you install MSVC.
+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 described above, install AMDs HIP package **AFTER** you install MSVC
+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)
+- [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`).

docs/faq.md

@@ -14,6 +14,10 @@ curl -fsSL https://ollama.com/install.sh | sh
 
 Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.
 
+## Is my GPU compatible with Ollama?
+
+Please refer to the [GPU docs](./gpu.md).
+
 ## How can I specify the context window size?
 
 By default, Ollama uses a context window size of 2048 tokens.
@@ -95,6 +99,37 @@ Ollama binds 127.0.0.1 port 11434 by default. Change the bind address with the `
 
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
+## How can I use Ollama with a proxy server?
+
+Ollama runs an HTTP server and can be exposed using a proxy server such as Nginx. To do so, configure the proxy to forward requests and optionally set required headers (if not exposing Ollama on the network). For example, with Nginx:
+
+```
+server {
+    listen 80;
+    server_name example.com;  # Replace with your domain or IP
+    location / {
+        proxy_pass http://localhost:11434;
+        proxy_set_header Host localhost:11434;
+    }
+}
+```
+
+## How can I use Ollama with ngrok?
+
+Ollama can be accessed using a range of tools for tunneling tools. For example with Ngrok:
+
+```
+ngrok http 11434 --host-header="localhost:11434"
+```
+
+## How can I use Ollama with Cloudflare Tunnel?
+
+To use Ollama with Cloudflare Tunnel, use the `--url` and `--http-host-header` flags:
+
+```
+cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11434"
+```
+
 ## How can I allow additional web origins to access Ollama?
 
 Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
@@ -119,7 +154,7 @@ No. Ollama runs locally, and conversation data does not leave your machine.
 
 ## How can I use Ollama in Visual Studio Code?
 
-There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/jmorganca/ollama#extensions--plugins) at the bottom of the main repository readme.
+There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/ollama/ollama#extensions--plugins) at the bottom of the main repository readme.
 
 ## How do I use Ollama behind a proxy?
 

docs/gpu.md

@@ -0,0 +1,102 @@
+# GPU
+## Nvidia
+Ollama supports Nvidia GPUs with compute capability 5.0+.
+
+Check your compute compatibility to see if your card is supported:
+[https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)
+
+| Compute Capability | Family              | Cards                                                                                                       |
+| ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
+| 9.0                | NVIDIA              | `H100`                                                                                                      |
+| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080` `RTX 4070 Ti` `RTX 4060 Ti`                                                           |
+|                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
+| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
+|                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
+| 8.0                | NVIDIA              | `A100` `A30`                                                                                                |
+| 7.5                | GeForce GTX/RTX     | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060`                                    |
+|                    | NVIDIA Professional | `T4` `RTX 5000` `RTX 4000` `RTX 3000` `T2000` `T1200` `T1000` `T600` `T500`                                 |
+|                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
+| 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
+| 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
+|                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
+|                    | Tesla               | `P40` `P4`                                                                                                  |
+| 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |
+| 5.2                | GeForce GTX         | `GTX TITAN X` `GTX 980 Ti` `GTX 980` `GTX 970` `GTX 960` `GTX 950`                                          |
+|                    | Quadro              | `M6000 24GB` `M6000` `M5000` `M5500M` `M4000` `M2200` `M2000` `M620`                                        |
+|                    | Tesla               | `M60` `M40`                                                                                                 |
+| 5.0                | GeForce GTX         | `GTX 750 Ti` `GTX 750` `NVS 810`                                                                            |
+|                    | Quadro              | `K2200` `K1200` `K620` `M1200` `M520` `M5000M` `M4000M` `M3000M` `M2000M` `M1000M` `K620M` `M600M` `M500M`  |
+
+
+### GPU Selection
+
+If you have multiple NVIDIA GPUs in your system and want to limit Ollama to use
+a subset, you can set `CUDA_VISIBLE_DEVICES` to a comma separated list of GPUs.
+Numeric IDs may be used, however ordering may vary, so UUIDs are more reliable.
+You can discover the UUID of your GPUs by running `nvidia-smi -L` If you want to
+ignore the GPUs and force CPU usage, use an invalid GPU ID (e.g., "-1")
+
+### Laptop Suspend Resume
+
+On linux, after a suspend/resume cycle, sometimes Ollama will fail to discover
+your NVIDIA GPU, and fallback to running on the CPU.  You can workaround this
+driver bug by reloading the NVIDIA UVM driver with `sudo rmmod nvidia_uvm &&
+sudo modprobe nvidia_uvm`
+
+## AMD Radeon
+Ollama supports the following AMD GPUs:
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
+| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
+
+### Overrides
+Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
+some cases you can force the system to try to use a similar LLVM target that is
+close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
+however, ROCm does not currently support this target. The closest support is
+`gfx1030`.  You can use the environment variable `HSA_OVERRIDE_GFX_VERSION` with
+`x.y.z` syntax.  So for example, to force the system to run on the RX 5400, you
+would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
+server.  If you have an unsupported AMD GPU you can experiment using the list of
+supported types below.
+
+At this time, the known supported GPU types are the following LLVM Targets.
+This table shows some example GPUs that map to these LLVM targets:
+| **LLVM Target** | **An Example GPU** |
+|-----------------|---------------------|
+| gfx900 | Radeon RX Vega 56 |
+| gfx906 | Radeon Instinct MI50 |
+| gfx908 | Radeon Instinct MI100 |
+| gfx90a | Radeon Instinct MI210 |
+| gfx940 | Radeon Instinct MI300 |
+| gfx941 | |
+| gfx942 | |
+| gfx1030 | Radeon PRO V620 |
+| gfx1100 | Radeon PRO W7900 |
+| gfx1101 | Radeon PRO W7700 |
+| gfx1102 | Radeon RX 7600 |
+
+AMD is working on enhancing ROCm v6 to broaden support for families of GPUs in a
+future release which should increase support for more GPUs.
+
+Reach out on [Discord](https://discord.gg/ollama) or file an
+[issue](https://github.com/ollama/ollama/issues) for additional help.
+
+### GPU Selection
+
+If you have multiple AMD GPUs in your system and want to limit Ollama to use a
+subset, you can set `HIP_VISIBLE_DEVICES` to a comma separated list of GPUs.
+You can see the list of devices with `rocminfo`.  If you want to ignore the GPUs
+and force CPU usage, use an invalid GPU ID (e.g., "-1")
+
+### Container Permission
+
+In some Linux distributions, SELinux can prevent containers from
+accessing the AMD GPU devices.  On the host system you can run 
+`sudo setsebool container_use_devices=1` to allow containers to use devices.
+
+### Metal (Apple GPUs)
+Ollama supports GPU acceleration on Apple devices via the Metal API.

docs/modelfile.md

@@ -113,7 +113,7 @@ FROM llama2
 ```
 
 A list of available base models:
-<https://github.com/jmorganca/ollama#model-library>
+<https://github.com/ollama/ollama#model-library>
 
 #### Build from a `bin` file
 
@@ -131,7 +131,7 @@ The `PARAMETER` instruction defines a parameter that can be set when the model i
 PARAMETER <parameter> <parametervalue>
 ```
 
-### Valid Parameters and Values
+#### Valid Parameters and Values
 
 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
@@ -139,9 +139,6 @@ PARAMETER <parameter> <parametervalue>
 | mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
 | mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
-| num_gqa        | The number of GQA groups in the transformer layer. Required for some models, for example it is 8 for llama2:70b                                                                                                                                         | int        | num_gqa 1            |
-| num_gpu        | The number of layers to send to the GPU(s). On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                            | int        | num_gpu 50           |
-| num_thread     | Sets the number of threads to use during computation. By default, Ollama will detect this for optimal performance. It is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). | int        | num_thread 8         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
@@ -183,7 +180,7 @@ SYSTEM """<system message>"""
 
 ### ADAPTER
 
-The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+The `ADAPTER` instruction is an optional instruction that specifies any LoRA adapter that should apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
 
 ```modelfile
 ADAPTER ./ollama-lora.bin
@@ -201,7 +198,22 @@ LICENSE """
 
 ### MESSAGE
 
-The `MESSAGE` instruction allows you to specify a message history for the model to use when responding:
+The `MESSAGE` instruction allows you to specify a message history for the model to use when responding. Use multiple iterations of the MESSAGE command to build up a conversation which will guide the model to answer in a similar way.
+
+```modelfile
+MESSAGE <role> <message>
+```
+
+#### Valid roles
+
+| Role      | Description                                                  |
+| --------- | ------------------------------------------------------------ |
+| system    | Alternate way of providing the SYSTEM message for the model. |
+| user      | An example message of what the user could have asked.        |
+| assistant | An example message of how the model should respond.          |
+
+
+#### Example conversation
 
 ```modelfile
 MESSAGE user Is Toronto in Canada?
@@ -212,6 +224,7 @@ MESSAGE user Is Ontario in Canada?
 MESSAGE assistant yes
 ```
 
+
 ## Notes
 
 - the **`Modelfile` is not case sensitive**. In the examples, uppercase instructions are used to make it easier to distinguish it from arguments.

docs/openai.md

@@ -1,6 +1,6 @@
 # OpenAI compatibility
 
-> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/jmorganca/ollama/blob/main/docs/api.md).
+> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).
 
 Ollama provides experimental compatibility with parts of the [OpenAI API](https://platform.openai.com/docs/api-reference) to help connect existing applications to Ollama.
 

docs/troubleshooting.md

@@ -67,49 +67,19 @@ You can see what features your CPU has with the following.
 cat /proc/cpuinfo| grep flags  | head -1
 ```
 
-## AMD Radeon GPU Support
+## Installing older or pre-release versions on Linux
 
-Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
-some cases you can force the system to try to use a similar LLVM target that is
-close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
-however, ROCm does not currently support this target. The closest support is
-`gfx1030`.  You can use the environment variable `HSA_OVERRIDE_GFX_VERSION` with
-`x.y.z` syntax.  So for example, to force the system to run on the RX 5400, you
-would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
-server.  If you have an unsupported AMD GPU you can experiment using the list of
-supported types below.
-
-At this time, the known supported GPU types are the following LLVM Targets.
-This table shows some example GPUs that map to these LLVM targets:
-| **LLVM Target** | **An Example GPU** |
-|-----------------|---------------------|
-| gfx900 | Radeon RX Vega 56 |
-| gfx906 | Radeon Instinct MI50 |
-| gfx908 | Radeon Instinct MI100 |
-| gfx90a | Radeon Instinct MI210 |
-| gfx940 | Radeon Instinct MI300 |
-| gfx941 | |
-| gfx942 | |
-| gfx1030 | Radeon PRO V620 |
-| gfx1100 | Radeon PRO W7900 |
-| gfx1101 | Radeon PRO W7700 |
-| gfx1102 | Radeon RX 7600 |
-
-AMD is working on enhancing ROCm v6 to broaden support for families of GPUs in a
-future release which should increase support for more GPUs.
-
-Reach out on [Discord](https://discord.gg/ollama) or file an
-[issue](https://github.com/ollama/ollama/issues) for additional help.
-
-## Installing older versions on Linux
-
-If you run into problems on Linux and want to install an older version you can tell the install script
-which version to install.
+If you run into problems on Linux and want to install an older version, or you'd
+like to try out a pre-release before it's officially released, you can tell the
+install script which version to install.
 
 ```sh
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.27" sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 ```
 
-## Known issues
+## Linux tmp noexec 
 
-* N/A
+If your system is configured with the "noexec" flag where Ollama stores its
+temporary executable files, you can specify an alternate location by setting
+OLLAMA_TMPDIR to a location writable by the user ollama runs as.  For example
+OLLAMA_TMPDIR=/usr/share/ollama/

docs/tutorials/langchainjs.md

@@ -18,7 +18,7 @@ const ollama = new Ollama({
   model: "llama2",
 });
 
-const answer = await ollama.call(`why is the sky blue?`);
+const answer = await ollama.invoke(`why is the sky blue?`);
 
 console.log(answer);
 ```

examples/go-chat/main.go

@@ -0,0 +1,51 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	messages := []api.Message{
+		api.Message{
+			Role:    "system",
+			Content: "Provide very brief, concise responses",
+		},
+		api.Message{
+			Role:    "user",
+			Content: "Name some unusual animals",
+		},
+		api.Message{
+			Role:    "assistant",
+			Content: "Monotreme, platypus, echidna",
+		},
+		api.Message{
+			Role:    "user",
+			Content: "which of these is the most dangerous?",
+		},
+	}
+
+	ctx := context.Background()
+	req := &api.ChatRequest{
+		Model:    "llama2",
+		Messages: messages,
+	}
+
+	respFunc := func(resp api.ChatResponse) error {
+		fmt.Print(resp.Message.Content)
+		return nil
+	}
+
+	err = client.Chat(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}

examples/go-generate-streaming/main.go

@@ -0,0 +1,40 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	// By default, GenerateRequest is streaming.
+	req := &api.GenerateRequest{
+		Model:  "gemma",
+		Prompt: "how many planets are there?",
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// Only print the response here; GenerateResponse has a number of other
+		// interesting fields you want to examine.
+
+		// In streaming mode, responses are partial so we call fmt.Print (and not
+		// Println) in order to avoid spurious newlines being introduced. The
+		// model will insert its own newlines if it wants.
+		fmt.Print(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println()
+}

examples/go-generate/main.go

@@ -0,0 +1,37 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	req := &api.GenerateRequest{
+		Model:  "gemma",
+		Prompt: "how many planets are there?",
+
+		// set streaming to false
+		Stream: new(bool),
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// Only print the response here; GenerateResponse has a number of other
+		// interesting fields you want to examine.
+		fmt.Println(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}

examples/go-multimodal/main.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	if len(os.Args) <= 1 {
+		log.Fatal("usage: <image name>")
+	}
+
+	imgData, err := os.ReadFile(os.Args[1])
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	req := &api.GenerateRequest{
+		Model:  "llava",
+		Prompt: "describe this image",
+		Images: []api.ImageData{imgData},
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// In streaming mode, responses are partial so we call fmt.Print (and not
+		// Println) in order to avoid spurious newlines being introduced. The
+		// model will insert its own newlines if it wants.
+		fmt.Print(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println()
+}

examples/go-pull-progress/main.go

@@ -0,0 +1,31 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	ctx := context.Background()
+
+	req := &api.PullRequest{
+		Model: "mistral",
+	}
+	progressFunc := func(resp api.ProgressResponse) error {
+		fmt.Printf("Progress: status=%v, total=%v, completed=%v\n", resp.Status, resp.Total, resp.Completed)
+		return nil
+	}
+
+	err = client.Pull(ctx, req, progressFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}

examples/langchain-python-rag-privategpt/README.md

@@ -1,6 +1,6 @@
 # PrivateGPT with Llama 2 uncensored
 
-https://github.com/jmorganca/ollama/assets/3325447/20cf8ec6-ff25-42c6-bdd8-9be594e3ce1b
+https://github.com/ollama/ollama/assets/3325447/20cf8ec6-ff25-42c6-bdd8-9be594e3ce1b
 
 > Note: this example is a slightly modified version of PrivateGPT using models such as Llama 2 Uncensored. All credit for PrivateGPT goes to Iván Martínez who is the creator of it, and you can find his GitHub repo [here](https://github.com/imartinez/privateGPT).
 

examples/modelfile-mario/readme.md

@@ -28,7 +28,7 @@ You are Mario from Super Mario Bros, acting as an assistant.
 What if you want to change its behaviour?
 
 - Try changing the prompt
-- Try changing the parameters [Docs](https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md)
+- Try changing the parameters [Docs](https://github.com/ollama/ollama/blob/main/docs/modelfile.md)
 - Try changing the model (e.g. An uncensored model by `FROM wizard-vicuna` this is the wizard-vicuna uncensored model )
 
 Once the changes are made,

examples/python-json-datagenerator/readme.md

@@ -1,6 +1,6 @@
 # JSON Output Example
 
-![llmjson 2023-11-10 15_31_31](https://github.com/jmorganca/ollama/assets/633681/e599d986-9b4a-4118-81a4-4cfe7e22da25)
+![llmjson 2023-11-10 15_31_31](https://github.com/ollama/ollama/assets/633681/e599d986-9b4a-4118-81a4-4cfe7e22da25)
 
 There are two python scripts in this example. `randomaddresses.py` generates random addresses from different countries. `predefinedschema.py` sets a template for the model to fill in.
 

examples/python-loganalysis/readme.md

@@ -1,6 +1,6 @@
 # Log Analysis example
 
-![loganalyzer 2023-11-10 08_53_29](https://github.com/jmorganca/ollama/assets/633681/ad30f1fc-321f-4953-8914-e30e24db9921)
+![loganalyzer 2023-11-10 08_53_29](https://github.com/ollama/ollama/assets/633681/ad30f1fc-321f-4953-8914-e30e24db9921)
 
 This example shows one possible way to create a log file analyzer. It uses the model **mattw/loganalyzer** which is based on **codebooga**, a 34b parameter model.
 

examples/typescript-functioncalling/readme.md

@@ -1,6 +1,6 @@
 # Function calling
 
-![function calling 2023-11-16 16_12_58](https://github.com/jmorganca/ollama/assets/633681/a0acc247-9746-45ab-b325-b65dfbbee4fb)
+![function calling 2023-11-16 16_12_58](https://github.com/ollama/ollama/assets/633681/a0acc247-9746-45ab-b325-b65dfbbee4fb)
 
 One of the features added to some models is 'function calling'. It's a bit of a confusing name. It's understandable if you think that means the model can call functions, but that's not what it means. Function calling simply means that the output of the model is formatted in JSON, using a preconfigured schema, and uses the expected types. Then your code can use the output of the model and call functions with it. Using the JSON format in Ollama, you can use any model for function calling. 
 

format/bytes.go

@@ -6,11 +6,15 @@ import (
 )
 
 const (
-	Byte     = 1
+	Byte = 1
+
 	KiloByte = Byte * 1000
 	MegaByte = KiloByte * 1000
 	GigaByte = MegaByte * 1000
 	TeraByte = GigaByte * 1000
+
+	KibiByte = Byte * 1024
+	MebiByte = KibiByte * 1024
 )
 
 func HumanBytes(b int64) string {
@@ -45,3 +49,14 @@ func HumanBytes(b int64) string {
 		return fmt.Sprintf("%d %s", int(value), unit)
 	}
 }
+
+func HumanBytes2(b uint64) string {
+	switch {
+	case b >= MebiByte:
+		return fmt.Sprintf("%.1f MiB", float64(b)/MebiByte)
+	case b >= KibiByte:
+		return fmt.Sprintf("%.1f KiB", float64(b)/KibiByte)
+	default:
+		return fmt.Sprintf("%d B", b)
+	}
+}

go.mod

@@ -1,4 +1,4 @@
-module github.com/jmorganca/ollama
+module github.com/ollama/ollama
 
 go 1.22
 
@@ -9,7 +9,7 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.9.1
-	github.com/golang/protobuf v1.5.0
+	github.com/golang/protobuf v1.5.0 // indirect
 	github.com/google/uuid v1.0.0
 	github.com/mitchellh/mapstructure v1.5.0
 	github.com/olekukonko/tablewriter v0.0.5
@@ -19,7 +19,10 @@ require (
 	golang.org/x/sync v0.3.0
 )
 
