Merge branch 'ollama:main' into main

2024-05-03 10:48:20 +08:00 · 2024-05-03 10:48:20 +08:00 · 76c3108984
commit 76c3108984
parent e9cb6f531b e9ae607ece
170 changed files with 48499 additions and 5327 deletions
--- a/.gitattributes
+++ b/.gitattributes
@ -0,0 +1 @@
+llm/ext_server/* linguist-vendored
--- a/.github/ISSUE_TEMPLATE/10_bug_report.yml
+++ b/.github/ISSUE_TEMPLATE/10_bug_report.yml
@ -0,0 +1,60 @@
+name: Bug report
+labels: [bug]
+description: Something isn't working right.
+body:
+  - type: textarea
+    id: description
+    attributes:
+      label: What is the issue?
+      description: What happened? What did you expect to happen?
+    validations:
+      required: true
+  - type: dropdown
+    id: os
+    attributes:
+      label: OS
+      description: Which operating system are you using?
+      multiple: true
+      options:
+        - Linux
+        - macOS
+        - Windows
+        - Docker
+        - WSL2
+    validations:
+      required: false
+  - type: dropdown
+    id: gpu
+    attributes:
+      label: GPU
+      description: Which GPU are you using?
+      multiple: true
+      options:
+        - Nvidia
+        - AMD
+        - Intel
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: dropdown
+    id: cpu
+    attributes:
+      label: CPU
+      description: Which CPU are you using?
+      multiple: true
+      options:
+        - Intel
+        - AMD
+        - Apple
+        - Other
+    validations:
+      required: false
+  - type: input
+    id: version
+    attributes:
+      label: Ollama version
+      description: What version of Ollama are you using? (`ollama --version`)
+      placeholder: e.g., 0.1.32
+    validations:
+      required: false
--- a/.github/ISSUE_TEMPLATE/20_feature_request.md
+++ b/.github/ISSUE_TEMPLATE/20_feature_request.md
@ -0,0 +1,6 @@
+---
+name: Feature request
+about: Request a new feature
+labels: feature request
+---
+
--- a/.github/ISSUE_TEMPLATE/30_model_request.md
+++ b/.github/ISSUE_TEMPLATE/30_model_request.md
@ -0,0 +1,5 @@
+---
+name: Model request
+about: Request support for a new model to be added to Ollama
+labels: model request
+---
--- a/.github/ISSUE_TEMPLATE/config.yml
+++ b/.github/ISSUE_TEMPLATE/config.yml
@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Help
+    url: https://discord.com/invite/ollama
+    about: Please join our Discord server for help using Ollama
+  - name: Troubleshooting
+    url: https://github.com/ollama/ollama/blob/main/docs/faq.md#faq
+    about: See the FAQ for common issues and solutions
--- a/.github/workflows/latest.yaml
+++ b/.github/workflows/latest.yaml
@ -0,0 +1,24 @@
+name: latest
+
+on:
+  release:
+    types: [released]
+
+jobs:
+  update-latest:
+    environment: release
+    runs-on: linux
+    steps:
+      - uses: actions/checkout@v4
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - name: Tag images as latest
+        env:
+          PUSH: "1"
+        shell: bash
+        run: |
+          export "VERSION=${GITHUB_REF_NAME#v}"
+          ./scripts/tag_latest.sh
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -0,0 +1,473 @@
+name: release
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  # Full build of the Mac assets
+  build-darwin:
+    runs-on: macos-12
+    environment: release
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: |
+          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
+      - name: key
+        env:
+          MACOS_SIGNING_KEY: ${{ secrets.MACOS_SIGNING_KEY }}
+          MACOS_SIGNING_KEY_PASSWORD: ${{ secrets.MACOS_SIGNING_KEY_PASSWORD }}
+        run: |
+          echo $MACOS_SIGNING_KEY | base64 --decode > certificate.p12
+          security create-keychain -p password build.keychain
+          security default-keychain -s build.keychain
+          security unlock-keychain -p password build.keychain
+          security import certificate.p12 -k build.keychain -P $MACOS_SIGNING_KEY_PASSWORD -T /usr/bin/codesign
+          security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k password build.keychain
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: Build Darwin
+        env:
+          APPLE_IDENTITY: ${{ secrets.APPLE_IDENTITY }}
+          APPLE_PASSWORD: ${{ secrets.APPLE_PASSWORD }}
+          APPLE_TEAM_ID: ${{ vars.APPLE_TEAM_ID }}
+          APPLE_ID: ${{ vars.APPLE_ID }}
+          SDKROOT: /Applications/Xcode_13.4.1.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
+          DEVELOPER_DIR: /Applications/Xcode_13.4.1.app/Contents/Developer
+        run: |
+          ./scripts/build_darwin.sh
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-darwin
+          path: |
+            dist/*arwin*
+            !dist/*-cov
+
+  # Windows builds take a long time to both install the dependencies and build, so parallelize
+  # CPU generation step
+  generate-windows-cpu:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$env:PATH"
+          go generate -x ./...
+        name: go generate
+      - uses: actions/upload-artifact@v4
+        with:
+          name: generate-windows-cpu
+          path: |
+            llm/build/**/bin/*
+            llm/build/**/*.a
+            dist/windows-amd64/**
+
+  # ROCm generation step
+  generate-windows-rocm:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install ROCm'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading AMD HIP Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP"
+      - name: 'Verify ROCm'
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$env:PATH"
+          $env:OLLAMA_SKIP_CPU_GENERATE="1"
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          go generate -x ./...
+        name: go generate
+      - name: 'gather rocm dependencies'
+        run: |
+          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          md "dist\deps\bin\rocblas\library"
+          cp "${HIP_PATH}\bin\hipblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas.dll" "dist\deps\bin\"
+          cp "${HIP_PATH}\bin\rocblas\library\*" "dist\deps\bin\rocblas\library\"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: generate-windows-rocm
+          path: |
+            llm/build/**/bin/*
+            dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-rocm-deps
+          path: dist/deps/*
+
+  # CUDA generation step
+  generate-windows-cuda:
+    environment: release
+    runs-on: windows
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install CUDA'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading CUDA Installer"
+          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          write-host "Installing CUDA"
+          Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
+          write-host "Completed CUDA"
+          $cudaPath=((resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0].path | split-path | split-path)
+          $cudaVer=($cudaPath | split-path -leaf ) -replace 'v(\d+).(\d+)', '$1_$2' 
+          echo "$cudaPath\bin" >> $env:GITHUB_PATH
+          echo "CUDA_PATH=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" >> $env:GITHUB_ENV
+      - name: 'Verify CUDA'
+        run: nvcc -V
+      - run: go get ./...
+      - name: go generate
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $cudabin=(get-command nvcc).source | split-path
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$cudabin;$env:PATH"
+          $env:OLLAMA_SKIP_CPU_GENERATE="1"
+          go generate -x ./...
+      - name: 'gather cuda dependencies'
+        run: |
+          $NVIDIA_DIR=(resolve-path 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*\bin\')[0]
+          md "dist\deps"
+          cp "${NVIDIA_DIR}\cudart64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublas64_*.dll" "dist\deps\"
+          cp "${NVIDIA_DIR}\cublasLt64_*.dll" "dist\deps\"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: generate-windows-cuda
+          path: |
+            llm/build/**/bin/*
+            dist/windows-amd64/**
+      - uses: actions/upload-artifact@v4
+        with:
+          name: windows-cuda-deps
+          path: dist/deps/*
+
+  # Import the prior generation steps and build the final windows assets
+  build-windows:
+    environment: release
+    runs-on: windows
+    needs:
+      - generate-windows-cuda
+      - generate-windows-rocm
+      - generate-windows-cpu
+    env:
+      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - uses: 'google-github-actions/auth@v2'
+        with:
+          project_id: 'ollama'
+          credentials_json: '${{ secrets.GOOGLE_SIGNING_CREDENTIALS }}'
+      - run: echo "${{ vars.OLLAMA_CERT }}" > ollama_inc.crt
+      - name: install Windows SDK 8.1 to get signtool
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading SDK"
+          Invoke-WebRequest -Uri "https://go.microsoft.com/fwlink/p/?LinkId=323507" -OutFile "${env:RUNNER_TEMP}\sdksetup.exe"
+          Start-Process "${env:RUNNER_TEMP}\sdksetup.exe" -ArgumentList @("/q") -NoNewWindow -Wait
+          write-host "Win SDK 8.1 installed"
+          gci -path 'C:\Program Files (x86)\Windows Kits\' -r -fi 'signtool.exe'
+      - name: install signing plugin
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading plugin"
+          Invoke-WebRequest -Uri "https://github.com/GoogleCloudPlatform/kms-integrations/releases/download/cng-v1.0/kmscng-1.0-windows-amd64.zip" -OutFile "${env:RUNNER_TEMP}\plugin.zip"
+          Expand-Archive -Path "${env:RUNNER_TEMP}\plugin.zip" -DestinationPath ${env:RUNNER_TEMP}\plugin\
+          write-host "Installing plugin"
+          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
+          write-host "plugin installed"
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - run: go get
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-cpu
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-cuda
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-cuda-deps
+      - uses: actions/download-artifact@v4
+        with:
+          name: windows-rocm-deps
+      - uses: actions/download-artifact@v4
+        with:
+          name: generate-windows-rocm
+      - run: dir llm/build
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$env:PATH"
+          $env:OLLAMA_SKIP_GENERATE="1"
+          & .\scripts\build_windows.ps1
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-windows
+          path: |
+            dist/OllamaSetup.exe
+            dist/ollama-windows-*.zip
+
+  # Linux x86 assets built using the container based build
+  build-linux-amd64:
+    environment: release
+    runs-on: linux
+    env:
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: amd64
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: |
+          ./scripts/build_linux.sh
+          ./scripts/build_docker.sh
+          mv dist/deps/* dist/
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-linux-amd64
+          path: |
+            dist/*linux*
+            !dist/*-cov
+
+  # Linux ARM assets built using the container based build
+  # (at present, docker isn't pre-installed on arm ubunutu images)
+  build-linux-arm64:
+    environment: release
+    runs-on: linux-arm64
+    env:
+      OLLAMA_SKIP_MANIFEST_CREATE: '1'
+      BUILD_ARCH: arm64
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - name: Set Version
+        shell: bash
+        run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+      - name: 'Install Docker'
+        run: |
+          # Add Docker's official GPG key:
+          env
+          uname -a
+          sudo apt-get update
+          sudo apt-get install -y ca-certificates curl
+          sudo install -m 0755 -d /etc/apt/keyrings
+          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
+          sudo chmod a+r /etc/apt/keyrings/docker.asc
+
+          # Add the repository to Apt sources:
+          echo \
+            "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
+            $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
+            sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
+          sudo apt-get update
+          sudo apt-get install -y docker-ce docker-ce-cli containerd.io
+          sudo usermod -aG docker $USER
+          sudo apt-get install acl
+          sudo setfacl --modify user:$USER:rw /var/run/docker.sock
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: |
+          ./scripts/build_linux.sh
+          ./scripts/build_docker.sh
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist-linux-arm64
+          path: |
+            dist/*linux*
+            !dist/*-cov
+
+  # Aggregate all the assets and ship a release
+  release:
+    needs:
+      - build-darwin
+      - build-windows
+      - build-linux-amd64
+      - build-linux-arm64
+    runs-on: linux
+    environment: release
+    permissions:
+      contents: write
+    env:
+      OLLAMA_SKIP_IMAGE_BUILD: '1'
+      PUSH: '1'
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set Version
+        shell: bash
+        run: |
+          echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
+          echo "RELEASE_VERSION=$(echo ${GITHUB_REF_NAME} | cut -f1 -d-)" >> $GITHUB_ENV
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - run: ./scripts/build_docker.sh
+      - name: Retrieve built artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: dist
+          pattern: dist-*
+          merge-multiple: true
+      - run: |
+          ls -lh dist/
+          (cd dist; sha256sum * > sha256sum.txt)
+          cat dist/sha256sum.txt
+      - uses: ncipollo/release-action@v1
+        with:
+          name: ${{ env.RELEASE_VERSION }}
+          allowUpdates: true
+          artifacts: 'dist/*'
+          draft: true
+          prerelease: true
+          omitBodyDuringUpdate: true
+          generateReleaseNotes: true
+          omitDraftDuringUpdate: true
+          omitPrereleaseDuringUpdate: true
+          replacesArtifacts: true
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@ -1,15 +1,51 @@
 name: test

+concurrency:
+  # For PRs, later CI runs preempt previous ones. e.g. a force push on a PR
+  # cancels running CI jobs and starts all new ones.
+  #
+  # For non-PR pushes, concurrency.group needs to be unique for every distinct
+  # CI run we want to have happen. Use run_id, which in practice means all
+  # non-PR CI runs will be allowed to run without preempting each other.
+  group: ${{ github.workflow }}-$${{ github.pull_request.number || github.run_id }}
+  cancel-in-progress: true
+
 on:
  pull_request:
    paths:
      - '**/*'
      - '!docs/**'
-      - '!examples/**'
      - '!README.md'

 jobs:
+  changes:
+    runs-on: ubuntu-latest
+    outputs:
+      GENERATE: ${{ steps.changes.outputs.GENERATE }}
+      GENERATE_CUDA: ${{ steps.changes.outputs.GENERATE_CUDA }}
+      GENERATE_ROCM: ${{ steps.changes.outputs.GENERATE_ROCM }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - id: changes
+        run: |
+          changed() {
+            git diff-tree -r --no-commit-id --name-only \
+              $(git merge-base ${{ github.event.pull_request.base.sha }} ${{ github.event.pull_request.head.sha }}) \
+              ${{ github.event.pull_request.head.sha }} \
+              | xargs python3 -c "import sys; print(any([x.startswith('$1') for x in sys.argv[1:]]))"
+          }
+
+          {
+            echo GENERATE=$(changed llm/)
+            echo GENERATE_CUDA=$(changed llm/)
+            echo GENERATE_ROCM=$(changed llm/)
+          } >>$GITHUB_OUTPUT
+
  generate:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE == 'True' }}
    strategy:
      matrix:
        os: [ubuntu-latest, macos-latest, windows-2019]
@ -26,26 +62,32 @@ jobs:
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v5
        with:
-          go-version: '1.22'
+          go-version-file: go.mod
          cache: true
      - run: go get ./...
      - run: |
          $gopath=(get-command go).source | split-path -parent
+          $gccpath=(get-command gcc).source | split-path -parent
          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
          cd $env:GITHUB_WORKSPACE
          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$env:PATH"
+          $env:PATH="$gopath;$gccpath;$env:PATH"
+          echo $env:PATH
          go generate -x ./...
        if: ${{ startsWith(matrix.os, 'windows-') }}
-        name: "Windows Go Generate"
+        name: 'Windows Go Generate'
      - run: go generate -x ./...
        if: ${{ ! startsWith(matrix.os, 'windows-') }}
-        name: "Unix Go Generate"
+        name: 'Unix Go Generate'
      - uses: actions/upload-artifact@v4
        with:
          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: |
+            llm/build/**/bin/*
+            llm/build/**/*.a
  generate-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
    strategy:
      matrix:
        cuda-version:
@ -62,7 +104,7 @@ jobs:
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v4
        with:
-          go-version: '1.22'
+          go-version-file: go.mod
          cache: true
      - run: go get ./...
      - run: |
@ -73,12 +115,16 @@ jobs:
      - uses: actions/upload-artifact@v4
        with:
          name: cuda-${{ matrix.cuda-version }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: |
+            llm/build/**/bin/*
+            dist/windows-amd64/**
  generate-rocm:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
    strategy:
      matrix:
        rocm-version:
-          - '6.0'
+          - '6.0.2'
    runs-on: linux
    container: rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}
    steps:
@ -91,7 +137,7 @@ jobs:
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v4
        with:
-          go-version: '1.22'
+          go-version-file: go.mod
          cache: true
      - run: go get ./...
      - run: |
@ -102,7 +148,89 @@ jobs:
      - uses: actions/upload-artifact@v4
        with:
          name: rocm-${{ matrix.rocm-version }}-libraries
-          path: llm/llama.cpp/build/**/lib/*
+          path: |
+            llm/build/**/bin/*
+            dist/windows-amd64/**
+
+  # ROCm generation step
+  generate-windows-rocm:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_ROCM == 'True' }}
+    runs-on: windows
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install ROCm'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading AMD HIP Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP"
+      - name: 'Verify ROCm'
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+      - run: go get ./...
+      - run: |
+          $gopath=(get-command go).source | split-path -parent
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$env:PATH"
+          $env:OLLAMA_SKIP_CPU_GENERATE="1"
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          go generate -x ./...
+        name: go generate
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
+
+  # CUDA generation step
+  generate-windows-cuda:
+    needs: [changes]
+    if: ${{ needs.changes.outputs.GENERATE_CUDA == 'True' }}
+    runs-on: windows
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+          cache: true
+      - name: 'Install CUDA'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "downloading CUDA Installer"
+          Invoke-WebRequest -Uri "https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe" -OutFile "${env:RUNNER_TEMP}\cuda-install.exe"
+          write-host "Installing CUDA"
+          Start-Process "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
+          write-host "Completed CUDA"
+          $cudaPath=((resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0].path | split-path | split-path)
+          $cudaVer=($cudaPath | split-path -leaf ) -replace 'v(\d+).(\d+)', '$1_$2' 
+          echo "$cudaPath\bin" >> $env:GITHUB_PATH
+          echo "CUDA_PATH=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" >> $env:GITHUB_ENV
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" >> $env:GITHUB_ENV
+      - name: 'Verify CUDA'
+        run: nvcc -V
+      - run: go get ./...
+      - name: go generate
+        run: |
+          $gopath=(get-command go).source | split-path -parent
+          $cudabin=(get-command nvcc).source | split-path
+          & "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Launch-VsDevShell.ps1"
+          cd $env:GITHUB_WORKSPACE
+          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
+          $env:PATH="$gopath;$cudabin;$env:PATH"
+          $env:OLLAMA_SKIP_CPU_GENERATE="1"
+          go generate -x ./...
+        env:
+          OLLAMA_SKIP_CPU_GENERATE: '1'
+      # TODO - do we need any artifacts?
+
  lint:
    strategy:
      matrix:
@ -125,24 +253,26 @@ jobs:
          submodules: recursive
      - uses: actions/setup-go@v5
        with:
-          go-version: '1.22'
+          go-version-file: go.mod
          cache: false
      - run: |
-          mkdir -p llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/linux/${{ matrix.arch }}/stub/lib/stub.so
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=x86_64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
+      - run: |
+          mkdir -p llm/build/linux/$ARCH/stub/bin
+          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
      - run: |
-          mkdir -p llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/darwin/${{ matrix.arch }}/stub/lib/stub.dylib
-          touch llm/llama.cpp/ggml-metal.metal
+          mkdir -p llm/build/darwin/$ARCH/stub/bin
+          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
        if: ${{ startsWith(matrix.os, 'macos-') }}
-      - run: |
-          mkdir -p llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/
-          touch llm/llama.cpp/build/windows/${{ matrix.arch }}/stub/lib/stub.dll
-        if: ${{ startsWith(matrix.os, 'windows-') }}
-      - uses: golangci/golangci-lint-action@v3
+      - uses: golangci/golangci-lint-action@v4
+        with:
+          args: --timeout 8m0s -v
  test:
-    needs: generate
    strategy:
      matrix:
        os: [ubuntu-latest, macos-latest, windows-2019]
@ -156,19 +286,31 @@ jobs:
    env:
      GOARCH: ${{ matrix.arch }}
      CGO_ENABLED: '1'
+      OLLAMA_CPU_TARGET: 'static'
    steps:
      - uses: actions/checkout@v4
        with:
          submodules: recursive
      - uses: actions/setup-go@v5
        with:
-          go-version: '1.22'
+          go-version-file: go.mod
          cache: true
-      - run: go get
-      - uses: actions/download-artifact@v4
-        with:
-          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
-          path: llm/llama.cpp/build
+      - run: |
+          case ${{ matrix.arch }} in
+            amd64) echo ARCH=x86_64 ;;
+            arm64) echo ARCH=arm64 ;;
+          esac >>$GITHUB_ENV
+        shell: bash
+      - run: |
+          mkdir -p llm/build/linux/$ARCH/stub/bin
+          touch llm/build/linux/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'ubuntu-') }}
+      - run: |
+          mkdir -p llm/build/darwin/$ARCH/stub/bin
+          touch llm/build/darwin/$ARCH/stub/bin/ollama_llama_server
+        if: ${{ startsWith(matrix.os, 'macos-') }}
+        shell: bash
+      - run: go generate ./...
      - run: go build
      - run: go test -v ./...
      - uses: actions/upload-artifact@v4
--- a/.gitignore
+++ b/.gitignore
@ -10,4 +10,6 @@ ggml-metal.metal
 *.exe
 .idea
 test_data
-*.crt
+*.crt
+llm/build
+__debug_bin*
--- a/.golangci.yaml
+++ b/.golangci.yaml
@ -15,13 +15,3 @@ linters:
    - misspell
    - nilerr
    - unused
-linters-settings:
-  errcheck:
-    # exclude the following functions since we don't generally
-    # need to be concerned with the returned errors
-    exclude-functions:
-      - encoding/binary.Read
-      - (*os.File).Seek
-      - (*bufio.Writer).WriteString
-      - (*github.com/spf13/pflag.FlagSet).Set
-      - (*github.com/jmorganca/ollama/llm.readSeekOffset).Seek
--- a/77
+++ b/77
@ -1,7 +1,8 @@
 ARG GOLANG_VERSION=1.22.1
 ARG CMAKE_VERSION=3.22.1
+# this CUDA_VERSION corresponds with the one specified in docs/gpu.md
 ARG CUDA_VERSION=11.3.1
-ARG ROCM_VERSION=6.0
+ARG ROCM_VERSION=6.0.2

