Remove llama.cpp submodule and shift new build to top

Move Go code out of llm package
image processing for llama3.2 (#6963 )
2024-10-18 16:39:22 -07:00 · 2024-10-18 16:38:59 -07:00 · 2024-10-18 16:12:35 -07:00 · 2024-10-17 15:03:09 -07:00 · 2024-10-17 13:43:20 -07:00 · 2024-10-17 11:59:52 -07:00
483 changed files with 154269 additions and 17450 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -3,7 +3,7 @@ ollama
 app
 macapp
 dist
-llm/llama.cpp
 .env
 .cache
 test_data
+llama/build
--- a/.gitattributes
+++ b/.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
--- a/.github/workflows/release.yaml
+++ b/.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
@@ -92,18 +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: |
-            llm/build/**/bin/*
-            llm/build/**/*.a
+            build/**/*
+            build/**/*.a
            dist/windows-amd64/**

  # ROCm generation step
@@ -157,14 +158,15 @@ 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
+          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
+          make -j $cores
+        name: make
      - name: 'gather rocm dependencies'
        run: |
          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
@@ -176,7 +178,7 @@ jobs:
        with:
          name: generate-windows-rocm
          path: |
-            llm/build/**/bin/*
+            build/**/*
            dist/windows-amd64/**
      - uses: actions/upload-artifact@v4
        with:
@@ -244,16 +246,17 @@ 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
      - name: 'gather cuda dependencies'
        run: |
          $NVIDIA_DIR=(resolve-path 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*\bin\')[0]
@@ -265,7 +268,7 @@ jobs:
        with:
          name: generate-windows-cuda-${{ matrix.cuda.version }}
          path: |
-            llm/build/**/bin/*
+            build/**/*
            dist/windows-amd64/**
      - uses: actions/upload-artifact@v4
        with:
@@ -273,7 +276,157 @@ jobs:
          path: dist/deps/*


-  # Import the prior generation steps and build the final windows assets
+  # windows arm64 generate, go build, and zip file (no installer)
+  # Output of this build is aggregated into the final x86 build
+  # for a unified windows installer
+  windows-arm64:
+    runs-on: windows-arm64
+    environment: release
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      # The current Windows arm64 beta image has effectively zero dev tools installed...
+      - name: Install git and gzip
+        run: |
+          Set-ExecutionPolicy Bypass -Scope Process -Force
+          [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072
+          iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))
+          choco install -y --no-progress git gzip
+          echo "C:\Program Files\Git\cmd" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\ProgramData\chocolatey\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      # 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-arm64
+          path: |
+            dist/windows-arm64/**
+            dist/windows-arm64-app.exe
+            dist/ollama-windows-arm64.zip
+
+  # Import the prior generation steps plus the full arm64 build, and build the final windows assets
  build-windows:
    environment: release
    runs-on: windows
@@ -281,6 +434,7 @@ jobs:
      - generate-windows-cuda
      - generate-windows-rocm
      - generate-windows-cpu
+      - windows-arm64
    env:
      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
    steps:
@@ -338,11 +492,15 @@ jobs:
      - uses: actions/download-artifact@v4
        with:
          name: generate-windows-rocm
-      - run: dir llm/build
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-arm64
+          path: dist
+      - run: dir build
      - run: |
          $gopath=(get-command go).source | split-path -parent
-          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
-          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"
          $env:OLLAMA_SKIP_GENERATE="1"
@@ -359,9 +517,7 @@ jobs:
    environment: release
    runs-on: linux
    env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
-      BUILD_ARCH: amd64
-      PUSH: '1'
+      PLATFORM: linux/amd64
    steps:
      - uses: actions/checkout@v4
        with:
@@ -369,14 +525,8 @@ jobs:
      - name: Set Version
        shell: bash
        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
      - run: |
          ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
      - uses: actions/upload-artifact@v4
        with:
          name: dist-linux-amd64
@@ -390,9 +540,7 @@ jobs:
    environment: release
    runs-on: linux-arm64
    env:
-      OLLAMA_SKIP_MANIFEST_CREATE: '1'
-      BUILD_ARCH: arm64
-      PUSH: '1'
+      PLATFORM: linux/arm64
    steps:
      - uses: actions/checkout@v4
        with:
@@ -421,14 +569,8 @@ jobs:
          sudo usermod -aG docker $USER
          sudo apt-get install acl
          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
      - run: |
          ./scripts/build_linux.sh
-          ./scripts/build_docker.sh
      - uses: actions/upload-artifact@v4
        with:
          name: dist-linux-arm64
@@ -436,6 +578,178 @@ jobs:
            dist/*linux*
            !dist/*-cov

+  # Container image build
+  build-container-image:
+    environment: release
+    strategy:
+      matrix:
+        runner:
+          - linux
+          - linux-arm64
+    runs-on: ${{ matrix.runner }}
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: 'Install Docker'
+        if: ${{ startsWith(matrix.runner, 'linux-arm64') }}
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y ca-certificates curl
+          sudo install -m 0755 -d /etc/apt/keyrings
+          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
+          sudo chmod a+r /etc/apt/keyrings/docker.asc
+          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
+            $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
+            sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
+          sudo apt-get update
+          sudo apt-get install -y docker-ce docker-ce-cli containerd.io
+          sudo usermod -aG docker $USER
+          sudo apt-get install acl
+          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          platforms: linux/${{ env.ARCH }}
+          build-args: |
+            GOFLAGS
+          outputs: type=image,name=${{ env.FINAL_IMAGE_REPO }},push-by-digest=true,name-canonical=true,push=true
+      - name: Export digest
+        run: |
+          mkdir -p /tmp/digests
+          digest="${{ steps.build.outputs.digest }}"
+          touch "/tmp/digests/${digest#sha256:}"
+      - name: Upload digest
+        uses: actions/upload-artifact@v4
+        with:
+          name: digests-${{ env.PLATFORM_PAIR }}
+          path: /tmp/digests/*
+          if-no-files-found: error
+          retention-days: 1
+  merge:
+    environment: release
+    runs-on: linux
+    needs:
+      - build-container-image
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Download digests
+        uses: actions/download-artifact@v4
+        with:
+          path: /tmp/digests
+          pattern: digests-*
+          merge-multiple: true
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          machine=$(uname -m)
+          case ${machine} in
+            x86_64) echo ARCH=amd64; echo PLATFORM_PAIR=linux-amd64 ;;
+            aarch64) echo ARCH=arm64; echo PLATFORM_PAIR=linux-arm64 ;;
+          esac >>$GITHUB_ENV
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Create manifest list and push
+        working-directory: /tmp/digests
+        run: |
+          docker buildx imagetools create $(jq -cr '.tags | map("-t " + .) | join(" ")' <<< "$DOCKER_METADATA_OUTPUT_JSON") \
+            $(printf '${{ env.FINAL_IMAGE_REPO }}@sha256:%s ' *)
+      - name: Inspect image
+        run: |
+          docker buildx imagetools inspect ${{ env.FINAL_IMAGE_REPO }}:${{ steps.meta.outputs.version }}          
+  build-container-image-rocm:
+    environment: release
+    runs-on: linux
+    env:
+      FINAL_IMAGE_REPO: ollama/ollama
+      ARCH: amd64
+      PLATFORM_PAIR: linux-amd64
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.FINAL_IMAGE_REPO }}
+          flavor: |
+            latest=false
+          tags: |
+            type=ref,enable=true,priority=600,prefix=0.0.0-pr,suffix=,event=pr
+            type=semver,pattern={{version}}
+      - name: Set Version
+        shell: bash
+        run: |
+          echo GOFLAGS="'-ldflags=-w -s \"-X=github.com/ollama/ollama/version.Version=${{ env.DOCKER_METADATA_OUTPUT_VERSION }}\" \"-X=github.com/ollama/ollama/server.mode=release\"'" >>$GITHUB_ENV
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Build and push by digest
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: "."
+          target: runtime-rocm
+          build-args: |
+            GOFLAGS
+          tags: ${{ env.FINAL_IMAGE_REPO }}:${{ env.DOCKER_METADATA_OUTPUT_VERSION}}-rocm
+          push: true
+
  # Aggregate all the assets and ship a release
  release:
    needs:
@@ -448,8 +762,6 @@ jobs:
    permissions:
      contents: write
    env:
-      OLLAMA_SKIP_IMAGE_BUILD: '1'
-      PUSH: '1'
      GH_TOKEN: ${{ github.token }}
    steps:
      - uses: actions/checkout@v4
@@ -458,12 +770,6 @@ jobs:
        run: |
          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ vars.DOCKER_USER }}
-          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
-      - run: ./scripts/build_docker.sh
      - name: Retrieve built artifact
        uses: actions/download-artifact@v4
        with:
@@ -474,8 +780,6 @@ jobs:
          ls -lh dist/
          (cd dist; find . -type f | xargs sha256sum > ../sha256sum.txt)
          mv sha256sum.txt dist/
-          mv dist/linux-???64 .
-          mv dist/linux-amd64-rocm .
          cat dist/sha256sum.txt
      - name: Create or update Release
        run: |
--- a/.github/workflows/test.yaml
+++ b/.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:
@@ -38,58 +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 .
-      - uses: actions/upload-artifact@v4
-        with:
-          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: |
-            llm/build/**/bin/*
-            llm/build/**/*.a
-  generate-cuda:
-    needs: [changes]
-    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
    strategy:
      matrix:
        cuda-version:
@@ -99,8 +51,6 @@ jobs:
    steps:
      - run: |
          apt-get update && apt-get install -y git build-essential curl
-          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
-            | tar -zx -C /usr --strip-components 1
        env:
          DEBIAN_FRONTEND: noninteractive
      - uses: actions/checkout@v4
@@ -111,18 +61,11 @@ jobs:
      - run: go get ./...
      - run: |
          git config --global --add safe.directory /__w/ollama/ollama
-          go generate -x ./...
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: cuda-${{ matrix.cuda-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**
-  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: ${{ needs.changes.outputs.RUNNERS == 'True' }}
    strategy:
      matrix:
        rocm-version:
@@ -132,8 +75,6 @@ jobs:
    steps:
      - run: |
          apt-get update && apt-get install -y git build-essential curl rocm-libs
-          curl -fsSL https://github.com/Kitware/CMake/releases/download/v3.28.1/cmake-3.28.1-linux-x86_64.tar.gz \
-            | tar -zx -C /usr --strip-components 1
        env:
          DEBIAN_FRONTEND: noninteractive
      - uses: actions/checkout@v4
@@ -144,20 +85,13 @@ jobs:
      - run: go get ./...
      - run: |
          git config --global --add safe.directory /__w/ollama/ollama
-          go generate -x ./...
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
-      - uses: actions/upload-artifact@v4
-        with:
-          name: rocm-${{ matrix.rocm-version }}-libraries
-          path: |
-            llm/build/**/bin/*
-            dist/windows-amd64/**
+          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
@@ -179,22 +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'
-      # TODO - do we need any artifacts?
+          $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
@@ -219,19 +152,59 @@ 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'
-      # TODO - do we need any artifacts?
+
+  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:
@@ -263,14 +236,6 @@ jobs:
            arm64) echo ARCH=arm64 ;;
          esac >>$GITHUB_ENV
        shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
      - uses: golangci/golangci-lint-action@v6
        with:
          args: --timeout 8m0s -v
@@ -288,9 +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:
@@ -301,23 +263,21 @@ jobs:
          cache: true
      - run: |
          case ${{ matrix.arch }} in
-            amd64) echo ARCH=x86_64 ;;
+            amd64) echo ARCH=amd64 ;;
            arm64) echo ARCH=arm64 ;;
          esac >>$GITHUB_ENV
        shell: bash
-      - run: |
-          mkdir -p llm/build/linux/$ARCH/stub/bin
-          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
-      - run: |
-          mkdir -p llm/build/darwin/$ARCH/stub/bin
-          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
-        if: ${{ startsWith(matrix.os, 'macos-') }}
-        shell: bash
-      - run: go generate ./...
      - run: go build
      - run: go test -v ./...
-      - uses: actions/upload-artifact@v4
+
+  patches:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.RUNNERS == 'True' }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
        with:
-          name: ${{ matrix.os }}-binaries
-          path: ollama
+          submodules: recursive
+      - name: Verify patches carry all the changes
+        run: |
+          make apply-patches sync && git diff --compact-summary --exit-code llama
--- a/.gitignore
+++ b/.gitignore
@@ -5,11 +5,14 @@
 .swp
 dist
 ollama
-ggml-metal.metal
 .cache
 *.exe
 .idea
 test_data
 *.crt
 llm/build
-__debug_bin*
+build/*/*/*
+!build/**/placeholder
+llama/build
+__debug_bin*
+llama/vendor
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,4 +0,0 @@
-[submodule "llama.cpp"]
-	path = llm/llama.cpp
-	url = https://github.com/ggerganov/llama.cpp.git
-	shallow = true
--- a/.golangci.yaml
+++ b/.golangci.yaml
@@ -32,6 +32,10 @@ linters:
 linters-settings:
  gci:
    sections: [standard, default, localmodule]
+  staticcheck:
+    checks:
+      - all
+      - -SA1019 # omit Deprecated check
 severity:
  default-severity: error
  rules:
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -18,7 +18,7 @@ See the [development documentation](./docs/development.md) for instructions on h

 * New features: new features (e.g. API fields, environment variables) add surface area to Ollama and make it harder to maintain in the long run as they cannot be removed without potentially breaking users in the future.
 * Refactoring: large code improvements are important, but can be harder or take longer to review and merge.
-* Documentation: small updates to fill in or dorrect missing documentation is helpful, however large documentation additions can be hard to maintain over time.
+* Documentation: small updates to fill in or correct missing documentation is helpful, however large documentation additions can be hard to maintain over time.

 ### Issues that may not be accepted

--- a/346
+++ b/346
@@ -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
@@ -6,199 +7,202 @@ 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:$CUDA_VERSION_11-devel-rockylinux8 AS cuda-11-build-server-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${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:$CUDA_VERSION_12-devel-rockylinux8 AS cuda-12-build-server-arm64
-ARG CMAKE_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${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.new --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:$PATH
-COPY --from=llm-code / /go/src/github.com/ollama/ollama/
-ARG OLLAMA_CUSTOM_CPU_DEFS
-ARG CGO_CFLAGS
-ENV GOARCH amd64 
-WORKDIR /go/src/github.com/ollama/ollama/llm/generate
-
-FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" bash gen_linux.sh
-FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" bash gen_linux.sh
-
-FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64
-ARG CMAKE_VERSION
-ARG GOLANG_VERSION
-COPY ./scripts/rh_linux_deps.sh /
-RUN CMAKE_VERSION=${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
-
-FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_CPU_TARGET="static" bash gen_linux.sh
-FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
-RUN --mount=type=cache,target=/root/.ccache \
-    OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" bash gen_linux.sh
-
-
-# Intermediate stage used for ./scripts/build_linux.sh
-FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
+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" ]
+
+### 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.new --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=${GOLANG_VERSION} sh /rh_linux_deps.sh
+RUN yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo && \
+    dnf config-manager --set-enabled appstream && \
+    dnf clean all && \
+    dnf install -y \
+    zsh \
+    cuda-toolkit-$(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/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 \
+    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
+        make -C llama -j $(expr $(nproc) / 2 ) ; \
+    else \
+        make -C llama -j 5 ; \
+    fi
+
+FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
+COPY . .
+ARG OLLAMA_SKIP_CUDA_GENERATE
+ARG OLLAMA_SKIP_CUDA_11_GENERATE
+ARG OLLAMA_SKIP_CUDA_12_GENERATE
+ARG CUDA_V11_ARCHITECTURES
+ARG CUDA_V12_ARCHITECTURES
+ARG OLLAMA_FAST_BUILD
+RUN --mount=type=cache,target=/root/.ccache \
+    make -C llama -j 8
+
+
+# Intermediate stages used for ./scripts/build_linux.sh
+FROM --platform=linux/amd64 centos:7 AS builder-amd64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
+ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
+ENV CGO_ENABLED 1
+ENV GOARCH amd64
+WORKDIR /go/src/github.com/ollama/ollama
+
+FROM --platform=linux/amd64 builder-amd64 AS build-amd64
+COPY . .
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-amd64 /go/src/github.com/ollama/ollama/build/ build/
+ARG GOFLAGS
+ARG CGO_CFLAGS
+ARG OLLAMA_SKIP_ROCM_GENERATE
+RUN --mount=type=cache,target=/root/.ccache \
+    go build -trimpath -o dist/linux-amd64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN if [ -z ${OLLAMA_SKIP_ROCM_GENERATE} ] ; then \
+    cd dist/linux-$GOARCH-rocm && \
+    tar -cf - . | pigz --best > ../ollama-linux-$GOARCH-rocm.tgz ;\
+    fi
+
+FROM --platform=linux/arm64 rockylinux:8 AS builder-arm64
+ARG CMAKE_VERSION
+ARG GOLANG_VERSION
+COPY ./scripts/rh_linux_deps.sh /
+RUN CMAKE_VERSION=${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=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 \
+    go build -trimpath -o dist/linux-arm64/bin/ollama .
+RUN cd dist/linux-$GOARCH && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+
+FROM --platform=linux/amd64 scratch AS dist-amd64
+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
+
+
+# Optimized container images do not cary nested payloads
+FROM --platform=linux/amd64 builder-amd64 AS container-build-amd64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
    go build -trimpath -o dist/linux-amd64/bin/ollama .