-require github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9
+require (
+	github.com/nlpodyssey/gopickle v0.3.0
+	github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9
+)
 
 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc // indirect
@@ -68,7 +71,7 @@ require (
 	golang.org/x/net v0.17.0 // indirect
 	golang.org/x/sys v0.13.0
 	golang.org/x/term v0.13.0
-	golang.org/x/text v0.13.0 // indirect
+	golang.org/x/text v0.14.0 // indirect
 	google.golang.org/protobuf v1.30.0
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -122,6 +122,8 @@ github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9Gz0M=
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
+github.com/nlpodyssey/gopickle v0.3.0 h1:BLUE5gxFLyyNOPzlXxt6GoHEMMxD0qhsE4p0CIQyoLw=
+github.com/nlpodyssey/gopickle v0.3.0/go.mod h1:f070HJ/yR+eLi5WmM1OXJEGaTpuJEUiib19olXgYha0=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
 github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9 h1:DV4iXjNn6fGeDl1AkZ1I0QB/0DBjrc7kPpxHrmuDzW4=
@@ -236,8 +238,8 @@ golang.org/x/term v0.13.0/go.mod h1:LTmsnFJwVN6bCy1rVCoS+qHT1HhALEFxKncY3WNNh4U=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.13.0 h1:ablQoSUd0tRdKxZewP80B+BaqeKJuVhuRxj/dkrun3k=
-golang.org/x/text v0.13.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
+golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
+golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=

gpu/amd_common.go

@@ -40,19 +40,17 @@ func amdSetVisibleDevices(ids []int, skip map[int]interface{}) {
 	// TODO - does sort order matter?
 	devices := []string{}
 	for i := range ids {
-		slog.Debug(fmt.Sprintf("i=%d", i))
 		if _, skipped := skip[i]; skipped {
-			slog.Debug("skipped")
 			continue
 		}
 		devices = append(devices, strconv.Itoa(i))
 	}
-	slog.Debug(fmt.Sprintf("devices=%v", devices))
 
 	val := strings.Join(devices, ",")
 	err := os.Setenv("HIP_VISIBLE_DEVICES", val)
 	if err != nil {
 		slog.Warn(fmt.Sprintf("failed to set env: %s", err))
+	} else {
+		slog.Info("Setting HIP_VISIBLE_DEVICES=" + val)
 	}
-	slog.Debug("HIP_VISIBLE_DEVICES=" + val)
 }

gpu/amd_linux.go

@@ -24,6 +24,9 @@ const (
 	GPUTotalMemoryFileGlob = "mem_banks/*/properties" // size_in_bytes line
 	GPUUsedMemoryFileGlob  = "mem_banks/*/used_memory"
 	RocmStandardLocation   = "/opt/rocm/lib"
+
+	// TODO find a better way to detect iGPU instead of minimum memory
+	IGPUMemLimit = 1024 * 1024 * 1024 // 512G is what they typically report, so anything less than 1G must be iGPU
 )
 
 var (
@@ -97,6 +100,8 @@ func AMDGetGPUInfo(resp *GpuInfo) {
 		return
 	}
 
+	updateLibPath(libDir)
+
 	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
 	if gfxOverride == "" {
 		supported, err := GetSupportedGFX(libDir)
@@ -110,7 +115,7 @@ func AMDGetGPUInfo(resp *GpuInfo) {
 			if !slices.Contains[[]string, string](supported, v.ToGFXString()) {
 				slog.Warn(fmt.Sprintf("amdgpu [%d] %s is not supported by %s %v", i, v.ToGFXString(), libDir, 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")
+				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
 				skip[i] = struct{}{}
 			} else {
 				slog.Info(fmt.Sprintf("amdgpu [%d] %s is supported", i, v.ToGFXString()))
@@ -140,14 +145,29 @@ func AMDGetGPUInfo(resp *GpuInfo) {
 	}
 }
 
+func updateLibPath(libDir string) {
+	ldPaths := []string{}
+	if val, ok := os.LookupEnv("LD_LIBRARY_PATH"); ok {
+		ldPaths = strings.Split(val, ":")
+	}
+	for _, d := range ldPaths {
+		if d == libDir {
+			return
+		}
+	}
+	val := strings.Join(append(ldPaths, libDir), ":")
+	slog.Debug("updated lib path", "LD_LIBRARY_PATH", val)
+	os.Setenv("LD_LIBRARY_PATH", val)
+}
+
 // Walk the sysfs nodes for the available GPUs and gather information from them
 // skipping over any devices in the skip map
 func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 	resp.memInfo.DeviceCount = 0
 	resp.memInfo.TotalMemory = 0
 	resp.memInfo.FreeMemory = 0
+	slog.Debug("discovering VRAM for amdgpu devices")
 	if len(ids) == 0 {
-		slog.Debug("discovering all amdgpu devices")
 		entries, err := os.ReadDir(AMDNodesSysfsDir)
 		if err != nil {
 			slog.Warn(fmt.Sprintf("failed to read amdgpu sysfs %s - %s", AMDNodesSysfsDir, err))
@@ -165,7 +185,7 @@ func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 			ids = append(ids, id)
 		}
 	}
-	slog.Debug(fmt.Sprintf("discovering amdgpu devices %v", ids))
+	slog.Debug(fmt.Sprintf("amdgpu devices %v", ids))
 
 	for _, id := range ids {
 		if _, skipped := skip[id]; skipped {
@@ -173,7 +193,8 @@ func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 		}
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
-		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id), GPUTotalMemoryFileGlob)
+		// Adjust for sysfs vs HIP ids
+		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id+1), GPUTotalMemoryFileGlob)
 		propFiles, err := filepath.Glob(propGlob)
 		if err != nil {
 			slog.Warn(fmt.Sprintf("error looking up total GPU memory: %s %s", propGlob, err))
@@ -205,6 +226,13 @@ func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 			}
 		}
 		if totalMemory == 0 {
+			slog.Warn(fmt.Sprintf("amdgpu [%d] reports zero total memory, skipping", id))
+			skip[id] = struct{}{}
+			continue
+		}
+		if totalMemory < IGPUMemLimit {
+			slog.Info(fmt.Sprintf("amdgpu [%d] appears to be an iGPU with %dM reported total memory, skipping", id, totalMemory/1024/1024))
+			skip[id] = struct{}{}
 			continue
 		}
 		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id), GPUUsedMemoryFileGlob)
@@ -232,8 +260,8 @@ func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 			}
 			usedMemory += used
 		}
-		slog.Info(fmt.Sprintf("[%d] amdgpu totalMemory %d", id, totalMemory))
-		slog.Info(fmt.Sprintf("[%d] amdgpu freeMemory  %d", id, (totalMemory - usedMemory)))
+		slog.Info(fmt.Sprintf("[%d] amdgpu totalMemory %dM", id, totalMemory/1024/1024))
+		slog.Info(fmt.Sprintf("[%d] amdgpu freeMemory  %dM", id, (totalMemory-usedMemory)/1024/1024))
 		resp.memInfo.DeviceCount++
 		resp.memInfo.TotalMemory += totalMemory
 		resp.memInfo.FreeMemory += (totalMemory - usedMemory)
@@ -282,7 +310,7 @@ func AMDValidateLibDir() (string, error) {
 	}
 
 	// If we already have a rocm dependency wired, nothing more to do
-	rocmTargetDir := filepath.Join(payloadsDir, "rocm")
+	rocmTargetDir := filepath.Clean(filepath.Join(payloadsDir, "..", "rocm"))
 	if rocmLibUsable(rocmTargetDir) {
 		return rocmTargetDir, nil
 	}
@@ -358,6 +386,8 @@ func AMDDriverVersion() (string, error) {
 }
 
 func AMDGFXVersions() map[int]Version {
+	// The amdgpu driver always exposes the host CPU as node 0, but we have to skip that and subtract one
+	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
 	res := map[int]Version{}
 	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
 	for _, match := range matches {
@@ -373,17 +403,20 @@ func AMDGFXVersions() map[int]Version {
 			continue
 		}
 
+		if i == 0 {
+			// Skipping the CPU
+			continue
+		}
+		// Align with HIP IDs (zero is first GPU, not CPU)
+		i -= 1
+
 		scanner := bufio.NewScanner(fp)
 		for scanner.Scan() {
 			line := strings.TrimSpace(scanner.Text())
 			if strings.HasPrefix(line, "gfx_target_version") {
 				ver := strings.Fields(line)
 				if len(ver) != 2 || len(ver[1]) < 5 {
-
-					if ver[1] == "0" {
-						// Silently skip the CPU
-						continue
-					} else {
+					if ver[1] != "0" {
 						slog.Debug("malformed " + line)
 					}
 					res[i] = Version{

gpu/assets.go

@@ -1,13 +1,17 @@
 package gpu
 
 import (
+	"errors"
 	"fmt"
 	"log/slog"
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
+	"syscall"
+	"time"
 )
 
 var (
@@ -18,24 +22,84 @@ var (
 func PayloadsDir() (string, error) {
 	lock.Lock()
 	defer lock.Unlock()
+	var err error
 	if payloadsDir == "" {
-		tmpDir, err := os.MkdirTemp("", "ollama")
-		if err != nil {
-			return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+		cleanupTmpDirs()
+		tmpDir := os.Getenv("OLLAMA_TMPDIR")
+		if tmpDir == "" {
+			tmpDir, err = os.MkdirTemp("", "ollama")
+			if err != nil {
+				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+			}
+		} else {
+			err = os.MkdirAll(tmpDir, 0755)
+			if err != nil {
+				return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
+			}
 		}
-		payloadsDir = tmpDir
+
+		// Track our pid so we can clean up orphaned tmpdirs
+		pidFilePath := filepath.Join(tmpDir, "ollama.pid")
+		pidFile, err := os.OpenFile(pidFilePath, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.ModePerm)
+		if err != nil {
+			return "", err
+		}
+		if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
+			return "", err
+		}
+
+		// We create a distinct subdirectory for payloads within the tmpdir
+		// This will typically look like /tmp/ollama3208993108/runners on linux
+		payloadsDir = filepath.Join(tmpDir, "runners")
 	}
 	return payloadsDir, nil
 }
 
+// Best effort to clean up prior tmpdirs
+func cleanupTmpDirs() {
+	dirs, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*"))
+	if err != nil {
+		return
+	}
+	for _, d := range dirs {
+		info, err := os.Stat(d)
+		if err != nil || !info.IsDir() {
+			continue
+		}
+		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
+		if err == nil {
+			pid, err := strconv.Atoi(string(raw))
+			if err == nil {
+				if proc, err := os.FindProcess(int(pid)); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+					// Another running ollama, ignore this tmpdir
+					continue
+				}
+			}
+		} else {
+			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
+		}
+		err = os.RemoveAll(d)
+		if err != nil {
+			slog.Debug(fmt.Sprintf("unable to cleanup stale tmpdir %s: %s", d, err))
+		}
+	}
+}
+
 func Cleanup() {
 	lock.Lock()
 	defer lock.Unlock()
 	if payloadsDir != "" {
-		slog.Debug("cleaning up", "dir", payloadsDir)
-		err := os.RemoveAll(payloadsDir)
+		// We want to fully clean up the tmpdir parent of the payloads dir
+		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
+		slog.Debug("cleaning up", "dir", tmpDir)
+		err := os.RemoveAll(tmpDir)
 		if err != nil {
-			slog.Warn("failed to clean up", "dir", payloadsDir, "err", err)
+			// On windows, if we remove too quickly the llama.dll may still be in-use and fail to remove
+			time.Sleep(1000 * time.Millisecond)
+			err = os.RemoveAll(tmpDir)
+			if err != nil {
+				slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
+			}
 		}
 	}
 }

gpu/gpu.go

@@ -20,20 +20,27 @@ import (
 	"strings"
 	"sync"
 	"unsafe"
+
+	"github.com/ollama/ollama/format"
 )
 
 type handles struct {
-	cuda *C.cuda_handle_t
+	nvml   *C.nvml_handle_t
+	cudart *C.cudart_handle_t
 }
 
+const (
+	cudaMinimumMemory = 457 * format.MebiByte
+	rocmMinimumMemory = 457 * format.MebiByte
+)
+
 var gpuMutex sync.Mutex
-var gpuHandles *handles = nil
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
 var CudaComputeMin = [2]C.int{5, 0}
 
 // Possible locations for the nvidia-ml library
-var CudaLinuxGlobs = []string{
+var NvmlLinuxGlobs = []string{
 	"/usr/local/cuda/lib64/libnvidia-ml.so*",
 	"/usr/lib/x86_64-linux-gnu/nvidia/current/libnvidia-ml.so*",
 	"/usr/lib/x86_64-linux-gnu/libnvidia-ml.so*",
@@ -41,49 +48,98 @@ var CudaLinuxGlobs = []string{
 	"/usr/lib/wsl/drivers/*/libnvidia-ml.so*",
 	"/opt/cuda/lib64/libnvidia-ml.so*",
 	"/usr/lib*/libnvidia-ml.so*",
-	"/usr/local/lib*/libnvidia-ml.so*",
 	"/usr/lib/aarch64-linux-gnu/nvidia/current/libnvidia-ml.so*",
 	"/usr/lib/aarch64-linux-gnu/libnvidia-ml.so*",
+	"/usr/local/lib*/libnvidia-ml.so*",
 
 	// TODO: are these stubs ever valid?
 	"/opt/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so*",
 }
 
-var CudaWindowsGlobs = []string{
+var NvmlWindowsGlobs = []string{
 	"c:\\Windows\\System32\\nvml.dll",
 }
 
+var CudartLinuxGlobs = []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 CudartWindowsGlobs = []string{
+	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
+}
+
+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
 // Note: gpuMutex must already be held
-func initGPUHandles() {
+func initGPUHandles() *handles {
 
 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing
 
-	gpuHandles = &handles{nil}
-	var cudaMgmtName string
-	var cudaMgmtPatterns []string
+	gpuHandles := &handles{nil, nil}
+	var nvmlMgmtName string
+	var nvmlMgmtPatterns []string
+	var cudartMgmtName string
+	var cudartMgmtPatterns []string
+
+	tmpDir, _ := PayloadsDir()
 	switch runtime.GOOS {
 	case "windows":
-		cudaMgmtName = "nvml.dll"
-		cudaMgmtPatterns = make([]string, len(CudaWindowsGlobs))
-		copy(cudaMgmtPatterns, CudaWindowsGlobs)
+		nvmlMgmtName = "nvml.dll"
+		nvmlMgmtPatterns = make([]string, len(NvmlWindowsGlobs))
+		copy(nvmlMgmtPatterns, NvmlWindowsGlobs)
+		cudartMgmtName = "cudart64_*.dll"
+		localAppData := os.Getenv("LOCALAPPDATA")
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
 	case "linux":
-		cudaMgmtName = "libnvidia-ml.so"
-		cudaMgmtPatterns = make([]string, len(CudaLinuxGlobs))
-		copy(cudaMgmtPatterns, CudaLinuxGlobs)
+		nvmlMgmtName = "libnvidia-ml.so"
+		nvmlMgmtPatterns = make([]string, len(NvmlLinuxGlobs))
+		copy(nvmlMgmtPatterns, NvmlLinuxGlobs)
+		cudartMgmtName = "libcudart.so*"
+		if tmpDir != "" {
+			// TODO - add "payloads" for subprocess
+			cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", cudartMgmtName)}
+		}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartLinuxGlobs...)
 	default:
-		return
+		return gpuHandles
 	}
 
 	slog.Info("Detecting GPU type")
-	cudaLibPaths := FindGPULibs(cudaMgmtName, cudaMgmtPatterns)
-	if len(cudaLibPaths) > 0 {
-		cuda := LoadCUDAMgmt(cudaLibPaths)
-		if cuda != nil {
-			slog.Info("Nvidia GPU detected")
-			gpuHandles.cuda = cuda
-			return
+	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
+	if len(cudartLibPaths) > 0 {
+		cudart := LoadCUDARTMgmt(cudartLibPaths)
+		if cudart != nil {
+			slog.Info("Nvidia GPU detected via cudart")
+			gpuHandles.cudart = cudart
+			return gpuHandles
 		}
 	}
+
+	// TODO once we build confidence, remove this and the gpu_info_nvml.[ch] files
+	nvmlLibPaths := FindGPULibs(nvmlMgmtName, nvmlMgmtPatterns)
+	if len(nvmlLibPaths) > 0 {
+		nvml := LoadNVMLMgmt(nvmlLibPaths)
+		if nvml != nil {
+			slog.Info("Nvidia GPU detected via nvidia-ml")
+			gpuHandles.nvml = nvml
+			return gpuHandles
+		}
+	}
+	return gpuHandles
 }
 
 func GetGPUInfo() GpuInfo {
@@ -91,9 +147,16 @@ func GetGPUInfo() GpuInfo {
 	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
 	gpuMutex.Lock()
 	defer gpuMutex.Unlock()
-	if gpuHandles == nil {
-		initGPUHandles()
-	}
+
+	gpuHandles := initGPUHandles()
+	defer func() {
+		if gpuHandles.nvml != nil {
+			C.nvml_release(*gpuHandles.nvml)
+		}
+		if gpuHandles.cudart != nil {
+			C.cudart_release(*gpuHandles.cudart)
+		}
+	}()
 