 # Copy the minimal context we need to run the generate scripts
 FROM scratch AS llm-code
@ -14,20 +15,20 @@ ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh

 FROM --platform=linux/arm64 nvidia/cuda:$CUDA_VERSION-devel-rockylinux8 AS cuda-build-arm64
 ARG CMAKE_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh

 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS rocm-build-amd64
 ARG CMAKE_VERSION
@ -35,18 +36,18 @@ COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
 ENV LIBRARY_PATH /opt/amdgpu/lib64
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
 ARG CGO_CFLAGS
 ARG AMDGPU_TARGETS
-RUN OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_SKIP_CPU_GENERATE=1 sh gen_linux.sh
 RUN mkdir /tmp/scratch && \
-    for dep in $(cat /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/x86_64/rocm*/lib/deps.txt) ; do \
+    for dep in $(zcat /go/src/github.com/ollama/ollama/llm/build/linux/x86_64/rocm*/bin/deps.txt.gz) ; do \
        cp ${dep} /tmp/scratch/ || exit 1 ; \
    done && \
    (cd /opt/rocm/lib && tar cf - rocblas/library) | (cd /tmp/scratch/ && tar xf - ) && \
-    mkdir -p /go/src/github.com/jmorganca/ollama/dist/deps/ && \
-    (cd /tmp/scratch/ && tar czvf /go/src/github.com/jmorganca/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )
+    mkdir -p /go/src/github.com/ollama/ollama/dist/deps/ && \
+    (cd /tmp/scratch/ && tar czvf /go/src/github.com/ollama/ollama/dist/deps/ollama-linux-amd64-rocm.tgz . )


 FROM --platform=linux/amd64 centos:7 AS cpu-builder-amd64
@ -55,41 +56,48 @@ ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate

+FROM --platform=linux/amd64 cpu-builder-amd64 AS static-build-amd64
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu-build-amd64
-RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx-build-amd64
-RUN OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh
 FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64
-RUN OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh

-FROM --platform=linux/arm64 centos:7 AS cpu-build-arm64
+FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64
 ARG CMAKE_VERSION
 ARG GOLANG_VERSION
 COPY ./scripts/rh_linux_deps.sh /
 RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh
 ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH
-COPY --from=llm-code / /go/src/github.com/jmorganca/ollama/
-WORKDIR /go/src/github.com/jmorganca/ollama/llm/generate
-# Note, we only build the "base" CPU variant on arm since avx/avx2 are x86 features
+COPY --from=llm-code / /go/src/github.com/ollama/ollama/
 ARG OLLAMA_CUSTOM_CPU_DEFS
 ARG CGO_CFLAGS
-RUN OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+WORKDIR /go/src/github.com/ollama/ollama/llm/generate
+
+FROM --platform=linux/arm64 cpu-builder-arm64 AS static-build-arm64
+RUN OLLAMA_CPU_TARGET="static" sh gen_linux.sh
+FROM --platform=linux/arm64 cpu-builder-arm64 AS cpu-build-arm64
+RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu" sh gen_linux.sh
+

 # Intermediate stage used for ./scripts/build_linux.sh
 FROM --platform=linux/amd64 cpu-build-amd64 AS build-amd64
 ENV CGO_ENABLED 1
-WORKDIR /go/src/github.com/jmorganca/ollama
+WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=cpu_avx-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=cpu_avx2-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=cuda-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
-COPY --from=rocm-build-amd64 /go/src/github.com/jmorganca/ollama/dist/deps/ ./dist/deps/
+COPY --from=static-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cpu_avx-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cpu_avx2-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=rocm-build-amd64 /go/src/github.com/ollama/ollama/dist/deps/ ./dist/deps/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN go build -trimpath .
@ -98,9 +106,10 @@ RUN go build -trimpath .
 FROM --platform=linux/arm64 cpu-build-arm64 AS build-arm64
 ENV CGO_ENABLED 1
 ARG GOLANG_VERSION
-WORKDIR /go/src/github.com/jmorganca/ollama
+WORKDIR /go/src/github.com/ollama/ollama
 COPY . .
-COPY --from=cuda-build-arm64 /go/src/github.com/jmorganca/ollama/llm/llama.cpp/build/linux/ llm/llama.cpp/build/linux/
+COPY --from=static-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
+COPY --from=cuda-build-arm64 /go/src/github.com/ollama/ollama/llm/build/linux/ llm/build/linux/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN go build -trimpath .
@ -108,15 +117,15 @@ RUN go build -trimpath .
 # Runtime stages
 FROM --platform=linux/amd64 ubuntu:22.04 as runtime-amd64
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-amd64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-amd64 /go/src/github.com/ollama/ollama/ollama /bin/ollama
 FROM --platform=linux/arm64 ubuntu:22.04 as runtime-arm64
 RUN apt-get update && apt-get install -y ca-certificates
-COPY --from=build-arm64 /go/src/github.com/jmorganca/ollama/ollama /bin/ollama
+COPY --from=build-arm64 /go/src/github.com/ollama/ollama/ollama /bin/ollama

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

--- a/README.md
+++ b/README.md
@ -1,5 +1,5 @@
 <div align="center">
-  <img alt="ollama" height="200px" src="https://github.com/jmorganca/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
+ <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
 </div>

 # Ollama
@ -22,7 +22,7 @@ Get up and running with large language models locally.
 curl -fsSL https://ollama.com/install.sh | sh
 ```

-[Manual install instructions](https://github.com/jmorganca/ollama/blob/main/docs/linux.md)
+[Manual install instructions](https://github.com/ollama/ollama/blob/main/docs/linux.md)

 ### Docker

@ -35,10 +35,10 @@ The official [Ollama Docker image](https://hub.docker.com/r/ollama/ollama) `olla

 ## Quickstart

-To run and chat with [Llama 2](https://ollama.com/library/llama2):
+To run and chat with [Llama 3](https://ollama.com/library/llama3):

 ```
-ollama run llama2
+ollama run llama3
 ```

 ## Model library
@ -49,21 +49,18 @@ Here are some example models that can be downloaded:

 | Model              | Parameters | Size  | Download                       |
 | ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 2            | 7B         | 3.8GB | `ollama run llama2`            |
+| Llama 3            | 8B         | 4.7GB | `ollama run llama3`            |
+| Llama 3            | 70B        | 40GB  | `ollama run llama3:70b`        |
+| Phi-3              | 3.8B       | 2.3GB | `ollama run phi3`              |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
-| Dolphin Phi        | 2.7B       | 1.6GB | `ollama run dolphin-phi`       |
-| Phi-2              | 2.7B       | 1.7GB | `ollama run phi`               |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`       |
 | Starling           | 7B         | 4.1GB | `ollama run starling-lm`       |
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`         |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored` |
-| Llama 2 13B        | 13B        | 7.3GB | `ollama run llama2:13b`        |
-| Llama 2 70B        | 70B        | 39GB  | `ollama run llama2:70b`        |
-| Orca Mini          | 3B         | 1.9GB | `ollama run orca-mini`         |
-| Vicuna             | 7B         | 3.8GB | `ollama run vicuna`            |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
 | Gemma              | 2B         | 1.4GB | `ollama run gemma:2b`          |
 | Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
+| Solar              | 10.7B      | 6.1GB | `ollama run solar`             |

 > Note: You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.

@ -97,16 +94,16 @@ See the [guide](docs/import.md) on importing models for more information.

 ### Customize a prompt

-Models from the Ollama library can be customized with a prompt. For example, to customize the `llama2` model:
+Models from the Ollama library can be customized with a prompt. For example, to customize the `llama3` model:

 ```
-ollama pull llama2
+ollama pull llama3
 ```

 Create a `Modelfile`:

 ```
-FROM llama2
+FROM llama3

 # set the temperature to 1 [higher is more creative, lower is more coherent]
 PARAMETER temperature 1
@ -141,7 +138,7 @@ ollama create mymodel -f ./Modelfile
 ### Pull a model

 ```
-ollama pull llama2
+ollama pull llama3
 ```

 > This command can also be used to update a local model. Only the diff will be pulled.
@ -149,13 +146,13 @@ ollama pull llama2
 ### Remove a model

 ```
-ollama rm llama2
+ollama rm llama3
 ```

 ### Copy a model

 ```
-ollama cp llama2 my-llama2
+ollama cp llama3 my-model
 ```

 ### Multiline input
@ -176,10 +173,10 @@ I'm a basic program that prints the famous "Hello, world!" message to the consol
 The image features a yellow smiley face, which is likely the central focus of the picture.
 ```

-### Pass in prompt as arguments
+### Pass the prompt as an argument

 ```
-$ ollama run llama2 "Summarize this file: $(cat README.md)"
+$ ollama run llama3 "Summarize this file: $(cat README.md)"
 Ollama is a lightweight, extensible framework for building and running language models on the local machine. It provides a simple API for creating, running, and managing models, as well as a library of pre-built models that can be easily used in a variety of applications.
 ```

@ -213,7 +210,7 @@ Then build the binary:
 go build .
 ```

-More detailed instructions can be found in the [developer guide](https://github.com/jmorganca/ollama/blob/main/docs/development.md)
+More detailed instructions can be found in the [developer guide](https://github.com/ollama/ollama/blob/main/docs/development.md)

 ### Running local builds

@ -226,7 +223,7 @@ Next, start the server:
 Finally, in a separate shell, run a model:

 ```
-./ollama run llama2
+./ollama run llama3
 ```

