Remove llama.cpp submodule and shift new build to top

---------

Co-authored-by: Bill Wang <bill.wang@bill.wang>

.dockerignore

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

.gitattributes

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

.github/workflows/release.yaml

@@ -48,8 +48,8 @@ jobs:
         with:
           name: dist-darwin
           path: |
-            dist/*arwin*
-            !dist/*-cov
+            dist/Ollama-darwin.zip
+            dist/ollama-darwin
 
   # Windows builds take a long time to both install the dependencies and build, so parallelize
   # CPU generation step
@@ -92,19 +92,19 @@ jobs:
       - run: go get ./...
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
-          go generate -x ./...
-        name: go generate
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
+        name: make
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cpu
           path: |
             build/**/*
             build/**/*.a
-            llm/build/**/*.a
             dist/windows-amd64/**
 
   # ROCm generation step
@@ -158,31 +158,21 @@ jobs:
       - 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
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$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\"
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
+        name: make
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-rocm
           path: |
             build/**/*
             dist/windows-amd64/**
-      - uses: actions/upload-artifact@v4
-        with:
-          name: windows-rocm-deps
-          path: dist/deps/*
 
   # CUDA generation step
   generate-windows-cuda:
@@ -245,36 +235,175 @@ jobs:
       - name: 'Verify CUDA'
         run: nvcc -V
       - run: go get ./...
-      - name: go generate
+      - name: make
         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
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$cudabin;$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\"
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
       - uses: actions/upload-artifact@v4
         with:
           name: generate-windows-cuda-${{ matrix.cuda.version }}
           path: |
             build/**/*
             dist/windows-amd64/**
+
+  # windows arm64 generate, go build, and zip file (no installer)
+  # Output of this build is aggregated into the final x86 build
+  # for a unified windows installer
+  windows-arm64:
+    runs-on: windows-arm64
+    environment: release
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      # The current Windows arm64 beta image has effectively zero dev tools installed...
+      - name: Install git and gzip
+        run: |
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
+          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
+          choco install -y --no-progress git gzip
+          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      # pacman is buggy on win arm64, so we avoid using it, but rely on the binary artifacts
+      # we download the sfx (7zip bundle) which isn't fully set up, but the binaries we need to build work
+      - name: Install msys2 x64
+        run: |
+          $url="https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-base-x86_64-20240727.sfx.exe"
+          write-host "Downloading MSYS2"
+          Invoke-WebRequest -Uri "$url" -outfile "${env:RUNNER_TEMP}\msys2.exe"
+          write-host "Installing msys2"
+          Start-Process "${env:RUNNER_TEMP}\msys2.exe" -ArgumentList @(
+              '-y', '-oC:\'
+              ) -NoNewWindow -Wait
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      # since pacman isn't reliable, we just download the tar file and extract directly
+      - name: Downloading and extracting msys2 make tar file
+        run: |
+          $url="https://mirror.msys2.org/msys/x86_64/make-4.4.1-2-x86_64.pkg.tar.zst"
+          write-host "Downloading make"
+          Invoke-WebRequest -Uri "$url" -outfile c:\msys64\make.tar.zst
+          cd c:\msys64; tar -xf make.tar.zst
+          rm c:\msys64\make.tar.zst
+      - name: Verify Make works properly
+        run: |
+          echo $env:PATH
+          make --version
+      - name: Install Visual Studio 2022
+        run: |
+          $components = @(
+            "Microsoft.VisualStudio.Component.CoreEditor",
+            "Microsoft.VisualStudio.Workload.CoreEditor",
+            "Microsoft.VisualStudio.Component.Roslyn.Compiler",
+            "Microsoft.Component.MSBuild",
+            "Microsoft.VisualStudio.Component.TextTemplating",
+            "Microsoft.VisualStudio.Component.Debugger.JustInTime",
+            "Microsoft.VisualStudio.Component.VC.CoreIde",
+            "Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
+            "Microsoft.VisualStudio.Component.Windows11SDK.22621",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64EC",
+            "Microsoft.VisualStudio.Component.VC.Tools.ARM64",
+            "Microsoft.VisualStudio.Component.VC.ATL",
+            "Microsoft.VisualStudio.Component.VC.ATL.ARM64",
+            "Microsoft.VisualStudio.Component.Graphics",
+            "Microsoft.VisualStudio.Component.VC.Redist.14.Latest",
+            "Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Core",
+            "Microsoft.VisualStudio.Component.Windows11Sdk.WindowsPerformanceToolkit",
+            "Microsoft.VisualStudio.Component.CppBuildInsights",
+            "Microsoft.VisualStudio.Component.VC.DiagnosticTools",
+            "Microsoft.VisualStudio.ComponentGroup.WebToolsExtensions.CMake",
+            "Microsoft.VisualStudio.Component.VC.CMake.Project",
+            "Microsoft.VisualStudio.Component.VC.ASAN",
+            "Microsoft.VisualStudio.Component.Vcpkg",
+            "Microsoft.VisualStudio.Workload.NativeDesktop"
+          )
+          $config = @{
+                "version" = "1.0"
+                "components"  = $components
+                "extensions"  = @()
+            }
+          $configPath = "${env:RUNNER_TEMP}\vsconfig"
+          $config | ConvertTo-Json | Out-File -FilePath $configPath
+          $bootstrapperFilePath = "${env:RUNNER_TEMP}\vs_community.exe"
+          write-host "Downloading Visual Studio 2022"
+          Invoke-WebRequest -Uri "https://aka.ms/vs/17/release/vs_community.exe" -outfile $bootstrapperFilePath
+          $bootstrapperArgumentList = ('/c', $bootstrapperFilePath, '--config', $configPath, '--quiet', '--wait' )
+          write-host "Installing Visual Studio 2022"
+          $process = Start-Process -FilePath cmd.exe -ArgumentList $bootstrapperArgumentList -Wait -PassThru
+          $exitCode = $process.ExitCode
+          write-host $exitCode
+      # pacman in mingw/msys2 is ~broken on windows arm right now - hangs consistently during attempts to install
+      # so we'll use this alternative GCC binary
+      - name: Install llvm-mingw GCC
+        run: |
+          $gcc_url="https://github.com/mstorsjo/llvm-mingw/releases/download/20240619/llvm-mingw-20240619-ucrt-aarch64.zip"
+          write-host "Downloading llvm-mingw"
+          Invoke-WebRequest -Uri "${gcc_url}" -OutFile "${env:RUNNER_TEMP}\gcc.zip"
+          write-host "Unpacking llvm-mingw"
+          expand-archive -path "${env:RUNNER_TEMP}\gcc.zip" -destinationpath "c:\"
+          mv c:\llvm-mingw-* c:\llvm-mingw
+          echo "c:\llvm-mingw\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Verify GCC
+        run: |
+          echo $env:PATH
+          gcc --version
+      - uses: actions/checkout@v4
+      - name: Set Version
+        run: |
+          $ver=${env:GITHUB_REF_NAME}.trim("v")
+          echo VERSION=$ver | Out-File -FilePath ${env:GITHUB_ENV} -Encoding utf8 -Append
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" | Out-File -FilePath ollama_inc.crt -Encoding utf8
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
+          import-module 'C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\Program Files\Microsoft Visual Studio\2022\Community' -skipautomaticlocation
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
+          $env:ARCH="arm64"
+          .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
+        name: 'Windows Build'
       - uses: actions/upload-artifact@v4
         with:
-          name: windows-cuda-deps-${{ matrix.cuda.version }}
-          path: dist/deps/*
+          name: windows-arm64
+          path: |
+            dist/windows-arm64/**
+            dist/windows-arm64-app.exe
+            dist/ollama-windows-arm64.zip
 
-
-  # Import the prior generation steps and build the final windows assets
+  # Import the prior generation steps plus the full arm64 build, and build the final windows assets
   build-windows:
     environment: release
     runs-on: windows
@@ -282,6 +411,7 @@ jobs:
       - generate-windows-cuda
       - generate-windows-rocm
       - generate-windows-cpu
+      - windows-arm64
     env:
       KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
     steps:
@@ -327,26 +457,22 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-11
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-rocm-deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-arm64
+          path: dist
       - run: dir build
       - run: |
           $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          cd $env:GITHUB_WORKSPACE
+          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
+          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_GENERATE="1"
+          $env:ARCH="amd64"
           & .\scripts\build_windows.ps1
       - uses: actions/upload-artifact@v4
         with:

.github/workflows/test.yaml

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

.gitignore

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

.gitmodules

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

Dockerfile

@@ -1,3 +1,4 @@
+# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
 ARG GOLANG_VERSION=1.22.5
 ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION_11=11.3.1
@@ -6,176 +7,134 @@ ARG CUDA_VERSION_12=12.4.0
 ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
 
-# Copy the minimal context we need to run the generate scripts
-FROM scratch AS llm-code
-COPY .git .git
-COPY .gitmodules .gitmodules
-COPY llm llm
-
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_11-devel-centos7 AS cuda-11-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
-
-FROM --platform=linux/amd64 nvidia/cuda:$CUDA_VERSION_12-devel-centos7 AS cuda-12-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
-
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-runner-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V11_ARCHITECTURES
-ENV GOARCH=arm64
-RUN OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V11_ARCHITECTURES}" \
-    CUDA_VARIANT="_v11" \
-    bash gen_linux.sh
-
-FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-runner-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG CUDA_V12_ARCHITECTURES
-ENV GOARCH=arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 \
-    OLLAMA_SKIP_CPU_GENERATE=1 \
-    CMAKE_CUDA_ARCHITECTURES="${CUDA_V12_ARCHITECTURES}" \
-    CUDA_VARIANT="_v12" \
-    OLLAMA_CUSTOM_CUDA_DEFS="-DGGML_CUDA_USE_GRAPHS=on" \
-    bash gen_linux.sh
-
-
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-ENV LIBRARY_PATH=/opt/amdgpu/lib64
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-ARG CGO_CFLAGS
-ARG AMDGPU_TARGETS
-ENV GOARCH=amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 bash gen_linux.sh
-RUN mkdir -p ../../dist/linux-amd64-rocm/lib/ollama && \
-    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd ../../dist/linux-amd64-rocm/lib/ollama && tar xf - )
-
-FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
+### To create a local image for building linux binaries on mac or windows with efficient incremental builds
+#
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
+# docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
+#
+### Then incremental builds will be much faster in this container
+#
+# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
+#
+FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:/usr/local/cuda/bin:$PATH
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH=amd64
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+# TODO intel oneapi goes here...
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
-
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
+### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
+# Note: this does not contain jetson variants
+#
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
+# docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
+#
+FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+ARG CUDA_VERSION_11
+ARG CUDA_VERSION_12
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
-ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH=arm64
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(echo ${CUDA_VERSION_11} | cut -f1-2 -d. | sed -e "s/\./-/g") \
+    cuda-toolkit-$(echo ${CUDA_VERSION_12} | cut -f1-2 -d. | sed -e "s/\./-/g")
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH:/usr/local/cuda/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
+ENV LIBRARY_PATH=/usr/local/cuda/lib64/stubs:/opt/amdgpu/lib64
+ENV GOARCH amd64
+ENV CGO_ENABLED 1
+WORKDIR /go/src/github.com/ollama/ollama/
+ENTRYPOINT [ "zsh" ]
 
-FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
+FROM --platform=linux/amd64 unified-builder-amd64 AS runners-amd64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG OLLAMA_SKIP_ROCM_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -C llama -j $(expr $(nproc) / 2 ) ; \
+    else \
+        make -C llama -j 5 ; \
+    fi
+
+FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
+    make -C llama -j 8
 
 
 # Intermediate stages used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
-ENV CGO_ENABLED=1
+FROM --platform=linux/amd64 centos:7 AS builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH amd64
 WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/amd64 builder-amd64 AS build-amd64
 COPY . .
-COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
+ARG OLLAMA_SKIP_ROCM_GENERATE
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 RUN cd dist/linux-$GOARCH && \
     tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
-RUN cd dist/linux-$GOARCH-rocm && \
-    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi
 
-FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
-ENV CGO_ENABLED=1
+FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
+ARG CMAKE_VERSION
 ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH arm64
 WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/arm64 builder-arm64 AS build-arm64
 COPY . .
-COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/ llm/build/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
@@ -187,11 +146,11 @@ FROM --platform=linux/amd64 scratch AS dist-amd64
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
 FROM --platform=linux/arm64 scratch AS dist-arm64
 COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
-FROM dist-$TARGETARCH as dist
+FROM dist-$TARGETARCH AS dist
 
 
 # Optimized container images do not cary nested payloads
-FROM --platform=linux/amd64 static-build-amd64 AS container-build-amd64
+FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -199,7 +158,7 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-amd64/bin/ollama .
 
-FROM --platform=linux/arm64 static-build-arm64 AS container-build-arm64
+FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
 ARG GOFLAGS
@@ -207,48 +166,52 @@ ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
     go build -trimpath -o dist/linux-arm64/bin/ollama .
 