 	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
 	cpuVariant := GetCPUVariant()
@@ -103,28 +166,50 @@ func GetGPUInfo() GpuInfo {
 
 	var memInfo C.mem_info_t
 	resp := GpuInfo{}
-	if gpuHandles.cuda != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
-		C.cuda_check_vram(*gpuHandles.cuda, &memInfo)
+	if gpuHandles.nvml != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
+		C.nvml_check_vram(*gpuHandles.nvml, &memInfo)
 		if memInfo.err != nil {
-			slog.Info(fmt.Sprintf("error looking up CUDA GPU memory: %s", C.GoString(memInfo.err)))
+			slog.Info(fmt.Sprintf("[nvidia-ml] error looking up NVML GPU memory: %s", C.GoString(memInfo.err)))
 			C.free(unsafe.Pointer(memInfo.err))
 		} else if memInfo.count > 0 {
 			// Verify minimum compute capability
-			var cc C.cuda_compute_capability_t
-			C.cuda_compute_capability(*gpuHandles.cuda, &cc)
+			var cc C.nvml_compute_capability_t
+			C.nvml_compute_capability(*gpuHandles.nvml, &cc)
 			if cc.err != nil {
-				slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err)))
+				slog.Info(fmt.Sprintf("[nvidia-ml] error looking up NVML GPU compute capability: %s", C.GoString(cc.err)))
 				C.free(unsafe.Pointer(cc.err))
 			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
-				slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
+				slog.Info(fmt.Sprintf("[nvidia-ml] NVML CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
 				resp.Library = "cuda"
+				resp.MinimumMemory = cudaMinimumMemory
 			} else {
-				slog.Info(fmt.Sprintf("CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
+				slog.Info(fmt.Sprintf("[nvidia-ml] CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
+			}
+		}
+	} else if gpuHandles.cudart != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
+		C.cudart_check_vram(*gpuHandles.cudart, &memInfo)
+		if memInfo.err != nil {
+			slog.Info(fmt.Sprintf("[cudart] error looking up CUDART GPU memory: %s", C.GoString(memInfo.err)))
+			C.free(unsafe.Pointer(memInfo.err))
+		} else if memInfo.count > 0 {
+			// Verify minimum compute capability
+			var cc C.cudart_compute_capability_t
+			C.cudart_compute_capability(*gpuHandles.cudart, &cc)
+			if cc.err != nil {
+				slog.Info(fmt.Sprintf("[cudart] error looking up CUDA compute capability: %s", C.GoString(cc.err)))
+				C.free(unsafe.Pointer(cc.err))
+			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
+				slog.Info(fmt.Sprintf("[cudart] CUDART CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
+				resp.Library = "cuda"
+				resp.MinimumMemory = cudaMinimumMemory
+			} else {
+				slog.Info(fmt.Sprintf("[cudart] CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
 			}
 		}
 	} else {
 		AMDGetGPUInfo(&resp)
 		if resp.Library != "" {
+			resp.MinimumMemory = rocmMinimumMemory
 			return resp
 		}
 	}
@@ -158,7 +243,7 @@ func getCPUMem() (memInfo, error) {
 	return ret, nil
 }
 
-func CheckVRAM() (int64, error) {
+func CheckVRAM() (uint64, error) {
 	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
 	if userLimit != "" {
 		avail, err := strconv.ParseInt(userLimit, 10, 64)
@@ -166,19 +251,11 @@ func CheckVRAM() (int64, error) {
 			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
 		}
 		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
-		return avail, nil
+		return uint64(avail), nil
 	}
 	gpuInfo := GetGPUInfo()
 	if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
-		// leave 10% or 1024MiB of VRAM free per GPU to handle unaccounted for overhead
-		overhead := gpuInfo.FreeMemory / 10
-		gpus := uint64(gpuInfo.DeviceCount)
-		if overhead < gpus*1024*1024*1024 {
-			overhead = gpus * 1024 * 1024 * 1024
-		}
-		avail := int64(gpuInfo.FreeMemory - overhead)
-		slog.Debug(fmt.Sprintf("%s detected %d devices with %dM available memory", gpuInfo.Library, gpuInfo.DeviceCount, avail/1024/1024))
-		return avail, nil
+		return gpuInfo.FreeMemory, nil
 	}
 
 	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation
@@ -238,15 +315,32 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 	return gpuLibPaths
 }
 
-func LoadCUDAMgmt(cudaLibPaths []string) *C.cuda_handle_t {
-	var resp C.cuda_init_resp_t
+func LoadNVMLMgmt(nvmlLibPaths []string) *C.nvml_handle_t {
+	var resp C.nvml_init_resp_t
 	resp.ch.verbose = getVerboseState()
-	for _, libPath := range cudaLibPaths {
+	for _, libPath := range nvmlLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
-		C.cuda_init(lib, &resp)
+		C.nvml_init(lib, &resp)
 		if resp.err != nil {
-			slog.Info(fmt.Sprintf("Unable to load CUDA management library %s: %s", libPath, C.GoString(resp.err)))
+			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
+			C.free(unsafe.Pointer(resp.err))
+		} else {
+			return &resp.ch
+		}
+	}
+	return nil
+}
+
+func LoadCUDARTMgmt(cudartLibPaths []string) *C.cudart_handle_t {
+	var resp C.cudart_init_resp_t
+	resp.ch.verbose = getVerboseState()
+	for _, libPath := range cudartLibPaths {
+		lib := C.CString(libPath)
+		defer C.free(unsafe.Pointer(lib))
+		C.cudart_init(lib, &resp)
+		if resp.err != nil {
+			slog.Info(fmt.Sprintf("Unable to load cudart CUDA management library %s: %s", libPath, C.GoString(resp.err)))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return &resp.ch

gpu/gpu_darwin.go

@@ -17,7 +17,7 @@ import (
 )
 
 // CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
-func CheckVRAM() (int64, error) {
+func CheckVRAM() (uint64, error) {
 	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
 	if userLimit != "" {
 		avail, err := strconv.ParseInt(userLimit, 10, 64)
@@ -25,15 +25,15 @@ func CheckVRAM() (int64, error) {
 			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
 		}
 		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
-		return avail, nil
+		return uint64(avail), nil
 	}
 
 	if runtime.GOARCH == "amd64" {
 		// gpu not supported, this may not be metal
 		return 0, nil
 	}
-	recommendedMaxVRAM := int64(C.getRecommendedMaxVRAM())
-	return recommendedMaxVRAM, nil
+
+	return uint64(C.getRecommendedMaxVRAM()), nil
 }
 
 func GetGPUInfo() GpuInfo {
@@ -53,8 +53,8 @@ func GetGPUInfo() GpuInfo {
 
 func getCPUMem() (memInfo, error) {
 	return memInfo{
-		TotalMemory: 0,
+		TotalMemory: uint64(C.getPhysicalMemory()),
 		FreeMemory:  0,
-		DeviceCount: 0,
+		DeviceCount: 1,
 	}, nil
 }

gpu/gpu_info.h

@@ -52,7 +52,8 @@ void cpu_check_ram(mem_info_t *resp);
 }
 #endif
 
-#include "gpu_info_cuda.h"
+#include "gpu_info_nvml.h"
+#include "gpu_info_cudart.h"
 
 #endif  // __GPU_INFO_H__
 #endif  // __APPLE__

gpu/gpu_info_cudart.c

@@ -0,0 +1,200 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include <string.h>
+#include "gpu_info_cudart.h"
+
+void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
+  cudartReturn_t ret;
+  resp->err = NULL;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+      {"cudaSetDevice", (void *)&resp->ch.cudaSetDevice},
+      {"cudaDeviceSynchronize", (void *)&resp->ch.cudaDeviceSynchronize},
+      {"cudaDeviceReset", (void *)&resp->ch.cudaDeviceReset},
+      {"cudaMemGetInfo", (void *)&resp->ch.cudaMemGetInfo},
+      {"cudaGetDeviceCount", (void *)&resp->ch.cudaGetDeviceCount},
+      {"cudaDeviceGetAttribute", (void *)&resp->ch.cudaDeviceGetAttribute},
+      {"cudaDriverGetVersion", (void *)&resp->ch.cudaDriverGetVersion},
+      {NULL, NULL},
+  };
+
+  resp->ch.handle = LOAD_LIBRARY(cudart_lib_path, RTLD_LAZY);
+  if (!resp->ch.handle) {
+    char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", cudart_lib_path, msg);
+    snprintf(buf, buflen,
+            "Unable to load %s library to query for Nvidia GPUs: %s",
+            cudart_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  // TODO once we've squashed the remaining corner cases remove this log
+  LOG(resp->ch.verbose, "wiring cudart library functions in %s\n", cudart_lib_path);
+  
+  for (i = 0; l[i].s != NULL; i++) {
+    // TODO once we've squashed the remaining corner cases remove this log
+    LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
+
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!l[i].p) {
+      char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
+      resp->ch.handle = NULL;
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+              msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->ch.cudaSetDevice)(0);
+  if (ret != CUDART_SUCCESS) {
+    LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+      resp->err = strdup("your nvidia driver is too old or missing, please upgrade to run ollama");
+      return;
+    }
+    snprintf(buf, buflen, "cudart init failure: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  int version = 0;
+  cudartDriverVersion_t driverVersion;
+  driverVersion.major = 0;
+  driverVersion.minor = 0;
+
+  // Report driver version if we're in verbose mode, ignore errors
+  ret = (*resp->ch.cudaDriverGetVersion)(&version);
+  if (ret != CUDART_SUCCESS) {
+    LOG(resp->ch.verbose, "cudaDriverGetVersion failed: %d\n", ret);
+  } else {
+    driverVersion.major = version / 1000;
+    driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
+    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
+  }
+}
+
+
+void cudart_check_vram(cudart_handle_t h, mem_info_t *resp) {
+  resp->err = NULL;
+  cudartMemory_t memInfo = {0,0,0};
+  cudartReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("cudart handle isn't initialized");
+    return;
+  }
+
+  // cudaGetDeviceCount takes int type, resp-> count is uint
+  int deviceCount;
+  ret = (*h.cudaGetDeviceCount)(&deviceCount);
+  if (ret != CUDART_SUCCESS) {
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  } else {
+    resp->count = (unsigned int)deviceCount;
+  }
+
+  resp->total = 0;
+  resp->free = 0;
+  for (i = 0; i < resp-> count; i++) {  
+    ret = (*h.cudaSetDevice)(i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device failed to initialize");
+      resp->err = strdup(buf);
+      return;
+    }
+    ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
+      resp->err = strdup(buf);
+      return;
+    }
+
+    LOG(h.verbose, "[%d] CUDA totalMem %lu\n", i, memInfo.total);
+    LOG(h.verbose, "[%d] CUDA freeMem %lu\n", i, memInfo.free);
+
+    resp->total += memInfo.total;
+    resp->free += memInfo.free;
+  }
+}
+
+void cudart_compute_capability(cudart_handle_t h, cudart_compute_capability_t *resp) {
+  resp->err = NULL;
+  resp->major = 0;
+  resp->minor = 0;
+  int major = 0;
+  int minor = 0;
+  cudartReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  if (h.handle == NULL) {
+    resp->err = strdup("cudart handle not initialized");
+    return;
+  }
+
+  int devices;
+  ret = (*h.cudaGetDeviceCount)(&devices);
+  if (ret != CUDART_SUCCESS) {
+    snprintf(buf, buflen, "unable to get cudart device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; i < devices; i++) {
+    ret = (*h.cudaSetDevice)(i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "cudart device failed to initialize");
+      resp->err = strdup(buf);
+      return;
+    }
+
+    ret = (*h.cudaDeviceGetAttribute)(&major, cudartDevAttrComputeCapabilityMajor, i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+    ret = (*h.cudaDeviceGetAttribute)(&minor, cudartDevAttrComputeCapabilityMinor, i);
+    if (ret != CUDART_SUCCESS) {
+      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
+      resp->err = strdup(buf);
+      return;
+    }
+      
+    // Report the lowest major.minor we detect as that limits our compatibility
+    if (resp->major == 0 || resp->major > major ) {
+      resp->major = major;
+      resp->minor = minor;
+    } else if ( resp->major == major && resp->minor > minor ) {
+      resp->minor = minor;
+    }
+  }
+}
+
+void cudart_release(cudart_handle_t h) {
+  LOG(h.verbose, "releasing cudart library\n");
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
+
+#endif  // __APPLE__

gpu/gpu_info_cudart.h

@@ -0,0 +1,61 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_CUDART_H__
+#define __GPU_INFO_CUDART_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum cudartReturn_enum {
+  CUDART_SUCCESS = 0,
+  CUDART_UNSUPPORTED = 1,
+  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  // Other values omitted for now...
+} cudartReturn_t;
+
+typedef enum cudartDeviceAttr_enum {
+  cudartDevAttrComputeCapabilityMajor = 75,
+  cudartDevAttrComputeCapabilityMinor = 76,
+} cudartDeviceAttr_t;
+
+typedef void *cudartDevice_t;  // Opaque is sufficient
+typedef struct cudartMemory_st {
+  size_t total;
+  size_t free;
+  size_t used;
+} cudartMemory_t;
+
+typedef struct cudartDriverVersion {
+  int major;
+  int minor;
+} cudartDriverVersion_t;
+
+typedef struct cudart_handle {
+  void *handle;
+  uint16_t verbose;
+  cudartReturn_t (*cudaSetDevice)(int device);
+  cudartReturn_t (*cudaDeviceSynchronize)(void);
+  cudartReturn_t (*cudaDeviceReset)(void);
+  cudartReturn_t (*cudaMemGetInfo)(size_t *, size_t *);
+  cudartReturn_t (*cudaGetDeviceCount)(int *);
+  cudartReturn_t (*cudaDeviceGetAttribute)(int* value, cudartDeviceAttr_t attr, int device);
+  cudartReturn_t (*cudaDriverGetVersion) (int *driverVersion);
+} cudart_handle_t;
+
+typedef struct cudart_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  cudart_handle_t ch;
+} cudart_init_resp_t;
+
+typedef struct cudart_compute_capability {
+  char *err;
+  int major;
+  int minor;
+} cudart_compute_capability_t;
+
+
+void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
+void cudart_check_vram(cudart_handle_t ch, mem_info_t *resp);
+void cudart_compute_capability(cudart_handle_t th, cudart_compute_capability_t *cc);
+void cudart_release(cudart_handle_t ch);
+
+#endif  // __GPU_INFO_CUDART_H__
+#endif  // __APPLE__

gpu/gpu_info_darwin.h

@@ -1,3 +1,4 @@
 #import <Metal/Metal.h>
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
+uint64_t getPhysicalMemory();

gpu/gpu_info_darwin.m

@@ -1,11 +1,13 @@
-//go:build darwin
+// go:build darwin
 #include "gpu_info_darwin.h"
 
-uint64_t getRecommendedMaxVRAM()
-{
-	id<MTLDevice> device = MTLCreateSystemDefaultDevice();
-	uint64_t result = device.recommendedMaxWorkingSetSize;
-	CFRelease(device);
-	return result;
+uint64_t getRecommendedMaxVRAM() {
+  id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+  uint64_t result = device.recommendedMaxWorkingSetSize;
+  CFRelease(device);
+  return result;
 }
 
+uint64_t getPhysicalMemory() {
+  return [[NSProcessInfo processInfo] physicalMemory];
+}

gpu/gpu_info_nvml.c

@@ -1,10 +1,10 @@
 #ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
 
-#include "gpu_info_cuda.h"
-
 #include <string.h>
 
-void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp) {
+#include "gpu_info_nvml.h"
+
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
   nvmlReturn_t ret;
   resp->err = NULL;
   const int buflen = 256;
@@ -30,20 +30,20 @@ void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp) {
       {NULL, NULL},
   };
 
-  resp->ch.handle = LOAD_LIBRARY(cuda_lib_path, RTLD_LAZY);
+  resp->ch.handle = LOAD_LIBRARY(nvml_lib_path, RTLD_LAZY);
   if (!resp->ch.handle) {
     char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", cuda_lib_path, msg);
+    LOG(resp->ch.verbose, "library %s load err: %s\n", nvml_lib_path, msg);
     snprintf(buf, buflen,
              "Unable to load %s library to query for Nvidia GPUs: %s",
-             cuda_lib_path, msg);
+             nvml_lib_path, msg);
     free(msg);
     resp->err = strdup(buf);
     return;
   }
 
   // TODO once we've squashed the remaining corner cases remove this log
-  LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", cuda_lib_path);
+  LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", nvml_lib_path);
   
   for (i = 0; l[i].s != NULL; i++) {
     // TODO once we've squashed the remaining corner cases remove this log
@@ -82,7 +82,7 @@ void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp) {
   }
 }
 
-void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
+void nvml_check_vram(nvml_handle_t h, mem_info_t *resp) {
   resp->err = NULL;
   nvmlDevice_t device;
   nvmlMemory_t memInfo = {0};
@@ -92,7 +92,7 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
   int i;
 
   if (h.handle == NULL) {
-    resp->err = strdup("nvml handle sn't initialized");
+    resp->err = strdup("nvml handle isn't initialized");
     return;
   }
 
@@ -156,14 +156,14 @@ void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
     }
 
     LOG(h.verbose, "[%d] CUDA totalMem %ld\n", i, memInfo.total);
-    LOG(h.verbose, "[%d] CUDA usedMem %ld\n", i, memInfo.used);
+    LOG(h.verbose, "[%d] CUDA freeMem %ld\n", i, memInfo.free);
 
     resp->total += memInfo.total;
     resp->free += memInfo.free;
   }
 }
 
-void cuda_compute_capability(cuda_handle_t h, cuda_compute_capability_t *resp) {
+void nvml_compute_capability(nvml_handle_t h, nvml_compute_capability_t *resp) {
   resp->err = NULL;
   resp->major = 0;
   resp->minor = 0;
@@ -211,4 +211,11 @@ void cuda_compute_capability(cuda_handle_t h, cuda_compute_capability_t *resp) {
     }
   }
 }
+
+void nvml_release(nvml_handle_t h) {
+  LOG(h.verbose, "releasing nvml library\n");
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
+
 #endif  // __APPLE__

gpu/gpu_info_nvml.h

@@ -1,6 +1,6 @@
 #ifndef __APPLE__
-#ifndef __GPU_INFO_CUDA_H__
-#define __GPU_INFO_CUDA_H__
+#ifndef __GPU_INFO_NVML_H__
+#define __GPU_INFO_NVML_H__
 #include "gpu_info.h"
 
 // Just enough typedef's to dlopen/dlsym for memory information
@@ -20,7 +20,7 @@ typedef enum nvmlBrandType_enum
     NVML_BRAND_UNKNOWN          = 0,
 } nvmlBrandType_t;
 
-typedef struct cuda_handle {
+typedef struct nvml_handle {
   void *handle;
   uint16_t verbose;
   nvmlReturn_t (*nvmlInit_v2)(void);
@@ -35,22 +35,23 @@ typedef struct cuda_handle {
   nvmlReturn_t (*nvmlDeviceGetVbiosVersion) (nvmlDevice_t device, char* version, unsigned int  length);
   nvmlReturn_t (*nvmlDeviceGetBoardPartNumber) (nvmlDevice_t device, char* partNumber, unsigned int  length);
   nvmlReturn_t (*nvmlDeviceGetBrand) (nvmlDevice_t device, nvmlBrandType_t* type);
-} cuda_handle_t;
+} nvml_handle_t;
 
-typedef struct cuda_init_resp {
+typedef struct nvml_init_resp {
   char *err;  // If err is non-null handle is invalid
-  cuda_handle_t ch;
-} cuda_init_resp_t;
+  nvml_handle_t ch;
+} nvml_init_resp_t;
 
-typedef struct cuda_compute_capability {
+typedef struct nvml_compute_capability {
   char *err;
   int major;
   int minor;
-} cuda_compute_capability_t;
+} nvml_compute_capability_t;
 
-void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp);
-void cuda_check_vram(cuda_handle_t ch, mem_info_t *resp);
-void cuda_compute_capability(cuda_handle_t ch, cuda_compute_capability_t *cc);
+void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
+void nvml_check_vram(nvml_handle_t ch, mem_info_t *resp);
+void nvml_compute_capability(nvml_handle_t ch, nvml_compute_capability_t *cc);
+void nvml_release(nvml_handle_t ch);
 
-#endif  // __GPU_INFO_CUDA_H__
+#endif  // __GPU_INFO_NVML_H__
 #endif  // __APPLE__

gpu/types.go

@@ -14,6 +14,9 @@ type GpuInfo struct {
 	// Optional variant to select (e.g. versions, cpu feature flags)
 	Variant string `json:"variant,omitempty"`
 