 ## REST API
@ -237,7 +234,7 @@ Ollama has a REST API for running and managing models.

 ```
 curl http://localhost:11434/api/generate -d '{
-  "model": "llama2",
+  "model": "llama3",
  "prompt":"Why is the sky blue?"
 }'
 ```
@ -246,7 +243,7 @@ curl http://localhost:11434/api/generate -d '{

 ```
 curl http://localhost:11434/api/chat -d '{
-  "model": "mistral",
+  "model": "llama3",
  "messages": [
    { "role": "user", "content": "why is the sky blue?" }
  ]
@ -259,27 +256,45 @@ See the [API documentation](./docs/api.md) for all endpoints.

 ### Web & Desktop

- [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
+- [Open WebUI](https://github.com/open-webui/open-webui)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
+- [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
+- [LibreChat](https://github.com/danny-avila/LibreChat)
+- [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
+- [Saddle](https://github.com/jikkuatwork/saddle)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
+- [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
- [Open WebUI](https://github.com/open-webui/open-webui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
 - [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
 - [Ollama-SwiftUI](https://github.com/kghandour/Ollama-SwiftUI)
+- [Dify.AI](https://github.com/langgenius/dify)
 - [MindMac](https://mindmac.app)
 - [NextJS Web Interface for Ollama](https://github.com/jakobhoeg/nextjs-ollama-llm-ui)
 - [Msty](https://msty.app)
 - [Chatbox](https://github.com/Bin-Huang/Chatbox)
 - [WinForm Ollama Copilot](https://github.com/tgraupmann/WinForm_Ollama_Copilot)
 - [NextChat](https://github.com/ChatGPTNextWeb/ChatGPT-Next-Web) with [Get Started Doc](https://docs.nextchat.dev/models/ollama)
+- [Alpaca WebUI](https://github.com/mmo80/alpaca-webui)
+- [OllamaGUI](https://github.com/enoch1118/ollamaGUI)
+- [OpenAOE](https://github.com/InternLM/OpenAOE)
 - [Odin Runes](https://github.com/leonid20000/OdinRunes)
 - [LLM-X: Progressive Web App](https://github.com/mrdjohnson/llm-x)
+- [AnythingLLM (Docker + MacOs/Windows/Linux native app)](https://github.com/Mintplex-Labs/anything-llm)
+- [Ollama Basic Chat: Uses HyperDiv Reactive UI](https://github.com/rapidarchitect/ollama_basic_chat)
+- [Ollama-chats RPG](https://github.com/drazdra/ollama-chats)
+- [QA-Pilot: Chat with Code Repository](https://github.com/reid41/QA-Pilot)
+- [ChatOllama: Open Source Chatbot based on Ollama with Knowledge Bases](https://github.com/sugarforever/chat-ollama)
+- [CRAG Ollama Chat: Simple Web Search with Corrective RAG](https://github.com/Nagi-ovo/CRAG-Ollama-Chat)
+- [RAGFlow: Open-source Retrieval-Augmented Generation engine based on deep document understanding](https://github.com/infiniflow/ragflow)
+- [chat: chat web app for teams](https://github.com/swuecho/chat)
+- [Lobe Chat](https://github.com/lobehub/lobe-chat) with [Integrating Doc](https://lobehub.com/docs/self-hosting/examples/ollama)
+- [Ollama RAG Chatbot: Local Chat with multiple PDFs using Ollama and RAG.](https://github.com/datvodinh/rag-chatbot.git)

 ### Terminal

@ -288,18 +303,25 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Emacs client](https://github.com/zweifisch/ollama)
 - [gen.nvim](https://github.com/David-Kunz/gen.nvim)
 - [ollama.nvim](https://github.com/nomnivore/ollama.nvim)
+- [ollero.nvim](https://github.com/marco-souza/ollero.nvim)
 - [ollama-chat.nvim](https://github.com/gerazov/ollama-chat.nvim)
 - [ogpt.nvim](https://github.com/huynle/ogpt.nvim)
 - [gptel Emacs client](https://github.com/karthink/gptel)
 - [Oatmeal](https://github.com/dustinblackman/oatmeal)
 - [cmdh](https://github.com/pgibler/cmdh)
+- [ooo](https://github.com/npahlfer/ooo)
+- [shell-pilot](https://github.com/reid41/shell-pilot)
 - [tenere](https://github.com/pythops/tenere)
 - [llm-ollama](https://github.com/taketwo/llm-ollama) for [Datasette's LLM CLI](https://llm.datasette.io/en/stable/).
+- [typechat-cli](https://github.com/anaisbetts/typechat-cli)
 - [ShellOracle](https://github.com/djcopley/ShellOracle)
+- [tlm](https://github.com/yusufcanb/tlm)
+- [podman-ollama](https://github.com/ericcurtin/podman-ollama)

 ### Database

- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md)
+- [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
+- [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)

 ### Package managers

@ -312,7 +334,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LlamaIndex](https://gpt-index.readthedocs.io/en/stable/examples/llm/ollama.html)
- [LangChain4j](https://github.com/langchain4j/langchain4j/tree/main/langchain4j-ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaSharp for .NET](https://github.com/awaescher/OllamaSharp)
 - [Ollama for Ruby](https://github.com/gbaptista/ollama-ai)
@ -329,6 +350,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama for R - rollama](https://github.com/JBGruber/rollama)
 - [Ollama-ex for Elixir](https://github.com/lebrunel/ollama-ex)
 - [Ollama Connector for SAP ABAP](https://github.com/b-tocs/abap_btocs_ollama)
+- [Testcontainers](https://testcontainers.com/modules/ollama/)

 ### Mobile

@ -350,9 +372,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Rivet plugin](https://github.com/abrenneke/rivet-plugin-ollama)
 - [Llama Coder](https://github.com/ex3ndr/llama-coder) (Copilot alternative using Ollama)
 - [Obsidian BMO Chatbot plugin](https://github.com/longy2k/obsidian-bmo-chatbot)
+- [Cliobot](https://github.com/herval/cliobot) (Telegram bot with Ollama support)
 - [Copilot for Obsidian plugin](https://github.com/logancyang/obsidian-copilot)
 - [Obsidian Local GPT plugin](https://github.com/pfrankov/obsidian-local-gpt)
 - [Open Interpreter](https://docs.openinterpreter.com/language-model-setup/local-models/ollama)
 - [twinny](https://github.com/rjmacarthy/twinny) (Copilot and Copilot chat alternative using Ollama)
 - [Wingman-AI](https://github.com/RussellCanfield/wingman-ai) (Copilot code and chat alternative using Ollama and HuggingFace)
 - [Page Assist](https://github.com/n4ze3m/page-assist) (Chrome Extension)
+- [AI Telegram Bot](https://github.com/tusharhero/aitelegrambot) (Telegram bot using Ollama in backend)
+- [AI ST Completion](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (Sublime Text 4 AI assistant plugin with Ollama support)
+- [Discord-Ollama Chat Bot](https://github.com/kevinthedang/discord-ollama) (Generalized TypeScript Discord Bot w/ Tuning Documentation)
+
+### Supported backends 
+- [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov. 
--- a/api/client.go
+++ b/api/client.go
@ -1,3 +1,9 @@
+// Package api implements the client-side API for code wishing to interact
+// with the ollama service. The methods of the [Client] type correspond to
+// the ollama REST API as described in https://github.com/ollama/ollama/blob/main/docs/api.md
+//
+// The ollama command-line client itself uses this package to interact with
+// the backend service.
 package api

 import (
@ -5,7 +11,6 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"net"
@ -13,12 +18,15 @@ import (
 	"net/url"
 	"os"
 	"runtime"
+	"strconv"
 	"strings"

-	"github.com/jmorganca/ollama/format"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/version"
 )

+// Client encapsulates client state for interacting with the ollama
+// service. Use [ClientFromEnvironment] to create new Clients.
 type Client struct {
 	base *url.URL
 	http *http.Client
@ -40,13 +48,46 @@ func checkError(resp *http.Response, body []byte) error {
 	return apiError
 }

+// ClientFromEnvironment creates a new [Client] using configuration from the
+// environment variable OLLAMA_HOST, which points to the network host and
+// port on which the ollama service is listenting. The format of this variable
+// is:
+//
+//	<scheme>://<host>:<port>
+//
+// If the variable is not specified, a default ollama host and port will be
+// used.
 func ClientFromEnvironment() (*Client, error) {
+	ollamaHost, err := GetOllamaHost()
+	if err != nil {
+		return nil, err
+	}
+
+	return &Client{
+		base: &url.URL{
+			Scheme: ollamaHost.Scheme,
+			Host:   net.JoinHostPort(ollamaHost.Host, ollamaHost.Port),
+		},
+		http: http.DefaultClient,
+	}, nil
+}
+
+type OllamaHost struct {
+	Scheme string
+	Host   string
+	Port   string
+}
+
+func GetOllamaHost() (OllamaHost, error) {
 	defaultPort := "11434"

-	scheme, hostport, ok := strings.Cut(os.Getenv("OLLAMA_HOST"), "://")
+	hostVar := os.Getenv("OLLAMA_HOST")
+	hostVar = strings.TrimSpace(strings.Trim(strings.TrimSpace(hostVar), "\"'"))
+
+	scheme, hostport, ok := strings.Cut(hostVar, "://")
 	switch {
 	case !ok:
-		scheme, hostport = "http", os.Getenv("OLLAMA_HOST")
+		scheme, hostport = "http", hostVar
 	case scheme == "http":
 		defaultPort = "80"
 	case scheme == "https":
@ -66,15 +107,24 @@ func ClientFromEnvironment() (*Client, error) {
 		}
 	}

-	return &Client{
-		base: &url.URL{
-			Scheme: scheme,
-			Host:   net.JoinHostPort(host, port),
-		},
-		http: http.DefaultClient,
+	if portNum, err := strconv.ParseInt(port, 10, 32); err != nil || portNum > 65535 || portNum < 0 {
+		return OllamaHost{}, ErrInvalidHostPort
+	}
+
+	return OllamaHost{
+		Scheme: scheme,
+		Host:   host,
+		Port:   port,
 	}, nil
 }

+func NewClient(base *url.URL, http *http.Client) *Client {
+	return &Client{
+		base: base,
+		http: http,
+	}
+}
+
 func (c *Client) do(ctx context.Context, method, path string, reqData, respData any) error {
 	var reqBody io.Reader
 	var data []byte
@ -191,8 +241,14 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f
 	return nil
 }

+// GenerateResponseFunc is a function that [Client.Generate] invokes every time
+// a response is received from the service. If this function returns an error,
+// [Client.Generate] will stop generating and return this error.
 type GenerateResponseFunc func(GenerateResponse) error

+// Generate generates a response for a given prompt. The req parameter should
+// be populated with prompt details. fn is called for each response (there may
+// be multiple responses, e.g. in case streaming is enabled).
 func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn GenerateResponseFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/generate", req, func(bts []byte) error {
 		var resp GenerateResponse
@ -204,8 +260,15 @@ func (c *Client) Generate(ctx context.Context, req *GenerateRequest, fn Generate
 	})
 }

+// ChatResponseFunc is a function that [Client.Chat] invokes every time
+// a response is received from the service. If this function returns an error,
+// [Client.Chat] will stop generating and return this error.
 type ChatResponseFunc func(ChatResponse) error

+// Chat generates the next message in a chat. [ChatRequest] may contain a
+// sequence of messages which can be used to maintain chat history with a model.
+// fn is called for each response (there may be multiple responses, e.g. if case
+// streaming is enabled).
 func (c *Client) Chat(ctx context.Context, req *ChatRequest, fn ChatResponseFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/chat", req, func(bts []byte) error {
 		var resp ChatResponse
@ -217,8 +280,14 @@ func (c *Client) Chat(ctx context.Context, req *ChatRequest, fn ChatResponseFunc
 	})
 }

+// PullProgressFunc is a function that [Client.Pull] invokes every time there
+// is progress with a "pull" request sent to the service. If this function
+// returns an error, [Client.Pull] will stop the process and return this error.
 type PullProgressFunc func(ProgressResponse) error

+// Pull downloads a model from the ollama library. fn is called each time
+// progress is made on the request and can be used to display a progress bar,
+// etc.
 func (c *Client) Pull(ctx context.Context, req *PullRequest, fn PullProgressFunc) error {
 	return c.stream(ctx, http.MethodPost, "/api/pull", req, func(bts []byte) error {
 		var resp ProgressResponse
@ -301,18 +370,7 @@ func (c *Client) Embeddings(ctx context.Context, req *EmbeddingRequest) (*Embedd
 }

 func (c *Client) CreateBlob(ctx context.Context, digest string, r io.Reader) error {
-	if err := c.do(ctx, http.MethodHead, fmt.Sprintf("/api/blobs/%s", digest), nil, nil); err != nil {
-		var statusError StatusError
-		if !errors.As(err, &statusError) || statusError.StatusCode != http.StatusNotFound {
-			return err
-		}
-
-		if err := c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil); err != nil {
-			return err
-		}
-	}
-
-	return nil
+	return c.do(ctx, http.MethodPost, fmt.Sprintf("/api/blobs/%s", digest), r, nil)
 }

 func (c *Client) Version(ctx context.Context) (string, error) {
--- a/api/client_test.go
+++ b/api/client_test.go
@ -1,6 +1,12 @@
 package api

-import "testing"
+import (
+	"fmt"
+	"net"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)

 func TestClientFromEnvironment(t *testing.T) {
 	type testCase struct {
@ -40,4 +46,40 @@ func TestClientFromEnvironment(t *testing.T) {
 			}
 		})
 	}
+
+	hostTestCases := map[string]*testCase{
+		"empty":               {value: "", expect: "127.0.0.1:11434"},
+		"only address":        {value: "1.2.3.4", expect: "1.2.3.4:11434"},
+		"only port":           {value: ":1234", expect: ":1234"},
+		"address and port":    {value: "1.2.3.4:1234", expect: "1.2.3.4:1234"},
+		"hostname":            {value: "example.com", expect: "example.com:11434"},
+		"hostname and port":   {value: "example.com:1234", expect: "example.com:1234"},
+		"zero port":           {value: ":0", expect: ":0"},
+		"too large port":      {value: ":66000", err: ErrInvalidHostPort},
+		"too small port":      {value: ":-1", err: ErrInvalidHostPort},
+		"ipv6 localhost":      {value: "[::1]", expect: "[::1]:11434"},
+		"ipv6 world open":     {value: "[::]", expect: "[::]:11434"},
+		"ipv6 no brackets":    {value: "::1", expect: "[::1]:11434"},
+		"ipv6 + port":         {value: "[::1]:1337", expect: "[::1]:1337"},
+		"extra space":         {value: " 1.2.3.4 ", expect: "1.2.3.4:11434"},
+		"extra quotes":        {value: "\"1.2.3.4\"", expect: "1.2.3.4:11434"},
+		"extra space+quotes":  {value: " \" 1.2.3.4 \" ", expect: "1.2.3.4:11434"},
+		"extra single quotes": {value: "'1.2.3.4'", expect: "1.2.3.4:11434"},
+	}
+
+	for k, v := range hostTestCases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_HOST", v.value)
+
+			oh, err := GetOllamaHost()
+			if err != v.err {
+				t.Fatalf("expected %s, got %s", v.err, err)
+			}
+
+			if err == nil {
+				host := net.JoinHostPort(oh.Host, oh.Port)
+				assert.Equal(t, v.expect, host, fmt.Sprintf("%s: expected %s, got %s", k, v.expect, host))
+			}
+		})
+	}
 }
--- a/api/types.go
+++ b/api/types.go
@ -2,6 +2,7 @@ package api

 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"math"
 	"os"
@ -33,18 +34,46 @@ func (e StatusError) Error() string {

 type ImageData []byte

+// GenerateRequest describes a request sent by [Client.Generate]. While you
+// have to specify the Model and Prompt fields, all the other fields have
+// reasonable defaults for basic uses.
 type GenerateRequest struct {
-	Model     string      `json:"model"`
-	Prompt    string      `json:"prompt"`
-	System    string      `json:"system"`
-	Template  string      `json:"template"`
-	Context   []int       `json:"context,omitempty"`
-	Stream    *bool       `json:"stream,omitempty"`
-	Raw       bool        `json:"raw,omitempty"`
-	Format    string      `json:"format"`
-	KeepAlive *Duration   `json:"keep_alive,omitempty"`
-	Images    []ImageData `json:"images,omitempty"`
+	// Model is the model name; it should be a name familiar to Ollama from
+	// the library at https://ollama.com/library
+	Model string `json:"model"`

+	// Prompt is the textual prompt to send to the model.
+	Prompt string `json:"prompt"`
+
+	// System overrides the model's default system message/prompt.
+	System string `json:"system"`
+
+	// Template overrides the model's default prompt template.
+	Template string `json:"template"`
+
+	// Context is the context parameter returned from a previous call to
+	// Generate call. It can be used to keep a short conversational memory.
+	Context []int `json:"context,omitempty"`
+
+	// Stream specifies whether the response is streaming; it is true by default.
+	Stream *bool `json:"stream,omitempty"`
+
+	// Raw set to true means that no formatting will be applied to the prompt.
+	Raw bool `json:"raw,omitempty"`
+
+	// Format specifies the format to return a response in.
+	Format string `json:"format"`
+
+	// KeepAlive controls how long the model will stay loaded in memory following
+	// this request.
+	KeepAlive *Duration `json:"keep_alive,omitempty"`
+
+	// Images is an optional list of base64-encoded images accompanying this
+	// request, for multimodal models.
+	Images []ImageData `json:"images,omitempty"`
+
+	// Options lists model-specific options. For example, temperature can be
+	// set through this field, if the model supports it.
 	Options map[string]interface{} `json:"options"`
 }

@ -109,21 +138,24 @@ type Options struct {

 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA            bool    `json:"numa,omitempty"`
-	NumCtx             int     `json:"num_ctx,omitempty"`
-	NumBatch           int     `json:"num_batch,omitempty"`
-	NumGQA             int     `json:"num_gqa,omitempty"`
-	NumGPU             int     `json:"num_gpu,omitempty"`
-	MainGPU            int     `json:"main_gpu,omitempty"`
-	LowVRAM            bool    `json:"low_vram,omitempty"`
-	F16KV              bool    `json:"f16_kv,omitempty"`
-	LogitsAll          bool    `json:"logits_all,omitempty"`
-	VocabOnly          bool    `json:"vocab_only,omitempty"`
-	UseMMap            bool    `json:"use_mmap,omitempty"`
-	UseMLock           bool    `json:"use_mlock,omitempty"`
-	RopeFrequencyBase  float32 `json:"rope_frequency_base,omitempty"`
+	UseNUMA   bool `json:"numa,omitempty"`
+	NumCtx    int  `json:"num_ctx,omitempty"`
+	NumBatch  int  `json:"num_batch,omitempty"`
+	NumGQA    int  `json:"num_gqa,omitempty"`
+	NumGPU    int  `json:"num_gpu,omitempty"`
+	MainGPU   int  `json:"main_gpu,omitempty"`
+	LowVRAM   bool `json:"low_vram,omitempty"`
+	F16KV     bool `json:"f16_kv,omitempty"`
+	LogitsAll bool `json:"logits_all,omitempty"`
+	VocabOnly bool `json:"vocab_only,omitempty"`
+	UseMMap   bool `json:"use_mmap,omitempty"`
+	UseMLock  bool `json:"use_mlock,omitempty"`
+	NumThread int  `json:"num_thread,omitempty"`
+
+	// Unused: RopeFrequencyBase is ignored. Instead the value in the model will be used
+	RopeFrequencyBase float32 `json:"rope_frequency_base,omitempty"`
+	// Unused: RopeFrequencyScale is ignored. Instead the value in the model will be used
 	RopeFrequencyScale float32 `json:"rope_frequency_scale,omitempty"`
-	NumThread          int     `json:"num_thread,omitempty"`
 }

 type EmbeddingRequest struct {
@ -139,10 +171,11 @@ type EmbeddingResponse struct {
 }

 type CreateRequest struct {
-	Model     string `json:"model"`
-	Path      string `json:"path"`
-	Modelfile string `json:"modelfile"`
-	Stream    *bool  `json:"stream,omitempty"`
+	Model        string `json:"model"`
+	Path         string `json:"path"`
+	Modelfile    string `json:"modelfile"`
+	Stream       *bool  `json:"stream,omitempty"`
+	Quantization string `json:"quantization,omitempty"`

 	// Name is deprecated, see Model
 	Name string `json:"name"`
@ -275,7 +308,8 @@ func (m *Metrics) Summary() {
 	}
 }

-var ErrInvalidOpts = fmt.Errorf("invalid options")
+var ErrInvalidOpts = errors.New("invalid options")
+var ErrInvalidHostPort = errors.New("invalid port specified in OLLAMA_HOST")

 func (opts *Options) FromMap(m map[string]interface{}) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
@ -363,8 +397,10 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
-		NumPredict:       -1,
-		NumKeep:          0,
+		NumPredict: -1,
+
+		// set a minimal num_keep to avoid issues on context shifts
+		NumKeep:          4,
 		Temperature:      0.8,
 		TopK:             40,
 		TopP:             0.9,
@ -382,18 +418,16 @@ func DefaultOptions() Options {

 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:             2048,
-			RopeFrequencyBase:  10000.0,
-			RopeFrequencyScale: 1.0,
-			NumBatch:           512,
-			NumGPU:             -1, // -1 here indicates that NumGPU should be set dynamically
-			NumGQA:             1,
-			NumThread:          0, // let the runtime decide
-			LowVRAM:            false,
-			F16KV:              true,
-			UseMLock:           false,
-			UseMMap:            true,
-			UseNUMA:            false,
+			NumCtx:    2048,
+			NumBatch:  512,
+			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
+			NumGQA:    1,
+			NumThread: 0, // let the runtime decide
+			LowVRAM:   false,
+			F16KV:     true,
+			UseMLock:  false,
+			UseMMap:   true,
+			UseNUMA:   false,
 		},
 	}
 }
--- a/api/types_test.go
+++ b/api/types_test.go
@ -0,0 +1,50 @@
+package api
+
+import (
+	"encoding/json"
+	"math"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestKeepAliveParsingFromJSON(t *testing.T) {
+	tests := []struct {
+		name string
+		req  string
+		exp  *Duration
+	}{
+		{
+			name: "Positive Integer",
+			req:  `{ "keep_alive": 42 }`,
+			exp:  &Duration{42 * time.Second},
+		},
+		{
+			name: "Positive Integer String",
+			req:  `{ "keep_alive": "42m" }`,
+			exp:  &Duration{42 * time.Minute},
+		},
+		{
+			name: "Negative Integer",
+			req:  `{ "keep_alive": -1 }`,
+			exp:  &Duration{math.MaxInt64},
+		},
+		{
+			name: "Negative Integer String",
+			req:  `{ "keep_alive": "-1m" }`,
+			exp:  &Duration{math.MaxInt64},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var dec ChatRequest
+			err := json.Unmarshal([]byte(test.req), &dec)
+			require.NoError(t, err)
+
+			assert.Equal(t, test.exp, dec.KeepAlive)
+		})
+	}
+}
--- a/app/lifecycle/lifecycle.go
+++ b/app/lifecycle/lifecycle.go
@ -9,8 +9,8 @@ import (
 	"os/signal"
 	"syscall"

-	"github.com/jmorganca/ollama/app/store"
-	"github.com/jmorganca/ollama/app/tray"
+	"github.com/ollama/ollama/app/store"
+	"github.com/ollama/ollama/app/tray"
 )

 func Run() {
--- a/app/lifecycle/server.go
+++ b/app/lifecycle/server.go
@ -11,7 +11,7 @@ import (
 	"path/filepath"
 	"time"

-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )

 func getCLIFullPath(command string) string {
@ -43,37 +43,36 @@ func getCLIFullPath(command string) string {
 	return command
 }

-func SpawnServer(ctx context.Context, command string) (chan int, error) {
-	done := make(chan int)
-
-	logDir := filepath.Dir(ServerLogFile)
-	_, err := os.Stat(logDir)
-	if errors.Is(err, os.ErrNotExist) {
-		if err := os.MkdirAll(logDir, 0o755); err != nil {
-			return done, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
-		}
-	}
-
+func start(ctx context.Context, command string) (*exec.Cmd, error) {
 	cmd := getCmd(ctx, getCLIFullPath(command))
-	// send stdout and stderr to a file
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
-		return done, fmt.Errorf("failed to spawn server stdout pipe %s", err)
+		return nil, fmt.Errorf("failed to spawn server stdout pipe: %w", err)
 	}
 	stderr, err := cmd.StderrPipe()
 	if err != nil {
-		return done, fmt.Errorf("failed to spawn server stderr pipe %s", err)
-	}
-	stdin, err := cmd.StdinPipe()
-	if err != nil {
-		return done, fmt.Errorf("failed to spawn server stdin pipe %s", err)
+		return nil, fmt.Errorf("failed to spawn server stderr pipe: %w", err)
 	}

 	// TODO - rotation
 	logFile, err := os.OpenFile(ServerLogFile, os.O_APPEND|os.O_WRONLY|os.O_CREATE, 0755)
 	if err != nil {
-		return done, fmt.Errorf("failed to create server log %w", err)
+		return nil, fmt.Errorf("failed to create server log: %w", err)
 	}
+
+	logDir := filepath.Dir(ServerLogFile)
+	_, err = os.Stat(logDir)
+	if err != nil {
+		if !errors.Is(err, os.ErrNotExist) {
+			return nil, fmt.Errorf("stat ollama server log dir %s: %v", logDir, err)
+
+		}
+
+		if err := os.MkdirAll(logDir, 0o755); err != nil {
+			return nil, fmt.Errorf("create ollama server log dir %s: %v", logDir, err)
+		}
+	}
+
 	go func() {
 		defer logFile.Close()
 		io.Copy(logFile, stdout) //nolint:errcheck
@ -83,21 +82,67 @@ func SpawnServer(ctx context.Context, command string) (chan int, error) {
 		io.Copy(logFile, stderr) //nolint:errcheck
 	}()

+	// Re-wire context done behavior to attempt a graceful shutdown of the server
+	cmd.Cancel = func() error {
+		if cmd.Process != nil {
+			err := terminate(cmd)
+			if err != nil {
+				slog.Warn("error trying to gracefully terminate server", "err", err)
+				return cmd.Process.Kill()
+			}
+
+			tick := time.NewTicker(10 * time.Millisecond)
+			defer tick.Stop()
+
+			for {
+				select {
+				case <-tick.C:
+					exited, err := isProcessExited(cmd.Process.Pid)
+					if err != nil {
+						return err
+					}
+
+					if exited {
+						return nil
+					}
+				case <-time.After(5 * time.Second):
+					slog.Warn("graceful server shutdown timeout, killing", "pid", cmd.Process.Pid)
+					return cmd.Process.Kill()
+				}
+			}
+		}
+		return nil
+	}
+
 	// run the command and wait for it to finish
 	if err := cmd.Start(); err != nil {
-		return done, fmt.Errorf("failed to start server %w", err)
+		return nil, fmt.Errorf("failed to start server %w", err)
 	}
 	if cmd.Process != nil {
 		slog.Info(fmt.Sprintf("started ollama server with pid %d", cmd.Process.Pid))
 	}
 	slog.Info(fmt.Sprintf("ollama server logs %s", ServerLogFile))

+	return cmd, nil
+}
+
+func SpawnServer(ctx context.Context, command string) (chan int, error) {
+	done := make(chan int)
+
 	go func() {
 		// Keep the server running unless we're shuttind down the app
 		crashCount := 0
 		for {
+			slog.Info("starting server...")
+			cmd, err := start(ctx, command)
+			if err != nil {
+				crashCount++
+				slog.Error(fmt.Sprintf("failed to start server %s", err))
+				time.Sleep(500 * time.Millisecond * time.Duration(crashCount))
+				continue
+			}
+
 			cmd.Wait() //nolint:errcheck
-			stdin.Close()
 			var code int
 			if cmd.ProcessState != nil {
 				code = cmd.ProcessState.ExitCode()
@ -105,21 +150,18 @@ func SpawnServer(ctx context.Context, command string) (chan int, error) {

 			select {
 			case <-ctx.Done():
-				slog.Debug(fmt.Sprintf("server shutdown with exit code %d", code))
+				slog.Info(fmt.Sprintf("server shutdown with exit code %d", code))
 				done <- code
 				return
 			default:
 				crashCount++
 				slog.Warn(fmt.Sprintf("server crash %d - exit code %d - respawning", crashCount, code))
-				time.Sleep(500 * time.Millisecond)
-				if err := cmd.Start(); err != nil {
-					slog.Error(fmt.Sprintf("failed to restart server %s", err))
-					// Keep trying, but back off if we keep failing
-					time.Sleep(time.Duration(crashCount) * time.Second)
-				}
+				time.Sleep(500 * time.Millisecond * time.Duration(crashCount))
+				break
 			}
 		}
 	}()
+
 	return done, nil
 }

--- a/app/lifecycle/server_unix.go
+++ b/app/lifecycle/server_unix.go
@ -4,9 +4,35 @@ package lifecycle

 import (
 	"context"
+	"errors"
+	"fmt"
+	"os"
 	"os/exec"
+	"syscall"
 )

 func getCmd(ctx context.Context, cmd string) *exec.Cmd {
 	return exec.CommandContext(ctx, cmd, "serve")
 }
+
+func terminate(cmd *exec.Cmd) error {
+	return cmd.Process.Signal(os.Interrupt)
+}
+
+func isProcessExited(pid int) (bool, error) {
+	proc, err := os.FindProcess(pid)
+	if err != nil {
+		return false, fmt.Errorf("failed to find process: %v", err)
+	}
+
+	err = proc.Signal(syscall.Signal(0))
+	if err != nil {
+		if errors.Is(err, os.ErrProcessDone) || errors.Is(err, syscall.ESRCH) {
+			return true, nil
+		}
+
+		return false, fmt.Errorf("error signaling process: %v", err)
+	}
+
+	return false, nil
+}
--- a/app/lifecycle/server_windows.go
+++ b/app/lifecycle/server_windows.go
@ -2,12 +2,88 @@ package lifecycle

 import (
 	"context"
+	"fmt"
 	"os/exec"
 	"syscall"
+
+	"golang.org/x/sys/windows"
 )

 func getCmd(ctx context.Context, exePath string) *exec.Cmd {
 	cmd := exec.CommandContext(ctx, exePath, "serve")
-	cmd.SysProcAttr = &syscall.SysProcAttr{HideWindow: true, CreationFlags: 0x08000000}
+	cmd.SysProcAttr = &syscall.SysProcAttr{
+		HideWindow:    true,
+		CreationFlags: windows.CREATE_NEW_PROCESS_GROUP,
+	}
+
 	return cmd
 }
+
+func terminate(cmd *exec.Cmd) error {
+	dll, err := windows.LoadDLL("kernel32.dll")
+	if err != nil {
+		return err
+	}
+	defer dll.Release() // nolint: errcheck
+
+	pid := cmd.Process.Pid
+
+	f, err := dll.FindProc("AttachConsole")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err := f.Call(uintptr(pid))
+	if r1 == 0 && err != syscall.ERROR_ACCESS_DENIED {
+		return err
+	}
+
+	f, err = dll.FindProc("SetConsoleCtrlHandler")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err = f.Call(0, 1)
+	if r1 == 0 {
+		return err
+	}
+
+	f, err = dll.FindProc("GenerateConsoleCtrlEvent")
+	if err != nil {
+		return err
+	}
+
+	r1, _, err = f.Call(windows.CTRL_BREAK_EVENT, uintptr(pid))
+	if r1 == 0 {
+		return err
+	}
+
+	r1, _, err = f.Call(windows.CTRL_C_EVENT, uintptr(pid))