+# For amd64 container images, filter out cuda/rocm to minimize size
+FROM runners-amd64 AS runners-cuda-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_hipblas.so \
+    ./dist/linux-amd64/lib/ollama/runners/rocm*
+
+FROM runners-amd64 AS runners-rocm-amd64
+RUN rm -rf \
+    ./dist/linux-amd64/lib/ollama/libggml_cuda*.so \
+    ./dist/linux-amd64/lib/ollama/libcu*.so* \
+    ./dist/linux-amd64/lib/ollama/runners/cuda*
+
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
 
 FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
-COPY --from=cpu-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-COPY --from=cuda-11-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-COPY --from=cuda-12-build-runner-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
 
 # ROCm libraries larger so we keep it distinct from the CPU/CUDA image
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
 # Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
 # across releases
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
 RUN apt-get update && \
     apt-get install -y ca-certificates && \
-    apt-get clean && rm -rf /var/lib/apt/lists/*
+    rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-COPY --from=cpu-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
 EXPOSE 11434
-ENV OLLAMA_HOST=0.0.0.0
+ENV OLLAMA_HOST 0.0.0.0
 
 ENTRYPOINT ["/bin/ollama"]
 CMD ["serve"]
 
 FROM runtime-$TARGETARCH
 EXPOSE 11434
-ENV OLLAMA_HOST=0.0.0.0
+ENV OLLAMA_HOST 0.0.0.0
 ENV PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
 ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility

Makefile

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

README.md

@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla
 
 ## Quickstart
 
-To run and chat with [Llama 3.1](https://ollama.com/library/llama3.1):
+To run and chat with [Llama 3.2](https://ollama.com/library/llama3.2):
 
 ```
-ollama run llama3.1
+ollama run llama3.2
 ```
 
 ## Model library
@@ -49,6 +49,8 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
+| Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`          |
+| Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`       |
 | Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
 | Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
@@ -99,16 +101,16 @@ See the [guide](docs/import.md) on importing models for more information.
 
 ### Customize a prompt
 
-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.1` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3.2` model:
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 Create a `Modelfile`:
 
 ```
-FROM llama3.1
+FROM llama3.2
 
 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@@ -143,7 +145,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 > This command can also be used to update a local model. Only the diff will be pulled.
@@ -151,13 +153,13 @@ ollama pull llama3.1
 ### Remove a model
 
 ```
-ollama rm llama3.1
+ollama rm llama3.2
 ```
 
 ### Copy a model
 
 ```
-ollama cp llama3.1 my-model
+ollama cp llama3.2 my-model
 ```
 
 ### Multiline input
@@ -181,14 +183,14 @@ The image features a yellow smiley face, which is likely the central focus of th
 ### Pass the prompt as an argument
 
 ```
-$ ollama run llama3.1 "Summarize this file: $(cat README.md)"
+$ ollama run llama3.2 "Summarize this file: $(cat README.md)"
  Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```
 
 ### Show model information
 
 ```
-ollama show llama3.1
+ollama show llama3.2
 ```
 
 ### List models on your computer
@@ -197,6 +199,18 @@ ollama show llama3.1
 ollama list
 ```
 
+### List which models are currently loaded
+
+```
+ollama ps
+```
+
+### Stop a model which is currently running
+
+```
+ollama stop llama3.2
+```
+
 ### Start Ollama
 
 `ollama serve` is used when you want to start ollama without running the desktop application.
@@ -216,7 +230,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:
 
 ```
-./ollama run llama3.1
+./ollama run llama3.2
 ```
 
 ## REST API
@@ -227,7 +241,7 @@ Ollama has a REST API for running and managing models.
 
 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt":"Why is the sky blue?"
 }'
 ```
@@ -236,7 +250,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     { "role": "user", "content": "why is the sky blue?" }
   ]
@@ -313,6 +327,10 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
 - [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)
+- [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
+- [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG on Mac/Windows/Linux)
+- [G1](https://github.com/bklieger-groq/g1) (Prototype of using prompting strategies to improve the LLM's reasoning through o1-like reasoning chains.)
+- [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 
 ### Terminal
 
@@ -359,13 +377,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Libraries
 
-- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/integrations/chat/ollama/) with [example](https://js.langchain.com/docs/tutorials/local_rag/)
 - [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
 - [crewAI](https://github.com/crewAIInc/crewAI)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
-- [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
+- [LlamaIndex](https://docs.llamaindex.ai/en/stable/examples/llm/ollama/) and [LlamaIndexTS](https://ts.llamaindex.ai/modules/llms/available_llms/ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
@@ -394,11 +412,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
 - [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
 - [Agents-Flex for Java](https://github.com/agents-flex/agents-flex) with [example](https://github.com/agents-flex/agents-flex/tree/main/agents-flex-llm/agents-flex-llm-ollama/src/test/java/com/agentsflex/llm/ollama)
+- [Ollama for Swift](https://github.com/mattt/ollama-swift)
 
 ### Mobile
 
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
+- [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 
 ### Extensions & Plugins

app/ollama.iss

@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64 arm64
-ArchitecturesInstallIn64BitMode=x64 arm64
+ArchitecturesAllowed=x64compatible arm64
+ArchitecturesInstallIn64BitMode=x64compatible arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@@ -48,6 +48,7 @@ OutputDir=..\dist\
 SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
+RestartIfNeededByRun=no
 
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
@@ -86,12 +87,21 @@ Name: "english"; MessagesFile: "compiler:Default.isl"
 DialogFontSize=12
 
 [Files]
-Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
-Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windows-{#ARCH}\lib\ollama\runners\*"; DestDir: "{app}\lib\ollama\runners"; Flags: ignoreversion 64bit recursesubdirs
+#if DirExists("..\dist\windows-amd64")
+Source: "..\dist\windows-amd64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: not IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\ollama.exe"; DestDir: "{app}"; Check: not IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: not IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
+
+#if DirExists("..\dist\windows-arm64")
+Source: "..\dist\windows-arm64\vc_redist.arm64.exe"; DestDir: "{tmp}"; Check: IsArm64() and vc_redist_needed(); Flags: deleteafterinstall
+Source: "..\dist\windows-arm64-app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ;Check: IsArm64();  Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\ollama.exe"; DestDir: "{app}"; Check: IsArm64(); Flags: ignoreversion 64bit
+Source: "..\dist\windows-arm64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Check: IsArm64(); Flags: ignoreversion 64bit recursesubdirs
+#endif
+
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-Source: "..\dist\windows-amd64\lib\ollama\*"; DestDir: "{app}\lib\ollama\"; Flags: ignoreversion recursesubdirs
 
 [Icons]
 Name: "{group}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
@@ -99,6 +109,9 @@ Name: "{userstartup}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilen
 Name: "{userprograms}\{#MyAppName}"; Filename: "{app}\{#MyAppExeName}"; IconFilename: "{app}\app.ico"
 
 [Run]
+#if DirExists("..\dist\windows-arm64")
+Filename: "{tmp}\vc_redist.arm64.exe"; Parameters: "/install /passive /norestart"; Check: IsArm64() and vc_redist_needed(); StatusMsg: "Installing VC++ Redistributables..."; Flags: waituntilterminated
+#endif
 Filename: "{cmd}"; Parameters: "/C set PATH={app};%PATH% & ""{app}\{#MyAppExeName}"""; Flags: postinstall nowait runhidden
 
 [UninstallRun]
@@ -129,7 +142,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi
 
 
 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.1
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3.2
 ;ClickFinish=%n
 
 [Registry]
@@ -154,3 +167,39 @@ begin
   { Pos() returns 0 if not found }
   Result := Pos(';' + ExpandConstant(Param) + ';', ';' + OrigPath + ';') = 0;
 end;
+
+{ --- VC Runtime libraries discovery code - Only install vc_redist if it isn't already installed ----- }
+const VCRTL_MIN_V1 = 14;
+const VCRTL_MIN_V2 = 40;
+const VCRTL_MIN_V3 = 33807;
+const VCRTL_MIN_V4 = 0;
+
+ // check if the minimum required vc redist is installed (by looking the registry)
+function vc_redist_needed (): Boolean;
+var
+  sRegKey: string;
+  v1: Cardinal;
+  v2: Cardinal;
+  v3: Cardinal;
+  v4: Cardinal;
+begin
+  sRegKey := 'SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\arm64';
+  if (RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Major', v1)  and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Minor', v2) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'Bld', v3) and
+      RegQueryDWordValue (HKEY_LOCAL_MACHINE, sRegKey, 'RBld', v4)) then
+  begin
+    Log ('VC Redist version: ' + IntToStr (v1) +
+        '.' + IntToStr (v2) + '.' + IntToStr (v3) +
+        '.' + IntToStr (v4));
+    { Version info was found. Return true if later or equal to our
+       minimal required version RTL_MIN_Vx }
+    Result := not (
+        (v1 > VCRTL_MIN_V1) or ((v1 = VCRTL_MIN_V1) and
+         ((v2 > VCRTL_MIN_V2) or ((v2 = VCRTL_MIN_V2) and
+          ((v3 > VCRTL_MIN_V3) or ((v3 = VCRTL_MIN_V3) and
+           (v4 >= VCRTL_MIN_V4)))))));
+  end
+  else
+    Result := TRUE;
+end;

app/ollama_welcome.ps1

@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3.1"
+write-host "`tollama run llama3.2"
 write-host ""

cmd/cmd.go

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

cmd/cmd_test.go

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

cmd/interactive.go

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

convert/convert.go

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

convert/convert_bert.go

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

convert/convert_gemma.go

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

convert/convert_gemma2.go

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

convert/convert_gemma2_adapter.go

@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
+func (p *gemma2Adapter) KV(baseKV fileutils.KV) fileutils.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv
 }
 
-func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),

convert/convert_llama.go

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

convert/convert_llama_adapter.go

@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )
 
 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
+func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
 	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
 	return kv
 }
 
-func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
-	var out []llm.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
+	var out []fileutils.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
 			t.SetRepacker(p.repack)
 		}
 
-		out = append(out, llm.Tensor{
+		out = append(out, fileutils.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,

convert/convert_mixtral.go

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

convert/convert_phi3.go

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

convert/convert_test.go

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

discover/README.md

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

discover/amd_common.go

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

discover/amd_hip_windows.go

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

discover/amd_linux.go

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

discover/amd_windows.go

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

discover/cpu_common.go

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

discover/cuda_common.go

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

discover/gpu.go

@@ -1,6 +1,6 @@
 //go:build linux || windows
 
-package gpu
+package discover
 
 /*
 #cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
@@ -54,6 +54,13 @@ var (
 	nvmlLibPath   string
 	rocmGPUs      []RocmGPUInfo
 	oneapiGPUs    []OneapiGPUInfo
+
+	// If any discovered GPUs are incompatible, report why
+	unsupportedGPUs []UnsupportedGPUInfo
+
+	// Keep track of errors during bootstrapping so that if GPUs are missing
+	// they expected to be present this may explain why
+	bootstrapErrors []error
 )
 
 // With our current CUDA compile flags, older than 5.0 will not work properly
@@ -70,16 +77,17 @@ func initCudaHandles() *cudaHandles {
 
 	cHandles := &cudaHandles{}
 	// Short Circuit if we already know which library to use
+	// ignore bootstrap errors in this case since we already recorded them
 	if nvmlLibPath != "" {
-		cHandles.nvml, _ = LoadNVMLMgmt([]string{nvmlLibPath})
+		cHandles.nvml, _, _ = loadNVMLMgmt([]string{nvmlLibPath})
 		return cHandles
 	}
 	if nvcudaLibPath != "" {
-		cHandles.deviceCount, cHandles.nvcuda, _ = LoadNVCUDAMgmt([]string{nvcudaLibPath})
+		cHandles.deviceCount, cHandles.nvcuda, _, _ = loadNVCUDAMgmt([]string{nvcudaLibPath})
 		return cHandles
 	}
 	if cudartLibPath != "" {
-		cHandles.deviceCount, cHandles.cudart, _ = LoadCUDARTMgmt([]string{cudartLibPath})
+		cHandles.deviceCount, cHandles.cudart, _, _ = loadCUDARTMgmt([]string{cudartLibPath})
 		return cHandles
 	}
 