+	// MinimumMemory represents the minimum memory required to use the GPU
+	MinimumMemory uint64 `json:"-"`
+
 	// TODO add other useful attributes about the card here for discovery information
 }
 

integration/README.md

@@ -0,0 +1,11 @@
+# Integration Tests
+
+This directory contains integration tests to exercise Ollama end-to-end to verify behavior
+
+By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...`
+
+
+The integration tests have 2 modes of operating.
+
+1. By default, they will start the server on a random port, run the tests, and then shutdown the server.
+2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote

integration/basic_test.go

@@ -0,0 +1,28 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestOrcaMiniBlueSky(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "orca-mini",
+		Prompt: "why is the sky blue?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{"rayleigh", "scattering"})
+}

integration/context_test.go

@@ -0,0 +1,29 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestContextExhaustion(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute) // TODO maybe shorter?
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "llama2",
+		Prompt: "Write me a story with a ton of emojis?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+			"num_ctx":     128,
+		},
+	}
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{"once", "upon", "lived"})
+}

integration/llm_image_test.go

@@ -1,49 +1,38 @@
 //go:build integration
 
-package server
+package integration
 
 import (
 	"context"
 	"encoding/base64"
-	"log"
-	"os"
-	"strings"
+	"net/http"
 	"testing"
 	"time"
 
-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/llm"
-	"github.com/stretchr/testify/assert"
+	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
 )
 
 func TestIntegrationMultimodal(t *testing.T) {
-	SkipIFNoTestData(t)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{
-		Model:   "llava:7b",
-		Prompt:  "what does the text in this image say?",
-		Options: map[string]interface{}{},
+		Model:  "llava:7b",
+		Prompt: "what does the text in this image say?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"seed":        42,
+			"temperature": 0.0,
+		},
 		Images: []api.ImageData{
 			image,
 		},
 	}
+
 	resp := "the ollamas"
-	workDir, err := os.MkdirTemp("", "ollama")
-	require.NoError(t, err)
-	defer os.RemoveAll(workDir)
-	require.NoError(t, llm.Init(workDir))
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*60)
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
-	opts := api.DefaultOptions()
-	opts.Seed = 42
-	opts.Temperature = 0.0
-	model, llmRunner := PrepareModelForPrompts(t, req.Model, opts)
-	defer llmRunner.Close()
-	response := OneShotPromptResponse(t, ctx, req, model, llmRunner)
-	log.Print(response)
-	assert.Contains(t, strings.ToLower(response), resp)
+	GenerateTestHelper(ctx, t, &http.Client{}, req, []string{resp})
 }
 
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb

integration/llm_test.go

@@ -0,0 +1,69 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"net/http"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// TODO - this would ideally be in the llm package, but that would require some refactoring of interfaces in the server
+//        package to avoid circular dependencies
+
+var (
+	stream = false
+	req    = [2]api.GenerateRequest{
+		{
+			Model:  "orca-mini",
+			Prompt: "why is the ocean blue?",
+			Stream: &stream,
+			Options: map[string]interface{}{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		}, {
+			Model:  "orca-mini",
+			Prompt: "what is the origin of the us thanksgiving holiday?",
+			Stream: &stream,
+			Options: map[string]interface{}{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		},
+	}
+	resp = [2][]string{
+		[]string{"sunlight"},
+		[]string{"england", "english", "massachusetts", "pilgrims"},
+	}
+)
+
+func TestIntegrationSimpleOrcaMini(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	GenerateTestHelper(ctx, t, &http.Client{}, req[0], resp[0])
+}
+
+// TODO
+// The server always loads a new runner and closes the old one, which forces serial execution
+// At present this test case fails with concurrency problems.  Eventually we should try to
+// get true concurrency working with n_parallel support in the backend
+func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
+	var wg sync.WaitGroup
+	wg.Add(len(req))
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	for i := 0; i < len(req); i++ {
+		go func(i int) {
+			defer wg.Done()
+			GenerateTestHelper(ctx, t, &http.Client{}, req[i], resp[i])
+		}(i)
+	}
+	wg.Wait()
+}
+
+// TODO - create a parallel test with 2 different models once we support concurrency

integration/utils_test.go

@@ -0,0 +1,265 @@
+//go:build integration
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"math/rand"
+	"net"
+	"net/http"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/app/lifecycle"
+	"github.com/stretchr/testify/assert"
+)
+
+func FindPort() string {
+	port := 0
+	if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
+		var l *net.TCPListener
+		if l, err = net.ListenTCP("tcp", a); err == nil {
+			port = l.Addr().(*net.TCPAddr).Port
+			l.Close()
+		}
+	}
+	if port == 0 {
+		port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+	}
+	return strconv.Itoa(port)
+}
+
+func GetTestEndpoint() (string, string) {
+	defaultPort := "11434"
+	ollamaHost := os.Getenv("OLLAMA_HOST")
+
+	scheme, hostport, ok := strings.Cut(ollamaHost, "://")
+	if !ok {
+		scheme, hostport = "http", ollamaHost
+	}
+
+	// trim trailing slashes
+	hostport = strings.TrimRight(hostport, "/")
+
+	host, port, err := net.SplitHostPort(hostport)
+	if err != nil {
+		host, port = "127.0.0.1", defaultPort
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
+	}
+
+	if os.Getenv("OLLAMA_TEST_EXISTING") == "" && port == defaultPort {
+		port = FindPort()
+	}
+
+	url := fmt.Sprintf("%s:%s", host, port)
+	slog.Info("server connection", "url", url)
+	return scheme, url
+}
+
+// TODO make fanicier, grab logs, etc.
+var serverMutex sync.Mutex
+var serverReady bool
+
+func StartServer(ctx context.Context, ollamaHost string) error {
+	// Make sure the server has been built
+	CLIName, err := filepath.Abs("../ollama")
+	if err != nil {
+		return err
+	}
+
+	if runtime.GOOS == "windows" {
+		CLIName += ".exe"
+	}
+	_, err = os.Stat(CLIName)
+	if err != nil {
+		return fmt.Errorf("CLI missing, did you forget to build first?  %w", err)
+	}
+	serverMutex.Lock()
+	defer serverMutex.Unlock()
+	if serverReady {
+		return nil
+	}
+
+	if tmp := os.Getenv("OLLAMA_HOST"); tmp != ollamaHost {
+		slog.Info("setting env", "OLLAMA_HOST", ollamaHost)
+		os.Setenv("OLLAMA_HOST", ollamaHost)
+	}
+
+	slog.Info("starting server", "url", ollamaHost)
+	done, err := lifecycle.SpawnServer(ctx, "../ollama")
+	if err != nil {
+		return fmt.Errorf("failed to start server: %w", err)
+	}
+
+	go func() {
+		<-ctx.Done()
+		serverMutex.Lock()
+		defer serverMutex.Unlock()
+		exitCode := <-done
+		if exitCode > 0 {
+			slog.Warn("server failure", "exit", exitCode)
+		}
+		serverReady = false
+	}()
+
+	// TODO wait only long enough for the server to be responsive...
+	time.Sleep(500 * time.Millisecond)
+
+	serverReady = true
+	return nil
+}
+
+func PullIfMissing(ctx context.Context, client *http.Client, scheme, testEndpoint, modelName string) error {
+	slog.Info("checking status of model", "model", modelName)
+	showReq := &api.ShowRequest{Name: modelName}
+	requestJSON, err := json.Marshal(showReq)
+	if err != nil {
+		return err
+	}
+
+	req, err := http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/show", bytes.NewReader(requestJSON))
+	if err != nil {
+		return err
+	}
+
+	// Make the request with the HTTP client
+	response, err := client.Do(req.WithContext(ctx))
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode == 200 {
+		slog.Info("model already present", "model", modelName)
+		return nil
+	}
+	slog.Info("model missing", "status", response.StatusCode)
+
+	pullReq := &api.PullRequest{Name: modelName, Stream: &stream}
+	requestJSON, err = json.Marshal(pullReq)
+	if err != nil {
+		return err
+	}
+
+	req, err = http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/pull", bytes.NewReader(requestJSON))
+	if err != nil {
+		return err
+	}
+	slog.Info("pulling", "model", modelName)
+
+	response, err = client.Do(req.WithContext(ctx))
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode != 200 {
+		return fmt.Errorf("failed to pull model") // TODO more details perhaps
+	}
+	slog.Info("model pulled", "model", modelName)
+	return nil
+}
+
+var serverProcMutex sync.Mutex
+
+func GenerateTestHelper(ctx context.Context, t *testing.T, client *http.Client, genReq api.GenerateRequest, anyResp []string) {
+
+	// TODO maybe stuff in an init routine?
+	lifecycle.InitLogging()
+
+	requestJSON, err := json.Marshal(genReq)
+	if err != nil {
+		t.Fatalf("Error serializing request: %v", err)
+	}
+	defer func() {
+		if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
+			defer serverProcMutex.Unlock()
+			if t.Failed() {
+				fp, err := os.Open(lifecycle.ServerLogFile)
+				if err != nil {
+					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					return
+				}
+				data, err := io.ReadAll(fp)
+				if err != nil {
+					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					return
+				}
+				slog.Warn("SERVER LOG FOLLOWS")
+				os.Stderr.Write(data)
+				slog.Warn("END OF SERVER")
+			}
+			err = os.Remove(lifecycle.ServerLogFile)
+			if err != nil && !os.IsNotExist(err) {
+				slog.Warn("failed to cleanup", "logfile", lifecycle.ServerLogFile, "error", err)
+			}
+		}
+	}()
+	scheme, testEndpoint := GetTestEndpoint()
+
+	if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
+		serverProcMutex.Lock()
+		fp, err := os.CreateTemp("", "ollama-server-*.log")
+		if err != nil {
+			t.Fatalf("failed to generate log file: %s", err)
+		}
+		lifecycle.ServerLogFile = fp.Name()
+		fp.Close()
+		assert.NoError(t, StartServer(ctx, testEndpoint))
+	}
+
+	err = PullIfMissing(ctx, client, scheme, testEndpoint, genReq.Model)
+	if err != nil {
+		t.Fatalf("Error pulling model: %v", err)
+	}
+
+	// Make the request and get the response
+	req, err := http.NewRequest("POST", scheme+"://"+testEndpoint+"/api/generate", bytes.NewReader(requestJSON))
+	if err != nil {
+		t.Fatalf("Error creating request: %v", err)
+	}
+
+	// Set the content type for the request
+	req.Header.Set("Content-Type", "application/json")
+
+	// Make the request with the HTTP client
+	response, err := client.Do(req.WithContext(ctx))
+	if err != nil {
+		t.Fatalf("Error making request: %v", err)
+	}
+	defer response.Body.Close()
+	body, err := io.ReadAll(response.Body)
+	assert.NoError(t, err)
+	assert.Equal(t, response.StatusCode, 200, string(body))
+
+	// Verify the response is valid JSON
+	var payload api.GenerateResponse
+	err = json.Unmarshal(body, &payload)
+	if err != nil {
+		assert.NoError(t, err, body)
+	}
+
+	// Verify the response contains the expected data
+	atLeastOne := false
+	for _, resp := range anyResp {
+		if strings.Contains(strings.ToLower(payload.Response), resp) {
+			atLeastOne = true
+			break
+		}
+	}
+	assert.True(t, atLeastOne, "none of %v found in %s", anyResp, payload.Response)
+}

llm/dyn_ext_server.c

@@ -1,142 +0,0 @@
-#include "dyn_ext_server.h"
-
-#include <stdio.h>
-#include <string.h>
-
-#ifdef __linux__
-#include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
-#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
-#define LOAD_ERR() strdup(dlerror())
-#define UNLOAD_LIBRARY(handle) dlclose(handle)
-#elif _WIN32
-#include <windows.h>
-#define LOAD_LIBRARY(lib, flags) LoadLibrary(lib)
-#define LOAD_SYMBOL(handle, sym) GetProcAddress(handle, sym)
-#define UNLOAD_LIBRARY(handle) FreeLibrary(handle)
-#define LOAD_ERR() ({\
-  LPSTR messageBuffer = NULL; \
-  size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, \
-                                 NULL, GetLastError(), MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&messageBuffer, 0, NULL); \
-  char *resp = strdup(messageBuffer); \
-  LocalFree(messageBuffer); \
-  resp; \
-})
-#else
-#include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
-#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
-#define LOAD_ERR() strdup(dlerror())
-#define UNLOAD_LIBRARY(handle) dlclose(handle)
-#endif
-
-void dyn_init(const char *libPath, struct dynamic_llama_server *s,
-                       ext_server_resp_t *err) {
-  int i = 0;
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"llama_server_init", (void *)&s->llama_server_init},
-      {"llama_server_start", (void *)&s->llama_server_start},
-      {"llama_server_stop", (void *)&s->llama_server_stop},
-      {"llama_server_completion", (void *)&s->llama_server_completion},
-      {"llama_server_completion_next_result",
-       (void *)&s->llama_server_completion_next_result},
-      {"llama_server_completion_cancel",
-       (void *)&s->llama_server_completion_cancel},
-      {"llama_server_release_task_result",
-       (void *)&s->llama_server_release_task_result},
-      {"llama_server_tokenize", (void *)&s->llama_server_tokenize},
-      {"llama_server_detokenize", (void *)&s->llama_server_detokenize},
-      {"llama_server_embedding", (void *)&s->llama_server_embedding},
-      {"llama_server_release_json_resp",
-       (void *)&s->llama_server_release_json_resp},
-      {"", NULL},
-  };
-
-  printf("loading library %s\n", libPath);
-  s->handle = LOAD_LIBRARY(libPath, RTLD_LOCAL|RTLD_NOW);
-  if (!s->handle) {
-    err->id = -1;
-    char *msg = LOAD_ERR();
-    snprintf(err->msg, err->msg_len,
-             "Unable to load dynamic server library: %s", msg);
-    free(msg);
-    return;
-  }
-
-  for (i = 0; l[i].p != NULL; i++) {
-    *l[i].p = LOAD_SYMBOL(s->handle, l[i].s);
-    if (!l[i].p) {
-      UNLOAD_LIBRARY(s->handle);
-      err->id = -1;
-      char *msg = LOAD_ERR();
-      snprintf(err->msg, err->msg_len, "symbol lookup for %s failed: %s",
-               l[i].s, msg);
-      free(msg);
-      return;
-    }
-  }
-}
-
-inline void dyn_llama_server_init(struct dynamic_llama_server s,
-                                           ext_server_params_t *sparams,
-                                           ext_server_resp_t *err) {
-  s.llama_server_init(sparams, err);
-}
-
-inline void dyn_llama_server_start(struct dynamic_llama_server s) {
-  s.llama_server_start();
-}
-
-inline void dyn_llama_server_stop(struct dynamic_llama_server s) {
-  s.llama_server_stop();
-}
-
-inline void dyn_llama_server_completion(struct dynamic_llama_server s,
-                                                 const char *json_req,
-                                                 ext_server_resp_t *resp) {
-  s.llama_server_completion(json_req, resp);
-}
-
-inline void dyn_llama_server_completion_next_result(
-    struct dynamic_llama_server s, const int task_id,
-    ext_server_task_result_t *result) {
-  s.llama_server_completion_next_result(task_id, result);
-}
-
-inline void dyn_llama_server_completion_cancel(
-    struct dynamic_llama_server s, const int task_id, ext_server_resp_t *err) {
-  s.llama_server_completion_cancel(task_id, err);
-}
-inline void dyn_llama_server_release_task_result(
-    struct dynamic_llama_server s, ext_server_task_result_t *result) {
-  s.llama_server_release_task_result(result);
-}
-
-inline void dyn_llama_server_tokenize(struct dynamic_llama_server s,
-                                               const char *json_req,
-                                               char **json_resp,
-                                               ext_server_resp_t *err) {
-  s.llama_server_tokenize(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_detokenize(struct dynamic_llama_server s,
-                                                 const char *json_req,
-                                                 char **json_resp,
-                                                 ext_server_resp_t *err) {
-  s.llama_server_detokenize(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_embedding(struct dynamic_llama_server s,
-                                                const char *json_req,
-                                                char **json_resp,
-                                                ext_server_resp_t *err) {
-  s.llama_server_embedding(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_release_json_resp(
-    struct dynamic_llama_server s, char **json_resp) {
-  s.llama_server_release_json_resp(json_resp);
-}