-# Intermediate stage used for ./scripts/build_linux.sh
-FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
-ENV CGO_ENABLED 1
-ARG GOLANG_VERSION
+FROM --platform=linux/arm64 builder-arm64 AS container-build-arm64
 WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-11-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
-COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
-COPY --from=cuda-12-build-server-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
    go build -trimpath -o dist/linux-arm64/bin/ollama .

-# Strip out ROCm dependencies to keep the primary image lean
-FROM --platform=linux/amd64 ubuntu:22.04 as amd64-libs-without-rocm
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /scratch/
-RUN cd /scratch/ollama/ && rm -rf rocblas libamd* libdrm* libroc* libhip* libhsa* 
+# 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*

-# Runtime stages
-FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
-COPY --from=amd64-libs-without-rocm /scratch/ /lib/
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+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/arm64 ubuntu:22.04 as runtime-arm64
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
-RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/
+
+FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
+COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+
+# ROCm libraries larger so we keep it distinct from the CPU/CUDA image
+FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
+# Frontload the rocm libraries which are large, and rarely change to increase chance of a common layer
+# across releases
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
+RUN apt-get update && \
+    apt-get install -y ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
+COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/

-# Radeon images are much larger so we keep it distinct from the CPU/CUDA image
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete as runtime-rocm
-RUN update-pciids
-COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
-RUN ln -s /opt/rocm/lib /lib/ollama
 EXPOSE 11434
 ENV OLLAMA_HOST 0.0.0.0

--- a/4
+++ b/4
@@ -0,0 +1,4 @@
+GOALS := $(or $(MAKECMDGOALS),all)
+.PHONY: $(GOALS)
+$(GOALS):
+	$(MAKE) -C llama $@
--- a/README.md
+++ b/README.md
@@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla

 ## Quickstart

-To run and chat with [Llama 3.1](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?" }
  ]
@@ -295,13 +309,28 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [AiLama](https://github.com/zeyoyt/ailama) (A Discord User App that allows you to interact with Ollama anywhere in discord )
 - [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
+- [Painting Droid](https://github.com/mateuszmigas/painting-droid) (Painting app with AI integrations)
 - [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
 - [AI Studio](https://github.com/MindWorkAI/AI-Studio)
 - [Sidellama](https://github.com/gyopak/sidellama) (browser-based LLM client)
 - [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
 - [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
 - [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
+- [Go-CREW](https://www.jonathanhecl.com/go-crew/) (Powerful Offline RAG in Golang)
+- [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
+- [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
+- [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
+- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
+- [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
+- [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
+- [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
+- [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)
+- [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

@@ -326,6 +355,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 - [gollama](https://github.com/sammcj/gollama)
 - [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
+- [Ollama Mixture of Experts (MOE) in 50 lines of code](https://github.com/rapidarchitect/ollama_moe)
+- [vim-intelligence-bridge](https://github.com/pepo-ec/vim-intelligence-bridge) Simple interaction of "Ollama" with the Vim editor
+
+### Apple Vision Pro
+- [Enchanted](https://github.com/AugustDev/enchanted)

 ### Database

@@ -335,23 +369,28 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Package managers

 - [Pacman](https://archlinux.org/packages/extra/x86_64/ollama/)
+- [Gentoo](https://github.com/gentoo/guru/tree/master/app-misc/ollama)
 - [Helm Chart](https://artifacthub.io/packages/helm/ollama-helm/ollama)
 - [Guix channel](https://codeberg.org/tusharhero/ollama-guix)
+- [Nix package](https://search.nixos.org/packages?channel=24.05&show=ollama&from=0&size=50&sort=relevance&type=packages&query=ollama)
+- [Flox](https://flox.dev/blog/ollama-part-one)

 ### Libraries

- [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/modules/model_io/models/llms/integrations/ollama) with [example](https://js.langchain.com/docs/use_cases/question_answering/local_retrieval_qa)
+- [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)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
 - [Ollama-rs for Rust](https://github.com/pepperoni21/ollama-rs)
 - [Ollama-hpp for C++](https://github.com/jmont-dev/ollama-hpp)
- [Ollama4j for Java](https://github.com/amithkoujalgi/ollama4j)
+- [Ollama4j for Java](https://github.com/ollama4j/ollama4j)
 - [ModelFusion Typescript Library](https://modelfusion.dev/integration/model-provider/ollama)
 - [OllamaKit for Swift](https://github.com/kevinhermawan/OllamaKit)
 - [Ollama for Dart](https://github.com/breitburg/dart-ollama)
@@ -368,11 +407,18 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
 - [LlamaScript](https://github.com/Project-Llama/llamascript)
+- [Gollm](https://docs.gollm.co/examples/ollama-example)
+- [Ollamaclient for Golang](https://github.com/xyproto/ollamaclient)
+- [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
+- [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
+- [Agents-Flex for Java](https://github.com/agents-flex/agents-flex) with [example](https://github.com/agents-flex/agents-flex/tree/main/agents-flex-llm/agents-flex-llm-ollama/src/test/java/com/agentsflex/llm/ollama)

 ### Mobile

 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
+- [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

@@ -397,11 +443,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and Hugging Face)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
+- [Plasmoid Ollama Control](https://github.com/imoize/plasmoid-ollamacontrol) (KDE Plasma extension that allows you to quickly manage/control Ollama model)
 - [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
 - [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
 - [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
 - [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 - [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
+- [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
+- [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
+- [QodeAssist](https://github.com/Palm1r/QodeAssist) (AI-powered coding assistant plugin for Qt Creator)
+- [Obsidian Quiz Generator plugin](https://github.com/ECuiDev/obsidian-quiz-generator)

 ### Supported backends

--- a/api/types.go
+++ b/api/types.go
@@ -296,15 +296,17 @@ type EmbeddingResponse struct {
 // CreateRequest is the request passed to [Client.Create].
 type CreateRequest struct {
 	Model     string `json:"model"`
-	Path      string `json:"path"`
 	Modelfile string `json:"modelfile"`
 	Stream    *bool  `json:"stream,omitempty"`
 	Quantize  string `json:"quantize,omitempty"`

-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`

-	// Quantization is deprecated, see Quantize
+	// Deprecated: set the file content with Modelfile instead
+	Path string `json:"path"`
+
+	// Deprecated: use Quantize instead
 	Quantization string `json:"quantization,omitempty"`
 }

@@ -312,7 +314,7 @@ type CreateRequest struct {
 type DeleteRequest struct {
 	Model string `json:"model"`

-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }

@@ -327,7 +329,7 @@ type ShowRequest struct {

 	Options map[string]interface{} `json:"options"`

-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }

@@ -359,7 +361,7 @@ type PullRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`

-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }

@@ -380,7 +382,7 @@ type PushRequest struct {
 	Password string `json:"password"`
 	Stream   *bool  `json:"stream,omitempty"`

-	// Name is deprecated, see Model
+	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
 }

--- a/app/ollama.iss
+++ b/app/ollama.iss
@@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64 arm64
-ArchitecturesInstallIn64BitMode=x64 arm64
+ArchitecturesAllowed=x64compatible arm64
+ArchitecturesInstallIn64BitMode=x64compatible arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@@ -48,6 +48,7 @@ OutputDir=..\dist\
 SetupLogging=yes
 CloseApplications=yes
 RestartApplications=no
+RestartIfNeededByRun=no

 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp
@@ -86,12 +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;
--- a/app/ollama_welcome.ps1
+++ b/app/ollama_welcome.ps1
@@ -4,5 +4,5 @@ write-host "Welcome to Ollama!"
 write-host ""
 write-host "Run your first model:"
 write-host ""
-write-host "`tollama run llama3.1"
+write-host "`tollama run llama3.2"
 write-host ""
--- a/build/darwin/amd64/placeholder
+++ b/build/darwin/amd64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/darwin/arm64/placeholder
+++ b/build/darwin/arm64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/embed_darwin_amd64.go
+++ b/build/embed_darwin_amd64.go
@@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/amd64/*
+var EmbedFS embed.FS
--- a/build/embed_darwin_arm64.go
+++ b/build/embed_darwin_arm64.go
@@ -0,0 +1,8 @@
+package build
+
+import "embed"
+
+// Darwin payloads separated by architecture to avoid duplicate payloads when cross compiling
+
+//go:embed darwin/arm64/*
+var EmbedFS embed.FS
--- a/build/embed_linux.go
+++ b/build/embed_linux.go
@@ -0,0 +1,6 @@
+package build
+
+import "embed"
+
+//go:embed linux/*
+var EmbedFS embed.FS
--- a/build/embed_unused.go
+++ b/build/embed_unused.go
@@ -0,0 +1,8 @@
+//go:build !linux && !darwin
+
+package build
+
+import "embed"
+
+// unused on windows
+var EmbedFS embed.FS
--- a/build/linux/amd64/placeholder
+++ b/build/linux/amd64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/build/linux/arm64/placeholder
+++ b/build/linux/arm64/placeholder
@@ -0,0 +1 @@
+This is here to make sure the build/ directory exists for the go:embed command
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -2,6 +2,7 @@ package cmd

 import (
 	"archive/zip"
+	"bufio"
 	"bytes"
 	"context"
 	"crypto/ed25519"
@@ -20,7 +21,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"runtime"
-	"slices"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -344,6 +345,39 @@ func (w *progressWriter) Write(p []byte) (n int, err error) {
 	return len(p), nil
 }

+func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
+	p := progress.NewProgress(os.Stderr)
+	defer p.StopAndClear()
+
+	spinner := progress.NewSpinner("")
+	p.Add("", spinner)
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	req := &api.GenerateRequest{
+		Model:     opts.Model,
+		KeepAlive: opts.KeepAlive,
+	}
+
+	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+}
+
+func StopHandler(cmd *cobra.Command, args []string) error {
+	opts := &runOptions{
+		Model:     args[0],
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := loadOrUnloadModel(cmd, opts); err != nil {
+		if strings.Contains(err.Error(), "not found") {
+			return fmt.Errorf("couldn't find model \"%s\" to stop", args[0])
+		}
+	}
+	return nil
+}
+
 func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true

@@ -418,11 +452,11 @@ 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 {
-		if err := loadModel(cmd, &opts); err != nil {
+		if err := loadOrUnloadModel(cmd, &opts); err != nil {
 			return err
 		}

@@ -578,7 +612,7 @@ func ListHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()

@@ -613,7 +647,15 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
 				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
 			}
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+
+			var until string
+			delta := time.Since(m.ExpiresAt)
+			if delta > 0 {
+				until = "Stopping..."
+			} else {
+				until = format.HumanTime(m.ExpiresAt, "Never")
+			}
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, until})
 		}
 	}

@@ -624,7 +666,7 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 	table.SetHeaderLine(false)
 	table.SetBorder(false)
 	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
+	table.SetTablePadding("    ")
 	table.AppendBulk(data)
 	table.Render()

@@ -637,6 +679,17 @@ 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 {
@@ -720,122 +773,89 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}

-	showInfo(resp)
-
-	return nil
+	return showInfo(resp, os.Stdout)
 }

-func showInfo(resp *api.ShowResponse) {
-	arch := resp.ModelInfo["general.architecture"].(string)
+func showInfo(resp *api.ShowResponse, w io.Writer) error {
+	tableRender := func(header string, rows func() [][]string) {
+		fmt.Fprintln(w, " ", header)
+		table := tablewriter.NewWriter(w)
+		table.SetAlignment(tablewriter.ALIGN_LEFT)
+		table.SetBorder(false)
+		table.SetNoWhiteSpace(true)
+		table.SetTablePadding("    ")

-	modelData := [][]string{
-		{"arch", arch},
-		{"parameters", resp.Details.ParameterSize},
-		{"quantization", resp.Details.QuantizationLevel},
-		{"context length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64))},
-		{"embedding length", fmt.Sprintf("%v", resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64))},
+		switch header {
+		case "Template", "System", "License":
+			table.SetColWidth(100)
+		}
+
+		table.AppendBulk(rows())
+		table.Render()
+		fmt.Fprintln(w)
 	}

-	mainTableData := [][]string{
-		{"Model"},
-		{renderSubTable(modelData, false)},
-	}
+	tableRender("Model", func() (rows [][]string) {
+		if resp.ModelInfo != nil {
+			arch := resp.ModelInfo["general.architecture"].(string)
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
+		} else {
+			rows = append(rows, []string{"", "architecture", resp.Details.Family})
+			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
+		}
+		rows = append(rows, []string{"", "quantization", resp.Details.QuantizationLevel})
+		return
+	})

 	if resp.ProjectorInfo != nil {
-		projectorData := [][]string{
-			{"arch", "clip"},
-			{"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))},
-		}
-
-		if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok {
-			projectorData = append(projectorData, []string{"projector type", projectorType.(string)})
-		}
-
-		projectorData = append(projectorData,
-			[]string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))},
-			[]string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))},
-		)
-
-		mainTableData = append(mainTableData,
-			[]string{"Projector"},
-			[]string{renderSubTable(projectorData, false)},
-		)
+		tableRender("Projector", func() (rows [][]string) {
+			arch := resp.ProjectorInfo["general.architecture"].(string)
+			rows = append(rows, []string{"", "architecture", arch})
+			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))})
+			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.embedding_length", arch)].(float64), 'f', -1, 64)})
+			rows = append(rows, []string{"", "dimensions", strconv.FormatFloat(resp.ProjectorInfo[fmt.Sprintf("%s.vision.projection_dim", arch)].(float64), 'f', -1, 64)})
+			return
+		})
 	}

 	if resp.Parameters != "" {
-		mainTableData = append(mainTableData, []string{"Parameters"}, []string{formatParams(resp.Parameters)})
+		tableRender("Parameters", func() (rows [][]string) {
+			scanner := bufio.NewScanner(strings.NewReader(resp.Parameters))
+			for scanner.Scan() {
+				if text := scanner.Text(); text != "" {
+					rows = append(rows, append([]string{""}, strings.Fields(text)...))
+				}
+			}
+			return
+		})
+	}
+
+	head := func(s string, n int) (rows [][]string) {
+		scanner := bufio.NewScanner(strings.NewReader(s))
+		for scanner.Scan() && (len(rows) < n || n < 0) {
+			if text := scanner.Text(); text != "" {
+				rows = append(rows, []string{"", strings.TrimSpace(text)})
+			}
+		}
+		return
 	}

 	if resp.System != "" {
-		mainTableData = append(mainTableData, []string{"System"}, []string{renderSubTable(twoLines(resp.System), true)})
+		tableRender("System", func() [][]string {
+			return head(resp.System, 2)
+		})
 	}

 	if resp.License != "" {
-		mainTableData = append(mainTableData, []string{"License"}, []string{renderSubTable(twoLines(resp.License), true)})
+		tableRender("License", func() [][]string {
+			return head(resp.License, 2)
+		})
 	}

-	table := tablewriter.NewWriter(os.Stdout)
-	table.SetAutoWrapText(false)
-	table.SetBorder(false)
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range mainTableData {
-		table.Append(v)
-	}
-
-	table.Render()
-}
-
-func renderSubTable(data [][]string, file bool) string {
-	var buf bytes.Buffer
-	table := tablewriter.NewWriter(&buf)
-	table.SetAutoWrapText(!file)
-	table.SetBorder(false)
-	table.SetNoWhiteSpace(true)
-	table.SetTablePadding("\t")
-	table.SetAlignment(tablewriter.ALIGN_LEFT)
-
-	for _, v := range data {
-		table.Append(v)
-	}
-
-	table.Render()
-
-	renderedTable := buf.String()
-	lines := strings.Split(renderedTable, "\n")
-	for i, line := range lines {
-		lines[i] = "\t" + line
-	}
-
-	return strings.Join(lines, "\n")
-}
-
-func twoLines(s string) [][]string {
-	lines := strings.Split(s, "\n")
-	res := [][]string{}
-
-	count := 0
-	for _, line := range lines {
-		line = strings.TrimSpace(line)
-		if line != "" {
-			count++
-			res = append(res, []string{line})
-			if count == 2 {
-				return res
-			}
-		}
-	}
-	return res
-}
-
-func formatParams(s string) string {
-	lines := strings.Split(s, "\n")
-	table := [][]string{}
-
-	for _, line := range lines {
-		table = append(table, strings.Fields(line))
-	}
-	return renderSubTable(table, false)
+	return nil
 }

 func CopyHandler(cmd *cobra.Command, args []string) error {
@@ -1325,6 +1345,15 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("insecure", false, "Use an insecure registry")
 	runCmd.Flags().Bool("nowordwrap", false, "Don't wrap words to the next line automatically")
 	runCmd.Flags().String("format", "", "Response format (e.g. json)")
+
+	stopCmd := &cobra.Command{
+		Use:     "stop MODEL",
+		Short:   "Stop a running model",
+		Args:    cobra.ExactArgs(1),
+		PreRunE: checkServerHeartbeat,
+		RunE:    StopHandler,
+	}
+
 	serveCmd := &cobra.Command{
 		Use:     "serve",
 		Aliases: []string{"start"},
@@ -1392,6 +1421,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
@@ -1418,6 +1448,8 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
+				envVars["OLLAMA_GPU_OVERHEAD"],
+				envVars["OLLAMA_LOAD_TIMEOUT"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)
@@ -1429,6 +1461,7 @@ func NewCLI() *cobra.Command {
 		createCmd,
 		showCmd,
 		runCmd,
+		stopCmd,
 		pullCmd,
 		pushCmd,
 		listCmd,
--- a/cmd/cmd_test.go
+++ b/cmd/cmd_test.go
@@ -0,0 +1,272 @@
+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"
+)
+
+func TestShowInfo(t *testing.T) {
+	t.Run("bare details", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+`
+
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("bare model info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			ModelInfo: map[string]any{
+				"general.architecture":    "test",
+				"general.parameter_count": float64(7_000_000_000),
+				"test.context_length":     float64(0),
+				"test.embedding_length":   float64(0),
+			},
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture        test    
+    parameters          7B      
+    context length      0       
+    embedding length    0       
+    quantization        FP16    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("parameters", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Parameters: `
+			stop never
+			stop gonna
+			stop give
+			stop you
+			stop up
+			temperature 99`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Parameters
+    stop           never    
+    stop           gonna    
+    stop           give     
+    stop           you      
+    stop           up       
+    temperature    99       
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("project info", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			ProjectorInfo: map[string]any{
+				"general.architecture":         "clip",
+				"general.parameter_count":      float64(133_700_000),
+				"clip.vision.embedding_length": float64(0),
+				"clip.vision.projection_dim":   float64(0),
+			},
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  Projector
+    architecture        clip       
+    parameters          133.70M    
+    embedding length    0          
+    dimensions          0          
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("system", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			System: `You are a pirate!
+Ahoy, matey!
+Weigh anchor!
+			`,
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  System
+    You are a pirate!    
+    Ahoy, matey!         
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("license", func(t *testing.T) {
+		var b bytes.Buffer
+		license, err := os.ReadFile(filepath.Join("..", "LICENSE"))
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			License: string(license),
+		}, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture    test    
+    parameters      7B      
+    quantization    FP16    
+
+  License
+    MIT License             
+    Copyright (c) Ollama    
+
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+}
+
+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)
+	}
+}
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@@ -18,7 +18,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/parser"
-	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
 )
@@ -31,26 +30,6 @@ const (
 	MultilineSystem
 )