@@ -102,18 +110,21 @@ func initCudaHandles() *cudaHandles {
 	if len(NvmlGlobs) > 0 {
 		nvmlLibPaths := FindGPULibs(NvmlMgmtName, NvmlGlobs)
 		if len(nvmlLibPaths) > 0 {
-			nvml, libPath := LoadNVMLMgmt(nvmlLibPaths)
+			nvml, libPath, err := loadNVMLMgmt(nvmlLibPaths)
 			if nvml != nil {
 				slog.Debug("nvidia-ml loaded", "library", libPath)
 				cHandles.nvml = nvml
 				nvmlLibPath = libPath
 			}
+			if err != nil {
+				bootstrapErrors = append(bootstrapErrors, err)
+			}
 		}
 	}
 
 	nvcudaLibPaths := FindGPULibs(NvcudaMgmtName, nvcudaMgmtPatterns)
 	if len(nvcudaLibPaths) > 0 {
-		deviceCount, nvcuda, libPath := LoadNVCUDAMgmt(nvcudaLibPaths)
+		deviceCount, nvcuda, libPath, err := loadNVCUDAMgmt(nvcudaLibPaths)
 		if nvcuda != nil {
 			slog.Debug("detected GPUs", "count", deviceCount, "library", libPath)
 			cHandles.nvcuda = nvcuda
@@ -121,11 +132,14 @@ func initCudaHandles() *cudaHandles {
 			nvcudaLibPath = libPath
 			return cHandles
 		}
+		if err != nil {
+			bootstrapErrors = append(bootstrapErrors, err)
+		}
 	}
 
 	cudartLibPaths := FindGPULibs(CudartMgmtName, cudartMgmtPatterns)
 	if len(cudartLibPaths) > 0 {
-		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
+		deviceCount, cudart, libPath, err := loadCUDARTMgmt(cudartLibPaths)
 		if cudart != nil {
 			slog.Debug("detected GPUs", "library", libPath, "count", deviceCount)
 			cHandles.cudart = cudart
@@ -133,6 +147,9 @@ func initCudaHandles() *cudaHandles {
 			cudartLibPath = libPath
 			return cHandles
 		}
+		if err != nil {
+			bootstrapErrors = append(bootstrapErrors, err)
+		}
 	}
 
 	return cHandles
@@ -143,14 +160,19 @@ func initOneAPIHandles() *oneapiHandles {
 	oHandles := &oneapiHandles{}
 
 	// Short Circuit if we already know which library to use
+	// ignore bootstrap errors in this case since we already recorded them
 	if oneapiLibPath != "" {
-		oHandles.deviceCount, oHandles.oneapi, _ = LoadOneapiMgmt([]string{oneapiLibPath})
+		oHandles.deviceCount, oHandles.oneapi, _, _ = loadOneapiMgmt([]string{oneapiLibPath})
 		return oHandles
 	}
 
 	oneapiLibPaths := FindGPULibs(OneapiMgmtName, OneapiGlobs)
 	if len(oneapiLibPaths) > 0 {
-		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath = LoadOneapiMgmt(oneapiLibPaths)
+		var err error
+		oHandles.deviceCount, oHandles.oneapi, oneapiLibPath, err = loadOneapiMgmt(oneapiLibPaths)
+		if err != nil {
+			bootstrapErrors = append(bootstrapErrors, err)
+		}
 	}
 
 	return oHandles
@@ -197,6 +219,7 @@ func GetGPUInfo() GpuInfoList {
 
 	if !bootstrapped {
 		slog.Info("looking for compatible GPUs")
+		bootstrapErrors = []error{}
 		needRefresh = false
 		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
@@ -205,27 +228,34 @@ func GetGPUInfo() GpuInfoList {
 		if err != nil {
 			slog.Warn("error looking up system memory", "error", err)
 		}
+		depPath := LibraryDir()
+		details, err := GetCPUDetails()
+		if err != nil {
+			slog.Warn("failed to lookup CPU details", "error", err)
+		}
 		cpus = []CPUInfo{
 			{
 				GpuInfo: GpuInfo{
-					memInfo: mem,
-					Library: "cpu",
-					Variant: cpuCapability.String(),
-					ID:      "0",
+					memInfo:        mem,
+					Library:        "cpu",
+					Variant:        cpuCapability.String(),
+					ID:             "0",
+					DependencyPath: depPath,
 				},
+				CPUs: details,
 			},
 		}
 
 		// Fallback to CPU mode if we're lacking required vector extensions on x86
 		if cpuCapability < GPURunnerCPUCapability && runtime.GOARCH == "amd64" {
-			slog.Warn("CPU does not have minimum vector extensions, GPU inference disabled", "required", GPURunnerCPUCapability, "detected", cpuCapability)
+			err := fmt.Errorf("CPU does not have minimum vector extensions, GPU inference disabled.  Required:%s  Detected:%s", GPURunnerCPUCapability, cpuCapability)
+			slog.Warn(err.Error())
+			bootstrapErrors = append(bootstrapErrors, err)
 			bootstrapped = true
 			// No need to do any GPU discovery, since we can't run on them
 			return GpuInfoList{cpus[0].GpuInfo}
 		}
 
-		depPath := LibraryDir()
-
 		// Load ALL libraries
 		cHandles = initCudaHandles()
 
@@ -252,10 +282,6 @@ func GetGPUInfo() GpuInfoList {
 					C.free(unsafe.Pointer(memInfo.err))
 					continue
 				}
-				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
-					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
-					continue
-				}
 				gpuInfo.TotalMemory = uint64(memInfo.total)
 				gpuInfo.FreeMemory = uint64(memInfo.free)
 				gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
@@ -278,6 +304,15 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
 				gpuInfo.Variant = variant
 
+				if memInfo.major < CudaComputeMin[0] || (memInfo.major == CudaComputeMin[0] && memInfo.minor < CudaComputeMin[1]) {
+					unsupportedGPUs = append(unsupportedGPUs,
+						UnsupportedGPUInfo{
+							GpuInfo: gpuInfo.GpuInfo,
+						})
+					slog.Info(fmt.Sprintf("[%d] CUDA GPU is too old. Compute Capability detected: %d.%d", i, memInfo.major, memInfo.minor))
+					continue
+				}
+
 				// query the management library as well so we can record any skew between the two
 				// which represents overhead on the GPU we must set aside on subsequent updates
 				if cHandles.nvml != nil {
@@ -340,7 +375,10 @@ func GetGPUInfo() GpuInfoList {
 			}
 		}
 
-		rocmGPUs = AMDGetGPUInfo()
+		rocmGPUs, err = AMDGetGPUInfo()
+		if err != nil {
+			bootstrapErrors = append(bootstrapErrors, err)
+		}
 		bootstrapped = true
 		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
 			slog.Info("no compatible GPUs were discovered")
@@ -525,92 +563,114 @@ func FindGPULibs(baseLibName string, defaultPatterns []string) []string {
 	return gpuLibPaths
 }
 
-func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
+// Bootstrap the runtime library
+// Returns: num devices, handle, libPath, error
+func loadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string, error) {
 	var resp C.cudart_init_resp_t
 	resp.ch.verbose = getVerboseState()
+	var err error
 	for _, libPath := range cudartLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.cudart_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load cudart", "library", libPath, "error", C.GoString(resp.err))
+			err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
+			slog.Debug(err.Error())
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return int(resp.num_devices), &resp.ch, libPath
+			err = nil
+			return int(resp.num_devices), &resp.ch, libPath, err
 		}
 	}
-	return 0, nil, ""
+	return 0, nil, "", err
 }
 
-func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
+// Bootstrap the driver library
+// Returns: num devices, handle, libPath, error
+func loadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string, error) {
 	var resp C.nvcuda_init_resp_t
 	resp.ch.verbose = getVerboseState()
+	var err error
 	for _, libPath := range nvcudaLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
 			// Decide what log level based on the type of error message to help users understand why
-			msg := C.GoString(resp.err)
 			switch resp.cudaErr {
 			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
-				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
+				err = fmt.Errorf("version mismatch between driver and cuda driver library - reboot or upgrade may be required: library %s", libPath)
+				slog.Warn(err.Error())
 			case C.CUDA_ERROR_NO_DEVICE:
-				slog.Info("no nvidia devices detected", "library", libPath)
+				err = fmt.Errorf("no nvidia devices detected by library %s", libPath)
+				slog.Info(err.Error())
 			case C.CUDA_ERROR_UNKNOWN:
-				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
-				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
+				err = fmt.Errorf("unknown error initializing cuda driver library %s: %s. see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information", libPath, C.GoString(resp.err))
+				slog.Warn(err.Error())
 			default:
+				msg := C.GoString(resp.err)
 				if strings.Contains(msg, "wrong ELF class") {
 					slog.Debug("skipping 32bit library", "library", libPath)
 				} else {
-					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
+					err = fmt.Errorf("Unable to load cudart library %s: %s", libPath, C.GoString(resp.err))
+					slog.Info(err.Error())
 				}
 			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return int(resp.num_devices), &resp.ch, libPath
+			err = nil
+			return int(resp.num_devices), &resp.ch, libPath, err
 		}
 	}
-	return 0, nil, ""
+	return 0, nil, "", err
 }
 
-func LoadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string) {
+// Bootstrap the management library
+// Returns: handle, libPath, error
+func loadNVMLMgmt(nvmlLibPaths []string) (*C.nvml_handle_t, string, error) {
 	var resp C.nvml_init_resp_t
 	resp.ch.verbose = getVerboseState()
+	var err error
 	for _, libPath := range nvmlLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.nvml_init(lib, &resp)
 		if resp.err != nil {
-			slog.Info(fmt.Sprintf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err)))
+			err = fmt.Errorf("Unable to load NVML management library %s: %s", libPath, C.GoString(resp.err))
+			slog.Info(err.Error())
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return &resp.ch, libPath
+			err = nil
+			return &resp.ch, libPath, err
 		}
 	}
-	return nil, ""
+	return nil, "", err
 }
 
-func LoadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string) {
+// bootstrap the Intel GPU library
+// Returns: num devices, handle, libPath, error
+func loadOneapiMgmt(oneapiLibPaths []string) (int, *C.oneapi_handle_t, string, error) {
 	var resp C.oneapi_init_resp_t
 	num_devices := 0
 	resp.oh.verbose = getVerboseState()
+	var err error
 	for _, libPath := range oneapiLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
 		C.oneapi_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load oneAPI management library", "library", libPath, "error", C.GoString(resp.err))
+			err = fmt.Errorf("Unable to load oneAPI management library %s: %s", libPath, C.GoString(resp.err))
+			slog.Debug(err.Error())
 			C.free(unsafe.Pointer(resp.err))
 		} else {
+			err = nil
 			for i := range resp.oh.num_drivers {
 				num_devices += int(C.oneapi_get_device_count(resp.oh, C.int(i)))
 			}
-			return num_devices, &resp.oh, libPath
+			return num_devices, &resp.oh, libPath, err
 		}
 	}
-	return 0, nil, ""
+	return 0, nil, "", err
 }
 
 func getVerboseState() C.uint16_t {
@@ -668,3 +728,23 @@ func LibraryDir() string {
 	slog.Warn("unable to locate gpu dependency libraries")
 	return ""
 }
+
+func GetSystemInfo() SystemInfo {
+	gpus := GetGPUInfo()
+	gpuMutex.Lock()
+	defer gpuMutex.Unlock()
+	discoveryErrors := []string{}
+	for _, err := range bootstrapErrors {
+		discoveryErrors = append(discoveryErrors, err.Error())
+	}
+	if len(gpus) == 1 && gpus[0].Library == "cpu" {
+		gpus = []GpuInfo{}
+	}
+
+	return SystemInfo{
+		System:          cpus[0],
+		GPUs:            gpus,
+		UnsupportedGPUs: unsupportedGPUs,
+		DiscoveryErrors: discoveryErrors,
+	}
+}

discover/gpu_darwin.go

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

discover/gpu_linux.go

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

discover/gpu_oneapi.go

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

discover/gpu_test.go

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

discover/gpu_windows.go

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

discover/types.go

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

docs/api.md

@@ -69,7 +69,7 @@ Enable JSON mode by setting the `format` parameter to `json`. This will structur
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?"
 }'
 ```
@@ -80,7 +80,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "response": "The",
   "done": false
@@ -102,7 +102,7 @@ To calculate how fast the response is generated in tokens per second (token/s),
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "",
   "done": true,
@@ -124,7 +124,7 @@ A response can be received in one reply when streaming is off.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "stream": false
 }'
@@ -136,7 +136,7 @@ If `stream` is set to `false`, the response will be a single JSON object:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -194,7 +194,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "What color is the sky at different times of the day? Respond using JSON",
   "format": "json",
   "stream": false
@@ -205,7 +205,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-11-09T21:07:55.186497Z",
   "response": "{\n\"morning\": {\n\"color\": \"blue\"\n},\n\"noon\": {\n\"color\": \"blue-gray\"\n},\n\"afternoon\": {\n\"color\": \"warm gray\"\n},\n\"evening\": {\n\"color\": \"orange\"\n}\n}\n",
   "done": true,
@@ -327,7 +327,7 @@ If you want to set custom options for the model at runtime rather than in the Mo
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "stream": false,
   "options": {
@@ -368,7 +368,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "response": "The sky is blue because it is the color of the sky.",
   "done": true,
@@ -390,7 +390,7 @@ If an empty prompt is provided, the model will be loaded into memory.
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1"
+  "model": "llama3.2"
 }'
 ```
 
@@ -400,13 +400,40 @@ A single JSON object is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-18T19:52:07.071755Z",
   "response": "",
   "done": true
 }
 ```
 
+#### Unload a model
+
+If an empty prompt is provided and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "llama3.2",
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.2",
+  "created_at": "2024-09-12T03:54:03.516566Z",
+  "response": "",
+  "done": true,
+  "done_reason": "unload"
+}
+```
+
 ## Generate a chat completion
 