llm/dyn_ext_server.go

@@ -1,368 +0,0 @@
-package llm
-
-/*
-#cgo CFLAGS: -I${SRCDIR}/ext_server -I${SRCDIR}/llama.cpp -I${SRCDIR}/llama.cpp/common -I${SRCDIR}/llama.cpp/examples/server
-#cgo CFLAGS: -DNDEBUG -DLLAMA_SERVER_LIBRARY=1 -D_XOPEN_SOURCE=600 -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
-#cgo CFLAGS: -Wmissing-noreturn -Wextra -Wcast-qual -Wno-unused-function -Wno-array-bounds
-#cgo CPPFLAGS: -Ofast -Wextra -Wno-unused-function -Wno-unused-variable -Wno-deprecated-declarations
-#cgo darwin CFLAGS: -D_DARWIN_C_SOURCE
-#cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
-#cgo darwin CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
-#cgo darwin LDFLAGS: -lc++ -framework Accelerate
-#cgo darwin LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-#cgo linux CFLAGS: -D_GNU_SOURCE
-#cgo linux LDFLAGS: -lrt -ldl -lstdc++ -lm
-#cgo linux windows LDFLAGS: -lpthread
-
-#include <stdlib.h>
-#include "dyn_ext_server.h"
-
-*/
-import "C"
-
-import (
-	"bytes"
-	"context"
-	"encoding/json"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"strings"
-	"sync"
-	"time"
-	"unsafe"
-
-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/gpu"
-)
-
-type dynExtServer struct {
-	s       C.struct_dynamic_llama_server
-	options api.Options
-}
-
-// Note: current implementation does not support concurrent instantiations
-var mutex sync.Mutex
-
-func newExtServerResp(len C.size_t) C.ext_server_resp_t {
-	var resp C.ext_server_resp_t
-	resp.msg_len = len
-	bytes := make([]byte, len)
-	resp.msg = (*C.char)(C.CBytes(bytes))
-	return resp
-}
-
-func freeExtServerResp(resp C.ext_server_resp_t) {
-	if resp.msg_len == 0 {
-		return
-	}
-	C.free(unsafe.Pointer(resp.msg))
-}
-
-func extServerResponseToErr(resp C.ext_server_resp_t) error {
-	return fmt.Errorf(C.GoString(resp.msg))
-}
-
-// Note: current implementation does not support concurrent instantiations
-var llm *dynExtServer
-
-func newDynExtServer(library, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
-	if !mutex.TryLock() {
-		slog.Info("concurrent llm servers not yet supported, waiting for prior server to complete")
-		mutex.Lock()
-	}
-	gpu.UpdatePath(filepath.Dir(library))
-	libPath := C.CString(library)
-	defer C.free(unsafe.Pointer(libPath))
-	resp := newExtServerResp(512)
-	defer freeExtServerResp(resp)
-	var srv C.struct_dynamic_llama_server
-	C.dyn_init(libPath, &srv, &resp)
-	if resp.id < 0 {
-		mutex.Unlock()
-		return nil, fmt.Errorf("Unable to load dynamic library: %s", C.GoString(resp.msg))
-	}
-	llm = &dynExtServer{
-		s:       srv,
-		options: opts,
-	}
-	slog.Info(fmt.Sprintf("Loading Dynamic llm server: %s", library))
-
-	var sparams C.ext_server_params_t
-	sparams.model = C.CString(model)
-	defer C.free(unsafe.Pointer(sparams.model))
-
-	sparams.embedding = true
-	sparams.n_ctx = C.uint(opts.NumCtx)
-	sparams.n_batch = C.uint(opts.NumBatch)
-	sparams.n_gpu_layers = C.int(opts.NumGPU)
-	sparams.main_gpu = C.int(opts.MainGPU)
-	sparams.n_parallel = 1 // TODO - wire up concurrency
-
-	// Always use the value encoded in the model
-	sparams.rope_freq_base = 0.0
-	sparams.rope_freq_scale = 0.0
-	sparams.memory_f16 = C.bool(opts.F16KV)
-	sparams.use_mlock = C.bool(opts.UseMLock)
-	sparams.use_mmap = C.bool(opts.UseMMap)
-
-	if opts.UseNUMA {
-		sparams.numa = C.int(1)
-	} else {
-		sparams.numa = C.int(0)
-	}
-
-	sparams.lora_adapters = nil
-	for i := 0; i < len(adapters); i++ {
-		la := (*C.ext_server_lora_adapter_t)(C.malloc(C.sizeof_ext_server_lora_adapter_t))
-		defer C.free(unsafe.Pointer(la))
-		la.adapter = C.CString(adapters[i])
-		defer C.free(unsafe.Pointer(la.adapter))
-		la.scale = C.float(1.0) // TODO expose scale/weights up through ollama UX
-		la.next = nil
-		if i == 0 {
-			sparams.lora_adapters = la
-		} else {
-			tmp := sparams.lora_adapters
-			for ; tmp.next != nil; tmp = tmp.next {
-			}
-			tmp.next = la
-		}
-	}
-
-	if len(projectors) > 0 {
-		// TODO: applying multiple projectors is not supported by the llama.cpp server yet
-		sparams.mmproj = C.CString(projectors[0])
-		defer C.free(unsafe.Pointer(sparams.mmproj))
-	} else {
-		sparams.mmproj = nil
-	}
-
-	sparams.n_threads = C.uint(opts.NumThread)
-
-	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
-		sparams.verbose_logging = C.bool(true)
-	} else {
-		sparams.verbose_logging = C.bool(false)
-	}
-
-	slog.Info("Initializing llama server")
-	slog.Debug(fmt.Sprintf("server params: %+v", sparams))
-	initResp := newExtServerResp(128)
-	defer freeExtServerResp(initResp)
-	C.dyn_llama_server_init(llm.s, &sparams, &initResp)
-	if initResp.id < 0 {
-		mutex.Unlock()
-		err := extServerResponseToErr(initResp)
-		slog.Debug(fmt.Sprintf("failure during initialization: %s", err))
-		return nil, err
-	}
-
-	slog.Info("Starting llama main loop")
-	C.dyn_llama_server_start(llm.s)
-	return llm, nil
-}
-
-func (llm *dynExtServer) Predict(ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-
-	if len(predict.Images) > 0 {
-		slog.Info(fmt.Sprintf("loaded %d images", len(predict.Images)))
-	}
-
-	request := map[string]any{
-		"prompt":            predict.Prompt,
-		"stream":            true,
-		"n_predict":         predict.Options.NumPredict,
-		"n_keep":            predict.Options.NumKeep,
-		"temperature":       predict.Options.Temperature,
-		"top_k":             predict.Options.TopK,
-		"top_p":             predict.Options.TopP,
-		"tfs_z":             predict.Options.TFSZ,
-		"typical_p":         predict.Options.TypicalP,
-		"repeat_last_n":     predict.Options.RepeatLastN,
-		"repeat_penalty":    predict.Options.RepeatPenalty,
-		"presence_penalty":  predict.Options.PresencePenalty,
-		"frequency_penalty": predict.Options.FrequencyPenalty,
-		"mirostat":          predict.Options.Mirostat,
-		"mirostat_tau":      predict.Options.MirostatTau,
-		"mirostat_eta":      predict.Options.MirostatEta,
-		"penalize_nl":       predict.Options.PenalizeNewline,
-		"seed":              predict.Options.Seed,
-		"stop":              predict.Options.Stop,
-		"image_data":        predict.Images,
-		"cache_prompt":      true,
-	}
-
-	if predict.Format == "json" {
-		request["grammar"] = jsonGrammar
-	}
-
-	retryDelay := 100 * time.Microsecond
-	for retries := 0; retries < maxRetries; retries++ {
-		if retries > 0 {
-			time.Sleep(retryDelay) // wait before retrying
-			retryDelay *= 2        // exponential backoff
-		}
-
-		// Handling JSON marshaling with special characters unescaped.
-		buffer := &bytes.Buffer{}
-		enc := json.NewEncoder(buffer)
-		enc.SetEscapeHTML(false)
-
-		if err := enc.Encode(request); err != nil {
-			return fmt.Errorf("failed to marshal data: %w", err)
-		}
-
-		req := C.CString(buffer.String())
-		defer C.free(unsafe.Pointer(req))
-
-		C.dyn_llama_server_completion(llm.s, req, &resp)
-		if resp.id < 0 {
-			return extServerResponseToErr(resp)
-		}
-
-		retryNeeded := false
-	out:
-		for {
-			select {
-			case <-ctx.Done():
-				// This handles the request cancellation
-				C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
-				if resp.id < 0 {
-					return extServerResponseToErr(resp)
-				} else {
-					return nil
-				}
-			default:
-				var result C.ext_server_task_result_t
-				C.dyn_llama_server_completion_next_result(llm.s, resp.id, &result)
-				json_resp := C.GoString(result.json_resp)
-				C.dyn_llama_server_release_task_result(llm.s, &result)
-
-				var p prediction
-				if err := json.Unmarshal([]byte(json_resp), &p); err != nil {
-					C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
-					if resp.id < 0 {
-						return fmt.Errorf("error unmarshaling llm prediction response: %w and cancel %s", err, C.GoString(resp.msg))
-					} else {
-						return fmt.Errorf("error unmarshaling llm prediction response: %w", err)
-					}
-				}
-
-				if bool(result.error) && strings.Contains(json_resp, "slot unavailable") {
-					retryNeeded = true
-					// task will already be canceled
-					break out
-				}
-
-				if p.Content != "" {
-					fn(PredictResult{
-						Content: p.Content,
-					})
-				}
-
-				if p.Stop || bool(result.stop) {
-					fn(PredictResult{
-						Done:               true,
-						PromptEvalCount:    p.Timings.PromptN,
-						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
-						EvalCount:          p.Timings.PredictedN,
-						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
-					})
-					return nil
-				}
-			}
-		}
-		if !retryNeeded {
-			return nil // success
-		}
-	}
-
-	// should never reach here ideally
-	return fmt.Errorf("max retries exceeded")
-}
-
-func (llm *dynExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
-	data, err := json.Marshal(TokenizeRequest{Content: prompt})
-	if err != nil {
-		return nil, fmt.Errorf("marshaling encode data: %w", err)
-	}
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_tokenize(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return nil, extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var encoded TokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &encoded); err2 != nil {
-		return nil, fmt.Errorf("unmarshal encode response: %w", err2)
-	}
-
-	return encoded.Tokens, err
-}
-
-func (llm *dynExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
-	if len(tokens) == 0 {
-		return "", nil
-	}
-	data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
-	if err != nil {
-		return "", fmt.Errorf("marshaling decode data: %w", err)
-	}
-
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_detokenize(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return "", extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var decoded DetokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &decoded); err2 != nil {
-		return "", fmt.Errorf("unmarshal encode response: %w", err2)
-	}
-
-	return decoded.Content, err
-}
-
-func (llm *dynExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
-	data, err := json.Marshal(TokenizeRequest{Content: input})
-	if err != nil {
-		return nil, fmt.Errorf("error marshaling embed data: %w", err)
-	}
-
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_embedding(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return nil, extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var embedding EmbeddingResponse
-	if err := json.Unmarshal([]byte(C.GoString(json_resp)), &embedding); err != nil {
-		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
-	}
-
-	return embedding.Embedding, nil
-}
-
-func (llm *dynExtServer) Close() {
-	C.dyn_llama_server_stop(llm.s)
-	mutex.Unlock()
-}

llm/dyn_ext_server.h

@@ -1,74 +0,0 @@
-#include <stdlib.h>
-
-#include "ext_server.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-struct dynamic_llama_server {
-  void *handle;
-  void (*llama_server_init)(ext_server_params_t *sparams,
-                            ext_server_resp_t *err);
-  void (*llama_server_start)();
-  void (*llama_server_stop)();
-  void (*llama_server_completion)(const char *json_req,
-                                  ext_server_resp_t *resp);
-  void (*llama_server_completion_next_result)(const int task_id,
-                                              ext_server_task_result_t *result);
-  void (*llama_server_completion_cancel)(const int task_id,
-                                         ext_server_resp_t *err);
-  void (*llama_server_release_task_result)(ext_server_task_result_t *result);
-  void (*llama_server_tokenize)(const char *json_req, char **json_resp,
-                                ext_server_resp_t *err);
-  void (*llama_server_detokenize)(const char *json_req, char **json_resp,
-                                  ext_server_resp_t *err);
-  void (*llama_server_embedding)(const char *json_req, char **json_resp,
-                                 ext_server_resp_t *err);
-  void (*llama_server_release_json_resp)(char **json_resp);
-};
-
-void dyn_init(const char *libPath, struct dynamic_llama_server *s,
-                       ext_server_resp_t *err);
-
-// No good way to call C function pointers from Go so inline the indirection
-void dyn_llama_server_init(struct dynamic_llama_server s,
-                                    ext_server_params_t *sparams,
-                                    ext_server_resp_t *err);
-
-void dyn_llama_server_start(struct dynamic_llama_server s);
-
-void dyn_llama_server_stop(struct dynamic_llama_server s);
-
-void dyn_llama_server_completion(struct dynamic_llama_server s,
-                                          const char *json_req,
-                                          ext_server_resp_t *resp);
-
-void dyn_llama_server_completion_next_result(
-    struct dynamic_llama_server s, const int task_id,
-    ext_server_task_result_t *result);
-
-void dyn_llama_server_completion_cancel(struct dynamic_llama_server s,
-                                                 const int task_id,
-                                                 ext_server_resp_t *err);
-
-void dyn_llama_server_release_task_result(
-    struct dynamic_llama_server s, ext_server_task_result_t *result);
-
-void dyn_llama_server_tokenize(struct dynamic_llama_server s,
-                                        const char *json_req, char **json_resp,
-                                        ext_server_resp_t *err);
-
-void dyn_llama_server_detokenize(struct dynamic_llama_server s,
-                                          const char *json_req,
-                                          char **json_resp,
-                                          ext_server_resp_t *err);
-
-void dyn_llama_server_embedding(struct dynamic_llama_server s,
-                                         const char *json_req, char **json_resp,
-                                         ext_server_resp_t *err);
-void dyn_llama_server_release_json_resp(struct dynamic_llama_server s,
-                                                 char **json_resp);
-
-#ifdef __cplusplus
-}
-#endif

llm/ext_server/CMakeLists.txt

@@ -1,25 +1,14 @@
-# Ollama specific CMakefile to include in llama.cpp/examples/server
 
-set(TARGET ext_server)
+set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
-    add_library(${TARGET} SHARED ../../../ext_server/ext_server.cpp ../../llama.cpp)
-else()
-    add_library(${TARGET} STATIC ../../../ext_server/ext_server.cpp ../../llama.cpp)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_include_directories(${TARGET} PRIVATE ../..)
-target_include_directories(${TARGET} PRIVATE ../../..)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-target_compile_definitions(${TARGET} PUBLIC LLAMA_SERVER_LIBRARY=1)
-target_link_libraries(${TARGET} PRIVATE ggml llava common )
-set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>)
-install(TARGETS ext_server LIBRARY)
-
-if (CUDAToolkit_FOUND)
-    target_include_directories(${TARGET} PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-    if (WIN32)
-        target_link_libraries(${TARGET} PRIVATE nvml)
-    endif()
-endif()
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

llm/ext_server/README.md

@@ -1,18 +0,0 @@
-# Extern C Server
-
-This directory contains a thin facade we layer on top of the Llama.cpp server to
-expose `extern C` interfaces to access the functionality through direct API
-calls in-process.  The llama.cpp code uses compile time macros to configure GPU
-type along with other settings.  During the `go generate ./...` execution, the
-build will generate one or more copies of the llama.cpp `extern C` server based
-on what GPU libraries are detected to support multiple GPU types as well as CPU
-only support. The Ollama go build then embeds these different servers to support
-different GPUs and settings at runtime.
-
-If you are making changes to the code in this directory, make sure to disable
-caching during your go build to ensure you pick up your changes.  A typical
-iteration cycle from the top of the source tree looks like:
-
-```
-go generate ./... && go build -a .
-```

llm/ext_server/ext_server.cpp

@@ -1,381 +0,0 @@
-#include "ext_server.h"
-#include <atomic>
-
-// Necessary evil since the server types are not defined in a header
-#include "server.cpp"
-
-// Low level API access to verify GPU access
-#if defined(GGML_USE_CUBLAS)
-#if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define cudaGetDevice hipGetDevice
-#define cudaError_t hipError_t
-#define cudaSuccess hipSuccess
-#define cudaGetErrorString hipGetErrorString
-#else
-#include <cuda_runtime.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-#endif // defined(GGML_USE_HIPBLAS)
-#endif // GGML_USE_CUBLAS
-
-// Expose the llama server as a callable extern "C" API
-server_context *llama = NULL;
-std::thread ext_server_thread;
-bool shutting_down = false;
-std::atomic_int recv_counter;
-
-// RAII wrapper for tracking in-flight recv calls
-class atomicRecv {
-  public:
-    atomicRecv(std::atomic<int> &atomic) : atomic(atomic) {
-      ++this->atomic;
-    }
-    ~atomicRecv() {
-      --this->atomic;
-    }
-  private:
-    std::atomic<int> &atomic;
-};
- 
-void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
-  recv_counter = 0;
-  assert(err != NULL && sparams != NULL);
-  log_set_target(stderr);
-  if (!sparams->verbose_logging) {
-    server_verbose = true;
-    log_disable();
-  }
-
-  LOG_TEE("system info: %s\n", llama_print_system_info());
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    llama = new server_context;
-    gpt_params params;
-    params.n_ctx = sparams->n_ctx;
-    params.n_batch = sparams->n_batch;
-    if (sparams->n_threads > 0) {
-      params.n_threads = sparams->n_threads;
-    }
-    params.n_parallel = sparams->n_parallel;
-    params.rope_freq_base = sparams->rope_freq_base;
-    params.rope_freq_scale = sparams->rope_freq_scale;
-
-    if (sparams->memory_f16) {
-      params.cache_type_k = "f16";
-      params.cache_type_v = "f16";
-    } else {
-      params.cache_type_k = "f32";
-      params.cache_type_v = "f32";
-    }
-
-    params.n_gpu_layers = sparams->n_gpu_layers;
-    params.main_gpu = sparams->main_gpu;
-    params.use_mlock = sparams->use_mlock;
-    params.use_mmap = sparams->use_mmap;
-    params.numa = (ggml_numa_strategy)sparams->numa;
-    params.embedding = sparams->embedding;
-    if (sparams->model != NULL) {
-      params.model = sparams->model;
-    }
-
-    if (sparams->lora_adapters != NULL) {
-      for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL;
-          la = la->next) {
-        params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
-      }
-
-      params.use_mmap = false;
-    }
-
-    if (sparams->mmproj != NULL) {
-      params.mmproj = std::string(sparams->mmproj);
-    }
-
-#if defined(GGML_USE_CUBLAS)
-    // Before attempting to init the backend which will assert on error, verify the CUDA/ROCM GPU is accessible
-    LOG_TEE("Performing pre-initialization of GPU\n");
-    int id;
-    cudaError_t cudaErr = cudaGetDevice(&id);
-    if (cudaErr != cudaSuccess) {
-      err->id = -1;
-      snprintf(err->msg, err->msg_len, "Unable to init GPU: %s", cudaGetErrorString(cudaErr));
-      return;
-    }
-#endif
-
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    // load the model
-    if (!llama->load_model(params)) {
-      // TODO - consider modifying the logging logic or patching load_model so
-      // we can capture more detailed error messages and pass them back to the
-      // caller for better UX
-      err->id = -1;
-      snprintf(err->msg, err->msg_len, "error loading model %s",
-               params.model.c_str());
-      return;
-    }
-
-    llama->init();
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len,
-             "Unknown exception initializing llama server");
-  }
-}
-
-void llama_server_start() {
-  assert(llama != NULL);
-  // TODO mutex to protect thread creation
-  ext_server_thread = std::thread([&]() {
-    try {
-      LOG_TEE("llama server main loop starting\n");
-      ggml_time_init();
-      llama->queue_tasks.on_new_task(std::bind(
-        &server_context::process_single_task, llama, std::placeholders::_1));
-      llama->queue_tasks.on_finish_multitask(std::bind(
-        &server_context::on_finish_multitask, llama, std::placeholders::_1));
-      llama->queue_tasks.on_run_slots(std::bind(
-        &server_context::update_slots, llama));
-      llama->queue_results.on_multitask_update(std::bind(
-          &server_queue::update_multitask,
-          &llama->queue_tasks,
-          std::placeholders::_1,
-          std::placeholders::_2,
-          std::placeholders::_3
-        ));
-      llama->queue_tasks.start_loop();
-    } catch (std::exception &e) {
-      LOG_TEE("caught exception in llama server main loop: %s\n", e.what());
-    } catch (...) {
-      LOG_TEE("caught unknown exception in llama server main loop\n");
-    }
-    LOG_TEE("\nllama server shutting down\n");
-    llama_backend_free();
-  });
-}
-
-void llama_server_stop() {
-  assert(llama != NULL);
-  // Shutdown any in-flight requests and block incoming requests.
-  LOG_TEE("\ninitiating shutdown - draining remaining tasks...\n");
-  shutting_down = true;
-
-  while (recv_counter.load() > 0) {
-    std::this_thread::sleep_for(std::chrono::milliseconds(50));
-  }
-
-  // This may take a while for any pending tasks to drain
-  // TODO - consider a timeout to cancel tasks if it's taking too long
-  llama->queue_tasks.terminate();
-  ext_server_thread.join();
-  delete llama;
-  llama = NULL;
-  LOG_TEE("llama server shutdown complete\n");
-  shutting_down = false;
-}
-
-void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
-  assert(llama != NULL && json_req != NULL && resp != NULL);
-  resp->id = -1;
-  resp->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    json data = json::parse(json_req);
-    resp->id = llama->queue_tasks.get_new_id();
-    llama->queue_results.add_waiting_task_id(resp->id);
-    llama->request_completion(resp->id, -1, data, false, false);
-  } catch (std::exception &e) {
-    snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
-  } catch (...) {
-    snprintf(resp->msg, resp->msg_len, "Unknown exception during completion");
-  }
-}
-
-void llama_server_completion_next_result(const int task_id,
-                                         ext_server_task_result_t *resp) {
-  assert(llama != NULL && resp != NULL);
-  resp->id = -1;
-  resp->stop = false;
-  resp->error = false;
-  resp->json_resp = NULL;
-  std::string result_json;
-  try {
-    atomicRecv ar(recv_counter);
-    server_task_result result = llama->queue_results.recv(task_id);
-    result_json =
-        result.data.dump(-1, ' ', false, json::error_handler_t::replace);
-    resp->id = result.id;
-    resp->stop = result.stop;
-    resp->error = result.error;
-    if (result.error) {
-      LOG_TEE("next result cancel on error\n");
-      llama->request_cancel(task_id);
-      LOG_TEE("next result removing waiting tak ID: %d\n", task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-    } else if (result.stop) {
-      LOG_TEE("next result cancel on stop\n");
-      llama->request_cancel(task_id);
-      LOG_TEE("next result removing waiting task ID: %d\n", task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-    } else if (shutting_down) {
-      LOG_TEE("aborting completion due to shutdown %d\n", task_id);
-      llama->request_cancel(task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-      resp->stop = true;
-    }
-  } catch (std::exception &e) {
-    resp->error = true;
-    resp->id = -1;
-    result_json = "{\"error\":\"exception " + std::string(e.what()) + "\"}";
-    LOG_TEE("llama server completion exception %s\n", e.what());
-  } catch (...) {
-    resp->error = true;
-    resp->id = -1;
-    result_json = "{\"error\":\"Unknown exception during completion\"}";
-    LOG_TEE("llama server completion unknown exception\n");
-  }
-  const std::string::size_type size = result_json.size() + 1;
-  resp->json_resp = new char[size];
-  snprintf(resp->json_resp, size, "%s", result_json.c_str());
-}
-
-void llama_server_release_task_result(ext_server_task_result_t *result) {
-  if (result == NULL || result->json_resp == NULL) {
-    return;
-  }
-  delete[] result->json_resp;
-}
-
-void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err) {
-  assert(llama != NULL && err != NULL);
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    llama->request_cancel(task_id);
-    llama->queue_results.remove_waiting_task_id(task_id);
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len,
-             "Unknown exception completion cancel in llama server");
-  }
-}
-
-void llama_server_tokenize(const char *json_req, char **json_resp,
-                           ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    std::vector<llama_token> tokens;
-    if (body.count("content") != 0) {
-      tokens = llama->tokenize(body["content"], false);
-    }
-    const json data = format_tokenizer_response(tokens);
-    std::string result_json = data.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during tokenize");
-  }
-}
-
-void llama_server_release_json_resp(char **json_resp) {
-  if (json_resp == NULL || *json_resp == NULL) {
-    return;
-  }
-  delete[] *json_resp;
-}
-
-void llama_server_detokenize(const char *json_req, char **json_resp,
-                             ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    std::string content;
-    if (body.count("tokens") != 0) {
-      const std::vector<llama_token> tokens = body["tokens"];
-      content = tokens_to_str(llama->ctx, tokens.cbegin(), tokens.cend());
-    }
-    const json data = format_detokenized_response(content);
-    std::string result_json = data.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during detokenize");
-  }
-}
-
-void llama_server_embedding(const char *json_req, char **json_resp,
-                            ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    json prompt;
-    if (body.count("content") != 0) {
-      prompt = body["content"];
-    } else {
-      prompt = "";
-    }
-    const int task_id = llama->queue_tasks.get_new_id();
-    llama->queue_results.add_waiting_task_id(task_id);
-    llama->request_completion(task_id, -1, {{"prompt", prompt}, {"n_predict", 0}}, false, true);
-    atomicRecv ar(recv_counter);
-    server_task_result result = llama->queue_results.recv(task_id);
-    std::string result_json = result.data.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-    llama->queue_results.remove_waiting_task_id(task_id);
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during embedding");
-  }
-}