-func loadModel(cmd *cobra.Command, opts *runOptions) error {
-	p := progress.NewProgress(os.Stderr)
-	defer p.StopAndClear()
-
-	spinner := progress.NewSpinner("")
-	p.Add("", spinner)
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
-	chatReq := &api.ChatRequest{
-		Model:     opts.Model,
-		KeepAlive: opts.KeepAlive,
-	}
-
-	return client.Chat(cmd.Context(), chatReq, func(api.ChatResponse) error { return nil })
-}
-
 func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 	usage := func() {
 		fmt.Fprintln(os.Stderr, "Available Commands:")
@@ -217,7 +196,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Model = args[1]
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
-			if err := loadModel(cmd, &opts); err != nil {
+			if err := loadOrUnloadModel(cmd, &opts); err != nil {
 				return err
 			}
 			continue
@@ -371,7 +350,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {

 				switch args[1] {
 				case "info":
-					showInfo(resp)
+					_ = showInfo(resp, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -463,13 +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
 			}
@@ -522,28 +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 {
@@ -563,10 +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
 		}
@@ -574,7 +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) {
--- a/convert/convert.go
+++ b/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
@@ -208,14 +208,18 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}

-	if vocabSize := int(p.VocabSize); vocabSize > len(t.Vocabulary.Tokens) {
-		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", p.VocabSize, "actual", len(t.Vocabulary.Tokens))
+	vocabSize := int(p.VocabSize)
+	switch {
+	case vocabSize > len(t.Vocabulary.Tokens):
+		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
 			t.Vocabulary.Tokens = append(t.Vocabulary.Tokens, fmt.Sprintf("[PAD%d]", i))
 			t.Vocabulary.Scores = append(t.Vocabulary.Scores, -1)
 			t.Vocabulary.Types = append(t.Vocabulary.Types, tokenTypeUserDefined)
 		}
-	} else {
+	case vocabSize < len(t.Vocabulary.Tokens):
+		return fmt.Errorf("vocabulary is larger than expected '%d' instead of '%d'", len(t.Vocabulary.Tokens), vocabSize)
+	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}

--- a/convert/convert_bert.go
+++ b/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(),
--- a/convert/convert_gemma.go
+++ b/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(),
--- a/convert/convert_gemma2.go
+++ b/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
@@ -34,10 +34,20 @@ func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
 }

 func (p *gemma2Model) Replacements() []string {
-	return append(
-		p.gemmaModel.Replacements(),
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
 		"post_attention_layernorm", "post_attention_norm",
 		"pre_feedforward_layernorm", "ffn_norm",
 		"post_feedforward_layernorm", "post_ffw_norm",
-	)
+	}
 }
--- a/convert/convert_gemma2_adapter.go
+++ b/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(),
--- a/convert/convert_llama.go
+++ b/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(),
--- a/convert/convert_llama_adapter.go
+++ b/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,
--- a/convert/convert_mixtral.go
+++ b/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()...),
--- a/convert/convert_phi3.go
+++ b/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(),
--- a/convert/convert_test.go
+++ b/convert/convert_test.go
@@ -15,14 +15,21 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
+	"strings"
 	"testing"

 	"golang.org/x/exp/maps"

-	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/fileutils"
 )

-func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, llm.Tensors) {
+type tensorData struct {
+	Offsets []int  `json:"data_offsets"`
+	Type    string `json:"dtype"`
+	Shape   []int  `json:"shape"`
+}
+
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, *fileutils.Tensors) {
 	t.Helper()

 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -41,7 +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)
 	}
@@ -53,7 +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 {
@@ -89,13 +96,14 @@ func TestMain(m *testing.M) {
 	os.Exit(m.Run())
 }

-func TestConvertFull(t *testing.T) {
+func TestConvertModel(t *testing.T) {
 	cases := []string{
 		"Meta-Llama-3-8B-Instruct",
 		"Meta-Llama-3.1-8B-Instruct",
 		"Mistral-7B-Instruct-v0.2",
 		"Mixtral-8x7B-Instruct-v0.1",
 		"gemma-2b-it",
+		"gemma-2-2b-it",
 		// microsoft/Phi-3-mini-128-instruct@d548c233192db00165d842bf8edff054bb3212f8
 		"Phi-3-mini-128k-instruct",
 		"all-MiniLM-L6-v2",
@@ -140,6 +148,36 @@ func TestConvertFull(t *testing.T) {
 	}
 }

+func TestConvertInvalidTensorNames(t *testing.T) {
+	f, err := os.CreateTemp(t.TempDir(), "testmodel")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	tempDir := t.TempDir()
+
+	td := map[string]*tensorData{}
+	offset := 4096
+
+	td["model.layers.0.self_attn.q_proj.weight"] = &tensorData{
+		Offsets: []int{0, offset},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	td["blk.0.attn_q.weight"] = &tensorData{
+		Offsets: []int{offset, offset * 2},
+		Type:    "F32",
+		Shape:   []int{4096, 4096},
+	}
+	generateSafetensorTestData(t, tempDir, td)
+
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || !strings.HasPrefix(err.Error(), "duplicate tensor name") {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
 func TestConvertInvalidDatatype(t *testing.T) {
 	f, err := os.CreateTemp(t.TempDir(), "testmodel")
 	if err != nil {
@@ -148,23 +186,10 @@ func TestConvertInvalidDatatype(t *testing.T) {
 	defer f.Close()

 	tempDir := t.TempDir()
-	generateSafetensorTestData(t, tempDir)
-
-	err = ConvertModel(os.DirFS(tempDir), f)
-	if err == nil || err.Error() != "unsupported safetensors model" {
-		t.Errorf("expected error but didn't get one")
-	}
-}
-
-func generateSafetensorTestData(t *testing.T, tempDir string) {
-	type tensorData struct {
-		Offsets []int  `json:"data_offsets"`
-		Type    string `json:"dtype"`
-		Shape   []int  `json:"shape"`
-	}
-	offset := 4096 * 14336

 	td := map[string]*tensorData{}
+	offset := 4096 * 14336
+
 	td["model.layers.0.mlp.down_proj.weight"] = &tensorData{
 		Offsets: []int{0, offset},
 		Type:    "I8",
@@ -175,8 +200,16 @@ func generateSafetensorTestData(t *testing.T, tempDir string) {
 		Type:    "U8",
 		Shape:   []int{},
 	}
+	generateSafetensorTestData(t, tempDir, td)

-	data, err := json.Marshal(td)
+	err = ConvertModel(os.DirFS(tempDir), f)
+	if err == nil || err.Error() != "unsupported safetensors model" {
+		t.Errorf("expected error but didn't get one")
+	}
+}
+
+func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[string]*tensorData) {
+	data, err := json.Marshal(tensorData)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -297,7 +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)
 			}
@@ -322,11 +355,6 @@ func TestConvertAdapter(t *testing.T) {
 }

 func generateLoraTestData(t *testing.T, tempDir string) {
-	type tensorData struct {
-		Offsets []int  `json:"data_offsets"`
-		Type    string `json:"dtype"`
-		Shape   []int  `json:"shape"`
-	}
 	offset := 4096 * 8 * 4

 	td := map[string]*tensorData{"__metadata__": nil}
--- a/convert/reader_safetensors.go
+++ b/convert/reader_safetensors.go
@@ -49,12 +49,19 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 		keys := maps.Keys(headers)
 		slices.Sort(keys)