 ```shell
@@ -445,7 +472,7 @@ Send a chat message with a streaming response.
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -461,7 +488,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -476,7 +503,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 4883583458,
@@ -494,7 +521,7 @@ Final response:
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -509,7 +536,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -533,7 +560,7 @@ Send a chat message with a conversation history. You can use this same approach
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -557,7 +584,7 @@ A stream of JSON objects is returned:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T08:52:19.385406455-07:00",
   "message": {
     "role": "assistant",
@@ -571,7 +598,7 @@ Final response:
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
   "done": true,
   "total_duration": 8113331500,
@@ -629,7 +656,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```shell
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -647,7 +674,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2023-12-12T14:13:43.416799Z",
   "message": {
     "role": "assistant",
@@ -669,7 +696,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "messages": [
     {
       "role": "user",
@@ -708,7 +735,7 @@ curl http://localhost:11434/api/chat -d '{
 
 ```json
 {
-  "model": "llama3.1",
+  "model": "llama3.2",
   "created_at": "2024-07-22T20:33:28.123648Z",
   "message": {
     "role": "assistant",
@@ -736,6 +763,64 @@ curl http://localhost:11434/api/chat -d '{
 }
 ```
 
+#### Load a model
+
+If the messages array is empty, the model will be loaded into memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.2",
+  "messages": []
+}'
+```
+
+##### Response
+```json
+{
+  "model": "llama3.2",
+  "created_at":"2024-09-12T21:17:29.110811Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "load",
+  "done": true
+}
+```
+
+#### Unload a model
+
+If the messages array is empty and the `keep_alive` parameter is set to `0`, a model will be unloaded from memory.
+
+##### Request
+
+```
+curl http://localhost:11434/api/chat -d '{
+  "model": "llama3.2",
+  "messages": [],
+  "keep_alive": 0
+}'
+```
+
+##### Response
+
+A single JSON object is returned:
+
+```json
+{
+  "model": "llama3.2",
+  "created_at":"2024-09-12T21:33:17.547535Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
+  "done_reason": "unload",
+  "done": true
+}
+```
+
 ## Create a Model
 
 ```shell
@@ -904,7 +989,7 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama3.1"
+  "name": "llama3.2"
 }'
 ```
 
@@ -965,7 +1050,7 @@ Copy a model. Creates a model with another name from an existing model.
 
 ```shell
 curl http://localhost:11434/api/copy -d '{
-  "source": "llama3.1",
+  "source": "llama3.2",
   "destination": "llama3-backup"
 }'
 ```
@@ -1020,7 +1105,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where
 
 ```shell
 curl http://localhost:11434/api/pull -d '{
-  "name": "llama3.1"
+  "name": "llama3.2"
 }'
 ```
 

docs/development.md

@@ -2,15 +2,13 @@
 
 Install required tools:
 
-- cmake version 3.24 or higher
 - go version 1.22 or higher
 - gcc version 11.4.0 or higher
 
+
 ### MacOS
 
-```bash
-brew install go cmake gcc
-```
+[Download Go](https://go.dev/dl/)
 
 Optionally enable debugging and more verbose logging:
 
@@ -22,10 +20,10 @@ export CGO_CFLAGS="-g"
 export OLLAMA_DEBUG=1
 ```
 
-Get the required libraries and build the native LLM code:
+Get the required libraries and build the native LLM code:  (Adjust the job count based on your number of processors for a faster build)
 
 ```bash
-go generate ./...
+make -j 5
 ```
 
 Then build ollama:
@@ -40,13 +38,17 @@ Now you can run `ollama`:
 ./ollama
 ```
 
+#### Xcode 15 warnings
+
+If you are using Xcode newer than version 14, you may see a warning during `go build` about `ld: warning: ignoring duplicate libraries: '-lobjc'` due to Golang issue https://github.com/golang/go/issues/67799 which can be safely ignored.  You can suppress the warning with `export CGO_LDFLAGS="-Wl,-no_warn_duplicate_libraries"`
+
 ### Linux
 
 #### Linux CUDA (NVIDIA)
 
 _Your operating system distribution may already have packages for NVIDIA CUDA. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
 
-Install `cmake` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
+Install `make`, `gcc` and `golang` as well as [NVIDIA CUDA](https://developer.nvidia.com/cuda-downloads)
 development and runtime packages.
 
 Typically the build scripts will auto-detect CUDA, however, if your Linux distro
@@ -55,10 +57,10 @@ specifying an environment variable `CUDA_LIB_DIR` to the location of the shared
 libraries, and `CUDACXX` to the location of the nvcc compiler. You can customize
 a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
 
-Then generate dependencies:
+Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
 
 ```
-go generate ./...
+make -j 5
 ```
 
 Then build the binary:
@@ -71,7 +73,7 @@ go build .
 
 _Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_
 
-Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `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 `make`, `gcc`, 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
@@ -80,8 +82,10 @@ install (typically `/opt/rocm`), and `CLBlast_DIR` to the location of the
 CLBlast install (typically `/usr/lib/cmake/CLBlast`). You can also customize
 the AMD GPU targets by setting AMDGPU_TARGETS (e.g. `AMDGPU_TARGETS="gfx1101;gfx1102"`)
 
+Then generate dependencies:  (Adjust the job count based on your number of processors for a faster build)
+
 ```
-go generate ./...
+make -j 5
 ```
 
 Then build the binary:
@@ -94,57 +98,75 @@ ROCm requires elevated privileges to access the GPU at runtime. On most distros
 
 #### Advanced CPU Settings
 
-By default, running `go generate ./...` will compile a few different variations
+By default, running `make` will compile a few different variations
 of the LLM library based on common CPU families and vector math capabilities,
 including a lowest-common-denominator which should run on almost any 64 bit CPU
 somewhat slowly. At runtime, Ollama will auto-detect the optimal variation to
-load. If you would like to build a CPU-based build customized for your
-processor, you can set `OLLAMA_CUSTOM_CPU_DEFS` to the llama.cpp flags you would
-like to use. For example, to compile an optimized binary for an Intel i9-9880H,
-you might use:
+load. 
 
-```
-OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
-go build .
-```
+Custom CPU settings are not currently supported in the new Go server build but will be added back after we complete the transition.
 
 #### Containerized Linux Build
 
-If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
+If you have Docker available, you can build linux binaries with `OLLAMA_NEW_RUNNERS=1 ./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
 
 ### Windows
 
-Note: The Windows build for Ollama is still under development.
+The following tools are required as a minimal development environment to build CPU inference support.
 
-First, install required tools:
-
-- MSVC toolchain - C/C++ and cmake as minimal requirements
 - Go version 1.22 or higher
-- MinGW (pick one variant) with GCC.
-  - [MinGW-w64](https://www.mingw-w64.org/)
+  - https://go.dev/dl/
+- Git
+  - https://git-scm.com/download/win
+- GCC and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
   - [MSYS2](https://www.msys2.org/)
-- The `ThreadJob` Powershell module: `Install-Module -Name ThreadJob -Scope CurrentUser`
+    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-ucrt-x86_64-gcc make` to install the required tools
+  - Assuming you used the default install prefix for msys2 above, add `c:\msys64\ucrt64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
 
 Then, build the `ollama` binary:
 
 ```powershell
 $env:CGO_ENABLED="1"
-go generate ./...
+make -j 8
 go build .
 ```
 
+#### GPU Support
+
+The GPU tools require the Microsoft native build tools.  To build either CUDA or ROCm, you must first install MSVC via Visual Studio:
+
+- Make sure to select `Desktop development with C++` as a Workload during the Visual Studio install
+- You must complete the Visual Studio install and run it once **BEFORE** installing CUDA or ROCm for the tools to properly register
+- Add the location of the **64 bit (x64)** compiler (`cl.exe`) to your `PATH`
+- Note: the default Developer Shell may configure the 32 bit (x86) compiler which will lead to build failures.  Ollama requires a 64 bit toolchain.
+
 #### Windows CUDA (NVIDIA)
 
-In addition to the common Windows development tools described above, install CUDA after installing MSVC.
+In addition to the common Windows development tools and MSVC described above:
 
 - [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 installing MSVC.
+In addition to the common Windows development tools and MSVC described above:
 
 - [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
-- [Strawberry Perl](https://strawberryperl.com/)
 
-Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
+#### Windows arm64
+
+The default `Developer PowerShell for VS 2022` may default to x86 which is not what you want.  To ensure you get an arm64 development environment, start a plain PowerShell terminal and run:
+
+```powershell
+import-module 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\Common7\\Tools\\Microsoft.VisualStudio.DevShell.dll'
+Enter-VsDevShell -Arch arm64 -vsinstallpath 'C:\\Program Files\\Microsoft Visual Studio\\2022\\Community' -skipautomaticlocation
+```
+
+You can confirm with `write-host $env:VSCMD_ARG_TGT_ARCH`
+
+Follow the instructions at https://www.msys2.org/wiki/arm64/ to set up an arm64 msys2 environment.  Ollama requires gcc and mingw32-make to compile, which is not currently available on Windows arm64, but a gcc compatibility adapter is available via `mingw-w64-clang-aarch64-gcc-compat`. At a minimum you will need to install the following:
+
+```
+pacman -S mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-gcc-compat mingw-w64-clang-aarch64-make make
+```
+
+You will need to ensure your PATH includes go, cmake, gcc and clang mingw32-make to build ollama from source. (typically `C:\msys64\clangarm64\bin\`)

docs/docker.md

@@ -63,7 +63,7 @@ docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 114
 Now you can run a model:
 
 ```
-docker exec -it ollama ollama run llama3.1
+docker exec -it ollama ollama run llama3.2
 ```
 
 ### Try different models

docs/faq.md

@@ -32,7 +32,7 @@ When using the API, specify the `num_ctx` parameter:
 
 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3.1",
+  "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
     "num_ctx": 4096
@@ -232,14 +232,18 @@ curl http://localhost:11434/api/chat -d '{"model": "mistral"}'
 
 To preload a model using the CLI, use the command:
 ```shell
-ollama run llama3.1 ""
+ollama run llama3.2 ""
 ```
 
 ## How do I keep a model loaded in memory or make it unload immediately?
 
-By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you are making numerous requests to the LLM. You may, however, want to free up the memory before the 5 minutes have elapsed or keep the model loaded indefinitely. Use the `keep_alive` parameter with either the `/api/generate` and `/api/chat` API endpoints to control how long the model is left in memory.
+By default models are kept in memory for 5 minutes before being unloaded. This allows for quicker response times if you're making numerous requests to the LLM. If you want to immediately unload a model from memory, use the `ollama stop` command:
 
-The `keep_alive` parameter can be set to:
+```shell
+ollama stop llama3.2
+```
+
+If you're using the API, use the `keep_alive` parameter with the `/api/generate` and `/api/chat` endpoints to set the amount of time that a model stays in memory. The `keep_alive` parameter can be set to:
 * a duration string (such as "10m" or "24h")
 * a number in seconds (such as 3600)
 * any negative number which will keep the model loaded in memory (e.g. -1 or "-1m")
@@ -247,17 +251,17 @@ The `keep_alive` parameter can be set to:
 
 For example, to preload a model and leave it in memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": -1}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": -1}'
 ```
 
 To unload the model and free up memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3.1", "keep_alive": 0}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": 0}'
 ```
 
-Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
+Alternatively, you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. The `OLLAMA_KEEP_ALIVE` variable uses the same parameter types as the `keep_alive` parameter types mentioned above. Refer to the section explaining [how to configure the Ollama server](#how-do-i-configure-ollama-server) to correctly set the environment variable.
 
-If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` API parameter with the `/api/generate` or `/api/chat` API endpoints.
+The `keep_alive` API parameter with the `/api/generate` and `/api/chat` API endpoints will override the `OLLAMA_KEEP_ALIVE` setting.
 