llm/ext_server/ext_server.h

@@ -1,95 +0,0 @@
-#if defined(LLAMA_SERVER_LIBRARY)
-#ifndef LLAMA_SERVER_H
-#define LLAMA_SERVER_H
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-#include <stdio.h>
-
-int __main(int argc, char **argv);
-
-// This exposes extern C entrypoints into the llama_server
-// To enable the server compile with LLAMA_SERVER_LIBRARY
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-typedef struct ext_server_resp {
-  int id;          // < 0 on error
-  size_t msg_len;  // caller must allocate msg and set msg_len
-  char *msg;
-} ext_server_resp_t;
-
-// Allocated and freed by caller
-typedef struct ext_server_lora_adapter {
-  char *adapter;
-  float scale;
-  struct ext_server_lora_adapter *next;
-} ext_server_lora_adapter_t;
-
-// Allocated and freed by caller
-typedef struct ext_server_params {
-  char *model;
-  uint32_t n_ctx;         // token context window, 0 = from model
-  uint32_t n_batch;       // prompt processing maximum batch size
-  uint32_t n_threads;     // number of threads to use for generation
-  int32_t n_parallel;     // number of parallel sequences to decodewra
-  float rope_freq_base;   // RoPE base frequency, 0 = from model
-  float rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
-  bool memory_f16;        // use f16 instead of f32 for memory kv
-  int32_t n_gpu_layers;  // number of layers to store in VRAM (-1 - use default)
-  int32_t main_gpu;      // the GPU that is used for scratch and small tensors
-  bool use_mlock;        // force system to keep model in RAM
-  bool use_mmap;         // use mmap if possible
-  int numa;              // attempt optimizations that help on some NUMA systems
-  bool embedding;        // get only sentence embedding
-  ext_server_lora_adapter_t *lora_adapters;
-  char *mmproj;
-  bool verbose_logging;  // Enable verbose logging of the server
-} ext_server_params_t;
-
-typedef struct ext_server_task_result {
-  int id;
-  bool stop;
-  bool error;
-  char *json_resp;  // null terminated, memory managed by ext_server
-} ext_server_task_result_t;
-
-// Initialize the server once per process
-// err->id = 0 for success and err->msg[0] = NULL
-// err->id != 0 for failure, and err->msg contains error message
-void llama_server_init(ext_server_params_t *sparams, ext_server_resp_t *err);
-
-// Run the main loop, called once per init
-void llama_server_start();
-// Stop the main loop and free up resources allocated in init and start.  Init
-// must be called again to reuse
-void llama_server_stop();
-
-// json_req null terminated string, memory managed by caller
-// resp->id >= 0 on success (task ID)
-// resp->id < 0 on error, and resp->msg contains error message
-void llama_server_completion(const char *json_req, ext_server_resp_t *resp);
-
-// Caller must call llama_server_release_task_result to free resp->json_resp
-void llama_server_completion_next_result(const int task_id,
-                                         ext_server_task_result_t *result);
-void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err);
-void llama_server_release_task_result(ext_server_task_result_t *result);
-
-// Caller must call llama_server_releaes_json_resp to free json_resp if err.id <
-// 0
-void llama_server_tokenize(const char *json_req, char **json_resp,
-                           ext_server_resp_t *err);
-void llama_server_detokenize(const char *json_req, char **json_resp,
-                             ext_server_resp_t *err);
-void llama_server_embedding(const char *json_req, char **json_resp,
-                            ext_server_resp_t *err);
-void llama_server_release_json_resp(char **json_resp);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
-#endif  // LLAMA_SERVER_LIBRARY

llm/ext_server/utils.hpp

@@ -0,0 +1,655 @@
+// MIT License
+
+// Copyright (c) 2023 Georgi Gerganov
+
+// 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 <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+
+#include "../llava/clip.h"
+
+using json = nlohmann::json;
+
+extern bool server_verbose;
+extern bool server_log_json;
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+#if SERVER_VERBOSE != 1
+#define LOG_VERBOSE(MSG, ...)
+#else
+#define LOG_VERBOSE(MSG, ...)                                            \
+    do                                                                   \
+    {                                                                    \
+        if (server_verbose)                                              \
+        {                                                                \
+            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while (0)
+#endif
+
+#define LOG_ERROR(  MSG, ...) server_log("ERR",  __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE,
+    TASK_TYPE_METRICS
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// completion token output with probabilities
+struct completion_token_output {
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+struct token_translator {
+    llama_context * ctx;
+    std::string operator()(llama_token tok)                    const { return llama_token_to_piece(ctx, tok); }
+    std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); }
+};
+
+static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) {
+    std::stringstream ss_tid;
+    ss_tid << std::this_thread::get_id();
+    json log = nlohmann::ordered_json{
+        {"tid", ss_tid.str()},
+        {"timestamp", time(nullptr)},
+    };
+
+    if (server_log_json) {
+        log.merge_patch(
+                {
+                        {"level",     level},
+                        {"function",  function},
+                        {"line",      line},
+                        {"msg",       message},
+                });
+        if (!extra.empty()) {
+            log.merge_patch(extra);
+        }
+
+        std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
+    } else {
+        char buf[1024];
+        snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
+
+        if (!extra.empty()) {
+            log.merge_patch(extra);
+        }
+        std::stringstream ss;
+        ss << buf << " |";
+        for (const auto& el : log.items())
+        {
+            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
+            ss << " " << el.key() << "=" << value;
+        }
+
+        const std::string str = ss.str();
+        printf("%.*s\n", (int)str.size(), str.data());
+        fflush(stdout);
+    }
+}
+
+//
+// server utils
+//
+
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value) {
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
+// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
+inline bool verify_custom_template(const std::string & tmpl) {
+    llama_chat_message chat[] = {{"user", "test"}};
+    std::vector<char> buf(1);
+    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
+    return res >= 0;
+}
+
+// Format given chat. If tmpl is empty, we take the template from model metadata
+inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
+    size_t alloc_size = 0;
+    // vector holding all allocated string to be passed to llama_chat_apply_template
+    std::vector<std::string> str(messages.size() * 2);
+    std::vector<llama_chat_message> chat(messages.size());
+
+    for (size_t i = 0; i < messages.size(); ++i) {
+        auto &curr_msg = messages[i];
+        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
+        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
+        alloc_size     += str[i*2 + 1].length();
+        chat[i].role    = str[i*2 + 0].c_str();
+        chat[i].content = str[i*2 + 1].c_str();
+    }
+
+    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
+    std::vector<char> buf(alloc_size * 2);
+
+    // run the first time to get the total output length
+    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+
+    // if it turns out that our buffer is too small, we resize it
+    if ((size_t) res > buf.size()) {
+        buf.resize(res);
+        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+    }
+
+    std::string formatted_chat(buf.data(), res);
+    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
+
+    return formatted_chat;
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    bool running;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_run_slots;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+            LOG_VERBOSE("new task id", {{"new_id", task.id}});
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        int new_id = id++;
+        LOG_VERBOSE("new task id", {{"new_id", new_id}});
+        return new_id;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask when it is finished
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when all slots data is ready to be processed
+    void on_run_slots(std::function<void(void)> callback) {
+        callback_run_slots = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // end the start_loop routine
+    void terminate() {
+        {
+            std::unique_lock<std::mutex> lock(mutex_tasks);
+            running = false;
+        }
+        condition_tasks.notify_all();
+    }
+
+    /**
+     * Main loop consists of these steps:
+     * - Wait until a new task arrives
+     * - Process the task (i.e. maybe copy data into slot)
+     * - Check if multitask is finished
+     * - Run all slots
+     */
+    void start_loop() {
+        running = true;
+        while (true) {
+            LOG_VERBOSE("new task may arrive", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("update_multitasks", {});
+                // check if we have any finished multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is processed, slots data is now ready
+                LOG_VERBOSE("callback_run_slots", {});
+                callback_run_slots();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    if (!running) {
+                        LOG_VERBOSE("ending start_loop", {});
+                        return;
+                    }
+                    condition_tasks.wait(lock, [&]{
+                        return (!queue_tasks.empty() || !running);
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    // add the task_id to the list of tasks waiting for response
+    void add_waiting_task_id(int task_id) {
+        LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    // when the request is finished, we can remove task associated with it
+    void remove_waiting_task_id(int task_id) {
+        LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
+        }
+
+        // should never reach here
+    }
+
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
+    }
+
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {{"task_id", result.id}});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
+
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
+                queue_results.push_back(result);
+                condition_results.notify_all();
+                return;
+            }
+        }
+    }
+};
+
+//
+// base64 utils (TODO: move to common in the future)
+//
+
+static const std::string base64_chars =
+             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+             "abcdefghijklmnopqrstuvwxyz"
+             "0123456789+/";
+
+static inline bool is_base64(uint8_t c)
+{
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
+    int i = 0;
+    int j = 0;
+    int in_ = 0;
+
+    int in_len = encoded_string.size();
+
+    uint8_t char_array_4[4];
+    uint8_t char_array_3[3];
+
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
+        char_array_4[i++] = encoded_string[in_]; in_++;
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
+                char_array_4[i] = base64_chars.find(char_array_4[i]);
+            }
+
+            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+            for (i = 0; (i < 3); i++)
+            {
+                ret.push_back(char_array_3[i]);
+            }
+            i = 0;
+        }
+    }
+
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
+            char_array_4[j] = 0;
+        }
+
+        for (j = 0; j <4; j++)
+        {
+            char_array_4[j] = base64_chars.find(char_array_4[j]);
+        }
+
+        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+        for (j = 0; (j < i - 1); j++)
+        {
+            ret.push_back(char_array_3[j]);
+        }
+    }
+
+    return ret;
+}
+
+//
+// random string / id
+//
+
+static std::string random_string()
+{
+    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
+
+    std::random_device rd;
+    std::mt19937 generator(rd());
+
+    std::string result(32, ' ');
+
+    for (int i = 0; i < 32; ++i) {
+        result[i] = str[generator() % str.size()];
+    }
+
+    return result;
+}
+
+static std::string gen_chatcmplid()
+{
+    std::stringstream chatcmplid;
+    chatcmplid << "chatcmpl-" << random_string();
+    return chatcmplid.str();
+}
+
+//
+// other common utils
+//
+
+static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
+{
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
+    {
+    }
+    return i;
+}
+
+static bool ends_with(const std::string &str, const std::string &suffix)
+{
+    return str.size() >= suffix.size() &&
+           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+}
+
+static size_t find_partial_stop_string(const std::string &stop,
+                                       const std::string &text)
+{
+    if (!text.empty() && !stop.empty())
+    {
+        const char text_last_char = text.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
+        {
+            if (stop[char_index] == text_last_char)
+            {
+                const std::string current_partial = stop.substr(0, char_index + 1);
+                if (ends_with(text, current_partial))
+                {
+                    return text.size() - char_index - 1;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
+
+// TODO: reuse llama_detokenize
+template <class Iter>
+static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
+{
+    std::string ret;
+    for (; begin != end; ++begin)
+    {
+        ret += llama_token_to_piece(ctx, *begin);
+    }
+    return ret;
+}
+
+// format incomplete utf-8 multibyte character for output
+static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
+{
+    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
+    // if the size is 1 and first bit is 1, meaning it's a partial character
+    //   (size > 1 meaning it's already a known token)
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
+    {
+        std::stringstream ss;
+        ss << std::hex << (out[0] & 0xff);
+        std::string res(ss.str());
+        out = "byte: \\x" + res;
+    }
+    return out;
+}
+
+// convert a vector of completion_token_output to json
+static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
+{
+    json out = json::array();
+    for (const auto &prob : probs)
+    {
+        json probs_for_token = json::array();
+        for (const auto &p : prob.probs)
+        {
+            std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
+            probs_for_token.push_back(json
+            {
+                {"tok_str", tok_str},
+                {"prob",    p.prob},
+            });
+        }
+        std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
+        out.push_back(json{
+            {"content", tok_str},
+            {"probs",   probs_for_token},
+        });
+    }
+    return out;
+}

llm/generate/gen_common.sh

@@ -14,7 +14,7 @@ init_vars() {
 
     LLAMACPP_DIR=../llama.cpp
     CMAKE_DEFS=""
-    CMAKE_TARGETS="--target ext_server"
+    CMAKE_TARGETS="--target ollama_llama_server"
     if echo "${CGO_CFLAGS}" | grep -- '-g' >/dev/null; then
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE=on -DLLAMA_GPROF=on -DLLAMA_SERVER_VERBOSE=on ${CMAKE_DEFS}"
     else
@@ -39,7 +39,7 @@ init_vars() {
     *)
         ;;
     esac
-    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then 
+    if [ -z "${CMAKE_CUDA_ARCHITECTURES}" ] ; then
         CMAKE_CUDA_ARCHITECTURES="50;52;61;70;75;80"
     fi
 }
@@ -61,8 +61,8 @@ git_module_setup() {
 
 apply_patches() {
     # Wire up our CMakefile
-    if ! grep ollama ${LLAMACPP_DIR}/examples/server/CMakeLists.txt; then
-        echo 'include (../../../ext_server/CMakeLists.txt) # ollama' >>${LLAMACPP_DIR}/examples/server/CMakeLists.txt
+    if ! grep ollama ${LLAMACPP_DIR}/CMakeLists.txt; then
+        echo 'add_subdirectory(../ext_server ext_server) # ollama' >>${LLAMACPP_DIR}/CMakeLists.txt
     fi
 
     if [ -n "$(ls -A ../patches/*.diff)" ]; then
@@ -76,35 +76,29 @@ apply_patches() {
             (cd ${LLAMACPP_DIR} && git apply ${patch})
         done
     fi
-
-    # Avoid duplicate main symbols when we link into the cgo binary
-    sed -e 's/int main(/int __main(/g' <${LLAMACPP_DIR}/examples/server/server.cpp >${LLAMACPP_DIR}/examples/server/server.cpp.tmp &&
-        mv ${LLAMACPP_DIR}/examples/server/server.cpp.tmp ${LLAMACPP_DIR}/examples/server/server.cpp
 }
 
 build() {
     cmake -S ${LLAMACPP_DIR} -B ${BUILD_DIR} ${CMAKE_DEFS}
     cmake --build ${BUILD_DIR} ${CMAKE_TARGETS} -j8
-    mkdir -p ${BUILD_DIR}/lib/
-    g++ -fPIC -g -shared -o ${BUILD_DIR}/lib/libext_server.${LIB_EXT} \
-        ${GCC_ARCH} \
-        ${WHOLE_ARCHIVE} ${BUILD_DIR}/examples/server/libext_server.a ${NO_WHOLE_ARCHIVE} \
-        ${BUILD_DIR}/common/libcommon.a \
-        ${BUILD_DIR}/libllama.a \
-        -Wl,-rpath,\$ORIGIN \
-        -lpthread -ldl -lm \
-        ${EXTRA_LIBS}
 }
 