+	if r1 == 0 {
+		return err
+	}
+
+	return nil
+}
+
+const STILL_ACTIVE = 259
+
+func isProcessExited(pid int) (bool, error) {
+	hProcess, err := windows.OpenProcess(windows.PROCESS_QUERY_INFORMATION, false, uint32(pid))
+	if err != nil {
+		return false, fmt.Errorf("failed to open process: %v", err)
+	}
+	defer windows.CloseHandle(hProcess) // nolint: errcheck
+
+	var exitCode uint32
+	err = windows.GetExitCodeProcess(hProcess, &exitCode)
+	if err != nil {
+		return false, fmt.Errorf("failed to get exit code: %v", err)
+	}
+
+	if exitCode == STILL_ACTIVE {
+		return false, nil
+	}
+
+	return true, nil
+}
--- a/app/lifecycle/updater.go
+++ b/app/lifecycle/updater.go
@ -18,8 +18,8 @@ import (
 	"strings"
 	"time"

-	"github.com/jmorganca/ollama/auth"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/auth"
+	"github.com/ollama/ollama/version"
 )

 var (
--- a/app/main.go
+++ b/app/main.go
@ -4,7 +4,7 @@ package main
 // go build -ldflags="-H windowsgui" .

 import (
-	"github.com/jmorganca/ollama/app/lifecycle"
+	"github.com/ollama/ollama/app/lifecycle"
 )

 func main() {
--- a/app/ollama.iss
+++ b/app/ollama.iss
@ -28,8 +28,8 @@ AppPublisher={#MyAppPublisher}
 AppPublisherURL={#MyAppURL}
 AppSupportURL={#MyAppURL}
 AppUpdatesURL={#MyAppURL}
-ArchitecturesAllowed=x64
-ArchitecturesInstallIn64BitMode=x64
+ArchitecturesAllowed=x64 arm64
+ArchitecturesInstallIn64BitMode=x64 arm64
 DefaultDirName={localappdata}\Programs\{#MyAppName}
 DefaultGroupName={#MyAppName}
 DisableProgramGroupPage=yes
@ -88,15 +88,12 @@ DialogFontSize=12
 [Files]
 Source: ".\app.exe"; DestDir: "{app}"; DestName: "{#MyAppExeName}" ; Flags: ignoreversion 64bit
 Source: "..\ollama.exe"; DestDir: "{app}"; Flags: ignoreversion 64bit
-Source: "..\dist\windeps\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\windows-{#ARCH}\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64bit
+Source: "..\dist\windows-{#ARCH}\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-; Assumes v5.7, may need adjustments for v6
-#if GetEnv("HIP_PATH") != ""
-  Source: "{#GetEnv('HIP_PATH')}\bin\hipblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
-  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
-  ; amdhip64.dll dependency comes from the driver and must be installed already
-  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas\library\*"; DestDir: "{app}\rocm\rocblas\library\"; Flags: ignoreversion
+#if DirExists("..\dist\windows-amd64\rocm")
+  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
 #endif


@ -132,7 +129,7 @@ SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or fi


 ;FinishedHeadingLabel=Run your first model
-;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama2
+;FinishedLabel=%nRun this command in a PowerShell or cmd terminal.%n%n%n    ollama run llama3
 ;ClickFinish=%n

 [Registry]
--- a/app/tray/tray.go
+++ b/app/tray/tray.go
@ -4,8 +4,8 @@ import (
 	"fmt"
 	"runtime"

-	"github.com/jmorganca/ollama/app/assets"
-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/assets"
+	"github.com/ollama/ollama/app/tray/commontray"
 )

 func NewTray() (commontray.OllamaTray, error) {
@ -24,10 +24,5 @@ func NewTray() (commontray.OllamaTray, error) {
 		return nil, fmt.Errorf("failed to load icon %s: %w", iconName, err)
 	}

-	tray, err := InitPlatformTray(icon, updateIcon)
-	if err != nil {
-		return nil, err
-	}
-
-	return tray, nil
+	return InitPlatformTray(icon, updateIcon)
 }
--- a/app/tray/tray_nonwindows.go
+++ b/app/tray/tray_nonwindows.go
@ -5,7 +5,7 @@ package tray
 import (
 	"fmt"

-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/commontray"
 )

 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
--- a/app/tray/tray_windows.go
+++ b/app/tray/tray_windows.go
@ -1,8 +1,8 @@
 package tray

 import (
-	"github.com/jmorganca/ollama/app/tray/commontray"
-	"github.com/jmorganca/ollama/app/tray/wintray"
+	"github.com/ollama/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/wintray"
 )

 func InitPlatformTray(icon, updateIcon []byte) (commontray.OllamaTray, error) {
--- a/app/tray/wintray/tray.go
+++ b/app/tray/wintray/tray.go
@ -13,7 +13,7 @@ import (
 	"sync"
 	"unsafe"

-	"github.com/jmorganca/ollama/app/tray/commontray"
+	"github.com/ollama/ollama/app/tray/commontray"
 	"golang.org/x/sys/windows"
 )

--- a/auth/auth.go
+++ b/auth/auth.go
@ -10,12 +10,44 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
+	"strings"

 	"golang.org/x/crypto/ssh"
 )

 const defaultPrivateKey = "id_ed25519"

+func keyPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Join(home, ".ollama", defaultPrivateKey), nil
+}
+
+func GetPublicKey() (string, error) {
+	keyPath, err := keyPath()
+	if err != nil {
+		return "", err
+	}
+
+	privateKeyFile, err := os.ReadFile(keyPath)
+	if err != nil {
+		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
+		return "", err
+	}
+
+	privateKey, err := ssh.ParsePrivateKey(privateKeyFile)
+	if err != nil {
+		return "", err
+	}
+
+	publicKey := ssh.MarshalAuthorizedKey(privateKey.PublicKey())
+
+	return strings.TrimSpace(string(publicKey)), nil
+}
+
 func NewNonce(r io.Reader, length int) (string, error) {
 	nonce := make([]byte, length)
 	if _, err := io.ReadFull(r, nonce); err != nil {
@ -26,13 +58,11 @@ func NewNonce(r io.Reader, length int) (string, error) {
 }

 func Sign(ctx context.Context, bts []byte) (string, error) {
-	home, err := os.UserHomeDir()
+	keyPath, err := keyPath()
 	if err != nil {
 		return "", err
 	}

-	keyPath := filepath.Join(home, ".ollama", defaultPrivateKey)
-
 	privateKeyFile, err := os.ReadFile(keyPath)
 	if err != nil {
 		slog.Info(fmt.Sprintf("Failed to load private key: %v", err))
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@ -17,6 +17,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strings"
 	"syscall"
@ -30,12 +31,15 @@ import (
 	"golang.org/x/exp/slices"
 	"golang.org/x/term"

-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/format"
-	"github.com/jmorganca/ollama/parser"
-	"github.com/jmorganca/ollama/progress"
-	"github.com/jmorganca/ollama/server"
-	"github.com/jmorganca/ollama/version"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/auth"
+	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/server"
+	"github.com/ollama/ollama/types/errtypes"
+	"github.com/ollama/ollama/types/model"
+	"github.com/ollama/ollama/version"
 )

 func CreateHandler(cmd *cobra.Command, args []string) error {
@ -53,14 +57,13 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()

-	bars := make(map[string]*progress.Bar)
-
-	modelfile, err := os.ReadFile(filename)
+	modelfile, err := os.Open(filename)
 	if err != nil {
 		return err
 	}
+	defer modelfile.Close()

-	commands, err := parser.Parse(bytes.NewReader(modelfile))
+	commands, err := parser.Parse(modelfile)
 	if err != nil {
 		return err
 	}
@ -74,10 +77,10 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	spinner := progress.NewSpinner(status)
 	p.Add(status, spinner)

-	for _, c := range commands {
-		switch c.Name {
+	for i := range commands {
+		switch commands[i].Name {
 		case "model", "adapter":
-			path := c.Args
+			path := commands[i].Args
 			if path == "~" {
 				path = home
 			} else if strings.HasPrefix(path, "~/") {
@ -89,77 +92,22 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 			}

 			fi, err := os.Stat(path)
-			if errors.Is(err, os.ErrNotExist) && c.Name == "model" {
+			if errors.Is(err, os.ErrNotExist) && commands[i].Name == "model" {
 				continue
 			} else if err != nil {
 				return err
 			}

-			// TODO make this work w/ adapters
 			if fi.IsDir() {
-				tf, err := os.CreateTemp("", "ollama-tf")
+				// this is likely a safetensors or pytorch directory
+				// TODO make this work w/ adapters
+				tempfile, err := tempZipFiles(path)
 				if err != nil {
 					return err
 				}
-				defer os.RemoveAll(tf.Name())
+				defer os.RemoveAll(tempfile)

-				zf := zip.NewWriter(tf)
-
-				files, err := filepath.Glob(filepath.Join(path, "model-*.safetensors"))
-				if err != nil {
-					return err
-				}
-
-				if len(files) == 0 {
-					return fmt.Errorf("no safetensors files were found in '%s'", path)
-				}
-
-				// add the safetensor config file + tokenizer
-				files = append(files, filepath.Join(path, "config.json"))
-				files = append(files, filepath.Join(path, "added_tokens.json"))
-				files = append(files, filepath.Join(path, "tokenizer.model"))
-
-				for _, fn := range files {
-					f, err := os.Open(fn)
-					if os.IsNotExist(err) && strings.HasSuffix(fn, "added_tokens.json") {
-						continue
-					} else if err != nil {
-						return err
-					}
-
-					fi, err := f.Stat()
-					if err != nil {
-						return err
-					}
-
-					h, err := zip.FileInfoHeader(fi)
-					if err != nil {
-						return err
-					}
-
-					h.Name = filepath.Base(fn)
-					h.Method = zip.Store
-
-					w, err := zf.CreateHeader(h)
-					if err != nil {
-						return err
-					}
-
-					_, err = io.Copy(w, f)
-					if err != nil {
-						return err
-					}
-
-				}
-
-				if err := zf.Close(); err != nil {
-					return err
-				}
-
-				if err := tf.Close(); err != nil {
-					return err
-				}
-				path = tf.Name()
+				path = tempfile
 			}

 			digest, err := createBlob(cmd, client, path)
@ -167,10 +115,11 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}

-			modelfile = bytes.ReplaceAll(modelfile, []byte(c.Args), []byte("@"+digest))
+			commands[i].Args = "@"+digest
 		}
 	}

+	bars := make(map[string]*progress.Bar)
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
 			spinner.Stop()
@ -194,7 +143,9 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}

-	request := api.CreateRequest{Name: args[0], Modelfile: string(modelfile)}
+	quantization, _ := cmd.Flags().GetString("quantization")
+
+	request := api.CreateRequest{Name: args[0], Modelfile: parser.Format(commands), Quantization: quantization}
 	if err := client.Create(cmd.Context(), &request, fn); err != nil {
 		return err
 	}
@ -202,6 +153,114 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }

+func tempZipFiles(path string) (string, error) {
+	tempfile, err := os.CreateTemp("", "ollama-tf")
+	if err != nil {
+		return "", err
+	}
+	defer tempfile.Close()
+
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
+	detectContentType := func(path string) (string, error) {
+		f, err := os.Open(path)
+		if err != nil {
+			return "", err
+		}
+		defer f.Close()
+
+		var b bytes.Buffer
+		b.Grow(512)
+
+		if _, err := io.CopyN(&b, f, 512); err != nil && !errors.Is(err, io.EOF) {
+			return "", err
+		}
+
+		contentType, _, _ := strings.Cut(http.DetectContentType(b.Bytes()), ";")
+		return contentType, nil
+	}
+
+	glob := func(pattern, contentType string) ([]string, error) {
+		matches, err := filepath.Glob(pattern)
+		if err != nil {
+			return nil, err
+		}
+
+		for _, safetensor := range matches {
+			if ct, err := detectContentType(safetensor); err != nil {
+				return nil, err
+			} else if ct != contentType {
+				return nil, fmt.Errorf("invalid content type: expected %s for %s", ct, safetensor)
+			}
+		}
+
+		return matches, nil
+	}
+
+	var files []string
+	if st, _ := glob(filepath.Join(path, "model*.safetensors"), "application/octet-stream"); len(st) > 0 {
+		// safetensors files might be unresolved git lfs references; skip if they are
+		// covers model-x-of-y.safetensors, model.fp32-x-of-y.safetensors, model.safetensors
+		files = append(files, st...)
+	} else if pt, _ := glob(filepath.Join(path, "pytorch_model*.bin"), "application/zip"); len(pt) > 0 {
+		// pytorch files might also be unresolved git lfs references; skip if they are
+		// covers pytorch_model-x-of-y.bin, pytorch_model.fp32-x-of-y.bin, pytorch_model.bin
+		files = append(files, pt...)
+	} else if pt, _ := glob(filepath.Join(path, "consolidated*.pth"), "application/octet-stream"); len(pt) > 0 {
+		// pytorch files might also be unresolved git lfs references; skip if they are
+		// covers consolidated.x.pth, consolidated.pth
+		files = append(files, pt...)
+	} else {
+		return "", errors.New("no safetensors or torch files found")
+	}
+
+	// add configuration files, json files are detected as text/plain
+	js, err := glob(filepath.Join(path, "*.json"), "text/plain")
+	if err != nil {
+		return "", err
+	}
+	files = append(files, js...)
+
+	if tks, _ := glob(filepath.Join(path, "tokenizer.model"), "application/octet-stream"); len(tks) > 0 {
+		// add tokenizer.model if it exists, tokenizer.json is automatically picked up by the previous glob
+		// tokenizer.model might be a unresolved git lfs reference; error if it is
+		files = append(files, tks...)
+	} else if tks, _ := glob(filepath.Join(path, "**/tokenizer.model"), "text/plain"); len(tks) > 0 {
+		// some times tokenizer.model is in a subdirectory (e.g. meta-llama/Meta-Llama-3-8B)
+		files = append(files, tks...)
+	}
+
+	for _, file := range files {
+		f, err := os.Open(file)
+		if err != nil {
+			return "", err
+		}
+		defer f.Close()
+
+		fi, err := f.Stat()
+		if err != nil {
+			return "", err
+		}
+
+		zfi, err := zip.FileInfoHeader(fi)
+		if err != nil {
+			return "", err
+		}
+
+		zf, err := zipfile.CreateHeader(zfi)
+		if err != nil {
+			return "", err
+		}
+
+		if _, err := io.Copy(zf, f); err != nil {
+			return "", err
+		}
+	}
+
+	return tempfile.Name(), nil
+}
+
 func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, error) {
 	bin, err := os.Open(path)
 	if err != nil {
@ -213,7 +272,10 @@ func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, er
 	if _, err := io.Copy(hash, bin); err != nil {
 		return "", err
 	}
-	bin.Seek(0, io.SeekStart)
+
+	if _, err := bin.Seek(0, io.SeekStart); err != nil {
+		return "", err
+	}

 	digest := fmt.Sprintf("sha256:%x", hash.Sum(nil))
 	if err = client.CreateBlob(cmd.Context(), digest, bin); err != nil {
@ -293,6 +355,47 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	return generateInteractive(cmd, opts)
 }

+func errFromUnknownKey(unknownKeyErr error) error {
+	// find SSH public key in the error message
+	sshKeyPattern := `ssh-\w+ [^\s"]+`
+	re := regexp.MustCompile(sshKeyPattern)
+	matches := re.FindStringSubmatch(unknownKeyErr.Error())
+
+	if len(matches) > 0 {
+		serverPubKey := matches[0]
+
+		localPubKey, err := auth.GetPublicKey()
+		if err != nil {
+			return unknownKeyErr
+		}
+
+		if runtime.GOOS == "linux" && serverPubKey != localPubKey {
+			// try the ollama service public key
+			svcPubKey, err := os.ReadFile("/usr/share/ollama/.ollama/id_ed25519.pub")
+			if err != nil {
+				return unknownKeyErr
+			}
+			localPubKey = strings.TrimSpace(string(svcPubKey))
+		}
+
+		// check if the returned public key matches the local public key, this prevents adding a remote key to the user's account
+		if serverPubKey != localPubKey {
+			return unknownKeyErr
+		}
+
+		var msg strings.Builder
+		msg.WriteString(unknownKeyErr.Error())
+		msg.WriteString("\n\nYour ollama key is:\n")
+		msg.WriteString(localPubKey)
+		msg.WriteString("\nAdd your key at:\n")
+		msg.WriteString("https://ollama.com/settings/keys")
+
+		return errors.New(msg.String())
+	}
+
+	return unknownKeyErr
+}
+
 func PushHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@ -340,6 +443,20 @@ func PushHandler(cmd *cobra.Command, args []string) error {

 	request := api.PushRequest{Name: args[0], Insecure: insecure}
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
+		if spinner != nil {
+			spinner.Stop()
+		}
+		if strings.Contains(err.Error(), "access denied") {
+			return errors.New("you are not authorized to push to this namespace, create the model under a namespace you own")
+		}
+		host := model.ParseName(args[0]).Host
+		isOllamaHost := strings.HasSuffix(host, ".ollama.ai") || strings.HasSuffix(host, ".ollama.com")
+		if strings.Contains(err.Error(), errtypes.UnknownOllamaKeyErrMsg) && isOllamaHost {
+			// the user has not added their ollama key to ollama.com
+			// re-throw an error with a more user-friendly message
+			return errFromUnknownKey(err)
+		}
+
 		return err
 	}

@ -767,19 +884,17 @@ func generate(cmd *cobra.Command, opts runOptions) error {
 }

 func RunServer(cmd *cobra.Command, _ []string) error {
-	host, port, err := net.SplitHostPort(strings.Trim(os.Getenv("OLLAMA_HOST"), "\"'"))
+	// retrieve the OLLAMA_HOST environment variable
+	ollamaHost, err := api.GetOllamaHost()
 	if err != nil {
-		host, port = "127.0.0.1", "11434"
-		if ip := net.ParseIP(strings.Trim(os.Getenv("OLLAMA_HOST"), "[]")); ip != nil {
-			host = ip.String()
-		}
+		return err
 	}

 	if err := initializeKeypair(); err != nil {
 		return err
 	}

-	ln, err := net.Listen("tcp", net.JoinHostPort(host, port))
+	ln, err := net.Listen("tcp", net.JoinHostPort(ollamaHost.Host, ollamaHost.Port))
 	if err != nil {
 		return err
 	}
@ -900,8 +1015,7 @@ func NewCLI() *cobra.Command {
 	cobra.EnableCommandSorting = false

 	if runtime.GOOS == "windows" {
-		// Enable colorful ANSI escape code in Windows terminal (disabled by default)
-		console.ConsoleFromFile(os.Stdout) //nolint:errcheck
+		console.ConsoleFromFile(os.Stdin) //nolint:errcheck
 	}

 	rootCmd := &cobra.Command{
@ -933,6 +1047,7 @@ func NewCLI() *cobra.Command {
 	}

 	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile (default \"Modelfile\")")
+	createCmd.Flags().StringP("quantization", "q", "", "Quantization level.")

 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@ -970,9 +1085,11 @@ func NewCLI() *cobra.Command {
 	serveCmd.SetUsageTemplate(serveCmd.UsageTemplate() + `
 Environment Variables:

-    OLLAMA_HOST       The host:port to bind to (default "127.0.0.1:11434")
-    OLLAMA_ORIGINS    A comma separated list of allowed origins.
-    OLLAMA_MODELS     The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_HOST         The host:port to bind to (default "127.0.0.1:11434")
+    OLLAMA_ORIGINS      A comma separated list of allowed origins.
+    OLLAMA_MODELS       The path to the models directory (default is "~/.ollama/models")
+    OLLAMA_KEEP_ALIVE   The duration that models stay loaded in memory (default is "5m")
+    OLLAMA_DEBUG        Set to 1 to enable additional debug logging
 `)

 	pullCmd := &cobra.Command{
@ -1003,7 +1120,7 @@ Environment Variables:
 		RunE:    ListHandler,
 	}
 	copyCmd := &cobra.Command{
-		Use:     "cp SOURCE TARGET",
+		Use:     "cp SOURCE DESTINATION",
 		Short:   "Copy a model",
 		Args:    cobra.ExactArgs(2),
 		PreRunE: checkServerHeartbeat,
--- a/cmd/interactive.go
+++ b/cmd/interactive.go
@ -14,9 +14,9 @@ import (
 	"github.com/spf13/cobra"
 	"golang.org/x/exp/slices"

-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/progress"
-	"github.com/jmorganca/ollama/readline"
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/progress"
+	"github.com/ollama/ollama/readline"
 )

 type MultilineState int
@ -94,6 +94,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /show           Show model information")
 		fmt.Fprintln(os.Stderr, "  /load <model>   Load a session or model")
 		fmt.Fprintln(os.Stderr, "  /save <model>   Save your current session")
+		fmt.Fprintln(os.Stderr, "  /clear          Clear session context")
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
@ -280,6 +281,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			}
 			fmt.Printf("Created new model '%s'\n", args[1])
 			continue
+		case strings.HasPrefix(line, "/clear"):
+			opts.Messages = []api.Message{}
+			fmt.Println("Cleared session context")
+			continue
 		case strings.HasPrefix(line, "/set"):
 			args := strings.Fields(line)
 			if len(args) > 1 {
@ -295,10 +300,14 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 					opts.WordWrap = false
 					fmt.Println("Set 'nowordwrap' mode.")
 				case "verbose":
-					cmd.Flags().Set("verbose", "true")
+					if err := cmd.Flags().Set("verbose", "true"); err != nil {
+						return err
+					}
 					fmt.Println("Set 'verbose' mode.")
 				case "quiet":
-					cmd.Flags().Set("verbose", "false")
+					if err := cmd.Flags().Set("verbose", "false"); err != nil {
+						return err
+					}
 					fmt.Println("Set 'quiet' mode.")
 				case "format":
 					if len(args) < 3 || args[2] != "json" {
--- a/cmd/interactive_test.go
+++ b/cmd/interactive_test.go
@ -7,7 +7,7 @@ import (

 	"github.com/stretchr/testify/assert"

-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )

 func TestExtractFilenames(t *testing.T) {
--- a/cmd/start_darwin.go
+++ b/cmd/start_darwin.go
@ -7,7 +7,7 @@ import (
 	"os/exec"
 	"strings"

-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )

 func startApp(ctx context.Context, client *api.Client) error {
--- a/cmd/start_default.go
+++ b/cmd/start_default.go
@ -6,7 +6,7 @@ import (
 	"context"
 	"fmt"

-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )

 func startApp(ctx context.Context, client *api.Client) error {
--- a/cmd/start_windows.go
+++ b/cmd/start_windows.go
@ -10,7 +10,7 @@ import (
 	"strings"
 	"syscall"

-	"github.com/jmorganca/ollama/api"
+	"github.com/ollama/ollama/api"
 )

 func startApp(ctx context.Context, client *api.Client) error {
--- a/convert/convert.go
+++ b/convert/convert.go
@ -1,166 +1,88 @@
 package convert

 import (
-	"bytes"
 	"cmp"
 	"encoding/binary"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
 	"os"
 	"path/filepath"
-	"regexp"
 	"slices"
+	"strings"

-	"github.com/mitchellh/mapstructure"
 	"google.golang.org/protobuf/proto"

-	"github.com/jmorganca/ollama/convert/sentencepiece"
-	"github.com/jmorganca/ollama/llm"
+	"github.com/ollama/ollama/convert/sentencepiece"
+	"github.com/ollama/ollama/llm"
 )

 type Params struct {
-	Architectures    []string `json:"architectures"`
-	VocabSize        int      `json:"vocab_size"`
-	HiddenSize       int      `json:"hidden_size"`       // n_embd
-	HiddenLayers     int      `json:"num_hidden_layers"` // n_layer
-	ContextSize      int      `json:"max_position_embeddings"`
-	IntermediateSize int      `json:"intermediate_size"`
-	AttentionHeads   int      `json:"num_attention_heads"` // n_head
-	KeyValHeads      int      `json:"num_key_value_heads"`
-	NormEPS          float64  `json:"rms_norm_eps"`
-	RopeFreqBase     float64  `json:"rope_theta"`
-	BoSTokenID       int      `json:"bos_token_id"`
-	EoSTokenID       int      `json:"eos_token_id"`
+	Architectures     []string `json:"architectures"`
+	VocabSize         int      `json:"vocab_size"`
+	HiddenSize        int      `json:"hidden_size"`       // n_embd
+	HiddenLayers      int      `json:"num_hidden_layers"` // n_layer
+	ContextSize       int      `json:"max_position_embeddings"`
+	IntermediateSize  int      `json:"intermediate_size"`
+	AttentionHeads    int      `json:"num_attention_heads"` // n_head
+	KeyValHeads       int      `json:"num_key_value_heads"`
+	NormEPS           float64  `json:"rms_norm_eps"`
+	BoSTokenID        int      `json:"bos_token_id"`
+	EoSTokenID        int      `json:"eos_token_id"`
+	HeadDimension     int      `json:"head_dim"`
+	PaddingTokenID    int      `json:"pad_token_id"`
+	RopeFrequencyBase float64  `json:"rope_theta"`
+
+	Experts     int `json:"num_local_experts"`
+	ExpertsUsed int `json:"num_experts_per_tok"`
+
+	ByteOrder
 }

-type MetaData struct {
-	Type    string `mapstructure:"dtype"`
-	Shape   []int  `mapstructure:"shape"`
-	Offsets []int  `mapstructure:"data_offsets"`
+type ByteOrder interface {
+	binary.ByteOrder
+	binary.AppendByteOrder
 }

-func ReadSafeTensors(fn string, offset uint64) ([]llm.Tensor, uint64, error) {
-	f, err := os.Open(fn)
-	if err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-	defer f.Close()
-
-	var jsonSize uint64
-	binary.Read(f, binary.LittleEndian, &jsonSize)
-
-	buf := make([]byte, jsonSize)
-	_, err = io.ReadFull(f, buf)
-	if err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-
-	d := json.NewDecoder(bytes.NewBuffer(buf))
-	d.UseNumber()
-	var parsed map[string]interface{}
-	if err = d.Decode(&parsed); err != nil {
-		return []llm.Tensor{}, 0, err
-	}
-
-	var keys []string
-	for k := range parsed {
-		keys = append(keys, k)
-	}
-
-	slices.Sort(keys)
-
-	slog.Info("converting layers")
-
-	var tensors []llm.Tensor
-	for _, k := range keys {
-		vals := parsed[k].(map[string]interface{})
-		var data MetaData
-		if err = mapstructure.Decode(vals, &data); err != nil {
-			return []llm.Tensor{}, 0, err
-		}
-
-		var size uint64
-		var kind uint32
-		switch len(data.Shape) {
-		case 0:
-			// metadata
-			continue
-		case 1:
-			// convert to float32
-			kind = 0
-			size = uint64(data.Shape[0] * 4)
-		case 2:
-			// convert to float16
-			kind = 1
-			size = uint64(data.Shape[0] * data.Shape[1] * 2)
-		}
-
-		ggufName, err := GetTensorName(k)
-		if err != nil {
-			slog.Error("%v", err)
-			return []llm.Tensor{}, 0, err
-		}
-
-		shape := []uint64{0, 0, 0, 0}
-		for i := range data.Shape {
-			shape[i] = uint64(data.Shape[i])
-		}
-
-		t := llm.Tensor{
-			Name:          ggufName,
-			Kind:          kind,
-			Offset:        offset,
-			Shape:         shape[:],
-			FileName:      fn,
-			OffsetPadding: 8 + jsonSize,
-			FileOffsets:   []uint64{uint64(data.Offsets[0]), uint64(data.Offsets[1])},
-		}
-		slog.Debug(fmt.Sprintf("%v", t))
-		tensors = append(tensors, t)
-		offset += size
-	}
-	return tensors, offset, nil
+type ModelArch interface {
+	GetTensors() error
+	LoadVocab() error
+	WriteGGUF() (string, error)
 }

-func GetSafeTensors(dirpath string) ([]llm.Tensor, error) {
-	var tensors []llm.Tensor
-	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
-	if err != nil {
-		return []llm.Tensor{}, err
-	}
-
-	var offset uint64
-	for _, f := range files {
-		var t []llm.Tensor
-		var err error
-		t, offset, err = ReadSafeTensors(f, offset)
-		if err != nil {
-			slog.Error("%v", err)
-			return []llm.Tensor{}, err
-		}
-		tensors = append(tensors, t...)
-	}
-	return tensors, nil
+type ModelFormat interface {
+	GetLayerName(string) (string, error)
+	GetTensors(string, *Params) ([]llm.Tensor, error)
+	GetParams(string) (*Params, error)
+	GetModelArch(string, string, *Params) (ModelArch, error)
 }

-func GetParams(dirpath string) (*Params, error) {
-	f, err := os.Open(filepath.Join(dirpath, "config.json"))
-	if err != nil {
-		return nil, err
-	}
-	defer f.Close()
+type ModelData struct {
+	Path    string
+	Name    string
+	Params  *Params
+	Vocab   *Vocab
+	Tensors []llm.Tensor
+	Format  ModelFormat
+}

-	var params Params
-
-	d := json.NewDecoder(f)
-	err = d.Decode(&params)
+func GetModelFormat(dirname string) (ModelFormat, error) {
+	files, err := filepath.Glob(filepath.Join(dirname, "*"))
 	if err != nil {
 		return nil, err
 	}

-	return &params, nil
+	for _, fn := range files {
+		slog.Debug(fmt.Sprintf("file = %s", fn))
+		if strings.HasSuffix(fn, ".safetensors") {
+			return &SafetensorFormat{}, nil
+		} else if strings.HasSuffix(fn, ".bin") {
+			slog.Debug("model is torch")
+			return &TorchFormat{}, nil
+		}
+	}
+
+	return nil, fmt.Errorf("couldn't determine model format")
 }

 // Details on gguf's tokenizer can be found at:
@ -171,7 +93,7 @@ type Vocab struct {
 	Types  []int32
 }

-func LoadTokens(dirpath string) (*Vocab, error) {
+func LoadSentencePieceTokens(dirpath string, params *Params) (*Vocab, error) {
 	slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
 	in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
 	if err != nil {
@ -196,6 +118,14 @@ func LoadTokens(dirpath string) (*Vocab, error) {
 		v.Tokens = append(v.Tokens, p.GetPiece())
 		v.Scores = append(v.Scores, p.GetScore())
 		t := p.GetType()
+		switch t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN:
+		case sentencepiece.ModelProto_SentencePiece_CONTROL:
+		case sentencepiece.ModelProto_SentencePiece_UNUSED:
+		case sentencepiece.ModelProto_SentencePiece_BYTE:
+		default:
+			t = sentencepiece.ModelProto_SentencePiece_NORMAL
+		}
 		v.Types = append(v.Types, int32(t))
 	}

@ -243,89 +173,15 @@ func LoadTokens(dirpath string) (*Vocab, error) {
 	}
 	slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))

-	return v, nil
-}
-
-func GetTensorName(n string) (string, error) {
-	tMap := map[string]string{
-		"model.embed_tokens.weight":                           "token_embd.weight",
-		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
-		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
-		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
-		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
-		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
-		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
-		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
-		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
-		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
-		"lm_head.weight":    "output.weight",
-		"model.norm.weight": "output_norm.weight",
-	}
-
-	v, ok := tMap[n]
-	if ok {
-		return v, nil
-	}
-
-	// quick hack to rename the layers to gguf format
-	for k, v := range tMap {
-		re := regexp.MustCompile(k)