 ## How do I manage the maximum number of requests the Ollama server can queue?
 

docs/modelfile.md

@@ -50,7 +50,7 @@ INSTRUCTION arguments
 An example of a `Modelfile` creating a mario blueprint:
 
 ```modelfile
-FROM llama3.1
+FROM llama3.2
 # sets the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
 # sets the context window size to 4096, this controls how many tokens the LLM can use as context to generate the next token
@@ -72,10 +72,10 @@ More examples are available in the [examples directory](../examples).
 To view the Modelfile of a given model, use the `ollama show --modelfile` command.
 
   ```bash
-  > ollama show --modelfile llama3.1
+  > ollama show --modelfile llama3.2
   # Modelfile generated by "ollama show"
   # To build a new Modelfile based on this one, replace the FROM line with:
-  # FROM llama3.1:latest
+  # FROM llama3.2:latest
   FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
   TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
 
@@ -103,7 +103,7 @@ FROM <model name>:<tag>
 #### Build from existing model
 
 ```modelfile
-FROM llama3.1
+FROM llama3.2
 ```
 
 A list of available base models:

docs/openai.md

@@ -25,7 +25,7 @@ chat_completion = client.chat.completions.create(
             'content': 'Say this is a test',
         }
     ],
-    model='llama3.1',
+    model='llama3.2',
 )
 
 response = client.chat.completions.create(
@@ -37,7 +37,7 @@ response = client.chat.completions.create(
                 {"type": "text", "text": "What's in this image?"},
                 {
                     "type": "image_url",
-                    "image_url": "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",
+                    "image_url": "data:image/png;base64,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",
                 },
             ],
         }
@@ -46,13 +46,13 @@ response = client.chat.completions.create(
 )
 
 completion = client.completions.create(
-    model="llama3.1",
+    model="llama3.2",
     prompt="Say this is a test",
 )
 
 list_completion = client.models.list()
 
-model = client.models.retrieve("llama3.1")
+model = client.models.retrieve("llama3.2")
 
 embeddings = client.embeddings.create(
     model="all-minilm",
@@ -74,7 +74,7 @@ const openai = new OpenAI({
 
 const chatCompletion = await openai.chat.completions.create({
     messages: [{ role: 'user', content: 'Say this is a test' }],
-    model: 'llama3.1',
+    model: 'llama3.2',
 })
 
 const response = await openai.chat.completions.create({
@@ -86,7 +86,7 @@ const response = await openai.chat.completions.create({
             { type: "text", text: "What's in this image?" },
             {
             type: "image_url",
-            image_url: "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",
+            image_url: "data:image/png;base64,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",
             },
         ],
         },
@@ -94,13 +94,13 @@ const response = await openai.chat.completions.create({
 })
 
 const completion = await openai.completions.create({
-    model: "llama3.1",
+    model: "llama3.2",
     prompt: "Say this is a test.",
 })
 
 const listCompletion = await openai.models.list()
 
-const model = await openai.models.retrieve("llama3.1")
+const model = await openai.models.retrieve("llama3.2")
 