-compress_libs() {
+compress() {
     echo "Compressing payloads to reduce overall binary size..."
     pids=""
-    rm -rf ${BUILD_DIR}/lib/*.${LIB_EXT}*.gz
-    for lib in ${BUILD_DIR}/lib/*.${LIB_EXT}* ; do
-        gzip -n --best -f ${lib} &
+    rm -rf ${BUILD_DIR}/bin/*.gz
+    for f in ${BUILD_DIR}/bin/* ; do
+        gzip -n --best -f ${f} &
         pids+=" $!"
     done
-    echo 
+    # check for lib directory
+    if [ -d ${BUILD_DIR}/lib ]; then
+        for f in ${BUILD_DIR}/lib/* ; do
+            gzip -n --best -f ${f} &
+            pids+=" $!"
+        done
+    fi
+    echo
     for pid in ${pids}; do
         wait $pid
     done
@@ -113,7 +107,7 @@ compress_libs() {
 
 # Keep the local tree clean after we're done with the build
 cleanup() {
-    (cd ${LLAMACPP_DIR}/examples/server/ && git checkout CMakeLists.txt server.cpp)
+    (cd ${LLAMACPP_DIR}/ && git checkout CMakeLists.txt)
 
     if [ -n "$(ls -A ../patches/*.diff)" ]; then
         for patch in ../patches/*.diff; do

llm/generate/gen_darwin.sh

@@ -18,34 +18,31 @@ sign() {
     fi
 }
 
-# bundle_metal bundles ggml-common.h and ggml-metal.metal into a single file
-bundle_metal() {
-    grep -v '#include "ggml-common.h"' "${LLAMACPP_DIR}/ggml-metal.metal" | grep -v '#pragma once' > "${LLAMACPP_DIR}/ggml-metal.metal.temp"
-    echo '#define GGML_COMMON_IMPL_METAL' > "${LLAMACPP_DIR}/ggml-metal.metal"
-    cat "${LLAMACPP_DIR}/ggml-common.h" | grep -v '#pragma once' >> "${LLAMACPP_DIR}/ggml-metal.metal"
-    cat  "${LLAMACPP_DIR}/ggml-metal.metal.temp" >> "${LLAMACPP_DIR}/ggml-metal.metal"
-    rm "${LLAMACPP_DIR}/ggml-metal.metal.temp"
-}
-
-cleanup_metal() {
-    (cd ${LLAMACPP_DIR} && git checkout ggml-metal.metal)
-}
-
-COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DCMAKE_SYSTEM_NAME=Darwin"
+COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on"
 
 case "${GOARCH}" in
 "amd64")
     COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
 
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}_static"
+    echo "Building static library"
+    build
+
+
     #
     # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
     #
+    init_vars
     CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu"
     echo "Building LCD CPU"
     build
-    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu/lib/libext_server.dylib
-    compress_libs
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
 
     #
     # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
@@ -53,11 +50,11 @@ case "${GOARCH}" in
     #
     init_vars
     CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
     echo "Building AVX CPU"
     build
-    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx/lib/libext_server.dylib
-    compress_libs
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
 
     #
     # ~2013 CPU Dynamic library
@@ -65,22 +62,30 @@ case "${GOARCH}" in
     #
     init_vars
     CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2"
+    BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
     echo "Building AVX2 CPU"
     EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
     build
-    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/cpu_avx2/lib/libext_server.dylib
-    compress_libs
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
     ;;
 "arm64")
-    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/metal"
-    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
-    bundle_metal
+
+    # Static build for linking into the Go binary
+    init_vars
+    CMAKE_TARGETS="--target llama --target ggml"
+    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}_static"
+    echo "Building static library"
     build
-    sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/metal/lib/libext_server.dylib
-    compress_libs
-    cleanup_metal
+
+    init_vars
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+    BUILD_DIR="../build/darwin/${ARCH}/metal"
+    EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
+    build
+    sign ${BUILD_DIR}/bin/ollama_llama_server
+    compress
     ;;
 *)
     echo "GOARCH must be set"
@@ -90,3 +95,4 @@ case "${GOARCH}" in
 esac
 
 cleanup
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_linux.sh

@@ -26,6 +26,9 @@ amdGPUs() {
         "gfx908:xnack-"
         "gfx90a:xnack+"
         "gfx90a:xnack-"
+        "gfx940"
+        "gfx941"
+        "gfx942"
         "gfx1010"
         "gfx1012"
         "gfx1030"
@@ -54,16 +57,31 @@ init_vars
 git_module_setup
 apply_patches
 
+
+init_vars
 if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
+
+    if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "static" ]; then
+        # Static build for linking into the Go binary
+        init_vars
+        CMAKE_TARGETS="--target llama --target ggml"
+        CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        BUILD_DIR="../build/linux/${ARCH}_static"
+        echo "Building static library"
+        build
+    fi
+
+
     # Users building from source can tune the exact flags we pass to cmake for configuring
     # llama.cpp, and we'll build only 1 CPU variant in that case as the default.
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
+        init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
         CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
-        BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu"
+        BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
-        compress_libs
+        compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
         # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
@@ -80,37 +98,43 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
             #
             # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
             #
+            init_vars
             CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu"
+            BUILD_DIR="../build/linux/${ARCH}/cpu"
             echo "Building LCD CPU"
             build
-            compress_libs
+            compress
         fi
 
-        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
+        if [ "${ARCH}" == "x86_64" ]; then
             #
-            # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
-            # Approximately 400% faster than LCD on same CPU
+            # ARM chips in M1/M2/M3-based MACs and NVidia Tegra devices do not currently support avx extensions.
             #
-            init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
-            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu_avx"
-            echo "Building AVX CPU"
-            build
-            compress_libs
-        fi
+            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx" ]; then
+                #
+                # ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
+                # Approximately 400% faster than LCD on same CPU
+                #
+                init_vars
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
+                echo "Building AVX CPU"
+                build
+                compress
+            fi
 
-        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
-            #
-            # ~2013 CPU Dynamic library
-            # Approximately 10% faster than AVX on same CPU
-            #
-            init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
-            BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cpu_avx2"
-            echo "Building AVX2 CPU"
-            build
-            compress_libs
+            if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu_avx2" ]; then
+                #
+                # ~2013 CPU Dynamic library
+                # Approximately 10% faster than AVX on same CPU
+                #
+                init_vars
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
+                echo "Building AVX2 CPU"
+                build
+                compress
+            fi
         fi
     fi
 else
@@ -139,29 +163,38 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
     if [ -n "${CUDA_MAJOR}" ]; then
         CUDA_VARIANT=_v${CUDA_MAJOR}
     fi
-    CMAKE_DEFS="-DLLAMA_CUBLAS=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS}"
-    BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/cuda${CUDA_VARIANT}"
+    if [ "${ARCH}" == "arm64" ]; then
+        echo "ARM CPU detected - disabling unsupported AVX instructions"
+        
+        # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
+        #
+        # CUDA compute < 6.0 lacks proper FP16 support on ARM. 
+        # Disabling has minimal performance effect while maintaining compatibility. 
+        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+    fi
+    CMAKE_DEFS="-DLLAMA_CUDA=on -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS}"
+    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
     EXTRA_LIBS="-L${CUDA_LIB_DIR} -lcudart -lcublas -lcublasLt -lcuda"
     build
 
-    # Cary the CUDA libs as payloads to help reduce dependency burden on users
+    # Carry the CUDA libs as payloads to help reduce dependency burden on users
     #
     # TODO - in the future we may shift to packaging these separately and conditionally
     #        downloading them in the install script.
-    DEPS="$(ldd ${BUILD_DIR}/lib/libext_server.so )"
+    DEPS="$(ldd ${BUILD_DIR}/bin/ollama_llama_server )"
     for lib in libcudart.so libcublas.so libcublasLt.so ; do
         DEP=$(echo "${DEPS}" | grep ${lib} | cut -f1 -d' ' | xargs || true)
         if [ -n "${DEP}" -a -e "${CUDA_LIB_DIR}/${DEP}" ]; then
-            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/lib/"
+            cp "${CUDA_LIB_DIR}/${DEP}" "${BUILD_DIR}/bin/"
         elif [ -e "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" ]; then
-            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/lib/"
+            cp "${CUDA_LIB_DIR}/${lib}.${CUDA_MAJOR}" "${BUILD_DIR}/bin/"
         elif [ -e "${CUDART_LIB_DIR}/${lib}" ]; then
-            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/lib/"
+            cp -d ${CUDART_LIB_DIR}/${lib}* "${BUILD_DIR}/bin/"
         else
-            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/lib/"
+            cp -d "${CUDA_LIB_DIR}/${lib}*" "${BUILD_DIR}/bin/"
         fi
     done
-    compress_libs
+    compress
 
 fi
 
@@ -184,17 +217,24 @@ if [ -d "${ROCM_PATH}" ]; then
     fi
     init_vars
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
-    BUILD_DIR="${LLAMACPP_DIR}/build/linux/${ARCH}/rocm${ROCM_VARIANT}"
-    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
+    BUILD_DIR="../build/linux/${ARCH}/rocm${ROCM_VARIANT}"
+    EXTRA_LIBS="-L${ROCM_PATH}/lib -L/opt/amdgpu/lib/x86_64-linux-gnu/ -Wl,-rpath,\$ORIGIN/../../rocm/ -lhipblas -lrocblas -lamdhip64 -lrocsolver -lamd_comgr -lhsa-runtime64 -lrocsparse -ldrm -ldrm_amdgpu"
     build
 
     # Record the ROCM dependencies
-    rm -f "${BUILD_DIR}/lib/deps.txt"
-    touch "${BUILD_DIR}/lib/deps.txt"
-    for dep in $(ldd "${BUILD_DIR}/lib/libext_server.so" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
-        echo "${dep}" >> "${BUILD_DIR}/lib/deps.txt"
+    rm -f "${BUILD_DIR}/bin/deps.txt"
+    touch "${BUILD_DIR}/bin/deps.txt"
+    for dep in $(ldd "${BUILD_DIR}/bin/ollama_llama_server" | grep "=>" | cut -f2 -d= | cut -f2 -d' ' | grep -e rocm -e amdgpu -e libtinfo ); do
+        echo "${dep}" >> "${BUILD_DIR}/bin/deps.txt"
     done
-    compress_libs
+    # bomb out if for some reason we didn't get a few deps
+    if [ $(cat "${BUILD_DIR}/bin/deps.txt" | wc -l ) -lt 8 ] ; then
+        cat "${BUILD_DIR}/bin/deps.txt"
+        echo "ERROR: deps file short"
+        exit 1
+    fi
+    compress
 fi
 
 cleanup
+echo "go generate completed.  LLM runners: $(cd ${BUILD_DIR}/..; echo *)"

llm/generate/gen_windows.ps1

@@ -13,6 +13,9 @@ function amdGPUs {
         "gfx908:xnack-"
         "gfx90a:xnack+"
         "gfx90a:xnack-"
+        "gfx940"
+        "gfx941"
+        "gfx942"
         "gfx1010"
         "gfx1012"
         "gfx1030"
@@ -24,19 +27,13 @@ function amdGPUs {
 }
 
 function init_vars {
-    # Verify the environment is a Developer Shell for MSVC 2019
-    write-host $env:VSINSTALLDIR
-    if (($env:VSINSTALLDIR -eq $null)) {
-        Write-Error "`r`nBUILD ERROR - YOUR DEVELOPMENT ENVIRONMENT IS NOT SET UP CORRECTLY`r`nTo build Ollama you must run from an MSVC Developer Shell`r`nSee .\docs\development.md for instructions to set up your dev environment"
-        exit 1
-    }
     $script:SRC_DIR = $(resolve-path "..\..\")
     $script:llamacppDir = "../llama.cpp"
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
         "-DLLAMA_NATIVE=off"
         )
-    $script:cmakeTargets = @("ext_server")
+    $script:cmakeTargets = @("ollama_llama_server")
     $script:ARCH = "amd64" # arm not yet supported.
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on", "-DLLAMA_SERVER_VERBOSE=on", "-DCMAKE_BUILD_TYPE=RelWithDebInfo")
@@ -65,8 +62,12 @@ function init_vars {
     } else {
         $script:CMAKE_CUDA_ARCHITECTURES=$env:CMAKE_CUDA_ARCHITECTURES
     }
-    # Note: 10 Windows Kit signtool crashes with GCP's plugin
-    ${script:SignTool}="C:\Program Files (x86)\Windows Kits\8.1\bin\x64\signtool.exe"
+    # Note: Windows Kits 10 signtool crashes with GCP's plugin
+    if ($null -eq $env:SIGN_TOOL) {
+        ${script:SignTool}="C:\Program Files (x86)\Windows Kits\8.1\bin\x64\signtool.exe"
+    } else {
+        ${script:SignTool}=${env:SIGN_TOOL}
+    }
     if ("${env:KEY_CONTAINER}") {
         ${script:OLLAMA_CERT}=$(resolve-path "${script:SRC_DIR}\ollama_inc.crt")
     }
@@ -82,8 +83,8 @@ function git_module_setup {
 
 function apply_patches {
     # Wire up our CMakefile
-    if (!(Select-String -Path "${script:llamacppDir}/examples/server/CMakeLists.txt" -Pattern 'ollama')) {
-        Add-Content -Path "${script:llamacppDir}/examples/server/CMakeLists.txt" -Value 'include (../../../ext_server/CMakeLists.txt) # ollama'
+    if (!(Select-String -Path "${script:llamacppDir}/CMakeLists.txt" -Pattern 'ollama')) {
+        Add-Content -Path "${script:llamacppDir}/CMakeLists.txt" -Value 'add_subdirectory(../ext_server ext_server) # ollama'
     }
 
     # Apply temporary patches until fix is upstream
@@ -96,22 +97,15 @@ function apply_patches {
         }
 
         # Checkout each file
-        Set-Location -Path ${script:llamacppDir}
         foreach ($file in $filePaths) {
-            git checkout $file
+            git -C "${script:llamacppDir}" checkout $file
         }
     }
 
     # Apply each patch
     foreach ($patch in $patches) {
-        Set-Location -Path ${script:llamacppDir}
-        git apply $patch.FullName
+        git -C "${script:llamacppDir}" apply $patch.FullName
     }
-
-    # Avoid duplicate main symbols when we link into the cgo binary
-    $content = Get-Content -Path "${script:llamacppDir}/examples/server/server.cpp"
-    $content = $content -replace 'int main\(', 'int __main('
-    Set-Content -Path "${script:llamacppDir}/examples/server/server.cpp" -Value $content
 }
 
 function build {
@@ -119,26 +113,20 @@ function build {
     & cmake --version
     & cmake -S "${script:llamacppDir}" -B $script:buildDir $script:cmakeDefs
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-    write-host "building with: cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })"
+    write-host "building with: cmake --build $script:buildDir --config $script:config $($script:cmakeTargets | ForEach-Object { `"--target`", $_ })"
     & cmake --build $script:buildDir --config $script:config ($script:cmakeTargets | ForEach-Object { "--target", $_ })
     if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-}
-
-function install {
-    rm -ea 0 -recurse -force -path "${script:buildDir}/lib"
-    md "${script:buildDir}/lib" -ea 0 > $null
-    cp "${script:buildDir}/bin/${script:config}/ext_server.dll" "${script:buildDir}/lib"
-    cp "${script:buildDir}/bin/${script:config}/llama.dll" "${script:buildDir}/lib"
-    # Display the dll dependencies in the build log
-    if ($script:DUMPBIN -ne $null) {
-        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/${script:config}/ext_server.dll" | select-string ".dll"
+    # Rearrange output to be consistent between different generators
+    if ($null -ne ${script:config} -And (test-path -path "${script:buildDir}/bin/${script:config}" ) ) {
+        mv -force "${script:buildDir}/bin/${script:config}/*" "${script:buildDir}/bin/"
+        remove-item "${script:buildDir}/bin/${script:config}"
     }
 }
 
 function sign {
     if ("${env:KEY_CONTAINER}") {
-        write-host "Signing ${script:buildDir}/lib/*.dll"
-        foreach ($file in (get-childitem "${script:buildDir}/lib/*.dll")){
+        write-host "Signing ${script:buildDir}/bin/*.exe  ${script:buildDir}/bin/*.dll"
+        foreach ($file in @(get-childitem "${script:buildDir}/bin/*.exe") + @(get-childitem "${script:buildDir}/bin/*.dll")){
             & "${script:SignTool}" sign /v /fd sha256 /t http://timestamp.digicert.com /f "${script:OLLAMA_CERT}" `
                 /csp "Google Cloud KMS Provider" /kc "${env:KEY_CONTAINER}" $file
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -146,14 +134,20 @@ function sign {
     }
 }
 
-function compress_libs {
+function compress {
     if ($script:GZIP -eq $null) {
         write-host "gzip not installed, not compressing files"
         return
     }
+    write-host "Compressing binaries..."
+    $binaries = dir "${script:buildDir}/bin/*.exe"
+    foreach ($file in $binaries) {
+        & "$script:GZIP" --best -f $file
+    }
+
     write-host "Compressing dlls..."
-    $libs = dir "${script:buildDir}/lib/*.dll"
-    foreach ($file in $libs) {
+    $dlls = dir "${script:buildDir}/bin/*.dll"
+    foreach ($file in $dlls) {
         & "$script:GZIP" --best -f $file
     }
 }
@@ -168,14 +162,11 @@ function cleanup {
         }
 
         # Checkout each file
-        Set-Location -Path ${script:llamacppDir}
         foreach ($file in $filePaths) {            
-            git checkout $file
+            git -C "${script:llamacppDir}" checkout $file
         }
+        git -C "${script:llamacppDir}" checkout CMakeLists.txt
     }
-    Set-Location "${script:llamacppDir}/examples/server"
-    git checkout CMakeLists.txt server.cpp
-
 }
 
 init_vars
@@ -183,38 +174,64 @@ git_module_setup
 apply_patches
 
 # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
 # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
 # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
 
 $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
 
-init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
-$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu"
-write-host "Building LCD CPU"
-build
-install
-sign
-compress_libs
+if ($null -eq ${env:OLLAMA_SKIP_CPU_GENERATE}) {
 
+# GCC build for direct linking into the Go binary
 init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
-$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx"
-write-host "Building AVX CPU"
+# cmake will silently fallback to msvc compilers if mingw isn't in the path, so detect and fail fast
+# as we need this to be compiled by gcc for golang to be able to link with itx
+write-host "Checking for MinGW..."
+# error action ensures we exit on failure
+get-command gcc
+get-command mingw32-make
+$script:cmakeTargets = @("llama", "ggml")
+$script:cmakeDefs = @(
+    "-G", "MinGW Makefiles"
+    "-DCMAKE_C_COMPILER=gcc.exe",
+    "-DCMAKE_CXX_COMPILER=g++.exe",
+    "-DBUILD_SHARED_LIBS=off",
+    "-DLLAMA_NATIVE=off",
+    "-DLLAMA_AVX=off",
+    "-DLLAMA_AVX2=off",
+    "-DLLAMA_AVX512=off",
+    "-DLLAMA_F16C=off",
+    "-DLLAMA_FMA=off")
+$script:buildDir="../build/windows/${script:ARCH}_static"
+write-host "Building static library"
 build
-install
-sign
-compress_libs
 