+		names := make(map[string]struct{}, len(keys))
+
 		for _, key := range keys {
 			if value := headers[key]; value.Type != "" {
 				// bitsandbytes quantized models are unsupported
 				if len(value.Shape) == 0 {
 					return nil, errors.New("unsupported safetensors model")
 				}
+				ggufName := replacer.Replace(key)
+				if _, ok := names[ggufName]; ok {
+					return nil, fmt.Errorf("duplicate tensor name '%s' was found for this model", ggufName)
+				}
+				names[ggufName] = struct{}{}
 				ts = append(ts, safetensor{
 					fs:     fsys,
 					path:   p,
@@ -62,7 +69,7 @@ func parseSafetensors(fsys fs.FS, replacer *strings.Replacer, ps ...string) ([]T
 					offset: safetensorsPad(n, value.Offsets[0]),
 					size:   safetensorsPad(n, value.Offsets[1]) - safetensorsPad(n, value.Offsets[0]),
 					tensorBase: &tensorBase{
-						name:  replacer.Replace(key),
+						name:  ggufName,
 						shape: value.Shape,
 					},
 				})
--- a/convert/testdata/gemma-2-2b-it.json
+++ b/convert/testdata/gemma-2-2b-it.json
@@ -0,0 +1,312 @@
+{
+  "general.architecture": "gemma2",
+  "general.file_type": "1",
+  "general.quantization_version": "2",
+  "gemma2.block_count": "26",
+  "gemma2.context_length": "8192",
+  "gemma2.embedding_length": "2304",
+  "gemma2.feed_forward_length": "9216",
+  "gemma2.attention.head_count": "8",
+  "gemma2.attention.head_count_kv": "4",
+  "gemma2.attention.key_length": "256",
+  "gemma2.attention.value_length": "256",
+  "gemma2.attention.layer_norm_rms_epsilon": "1e-06",
+  "tokenizer.ggml.model": "llama",
+  "tokenizer.ggml.add_bos_token": "true",
+  "tokenizer.ggml.add_eos_token": "false",
+  "tokenizer.ggml.bos_token_id": "2",
+  "tokenizer.ggml.eos_token_id": "1",
+  "tokenizer.ggml.padding_token_id": "0",
+  "tokenizer.ggml.unknown_token_id": "3",
+  "tokenizer.ggml.scores": "0872465d173867d755d3ee728f882b9dc2057a0bfd596fe1e3d131522f1250d8",
+  "tokenizer.ggml.token_type": "8d40143b3477df77beea4139420335ede458bf5e14102f01b0170197b55da8d8",
+  "tokenizer.ggml.tokens": "c6e66de1841f04de8b8d236d461ab720a4c9b9b5414dc293a09c6e10eab45fda",
+  "token_embd.weight": "64a9d30707e659e2e673656d71f5aef7a9fb9fd83bb9a77558dfc5abbe218a05",
+  "blk.0.attn_k.weight": "d8b4437c5edb3cddf6af9987038e1bb2b191c4f0fce0e160d2abace717f5d5d7",
+  "blk.0.attn_norm.weight": "1eb73e3f7aa8e502f6ca31cd19efbb8e4fd9a89692e13e48ac8205545a7fa7e8",
+  "blk.0.attn_output.weight": "39e7b78e57d356a22dd89ce1c4d7163b970712ba756545e1703f97866cd2192e",
+  "blk.0.attn_q.weight": "795058e23b6109febd9d55c89e1eebe6af0714ec8c56fd86a160876a6135ffe8",
+  "blk.0.attn_v.weight": "0cd6e583d1887c020472e961bbb113fe5a0d23ae2f1c2c876fc366cdb7692b52",
+  "blk.0.ffn_down.weight": "51eb4d962189e945a84e94e0dc1aad3f8f90cc1a11e18029670afcd0ea0acb1b",
+  "blk.0.ffn_gate.weight": "9811a29b8ad48432925897ab21dfcb13c5cbd372aeccbbefca9b7866883b4ce3",
+  "blk.0.ffn_norm.weight": "92cbf4652ef503c1de5b10f2be00b3fcf00100980cb3baa8f3013a8d8bf3d851",
+  "blk.0.ffn_up.weight": "af87de21746879483ed1b374cdd76b19ba11ca2b6dbb1beba98efdf3be3e8077",
+  "blk.0.post_attention_norm.weight": "32e135f1f258ffe407018899e39af1725d59d66d60022b9a21575ba160e0357a",
+  "blk.0.post_ffw_norm.weight": "ba286f5ac11b07fbc986173708c66f1920427be5a6d108af38fa0a837c1c8eb6",
+  "blk.1.attn_k.weight": "51584435552051f7fade76beca582b3f7190cf7fc07adcf527c2774d4b1c3901",
+  "blk.1.attn_norm.weight": "6833104c7fbf35a7e799ae56c262b97fffa14789642aee14381b25acd21ed80a",
+  "blk.1.attn_output.weight": "14c39481369087bf292ac9a3ab2ef166f9fe376a9f90c246653213ef264febdc",
+  "blk.1.attn_q.weight": "443f64ae2229f857c69d6bebb7800b685786cb77884c3ae19d4286aeed081325",
+  "blk.1.attn_v.weight": "0df482de2038f1e4c8a7733ac0ddb69ad90759dab5968b942af0155588de4c4a",
+  "blk.1.ffn_down.weight": "66f30763a8bbbcaea609a0087ed75fadb5e771c06378dd2cea94cf17e492e8cf",
+  "blk.1.ffn_gate.weight": "a7151bff00a545fa18b2c92dcd2a14572ccf9beb957a6c494f1374e8ebe174c9",
+  "blk.1.ffn_norm.weight": "e197d71ea11b5276bc0167d2663b88089b3ff42b47ba91e85f6c5d95f6306435",
+  "blk.1.ffn_up.weight": "57c182e0b14cccd1350d388f0c616991702e74281db54637451b70f4ccc24f9b",
+  "blk.1.post_attention_norm.weight": "3c56f837168d784c2d8bac247c130bdca6610c095c8da4558c536ccad7605609",
+  "blk.1.post_ffw_norm.weight": "d2a51d320fd01069dd7ccaa7082f16a7faeb671885607d7900b10a89c354d0fa",
+  "blk.2.attn_k.weight": "bc103c818192de7ce36caaf89dc117be4df13fb902e0bd9a23c64edace5df9b6",
+  "blk.2.attn_norm.weight": "0f2503aa126083a5d6ac72481be1ef66c6014705b573682b35bd864e4749a3d5",
+  "blk.2.attn_output.weight": "05fcd4a1226e482f91803a266f72caca887a93e63c2d2ba5611ab3c68d38743a",
+  "blk.2.attn_q.weight": "6a10b5c2fd423d1e4c4fd60fa8c154a0159b6b2501ea79cae2ef19f45a674e5e",
+  "blk.2.attn_v.weight": "3cf891945a1f8ae7cc908a5c6b729ff5b70f4436c5ffdbf245cc0ed4cc19cd1b",
+  "blk.2.ffn_down.weight": "ea204fd04e0d2fc728a9861a459216bbfec629c152004ba625f52cd8837bd51e",
+  "blk.2.ffn_gate.weight": "3a3518729f1b8b64a82b8792f33987db5418fdb094be0263c68f146a5c38de54",
+  "blk.2.ffn_norm.weight": "754ede678b725de41a34b82f0edf7688b5c065be7c0d46df6f7ad9430d986884",
+  "blk.2.ffn_up.weight": "ffdcb88439f5828ffbd9fc844b03ff91637b790b9838097258cc3ae75935720c",
+  "blk.2.post_attention_norm.weight": "4b3f53b7ba26e8c36b2dfda3b7e5fc4b1065257cefdea235fc7df9af130ac2fd",
+  "blk.2.post_ffw_norm.weight": "e550369e26b8485e2b54ad34b34bc98af5494287dcc513c2c39cf1eaa5b89d07",
+  "blk.3.attn_k.weight": "89f24ea450e37d9e95757651a83205c085d81b354ee9489dd6310a391d8409f3",
+  "blk.3.attn_norm.weight": "24e2ea662b7cb822b4ca5cd61bc17f2709f406d990ec3b4a0dac1cc112db45cf",
+  "blk.3.attn_output.weight": "ac4dad69473c6e3fac56669212cadd8c34ecc5973d945972e974d94805334967",
+  "blk.3.attn_q.weight": "b6a9c9a7d4722b9096631c65de62228dfddca6e26edfe6af7fce01e116ef0f4c",
+  "blk.3.attn_v.weight": "f272a960a40093942309bc342a379984cbacec2d7bc64428db3f64e6b1887ed4",
+  "blk.3.ffn_down.weight": "c0188ba50d8228805982029c277fc0e87aa57473b8363037c648f6d006ff828a",
+  "blk.3.ffn_gate.weight": "a04aec1561ee6c0fbb18c3db49dc62fb533619cf697fd548cbf2279761aaec3b",
+  "blk.3.ffn_norm.weight": "bc053837d44087ec05eb5d9458357b2a5be787789b19cdbbdc694b57697f99a6",
+  "blk.3.ffn_up.weight": "b3ce8b274f20796d3b1a7c08ba27a919066f9de89a782faa544c4a8d6bea1382",
+  "blk.3.post_attention_norm.weight": "9c922dee7a7df5667289e2788e60170238239cee2dfdbbd9e435763f9f416718",
+  "blk.3.post_ffw_norm.weight": "b682544ac953ad2e0b49027ed8916f2e9d1aba5d1587bb4127ac703570c7a03a",
+  "blk.4.attn_k.weight": "143b0cbb4b787b95c2b6212374410e32173ccef2adb914908a2f89a7916de512",
+  "blk.4.attn_norm.weight": "5668f60491b780273745192662d02c9a92a4f692b29d16aa0bbc7413fec4f85b",
+  "blk.4.attn_output.weight": "b9f2bdb68be1e0cf66dd19f8fa2afb105910ad2ef394864cb32cea8f8944e0d5",
+  "blk.4.attn_q.weight": "ddcf1343dafbc2dfcd0b8741225af22fe4b54b2becce29240bd01c34265d126c",
+  "blk.4.attn_v.weight": "6dc7074366e7ed52d9f48c594dcc85bef738e096276cb99d28228c89eecc5b9c",
+  "blk.4.ffn_down.weight": "30334ffc59ce343cf2a1b973174acb7722823463adc07e19a99bd0f404bc9906",
+  "blk.4.ffn_gate.weight": "890f7c8af208d63b28db52c4b8c16c2288a382d87ff5a6a6d6b0a5b3bf27e6cd",
+  "blk.4.ffn_norm.weight": "ff0316cc7847221eb86a90c1ab441d4ee61553d410c66414a7755021b3b12448",
+  "blk.4.ffn_up.weight": "6af97d113f91564c636734f215e25ee602d48eb045458f300b3ec7582be0f41d",
+  "blk.4.post_attention_norm.weight": "69438f231e105e68216b078bdeb35a7cdc8b12c4e2845e18ecf4c8d361d6a321",
+  "blk.4.post_ffw_norm.weight": "0fd535da78bcf2b32c95b05b2b83dc49817393765be90d8cc1ed3d56f47b68ec",
+  "blk.5.attn_k.weight": "0166eb3c6d20dcf3d3c169e94caa8dee057535bb525e29f698fb6f8844f18a6c",
+  "blk.5.attn_norm.weight": "a7808f27f164023d5cde2be00fc23cac6c71aa0ddeb60bc23e12411b80087672",
+  "blk.5.attn_output.weight": "8b65b2027a0842b68c5308f91d6a31de9599d794157d77df8418b19f9e0d9334",
+  "blk.5.attn_q.weight": "966bc626ef2c2394d872087a41c126bb1b67d1d5f6de920204ef5e5b16c34003",
+  "blk.5.attn_v.weight": "9a362aef3f4437fbf0ef6e1ba785f3329c3db2960f93fe36547d2795e9c254ea",
+  "blk.5.ffn_down.weight": "63e53541d34197720c06f297aa8142ac6b6eec002c7987b296f26e8b1400f931",
+  "blk.5.ffn_gate.weight": "d9591fdd32f783e0fc26e20d5d587ee8971ac8ae2e4c818c6eac1c125c7c7f37",
+  "blk.5.ffn_norm.weight": "677334cc60ecce3a7f4ab3acda15d359353d7358872f614ad8914e3780e9fc6e",
+  "blk.5.ffn_up.weight": "a63764110e1c655ffbd55af0669b2dfe4cc29d0e198d33a8e5426461b08a85f7",
+  "blk.5.post_attention_norm.weight": "c55499f859b2c0a7f5cabceaae47309a5ad38bc29d0f4a8db81f1357023162a9",
+  "blk.5.post_ffw_norm.weight": "82752754665f842418f3e302cb5f43d1e0504dcd124c4b8ddb77018b2c793837",
+  "blk.6.attn_k.weight": "e20a5f0d6c807273c8d491439566b428497ac02097cf0aa55e33748c28e14be6",
+  "blk.6.attn_norm.weight": "2c6ba42fd3c73d72073ced03a32dd28d70a89ed9bbbc8fea1ba03a7ade951e6c",
+  "blk.6.attn_output.weight": "4de7c5c2f4a133a266e17ed8c14c52959466b54cc7ab9e19f789a33b4850f284",
+  "blk.6.attn_q.weight": "56462d921800e6b8cd2213fef04c4ff16d728905cb2f4c58e966d0a053a3b0ae",
+  "blk.6.attn_v.weight": "b758dcbff769d6240c2245ede1dbc62c4170a67c77458e866312589220fe29af",
+  "blk.6.ffn_down.weight": "582247fb3c2bf687cbe9413fe18d18ad47bef4b65df7d78905e10335c6134764",
+  "blk.6.ffn_gate.weight": "3035444d5286aefb7a6d04e55bc27e1fac7cf895cd5be02319a431b8e047b4ae",
+  "blk.6.ffn_norm.weight": "e582d24c66e01b96faa20ce6adfda3d8583b11e809bff89969927398175e369a",
+  "blk.6.ffn_up.weight": "6f4b7bbfedeacf61a4866ae0616c4ba6c9e856662e8f00ae6aaec7f52c53e7b4",
+  "blk.6.post_attention_norm.weight": "8fe51b50bd677d21586aecab0b565c4bf9fa68ad50bfe366f45e8fea3c657ca8",
+  "blk.6.post_ffw_norm.weight": "81ba3cb4c2bf5c546b86855b7a885d3fafededc67eb3a35cd3598b03c9e26e65",
+  "blk.7.attn_k.weight": "2e044179cdcae0946708c86bfea7aa0391e1f7e2a09b33fca035d384cc3ca758",
+  "blk.7.attn_norm.weight": "94b48c546b046803c60e75a3acb17a356b710735989938021b565f68df9b4985",
+  "blk.7.attn_output.weight": "65709b4ad7a581f4d75793d39d4032a359f6bcc0c3835205242a0b99e5b66824",
+  "blk.7.attn_q.weight": "8ded993c95d1f7caf201ceb6fa035cd6ed6d351b50b999fa9355dfee9486cb5b",
+  "blk.7.attn_v.weight": "c92d5e2d2d48397542bc03bea25bf39154075e66c5bb1ead85188505aa04ae91",
+  "blk.7.ffn_down.weight": "e8ba8fb57208805ef1dc23cd7c86e9a2d1fb7c52c3940d292cd5bb2eb24b3fac",
+  "blk.7.ffn_gate.weight": "f0f06d6a2e06c5ac252083bc61d05c814e6289d3f4e4a87d2f06918254c02c36",
+  "blk.7.ffn_norm.weight": "ebf8ef775f72624148e09d68a4332187a7a5020c521fe0623da1cd3485ad33e0",
+  "blk.7.ffn_up.weight": "a554adc4fc7122c247c77670e169916ba1794c787b5be30a2b36705138f1f746",
+  "blk.7.post_attention_norm.weight": "3aa6bc21d85c3a0c12b964e82b12feaedfdd13130c3cd2229228e24e0967ebdf",
+  "blk.7.post_ffw_norm.weight": "508bc7b19ee8ff08f0007c890133a462fc57c7e72b16ee8f6dd64def264ef876",
+  "blk.8.attn_k.weight": "363c8e74056642fe9e7c2f3f9769d57319cd3fa0a6022810189ab8d894322885",
+  "blk.8.attn_norm.weight": "685b49a1f1acb169f4df0bdd8e3de6943f3033cebad14b898a72000595610d92",
+  "blk.8.attn_output.weight": "7bde571e4efef1c6a6143f0526721dfb59e0a0ea0e1a3616a322b2eb937efa48",
+  "blk.8.attn_q.weight": "fc993dbc1074c28a0e1d85e5ab2f4ea6a9c6c1affe7ee56027000a275daed9b6",
+  "blk.8.attn_v.weight": "281e8791d3aef9b3864f1cb054da0ae0c2fef4ce0a58b1bad8bc136b2fa0f62b",
+  "blk.8.ffn_down.weight": "b1164a2578a7f87ed99c2bbc76c5dfbbbc6a1a803605391acc3f320fc989ffd7",
+  "blk.8.ffn_gate.weight": "6b39a3b3aaaa79aee61416b54d62160b9258042650e61c6b47bc77c2dd17daf3",
+  "blk.8.ffn_norm.weight": "17ea1362c72da27f12bc936500492035bdef3fd8f940cb12b57f37d42ba8ecb1",
+  "blk.8.ffn_up.weight": "bc3a7c47afc440d2bdf8fbe9ddf2c9220467472c60c8b4ded8c0f181470ec96c",
+  "blk.8.post_attention_norm.weight": "5c506204e00411ef9c8b4134d40eedcc19fffe68dd0af7d7cc49dcabf2dfac7e",
+  "blk.8.post_ffw_norm.weight": "002faec235c3678864e2901eed275ce4e9dc229164a91c9cd4c965142ba62305",
+  "blk.9.attn_k.weight": "0bab39d8c237f1b6d0010db40467142625a9e6f2e0e4c49a56c12b41e4e0b1fa",
+  "blk.9.attn_norm.weight": "de5f38e873b17f07aa7598831b89cc1cae2c9bc3eb2e042ee9af059d2563e84e",
+  "blk.9.attn_output.weight": "8a8184702c25a62df9ff309c0c7badc8587208523b2be3e8fa90ce7080573e6f",
+  "blk.9.attn_q.weight": "7c961b2431b09ddf95377acd07201cb91bf13d9cd3ae0f2c25c7d6a0358d9f50",
+  "blk.9.attn_v.weight": "e22d240cb4743067033e659cbf210ebe2ebbab3e1dea6ccbe5eaa982382ca038",
+  "blk.9.ffn_down.weight": "a426f81210f03d6ad53277416e1fdcdf37d8065e4817613edaf6c67a343426be",
+  "blk.9.ffn_gate.weight": "a82eba825cb77b8e64f85ff99ede2fc71bc9b01751eeb17e9e6c246ee12ea62e",
+  "blk.9.ffn_norm.weight": "1a97f9b1302a3a326d534c5c3fed2db6db0ae45fd0edd381a3e4fc1c75d81030",
+  "blk.9.ffn_up.weight": "5f20bac2bbf03bb42adb92fbf99561651e1edda57e0b61935ac7f6c08c0ed7cb",
+  "blk.9.post_attention_norm.weight": "9f9866d13988e1946b1e1c80d9374a92a6e3be33748f8eaed3e126d1e1a4c796",
+  "blk.9.post_ffw_norm.weight": "a6896dbf698db4dbbe5dbf12417d4fd80e9cad0c539c858892ec0aa5b046bb58",
+  "blk.10.attn_k.weight": "ca8446e5d21ecd4e6a70dca8d321be480be4fba94d70cba065205436feb44270",
+  "blk.10.attn_norm.weight": "4f41fe290e8f21f63b82151b6cce94bf7318d121468816b0c58af0ff7c1658ab",
+  "blk.10.attn_output.weight": "c626d2e9681c5c941bbde43dddfae1a8d4986bf2be4470857bc8e8bd7f869044",
+  "blk.10.attn_q.weight": "1e61b210a13a429977325cf15d781ab77d604cfa862f4270329cbd94237d5835",
+  "blk.10.attn_v.weight": "8ff8d3e3f058ec3b35ada1057f2ed59c06494d0e0be6a8dc3ff9edf9f0e1a115",
+  "blk.10.ffn_down.weight": "bcebc04219f8081a5f483e58103c0ddbbbc631a0a54fd6dd9d55778e041f70ee",
+  "blk.10.ffn_gate.weight": "7a23a1e620ef871384ddf9611ccdcfb893fbf013cc203ac8e72f745420f1eea0",
+  "blk.10.ffn_norm.weight": "e3a375e43c349a1c6c66c22328e513cc1af3137fe839e43dc8e9be2f65914fd7",
+  "blk.10.ffn_up.weight": "5d182e7c94369194fca5f19cbbe668a999911e57f3d363bc7fb6088428700cb9",
+  "blk.10.post_attention_norm.weight": "b841c6308296e8984f3c5f549c6e3a242f4b3e19141e1f54cc08de9c46759c09",
+  "blk.10.post_ffw_norm.weight": "9d66fa05b5c940208f634f5053d809094c99a2a10a1d1e8847c8281fbd99fb49",
+  "blk.11.attn_k.weight": "14adf24ebb2bb17b336ca81cec3e690fd854782f4440ca6c66cc1d7e7bf1c850",
+  "blk.11.attn_norm.weight": "2d2213f311f50414702b5b34f22aafb9d9a0b6787243e7578562583dc40ad195",
+  "blk.11.attn_output.weight": "de1f14cc2a7fff00cf11b229f0576999205f17b9536e97abc9d6de3cc79a7884",
+  "blk.11.attn_q.weight": "2bcc5c147524003109ece0be08b89ac8b25baa71416ffa76573c6c052ffc6eea",
+  "blk.11.attn_v.weight": "2e6ab8573070c22dc1e0d7aebe4d52123226dacf7822dcce06fadbb38fb036a4",
+  "blk.11.ffn_down.weight": "1b86902f4e36868421e5228b9445051f8290b292df22a6d1af836dcecc1f25c3",
+  "blk.11.ffn_gate.weight": "e756e8081bd0a16aea4a9ef5076ad102113524f7a3d50a3a77aaa7f7938b63e8",
+  "blk.11.ffn_norm.weight": "6913887267be227cf9d1991a3dd8db2e7e74bb9b5fbdfcb9ac954fd7d7b95b3b",
+  "blk.11.ffn_up.weight": "619a3ac0609ebdf42c3fb2b6e4b1db48df79e6dd8418d7ab8f1bbff13d8a6a50",
+  "blk.11.post_attention_norm.weight": "e4b4ba92cef7b6a78407e8ab1b0307d47dac6c3df7b6817e28038317ff662d7e",
+  "blk.11.post_ffw_norm.weight": "40aceeec58cb855f0c158c9cc217168fcd5d0e735567d587217b1d78df17bc5f",
+  "blk.12.attn_k.weight": "c54c5a4d4892522022d1aa2204cfc624f0b4042caa536e678967316293fe5cb1",
+  "blk.12.attn_norm.weight": "7cd2ef58298569ffdf244d9b390f3917245276c8206e5780af5f96d8c0bbb446",