 const embedding = await openai.embeddings.create({
   model: "all-minilm",
@@ -114,7 +114,7 @@ const embedding = await openai.embeddings.create({
 curl http://localhost:11434/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
-        "model": "llama3.1",
+        "model": "llama3.2",
         "messages": [
             {
                 "role": "system",
@@ -142,7 +142,7 @@ curl http://localhost:11434/v1/chat/completions \
           {
             "type": "image_url",
             "image_url": {
-               "url": "iVBORw0KGgoAAAANSUhEUgAAAG0AAABmCAYAAADBPx+VAAAACXBIWXMAAAsTAAALEwEAmpwYAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAA3VSURBVHgB7Z27r0zdG8fX743i1bi1ikMoFMQloXRpKFFIqI7LH4BEQ+NWIkjQuSWCRIEoULk0gsK1kCBI0IhrQVT7tz/7zZo888yz1r7MnDl7z5xvsjkzs2fP3uu71nNfa7lkAsm7d++Sffv2JbNmzUqcc8m0adOSzZs3Z+/XES4ZckAWJEGWPiCxjsQNLWmQsWjRIpMseaxcuTKpG/7HP27I8P79e7dq1ars/yL4/v27S0ejqwv+cUOGEGGpKHR37tzJCEpHV9tnT58+dXXCJDdECBE2Ojrqjh071hpNECjx4cMHVycM1Uhbv359B2F79+51586daxN/+pyRkRFXKyRDAqxEp4yMlDDzXG1NPnnyJKkThoK0VFd1ELZu3TrzXKxKfW7dMBQ6bcuWLW2v0VlHjx41z717927ba22U9APcw7Nnz1oGEPeL3m3p2mTAYYnFmMOMXybPPXv2bNIPpFZr1NHn4HMw0KRBjg9NuRw95s8PEcz/6DZELQd/09C9QGq5RsmSRybqkwHGjh07OsJSsYYm3ijPpyHzoiacg35MLdDSIS/O1yM778jOTwYUkKNHWUzUWaOsylE00MyI0fcnOwIdjvtNdW/HZwNLGg+sR1kMepSNJXmIwxBZiG8tDTpEZzKg0GItNsosY8USkxDhD0Rinuiko2gfL/RbiD2LZAjU9zKQJj8RDR0vJBR1/Phx9+PHj9Z7REF4nTZkxzX4LCXHrV271qXkBAPGfP/atWvu/PnzHe4C97F48eIsRLZ9+3a3f/9+87dwP1JxaF7/3r17ba+5l4EcaVo0lj3SBq5kGTJSQmLWMjgYNei2GPT1MuMqGTDEFHzeQSP2wi/jGnkmPJ/nhccs44jvDAxpVcxnq0F6eT8h4ni/iIWpR5lPyA6ETkNXoSukvpJAD3AsXLiwpZs49+fPn5ke4j10TqYvegSfn0OnafC+Tv9ooA/JPkgQysqQNBzagXY55nO/oa1F7qvIPWkRL12WRpMWUvpVDYmxAPehxWSe8ZEXL20sadYIozfmNch4QJPAfeJgW3rNsnzphBKNJM2KKODo1rVOMRYik5ETy3ix4qWNI81qAAirizgMIc+yhTytx0JWZuNI03qsrgWlGtwjoS9XwgUhWGyhUaRZZQNNIEwCiXD16tXcAHUs79co0vSD8rrJCIW98pzvxpAWyyo3HYwqS0+H0BjStClcZJT5coMm6D2LOF8TolGJtK9fvyZpyiC5ePFi9nc/oJU4eiEP0jVoAnHa9wyJycITMP78+eMeP37sXrx44d6+fdt6f82aNdkx1pg9e3Zb5W+RSRE+n+VjksQWifvVaTKFhn5O8my63K8Qabdv33b379/PiAP//vuvW7BggZszZ072/+TJk91YgkafPn166zXB1rQHFvouAWHq9z3SEevSUerqCn2/dDCeta2jxYbr69evk4MHDyY7d+7MjhMnTiTPnz9Pfv/+nfQT2ggpO2dMF8cghuoM7Ygj5iWCqRlGFml0QC/ftGmTmzt3rmsaKDsgBSPh0/8yPeLLBihLkOKJc0jp8H8vUzcxIA1k6QJ/c78tWEyj5P3o4u9+jywNPdJi5rAH9x0KHcl4Hg570eQp3+vHXGyrmEeigzQsQsjavXt38ujRo44LQuDDhw+TW7duRS1HGgMxhNXHgflaNTOsHyKvHK5Ijo2jbFjJBQK9YwFd6RVMzfgRBmEfP37suBBm/p49e1qjEP2mwTViNRo0VJWH1deMXcNK08uUjVUu7s/zRaL+oLNxz1bpANco4npUgX4G2eFbpDFyQoQxojBCpEGSytmOH8qrH5Q9vuzD6ofQylkCUmh8DBAr+q8JCyVNtWQIidKQE9wNtLSQnS4jDSsxNHogzFuQBw4cyM61UKVsjfr3ooBkPSqqQHesUPWVtzi9/vQi1T+rJj7WiTz4Pt/l3LxUkr5P2VYZaZ4URpsE+st/dujQoaBBYokbrz/8TJNQYLSonrPS9kUaSkPeZyj1AWSj+d+VBoy1pIWVNed8P0Ll/ee5HdGRhrHhR5GGN0r4LGZBaj8oFDJitBTJzIZgFcmU0Y8ytWMZMzJOaXUSrUs5RxKnrxmbb5YXO9VGUhtpXldhEUogFr3IzIsvlpmdosVcGVGXFWp2oU9kLFL3dEkSz6NHEY1sjSRdIuDFWEhd8KxFqsRi1uM/nz9/zpxnwlESONdg6dKlbsaMGS4EHFHtjFIDHwKOo46l4TxSuxgDzi+rE2jg+BaFruOX4HXa0Nnf1lwAPufZeF8/r6zD97WK2qFnGjBxTw5qNGPxT+5T/r7/7RawFC3j4vTp09koCxkeHjqbHJqArmH5UrFKKksnxrK7FuRIs8STfBZv+luugXZ2pR/pP9Ois4z+TiMzUUkUjD0iEi1fzX8GmXyuxUBRcaUfykV0YZnlJGKQpOiGB76x5GeWkWWJc3mOrK6S7xdND+W5N6XyaRgtWJFe13GkaZnKOsYqGdOVVVbGupsyA/l7emTLHi7vwTdirNEt0qxnzAvBFcnQF16xh/TMpUuXHDowhlA9vQVraQhkudRdzOnK+04ZSP3DUhVSP61YsaLtd/ks7ZgtPcXqPqEafHkdqa84X6aCeL7YWlv6edGFHb+ZFICPlljHhg0bKuk0CSvVznWsotRu433alNdFrqG45ejoaPCaUkWERpLXjzFL2Rpllp7PJU2a/v7Ab8N05/9t27Z16KUqoFGsxnI9EosS2niSYg9SpU6B4JgTrvVW1flt1sT+0ADIJU2maXzcUTraGCRaL1Wp9rUMk16PMom8QhruxzvZIegJjFU7LLCePfS8uaQdPny4jTTL0dbee5mYokQsXTIWNY46kuMbnt8Kmec+LGWtOVIl9cT1rCB0V8WqkjAsRwta93TbwNYoGKsUSChN44lgBNCoHLHzquYKrU6qZ8lolCIN0Rh6cP0Q3U6I6IXILYOQI513hJaSKAorFpuHXJNfVlpRtmYBk1Su1obZr5dnKAO+L10Hrj3WZW+E3qh6IszE37F6EB+68mGpvKm4eb9bFrlzrok7fvr0Kfv727dvWRmdVTJHw0qiiCUSZ6wCK+7XL/AcsgNyL74DQQ730sv78Su7+t/A36MdY0sW5o40ahslXr58aZ5HtZB8GH64m9EmMZ7FpYw4T6QnrZfgenrhFxaSiSGXtPnz57e9TkNZLvTjeqhr734CNtrK41L40sUQckmj1lGKQ0rC37x544r8eNXRpnVE3ZZY7zXo8NomiO0ZUCj2uHz58rbXoZ6gc0uA+F6ZeKS/jhRDUq8MKrTho9fEkihMmhxtBI1DxKFY9XLpVcSkfoi8JGnToZO5sU5aiDQIW716ddt7ZLYtMQlhECdBGXZZMWldY5BHm5xgAroWj4C0hbYkSc/jBmggIrXJWlZM6pSETsEPGqZOndr2uuuR5rF169a2HoHPdurUKZM4CO1WTPqaDaAd+GFGKdIQkxAn9RuEWcTRyN2KSUgiSgF5aWzPTeA/lN5rZubMmR2bE4SIC4nJoltgAV/dVefZm72AtctUCJU2CMJ327hxY9t7EHbkyJFseq+EJSY16RPo3Dkq1kkr7+q0bNmyDuLQcZBEPYmHVdOBiJyIlrRDq41YPWfXOxUysi5fvtyaj+2BpcnsUV/oSoEMOk2CQGlr4ckhBwaetBhjCwH0ZHtJROPJkyc7UjcYLDjmrH7ADTEBXFfOYmB0k9oYBOjJ8b4aOYSe7QkKcYhFlq3QYLQhSidNmtS2RATwy8YOM3EQJsUjKiaWZ+vZToUQgzhkHXudb/PW5YMHD9yZM2faPsMwoc7RciYJXbGuBqJ1UIGKKLv915jsvgtJxCZDubdXr165mzdvtr1Hz5LONA8jrUwKPqsmVesKa49S3Q4WxmRPUEYdTjgiUcfUwLx589ySJUva3oMkP6IYddq6HMS4o55xBJBUeRjzfa4Zdeg56QZ43LhxoyPo7Lf1kNt7oO8wWAbNwaYjIv5lhyS7kRf96dvm5Jah8vfvX3flyhX35cuX6HfzFHOToS1H4BenCaHvO8pr8iDuwoUL7tevX+b5ZdbBair0xkFIlFDlW4ZknEClsp/TzXyAKVOmmHWFVSbDNw1l1+4f90U6IY/q4V27dpnE9bJ+v87QEydjqx/UamVVPRG+mwkNTYN+9tjkwzEx+atCm/X9WvWtDtAb68Wy9LXa1UmvCDDIpPkyOQ5ZwSzJ4jMrvFcr0rSjOUh+GcT4LSg5ugkW1Io0/SCDQBojh0hPlaJdah+tkVYrnTZowP8iq1F1TgMBBauufyB33x1v+NWFYmT5KmppgHC+NkAgbmRkpD3yn9QIseXymoTQFGQmIOKTxiZIWpvAatenVqRVXf2nTrAWMsPnKrMZHz6bJq5jvce6QK8J1cQNgKxlJapMPdZSR64/UivS9NztpkVEdKcrs5alhhWP9NeqlfWopzhZScI6QxseegZRGeg5a8C3Re1Mfl1ScP36ddcUaMuv24iOJtz7sbUjTS4qBvKmstYJoUauiuD3k5qhyr7QdUHMeCgLa1Ear9NquemdXgmum4fvJ6w1lqsuDhNrg1qSpleJK7K3TF0Q2jSd94uSZ60kK1e3qyVpQK6PVWXp2/FC3mp6jBhKKOiY2h3gtUV64TWM6wDETRPLDfSakXmH3w8g9Jlug8ZtTt4kVF0kLUYYmCCtD/DrQ5YhMGbA9L3ucdjh0y8kOHW5gU/VEEmJTcL4Pz/f7mgoAbYkAAAAAElFTkSuQmCC"
+               "url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAG0AAABmCAYAAADBPx+VAAAACXBIWXMAAAsTAAALEwEAmpwYAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAA3VSURBVHgB7Z27r0zdG8fX743i1bi1ikMoFMQloXRpKFFIqI7LH4BEQ+NWIkjQuSWCRIEoULk0gsK1kCBI0IhrQVT7tz/7zZo888yz1r7MnDl7z5xvsjkzs2fP3uu71nNfa7lkAsm7d++Sffv2JbNmzUqcc8m0adOSzZs3Z+/XES4ZckAWJEGWPiCxjsQNLWmQsWjRIpMseaxcuTKpG/7HP27I8P79e7dq1ars/yL4/v27S0ejqwv+cUOGEGGpKHR37tzJCEpHV9tnT58+dXXCJDdECBE2Ojrqjh071hpNECjx4cMHVycM1Uhbv359B2F79+51586daxN/+pyRkRFXKyRDAqxEp4yMlDDzXG1NPnnyJKkThoK0VFd1ELZu3TrzXKxKfW7dMBQ6bcuWLW2v0VlHjx41z717927ba22U9APcw7Nnz1oGEPeL3m3p2mTAYYnFmMOMXybPPXv2bNIPpFZr1NHn4HMw0KRBjg9NuRw95s8PEcz/6DZELQd/09C9QGq5RsmSRybqkwHGjh07OsJSsYYm3ijPpyHzoiacg35MLdDSIS/O1yM778jOTwYUkKNHWUzUWaOsylE00MyI0fcnOwIdjvtNdW/HZwNLGg+sR1kMepSNJXmIwxBZiG8tDTpEZzKg0GItNsosY8USkxDhD0Rinuiko2gfL/RbiD2LZAjU9zKQJj8RDR0vJBR1/Phx9+PHj9Z7REF4nTZkxzX4LCXHrV271qXkBAPGfP/atWvu/PnzHe4C97F48eIsRLZ9+3a3f/9+87dwP1JxaF7/3r17ba+5l4EcaVo0lj3SBq5kGTJSQmLWMjgYNei2GPT1MuMqGTDEFHzeQSP2wi/jGnkmPJ/nhccs44jvDAxpVcxnq0F6eT8h4ni/iIWpR5lPyA6ETkNXoSukvpJAD3AsXLiwpZs49+fPn5ke4j10TqYvegSfn0OnafC+Tv9ooA/JPkgQysqQNBzagXY55nO/oa1F7qvIPWkRL12WRpMWUvpVDYmxAPehxWSe8ZEXL20sadYIozfmNch4QJPAfeJgW3rNsnzphBKNJM2KKODo1rVOMRYik5ETy3ix4qWNI81qAAirizgMIc+yhTytx0JWZuNI03qsrgWlGtwjoS9XwgUhWGyhUaRZZQNNIEwCiXD16tXcAHUs79co0vSD8rrJCIW98pzvxpAWyyo3HYwqS0+H0BjStClcZJT5coMm6D2LOF8TolGJtK9fvyZpyiC5ePFi9nc/oJU4eiEP0jVoAnHa9wyJycITMP78+eMeP37sXrx44d6+fdt6f82aNdkx1pg9e3Zb5W+RSRE+n+VjksQWifvVaTKFhn5O8my63K8Qabdv33b379/PiAP//vuvW7BggZszZ072/+TJk91YgkafPn166zXB1rQHFvouAWHq9z3SEevSUerqCn2/dDCeta2jxYbr69evk4MHDyY7d+7MjhMnTiTPnz9Pfv/+nfQT2ggpO2dMF8cghuoM7Ygj5iWCqRlGFml0QC/ftGmTmzt3rmsaKDsgBSPh0/8yPeLLBihLkOKJc0jp8H8vUzcxIA1k6QJ/c78tWEyj5P3o4u9+jywNPdJi5rAH9x0KHcl4Hg570eQp3+vHXGyrmEeigzQsQsjavXt38ujRo44LQuDDhw+TW7duRS1HGgMxhNXHgflaNTOsHyKvHK5Ijo2jbFjJBQK9YwFd6RVMzfgRBmEfP37suBBm/p49e1qjEP2mwTViNRo0VJWH1deMXcNK08uUjVUu7s/zRaL+oLNxz1bpANco4npUgX4G2eFbpDFyQoQxojBCpEGSytmOH8qrH5Q9vuzD6ofQylkCUmh8DBAr+q8JCyVNtWQIidKQE9wNtLSQnS4jDSsxNHogzFuQBw4cyM61UKVsjfr3ooBkPSqqQHesUPWVtzi9/vQi1T+rJj7WiTz4Pt/l3LxUkr5P2VYZaZ4URpsE+st/dujQoaBBYokbrz/8TJNQYLSonrPS9kUaSkPeZyj1AWSj+d+VBoy1pIWVNed8P0Ll/ee5HdGRhrHhR5GGN0r4LGZBaj8oFDJitBTJzIZgFcmU0Y8ytWMZMzJOaXUSrUs5RxKnrxmbb5YXO9VGUhtpXldhEUogFr3IzIsvlpmdosVcGVGXFWp2oU9kLFL3dEkSz6NHEY1sjSRdIuDFWEhd8KxFqsRi1uM/nz9/zpxnwlESONdg6dKlbsaMGS4EHFHtjFIDHwKOo46l4TxSuxgDzi+rE2jg+BaFruOX4HXa0Nnf1lwAPufZeF8/r6zD97WK2qFnGjBxTw5qNGPxT+5T/r7/7RawFC3j4vTp09koCxkeHjqbHJqArmH5UrFKKksnxrK7FuRIs8STfBZv+luugXZ2pR/pP9Ois4z+TiMzUUkUjD0iEi1fzX8GmXyuxUBRcaUfykV0YZnlJGKQpOiGB76x5GeWkWWJc3mOrK6S7xdND+W5N6XyaRgtWJFe13GkaZnKOsYqGdOVVVbGupsyA/l7emTLHi7vwTdirNEt0qxnzAvBFcnQF16xh/TMpUuXHDowhlA9vQVraQhkudRdzOnK+04ZSP3DUhVSP61YsaLtd/ks7ZgtPcXqPqEafHkdqa84X6aCeL7YWlv6edGFHb+ZFICPlljHhg0bKuk0CSvVznWsotRu433alNdFrqG45ejoaPCaUkWERpLXjzFL2Rpllp7PJU2a/v7Ab8N05/9t27Z16KUqoFGsxnI9EosS2niSYg9SpU6B4JgTrvVW1flt1sT+0ADIJU2maXzcUTraGCRaL1Wp9rUMk16PMom8QhruxzvZIegJjFU7LLCePfS8uaQdPny4jTTL0dbee5mYokQsXTIWNY46kuMbnt8Kmec+LGWtOVIl9cT1rCB0V8WqkjAsRwta93TbwNYoGKsUSChN44lgBNCoHLHzquYKrU6qZ8lolCIN0Rh6cP0Q3U6I6IXILYOQI513hJaSKAorFpuHXJNfVlpRtmYBk1Su1obZr5dnKAO+L10Hrj3WZW+E3qh6IszE37F6EB+68mGpvKm4eb9bFrlzrok7fvr0Kfv727dvWRmdVTJHw0qiiCUSZ6wCK+7XL/AcsgNyL74DQQ730sv78Su7+t/A36MdY0sW5o40ahslXr58aZ5HtZB8GH64m9EmMZ7FpYw4T6QnrZfgenrhFxaSiSGXtPnz57e9TkNZLvTjeqhr734CNtrK41L40sUQckmj1lGKQ0rC37x544r8eNXRpnVE3ZZY7zXo8NomiO0ZUCj2uHz58rbXoZ6gc0uA+F6ZeKS/jhRDUq8MKrTho9fEkihMmhxtBI1DxKFY9XLpVcSkfoi8JGnToZO5sU5aiDQIW716ddt7ZLYtMQlhECdBGXZZMWldY5BHm5xgAroWj4C0hbYkSc/jBmggIrXJWlZM6pSETsEPGqZOndr2uuuR5rF169a2HoHPdurUKZM4CO1WTPqaDaAd+GFGKdIQkxAn9RuEWcTRyN2KSUgiSgF5aWzPTeA/lN5rZubMmR2bE4SIC4nJoltgAV/dVefZm72AtctUCJU2CMJ327hxY9t7EHbkyJFseq+EJSY16RPo3Dkq1kkr7+q0bNmyDuLQcZBEPYmHVdOBiJyIlrRDq41YPWfXOxUysi5fvtyaj+2BpcnsUV/oSoEMOk2CQGlr4ckhBwaetBhjCwH0ZHtJROPJkyc7UjcYLDjmrH7ADTEBXFfOYmB0k9oYBOjJ8b4aOYSe7QkKcYhFlq3QYLQhSidNmtS2RATwy8YOM3EQJsUjKiaWZ+vZToUQgzhkHXudb/PW5YMHD9yZM2faPsMwoc7RciYJXbGuBqJ1UIGKKLv915jsvgtJxCZDubdXr165mzdvtr1Hz5LONA8jrUwKPqsmVesKa49S3Q4WxmRPUEYdTjgiUcfUwLx589ySJUva3oMkP6IYddq6HMS4o55xBJBUeRjzfa4Zdeg56QZ43LhxoyPo7Lf1kNt7oO8wWAbNwaYjIv5lhyS7kRf96dvm5Jah8vfvX3flyhX35cuX6HfzFHOToS1H4BenCaHvO8pr8iDuwoUL7tevX+b5ZdbBair0xkFIlFDlW4ZknEClsp/TzXyAKVOmmHWFVSbDNw1l1+4f90U6IY/q4V27dpnE9bJ+v87QEydjqx/UamVVPRG+mwkNTYN+9tjkwzEx+atCm/X9WvWtDtAb68Wy9LXa1UmvCDDIpPkyOQ5ZwSzJ4jMrvFcr0rSjOUh+GcT4LSg5ugkW1Io0/SCDQBojh0hPlaJdah+tkVYrnTZowP8iq1F1TgMBBauufyB33x1v+NWFYmT5KmppgHC+NkAgbmRkpD3yn9QIseXymoTQFGQmIOKTxiZIWpvAatenVqRVXf2nTrAWMsPnKrMZHz6bJq5jvce6QK8J1cQNgKxlJapMPdZSR64/UivS9NztpkVEdKcrs5alhhWP9NeqlfWopzhZScI6QxseegZRGeg5a8C3Re1Mfl1ScP36ddcUaMuv24iOJtz7sbUjTS4qBvKmstYJoUauiuD3k5qhyr7QdUHMeCgLa1Ear9NquemdXgmum4fvJ6w1lqsuDhNrg1qSpleJK7K3TF0Q2jSd94uSZ60kK1e3qyVpQK6PVWXp2/FC3mp6jBhKKOiY2h3gtUV64TWM6wDETRPLDfSakXmH3w8g9Jlug8ZtTt4kVF0kLUYYmCCtD/DrQ5YhMGbA9L3ucdjh0y8kOHW5gU/VEEmJTcL4Pz/f7mgoAbYkAAAAAElFTkSuQmCC"
             }
           }
         ]
@@ -154,13 +154,13 @@ curl http://localhost:11434/v1/chat/completions \
 curl http://localhost:11434/v1/completions \
     -H "Content-Type: application/json" \
     -d '{
-        "model": "llama3.1",
+        "model": "llama3.2",
         "prompt": "Say this is a test"
     }'
 
 curl http://localhost:11434/v1/models
 
-curl http://localhost:11434/v1/models/llama3.1
+curl http://localhost:11434/v1/models/llama3.2
 
 curl http://localhost:11434/v1/embeddings \
     -H "Content-Type: application/json" \
@@ -274,7 +274,7 @@ curl http://localhost:11434/v1/embeddings \
 Before using a model, pull it locally `ollama pull`:
 
 ```shell
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 ### Default model names
@@ -282,7 +282,7 @@ ollama pull llama3.1
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:
 
 ```
-ollama cp llama3.1 gpt-3.5-turbo
+ollama cp llama3.2 gpt-3.5-turbo
 ```
 
 Afterwards, this new model name can be specified the `model` field:

docs/template.md

@@ -33,7 +33,7 @@ Omitting a template in these models puts the responsibility of correctly templat
 To add templates in your model, you'll need to add a `TEMPLATE` command to the Modelfile. Here's an example using Meta's Llama 3.
 