-		newName := re.ReplaceAllString(n, v)
-		if newName != n {
-			return newName, nil
+	if params.VocabSize > len(v.Tokens) {
+		missingTokens := params.VocabSize - len(v.Tokens)
+		slog.Warn(fmt.Sprintf("vocab is missing %d tokens", missingTokens))
+		for cnt := 0; cnt < missingTokens; cnt++ {
+			v.Tokens = append(v.Tokens, fmt.Sprintf("<dummy%05d>", cnt+1))
+			v.Scores = append(v.Scores, -1)
+			v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
 		}
 	}

-	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
-}
-
-func WriteGGUF(name string, tensors []llm.Tensor, params *Params, vocab *Vocab) (string, error) {
-	c := llm.ContainerGGUF{
-		ByteOrder: binary.LittleEndian,
-	}
-
-	m := llm.NewGGUFModel(&c)
-	m.Tensors = tensors
-	m.KV["general.architecture"] = "llama"
-	m.KV["general.name"] = name
-	m.KV["llama.context_length"] = uint32(params.ContextSize)
-	m.KV["llama.embedding_length"] = uint32(params.HiddenSize)
-	m.KV["llama.block_count"] = uint32(params.HiddenLayers)
-	m.KV["llama.feed_forward_length"] = uint32(params.IntermediateSize)
-	m.KV["llama.rope.dimension_count"] = uint32(128)
-	m.KV["llama.attention.head_count"] = uint32(params.AttentionHeads)
-	m.KV["llama.attention.head_count_kv"] = uint32(params.KeyValHeads)
-	m.KV["llama.attention.layer_norm_rms_epsilon"] = float32(params.NormEPS)
-	m.KV["llama.rope.freq_base"] = float32(params.RopeFreqBase)
-	m.KV["general.file_type"] = uint32(1)
-	m.KV["tokenizer.ggml.model"] = "llama"
-
-	m.KV["tokenizer.ggml.tokens"] = vocab.Tokens
-	m.KV["tokenizer.ggml.scores"] = vocab.Scores
-	m.KV["tokenizer.ggml.token_type"] = vocab.Types
-
-	m.KV["tokenizer.ggml.bos_token_id"] = uint32(params.BoSTokenID)
-	m.KV["tokenizer.ggml.eos_token_id"] = uint32(params.EoSTokenID)
-	m.KV["tokenizer.ggml.unknown_token_id"] = uint32(0)
-	m.KV["tokenizer.ggml.add_bos_token"] = true
-	m.KV["tokenizer.ggml.add_eos_token"] = false
-
-	// llamacpp sets the chat template, however we don't need to set it since we pass it in through a layer
-	// m.KV["tokenizer.chat_template"] = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" // XXX removeme
-
-	c.V3.NumTensor = uint64(len(tensors))
-	c.V3.NumKV = uint64(len(m.KV))
-
-	f, err := os.CreateTemp("", "ollama-gguf")
-	if err != nil {
-		return "", err
-	}
-	defer f.Close()
-
-	err = m.Encode(f)
-	if err != nil {
-		return "", err
-	}
-
-	return f.Name(), nil
+	return v, nil
 }
--- a/convert/gemma.go
+++ b/convert/gemma.go
@ -0,0 +1,137 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type GemmaModel struct {
+	ModelData
+}
+
+func gemmaLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	slog.Debug(fmt.Sprintf("converting '%s'", r.t.Name))
+
+	data := make([]byte, r.end-r.start)
+	if err := binary.Read(f, r.bo, data); err != nil {
+		return err
+	}
+
+	tDataF32 := bfloat16.DecodeFloat32(data)
+
+	var err error
+	tDataF32, err = addOnes(tDataF32, int(r.t.Shape[0]))
+	if err != nil {
+		return err
+	}
+
+	if err := binary.Write(w, r.bo, tDataF32); err != nil {
+		return err
+	}
+	return nil
+}
+
+func addOnes(data []float32, vectorSize int) ([]float32, error) {
+	n := tensor.New(tensor.WithShape(vectorSize), tensor.WithBacking(data))
+	ones := tensor.Ones(tensor.Float32, vectorSize)
+
+	var err error
+	n, err = n.Add(ones)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	newN, err := native.SelectF32(n, 0)
+	if err != nil {
+		return []float32{}, err
+	}
+
+	var fullTensor []float32
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+
+	return fullTensor, nil
+}
+
+func (m *GemmaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug(fmt.Sprintf("Total tensors: %d", len(t)))
+
+	m.Tensors = []llm.Tensor{}
+	for _, l := range t {
+		if strings.HasSuffix(l.Name, "norm.weight") {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = gemmaLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *GemmaModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *GemmaModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "gemma",
+		"general.name":                           m.Name,
+		"gemma.context_length":                   uint32(m.Params.ContextSize),
+		"gemma.embedding_length":                 uint32(m.Params.HiddenSize),
+		"gemma.block_count":                      uint32(m.Params.HiddenLayers),
+		"gemma.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"gemma.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"gemma.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"gemma.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"gemma.attention.key_length":             uint32(m.Params.HeadDimension),
+		"gemma.attention.value_length":           uint32(m.Params.HeadDimension),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.padding_token_id": uint32(m.Params.PaddingTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(3),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}
--- a/convert/llama.go
+++ b/convert/llama.go
@ -0,0 +1,176 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type LlamaModel struct {
+	ModelData
+}
+
+func llamaLayerHandler(w io.Writer, r torchWriterTo) error {
+	slog.Debug(fmt.Sprintf("repacking layer '%s'", r.t.Name))
+
+	data := r.storage.(*pytorch.HalfStorage).Data
+	tData := make([]uint16, len(data))
+	for cnt, v := range data {
+		tData[cnt] = uint16(float16.Fromfloat32(v))
+	}
+
+	var err error
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	slog.Debug(fmt.Sprintf("heads = %d", heads))
+
+	tData, err = llamaRepack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	if err = binary.Write(w, r.bo, tData); err != nil {
+		return err
+	}
+	return nil
+}
+
+func llamaRepack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *LlamaModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			slog.Debug(fmt.Sprintf("setting handler for: %s", l.Name))
+			wt := l.WriterTo.(torchWriterTo)
+			wt.handler = llamaLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *LlamaModel) LoadVocab() error {
+	var v *Vocab
+	var err error
+
+	slog.Debug("loading vocab")
+	v, err = LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	slog.Debug("vocab loaded")
+
+	m.Vocab = v
+	return nil
+}
+
+func (m *LlamaModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.vocab_size":                       uint32(len(m.Vocab.Tokens)),
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	slog.Debug(fmt.Sprintf("gguf file = %s", f.Name()))
+
+	return f.Name(), nil
+}
--- a/convert/mistral.go
+++ b/convert/mistral.go
@ -0,0 +1,173 @@
+package convert
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"os"
+	"regexp"
+	"strings"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MistralModel struct {
+	ModelData
+}
+
+func mistralLayerHandler(w io.Writer, r safetensorWriterTo, f *os.File) error {
+	layerSize := r.end - r.start
+
+	var err error
+	tData := make([]uint16, layerSize/2)
+	if err = binary.Read(f, r.bo, tData); err != nil {
+		return err
+	}
+
+	var heads uint32
+	if strings.Contains(r.t.Name, "attn_q") {
+		heads = uint32(r.params.AttentionHeads)
+	} else if strings.Contains(r.t.Name, "attn_k") {
+		heads = uint32(r.params.KeyValHeads)
+		if heads == 0 {
+			heads = uint32(r.params.AttentionHeads)
+		}
+	} else {
+		return fmt.Errorf("unknown layer type")
+	}
+
+	tData, err = repack(tData, int(heads), r.t.Shape)
+	if err != nil {
+		return err
+	}
+
+	var buf []byte
+	for _, n := range tData {
+		buf = r.bo.AppendUint16(buf, n)
+	}
+
+	tempBuf := make([]uint16, len(tData))
+	tDataF32 := bfloat16.DecodeFloat32(buf)
+	for cnt, v := range tDataF32 {
+		tDataF16 := float16.Fromfloat32(v)
+		tempBuf[cnt] = uint16(tDataF16)
+	}
+
+	if err = binary.Write(w, r.bo, tempBuf); err != nil {
+		return err
+	}
+	return nil
+}
+
+func repack(data []uint16, heads int, shape []uint64) ([]uint16, error) {
+	n := tensor.New(tensor.WithShape(int(shape[0]), int(shape[1])), tensor.WithBacking(data))
+	origShape := n.Shape().Clone()
+
+	// reshape the tensor and swap axes 1 and 2 to unpack the layer for gguf
+	if err := n.Reshape(heads, 2, origShape[0]/heads/2, origShape[1]); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(origShape...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+	newN, err := native.SelectU16(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var fullTensor []uint16
+	for _, v := range newN {
+		fullTensor = append(fullTensor, v...)
+	}
+	return fullTensor, nil
+}
+
+func (m *MistralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = mistralLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *MistralModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *MistralModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":                   "llama",
+		"general.name":                           m.Name,
+		"llama.context_length":                   uint32(m.Params.ContextSize),
+		"llama.embedding_length":                 uint32(m.Params.HiddenSize),
+		"llama.block_count":                      uint32(m.Params.HiddenLayers),
+		"llama.feed_forward_length":              uint32(m.Params.IntermediateSize),
+		"llama.rope.dimension_count":             uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+		"llama.attention.head_count":             uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv":          uint32(m.Params.KeyValHeads),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+		"general.file_type":                      uint32(1),
+		"tokenizer.ggml.model":                   "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}
--- a/convert/mixtral.go
+++ b/convert/mixtral.go
@ -0,0 +1,96 @@
+package convert
+
+import (
+	"os"
+	"regexp"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type MixtralModel struct {
+	ModelData
+}
+
+func (m *MixtralModel) GetTensors() error {
+	t, err := m.Format.GetTensors(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+
+	m.Tensors = []llm.Tensor{}
+
+	pattern := `^blk\.[0-9]+\.attn_(?P<layer>q|k)\.weight$`
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return err
+	}
+
+	for _, l := range t {
+		matches := re.FindAllStringSubmatch(l.Name, -1)
+		if len(matches) > 0 {
+			wt := l.WriterTo.(safetensorWriterTo)
+			wt.handler = mistralLayerHandler
+			l.WriterTo = wt
+		}
+		m.Tensors = append(m.Tensors, l)
+	}
+
+	return nil
+}
+
+func (m *MixtralModel) LoadVocab() error {
+	v, err := LoadSentencePieceTokens(m.Path, m.Params)
+	if err != nil {
+		return err
+	}
+	m.Vocab = v
+	return nil
+}
+
+func (m *MixtralModel) WriteGGUF() (string, error) {
+	kv := llm.KV{
+		"general.architecture":          "llama",
+		"general.name":                  m.Name,
+		"llama.block_count":             uint32(m.Params.HiddenLayers),
+		"llama.context_length":          uint32(m.Params.ContextSize),
+		"llama.embedding_length":        uint32(m.Params.HiddenSize),
+		"llama.feed_forward_length":     uint32(m.Params.IntermediateSize),
+		"llama.attention.head_count":    uint32(m.Params.AttentionHeads),
+		"llama.attention.head_count_kv": uint32(m.Params.KeyValHeads),
+
+		"llama.rope.freq_base":                   float32(m.Params.RopeFrequencyBase),
+		"llama.attention.layer_norm_rms_epsilon": float32(m.Params.NormEPS),
+
+		"llama.expert_count":      uint32(m.Params.Experts),
+		"llama.expert_used_count": uint32(m.Params.ExpertsUsed),
+
+		"llama.vocab_size":           uint32(len(m.Vocab.Tokens)),
+		"llama.rope.dimension_count": uint32(m.Params.HiddenSize / m.Params.AttentionHeads),
+
+		"general.file_type":    uint32(1),
+		"tokenizer.ggml.model": "llama",
+
+		"tokenizer.ggml.tokens":     m.Vocab.Tokens,
+		"tokenizer.ggml.scores":     m.Vocab.Scores,
+		"tokenizer.ggml.token_type": m.Vocab.Types,
+
+		"tokenizer.ggml.bos_token_id":     uint32(m.Params.BoSTokenID),
+		"tokenizer.ggml.eos_token_id":     uint32(m.Params.EoSTokenID),
+		"tokenizer.ggml.unknown_token_id": uint32(0),
+		"tokenizer.ggml.add_bos_token":    true,
+		"tokenizer.ggml.add_eos_token":    false,
+	}
+
+	f, err := os.CreateTemp("", "ollama-gguf")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	mod := llm.NewGGUFV3(m.Params.ByteOrder)
+	if err := mod.Encode(f, kv, m.Tensors); err != nil {
+		return "", err
+	}
+
+	return f.Name(), nil
+}
--- a/convert/safetensors.go
+++ b/convert/safetensors.go
@ -0,0 +1,317 @@
+package convert
+
+import (
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"slices"
+
+	"github.com/d4l3k/go-bfloat16"
+	"github.com/mitchellh/mapstructure"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type safetensorWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	filename string
+
+	start, end, padding uint64
+	handler             func(w io.Writer, r safetensorWriterTo, f *os.File) error
+}
+
+type tensorMetaData struct {
+	Type    string `mapstructure:"dtype"`
+	Shape   []int  `mapstructure:"shape"`
+	Offsets []int  `mapstructure:"data_offsets"`
+}
+
+type SafetensorFormat struct{}
+
+func (m *SafetensorFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting tensor data")
+	var tensors []llm.Tensor
+	files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
+	if err != nil {
+		return nil, err
+	}
+
+	var offset uint64
+	for _, f := range files {
+		var t []llm.Tensor
+		var err error
+		t, offset, err = m.readTensors(f, offset, params)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, err
+		}
+		tensors = append(tensors, t...)
+	}
+	slog.Debug(fmt.Sprintf("all tensors = %d", len(tensors)))
+	return tensors, nil
+}
+
+func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params) ([]llm.Tensor, uint64, error) {
+	f, err := os.Open(fn)
+	if err != nil {
+		return nil, 0, err
+	}
+	defer f.Close()
+
+	var jsonSize uint64
+	if err := binary.Read(f, binary.LittleEndian, &jsonSize); err != nil {
+		return nil, 0, err
+	}
+
+	buf := make([]byte, jsonSize)
+	_, err = io.ReadFull(f, buf)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	d := json.NewDecoder(bytes.NewBuffer(buf))
+	d.UseNumber()
+	var parsed map[string]interface{}
+	if err = d.Decode(&parsed); err != nil {
+		return nil, 0, err
+	}
+
+	var keys []string
+	for k := range parsed {
+		keys = append(keys, k)
+	}
+
+	slices.Sort(keys)
+	slog.Info("converting layers")
+
+	var tensors []llm.Tensor
+	for _, k := range keys {
+		vals := parsed[k].(map[string]interface{})
+		var data tensorMetaData
+		if err = mapstructure.Decode(vals, &data); err != nil {
+			slog.Error("couldn't decode properly")
+			return nil, 0, err
+		}
+
+		var size uint64
+		var kind uint32
+		switch len(data.Shape) {
+		case 0:
+			// metadata
+			continue
+		case 1:
+			// convert to float32
+			kind = 0
+			size = uint64(data.Shape[0] * 4)
+		case 2:
+			// convert to float16
+			kind = 1
+			size = uint64(data.Shape[0] * data.Shape[1] * 2)
+		}
+
+		ggufName, err := m.GetLayerName(k)
+		if err != nil {
+			slog.Error("%v", err)
+			return nil, 0, err
+		}
+
+		shape := []uint64{0, 0, 0, 0}
+		for i := range data.Shape {
+			shape[i] = uint64(data.Shape[i])
+		}
+
+		t := llm.Tensor{
+			Name:   ggufName,
+			Kind:   kind,
+			Offset: offset,
+			Shape:  shape[:],
+		}
+
+		t.WriterTo = safetensorWriterTo{
+			t:        &t,
+			params:   params,
+			bo:       params.ByteOrder,
+			filename: fn,
+			start:    uint64(data.Offsets[0]),
+			end:      uint64(data.Offsets[1]),
+			padding:  8 + jsonSize,
+		}
+
+		offset += size
+		tensors = append(tensors, t)
+	}
+
+	slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
+	slog.Debug(fmt.Sprintf("offset = %d", offset))
+
+	return tensors, offset, nil
+}
+
+func (m *SafetensorFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+
+	var params Params
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	tMap := map[string]string{
+		"model.layers.(\\d+).input_layernorm.weight":                    "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":                      "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":                      "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":                        "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight":           "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":                   "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":                   "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":                   "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).block_sparse_moe.gate.weight":              "blk.$1.ffn_gate_inp.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
+		"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range tMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r safetensorWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	f, err := os.Open(r.filename)
+	if err != nil {
+		return 0, err
+	}
+	defer f.Close()
+
+	if _, err = f.Seek(int64(r.padding+r.start), 0); err != nil {
+		return 0, err
+	}
+
+	// use the handler if one is present
+	if r.handler != nil {
+		return 0, r.handler(w, r, f)
+	}
+
+	remaining := r.end - r.start
+
+	bufSize := uint64(10240)
+	var finished bool
+	for {
+		data := make([]byte, min(bufSize, remaining))
+
+		b, err := io.ReadFull(f, data)
+		remaining -= uint64(b)
+
+		if err == io.EOF || remaining <= 0 {
+			finished = true
+		} else if err != nil {
+			return 0, err
+		}
+
+		// convert bfloat16 -> ieee float32
+		tDataF32 := bfloat16.DecodeFloat32(data)
+
+		switch r.t.Kind {
+		case 0:
+			if err := binary.Write(w, r.bo, tDataF32); err != nil {
+				return 0, err
+			}
+		case 1:
+			// convert float32 -> float16
+			tempBuf := make([]uint16, len(data)/2)
+			for cnt, v := range tDataF32 {
+				tDataF16 := float16.Fromfloat32(v)
+				tempBuf[cnt] = uint16(tDataF16)
+			}
+			if err := binary.Write(w, r.bo, tempBuf); err != nil {
+				return 0, err
+			}
+		}
+		if finished {
+			break
+		}
+	}
+	return 0, nil
+}
+
+func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "MistralForCausalLM":
+			return &MistralModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "MixtralForCausalLM":
+			return &MixtralModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		case "GemmaForCausalLM":
+			return &GemmaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}
--- a/convert/torch.go
+++ b/convert/torch.go
@ -0,0 +1,286 @@
+package convert
+
+import (
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+
+	"github.com/nlpodyssey/gopickle/pytorch"
+	"github.com/nlpodyssey/gopickle/types"
+	"github.com/x448/float16"
+
+	"github.com/ollama/ollama/llm"
+)
+
+type torchWriterTo struct {
+	t *llm.Tensor
+
+	params *Params
+	bo     ByteOrder
+
+	storage pytorch.StorageInterface
+	handler func(w io.Writer, r torchWriterTo) error
+}
+
+type TorchFormat struct{}
+
+func (tf *TorchFormat) GetTensors(dirpath string, params *Params) ([]llm.Tensor, error) {
+	slog.Debug("getting torch tensors")
+
+	files, err := filepath.Glob(filepath.Join(dirpath, "pytorch_model-*.bin"))
+	if err != nil {
+		slog.Error("didn't find any torch files")
+		return nil, err
+	}
+
+	var offset uint64
+
+	var tensors []llm.Tensor
+	for _, fn := range files {
+		m, err := pytorch.Load(fn)
+		if err != nil {
+			slog.Error(fmt.Sprintf("error unpickling: %q", err))
+			return []llm.Tensor{}, err
+		}
+
+		for _, k := range m.(*types.Dict).Keys() {
+			if strings.HasSuffix(k.(string), "self_attn.rotary_emb.inv_freq") {
+				continue
+			}
+
+			t, _ := m.(*types.Dict).Get(k)
+			tshape := t.(*pytorch.Tensor).Size
+
+			var size uint64
+			var kind uint32
+			switch len(tshape) {
+			case 0:
+				continue
+			case 1:
+				// convert to float32
+				kind = 0
+				size = uint64(tshape[0] * 4)
+			case 2:
+				// convert to float16
+				kind = 1
+				size = uint64(tshape[0] * tshape[1] * 2)
+			}
+
+			ggufName, err := tf.GetLayerName(k.(string))
+			if err != nil {
+				slog.Error("%v", err)
+				return nil, err
+			}
+			slog.Debug(fmt.Sprintf("finding name for '%s' -> '%s'", k.(string), ggufName))
+
+			shape := []uint64{0, 0, 0, 0}
+			for i := range tshape {
+				shape[i] = uint64(tshape[i])
+			}
+
+			tensor := llm.Tensor{
+				Name:   ggufName,
+				Kind:   kind,
+				Offset: offset, // calculate the offset
+				Shape:  shape[:],
+			}
+
+			tensor.WriterTo = torchWriterTo{
+				t:       &tensor,
+				params:  params,
+				bo:      params.ByteOrder,
+				storage: t.(*pytorch.Tensor).Source,
+			}
+
+			tensors = append(tensors, tensor)
+			offset += size
+		}
+	}
+
+	return tensors, nil
+
+}
+
+func getAltParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "params.json"))
+	if err != nil {
+		slog.Error("no params.json")
+		return nil, err
+	}
+	defer f.Close()
+
+	type TorchParams struct {
+		HiddenSize     int     `json:"dim"`
+		AttentionHeads int     `json:"n_heads"`
+		KeyValHeads    int     `json:"n_kv_heads"`
+		HiddenLayers   int     `json:"n_layers"`
+		RopeTheta      int     `json:"rope_theta"`
+		NormEPS        float64 `json:"norm_eps"`
+	}
+
+	var tparams TorchParams
+
+	d := json.NewDecoder(f)
+	err = d.Decode(&tparams)
+	if err != nil {
+		return nil, err
+	}
+
+	params := &Params{
+		HiddenSize:     tparams.HiddenSize,
+		AttentionHeads: tparams.AttentionHeads,
+		KeyValHeads:    tparams.KeyValHeads,
+		HiddenLayers:   tparams.HiddenLayers,
+		NormEPS:        tparams.NormEPS,
+	}
+
+	switch {
+	case tparams.RopeTheta == 1000000:
+		// Codellama
+		params.ContextSize = 16384
+	case tparams.NormEPS == 1e-06:
+		// llama2
+		slog.Debug("Found llama2 - setting context size to 4096")
+		params.ContextSize = 4096
+	default:
+		params.ContextSize = 2048
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return params, nil
+}
+
+func (m *TorchFormat) GetParams(dirpath string) (*Params, error) {
+	f, err := os.Open(filepath.Join(dirpath, "config.json"))
+	if err != nil {
+		if os.IsNotExist(err) {
+			// try params.json instead
+			return getAltParams(dirpath)
+		} else {
+			return nil, err
+		}
+	}
+
+	var params Params
+	d := json.NewDecoder(f)
+	err = d.Decode(&params)
+	if err != nil {
+		return nil, err
+	}
+
+	params.ByteOrder = binary.LittleEndian
+	return &params, nil
+}
+
+func (m *TorchFormat) GetLayerName(n string) (string, error) {
+	directMap := map[string]string{
+		"tok_embeddings.weight":     "token_embd.weight",
+		"output.weight":             "output.weight",
+		"norm.weight":               "output_norm.weight",
+		"rope.freqs":                "rope_freqs.weight",
+		"model.embed_tokens.weight": "token_embd.weight",
+		"lm_head.weight":            "output.weight",
+		"model.norm.weight":         "output_norm.weight",
+	}
+
+	lMap := map[string]string{
+		"layers.(\\d+).attention_norm.weight":                 "blk.$1.attn_norm.weight",
+		"layers.(\\d+).attention_output_norm.weight":          "blk.$1.attn_norm.weight",
+		"layers.(\\d+).feed_forward.w2.weight":                "blk.$1.ffn_down.weight",
+		"layers.(\\d+).feed_forward.w1.weight":                "blk.$1.ffn_gate.weight",
+		"layers.(\\d+).feed_forward.w3.weight":                "blk.$1.ffn_up.weight",
+		"layers.(\\d+).ffn_norm.weight":                       "blk.$1.ffn_norm.weight",
+		"layers.(\\d+).attention.wk.weight":                   "blk.$1.attn_k.weight",
+		"layers.(\\d+).attention.wo.weight":                   "blk.$1.attn_output.weight",
+		"layers.(\\d+).attention.wq.weight":                   "blk.$1.attn_q.weight",
+		"layers.(\\d+).attention.wv.weight":                   "blk.$1.attn_v.weight",
+		"model.layers.(\\d+).input_layernorm.weight":          "blk.$1.attn_norm.weight",
+		"model.layers.(\\d+).mlp.down_proj.weight":            "blk.$1.ffn_down.weight",
+		"model.layers.(\\d+).mlp.gate_proj.weight":            "blk.$1.ffn_gate.weight",
+		"model.layers.(\\d+).mlp.up_proj.weight":              "blk.$1.ffn_up.weight",
+		"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
+		"model.layers.(\\d+).self_attn.k_proj.weight":         "blk.$1.attn_k.weight",
+		"model.layers.(\\d+).self_attn.o_proj.weight":         "blk.$1.attn_output.weight",
+		"model.layers.(\\d+).self_attn.q_proj.weight":         "blk.$1.attn_q.weight",
+		"model.layers.(\\d+).self_attn.v_proj.weight":         "blk.$1.attn_v.weight",
+	}
+
+	v, ok := directMap[n]
+	if ok {
+		return v, nil
+	}
+
+	// quick hack to rename the layers to gguf format
+	for k, v := range lMap {
+		re := regexp.MustCompile(k)
+		newName := re.ReplaceAllString(n, v)
+		if newName != n {
+			return newName, nil
+		}
+	}
+
+	return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
+}
+
+func (r torchWriterTo) WriteTo(w io.Writer) (n int64, err error) {
+	// use the handler if one is present
+	if r.handler != nil {
+		return 0, r.handler(w, r)
+	}
+
+	switch r.storage.(type) {
+	case *pytorch.FloatStorage:
+		slog.Warn(fmt.Sprintf("unexpected storage found for layer '%s'; skipping", r.t.Name))
+		return 0, nil
+	case *pytorch.HalfStorage:
+		switch r.t.Kind {
+		case 0:
+			data := r.storage.(*pytorch.HalfStorage).Data
+			slog.Debug(fmt.Sprintf("%35s F32 (%d)", r.t.Name, len(data)))
+			if err := binary.Write(w, r.bo, data); err != nil {
+				return 0, err
+			}
+		case 1:
+			data := r.storage.(*pytorch.HalfStorage).Data
+			tData := make([]uint16, len(data))
+			for cnt, v := range data {
+				tData[cnt] = uint16(float16.Fromfloat32(v))
+			}
+			slog.Debug(fmt.Sprintf("%35s F16 (%d)", r.t.Name, len(tData)))
+			if err := binary.Write(w, r.bo, tData); err != nil {
+				return 0, err
+			}
+		}
+	}
+
+	return 0, nil
+}
+
+func (m *TorchFormat) GetModelArch(name, dirPath string, params *Params) (ModelArch, error) {
+	switch len(params.Architectures) {
+	case 0:
+		return nil, fmt.Errorf("No architecture specified to convert")
+	case 1:
+		switch params.Architectures[0] {
+		case "LlamaForCausalLM":
+			return &LlamaModel{
+				ModelData{
+					Name:   name,
+					Path:   dirPath,
+					Params: params,
+					Format: m,
+				},
+			}, nil
+		default:
+			return nil, fmt.Errorf("Models based on '%s' are not yet supported", params.Architectures[0])
+		}
+	}
+
+	return nil, fmt.Errorf("Unknown error")
+}
--- a/docs/README.md
+++ b/docs/README.md
@ -3,7 +3,7 @@
 ### Getting Started
 * [Quickstart](../README.md#quickstart)
 * [Examples](../examples)
-* [Importing models](./import.md) from GGUF, Pytorch and Safetensors
+* [Importing models](./import.md)
 * [Linux Documentation](./linux.md)
 * [Windows Documentation](./windows.md)
 * [Docker Documentation](https://hub.docker.com/r/ollama/ollama)
--- a/docs/api.md
+++ b/docs/api.md
@ -90,7 +90,7 @@ The final response in the stream also includes additional data about the generat
 - `load_duration`: time spent in nanoseconds loading the model
 - `prompt_eval_count`: number of tokens in the prompt
 - `prompt_eval_duration`: time spent in nanoseconds evaluating the prompt
- `eval_count`: number of tokens the response
+- `eval_count`: number of tokens in the response
 - `eval_duration`: time in nanoseconds spent generating the response
 - `context`: an encoding of the conversation used in this response, this can be sent in the next request to keep a conversational memory
 - `response`: empty if the response was streamed, if not streamed, this will contain the full response
@ -394,7 +394,6 @@ Advanced parameters (optional):

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

--- a/docs/development.md
+++ b/docs/development.md
@ -51,7 +51,7 @@ Typically the build scripts will auto-detect CUDA, however, if your Linux distro
 or installation approach uses unusual paths, you can specify the location by
 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
-set set of target CUDA architectues by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")
+a set of target CUDA architectures by setting `CMAKE_CUDA_ARCHITECTURES` (e.g. "50;60;70")

 Then generate dependencies:

@ -69,7 +69,7 @@ go build .

 _Your operating system distribution may already have packages for AMD ROCm and CLBlast. Distro packages are often preferable, but instructions are distro-specific. Please consult distro-specific docs for dependencies if available!_

-Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html) development packages first, as well as `cmake` and `golang`.
+Install [CLBlast](https://github.com/CNugteren/CLBlast/blob/master/doc/installation.md) and [ROCm](https://rocm.docs.amd.com/en/latest/) development packages first, as well as `cmake` and `golang`.

 Typically the build scripts will auto-detect ROCm, however, if your Linux distro
 or installation approach uses unusual paths, you can specify the location by
@ -116,29 +116,30 @@ Note: The windows build for Ollama is still under development.

 Install required tools:

- MSVC toolchain - C/C++ and cmake as minimal requirements - You must build from a "Developer Shell" with the environment variables set
- go version 1.22 or higher
+- MSVC toolchain - C/C++ and cmake as minimal requirements
+- Go version 1.22 or higher
 - MinGW (pick one variant) with GCC.
-  - <https://www.mingw-w64.org/>
-  - <https://www.msys2.org/>
+  - [MinGW-w64](https://www.mingw-w64.org/)
+  - [MSYS2](https://www.msys2.org/)