+  "blk.12.attn_output.weight": "85495ef9cc8b3deb21f741bde463ff6493acae2be51f02ecdeef952cbdec3375",
+  "blk.12.attn_q.weight": "d19383f83fd119bfb8c0280c9515705c11d8e7d502019fcf8f49efeef0d106d0",
+  "blk.12.attn_v.weight": "869ac669ba49531d9128892a0e27cef15de508ff40cdf80cc1681dde50d09204",
+  "blk.12.ffn_down.weight": "578f39f8f9fc2f09138afc884a952d7cc3a9a31de4216acd10e88e19e0b75f8c",
+  "blk.12.ffn_gate.weight": "e29a0186bc6c4a0720246306e922d3a83f777dadcf4ac80bad468287031cc8b5",
+  "blk.12.ffn_norm.weight": "e1ee95c6584b5cb57fcf1db8ce2bcc03aff91eb389238c094a61c00dde93d1f2",
+  "blk.12.ffn_up.weight": "2a826f06d7cdfb3edc6ae250ff44363ef77a2a9cdf96313e23a331b99ebfa17d",
+  "blk.12.post_attention_norm.weight": "4bafc7699b948d5cbc0d3e09b418b06c6abc4651a61ada9609d9a2f21c7e5607",
+  "blk.12.post_ffw_norm.weight": "bbb8c34a7176bb1a49f9fe2bacca0bd26b673d52c0835b2e90fa11f2962f077f",
+  "blk.13.attn_k.weight": "ffeefccfe8255d1b694382012ff4134eee5fec9d9491c8d0ff0a13832d1a37e8",
+  "blk.13.attn_norm.weight": "35713726529e3887c4135a88e86e8a4d7270ba5b9f2d1ab462622fbf40a7cdce",
+  "blk.13.attn_output.weight": "0d60b7c5cd71190a9ef4b873b0f516be15447c32d83914db2794b14592b0b460",
+  "blk.13.attn_q.weight": "8296069e65bef794cefc61257fc65789b3cb22955e30f3df129205e5041b2222",
+  "blk.13.attn_v.weight": "ca0f4ab9d16a748fc643a5c0c7a19826a811bf2a4e7316a8c935d4bf0ce8abc6",
+  "blk.13.ffn_down.weight": "d5514e0c8e7b3ed1cbcc1605eb5be1733b6ab3514cf8a0508fc72f7d05ed8bcb",
+  "blk.13.ffn_gate.weight": "8108e517a82e08a3aefbbd267bfa50a1668f92a76273280ce8a6bc1f6dd61521",
+  "blk.13.ffn_norm.weight": "5fcb6132d2134bf1f835b904a99820fa501dbc57d2224129f7098bf3cabc1d36",
+  "blk.13.ffn_up.weight": "6d744b7cd390a3cae3aa350dd379b81246acd056a2259996b6aaadece8465ccc",
+  "blk.13.post_attention_norm.weight": "e08b14698912509790e9575b8676971fbb0a4d82d719367e3756c0d0c4ab8cc0",
+  "blk.13.post_ffw_norm.weight": "2b196e4450fc5f1e7367b2cf7fe33a15fe919fbcdd861d11002346f16e980535",
+  "blk.14.attn_k.weight": "120e5f48d7268dfd9ab5f4bc9cc57a7cec63ea9635f56b80d435eb22936e9483",
+  "blk.14.attn_norm.weight": "146367bcce4db72cc894419a2e0145a6f533507dd68e4739c10ee480308c401f",
+  "blk.14.attn_output.weight": "720fa0165e756876c5cb6ad9e2780dd910390933f3f8849e5add5da04266650b",
+  "blk.14.attn_q.weight": "f5183466f56219ca1aca52d8b82c2d966a4198fea40fdd6b39f4d8b06ca2a6dd",
+  "blk.14.attn_v.weight": "24f8ea3d5512cd37c43c8329cb0da0c90d1895aef763ac2dcee3fe5157ec50a2",
+  "blk.14.ffn_down.weight": "e29960965b384ae5ab3d898a4dbaa8fddd28fa0e477ac28bcac49dec12a5ac67",
+  "blk.14.ffn_gate.weight": "6d0d6a74bfe9692e8f8eedff0fc34fc4fa1c8687794f35f2e2b033ab2d7510b8",
+  "blk.14.ffn_norm.weight": "f7036c1a9a71e046c9d2af16e9218fda5dbb0f7241ab44747abed1f0f9d602ca",
+  "blk.14.ffn_up.weight": "7d69ea1424007ffc9c12247dd0308c616e93ac02a59ec341cfa48f92d6ce3b10",
+  "blk.14.post_attention_norm.weight": "65b9712834d9445d4236bec362f3fb795c20d60c541b3dc6dbb7914d9b493e41",
+  "blk.14.post_ffw_norm.weight": "9c6a8da2e4e437d5cfdf3b9097e9f8b64bf07946a048badec20f4d374613f38f",
+  "blk.15.attn_k.weight": "864bc618303a0e4ee67fb1d5e751de61e936cd51e96669dd86f8cd08f2305045",
+  "blk.15.attn_norm.weight": "f9f4187da6eeadc2fc5921d8fe669741697d16c13d71e4aaeb73b82f50dc577e",
+  "blk.15.attn_output.weight": "ce2419a0b097036b2a31f2f4ad731d5814bcc2ef4c511786e24471e5eefd273b",
+  "blk.15.attn_q.weight": "9539db5a970d11ebe99722d1e13fcd635e250033630811efe583d2f97778e4a9",
+  "blk.15.attn_v.weight": "1c834b48ccd88adaeabb7d8bcb6be0bcd6d5ac1354ce88fc28f19a1a96b81ab3",
+  "blk.15.ffn_down.weight": "bc1f97a65dde6fa2c1e5397afb612266944b343f2eaa868b635ddd25829f8a42",
+  "blk.15.ffn_gate.weight": "1b14529d57056b79037f6cb5008132e62cc35992353b38dda59572274623103b",
+  "blk.15.ffn_norm.weight": "9af77458de9ee55c66f93865759f9c2c398557f94f3fa8fa6af30543d7339cde",
+  "blk.15.ffn_up.weight": "41d524a26b61a9595816b4fd53cf57ef50a702e4ef32933ff6136dca9136a267",
+  "blk.15.post_attention_norm.weight": "c60a03cd0e63a7db5c80015e58e9b97ba2208caa19f66a6fef5c4447eca900ce",
+  "blk.15.post_ffw_norm.weight": "34f7f9f96769215bbc3d17084df091864aef96a6645b7d0b3b7d9bd92f1a4b0b",
+  "blk.16.attn_k.weight": "7e27240d9f3a8c6cf0f4a980113d43234f514eadc3e3e1792b86efb29ffb1a6d",
+  "blk.16.attn_norm.weight": "af798acc0899282a30448edec48223b3e8efda177090273e612d8eca5e377301",
+  "blk.16.attn_output.weight": "79df39a3709d3d53e84146291e0944a7a653d06705293d9ccb5648dceadb432c",
+  "blk.16.attn_q.weight": "db58a1c3b83ad294804e5fd7321005719e200659173466df5a52a182b80b7165",
+  "blk.16.attn_v.weight": "2af6d48cbaeb225b5c1a704f76abd89c8ab1521417695b112b4dcc2cbd39b74d",
+  "blk.16.ffn_down.weight": "fc1c813eb5e7da3d6194569d6cb21602fc6eff2dc8e1b0eb753f2d5df148189c",
+  "blk.16.ffn_gate.weight": "7a80bcbc42464bd55df4814a6edbd7b5c153e0428323bbe49de55e2d2add33e7",
+  "blk.16.ffn_norm.weight": "2041685ee926d30f3f2ae4ec35b5688f1cd834167a6359a7d4057eac804c58b2",
+  "blk.16.ffn_up.weight": "8da4b718973ac1d43b928829bc45e062fd101984d6c98dd825bd7c5d08ebfbe3",
+  "blk.16.post_attention_norm.weight": "975c48fe680a6167438a106140a8872eee7765191f152d80e3b8ddf47693e095",
+  "blk.16.post_ffw_norm.weight": "4de2d4d483acfe4fc77860ea929025df2f4e15c10729413f36a18c94eaa6d689",
+  "blk.17.attn_k.weight": "f937e61f0af8c4cd98ee742648eb60e02e579683e21d421071295a3b70aebaad",
+  "blk.17.attn_norm.weight": "c3270583ed28b7e423f5b170c59113234f258169b93a867d9274f4c10b7cb115",
+  "blk.17.attn_output.weight": "b8c1150e81e685e539a5dcf2c19047a24eba2b281fabe166674b1d71ef4612ea",
+  "blk.17.attn_q.weight": "c255100ae2011e7dc7e3bf3bc3ccd96d859fbb98581cae993d7b82c1ba8e8b39",
+  "blk.17.attn_v.weight": "5830bb0a555984c6485348067f70b5d22ae337c011aa9248dac2ff4c95944551",
+  "blk.17.ffn_down.weight": "8ff9a7cccaa3776434a9d895aae4fb5c36c736bf2ec98784226b4c234940fbb0",
+  "blk.17.ffn_gate.weight": "1b52876739712831c272911533da206f407b46034a1a4ae8a88c1f96b6bd5747",
+  "blk.17.ffn_norm.weight": "d0e16ba5e87c91b545334e022058c7d03849665c3b1a6298771b656531366b66",
+  "blk.17.ffn_up.weight": "4dd6211d01dbebbe21052708eddc242b082a58b5f18ed16479e17987c1d3432e",
+  "blk.17.post_attention_norm.weight": "6f49c775c7417dade77ba8268a0f8441c1e5ec28b5d7e4dc5ed07a04d04600c8",
+  "blk.17.post_ffw_norm.weight": "b91a0bb2e6679e9c9be06ad323adae441d00a3d673efb19d7c4954be2aa84b27",
+  "blk.18.attn_k.weight": "22b565ace1b4da8b33865a58625be1d90beea9891f29686a69fa9cf7c93217db",
+  "blk.18.attn_norm.weight": "3e0160d7063c8753de65d2356a66648e47d921efdc5c917efb8209892120f8db",
+  "blk.18.attn_output.weight": "e3180f0bb4ca90b31e9b08158db38e332de62dfbaefe34aa94cc316409331e09",
+  "blk.18.attn_q.weight": "f3a5a83614c3ba7ea41cdd5b1b0819a241ee2a951a381ce4a9e001d3f700ed8f",
+  "blk.18.attn_v.weight": "f3350a5984fb951fc738adcf78147e6d812ff1c576670c460cafc99c253c1654",
+  "blk.18.ffn_down.weight": "9e9d09b13a33525e14bdaee6efc65c551ac7cf7680e534b940ab122a3a7c1ac9",
+  "blk.18.ffn_gate.weight": "ebaec8b4b578a2e8d815baac12f1675c208f80c68074d5a18288a2e1a60680ee",
+  "blk.18.ffn_norm.weight": "33e7687c53a242f2f8dc7093a491c97b18d4a5a8c14d183f02bd586a770f05aa",
+  "blk.18.ffn_up.weight": "78a1816662378ce56cc870e705174492781897b3afd2d4d97a51f10f2f2987c1",
+  "blk.18.post_attention_norm.weight": "a58dde3f12df3e94cbc27d87c8ea86f89af8a388a506446ff6758f05399b05fc",
+  "blk.18.post_ffw_norm.weight": "cebf90cc143577d483cca27b032dfd82031ee59bdf17c0e2cf60a0a3ad5bf996",
+  "blk.19.attn_k.weight": "4683375d0599ac9e2232196aae1e90af13a14cae26e865465de5c8e257bb2055",
+  "blk.19.attn_norm.weight": "f3eba936bfb1814bbcb0a1d62739eb66daac839df8c9c836fe0e94860df88525",
+  "blk.19.attn_output.weight": "51c0f01d38a9dcfe9bdbc4643576fab164c1d9e4b7168b7695c0ee55e6965667",
+  "blk.19.attn_q.weight": "28d15b69b8416f2e7ddc88fe381cb1e2ef2ad705fb1c268139ba96498cc74848",
+  "blk.19.attn_v.weight": "6860f1cd720638e63a981fa2c0b4db900129826bcb9823c9ddf9fb8b1b9f3383",
+  "blk.19.ffn_down.weight": "bc7f2d7827ee01c2dd41401c7b3b1700ad3a4ff620e8bb734f92630d342dcc7f",
+  "blk.19.ffn_gate.weight": "54d03ef69ba373fc410fbca8f1e34a565d58e4296d9a035ff7e48340b9c848e7",
+  "blk.19.ffn_norm.weight": "9178fc796a340ee6e8128ca74c0cb6203d1adbed6927af4e5ac7863da57affc7",
+  "blk.19.ffn_up.weight": "a77bd708026c6e83ad5c79c223278e74621bcf74a9641c7818d96b595daaad20",
+  "blk.19.post_attention_norm.weight": "ae94aa26f4c411bf9496a6fd4a6df64ee589ee1ae9a04b531d45acc95721e582",
+  "blk.19.post_ffw_norm.weight": "9ad210700edeef12133bdcff04bf1c7f62b49f6f4a9ba483c7cdc59857c24a5c",
+  "blk.20.attn_k.weight": "e35bce1e9f4a7a09ef34721f57ea38cfca68c272f52d923fe50af8308f66cfaa",
+  "blk.20.attn_norm.weight": "644800f6926fd34f233795c4dec1151a295d2138ca8cac33e3e48167d26f8b41",
+  "blk.20.attn_output.weight": "8d3758cd236471741e1ad66c0710cb79077dc8c7a3a292d35bc551c0c5abe627",
+  "blk.20.attn_q.weight": "c333b1f0f6f956b5d73891df10b1a0321e55fc31c40d623a24e1f52caa6a998b",
+  "blk.20.attn_v.weight": "8562b418d0c4868a050fb19fa3fcaf50a8cf1c669f537d666c80c7b3a04714e1",
+  "blk.20.ffn_down.weight": "97efb608ac44cc804198faec3ee66eafe56ced6b7ca5359700c6f1df75b7205e",
+  "blk.20.ffn_gate.weight": "5c61151d86f28415c73c73d90ec088c646cbe5c1640197caf58eb501ba7db293",
+  "blk.20.ffn_norm.weight": "24bbe0a701afd4bbeea65b3edde712b3cbb2281043bbc43dbf250582453116ed",
+  "blk.20.ffn_up.weight": "e170cf68e249566aa99eb6f6b265679bf9a5a6b76830ba24e7e130c2515910c4",
+  "blk.20.post_attention_norm.weight": "e092d751cfe20dbf2d348358f3b38397bd83e4ed94d6bbaa6bbaddcd902b2ac4",
+  "blk.20.post_ffw_norm.weight": "219a18a47dcba76e669e4322223a5a9227bd3db1de3fbd3d3cfb22e54a783c5a",
+  "blk.21.attn_k.weight": "c3a095ebddb42c63824f1c98da65263dc88e4d790a26aa1632840b44f5cc7cb1",
+  "blk.21.attn_norm.weight": "ef8bbaded5fbc45ad9cf3985ae02174524e7090fe6362811124f942ef643bec7",
+  "blk.21.attn_output.weight": "668f018aba72baac6252aa3ad58569ddd55ab751a0dd8d7bcc9fb9b6efb4bf53",
+  "blk.21.attn_q.weight": "e759c65663089f3bbbd51847934c185e680c82f1249065d5d487da638e519e6d",
+  "blk.21.attn_v.weight": "2ff57762686cf9ba1f5a6be76503454b97556ce67f4ac98254bd0562231197ba",
+  "blk.21.ffn_down.weight": "3fd106556fb721b1c28ae3f4026bc83eb1b08ed910f2ba5f466c6b5f327d91cb",
+  "blk.21.ffn_gate.weight": "338022d882f4b6619e8054a6fb909696fa3eef3013cf69b65c3cacdfc5b9e42c",
+  "blk.21.ffn_norm.weight": "1e77660c23a3f9653ee721a863d1960f773d87437cabc4dc0a6e17ee3d4e5e44",
+  "blk.21.ffn_up.weight": "7d31b20fbc2e6eba8f350f170069dc36f0cb12f68fbc4206ec5022a74085ebcb",
+  "blk.21.post_attention_norm.weight": "9638bae8d8bdcd7ed68da282979cd84a07c41ff9cabcaea94ebc846a1803db23",
+  "blk.21.post_ffw_norm.weight": "d622ef11115fe0cbe04b727d5a3b6371e7f39bf08c8d5eb9bc6da52e3f3cfb9d",
+  "blk.22.attn_k.weight": "5c321cb29deffbe57de200dd206a62005f1e80acb86c4fd2349dd44c8d3594fd",
+  "blk.22.attn_norm.weight": "198d949705d7170a331d75889d8c7500c3635254dac2cc6aa4dc35d556584536",
+  "blk.22.attn_output.weight": "19805cd5d7025b457e5d41d70db8b3fd63c2dd0e4a94d3ef1704d50ef4e749e8",
+  "blk.22.attn_q.weight": "177836cd583fc87405975ddc21ebfebdaa090a0363799664c72caa3da851ae2c",
+  "blk.22.attn_v.weight": "fea255692483e30d0108f9e4e250eb3ed7dbda8d83f499b06519b8c223ae6096",
+  "blk.22.ffn_down.weight": "00cb8939f03e5817d6d412de8cf2c923c9568d5493e382cec7faf5718fb034eb",
+  "blk.22.ffn_gate.weight": "b0591065b91281b2fbd8a9567f3568d40479f680e1f0a29e27ae213f37642489",
+  "blk.22.ffn_norm.weight": "96b5c5d0737c2ceb8fc869f54adb9e5f46e28cb7b177c40f49fa926b923c00f8",
+  "blk.22.ffn_up.weight": "81f472185b24344ab0594ea8246cc6e200e0dc1cab4943e74fbe4ca19d5a9701",
+  "blk.22.post_attention_norm.weight": "27fa9aa6260aa3071e0391e1a1d49322dcb6e8072315b8a9b7064087108dbd06",
+  "blk.22.post_ffw_norm.weight": "f37e1dcd7f643d9545675ffe9dc527a11eba86eb204989c2f44f636b266d896a",
+  "blk.23.attn_k.weight": "5d82f36658a56c3f94d0bb2d61f65509c966fa6568f81812e0d3e338b380ef8c",
+  "blk.23.attn_norm.weight": "b7983f88d9cad88bc88a528923e6da592ad20e699965b223ebc10840fe1f4fec",
+  "blk.23.attn_output.weight": "59f97f80f430d71606aab0158a195aed29ccd3405e6c0a5c41c809be8eb01898",
+  "blk.23.attn_q.weight": "53ac4789fe958919cc02ea4222bcd64c0ea1b4baa54304bff46635bdf42f7490",
+  "blk.23.attn_v.weight": "ec8abe09b9e84dbb52c7a068094657c6d3c62fe551ba8d7c3a3f23da622e9756",
+  "blk.23.ffn_down.weight": "3cf547eccb1b82aa64f208cee9682d7f558ca84e0aead7d9d3d1420d90f3d992",
+  "blk.23.ffn_gate.weight": "366aa2486d911ba81eb519119e13807deacf7e9908bc1975a2a63e00d6b10124",
+  "blk.23.ffn_norm.weight": "6d1d4a4af34bb7dc090ac87d6457d398c3e0fb68bd2e2b60b099dc318b6cfac3",
+  "blk.23.ffn_up.weight": "53f76692e253f5d2420b3f200c731b9f3b7a83e379920b4a067c729b4674aa4d",
+  "blk.23.post_attention_norm.weight": "7c952fa0efa76b3f048c8c4c9e8dcb5e3724d231327eda6423a34d3f3d3367de",
+  "blk.23.post_ffw_norm.weight": "7ab188cfe61f0a91b40309a0ab6bfa99f19d0ff2a37b6ac10e5f0c7f44eb5270",
+  "blk.24.attn_k.weight": "225798792f9bfdd10eff0505ebe61e0aad0209c17b431f6044ee7968ffe8c198",
+  "blk.24.attn_norm.weight": "635e3c1ebf5219bbebfc40ef164bc32d2b726ef595a94da64ac524ae878e2915",
+  "blk.24.attn_output.weight": "482f5bb2db8d9ed22b253d9a3296333b239efe698e5992e5d77e7e12dc2a5cf5",
+  "blk.24.attn_q.weight": "43805bbccddb65d58fffc4be9b5c374d4e1df1395ec1e1ffb4bcff03e98d5adb",
+  "blk.24.attn_v.weight": "fa741af54b4a3b1775d32f59134756090c5df2e7345a12a2d8db94fe289667a7",
+  "blk.24.ffn_down.weight": "83c6351e3162626b276f524a57836144625c2556dbe321b57cbd8fd486a68fab",
+  "blk.24.ffn_gate.weight": "fbe66be0d84d12cea5176cc7eaef64382ffc7324cd9d6266a3342dc43442f2ac",
+  "blk.24.ffn_norm.weight": "77c1445a8639ad24938bdf0280233eea2362d47391421833dfa72ec756dfc1e8",
+  "blk.24.ffn_up.weight": "78235ac729ee23c1cf1ae543751e3af32776d8808cee6e529c2a625a1f027654",
+  "blk.24.post_attention_norm.weight": "161f71b6d07628d43e4ae51a4c9088ec6ca2db123a17986a14505d83fdd04dad",
+  "blk.24.post_ffw_norm.weight": "cf1ba692aa683368b02ac413e69b2521b98c69a5274eacbb54165b53bf38a8b2",
+  "blk.25.attn_k.weight": "057a56bd8c8d2b41608d1f71faa3052902152ddf85e47669ad950c1c3e77c33f",
+  "blk.25.attn_norm.weight": "b7179fe02c334da556ddcf6c1b502245639a728c4cbba8b552d8e1df4565ee9d",
+  "blk.25.attn_output.weight": "4fed8b05b08a0ff75ffd022701bbeb52f17b23d09332a1ddcba737244bd0d3b0",
+  "blk.25.attn_q.weight": "c52e99f5d38bf7538d6106a0bbf38ac6dc6296bca9a3f849afa384ea67b4af01",
+  "blk.25.attn_v.weight": "c49c23d8e1cfa6a8eb971eb69942204890c6d7d830dc8774c84b108a80598912",
+  "blk.25.ffn_down.weight": "c08d4dc8412b19fdc870c164b83c341b236ec6fe7bb4a9bcfe0dc100faa20286",
+  "blk.25.ffn_gate.weight": "1a4cb3f36735d59181721471452807903006539e5e1b5ceb4f72d1d7ae134127",
+  "blk.25.ffn_norm.weight": "8fd6bd0dcec5198761525a36992a57c9ec5e9da60a22092839a84ae8c4e87f26",
+  "blk.25.ffn_up.weight": "3a00f39bdd5f31dc5e3b281d2002e1ac4f2475d49a0ac1d7720a25b377dcd04a",
+  "blk.25.post_attention_norm.weight": "e5f31a648612c859b6d21c9ee426e87a86cb1973dfdd86276c767371d9cef5ad",
+  "blk.25.post_ffw_norm.weight": "553c3bd774922c99c2384380a142d019881d30dbf0fe3bf9430dabfb3f6cbd33",
+  "output_norm.weight": "49445c4585ab0a8135717a0bdb1cda4a062a030177d0119561d91542aec5744b"
+}
--- a/convert/tokenizer.go
+++ b/convert/tokenizer.go
@@ -100,8 +100,21 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 		}