 ```dockerfile
-FROM llama3.1
+FROM llama3.2
 
 TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>
 

docs/tutorials/langchainjs.md

@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";
 
 const ollama = new Ollama({
   baseUrl: "http://localhost:11434",
-  model: "llama3.1",
+  model: "llama3.2",
 });
 
 const answer = await ollama.invoke(`why is the sky blue?`);
@@ -23,7 +23,7 @@ const answer = await ollama.invoke(`why is the sky blue?`);
 console.log(answer);
 ```
 
-That will get us the same thing as if we ran `ollama run llama3.1 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
+That will get us the same thing as if we ran `ollama run llama3.2 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
 
 ```bash
 npm install cheerio

docs/windows.md

@@ -29,7 +29,7 @@ Ollama uses unicode characters for progress indication, which may render as unkn
 
 Here's a quick example showing API access from `powershell`
 ```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama3.1", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+(Invoke-WebRequest -method POST -Body '{"model":"llama3.2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
 ```
 
 ## Troubleshooting

envconfig/config.go

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

envconfig/config_test.go

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

examples/go-chat/main.go

@@ -35,7 +35,7 @@ func main() {
 
 	ctx := context.Background()
 	req := &api.ChatRequest{
-		Model:    "llama3.1",
+		Model:    "llama3.2",
 		Messages: messages,
 	}
 

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

@@ -4,10 +4,10 @@ This example provides an interface for asking questions to a PDF document.
 
 ## Setup
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
 ```
-ollama pull llama3.1
+ollama pull llama3.2
 ```
 
 2. Install the Python Requirements.

examples/langchain-python-rag-document/main.py

@@ -51,7 +51,7 @@ while True:
         template=template,
     )
 
-    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.2", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
     qa_chain = RetrievalQA.from_chain_type(
         llm,
         retriever=vectorstore.as_retriever(),

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

@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-rag-websummary/main.py

@@ -5,7 +5,7 @@ from langchain.chains.summarize import load_summarize_chain
 loader = WebBaseLoader("https://ollama.com/blog/run-llama2-uncensored-locally")
 docs = loader.load()
 
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3.2")
 chain = load_summarize_chain(llm, chain_type="stuff")
 
 result = chain.invoke(docs)

examples/langchain-python-simple/README.md

@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/langchain-python-simple/main.py

@@ -1,6 +1,6 @@
 from langchain.llms import Ollama
 
 input = input("What is your question?")
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3.2")
 res = llm.predict(input)
 print (res)

examples/modelfile-mario/Modelfile

@@ -1,4 +1,4 @@
-FROM llama3.1
+FROM llama3.2
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from super mario bros, acting as an assistant.

examples/modelfile-mario/readme.md

@@ -2,12 +2,12 @@
 
 # Example character: Mario
 
-This example shows how to create a basic character using Llama3.1 as the base model.
+This example shows how to create a basic character using Llama 3.2 as the base model.
 
 To run this example:
 
 1. Download the Modelfile
-2. `ollama pull llama3.1` to get the base model used in the model file.
+2. `ollama pull llama3.2` to get the base model used in the model file.
 3. `ollama create NAME -f ./Modelfile`
 4. `ollama run NAME`
 
@@ -18,7 +18,7 @@ Ask it some questions like "Who are you?" or "Is Peach in trouble again?"
 What the model file looks like:
 
 ```
-FROM llama3.1
+FROM llama3.2
 PARAMETER temperature 1
 SYSTEM """
 You are Mario from Super Mario Bros, acting as an assistant.

examples/python-grounded-factuality-rag-check/README.md

@@ -0,0 +1,93 @@
+# RAG Hallucination Checker using Bespoke-Minicheck
+
+This example allows the user to ask questions related to a document, which can be specified via an article url. Relevant chunks are retreived from the document and given to `llama3.2` as context to answer the question. Then each sentence in the answer is checked against the retrieved chunks using `bespoke-minicheck` to ensure that the answer does not contain hallucinations. 
+
+## Running the Example
+
+1. Ensure `all-minilm` (embedding) `llama3.2` (chat) and `bespoke-minicheck` (check) models installed:
+
+   ```bash
+   ollama pull all-minilm
+   ollama pull llama3.2
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies.
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the example:
+
+   ```bash
+   python main.py
+   ```
+
+## Expected Output
+
+```text
+Enter the URL of an article you want to chat with, or press Enter for default example:
+
+Loaded, chunked, and embedded text from https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt.
+
+Enter your question or type quit: Who is the CEO of openai?
+
+Retrieved chunks:
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence .
+
+OpenAI is releasing a new model called o1 , the first in a planned series of “ reasoning ” models that have been trained to answer more complex questions , faster than a human can . It ’ s being released alongside o1-mini , a smaller , cheaper version . And yes , if you ’ re steeped in AI rumors : this is , in fact , the extremely hyped Strawberry model . For OpenAI , o1 represents a step toward its broader goal of human-like artificial intelligence . More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+LLM Answer:
+The text does not mention the CEO of OpenAI. It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+
+LLM Claim: The text does not mention the CEO of OpenAI.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: It only discusses the release of a new model called o1 and some details about it, but does not provide information on the company's leadership.
+Is this claim supported by the context according to bespoke-minicheck? No
+```
+
+The second claim is unsupported since the text mentions the research lead. 
+
+Another tricky example:
+
+```text
+
+Enter your question or type quit: what sets o1 apart from gpt-4o?
+
+Retrieved chunks: 
+OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens . The training behind o1 is fundamentally different from its predecessors , OpenAI ’ s research lead , Jerry Tworek , tells me , though the company is being vague about the exact details . He says o1 “ has been trained using a completely new optimization algorithm and a new training dataset specifically tailored for it. ” Image : OpenAI OpenAI taught previous GPT models to mimic patterns from its training data .
+
+He says OpenAI also tested o1 against a qualifying exam for the International Mathematics Olympiad , and while GPT-4o only correctly solved only 13 percent of problems , o1 scored 83 percent . “ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world .
+
+More practically , it does a better job at writing code and solving multistep problems than previous models . But it ’ s also more expensive and slower to use than GPT-4o . OpenAI is calling this release of o1 a “ preview ” to emphasize how nascent it is . ChatGPT Plus and Team users get access to both o1-preview and o1-mini starting today , while Enterprise and Edu users will get access early next week . OpenAI says it plans to bring o1-mini access to all the free users of ChatGPT but hasn ’ t set a release date yet . Developer access to o1 is really expensive : In the API , o1-preview is $ 15 per 1 million input tokens , or chunks of text parsed by the model , and $ 60 per 1 million output tokens . For comparison , GPT-4o costs $ 5 per 1 million input tokens and $ 15 per 1 million output tokens .
+
+“ We can ’ t say we solved hallucinations ” In online programming contests known as Codeforces competitions , this new model reached the 89th percentile of participants , and OpenAI claims the next update of this model will perform “ similarly to PhD students on challenging benchmark tasks in physics , chemistry and biology. ” At the same time , o1 is not as capable as GPT-4o in a lot of areas . It doesn ’ t do as well on factual knowledge about the world . It also doesn ’ t have the ability to browse the web or process files and images . Still , the company believes it represents a brand-new class of capabilities . It was named o1 to indicate “ resetting the counter back to 1. ” “ I ’ m gon na be honest : I think we ’ re terrible at naming , traditionally , ” McGrew says .
+LLM Answer: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+* The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+* However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+
+LLM Claim: According to the text, several things set o1 apart from GPT-4o:
+
+* In online programming contests (Codeforces competitions), o1 scored 83% correct solutions compared to GPT-4o's 13%.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * The training behind o1 is "fundamentally different" from its predecessors, including a completely new optimization algorithm and a new training dataset specifically tailored for it.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+
+LLM Claim: * o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance.
+Is this claim supported by the context according to bespoke-minicheck? No
+
+LLM Claim: * However, o1 is also less capable than GPT-4o in some areas, particularly with regard to factual knowledge about the world.
+Is this claim supported by the context according to bespoke-minicheck? Yes
+```
+
+We see that the third claim "* o1 has been shown to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology, while GPT-4o does not have this level of performance." is not supported by the context. This is because the context only mentions that o1 "is claimed to perform" which is different from "has been shown to perform".

examples/python-grounded-factuality-rag-check/main.py

@@ -0,0 +1,137 @@
+import ollama
+import warnings
+from mattsollamatools import chunker
+from newspaper import Article
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+import nltk
+
+warnings.filterwarnings(
+    "ignore", category=FutureWarning, module="transformers.tokenization_utils_base"
+)
+nltk.download("punkt_tab", quiet=True)
+
+
+def getArticleText(url):
+    """Gets the text of an article from a URL.
+
+    Often there are a bunch of ads and menus on pages for a news article.
+    This uses newspaper3k to get just the text of just the article.
+    """
+    article = Article(url)
+    article.download()
+    article.parse()
+    return article.text
+
+
+def knn_search(question_embedding, embeddings, k=5):
+    """Performs K-nearest neighbors (KNN) search"""
+    X = np.array(
+        [item["embedding"] for article in embeddings for item in article["embeddings"]]
+    )
+    source_texts = [
+        item["source"] for article in embeddings for item in article["embeddings"]
+    ]
+
+    # Fit a KNN model on the embeddings
+    knn = NearestNeighbors(n_neighbors=k, metric="cosine")
+    knn.fit(X)
+
+    # Find the indices and distances of the k-nearest neighbors.
+    _, indices = knn.kneighbors(question_embedding, n_neighbors=k)
+
+    # Get the indices and source texts of the best matches
+    best_matches = [(indices[0][i], source_texts[indices[0][i]]) for i in range(k)]
+
+    return best_matches
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+if __name__ == "__main__":
+    allEmbeddings = []
+    default_url = "https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt"
+    user_input = input(
+        "Enter the URL of an article you want to chat with, or press Enter for default example: "
+    )
+    article_url = user_input.strip() if user_input.strip() else default_url
+    article = {}
+    article["embeddings"] = []
+    article["url"] = article_url
+    text = getArticleText(article_url)
+    chunks = chunker(text)
+
+    # Embed (batch) chunks using ollama
+    embeddings = ollama.embed(model="all-minilm", input=chunks)["embeddings"]
+
+    for chunk, embedding in zip(chunks, embeddings):
+        item = {}
+        item["source"] = chunk
+        item["embedding"] = embedding
+        item["sourcelength"] = len(chunk)
+        article["embeddings"].append(item)
+
+    allEmbeddings.append(article)
+
+    print(f"\nLoaded, chunked, and embedded text from {article_url}.\n")
+
+    while True:
+        # Input a question from the user
+        # For example, "Who is the chief research officer?"
+        question = input("Enter your question or type quit: ")
+
+        if question.lower() == "quit":
+            break
+
+        # Embed the user's question using ollama.embed
+        question_embedding = ollama.embed(model="all-minilm", input=question)[
+            "embeddings"
+        ]
+
+        # Perform KNN search to find the best matches (indices and source text)
+        best_matches = knn_search(question_embedding, allEmbeddings, k=4)
+
+        sourcetext = "\n\n".join([source_text for (_, source_text) in best_matches])
+
+        print(f"\nRetrieved chunks: \n{sourcetext}\n")
+
+        # Give the retreived chunks and question to the chat model
+        system_prompt = f"Only use the following information to answer the question. Do not use anything else: {sourcetext}"
+
+        ollama_response = ollama.generate(
+            model="llama3.2",
+            prompt=question,
+            system=system_prompt,
+            options={"stream": False},
+        )
+
+        answer = ollama_response["response"]
+        print(f"LLM Answer:\n{answer}\n")
+
+        # Check each sentence in the response for grounded factuality
+        if answer:
+            for claim in nltk.sent_tokenize(answer):
+                print(f"LLM Claim: {claim}")
+                print(
+                    f"Is this claim supported by the context according to bespoke-minicheck? {check(sourcetext, claim)}\n"
+                )

examples/python-grounded-factuality-rag-check/requirements.txt

@@ -0,0 +1,8 @@
+ollama
+lxml==5.3.0
+lxml_html_clean==0.2.2
+mattsollamatools==0.0.25
+newspaper3k==0.2.8
+nltk==3.9.1
+numpy==1.26.4
+scikit-learn==1.5.2

examples/python-grounded-factuality-simple-check/main.py

@@ -0,0 +1,53 @@
+"""Simple example to demonstrate how to use the bespoke-minicheck model."""
+
+import ollama
+
+# NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+def get_user_input(prompt):
+    user_input = input(prompt)
+    if not user_input:
+        exit()
+    print()
+    return user_input
+
+
+def main():
+    while True:
+        # Get a document from the user (e.g. "Ryan likes running and biking.")
+        document = get_user_input("Enter a document: ")
+        # Get a claim from the user (e.g. "Ryan likes to run.")
+        claim = get_user_input("Enter a claim: ")
+        # Check if the claim is supported by the document
+        grounded_factuality_check = check(document, claim)
+        print(
+            f"Is the claim supported by the document according to bespoke-minicheck? {grounded_factuality_check}"
+        )
+        print("\n\n")
+
+
+if __name__ == "__main__":
+    main()