 ```powershell
 $env:CGO_ENABLED="1"
-
 go generate ./...
-
 go build .
 ```

 #### Windows CUDA (NVIDIA)

-In addition to the common Windows development tools described above, install CUDA **AFTER** you install MSVC.
+In addition to the common Windows development tools described above, install CUDA after installing MSVC.

 - [NVIDIA CUDA](https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html)


 #### Windows ROCm (AMD Radeon)

-In addition to the common Windows development tools described above, install AMDs HIP package **AFTER** you install MSVC
+In addition to the common Windows development tools described above, install AMDs HIP package after installing MSVC.

- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+- [AMD HIP](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html)
+- [Strawberry Perl](https://strawberryperl.com/)
+
+Lastly, add `ninja.exe` included with MSVC to the system path (e.g. `C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja`).
--- a/docs/faq.md
+++ b/docs/faq.md
@ -14,6 +14,10 @@ curl -fsSL https://ollama.com/install.sh | sh

 Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.

+## Is my GPU compatible with Ollama?
+
+Please refer to the [GPU docs](./gpu.md).
+
 ## How can I specify the context window size?

 By default, Ollama uses a context window size of 2048 tokens.
@ -95,6 +99,37 @@ Ollama binds 127.0.0.1 port 11434 by default. Change the bind address with the `

 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.

+## How can I use Ollama with a proxy server?
+
+Ollama runs an HTTP server and can be exposed using a proxy server such as Nginx. To do so, configure the proxy to forward requests and optionally set required headers (if not exposing Ollama on the network). For example, with Nginx:
+
+```
+server {
+    listen 80;
+    server_name example.com;  # Replace with your domain or IP
+    location / {
+        proxy_pass http://localhost:11434;
+        proxy_set_header Host localhost:11434;
+    }
+}
+```
+
+## How can I use Ollama with ngrok?
+
+Ollama can be accessed using a range of tools for tunneling tools. For example with Ngrok:
+
+```
+ngrok http 11434 --host-header="localhost:11434"
+```
+
+## How can I use Ollama with Cloudflare Tunnel?
+
+To use Ollama with Cloudflare Tunnel, use the `--url` and `--http-host-header` flags:
+
+```
+cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11434"
+```
+
 ## How can I allow additional web origins to access Ollama?

 Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
@ -119,7 +154,7 @@ No. Ollama runs locally, and conversation data does not leave your machine.

 ## How can I use Ollama in Visual Studio Code?

-There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/jmorganca/ollama#extensions--plugins) at the bottom of the main repository readme.
+There is already a large collection of plugins available for VSCode as well as other editors that leverage Ollama. See the list of [extensions & plugins](https://github.com/ollama/ollama#extensions--plugins) at the bottom of the main repository readme.

 ## How do I use Ollama behind a proxy?

@ -193,3 +228,7 @@ To unload the model and free up memory use:
 ```shell
 curl http://localhost:11434/api/generate -d '{"model": "llama2", "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.
+
+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.
--- a/docs/gpu.md
+++ b/docs/gpu.md
@ -0,0 +1,102 @@
+# GPU
+## Nvidia
+Ollama supports Nvidia GPUs with compute capability 5.0+.
+
+Check your compute compatibility to see if your card is supported:
+[https://developer.nvidia.com/cuda-gpus](https://developer.nvidia.com/cuda-gpus)
+
+| Compute Capability | Family              | Cards                                                                                                       |
+| ------------------ | ------------------- | ----------------------------------------------------------------------------------------------------------- |
+| 9.0                | NVIDIA              | `H100`                                                                                                      |
+| 8.9                | GeForce RTX 40xx    | `RTX 4090` `RTX 4080` `RTX 4070 Ti` `RTX 4060 Ti`                                                           |
+|                    | NVIDIA Professional | `L4` `L40` `RTX 6000`                                                                                       |
+| 8.6                | GeForce RTX 30xx    | `RTX 3090 Ti` `RTX 3090` `RTX 3080 Ti` `RTX 3080` `RTX 3070 Ti` `RTX 3070` `RTX 3060 Ti` `RTX 3060`         |
+|                    | NVIDIA Professional | `A40` `RTX A6000` `RTX A5000` `RTX A4000` `RTX A3000` `RTX A2000` `A10` `A16` `A2`                          |
+| 8.0                | NVIDIA              | `A100` `A30`                                                                                                |
+| 7.5                | GeForce GTX/RTX     | `GTX 1650 Ti` `TITAN RTX` `RTX 2080 Ti` `RTX 2080` `RTX 2070` `RTX 2060`                                    |
+|                    | NVIDIA Professional | `T4` `RTX 5000` `RTX 4000` `RTX 3000` `T2000` `T1200` `T1000` `T600` `T500`                                 |
+|                    | Quadro              | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000`                                                                 |
+| 7.0                | NVIDIA              | `TITAN V` `V100` `Quadro GV100`                                                                             |
+| 6.1                | NVIDIA TITAN        | `TITAN Xp` `TITAN X`                                                                                        |
+|                    | GeForce GTX         | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050`                                     |
+|                    | Quadro              | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` |
+|                    | Tesla               | `P40` `P4`                                                                                                  |
+| 6.0                | NVIDIA              | `Tesla P100` `Quadro GP100`                                                                                 |
+| 5.2                | GeForce GTX         | `GTX TITAN X` `GTX 980 Ti` `GTX 980` `GTX 970` `GTX 960` `GTX 950`                                          |
+|                    | Quadro              | `M6000 24GB` `M6000` `M5000` `M5500M` `M4000` `M2200` `M2000` `M620`                                        |
+|                    | Tesla               | `M60` `M40`                                                                                                 |
+| 5.0                | GeForce GTX         | `GTX 750 Ti` `GTX 750` `NVS 810`                                                                            |
+|                    | Quadro              | `K2200` `K1200` `K620` `M1200` `M520` `M5000M` `M4000M` `M3000M` `M2000M` `M1000M` `K620M` `M600M` `M500M`  |
+
+
+### GPU Selection
+
+If you have multiple NVIDIA GPUs in your system and want to limit Ollama to use
+a subset, you can set `CUDA_VISIBLE_DEVICES` to a comma separated list of GPUs.
+Numeric IDs may be used, however ordering may vary, so UUIDs are more reliable.
+You can discover the UUID of your GPUs by running `nvidia-smi -L` If you want to
+ignore the GPUs and force CPU usage, use an invalid GPU ID (e.g., "-1")
+
+### Laptop Suspend Resume
+
+On linux, after a suspend/resume cycle, sometimes Ollama will fail to discover
+your NVIDIA GPU, and fallback to running on the CPU.  You can workaround this
+driver bug by reloading the NVIDIA UVM driver with `sudo rmmod nvidia_uvm &&
+sudo modprobe nvidia_uvm`
+
+## AMD Radeon
+Ollama supports the following AMD GPUs:
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
+| AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
+
+### Overrides
+Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
+some cases you can force the system to try to use a similar LLVM target that is
+close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
+however, ROCm does not currently support this target. The closest support is
+`gfx1030`.  You can use the environment variable `HSA_OVERRIDE_GFX_VERSION` with
+`x.y.z` syntax.  So for example, to force the system to run on the RX 5400, you
+would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
+server.  If you have an unsupported AMD GPU you can experiment using the list of
+supported types below.
+
+At this time, the known supported GPU types are the following LLVM Targets.
+This table shows some example GPUs that map to these LLVM targets:
+| **LLVM Target** | **An Example GPU** |
+|-----------------|---------------------|
+| gfx900 | Radeon RX Vega 56 |
+| gfx906 | Radeon Instinct MI50 |
+| gfx908 | Radeon Instinct MI100 |
+| gfx90a | Radeon Instinct MI210 |
+| gfx940 | Radeon Instinct MI300 |
+| gfx941 | |
+| gfx942 | |
+| gfx1030 | Radeon PRO V620 |
+| gfx1100 | Radeon PRO W7900 |
+| gfx1101 | Radeon PRO W7700 |
+| gfx1102 | Radeon RX 7600 |
+
+AMD is working on enhancing ROCm v6 to broaden support for families of GPUs in a
+future release which should increase support for more GPUs.
+
+Reach out on [Discord](https://discord.gg/ollama) or file an
+[issue](https://github.com/ollama/ollama/issues) for additional help.
+
+### GPU Selection
+
+If you have multiple AMD GPUs in your system and want to limit Ollama to use a
+subset, you can set `HIP_VISIBLE_DEVICES` to a comma separated list of GPUs.
+You can see the list of devices with `rocminfo`.  If you want to ignore the GPUs
+and force CPU usage, use an invalid GPU ID (e.g., "-1")
+
+### Container Permission
+
+In some Linux distributions, SELinux can prevent containers from
+accessing the AMD GPU devices.  On the host system you can run 
+`sudo setsebool container_use_devices=1` to allow containers to use devices.
+
+### Metal (Apple GPUs)
+Ollama supports GPU acceleration on Apple devices via the Metal API.
--- a/docs/modelfile.md
+++ b/docs/modelfile.md
@ -113,7 +113,7 @@ FROM llama2
 ```

 A list of available base models:
-<https://github.com/jmorganca/ollama#model-library>
+<https://github.com/ollama/ollama#model-library>

 #### Build from a `bin` file

@ -131,7 +131,7 @@ The `PARAMETER` instruction defines a parameter that can be set when the model i
 PARAMETER <parameter> <parametervalue>
 ```