 		if template, ok := p["chat_template"]; ok {
-			if err := json.Unmarshal(template, &t.Template); err != nil {
-				return nil, err
+			var s []struct {
+				Name     string `json:"name"`
+				Template string `json:"template"`
+			}
+			if err := json.Unmarshal(template, &t.Template); err == nil {
+				// noop
+			} else if err := json.Unmarshal(template, &s); err == nil {
+				for _, e := range s {
+					if e.Name == "default" {
+						t.Template = e.Template
+						break
+					}
+				}
+			} else {
+				return nil, fmt.Errorf("invalid chat_template: %w", err)
 			}
 		}

@@ -141,7 +154,6 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 }

 type tokenizer struct {
-	Version     string  `json:"version"`
 	AddedTokens []token `json:"added_tokens"`
 	Model       struct {
 		Type   string         `json:"type"`
@@ -239,7 +251,7 @@ func parseVocabulary(fsys fs.FS) (*Vocabulary, error) {
 		return pattern.Func(fsys)
 	}

-	return nil, errors.New("unknown tensor format")
+	return nil, errors.New("unknown tokenizer format")
 }

 type SpecialVocabulary struct {
--- a/convert/tokenizer_test.go
+++ b/convert/tokenizer_test.go
@@ -0,0 +1,208 @@
+package convert
+
+import (
+	"io"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+)
+
+func createTokenizerFS(t *testing.T, dir string, files map[string]io.Reader) fs.FS {
+	t.Helper()
+
+	for k, v := range files {
+		if err := func() error {
+			f, err := os.Create(filepath.Join(dir, k))
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+
+			if _, err := io.Copy(f, v); err != nil {
+				return err
+			}
+
+			return nil
+		}(); err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+	}
+
+	return os.DirFS(dir)
+}
+
+func TestParseTokenizer(t *testing.T) {
+	cases := []struct {
+		name              string
+		fsys              fs.FS
+		specialTokenTypes []string
+		want              *Tokenizer
+	}{
+		{
+			name: "string chat template",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"chat_template": "<default template>"
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{Model: "gpt2"},
+				Pre:        "default",
+				Template:   "<default template>",
+			},
+		},
+		{
+			name: "list chat template",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"chat_template": [
+						{
+							"name": "default",
+							"template": "<default template>"
+						},
+						{
+							"name": "tools",
+							"template": "<tools template>"
+						}
+					]
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{Model: "gpt2"},
+				Pre:        "default",
+				Template:   "<default template>",
+			},
+		},
+		{
+			name: "added tokens",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 999,
+							"content": "<unused999>",
+							"special": false
+						}
+					]
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<unused999>"},
+					Scores: []float32{999},
+					Types:  []int32{4},
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "added tokens overlap vocab",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<pad>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<pad>": 0
+						}
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<pad>"},
+					Scores: []float32{0},
+					Types:  []int32{3},
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "special token types",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"added_tokens": [
+						{
+							"id": 0,
+							"content": "<pad>",
+							"special": true
+						},
+						{
+							"id": 1,
+							"content": "<eos>",
+							"special": true
+						},
+						{
+							"id": 2,
+							"content": "<bos>",
+							"special": true
+						},
+						{
+							"id": 3,
+							"content": "<unk>",
+							"special": true
+						}
+					],
+					"model": {
+						"vocab": {
+							"<pad>": 0,
+							"<eos>": 1,
+							"<bos>": 2,
+							"<unk>": 3
+						}
+					}
+				}`),
+				"tokenizer_config.json": strings.NewReader(`{
+					"add_bos_token": true,
+					"add_eos_token": false,
+					"bos_token": "<bos>",
+					"eos_token": "<eos>",
+					"pad_token": "<pad>",
+					"unk_token": "<unk>"
+				}`),
+			}),
+			specialTokenTypes: []string{"pad", "eos", "bos", "unk"},
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model:  "gpt2",
+					Tokens: []string{"<pad>", "<eos>", "<bos>", "<unk>"},
+					Scores: []float32{0, 1, 2, 3},
+					Types:  []int32{3, 3, 3, 3},
+				},
+				SpecialVocabulary: []*SpecialVocabulary{
+					{Type: "pad", Content: "<pad>", ID: 0, AddToken: false},
+					{Type: "eos", Content: "<eos>", ID: 1, AddToken: false},
+					{Type: "bos", Content: "<bos>", ID: 2, AddToken: true},
+					{Type: "unk", Content: "<unk>", ID: 3, AddToken: false},
+				},
+				Pre: "default",
+			},
+		},
+	}
+
+	for _, tt := range cases {
+		t.Run(tt.name, func(t *testing.T) {
+			tokenizer, err := parseTokenizer(tt.fsys, tt.specialTokenTypes)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			if diff := cmp.Diff(tt.want, tokenizer); diff != "" {
+				t.Errorf("unexpected tokenizer (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
--- a/discover/README.md
+++ b/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.
--- a/discover/amd_common.go
+++ b/discover/amd_common.go
@@ -1,6 +1,6 @@
 //go:build linux || windows

-package gpu
+package discover

 import (
 	"errors"
--- a/discover/amd_hip_windows.go
+++ b/discover/amd_hip_windows.go
@@ -1,4 +1,4 @@
-package gpu
+package discover

 import (
 	"errors"
--- a/discover/amd_linux.go
+++ b/discover/amd_linux.go
@@ -1,10 +1,11 @@
-package gpu
+package discover

 import (
 	"bufio"
 	"errors"
 	"fmt"
 	"io"
+	"io/fs"
 	"log/slog"
 	"os"
 	"path/filepath"
@@ -46,10 +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
@@ -193,13 +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
@@ -269,19 +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",
@@ -299,6 +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
@@ -309,7 +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
 			}
 		}
@@ -319,8 +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
@@ -330,14 +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
@@ -357,9 +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
 	}
-	return resp
+	if err := verifyKFDDriverAccess(); err != nil {
+		err = fmt.Errorf("amdgpu devices detected but permission problems block access: %w", err)
+		slog.Error(err.Error())
+		return nil, err
+	}
+	return resp, nil
 }

 // Quick check for AMD driver so we can skip amdgpu discovery if not present
@@ -455,3 +500,19 @@ func getFreeMemory(usedFile string) (uint64, error) {
 	}
 	return usedMemory, nil
 }
+
+func verifyKFDDriverAccess() error {
+	// Verify we have permissions - either running as root, or we have group access to the driver
+	fd, err := os.OpenFile("/dev/kfd", os.O_RDWR, 0o666)
+	if err != nil {
+		if errors.Is(err, fs.ErrPermission) {
+			return fmt.Errorf("permissions not set up properly.  Either run ollama as root, or add you user account to the render group. %w", err)
+		} else if errors.Is(err, fs.ErrNotExist) {
+			// Container runtime failure?
+			return fmt.Errorf("kfd driver not loaded.  If running in a container, remember to include '--device /dev/kfd --device /dev/dri'")
+		}
+		return fmt.Errorf("failed to check permission on /dev/kfd: %w", err)
+	}
+	fd.Close()
+	return nil
+}
--- a/discover/amd_windows.go
+++ b/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) {
--- a/discover/cpu_common.go
+++ b/discover/cpu_common.go
@@ -1,4 +1,4 @@
-package gpu
+package discover

 import (
 	"os"
--- a/discover/cuda_common.go
+++ b/discover/cuda_common.go
@@ -1,6 +1,6 @@
 //go:build linux || windows

-package gpu
+package discover

 import (
 	"log/slog"
@@ -57,7 +57,7 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}

-	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 {
+	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
 		return "v11"
 	}
 	return "v12"
--- a/discover/gpu.go
+++ b/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
 	}

@@ -93,28 +101,30 @@ func initCudaHandles() *cudaHandles {
 		localAppData := os.Getenv("LOCALAPPDATA")
 		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", CudartMgmtName)}
 	}
-	tmpDir, _ := PayloadsDir()
-	if tmpDir != "" {
-		// TODO - add "payloads" for subprocess
-		cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", CudartMgmtName)}
+	libDir := LibraryDir()
+	if libDir != "" {
+		cudartMgmtPatterns = []string{filepath.Join(libDir, CudartMgmtName)}
 	}
 	cudartMgmtPatterns = append(cudartMgmtPatterns, CudartGlobs...)

 	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
@@ -122,11 +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
@@ -134,6 +147,9 @@ func initCudaHandles() *cudaHandles {
 			cudartLibPath = libPath
 			return cHandles
 		}
+		if err != nil {
+			bootstrapErrors = append(bootstrapErrors, err)
+		}
 	}

 	return cHandles
@@ -144,14 +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
@@ -198,6 +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
@@ -206,27 +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()

@@ -253,10 +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])
@@ -279,6 +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 {
@@ -341,7 +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")
@@ -526,92 +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 {
@@ -669,3 +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,
+	}
+}
--- a/discover/gpu_darwin.go
+++ b/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(),
+	}
+}
--- a/discover/gpu_info.h
+++ b/discover/gpu_info.h
--- a/discover/gpu_info_cudart.c
+++ b/discover/gpu_info_cudart.c
--- a/discover/gpu_info_cudart.h
+++ b/discover/gpu_info_cudart.h
--- a/discover/gpu_info_darwin.h
+++ b/discover/gpu_info_darwin.h
--- a/discover/gpu_info_darwin.m
+++ b/discover/gpu_info_darwin.m
--- a/discover/gpu_info_nvcuda.c
+++ b/discover/gpu_info_nvcuda.c
--- a/discover/gpu_info_nvcuda.h
+++ b/discover/gpu_info_nvcuda.h
--- a/discover/gpu_info_nvml.c
+++ b/discover/gpu_info_nvml.c
--- a/discover/gpu_info_nvml.h
+++ b/discover/gpu_info_nvml.h
--- a/discover/gpu_info_oneapi.c
+++ b/discover/gpu_info_oneapi.c
--- a/discover/gpu_info_oneapi.h
+++ b/discover/gpu_info_oneapi.h
--- a/discover/gpu_linux.go
+++ b/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
+}
--- a/discover/gpu_oneapi.go
+++ b/discover/gpu_oneapi.go
@@ -1,6 +1,6 @@
 //go:build linux || windows

-package gpu
+package discover

 import (
 	"log/slog"
--- a/discover/gpu_test.go
+++ b/discover/gpu_test.go
@@ -1,4 +1,4 @@
-package gpu
+package discover

 import (
 	"runtime"
--- a/discover/gpu_windows.go
+++ b/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
+}
--- a/discover/gpu_windows_test.go
+++ b/discover/gpu_windows_test.go
--- a/discover/types.go
+++ b/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
+}
--- a/docs/api.md
+++ b/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",
+  "model": "llama3.2",
  "prompt": "Why is the sky blue?"
 }'
 ```
@@ -80,7 +80,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "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",
+  "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",
+  "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",
+  "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",
+  "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",
+  "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",
+  "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",
+  "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"
+  "model": "llama3.2"
 }'
 ```

@@ -400,13 +400,40 @@ A single JSON object is returned:

 ```json
 {
-  "model": "llama3",
+  "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",
+  "model": "llama3.2",
  "messages": [
    {
      "role": "user",
@@ -461,7 +488,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "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",
+  "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",
+  "model": "llama3.2",
  "messages": [
    {
      "role": "user",
@@ -509,7 +536,7 @@ curl http://localhost:11434/api/chat -d '{

 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "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",
+  "model": "llama3.2",
  "messages": [
    {
      "role": "user",
@@ -557,7 +584,7 @@ A stream of JSON objects is returned:

 ```json
 {
-  "model": "llama3",
+  "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",
+  "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",
+  "model": "llama3.2",
  "messages": [
    {
      "role": "user",
@@ -647,7 +674,7 @@ curl http://localhost:11434/api/chat -d '{

 ```json
 {
-  "model": "registry.ollama.ai/library/llama3:latest",
+  "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"
+  "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",
+  "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"
+  "name": "llama3.2"
 }'
 ```

--- a/docs/development.md
+++ b/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\`)
--- a/docs/docker.md
+++ b/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
--- a/docs/faq.md
+++ b/docs/faq.md
@@ -32,7 +32,7 @@ When using the API, specify the `num_ctx` parameter:

 ```shell
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama3",
+  "model": "llama3.2",
  "prompt": "Why is the sky blue?",
  "options": {
    "num_ctx": 4096
@@ -194,6 +194,8 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e

 If a different directory needs to be used, set the environment variable `OLLAMA_MODELS` to the chosen directory.

+> Note: on Linux using the standard installer, the `ollama` user needs read and write access to the specified directory. To assign the directory to the `ollama` user run `sudo chown -R ollama:ollama <directory>`.
+
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.

 ## How can I use Ollama in Visual Studio Code?
@@ -230,14 +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")
@@ -245,17 +251,17 @@ The `keep_alive` parameter can be set to:

 For example, to preload a model and leave it in memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": -1}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "keep_alive": -1}'
 ```