examples/python-grounded-factuality-simple-check/readme.md

@@ -0,0 +1,54 @@
+# Simple Bespoke-Minicheck Example
+
+`bespoke-minicheck` is a model for checking if a claim is supported by a document. It is used through the **generate** endpoint, which is called in this example with a `prompt` that includes the expected formatting of the user input. 
+
+## Running the Example
+
+1. Ensure you have the `bespoke-minicheck` model installed:
+
+   ```bash
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies:
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the program:
+
+   ```bash
+   python main.py
+   ```
+
+4. Enter a document and a claim when prompted:
+
+   ```bash
+   Enter a document: Roses are red.
+
+   Enter a claim: Roses are blue. 
+   ```
+
+   The claim and document are then given to the `bespoke-minicheck` as inputs, which then generates a response (Yes or No) on whether the claim is supported by the document.
+
+   ```bash
+   Is the claim supported by the document according to bespoke-minicheck? No
+   ```
+
+## More Examples
+
+Document ([source](https://en.wikipedia.org/wiki/Apple_I)): 
+> The Apple Computer 1 (Apple-1[a]), later known predominantly as the Apple I(written with a Roman numeral),[b] is an 8-bit motherboard-only personal computer designed by Steve Wozniak[5][6] and released by the Apple Computer Company (now Apple Inc.) in 1976. The company was initially formed to sell the Apple I – its first product – and would later become the world's largest technology company.[7] The idea of starting a company and selling the computer came from Wozniak's friend and Apple co-founder Steve Jobs.[8][9] One of the main innovations of the Apple I was that it included video display terminal circuitry on its circuit board, allowing it to connect to a low-cost composite video monitor or television, instead of an expensive computer terminal, compared to most existing computers at the time.
+
+Claim: 
+>The Apple I is a 16-bit computer.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **No**
+
+Claim: 
+>Apple was originally called the Apple Computer Company.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **Yes**

examples/python-grounded-factuality-simple-check/requirements.txt

@@ -0,0 +1 @@
+ollama

examples/python-json-datagenerator/predefinedschema.py

@@ -2,7 +2,7 @@ import requests
 import json
 import random
 
-model = "llama3.1"
+model = "llama3.2"
 template = {
   "firstName": "",
   "lastName": "",

examples/python-json-datagenerator/randomaddresses.py

@@ -12,7 +12,7 @@ countries = [
     "France",
 ]
 country = random.choice(countries)
-model = "llama3.1"
+model = "llama3.2"
 
 prompt = f"generate one realistically believable sample data set of a persons first name, last name, address in {country}, and phone number. Do not use common names. Respond using JSON. Key names should have no backslashes, values should use plain ascii with no special characters."
 

examples/python-json-datagenerator/readme.md

@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/python-simplechat/client.py

@@ -2,7 +2,7 @@ import json
 import requests
 
 # NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
-model = "llama3.1"  # TODO: update this for whatever model you wish to use
+model = "llama3.2"  # TODO: update this for whatever model you wish to use
 
 
 def chat(messages):

examples/python-simplechat/readme.md

@@ -4,10 +4,10 @@ The **chat** endpoint is one of two ways to generate text from an LLM with Ollam
 
 ## Running the Example
 
-1. Ensure you have the `llama3.1` model installed:
+1. Ensure you have the `llama3.2` model installed:
 
    ```bash
-   ollama pull llama3.1
+   ollama pull llama3.2
    ```
 
 2. Install the Python Requirements.

examples/typescript-simplechat/client.ts

@@ -1,6 +1,6 @@
 import * as readline from "readline";
 
-const model = "llama3.1";
+const model = "llama3.2";
 type Message = {
   role: "assistant" | "user" | "system";
   content: string;

fileutils/README.md

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

fileutils/filetype.go

@@ -1,9 +1,11 @@
-package llm
+package fileutils
 
 import "fmt"
 
 type fileType uint32
 
+// TODO this should map over to the GGML CGO enum type
+
 const (
 	fileTypeF32 fileType = iota
 	fileTypeF16

fileutils/ggla.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"encoding/binary"
@@ -51,8 +51,8 @@ func (llm *ggla) KV() KV {
 	return llm.kv
 }
 
-func (llm *ggla) Tensors() Tensors {
-	return Tensors{
+func (llm *ggla) Tensors() *Tensors {
+	return &Tensors{
 		Items:  llm.tensors,
 		Offset: llm.tensorOffset,
 	}

fileutils/ggml.go

@@ -1,11 +1,14 @@
-package llm
+package fileutils
 
 import (
 	"encoding/binary"
 	"errors"
 	"fmt"
 	"io"
+	"os"
+	"slices"
 	"strings"
+	"sync"
 
 	"github.com/ollama/ollama/util/bufioutil"
 )
@@ -17,7 +20,7 @@ type GGML struct {
 
 type model interface {
 	KV() KV
-	Tensors() Tensors
+	Tensors() *Tensors
 }
 
 type KV map[string]any
@@ -123,25 +126,34 @@ func (kv KV) ChatTemplate() string {
 type Tensors struct {
 	Items  []*Tensor
 	Offset uint64
+
+	layers     map[string]Layer
+	layersOnce sync.Once
 }
 
-func (ts Tensors) Layers() map[string]Layer {
-	layers := make(map[string]Layer)
-	for _, t := range ts.Items {
-		parts := strings.Split(t.Name, ".")
-		if parts[0] == "blk" {
-			// join first and second part, e.g. blk.%d
-			parts = append([]string{fmt.Sprintf("%s.%s", parts[0], parts[1])}, parts[2:]...)
+func (ts *Tensors) Layers() map[string]Layer {
+	ts.layersOnce.Do(func() {
+		ts.layers = make(map[string]Layer)
+		for _, t := range ts.Items {
+			parts := strings.Split(t.Name, ".")
+			if index := slices.IndexFunc(parts, func(s string) bool { return s == "blk" || s == "mm" }); index != -1 {
+				if len(parts) > index+2 {
+					// blk and mm should have a number after them, join it
+					parts = append(
+						[]string{strings.Join(parts[:index+2], ".")},
+						parts[index+2:]...)
+				}
+			}
+
+			if _, ok := ts.layers[parts[0]]; !ok {
+				ts.layers[parts[0]] = make(Layer)
+			}
+
+			ts.layers[parts[0]][strings.Join(parts[1:], ".")] = t
 		}
+	})
 
-		if _, ok := layers[parts[0]]; !ok {
-			layers[parts[0]] = make(Layer)
-		}
-
-		layers[parts[0]][strings.Join(parts[1:], ".")] = t
-	}
-
-	return layers
+	return ts.layers
 }
 
 type Layer map[string]*Tensor
@@ -244,6 +256,8 @@ func (t Tensor) typeSize() uint64 {
 		return 8
 	case 29: // IQ1_M
 		return blockSize/8 + blockSize/16 + blockSize/32
+	case 30: // BF16
+		return 2
 	default:
 		return 0
 	}
@@ -475,3 +489,23 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 
 	return
 }
+
+// LoadModel will load a model from disk. The model must be in the GGML format.
+//
+// It collects array values for arrays with a size less than or equal to
+// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
+// the maxArraySize is negative, all arrays are collected.
+func LoadModel(model string, maxArraySize int) (*GGML, error) {
+	if _, err := os.Stat(model); err != nil {
+		return nil, err
+	}
+
+	f, err := os.Open(model)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	ggml, _, err := DecodeGGML(f, maxArraySize)
+	return ggml, err
+}

fileutils/ggml_test.go

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

fileutils/gguf.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"bytes"
@@ -110,8 +110,8 @@ func (llm *gguf) KV() KV {
 	return llm.kv
 }
 
-func (llm *gguf) Tensors() Tensors {
-	return Tensors{
+func (llm *gguf) Tensors() *Tensors {
+	return &Tensors{
 		Items:  llm.tensors,
 		Offset: llm.tensorOffset,
 	}

fileutils/memory.go

@@ -1,19 +1,20 @@
-package llm
+package fileutils
 
 import (
 	"fmt"
 	"log/slog"
+	"os"
 	"strconv"
 	"strings"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/discover"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
-	"github.com/ollama/ollama/gpu"
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus gpu.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
@@ -63,11 +64,13 @@ type MemoryEstimate struct {
 	memoryLayerOutput   uint64
 	graphFullOffload    uint64
 	graphPartialOffload uint64
+
+	projectorWeights, projectorGraph uint64
 }
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts api.Options) MemoryEstimate {
+func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string, opts api.Options) MemoryEstimate {
 	// Graph size for a partial offload, applies to all GPUs
 	var graphPartialOffload uint64
 
@@ -78,7 +81,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	var graphOffload uint64
 
 	// Projectors loaded into GPU0 only
-	var projectorSize uint64
+	var projectorWeights uint64
+	var projectorGraph uint64
 
 	// Conditional output size on GPU 0
 	var memoryLayerOutput uint64
@@ -103,7 +107,9 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)
 
 	for _, projector := range projectors {
-		projectorSize += projectorMemoryRequirements(projector)
+		weight, graph := projectorMemoryRequirements(projector)
+		projectorWeights += weight
+		projectorGraph += graph
 
 		// multimodal models require at least 2048 context
 		opts.NumCtx = max(opts.NumCtx, 2048)
@@ -149,7 +155,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	}
 
 	// Output layer handled at the end if we have space
-	gpuZeroOverhead := projectorSize
+	gpuZeroOverhead := projectorWeights + projectorGraph
 
 	// Reduce set of GPUs to only those that have sufficient space to fit overhead and at least one layer
 	var layerCount int
@@ -157,7 +163,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	gpuAllocations := make([]uint64, len(gpus))
 	type gs struct {
 		i int
-		g *gpu.GpuInfo
+		g *discover.GpuInfo
 	}
 	gpusWithSpace := []gs{}
 	for i := range gpus {
@@ -303,6 +309,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		memoryLayerOutput:   memoryLayerOutput,
 		graphFullOffload:    graphFullOffload,
 		graphPartialOffload: graphPartialOffload,
+		projectorWeights:    projectorWeights,
+		projectorGraph:      projectorGraph,
 	}
 
 	if gpus[0].Library == "cpu" {
@@ -321,9 +329,21 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	return estimate
 }
 
-func (m MemoryEstimate) log() {
+func (m MemoryEstimate) Log() {
 	overhead := envconfig.GpuOverhead()
-	slog.Info(
+
+	log := slog.With()
+	if m.projectorWeights > 0 {
+		log = log.With(
+			slog.Group(
+				"projector",
+				"weights", format.HumanBytes2(m.projectorWeights),
+				"graph", format.HumanBytes2(m.projectorGraph),
+			),
+		)
+	}
+
+	log.Info(
 		"offload to "+m.inferenceLibrary,
 		slog.Group(
 			"layers",
@@ -371,3 +391,52 @@ func (m MemoryEstimate) log() {
 		),
 	)
 }
+
+func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
+	file, err := os.Open(filename)
+	if err != nil {
+		return 0, 0
+	}
+	defer file.Close()
+
+	ggml, _, err := DecodeGGML(file, 0)
+	if err != nil {
+		return 0, 0
+	}
+
+	for _, layer := range ggml.Tensors().Layers() {
+		weights += layer.size()
+	}
+
+	switch arch := ggml.KV().Architecture(); arch {
+	case "mllama":
+		kv := func(n string) uint64 {
+			if v, ok := ggml.KV()[arch+".vision."+n].(uint32); ok {
+				return uint64(v)
+			}
+
+			return 0
+		}
+
+		imageSize := kv("image_size")
+
+		maxNumTiles := kv("max_num_tiles")
+		embeddingLength := kv("embedding_length")
+		headCount := kv("attention.head_count")
+
+		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
+		if _, ok := ggml.Tensors().Layers()["v"]["class_embd"]; ok {
+			numPatches++
+		}
+
+		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
+
+		graphSize = 4 * (8 +
+			imageSize*imageSize*kv("num_channels")*maxNumTiles +
+			embeddingLength*numPatches*maxNumTiles +
+			9*embeddingLength*numPaddedPatches*maxNumTiles +
+			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
+	}
+
+	return weights, graphSize
+}

fileutils/memory_test.go

@@ -1,4 +1,4 @@
-package llm
+package fileutils
 
 import (
 	"bytes"
@@ -10,7 +10,7 @@ import (
 	"github.com/stretchr/testify/require"
 
 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/discover"
 )
 
 func TestEstimateGPULayers(t *testing.T) {
@@ -50,7 +50,7 @@ func TestEstimateGPULayers(t *testing.T) {
 	}
 
 	// Simple CPU scenario
-	gpus := []gpu.GpuInfo{
+	gpus := []discover.GpuInfo{
 		{
 			Library: "cpu",
 		},
@@ -72,7 +72,7 @@ func TestEstimateGPULayers(t *testing.T) {
 
 	// Dual CUDA scenario with assymetry
 	gpuMinimumMemory := uint64(2048)
-	gpus = []gpu.GpuInfo{
+	gpus = []discover.GpuInfo{
 		{
 			Library:       "cuda",
 			MinimumMemory: gpuMinimumMemory,

go.mod

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

go.sum

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

gpu/gpu_linux.go

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

gpu/gpu_windows.go

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

integration/basic_test.go

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

integration/concurrency_test.go

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

integration/utils_test.go

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

llama/.gitignore

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

llama/Makefile

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