-init_vars
-$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
-$script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cpu_avx2"
-write-host "Building AVX2 CPU"
-build
-install
-sign
-compress_libs
+# remaining llama.cpp builds use MSVC 
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu"
+    write-host "Building LCD CPU"
+    build
+    sign
+    compress
+
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
+    write-host "Building AVX CPU"
+    build
+    sign
+    compress
+
+    init_vars
+    $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+    $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
+    write-host "Building AVX2 CPU"
+    build
+    sign
+    compress
+} else {
+    write-host "Skipping CPU generation step as requested"
+}
 
 if ($null -ne $script:CUDA_LIB_DIR) {
     # Then build cuda as a dynamically loaded library
@@ -224,13 +241,11 @@ if ($null -ne $script:CUDA_LIB_DIR) {
         $script:CUDA_VARIANT="_"+$script:CUDA_VERSION
     }
     init_vars
-    $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
-    $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUBLAS=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
-    write-host "Building CUDA"
+    $script:buildDir="../build/windows/${script:ARCH}/cuda$script:CUDA_VARIANT"
+    $script:cmakeDefs += @("-A", "x64", "-DLLAMA_CUDA=ON", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR", "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}")
     build
-    install
     sign
-    compress_libs
+    compress
 }
 
 if ($null -ne $env:HIP_PATH) {
@@ -240,12 +255,13 @@ if ($null -ne $env:HIP_PATH) {
     }
 
     init_vars
-    $script:buildDir="${script:llamacppDir}/build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
+    $script:buildDir="../build/windows/${script:ARCH}/rocm$script:ROCM_VARIANT"
     $script:cmakeDefs += @(
         "-G", "Ninja", 
         "-DCMAKE_C_COMPILER=clang.exe",
         "-DCMAKE_CXX_COMPILER=clang++.exe",
         "-DLLAMA_HIPBLAS=on",
+        "-DHIP_PLATFORM=amd",
         "-DLLAMA_AVX=on",
         "-DLLAMA_AVX2=off",
         "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
@@ -254,7 +270,7 @@ if ($null -ne $env:HIP_PATH) {
         )
 
     # Make sure the ROCm binary dir is first in the path
-    $env:PATH="$env:HIP_PATH\bin;$env:VSINSTALLDIR\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja;$env:PATH"
+    $env:PATH="$env:HIP_PATH\bin;$env:PATH"
 
     # We have to clobber the LIB var from the developer shell for clang to work properly
     $env:LIB=""
@@ -263,13 +279,13 @@ if ($null -ne $env:HIP_PATH) {
     build
     # Ninja doesn't prefix with config name
     ${script:config}=""
-    install
     if ($null -ne $script:DUMPBIN) {
-        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/${script:config}/ext_server.dll" | select-string ".dll"
+        & "$script:DUMPBIN" /dependents "${script:buildDir}/bin/ollama_llama_server.exe" | select-string ".dll"
     }
     sign
-    compress_libs
+    compress
 }
 
+
 cleanup
-write-host "`ngo generate completed"
+write-host "`ngo generate completed.  LLM runners: $(get-childitem -path ${script:SRC_DIR}\llm\build\windows\${script:ARCH})"

llm/generate/generate_darwin.go

@@ -1,3 +1,3 @@
 package generate
 
-//go:generate sh ./gen_darwin.sh
+//go:generate bash ./gen_darwin.sh

llm/ggla.go

@@ -7,16 +7,18 @@ import (
 	"slices"
 )
 
-type ContainerGGLA struct {
+type containerGGLA struct {
 	version uint32
 }
 
-func (c *ContainerGGLA) Name() string {
+func (c *containerGGLA) Name() string {
 	return "ggla"
 }
 
-func (c *ContainerGGLA) Decode(rso *readSeekOffset) (model, error) {
-	binary.Read(rso, binary.LittleEndian, &c.version)
+func (c *containerGGLA) Decode(rs io.ReadSeeker) (model, error) {
+	if err := binary.Read(rs, binary.LittleEndian, &c.version); err != nil {
+		return nil, err
+	}
 
 	switch c.version {
 	case 1:
@@ -24,58 +26,66 @@ func (c *ContainerGGLA) Decode(rso *readSeekOffset) (model, error) {
 		return nil, errors.New("invalid version")
 	}
 
-	model := newModelGGLA(c)
-	err := model.decode(rso)
+	model := newGGLA(c)
+	err := model.decode(rs)
 	return model, err
 }
 
-type ModelGGLA struct {
-	*ContainerGGLA
+type ggla struct {
+	*containerGGLA
 
 	kv      KV
-	tensors []Tensor
+	tensors []*Tensor
 }
 
-func newModelGGLA(container *ContainerGGLA) *ModelGGLA {
-	return &ModelGGLA{
-		ContainerGGLA: container,
+func newGGLA(container *containerGGLA) *ggla {
+	return &ggla{
+		containerGGLA: container,
 		kv:            make(KV),
 	}
 }
 
-func (m *ModelGGLA) decode(rso *readSeekOffset) error {
+func (llm *ggla) KV() KV {
+	return llm.kv
+}
+
+func (llm *ggla) Tensors() Tensors {
+	return llm.tensors
+}
+
+func (llm *ggla) decode(rs io.ReadSeeker) error {
 	var r uint32
-	if err := binary.Read(rso, binary.LittleEndian, &r); err != nil {
+	if err := binary.Read(rs, binary.LittleEndian, &r); err != nil {
 		return err
 	}
-	m.kv["r"] = r
+	llm.kv["r"] = r
 
 	var alpha uint32
-	if err := binary.Read(rso, binary.LittleEndian, &alpha); err != nil {
+	if err := binary.Read(rs, binary.LittleEndian, &alpha); err != nil {
 		return err
 	}
-	m.kv["alpha"] = alpha
+	llm.kv["alpha"] = alpha
 
 	for {
 		var dims uint32
-		if err := binary.Read(rso, binary.LittleEndian, &dims); err != nil {
+		if err := binary.Read(rs, binary.LittleEndian, &dims); err != nil {
 			return err
 		}
 
 		var namesize uint32
-		if err := binary.Read(rso, binary.LittleEndian, &namesize); err != nil {
+		if err := binary.Read(rs, binary.LittleEndian, &namesize); err != nil {
 			return err
 		}
 
 		var t Tensor
-		if err := binary.Read(rso, binary.LittleEndian, &t.Kind); err != nil {
+		if err := binary.Read(rs, binary.LittleEndian, &t.Kind); err != nil {
 			return err
 		}
 
 		t.Shape = make([]uint64, dims)
 		for i := 0; uint32(i) < dims; i++ {
 			var shape32 uint32
-			if err := binary.Read(rso, binary.LittleEndian, &shape32); err != nil {
+			if err := binary.Read(rs, binary.LittleEndian, &shape32); err != nil {
 				return err
 			}
 
@@ -87,66 +97,32 @@ func (m *ModelGGLA) decode(rso *readSeekOffset) error {
 		slices.Reverse(t.Shape)
 
 		name := make([]byte, namesize)
-		if err := binary.Read(rso, binary.LittleEndian, &name); err != nil {
+		if err := binary.Read(rs, binary.LittleEndian, &name); err != nil {
 			return err
 		}
 
 		t.Name = string(name)
 
-		if _, err := rso.Seek((rso.offset+31)&-32, io.SeekStart); err != nil {
+		offset, err := rs.Seek(0, io.SeekCurrent)
+		if err != nil {
 			return err
 		}
 
-		t.Offset = uint64(rso.offset)
-
-		if _, err := rso.Seek(int64(t.Size()), io.SeekCurrent); err != nil {
+		if _, err := rs.Seek((offset+31)&-32, io.SeekStart); err != nil {
 			return err
 		}
 
-		m.tensors = append(m.tensors, t)
+		offset, err = rs.Seek(0, io.SeekCurrent)
+		if err != nil {
+			return err
+		}
+
+		t.Offset = uint64(offset)
+
+		if _, err := rs.Seek(int64(t.size()), io.SeekCurrent); err != nil {
+			return err
+		}
+
+		llm.tensors = append(llm.tensors, &t)
 	}
 }
-
-func (m *ModelGGLA) KV() KV {
-	return m.kv
-}
-
-func (m *ModelGGLA) Tensor() []Tensor {
-	return m.tensors
-}
-
-func (*ModelGGLA) ModelFamily() string {
-	return "ggla"
-}
-
-func (*ModelGGLA) ModelType() string {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) FileType() string {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumLayers() uint32 {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumGQA() uint32 {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumEmbed() uint32 {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumHead() uint32 {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumHeadKv() uint32 {
-	panic("not implemented")
-}
-
-func (*ModelGGLA) NumCtx() uint32 {
-	panic("not implemented")
-}

llm/ggml.go

@@ -3,14 +3,14 @@ package llm
 import (
 	"encoding/binary"
 	"errors"
+	"fmt"
 	"io"
+	"strings"
 )
 
 type GGML struct {
 	container
 	model
-
-	Size int64
 }
 
 const (
@@ -90,20 +90,183 @@ func fileType(fileType uint32) string {
 }
 
 type model interface {
-	ModelFamily() string
-	ModelType() string
-	FileType() string
-	NumLayers() uint32
-	NumGQA() uint32
-	NumEmbed() uint32
-	NumHead() uint32
-	NumHeadKv() uint32
-	NumCtx() uint32
+	KV() KV
+	Tensors() Tensors
+}
+
+type KV map[string]any
+
+func (kv KV) u64(key string) uint64 {
+	switch v := kv[key].(type) {
+	case uint64:
+		return v
+	case uint32:
+		return uint64(v)
+	case float64:
+		return uint64(v)
+	default:
+		return 0
+	}
+}
+
+func (kv KV) Architecture() string {
+	if s, ok := kv["general.architecture"].(string); ok {
+		return s
+	}
+
+	return "unknown"
+}
+
+func (kv KV) ParameterCount() uint64 {
+	return kv.u64("general.parameter_count")
+}
+
+func (kv KV) FileType() string {
+	if u64 := kv.u64("general.file_type"); u64 > 0 {
+		return fileType(uint32(u64))
+	}
+
+	return "unknown"
+}
+
+func (kv KV) BlockCount() uint64 {
+	return kv.u64(fmt.Sprintf("%s.block_count", kv.Architecture()))
+}
+
+func (kv KV) HeadCount() uint64 {
+	return kv.u64(fmt.Sprintf("%s.attention.head_count", kv.Architecture()))
+}
+
+func (kv KV) HeadCountKV() uint64 {
+	if headCountKV := kv.u64(fmt.Sprintf("%s.attention.head_count_kv", kv.Architecture())); headCountKV > 0 {
+		return headCountKV
+	}
+
+	return 1
+}
+
+func (kv KV) GQA() uint64 {
+	return kv.HeadCount() / kv.HeadCountKV()
+}
+
+func (kv KV) EmbeddingLength() uint64 {
+	return kv.u64(fmt.Sprintf("%s.embedding_length", kv.Architecture()))
+}
+
+func (kv KV) ContextLength() uint64 {
+	return kv.u64(fmt.Sprintf("%s.context_length", kv.Architecture()))
+}
+
+type Tensors []*Tensor
+
+func (ts Tensors) Layers() map[string]Layer {
+	layers := make(map[string]Layer)
+	for _, t := range ts {
+		parts := strings.Split(t.Name, ".")
+		if parts[0] == "blk" {
+			parts = parts[1:]
+		}
+
+		if _, ok := layers[parts[0]]; !ok {
+			layers[parts[0]] = make(Layer)
+		}
+
+		layers[parts[0]][strings.Join(parts[1:], ".")] = t
+	}
+
+	return layers
+}
+
+type Layer map[string]*Tensor
+
+func (l Layer) size() (size uint64) {
+	for _, t := range l {
+		size += t.size()
+	}
+
+	return size
+}
+
+type Tensor struct {
+	Name   string `json:"name"`
+	Kind   uint32 `json:"kind"`
+	Offset uint64 `json:"-"`
+
+	// Shape is the number of elements in each dimension
+	Shape []uint64 `json:"shape"`
+
+	io.WriterTo `json:"-"`
+}
+
+func (t Tensor) blockSize() uint64 {
+	switch {
+	case t.Kind < 2:
+		return 1
+	case t.Kind < 10:
+		return 32
+	default:
+		return 256
+	}
+}
+
+func (t Tensor) typeSize() uint64 {
+	blockSize := t.blockSize()
+
+	switch t.Kind {
+	case 0: // FP32
+		return 4
+	case 1: // FP16
+		return 2
+	case 2: // Q4_0
+		return 2 + blockSize/2
+	case 3: // Q4_1
+		return 2 + 2 + blockSize/2
+	case 6: // Q5_0
+		return 2 + 4 + blockSize/2
+	case 7: // Q5_1
+		return 2 + 2 + 4 + blockSize/2
+	case 8: // Q8_0
+		return 2 + blockSize
+	case 9: // Q8_1
+		return 4 + 4 + blockSize
+	case 10: // Q2_K
+		return blockSize/16 + blockSize/4 + 2 + 2
+	case 11: // Q3_K
+		return blockSize/8 + blockSize/4 + 12 + 2
+	case 12: // Q4_K
+		return 2 + 2 + 12 + blockSize/2
+	case 13: // Q5_K
+		return 2 + 2 + 12 + blockSize/8 + blockSize/2
+	case 14: // Q6_K
+		return blockSize/2 + blockSize/4 + blockSize/16 + 2
+	case 15: // Q8_K
+		return 2 + blockSize + 2*blockSize/16
+	case 16: // IQ2_XXS
+		return 2 + 2*blockSize/8
+	case 17: // IQ2_XS
+		return 2 + 2*blockSize/8 + blockSize/32
+	case 18: // IQ3_XXS
+		return 2 + 3*blockSize/8
+	default:
+		return 0
+	}
+}
+
+func (t Tensor) parameters() uint64 {
+	var count uint64 = 1
+	for _, n := range t.Shape {
+		count *= n
+	}
+	return count
+}
+
+func (t Tensor) size() uint64 {
+	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
 type container interface {
 	Name() string
-	Decode(*readSeekOffset) (model, error)
+	Decode(io.ReadSeeker) (model, error)
 }
 
 const (
@@ -122,60 +285,102 @@ const (
 
 var ErrUnsupportedFormat = errors.New("unsupported model format")
 
-func DecodeGGML(r io.ReadSeeker) (*GGML, error) {
-	ro := readSeekOffset{ReadSeeker: r}
-
+func DecodeGGML(rs io.ReadSeeker) (*GGML, int64, error) {
 	var magic uint32
-	if err := binary.Read(&ro, binary.LittleEndian, &magic); err != nil {
-		return nil, err
+	if err := binary.Read(rs, binary.LittleEndian, &magic); err != nil {
+		return nil, 0, err
 	}
 
 	var c container
 	switch magic {
 	case FILE_MAGIC_GGML, FILE_MAGIC_GGMF, FILE_MAGIC_GGJT:
-		return nil, ErrUnsupportedFormat
+		return nil, 0, ErrUnsupportedFormat
 	case FILE_MAGIC_GGLA:
-		c = &ContainerGGLA{}
+		c = &containerGGLA{}
 	case FILE_MAGIC_GGUF_LE:
-		c = &ContainerGGUF{ByteOrder: binary.LittleEndian}
+		c = &containerGGUF{ByteOrder: binary.LittleEndian}
 	case FILE_MAGIC_GGUF_BE:
-		c = &ContainerGGUF{ByteOrder: binary.BigEndian}
+		c = &containerGGUF{ByteOrder: binary.BigEndian}
 	default:
-		return nil, errors.New("invalid file magic")
+		return nil, 0, errors.New("invalid file magic")
 	}
 
-	model, err := c.Decode(&ro)
+	model, err := c.Decode(rs)
 	if errors.Is(err, io.EOF) {
 		// noop
 	} else if err != nil {
-		return nil, err
+		return nil, 0, err
+	}
+
+	offset, err := rs.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return nil, 0, err
 	}
 
 	// final model type
 	return &GGML{
 		container: c,
 		model:     model,
-		Size:      ro.offset,
-	}, nil
+	}, offset, nil
 }
 
-type readSeekOffset struct {
-	io.ReadSeeker
-	offset int64
-}
+func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload uint64) {
+	embedding := llm.KV().EmbeddingLength()
+	heads := llm.KV().HeadCount()
+	headsKV := llm.KV().HeadCountKV()
+	vocab := uint64(len(llm.KV()["tokenizer.ggml.tokens"].([]any)))
 
-func (rso *readSeekOffset) Seek(offset int64, whence int) (int64, error) {
-	offset, err := rso.ReadSeeker.Seek(offset, whence)
-	if err != nil {
-		return 0, err
+	layers := llm.Tensors().Layers()
+
+	switch llm.KV().Architecture() {
+	case "llama":
+		fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
+
+		partialOffload = 4 * batch * embedding
+		partialOffload += max(
+			4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embedding/heads*headsKV),
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+		)
+
+		if ffnGateWeight, ok := layers["0"]["ffn_gate.0.weight"]; ok {
+			ffnGateWeight1 := ffnGateWeight.Shape[1]
+			fullOffload = 4 * batch * (2 + 3*embedding + context*(1+heads) + 2*headsKV + ffnGateWeight1)
+			partialOffload = max(
+				4*batch*(3+embedding/heads*headsKV+embedding+context*(1+heads)+ffnGateWeight1)+(embedding*embedding+3*embedding*headsKV*ffnGateWeight1)*9/16,
+				4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16),
+			)
+		}
+	case "gemma":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+	case "command-r":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(2+4*embedding+context*(1+heads)),
+		)
+
+		partialOffload = max(
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+			4*batch*(1+2*embedding+context*(1+heads))+4*embedding*context+embedding*embedding*9/16,
+		)
+	case "qwen2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(1+2*embedding+context+context*heads),
+		)
+
+		partialOffload = max(
+			4*batch*(embedding+vocab)+embedding*vocab*105/128,
+			4*(batch*(1+2*embedding+context*(1+heads))+embedding*(1+context)),
+		)
+	case "phi2":
+		fullOffload = max(
+			4*batch*(embedding+vocab),
+			4*batch*(1+4*embedding+context+context*heads),
+		)
+
+		partialOffload = 4*batch*(2*embedding+vocab) + embedding*vocab*105/128
 	}
 
-	rso.offset = offset
-	return offset, nil
-}
-
-func (rso *readSeekOffset) Read(p []byte) (int, error) {
-	n, err := rso.ReadSeeker.Read(p)
-	rso.offset += int64(n)
-	return n, err
+	return
 }