 To unload the model and free up memory use:
 ```shell
-curl http://localhost:11434/api/generate -d '{"model": "llama3", "keep_alive": 0}'
+curl http://localhost:11434/api/generate -d '{"model": "llama3.2", "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?

--- a/docs/gpu.md
+++ b/docs/gpu.md
@@ -10,7 +10,7 @@ Check your compute compatibility to see if your card is supported:
 | 9.0                | NVIDIA              | `H100`                                                                                                      |
 | 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080 SUPER` `RTX 4080` `RTX 4070 Ti SUPER` `RTX 4070 Ti` `RTX 4070 SUPER` `RTX 4070` `RTX 4060 Ti` `RTX 4060`  |
 |                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
-| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
+| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060` `RTX 3050 Ti` `RTX 3050`   |
 |                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
 | 8.0                | NVIDIA              | `A100` `A30`                                                                                                |
 | 7.5                | GeForce GTX/RTX     | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060`                                    |
--- a/docs/import.md
+++ b/docs/import.md
@@ -38,7 +38,7 @@ Ollama supports importing adapters based on several different model architecture

 You can create the adapter using a fine tuning framework or tool which can output adapters in the Safetensors format, such as:

-  * Hugging Face [fine tuning framework] (https://huggingface.co/docs/transformers/en/training)
+  * Hugging Face [fine tuning framework](https://huggingface.co/docs/transformers/en/training)
  * [Unsloth](https://github.com/unslothai/unsloth)
  * [MLX](https://github.com/ml-explore/mlx)

--- a/docs/linux.md
+++ b/docs/linux.md
@@ -1,44 +1,59 @@
-# Ollama on Linux
+# Linux

 ## Install

-Install Ollama running this one-liner:
+To install Ollama, run the following command:

->
-
-```bash
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```

-## AMD Radeon GPU support
-
-While AMD has contributed the `amdgpu` driver upstream to the official linux
-kernel source, the version is older and may not support all ROCm features. We
-recommend you install the latest driver from
-https://www.amd.com/en/support/linux-drivers for best support of your Radeon
-GPU.
-
 ## Manual install

-### Download `ollama`
+Download and extract the package:

-Download and extract the Linux package:
-
-```bash
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```

-If you have an AMD GPU, also download and extract the ROCm package into the same location
-```bash
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz | sudo tar zx -C /usr
+Start Ollama:
+
+```shell
+ollama serve
+```
+
+In another terminal, verify that Ollama is running:
+
+```shell
+ollama -v
+```
+
+### AMD GPU install
+
+If you have an AMD GPU, also download and extract the additional ROCm package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64-rocm.tgz -o ollama-linux-amd64-rocm.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64-rocm.tgz
+```
+
+### ARM64 install
+
+Download and extract the ARM64-specific package:
+
+```shell
+curl -L https://ollama.com/download/ollama-linux-arm64.tgz -o ollama-linux-arm64.tgz
+sudo tar -C /usr -xzf ollama-linux-arm64.tgz
 ```

 ### Adding Ollama as a startup service (recommended)

-Create a user for Ollama:
+Create a user and group for Ollama:

-```bash
-sudo useradd -r -s /bin/false -m -d /usr/share/ollama ollama
+```shell
+sudo useradd -r -s /bin/false -U -m -d /usr/share/ollama ollama
+sudo usermod -a -G ollama $(whoami)
 ```

 Create a service file in `/etc/systemd/system/ollama.service`:
@@ -54,6 +69,7 @@ User=ollama
 Group=ollama
 Restart=always
 RestartSec=3
+Environment="PATH=$PATH"

 [Install]
 WantedBy=default.target
@@ -61,46 +77,54 @@ WantedBy=default.target

 Then start the service:

-```bash
+```shell
 sudo systemctl daemon-reload
 sudo systemctl enable ollama
 ```

-### Install CUDA drivers (optional – for Nvidia GPUs)
+### Install CUDA drivers (optional)

 [Download and install](https://developer.nvidia.com/cuda-downloads) CUDA.

 Verify that the drivers are installed by running the following command, which should print details about your GPU:

-```bash
+```shell
 nvidia-smi
 ```

-### Install ROCm (optional - for Radeon GPUs)
-[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html)
+### Install AMD ROCm drivers (optional)

-Make sure to install ROCm v6
+[Download and Install](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html) ROCm v6.

 ### Start Ollama

-Start Ollama using `systemd`:
+Start Ollama and verify it is running:

-```bash
+```shell
 sudo systemctl start ollama
+sudo systemctl status ollama
 ```

-## Update
+> [!NOTE]
+> While AMD has contributed the `amdgpu` driver upstream to the official linux
+> kernel source, the version is older and may not support all ROCm features. We
+> recommend you install the latest driver from
+> https://www.amd.com/en/support/linux-drivers for best support of your Radeon
+> GPU.

-Update ollama by running the install script again:
+## Updating

-```bash
+Update Ollama by running the install script again:
+
+```shell
 curl -fsSL https://ollama.com/install.sh | sh
 ```

-Or by downloading the ollama binary:
+Or by re-downloading Ollama:

-```bash
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz | sudo tar zx -C /usr
+```shell
+curl -L https://ollama.com/download/ollama-linux-amd64.tgz -o ollama-linux-amd64.tgz
+sudo tar -C /usr -xzf ollama-linux-amd64.tgz
 ```

 ## Installing specific versions
@@ -109,15 +133,15 @@ Use `OLLAMA_VERSION` environment variable with the install script to install a s

 For example:

-```
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.1.32 sh
+```shell
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.3.9 sh
 ```

 ## Viewing logs

 To view logs of Ollama running as a startup service, run:

-```bash
+```shell
 journalctl -e -u ollama
 ```

@@ -125,7 +149,7 @@ journalctl -e -u ollama

 Remove the ollama service:

-```bash
+```shell
 sudo systemctl stop ollama
 sudo systemctl disable ollama
 sudo rm /etc/systemd/system/ollama.service
@@ -133,13 +157,13 @@ sudo rm /etc/systemd/system/ollama.service

 Remove the ollama binary from your bin directory (either `/usr/local/bin`, `/usr/bin`, or `/bin`):

-```bash
+```shell
 sudo rm $(which ollama)
 ```

 Remove the downloaded models and Ollama service user and group:

-```bash
+```shell
 sudo rm -r /usr/share/ollama
 sudo userdel ollama
 sudo groupdel ollama
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@@ -11,7 +11,7 @@ A model file is the blueprint to create and share models with Ollama.
 - [Examples](#examples)
 - [Instructions](#instructions)
  - [FROM (Required)](#from-required)
-    - [Build from llama3.1](#build-from-llama31)
+    - [Build from existing model](#build-from-existing-model)
    - [Build from a Safetensors model](#build-from-a-safetensors-model)
    - [Build from a GGUF file](#build-from-a-gguf-file)
  - [PARAMETER](#parameter)
@@ -50,7 +50,7 @@ INSTRUCTION arguments
 An example of a `Modelfile` creating a mario blueprint:

 ```modelfile
-FROM llama3
+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
+  > 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:latest
+  # FROM llama3.2:latest
  FROM /Users/pdevine/.ollama/models/blobs/sha256-00e1317cbf74d901080d7100f57580ba8dd8de57203072dc6f668324ba545f29
  TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

@@ -100,10 +100,10 @@ The `FROM` instruction defines the base model to use when creating a model.
 FROM <model name>:<tag>
 ```

-#### Build from llama3.1
+#### Build from existing model

 ```modelfile
-FROM llama3.1
+FROM llama3.2
 ```

 A list of available base models:
@@ -128,10 +128,10 @@ Currently supported model architectures:
 #### Build from a GGUF file

 ```modelfile
-FROM ./ollama-model.bin
+FROM ./ollama-model.gguf
 ```

-The GGUF bin file location should be specified as an absolute path or relative to the `Modelfile` location.
+The GGUF file location should be specified as an absolute path or relative to the `Modelfile` location.


 ### PARAMETER
@@ -208,7 +208,7 @@ Currently supported Safetensor adapters:
 #### GGUF adapter

 ```modelfile
-ADAPTER ./ollama-lora.bin
+ADAPTER ./ollama-lora.gguf
 ```

 ### LICENSE
--- a/docs/openai.md
+++ b/docs/openai.md
@@ -25,7 +25,7 @@ chat_completion = client.chat.completions.create(
            'content': 'Say this is a test',
        }
    ],
-    model='llama3',
+    model='llama3.2',
 )

 response = client.chat.completions.create(
@@ -46,13 +46,13 @@ response = client.chat.completions.create(
 )

 completion = client.completions.create(
-    model="llama3",
+    model="llama3.2",
    prompt="Say this is a test",
 )

 list_completion = client.models.list()

-model = client.models.retrieve("llama3")
+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',
+    model: 'llama3.2',
 })

 const response = await openai.chat.completions.create({
@@ -94,13 +94,13 @@ const response = await openai.chat.completions.create({
 })

 const completion = await openai.completions.create({
-    model: "llama3",
+    model: "llama3.2",
    prompt: "Say this is a test.",
 })

 const listCompletion = await openai.models.list()

-const model = await openai.models.retrieve("llama3")
+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",
+        "model": "llama3.2",
        "messages": [
            {
                "role": "system",
@@ -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",
+        "model": "llama3.2",
        "prompt": "Say this is a test"
    }'

 curl http://localhost:11434/v1/models

-curl http://localhost:11434/v1/models/llama3
+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
+ollama pull llama3.2
 ```

 ### Default model names
@@ -282,7 +282,7 @@ ollama pull llama3
 For tooling that relies on default OpenAI model names such as `gpt-3.5-turbo`, use `ollama cp` to copy an existing model name to a temporary name:

 ```
-ollama cp llama3 gpt-3.5-turbo
+ollama cp llama3.2 gpt-3.5-turbo
 ```

 Afterwards, this new model name can be specified the `model` field:
@@ -300,3 +300,28 @@ curl http://localhost:11434/v1/chat/completions \
        ]
    }'
 ```
+
+### Setting the context size
+
+The OpenAI API does not have a way of setting the context size for a model. If you need to change the context size, create a `Modelfile` which looks like:
+
+```modelfile
+FROM <some model>
+PARAMETER num_ctx <context size>
+```
+
+Use the `ollama create mymodel` command to create a new model with the updated context size. Call the API with the updated model name:
+
+```shell
+curl http://localhost:11434/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "mymodel",
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
--- a/docs/template.md
+++ b/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
+FROM llama3.2

 TEMPLATE """{{- if .System }}<|start_header_id|>system<|end_header_id|>

--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@@ -91,6 +91,17 @@ If none of those resolve the problem, gather additional information and file an
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`


+## AMD GPU Discovery
+
+On linux, AMD GPU access typically requires `video` and/or `render` group membership to access the `/dev/kfd` device.  If permissions are not set up correctly, Ollama will detect this and report an error in the server log.
+
+When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -ld /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the group assignments on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.
+
+If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
+- `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
+- `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported
+- Check dmesg for any errors from amdgpu or kfd drivers `sudo dmesg | grep -i amdgpu` and `sudo dmesg | grep -i kfd`
+
 ## Windows Terminal Errors

 Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
--- a/docs/tutorials/langchainjs.md
+++ b/docs/tutorials/langchainjs.md
@@ -15,7 +15,7 @@ import { Ollama } from "@langchain/community/llms/ollama";

 const ollama = new Ollama({
  baseUrl: "http://localhost:11434",
-  model: "llama3.1",
+  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
--- a/docs/windows.md
+++ b/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", "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
@@ -48,6 +48,9 @@ the explorer window by hitting `<cmd>+R` and type in:
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
 - `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories

+## Uninstall
+
+The Ollama Windows installer registers an Uninstaller application.  Under `Add or remove programs` in Windows Settings, you can uninstall Ollama.

 ## Standalone CLI

--- a/envconfig/config.go
+++ b/envconfig/config.go
@@ -30,9 +30,7 @@ func Host() *url.URL {
 		defaultPort = "443"
 	}

-	// trim trailing slashes
-	hostport = strings.TrimRight(hostport, "/")
-
+	hostport, path, _ := strings.Cut(hostport, "/")
 	host, port, err := net.SplitHostPort(hostport)
 	if err != nil {
 		host, port = "127.0.0.1", defaultPort
@@ -45,15 +43,13 @@ func Host() *url.URL {

 	if n, err := strconv.ParseInt(port, 10, 32); err != nil || n > 65535 || n < 0 {
 		slog.Warn("invalid port, using default", "port", port, "default", defaultPort)
-		return &url.URL{
-			Scheme: scheme,
-			Host:   net.JoinHostPort(host, defaultPort),
-		}
+		port = defaultPort
 	}

 	return &url.URL{
 		Scheme: scheme,
 		Host:   net.JoinHostPort(host, port),
+		Path:   path,
 	}
 }

@@ -116,6 +112,26 @@ func KeepAlive() (keepAlive time.Duration) {
 	return keepAlive
 }

+// LoadTimeout returns the duration for stall detection during model loads. LoadTimeout can be configured via the OLLAMA_LOAD_TIMEOUT environment variable.
+// Zero or Negative values are treated as infinite.
+// Default is 5 minutes.
+func LoadTimeout() (loadTimeout time.Duration) {
+	loadTimeout = 5 * time.Minute
+	if s := Var("OLLAMA_LOAD_TIMEOUT"); s != "" {
+		if d, err := time.ParseDuration(s); err == nil {
+			loadTimeout = d
+		} else if n, err := strconv.ParseInt(s, 10, 64); err == nil {
+			loadTimeout = time.Duration(n) * time.Second
+		}
+	}
+
+	if loadTimeout <= 0 {
+		return time.Duration(math.MaxInt64)
+	}
+
+	return loadTimeout
+}
+
 func Bool(k string) func() bool {
 	return func() bool {
 		if s := Var(k); s != "" {
@@ -144,6 +160,8 @@ 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 {
@@ -163,53 +181,6 @@ var (
 	HsaOverrideGfxVersion = String("HSA_OVERRIDE_GFX_VERSION")
 )

-func RunnersDir() (p string) {
-	if p := Var("OLLAMA_RUNNERS_DIR"); p != "" {
-		return p
-	}
-
-	if runtime.GOOS != "windows" {
-		return
-	}
-
-	defer func() {
-		if p == "" {
-			slog.Error("unable to locate llm runner directory. Set OLLAMA_RUNNERS_DIR to the location of 'ollama/runners'")
-		}
-	}()
-
-	// On Windows we do not carry the payloads inside the main executable
-	exe, err := os.Executable()
-	if err != nil {
-		return
-	}
-
-	cwd, err := os.Getwd()
-	if err != nil {
-		return
-	}
-
-	var paths []string
-	for _, root := range []string{filepath.Dir(exe), filepath.Join(filepath.Dir(exe), LibRelativeToExe()), cwd} {
-		paths = append(paths,
-			root,
-			filepath.Join(root, runtime.GOOS+"-"+runtime.GOARCH),
-			filepath.Join(root, "dist", runtime.GOOS+"-"+runtime.GOARCH),
-		)
-	}
-
-	// Try a few variations to improve developer experience when building from source in the local tree
-	for _, path := range paths {
-		candidate := filepath.Join(path, "lib", "ollama", "runners")
-		if _, err := os.Stat(candidate); err == nil {
-			p = candidate
-			break
-		}
-	}
-
-	return p
-}
-
 func Uint(key string, defaultValue uint) func() uint {
 	return func() uint {
 		if s := Var(key); s != "" {
@@ -235,6 +206,23 @@ var (
 	MaxVRAM = Uint("OLLAMA_MAX_VRAM", 0)
 )

+func Uint64(key string, defaultValue uint64) func() uint64 {
+	return func() uint64 {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseUint(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return n
+			}
+		}
+
+		return defaultValue
+	}
+}
+
+// Set aside VRAM per GPU
+var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
+
 type EnvVar struct {
 	Name        string
 	Value       any
@@ -245,9 +233,11 @@ func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
 		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
 		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary(), "Set LLM library to bypass autodetection"},
+		"OLLAMA_LOAD_TIMEOUT":      {"OLLAMA_LOAD_TIMEOUT", LoadTimeout(), "How long to allow model loads to stall before giving up (default \"5m\")"},
 		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners(), "Maximum number of loaded models per GPU"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueue(), "Maximum number of queued requests"},
 		"OLLAMA_MODELS":            {"OLLAMA_MODELS", Models(), "The path to the models directory"},
@@ -255,10 +245,23 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
 		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
-		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir(), "Location for runners"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_TMPDIR":            {"OLLAMA_TMPDIR", TmpDir(), "Location for temporary files"},
+		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
+
+		// Informational
+		"HTTP_PROXY":  {"HTTP_PROXY", String("HTTP_PROXY")(), "HTTP proxy"},
+		"HTTPS_PROXY": {"HTTPS_PROXY", String("HTTPS_PROXY")(), "HTTPS proxy"},
+		"NO_PROXY":    {"NO_PROXY", String("NO_PROXY")(), "No proxy"},
 	}
+
+	if runtime.GOOS != "windows" {
+		// Windows environment variables are case-insensitive so there's no need to duplicate them
+		ret["http_proxy"] = EnvVar{"http_proxy", String("http_proxy")(), "HTTP proxy"}
+		ret["https_proxy"] = EnvVar{"https_proxy", String("https_proxy")(), "HTTPS proxy"}
+		ret["no_proxy"] = EnvVar{"no_proxy", String("no_proxy")(), "No proxy"}
+	}
+
 	if runtime.GOOS != "darwin" {
 		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices(), "Set which NVIDIA devices are visible"}
 		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible"}
@@ -267,6 +270,7 @@ func AsMap() map[string]EnvVar {
 		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion(), "Override the gfx used for all detected AMD GPUs"}
 		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGPU(), "Enable experimental Intel GPU detection"}
 	}
+
 	return ret
 }

--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@@ -13,34 +13,35 @@ func TestHost(t *testing.T) {
 		value  string
 		expect string
 	}{
-		"empty":               {"", "127.0.0.1:11434"},
-		"only address":        {"1.2.3.4", "1.2.3.4:11434"},
-		"only port":           {":1234", ":1234"},
-		"address and port":    {"1.2.3.4:1234", "1.2.3.4:1234"},
-		"hostname":            {"example.com", "example.com:11434"},
-		"hostname and port":   {"example.com:1234", "example.com:1234"},
-		"zero port":           {":0", ":0"},
-		"too large port":      {":66000", ":11434"},
-		"too small port":      {":-1", ":11434"},
-		"ipv6 localhost":      {"[::1]", "[::1]:11434"},
-		"ipv6 world open":     {"[::]", "[::]:11434"},
-		"ipv6 no brackets":    {"::1", "[::1]:11434"},
-		"ipv6 + port":         {"[::1]:1337", "[::1]:1337"},
-		"extra space":         {" 1.2.3.4 ", "1.2.3.4:11434"},
-		"extra quotes":        {"\"1.2.3.4\"", "1.2.3.4:11434"},
-		"extra space+quotes":  {" \" 1.2.3.4 \" ", "1.2.3.4:11434"},
-		"extra single quotes": {"'1.2.3.4'", "1.2.3.4:11434"},
-		"http":                {"http://1.2.3.4", "1.2.3.4:80"},
-		"http port":           {"http://1.2.3.4:4321", "1.2.3.4:4321"},
-		"https":               {"https://1.2.3.4", "1.2.3.4:443"},
-		"https port":          {"https://1.2.3.4:4321", "1.2.3.4:4321"},
+		"empty":               {"", "http://127.0.0.1:11434"},
+		"only address":        {"1.2.3.4", "http://1.2.3.4:11434"},
+		"only port":           {":1234", "http://:1234"},
+		"address and port":    {"1.2.3.4:1234", "http://1.2.3.4:1234"},
+		"hostname":            {"example.com", "http://example.com:11434"},
+		"hostname and port":   {"example.com:1234", "http://example.com:1234"},
+		"zero port":           {":0", "http://:0"},
+		"too large port":      {":66000", "http://:11434"},
+		"too small port":      {":-1", "http://:11434"},
+		"ipv6 localhost":      {"[::1]", "http://[::1]:11434"},
+		"ipv6 world open":     {"[::]", "http://[::]:11434"},
+		"ipv6 no brackets":    {"::1", "http://[::1]:11434"},
+		"ipv6 + port":         {"[::1]:1337", "http://[::1]:1337"},
+		"extra space":         {" 1.2.3.4 ", "http://1.2.3.4:11434"},
+		"extra quotes":        {"\"1.2.3.4\"", "http://1.2.3.4:11434"},
+		"extra space+quotes":  {" \" 1.2.3.4 \" ", "http://1.2.3.4:11434"},
+		"extra single quotes": {"'1.2.3.4'", "http://1.2.3.4:11434"},
+		"http":                {"http://1.2.3.4", "http://1.2.3.4:80"},
+		"http port":           {"http://1.2.3.4:4321", "http://1.2.3.4:4321"},
+		"https":               {"https://1.2.3.4", "https://1.2.3.4:443"},
+		"https port":          {"https://1.2.3.4:4321", "https://1.2.3.4:4321"},
+		"proxy path":          {"https://example.com/ollama", "https://example.com:443/ollama"},
 	}

 	for name, tt := range cases {
 		t.Run(name, func(t *testing.T) {
 			t.Setenv("OLLAMA_HOST", tt.value)
-			if host := Host(); host.Host != tt.expect {
-				t.Errorf("%s: expected %s, got %s", name, tt.expect, host.Host)
+			if host := Host(); host.String() != tt.expect {
+				t.Errorf("%s: expected %s, got %s", name, tt.expect, host.String())
 			}
 		})
 	}
@@ -214,6 +215,40 @@ func TestKeepAlive(t *testing.T) {
 	}
 }

+func TestLoadTimeout(t *testing.T) {
+	defaultTimeout := 5 * time.Minute
+	cases := map[string]time.Duration{
+		"":       defaultTimeout,
+		"1s":     time.Second,
+		"1m":     time.Minute,
+		"1h":     time.Hour,
+		"5m0s":   defaultTimeout,
+		"1h2m3s": 1*time.Hour + 2*time.Minute + 3*time.Second,
+		"0":      time.Duration(math.MaxInt64),
+		"60":     60 * time.Second,
+		"120":    2 * time.Minute,
+		"3600":   time.Hour,
+		"-0":     time.Duration(math.MaxInt64),
+		"-1":     time.Duration(math.MaxInt64),
+		"-1m":    time.Duration(math.MaxInt64),
+		// invalid values
+		" ":   defaultTimeout,
+		"???": defaultTimeout,
+		"1d":  defaultTimeout,
+		"1y":  defaultTimeout,
+		"1w":  defaultTimeout,
+	}
+
+	for tt, expect := range cases {
+		t.Run(tt, func(t *testing.T) {
+			t.Setenv("OLLAMA_LOAD_TIMEOUT", tt)
+			if actual := LoadTimeout(); actual != expect {
+				t.Errorf("%s: expected %s, got %s", tt, expect, actual)
+			}
+		})
+	}
+}
+
 func TestVar(t *testing.T) {
 	cases := map[string]string{
 		"value":       "value",
--- a/examples/go-chat/main.go
+++ b/examples/go-chat/main.go
@@ -35,7 +35,7 @@ func main() {

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

--- a/examples/langchain-python-rag-document/README.md
+++ b/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.
--- a/examples/langchain-python-rag-document/main.py
+++ b/examples/langchain-python-rag-document/main.py
@@ -51,7 +51,7 @@ while True:
        template=template,
    )

-    llm = Ollama(model="llama3.1", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
+    llm = Ollama(model="llama3.2", callback_manager=CallbackManager([StreamingStdOutCallbackHandler()]))
    qa_chain = RetrievalQA.from_chain_type(
        llm,
        retriever=vectorstore.as_retriever(),
--- a/examples/langchain-python-rag-privategpt/requirements.txt
+++ b/examples/langchain-python-rag-privategpt/requirements.txt
@@ -1,6 +1,6 @@
 langchain==0.0.274
 gpt4all==1.0.8
-chromadb==0.4.7
+chromadb==0.5.0
 llama-cpp-python==0.1.81
 urllib3==2.0.4
 PyMuPDF==1.23.5
@@ -12,4 +12,4 @@ pandoc==2.3
 pypandoc==1.11
 tqdm==4.66.1
 sentence_transformers==2.2.2
-numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
+numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability
--- a/examples/langchain-python-rag-websummary/README.md
+++ b/examples/langchain-python-rag-websummary/README.md
@@ -4,10 +4,10 @@ This example summarizes the website, [https://ollama.com/blog/run-llama2-uncenso

 ## Running the Example

-1. Ensure you have the `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.
--- a/examples/langchain-python-rag-websummary/main.py
+++ b/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)
--- a/examples/langchain-python-simple/README.md
+++ b/examples/langchain-python-simple/README.md
@@ -4,10 +4,10 @@ This example is a basic "hello world" of using LangChain with Ollama.

 ## Running the Example

-1. Ensure you have the `llama3.1` 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.
--- a/examples/langchain-python-simple/main.py
+++ b/examples/langchain-python-simple/main.py
@@ -1,6 +1,6 @@
 from langchain.llms import Ollama

 input = input("What is your question?")
-llm = Ollama(model="llama3.1")
+llm = Ollama(model="llama3.2")
 res = llm.predict(input)
 print (res)
--- a/examples/modelfile-mario/Modelfile
+++ b/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.
--- a/examples/modelfile-mario/readme.md
+++ b/examples/modelfile-mario/readme.md
@@ -2,12 +2,12 @@

 # Example character: Mario

-This example shows how to create a basic character using Llama3.1 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.
--- a/examples/python-grounded-factuality-rag-check/README.md
+++ b/examples/python-grounded-factuality-rag-check/README.md
@@ -0,0 +1,93 @@
+# RAG Hallucination Checker using Bespoke-Minicheck
+
+This example allows the user to ask questions related to a document, which can be specified via an article url. Relevant chunks are retreived from the document and given to `llama3.2` 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".
--- a/examples/python-grounded-factuality-rag-check/main.py
+++ b/examples/python-grounded-factuality-rag-check/main.py
@@ -0,0 +1,137 @@
+import ollama
+import warnings
+from mattsollamatools import chunker
+from newspaper import Article
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+import nltk
+
+warnings.filterwarnings(
+    "ignore", category=FutureWarning, module="transformers.tokenization_utils_base"
+)
+nltk.download("punkt_tab", quiet=True)
+
+
+def getArticleText(url):
+    """Gets the text of an article from a URL.
+
+    Often there are a bunch of ads and menus on pages for a news article.
+    This uses newspaper3k to get just the text of just the article.
+    """
+    article = Article(url)
+    article.download()
+    article.parse()
+    return article.text
+
+
+def knn_search(question_embedding, embeddings, k=5):
+    """Performs K-nearest neighbors (KNN) search"""
+    X = np.array(
+        [item["embedding"] for article in embeddings for item in article["embeddings"]]
+    )
+    source_texts = [
+        item["source"] for article in embeddings for item in article["embeddings"]
+    ]
+
+    # Fit a KNN model on the embeddings
+    knn = NearestNeighbors(n_neighbors=k, metric="cosine")
+    knn.fit(X)
+
+    # Find the indices and distances of the k-nearest neighbors.
+    _, indices = knn.kneighbors(question_embedding, n_neighbors=k)
+
+    # Get the indices and source texts of the best matches
+    best_matches = [(indices[0][i], source_texts[indices[0][i]]) for i in range(k)]
+
+    return best_matches
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+if __name__ == "__main__":
+    allEmbeddings = []
+    default_url = "https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt"
+    user_input = input(
+        "Enter the URL of an article you want to chat with, or press Enter for default example: "
+    )
+    article_url = user_input.strip() if user_input.strip() else default_url
+    article = {}
+    article["embeddings"] = []
+    article["url"] = article_url
+    text = getArticleText(article_url)
+    chunks = chunker(text)
+
+    # Embed (batch) chunks using ollama
+    embeddings = ollama.embed(model="all-minilm", input=chunks)["embeddings"]
+
+    for chunk, embedding in zip(chunks, embeddings):
+        item = {}
+        item["source"] = chunk
+        item["embedding"] = embedding
+        item["sourcelength"] = len(chunk)
+        article["embeddings"].append(item)
+
+    allEmbeddings.append(article)
+
+    print(f"\nLoaded, chunked, and embedded text from {article_url}.\n")
+
+    while True:
+        # Input a question from the user
+        # For example, "Who is the chief research officer?"
+        question = input("Enter your question or type quit: ")
+
+        if question.lower() == "quit":
+            break
+
+        # Embed the user's question using ollama.embed
+        question_embedding = ollama.embed(model="all-minilm", input=question)[
+            "embeddings"
+        ]
+
+        # Perform KNN search to find the best matches (indices and source text)
+        best_matches = knn_search(question_embedding, allEmbeddings, k=4)
+
+        sourcetext = "\n\n".join([source_text for (_, source_text) in best_matches])
+
+        print(f"\nRetrieved chunks: \n{sourcetext}\n")
+
+        # Give the retreived chunks and question to the chat model
+        system_prompt = f"Only use the following information to answer the question. Do not use anything else: {sourcetext}"
+
+        ollama_response = ollama.generate(
+            model="llama3.2",
+            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"
+                )
--- a/examples/python-grounded-factuality-rag-check/requirements.txt
+++ b/examples/python-grounded-factuality-rag-check/requirements.txt
@@ -0,0 +1,8 @@
+ollama
+lxml==5.3.0
+lxml_html_clean==0.2.2
+mattsollamatools==0.0.25
+newspaper3k==0.2.8
+nltk==3.9.1
+numpy==1.26.4
+scikit-learn==1.5.2
--- a/examples/python-grounded-factuality-simple-check/main.py
+++ b/examples/python-grounded-factuality-simple-check/main.py
@@ -0,0 +1,53 @@
+"""Simple example to demonstrate how to use the bespoke-minicheck model."""
+
+import ollama
+
+# NOTE: ollama must be running for this to work, start the ollama app or run `ollama serve`
+
+
+def check(document, claim):
+    """Checks if the claim is supported by the document by calling bespoke-minicheck.
+
+    Returns Yes/yes if the claim is supported by the document, No/no otherwise.
+    Support for logits will be added in the future.
+
+    bespoke-minicheck's system prompt is defined as:
+      'Determine whether the provided claim is consistent with the corresponding
+      document. Consistency in this context implies that all information presented in the claim
+      is substantiated by the document. If not, it should be considered inconsistent. Please
+      assess the claim's consistency with the document by responding with either "Yes" or "No".'
+
+    bespoke-minicheck's user prompt is defined as:
+      "Document: {document}\nClaim: {claim}"
+    """
+    prompt = f"Document: {document}\nClaim: {claim}"
+    response = ollama.generate(
+        model="bespoke-minicheck", prompt=prompt, options={"num_predict": 2, "temperature": 0.0}
+    )
+    return response["response"].strip()
+
+
+def get_user_input(prompt):
+    user_input = input(prompt)
+    if not user_input:
+        exit()
+    print()
+    return user_input
+
+
+def main():
+    while True:
+        # Get a document from the user (e.g. "Ryan likes running and biking.")
+        document = get_user_input("Enter a document: ")
+        # Get a claim from the user (e.g. "Ryan likes to run.")
+        claim = get_user_input("Enter a claim: ")
+        # Check if the claim is supported by the document
+        grounded_factuality_check = check(document, claim)
+        print(
+            f"Is the claim supported by the document according to bespoke-minicheck? {grounded_factuality_check}"
+        )
+        print("\n\n")
+
+
+if __name__ == "__main__":
+    main()
--- a/examples/python-grounded-factuality-simple-check/readme.md
+++ b/examples/python-grounded-factuality-simple-check/readme.md
@@ -0,0 +1,54 @@
+# Simple Bespoke-Minicheck Example
+
+`bespoke-minicheck` is a model for checking if a claim is supported by a document. It is used through the **generate** endpoint, which is called in this example with a `prompt` that includes the expected formatting of the user input. 
+
+## Running the Example
+
+1. Ensure you have the `bespoke-minicheck` model installed:
+
+   ```bash
+   ollama pull bespoke-minicheck
+   ```
+
+2. Install the dependencies:
+
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. Run the program:
+
+   ```bash
+   python main.py
+   ```
+
+4. Enter a document and a claim when prompted:
+
+   ```bash
+   Enter a document: Roses are red.
+
+   Enter a claim: Roses are blue. 
+   ```
+
+   The claim and document are then given to the `bespoke-minicheck` as inputs, which then generates a response (Yes or No) on whether the claim is supported by the document.
+
+   ```bash
+   Is the claim supported by the document according to bespoke-minicheck? No
+   ```
+
+## More Examples
+
+Document ([source](https://en.wikipedia.org/wiki/Apple_I)): 
+> The Apple Computer 1 (Apple-1[a]), later known predominantly as the Apple I(written with a Roman numeral),[b] is an 8-bit motherboard-only personal computer designed by Steve Wozniak[5][6] and released by the Apple Computer Company (now Apple Inc.) in 1976. The company was initially formed to sell the Apple I – its first product – and would later become the world's largest technology company.[7] The idea of starting a company and selling the computer came from Wozniak's friend and Apple co-founder Steve Jobs.[8][9] One of the main innovations of the Apple I was that it included video display terminal circuitry on its circuit board, allowing it to connect to a low-cost composite video monitor or television, instead of an expensive computer terminal, compared to most existing computers at the time.
+
+Claim: 
+>The Apple I is a 16-bit computer.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **No**
+
+Claim: 
+>Apple was originally called the Apple Computer Company.
+
+Expected output:
+>Is the claim supported by the document according to bespoke-minicheck? **Yes**
--- a/examples/python-grounded-factuality-simple-check/requirements.txt
+++ b/examples/python-grounded-factuality-simple-check/requirements.txt
@@ -0,0 +1 @@
+ollama
--- a/examples/python-json-datagenerator/predefinedschema.py
+++ b/examples/python-json-datagenerator/predefinedschema.py
@@ -2,7 +2,7 @@ import requests
 import json
 import random

-model = "llama3.1"
+model = "llama3.2"
 template = {
  "firstName": "",
  "lastName": "",
--- a/examples/python-json-datagenerator/randomaddresses.py
+++ b/examples/python-json-datagenerator/randomaddresses.py
@@ -12,7 +12,7 @@ countries = [
    "France",
 ]
 country = random.choice(countries)
-model = "llama3.1"
+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."

--- a/examples/python-json-datagenerator/readme.md
+++ b/examples/python-json-datagenerator/readme.md
@@ -6,10 +6,10 @@ There are two python scripts in this example. `randomaddresses.py` generates ran

 ## Running the Example

-1. Ensure you have the `llama3.1` 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.
--- a/examples/python-loganalysis/Modelfile
+++ b/examples/python-loganalysis/Modelfile
@@ -4,5 +4,5 @@ SYSTEM """
 You are a log file analyzer. You will receive a set of lines from a log file for some software application, find the errors and other interesting aspects of the logs, and explain them so a new user can understand what they mean. If there are any steps they can do to resolve them, list the steps in your answer.
 """

-PARAMETER TEMPERATURE 0.3
+PARAMETER temperature 0.3

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