-### Valid Parameters and Values
+#### Valid Parameters and Values

 | Parameter      | Description                                                                                                                                                                                                                                             | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
@ -139,9 +139,6 @@ PARAMETER <parameter> <parametervalue>
 | mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                        | float      | mirostat_eta 0.1     |
 | mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                         | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                    | int        | num_ctx 4096         |
-| num_gqa        | The number of GQA groups in the transformer layer. Required for some models, for example it is 8 for llama2:70b                                                                                                                                         | int        | num_gqa 1            |
-| num_gpu        | The number of layers to send to the GPU(s). On macOS it defaults to 1 to enable metal support, 0 to disable.                                                                                                                                            | int        | num_gpu 50           |
-| num_thread     | Sets the number of threads to use during computation. By default, Ollama will detect this for optimal performance. It is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). | int        | num_thread 8         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                           | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                     | float      | repeat_penalty 1.1   |
 | temperature    | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8)                                                                                                                                     | float      | temperature 0.7      |
@ -183,7 +180,7 @@ SYSTEM """<system message>"""

 ### ADAPTER

-The `ADAPTER` instruction specifies the LoRA adapter to apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.
+The `ADAPTER` instruction is an optional instruction that specifies any LoRA adapter that should apply to the base model. The value of this instruction should be an absolute path or a path relative to the Modelfile and the file must be in a GGML file format. The adapter should be tuned from the base model otherwise the behaviour is undefined.

 ```modelfile
 ADAPTER ./ollama-lora.bin
@ -201,7 +198,22 @@ LICENSE """

 ### MESSAGE

-The `MESSAGE` instruction allows you to specify a message history for the model to use when responding:
+The `MESSAGE` instruction allows you to specify a message history for the model to use when responding. Use multiple iterations of the MESSAGE command to build up a conversation which will guide the model to answer in a similar way.
+
+```modelfile
+MESSAGE <role> <message>
+```
+
+#### Valid roles
+
+| Role      | Description                                                  |
+| --------- | ------------------------------------------------------------ |
+| system    | Alternate way of providing the SYSTEM message for the model. |
+| user      | An example message of what the user could have asked.        |
+| assistant | An example message of how the model should respond.          |
+
+
+#### Example conversation

 ```modelfile
 MESSAGE user Is Toronto in Canada?
@ -212,6 +224,7 @@ MESSAGE user Is Ontario in Canada?
 MESSAGE assistant yes
 ```

+
 ## Notes

 - the **`Modelfile` is not case sensitive**. In the examples, uppercase instructions are used to make it easier to distinguish it from arguments.
--- a/docs/openai.md
+++ b/docs/openai.md
@ -1,6 +1,6 @@
 # OpenAI compatibility

-> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/jmorganca/ollama/blob/main/docs/api.md).
+> **Note:** OpenAI compatibility is experimental and is subject to major adjustments including breaking changes. For fully-featured access to the Ollama API, see the Ollama [Python library](https://github.com/ollama/ollama-python), [JavaScript library](https://github.com/ollama/ollama-js) and [REST API](https://github.com/ollama/ollama/blob/main/docs/api.md).

 Ollama provides experimental compatibility with parts of the [OpenAI API](https://platform.openai.com/docs/api-reference) to help connect existing applications to Ollama.

--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@ -67,45 +67,19 @@ You can see what features your CPU has with the following.
 cat /proc/cpuinfo| grep flags  | head -1
 ```

-## AMD Radeon GPU Support
+## Installing older or pre-release versions on Linux

-Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
-some cases you can force the system to try to use a similar LLVM target that is
-close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
-however, ROCm does not currently support this target. The closest support is
-`gfx1030`.  You can use the environment variable `HSA_OVERRIDE_GFX_VERSION` with
-`x.y.z` syntax.  So for example, to force the system to run on the RX 5400, you
-would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
-server.  If you have an unsupported AMD GPU you can experiment using the list of
-supported types below.
-
-At this time, the known supported GPU types are the following LLVM Targets.
-This table shows some example GPUs that map to these LLVM targets:
-| **LLVM Target** | **An Example GPU** |
-|-----------------|---------------------|
-| gfx900 | Radeon RX Vega 56 |
-| gfx906 | Radeon Instinct MI50 |
-| gfx908 | Radeon Instinct MI100 |
-| gfx90a | Radeon Instinct MI210 |
-| gfx940 | Radeon Instinct MI300 |
-| gfx941 | |
-| gfx942 | |
-| gfx1030 | Radeon PRO V620 |
-| gfx1100 | Radeon PRO W7900 |
-| gfx1101 | Radeon PRO W7700 |
-| gfx1102 | Radeon RX 7600 |
-
-AMD is working on enhancing ROCm v6 to broaden support for families of GPUs in a
-future release which should increase support for more GPUs.
-
-Reach out on [Discord](https://discord.gg/ollama) or file an
-[issue](https://github.com/ollama/ollama/issues) for additional help.
-
-## Installing older versions on Linux
-
-If you run into problems on Linux and want to install an older version you can tell the install script
-which version to install.
+If you run into problems on Linux and want to install an older version, or you'd
+like to try out a pre-release before it's officially released, you can tell the
+install script which version to install.

 ```sh
-curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.27" sh
+curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
 ```
+
+## Linux tmp noexec 
+
+If your system is configured with the "noexec" flag where Ollama stores its
+temporary executable files, you can specify an alternate location by setting
+OLLAMA_TMPDIR to a location writable by the user ollama runs as.  For example
+OLLAMA_TMPDIR=/usr/share/ollama/
--- a/docs/tutorials/langchainjs.md
+++ b/docs/tutorials/langchainjs.md
@ -18,7 +18,7 @@ const ollama = new Ollama({
  model: "llama2",
 });

-const answer = await ollama.call(`why is the sky blue?`);
+const answer = await ollama.invoke(`why is the sky blue?`);

 console.log(answer);
 ```
--- a/docs/tutorials/langchainpy.md
+++ b/docs/tutorials/langchainpy.md
@ -17,10 +17,12 @@ Let's start by asking a simple question that we can get an answer to from the **
 Then we can create a model and ask the question:

 ```python
-from langchain.llms import Ollama
-ollama = Ollama(base_url='http://localhost:11434',
-model="llama2")
-print(ollama("why is the sky blue"))
+from langchain_community.llms import Ollama
+ollama = Ollama(
+    base_url='http://localhost:11434',
+    model="llama3"
+)
+print(ollama.invoke("why is the sky blue"))
 ```

 Notice that we are defining the model and the base URL for Ollama.
--- a/docs/tutorials/nvidia-jetson.md
+++ b/docs/tutorials/nvidia-jetson.md
@ -1,38 +1,15 @@
 # Running Ollama on NVIDIA Jetson Devices

-With some minor configuration, Ollama runs well on [NVIDIA Jetson Devices](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/). The following has been tested on [JetPack 5.1.2](https://developer.nvidia.com/embedded/jetpack).
+Ollama runs well on [NVIDIA Jetson Devices](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/) and should run out of the box with the standard installation instructions. 

-NVIDIA Jetson devices are Linux-based embedded AI computers that are purpose-built for AI applications.
-
-Jetsons have an integrated GPU that is wired directly to the memory controller of the machine. For this reason, the `nvidia-smi` command is unrecognized, and Ollama proceeds to operate in "CPU only"
-mode. This can be verified by using a monitoring tool like jtop.
-
-In order to address this, we simply pass the path to the Jetson's pre-installed CUDA libraries into `ollama serve` (while in a tmux session). We then hardcode the num_gpu parameters into a cloned
-version of our target model.
-
-Prerequisites:
-
- curl
- tmux
-
-Here are the steps:
+The following has been tested on [JetPack 5.1.2](https://developer.nvidia.com/embedded/jetpack), but should also work on JetPack 6.0.

 - Install Ollama via standard Linux command (ignore the 404 error): `curl https://ollama.com/install.sh | sh`
- Stop the Ollama service: `sudo systemctl stop ollama`
- Start Ollama serve in a tmux session called ollama_jetson and reference the CUDA libraries path: `tmux has-session -t ollama_jetson 2>/dev/null || tmux new-session -d -s ollama_jetson 
-'LD_LIBRARY_PATH=/usr/local/cuda/lib64 ollama serve'`
 - Pull the model you want to use (e.g. mistral): `ollama pull mistral`
- Create a new Modelfile specifically for enabling GPU support on the Jetson: `touch ModelfileMistralJetson`
- In the ModelfileMistralJetson file, specify the FROM model and the num_gpu PARAMETER as shown below:
-
-```
-FROM mistral
-PARAMETER num_gpu 999
-```
-
- Create a new model from your Modelfile: `ollama create mistral-jetson -f ./ModelfileMistralJetson`
- Run the new model: `ollama run mistral-jetson`
-
-If you run a monitoring tool like jtop you should now see that Ollama is using the Jetson's integrated GPU.
+- Start an interactive session: `ollama run mistral`

 And that's it!
+
+# Running Ollama in Docker
+
+When running GPU accelerated applications in Docker, it is highly recommended to use [dusty-nv jetson-containers repo](https://github.com/dusty-nv/jetson-containers).
--- a/docs/windows.md
+++ b/docs/windows.md
@ -1,47 +1,47 @@
-# Ollama Windows Preview
-
-Welcome to the Ollama Windows preview.
-
-No more WSL required!
-
-Ollama now runs as a native Windows application, including NVIDIA and AMD Radeon GPU support.
-After installing Ollama Windows Preview, Ollama will run in the background and
-the `ollama` command line is available in `cmd`, `powershell` or your favorite
-terminal application. As usual the Ollama [api](./api.md) will be served on
-`http://localhost:11434`.
-
-As this is a preview release, you should expect a few bugs here and there.  If
-you run into a problem you can reach out on
-[Discord](https://discord.gg/ollama), or file an 
-[issue](https://github.com/ollama/ollama/issues).
-Logs will often be helpful in dianosing the problem (see
-[Troubleshooting](#troubleshooting) below)
-
-## System Requirements
-
-* Windows 10 or newer, Home or Pro
-* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
-* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
-
-## API Access
-
-Here's a quick example showing API access from `powershell`
-```powershell
-(Invoke-WebRequest -method POST -Body '{"model":"llama2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
-```
-
-## Troubleshooting
-
-While we're in preview, `OLLAMA_DEBUG` is always enabled, which adds
-a "view logs" menu item to the app, and increses logging for the GUI app and
-server.
-
-Ollama on Windows stores files in a few different locations.  You can view them in
-the explorer window by hitting `<cmd>+R` and type in:
- `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
-    - *app.log* contains logs from the GUI application
-    - *server.log* contains the server logs
-    - *upgrade.log* contains log output for upgrades
- `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
- `explorer %HOMEPATH%\.ollama` contains models and configuration
- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
+# Ollama Windows Preview
+
+Welcome to the Ollama Windows preview.
+
+No more WSL required!
+
+Ollama now runs as a native Windows application, including NVIDIA and AMD Radeon GPU support.
+After installing Ollama Windows Preview, Ollama will run in the background and
+the `ollama` command line is available in `cmd`, `powershell` or your favorite
+terminal application. As usual the Ollama [api](./api.md) will be served on
+`http://localhost:11434`.
+
+As this is a preview release, you should expect a few bugs here and there.  If
+you run into a problem you can reach out on
+[Discord](https://discord.gg/ollama), or file an
+[issue](https://github.com/ollama/ollama/issues).
+Logs will often be helpful in diagnosing the problem (see
+[Troubleshooting](#troubleshooting) below)
+
+## System Requirements
+
+* Windows 10 or newer, Home or Pro
+* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
+* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card
+
+## API Access
+
+Here's a quick example showing API access from `powershell`
+```powershell
+(Invoke-WebRequest -method POST -Body '{"model":"llama2", "prompt":"Why is the sky blue?", "stream": false}' -uri http://localhost:11434/api/generate ).Content | ConvertFrom-json
+```
+
+## Troubleshooting
+
+While we're in preview, `OLLAMA_DEBUG` is always enabled, which adds
+a "view logs" menu item to the app, and increses logging for the GUI app and
+server.
+
+Ollama on Windows stores files in a few different locations.  You can view them in
+the explorer window by hitting `<cmd>+R` and type in:
+- `explorer %LOCALAPPDATA%\Ollama` contains logs, and downloaded updates
+    - *app.log* contains logs from the GUI application
+    - *server.log* contains the server logs
+    - *upgrade.log* contains log output for upgrades
+- `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
+- `explorer %HOMEPATH%\.ollama` contains models and configuration
+- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
--- a/examples/go-chat/main.go
+++ b/examples/go-chat/main.go
@ -0,0 +1,51 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	messages := []api.Message{
+		api.Message{
+			Role:    "system",
+			Content: "Provide very brief, concise responses",
+		},
+		api.Message{
+			Role:    "user",
+			Content: "Name some unusual animals",
+		},
+		api.Message{
+			Role:    "assistant",
+			Content: "Monotreme, platypus, echidna",
+		},
+		api.Message{
+			Role:    "user",
+			Content: "which of these is the most dangerous?",
+		},
+	}
+
+	ctx := context.Background()
+	req := &api.ChatRequest{
+		Model:    "llama2",
+		Messages: messages,
+	}
+
+	respFunc := func(resp api.ChatResponse) error {
+		fmt.Print(resp.Message.Content)
+		return nil
+	}
+
+	err = client.Chat(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}
--- a/examples/go-generate-streaming/main.go
+++ b/examples/go-generate-streaming/main.go
@ -0,0 +1,40 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	// By default, GenerateRequest is streaming.
+	req := &api.GenerateRequest{
+		Model:  "gemma",
+		Prompt: "how many planets are there?",
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// Only print the response here; GenerateResponse has a number of other
+		// interesting fields you want to examine.
+
+		// In streaming mode, responses are partial so we call fmt.Print (and not
+		// Println) in order to avoid spurious newlines being introduced. The
+		// model will insert its own newlines if it wants.
+		fmt.Print(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println()
+}
--- a/examples/go-generate/main.go
+++ b/examples/go-generate/main.go
@ -0,0 +1,37 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	req := &api.GenerateRequest{
+		Model:  "gemma",
+		Prompt: "how many planets are there?",
+
+		// set streaming to false
+		Stream: new(bool),
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// Only print the response here; GenerateResponse has a number of other
+		// interesting fields you want to examine.
+		fmt.Println(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}
--- a/examples/golang-simplegenerate/README.md
+++ b/examples/golang-simplegenerate/README.md
--- a/examples/golang-simplegenerate/main.go
+++ b/examples/golang-simplegenerate/main.go
--- a/examples/go-multimodal/main.go
+++ b/examples/go-multimodal/main.go
@ -0,0 +1,47 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	if len(os.Args) <= 1 {
+		log.Fatal("usage: <image name>")
+	}
+
+	imgData, err := os.ReadFile(os.Args[1])
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	req := &api.GenerateRequest{
+		Model:  "llava",
+		Prompt: "describe this image",
+		Images: []api.ImageData{imgData},
+	}
+
+	ctx := context.Background()
+	respFunc := func(resp api.GenerateResponse) error {
+		// In streaming mode, responses are partial so we call fmt.Print (and not
+		// Println) in order to avoid spurious newlines being introduced. The
+		// model will insert its own newlines if it wants.
+		fmt.Print(resp.Response)
+		return nil
+	}
+
+	err = client.Generate(ctx, req, respFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println()
+}
--- a/examples/go-pull-progress/main.go
+++ b/examples/go-pull-progress/main.go
@ -0,0 +1,31 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+)
+
+func main() {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	ctx := context.Background()
+
+	req := &api.PullRequest{
+		Model: "mistral",
+	}
+	progressFunc := func(resp api.ProgressResponse) error {
+		fmt.Printf("Progress: status=%v, total=%v, completed=%v\n", resp.Status, resp.Total, resp.Completed)
+		return nil
+	}
+
+	err = client.Pull(ctx, req, progressFunc)
+	if err != nil {
+		log.Fatal(err)
+	}
+}
--- a/examples/langchain-python-rag-privategpt/README.md
+++ b/examples/langchain-python-rag-privategpt/README.md
@ -1,6 +1,6 @@
 # PrivateGPT with Llama 2 uncensored

-https://github.com/jmorganca/ollama/assets/3325447/20cf8ec6-ff25-42c6-bdd8-9be594e3ce1b
+https://github.com/ollama/ollama/assets/3325447/20cf8ec6-ff25-42c6-bdd8-9be594e3ce1b

 > Note: this example is a slightly modified version of PrivateGPT using models such as Llama 2 Uncensored. All credit for PrivateGPT goes to Iván Martínez who is the creator of it, and you can find his GitHub repo [here](https://github.com/imartinez/privateGPT).

--- a/examples/modelfile-mario/readme.md
+++ b/examples/modelfile-mario/readme.md
@ -28,7 +28,7 @@ You are Mario from Super Mario Bros, acting as an assistant.
 What if you want to change its behaviour?

 - Try changing the prompt
- Try changing the parameters [Docs](https://github.com/jmorganca/ollama/blob/main/docs/modelfile.md)
+- Try changing the parameters [Docs](https://github.com/ollama/ollama/blob/main/docs/modelfile.md)
 - Try changing the model (e.g. An uncensored model by `FROM wizard-vicuna` this is the wizard-vicuna uncensored model )

 Once the changes are made,
--- a/examples/python-json-datagenerator/readme.md
+++ b/examples/python-json-datagenerator/readme.md
@ -1,6 +1,6 @@
 # JSON Output Example

-![llmjson 2023-11-10 15_31_31](https://github.com/jmorganca/ollama/assets/633681/e599d986-9b4a-4118-81a4-4cfe7e22da25)
+![llmjson 2023-11-10 15_31_31](https://github.com/ollama/ollama/assets/633681/e599d986-9b4a-4118-81a4-4cfe7e22da25)

 There are two python scripts in this example. `randomaddresses.py` generates random addresses from different countries. `predefinedschema.py` sets a template for the model to fill in.

--- a/examples/python-loganalysis/readme.md
+++ b/examples/python-loganalysis/readme.md
@ -1,6 +1,6 @@
 # Log Analysis example

-![loganalyzer 2023-11-10 08_53_29](https://github.com/jmorganca/ollama/assets/633681/ad30f1fc-321f-4953-8914-e30e24db9921)
+![loganalyzer 2023-11-10 08_53_29](https://github.com/ollama/ollama/assets/633681/ad30f1fc-321f-4953-8914-e30e24db9921)

 This example shows one possible way to create a log file analyzer. It uses the model **mattw/loganalyzer** which is based on **codebooga**, a 34b parameter model.

--- a/examples/typescript-functioncalling/readme.md
+++ b/examples/typescript-functioncalling/readme.md
@ -1,6 +1,6 @@
 # Function calling

-![function calling 2023-11-16 16_12_58](https://github.com/jmorganca/ollama/assets/633681/a0acc247-9746-45ab-b325-b65dfbbee4fb)
+![function calling 2023-11-16 16_12_58](https://github.com/ollama/ollama/assets/633681/a0acc247-9746-45ab-b325-b65dfbbee4fb)

 One of the features added to some models is 'function calling'. It's a bit of a confusing name. It's understandable if you think that means the model can call functions, but that's not what it means. Function calling simply means that the output of the model is formatted in JSON, using a preconfigured schema, and uses the expected types. Then your code can use the output of the model and call functions with it. Using the JSON format in Ollama, you can use any model for function calling. 

--- a/format/bytes.go
+++ b/format/bytes.go
@ -6,11 +6,16 @@ import (
 )

 const (
-	Byte     = 1
+	Byte = 1
+
 	KiloByte = Byte * 1000
 	MegaByte = KiloByte * 1000
 	GigaByte = MegaByte * 1000
 	TeraByte = GigaByte * 1000
+
+	KibiByte = Byte * 1024
+	MebiByte = KibiByte * 1024
+	GibiByte = MebiByte * 1024
 )

 func HumanBytes(b int64) string {
@ -45,3 +50,14 @@ func HumanBytes(b int64) string {
 		return fmt.Sprintf("%d %s", int(value), unit)
 	}
 }
+
+func HumanBytes2(b uint64) string {
+	switch {
+	case b >= MebiByte:
+		return fmt.Sprintf("%.1f MiB", float64(b)/MebiByte)
+	case b >= KibiByte:
+		return fmt.Sprintf("%.1f KiB", float64(b)/KibiByte)
+	default:
+		return fmt.Sprintf("%d B", b)
+	}
+}
--- a/go.mod
+++ b/go.mod
@ -1,4 +1,4 @@
-module github.com/jmorganca/ollama
+module github.com/ollama/ollama

 go 1.22

@ -9,7 +9,7 @@ require (
 	github.com/d4l3k/go-bfloat16 v0.0.0-20211005043715-690c3bdd05f1
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.9.1
-	github.com/golang/protobuf v1.5.0
+	github.com/golang/protobuf v1.5.0 // indirect
 	github.com/google/uuid v1.0.0
 	github.com/mitchellh/mapstructure v1.5.0
 	github.com/olekukonko/tablewriter v0.0.5
@ -19,7 +19,10 @@ require (
 	golang.org/x/sync v0.3.0
 )

-require github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9
+require (
+	github.com/nlpodyssey/gopickle v0.3.0
+	github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9
+)

 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc // indirect
@ -68,7 +71,7 @@ require (
 	golang.org/x/net v0.17.0 // indirect
 	golang.org/x/sys v0.13.0
 	golang.org/x/term v0.13.0
-	golang.org/x/text v0.13.0 // indirect
+	golang.org/x/text v0.14.0 // indirect
 	google.golang.org/protobuf v1.30.0
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )
--- a/go.sum
+++ b/go.sum
@ -122,6 +122,8 @@ github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/reflect2 v1.0.2 h1:xBagoLtFs94CBntxluKeaWgTMpvLxC4ur3nMaC9Gz0M=
 github.com/modern-go/reflect2 v1.0.2/go.mod h1:yWuevngMOJpCy52FWWMvUC8ws7m/LJsjYzDa0/r8luk=
+github.com/nlpodyssey/gopickle v0.3.0 h1:BLUE5gxFLyyNOPzlXxt6GoHEMMxD0qhsE4p0CIQyoLw=
+github.com/nlpodyssey/gopickle v0.3.0/go.mod h1:f070HJ/yR+eLi5WmM1OXJEGaTpuJEUiib19olXgYha0=
 github.com/olekukonko/tablewriter v0.0.5 h1:P2Ga83D34wi1o9J6Wh1mRuqd4mF/x/lgBS7N7AbDhec=
 github.com/olekukonko/tablewriter v0.0.5/go.mod h1:hPp6KlRPjbx+hW8ykQs1w3UBbZlj6HuIJcUGPhkA7kY=
 github.com/pdevine/tensor v0.0.0-20240228013915-64ccaa8d9ca9 h1:DV4iXjNn6fGeDl1AkZ1I0QB/0DBjrc7kPpxHrmuDzW4=
@ -236,8 +238,8 @@ golang.org/x/term v0.13.0/go.mod h1:LTmsnFJwVN6bCy1rVCoS+qHT1HhALEFxKncY3WNNh4U=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.13.0 h1:ablQoSUd0tRdKxZewP80B+BaqeKJuVhuRxj/dkrun3k=
-golang.org/x/text v0.13.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=
+golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
+golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
--- a/gpu/amd_common.go
+++ b/gpu/amd_common.go
@ -7,7 +7,7 @@ import (
 	"log/slog"
 	"os"
 	"path/filepath"
-	"strconv"
+	"runtime"
 	"strings"
 )

@ -35,24 +35,64 @@ func GetSupportedGFX(libDir string) ([]string, error) {
 	return ret, nil
 }

-func amdSetVisibleDevices(ids []int, skip map[int]interface{}) {
-	// Set the visible devices if not already set
-	// TODO - does sort order matter?
-	devices := []string{}
-	for i := range ids {
-		slog.Debug(fmt.Sprintf("i=%d", i))
-		if _, skipped := skip[i]; skipped {
-			slog.Debug("skipped")
+func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "rocm" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
 			continue
 		}
-		devices = append(devices, strconv.Itoa(i))
+		ids = append(ids, info.ID)
 	}
-	slog.Debug(fmt.Sprintf("devices=%v", devices))
-
-	val := strings.Join(devices, ",")
-	err := os.Setenv("HIP_VISIBLE_DEVICES", val)
-	if err != nil {
-		slog.Warn(fmt.Sprintf("failed to set env: %s", err))
-	}
-	slog.Debug("HIP_VISIBLE_DEVICES=" + val)
+	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
+}
+
+func commonAMDValidateLibDir() (string, error) {
+	// We try to favor system paths first, so that we can wire up the subprocess to use
+	// the system version.  Only use our bundled version if the system version doesn't work
+	// This gives users a more recovery options if versions have subtle problems at runtime
+
+	// Prefer explicit HIP env var
+	hipPath := os.Getenv("HIP_PATH")
+	if hipPath != "" {
+		hipLibDir := filepath.Join(hipPath, "bin")
+		if rocmLibUsable(hipLibDir) {
+			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
+			return hipLibDir, nil
+		}
+	}
+
+	// Scan the LD_LIBRARY_PATH or PATH
+	pathEnv := "LD_LIBRARY_PATH"
+	if runtime.GOOS == "windows" {
+		pathEnv = "PATH"
+	}
+
+	paths := os.Getenv(pathEnv)
+	for _, path := range filepath.SplitList(paths) {
+		d, err := filepath.Abs(path)
+		if err != nil {
+			continue
+		}
+		if rocmLibUsable(d) {
+			return d, nil
+		}
+	}
+
+	// Well known location(s)
+	if rocmLibUsable(RocmStandardLocation) {
+		return RocmStandardLocation, nil
+	}
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
+	}
+	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }
--- a/gpu/amd_hip_windows.go
+++ b/gpu/amd_hip_windows.go
@ -69,7 +69,7 @@ func NewHipLib() (*HipLib, error) {
 func (hl *HipLib) Release() {
 	err := windows.FreeLibrary(hl.dll)
 	if err != nil {
-		slog.Warn(fmt.Sprintf("failed to unload amdhip64.dll: %s", err))
+		slog.Warn("failed to unload amdhip64.dll", "error", err)
 	}
 	hl.dll = 0
 }
@ -98,7 +98,7 @@ func (hl *HipLib) HipGetDeviceCount() int {
 		return 0
 	}
 	if status != hipSuccess {
-		slog.Warn(fmt.Sprintf("failed call to hipGetDeviceCount: %d %s", status, err))
+		slog.Warn("failed call to hipGetDeviceCount", "status", status, "error", err)
 	}
 	return count
 }
--- a/gpu/amd_linux.go
+++ b/gpu/amd_linux.go
@ -11,6 +11,8 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+
+	"github.com/ollama/ollama/format"
 )

 // Discovery logic for AMD/ROCm GPUs
@ -32,14 +34,11 @@ var (
 )

 // Gather GPU information from the amdgpu driver if any supported GPUs are detected
-// HIP_VISIBLE_DEVICES will be set if we detect a mix of unsupported and supported devices
-// and the user hasn't already set this variable
-func AMDGetGPUInfo(resp *GpuInfo) {
-	// TODO - DRY this out with windows
+func AMDGetGPUInfo() []GpuInfo {
+	resp := []GpuInfo{}
 	if !AMDDetected() {
-		return
+		return resp
 	}
-	skip := map[int]interface{}{}

 	// Opportunistic logging of driver version to aid in troubleshooting
 	ver, err := AMDDriverVersion()
@ -47,142 +46,117 @@ func AMDGetGPUInfo(resp *GpuInfo) {
 		slog.Info("AMD Driver: " + ver)
 	} else {
 		// TODO - if we see users crash and burn with the upstreamed kernel this can be adjusted to hard-fail rocm support and fallback to CPU
-		slog.Warn(fmt.Sprintf("ollama recommends running the https://www.amd.com/en/support/linux-drivers: %s", err))
+		slog.Warn("ollama recommends running the https://www.amd.com/en/support/linux-drivers", "error", err)
 	}

-	// If the user has specified exactly which GPUs to use, look up their memory
-	visibleDevices := os.Getenv("HIP_VISIBLE_DEVICES")
-	if visibleDevices != "" {
-		ids := []int{}
-		for _, idStr := range strings.Split(visibleDevices, ",") {
-			id, err := strconv.Atoi(idStr)
-			if err != nil {
-				slog.Warn(fmt.Sprintf("malformed HIP_VISIBLE_DEVICES=%s %s", visibleDevices, err))
-			} else {
-				ids = append(ids, id)
-			}
-		}
-		amdProcMemLookup(resp, nil, ids)
-		return
-	}
-
-	// Gather GFX version information from all detected cards
-	gfx := AMDGFXVersions()
-	verStrings := []string{}
-	for i, v := range gfx {
-		verStrings = append(verStrings, v.ToGFXString())
-		if v.Major == 0 {
-			// Silently skip CPUs
-			skip[i] = struct{}{}
-			continue
-		}
-		if v.Major < 9 {
-			// TODO consider this a build-time setting if we can support 8xx family GPUs
-			slog.Warn(fmt.Sprintf("amdgpu [%d] too old %s", i, v.ToGFXString()))
-			skip[i] = struct{}{}
-		}
-	}
-	slog.Info(fmt.Sprintf("detected amdgpu versions %v", verStrings))
-
-	// Abort if all GPUs are skipped
-	if len(skip) >= len(gfx) {
-		slog.Info("all detected amdgpus are skipped, falling back to CPU")
-		return
-	}
-
-	// If we got this far, then we have at least 1 GPU that's a ROCm candidate, so make sure we have a lib
-	libDir, err := AMDValidateLibDir()
-	if err != nil {
-		slog.Warn(fmt.Sprintf("unable to verify rocm library, will use cpu: %s", err))
-		return
+	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
+	var visibleDevices []string
+	hipVD := os.Getenv("HIP_VISIBLE_DEVICES")   // zero based index only
+	rocrVD := os.Getenv("ROCR_VISIBLE_DEVICES") // zero based index or UUID, but consumer cards seem to not support UUID
+	gpuDO := os.Getenv("GPU_DEVICE_ORDINAL")    // zero based index
+	switch {
+	// TODO is this priorty order right?
+	case hipVD != "":
+		visibleDevices = strings.Split(hipVD, ",")
+	case rocrVD != "":
+		visibleDevices = strings.Split(rocrVD, ",")
+		// TODO - since we don't yet support UUIDs, consider detecting and reporting here
+		// all our test systems show GPU-XX indicating UUID is not supported
+	case gpuDO != "":
+		visibleDevices = strings.Split(gpuDO, ",")
 	}

 	gfxOverride := os.Getenv("HSA_OVERRIDE_GFX_VERSION")
-	if gfxOverride == "" {
-		supported, err := GetSupportedGFX(libDir)
+	var supported []string
+	libDir := ""
+
+	// The amdgpu driver always exposes the host CPU(s) first, but we have to skip them and subtract
+	// from the other IDs to get alignment with the HIP libraries expectations (zero is the first GPU, not the CPU)
+	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
+	cpuCount := 0
+	for _, match := range matches {
+		slog.Debug("evaluating amdgpu node " + match)
+		fp, err := os.Open(match)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("failed to lookup supported GFX types, falling back to CPU mode: %s", err))
-			return
-		}
-		slog.Debug(fmt.Sprintf("rocm supported GPU types %v", supported))
-
-		for i, v := range gfx {
-			if !slices.Contains[[]string, string](supported, v.ToGFXString()) {
-				slog.Warn(fmt.Sprintf("amdgpu [%d] %s is not supported by %s %v", i, v.ToGFXString(), libDir, supported))
-				// TODO - consider discrete markdown just for ROCM troubleshooting?
-				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")
-				skip[i] = struct{}{}
-			} else {
-				slog.Info(fmt.Sprintf("amdgpu [%d] %s is supported", i, v.ToGFXString()))
-			}
-		}
-	} else {
-		slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
-	}
-
-	if len(skip) >= len(gfx) {
-		slog.Info("all detected amdgpus are skipped, falling back to CPU")
-		return
-	}
-
-	ids := make([]int, len(gfx))
-	i := 0
-	for k := range gfx {
-		ids[i] = k
-		i++
-	}
-	amdProcMemLookup(resp, skip, ids)
-	if resp.memInfo.DeviceCount == 0 {
-		return
-	}
-	if len(skip) > 0 {
-		amdSetVisibleDevices(ids, skip)
-	}
-}
-
-// Walk the sysfs nodes for the available GPUs and gather information from them
-// skipping over any devices in the skip map
-func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
-	resp.memInfo.DeviceCount = 0
-	resp.memInfo.TotalMemory = 0
-	resp.memInfo.FreeMemory = 0
-	if len(ids) == 0 {
-		slog.Debug("discovering all amdgpu devices")
-		entries, err := os.ReadDir(AMDNodesSysfsDir)
-		if err != nil {
-			slog.Warn(fmt.Sprintf("failed to read amdgpu sysfs %s - %s", AMDNodesSysfsDir, err))
-			return
-		}
-		for _, node := range entries {
-			if !node.IsDir() {
-				continue
-			}
-			id, err := strconv.Atoi(node.Name())
-			if err != nil {
-				slog.Warn("malformed amdgpu sysfs node id " + node.Name())
-				continue
-			}
-			ids = append(ids, id)
-		}
-	}
-	slog.Debug(fmt.Sprintf("discovering amdgpu devices %v", ids))
-
-	for _, id := range ids {
-		if _, skipped := skip[id]; skipped {
+			slog.Debug("failed to open sysfs node", "file", match, "error", err)
 			continue
 		}
+		defer fp.Close()
+		nodeID, err := strconv.Atoi(filepath.Base(filepath.Dir(match)))
+		if err != nil {
+			slog.Debug("failed to parse node ID", "error", err)
+			continue
+		}
+
+		scanner := bufio.NewScanner(fp)
+		isCPU := false
+		var major, minor, patch uint64
+		for scanner.Scan() {
+			line := strings.TrimSpace(scanner.Text())
+			// Note: we could also use "cpu_cores_count X" where X is greater than zero to detect CPUs
+			if strings.HasPrefix(line, "gfx_target_version") {
+				ver := strings.Fields(line)
+
+				// Detect CPUs
+				if len(ver) == 2 && ver[1] == "0" {
+					slog.Debug("detected CPU " + match)
+					isCPU = true
+					break
+				}
+
+				if len(ver) != 2 || len(ver[1]) < 5 {
+					slog.Warn("malformed "+match, "gfx_target_version", line)
+					// If this winds up being a CPU, our offsets may be wrong
+					continue
+				}
+				l := len(ver[1])
+				var err1, err2, err3 error
+				patch, err1 = strconv.ParseUint(ver[1][l-2:l], 10, 32)
+				minor, err2 = strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
+				major, err3 = strconv.ParseUint(ver[1][:l-4], 10, 32)
+				if err1 != nil || err2 != nil || err3 != nil {
+					slog.Debug("malformed int " + line)
+					continue
+				}
+			}
+
+			// TODO - any other properties we want to extract and record?
+			// vendor_id + device_id -> pci lookup for "Name"
+			// Other metrics that may help us understand relative performance between multiple GPUs
+		}
+
+		if isCPU {
+			cpuCount++
+			continue
+		}
+
+		// CPUs are always first in the list
+		gpuID := nodeID - cpuCount
+
+		// 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 []GpuInfo{}
+		}
+
+		if int(major) < RocmComputeMin {
+			slog.Warn(fmt.Sprintf("amdgpu too old gfx%d%d%x", major, minor, patch), "gpu", gpuID)
+			continue
+		}
+
+		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
-		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id), GPUTotalMemoryFileGlob)
+		propGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUTotalMemoryFileGlob)
 		propFiles, err := filepath.Glob(propGlob)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("error looking up total GPU memory: %s %s", propGlob, err))
+			slog.Warn("error looking up total GPU memory", "glob", propGlob, "error", err)
 		}
 		// 1 or more memory banks - sum the values of all of them
 		for _, propFile := range propFiles {
 			fp, err := os.Open(propFile)
 			if err != nil {
-				slog.Warn(fmt.Sprintf("failed to open sysfs node file %s: %s", propFile, err))
+				slog.Warn("failed to open sysfs node", "file", propFile, "erroir", err)
 				continue
 			}
 			defer fp.Close()
@ -205,42 +179,113 @@ func amdProcMemLookup(resp *GpuInfo, skip map[int]interface{}, ids []int) {
 			}
 		}
 		if totalMemory == 0 {
+			slog.Warn("amdgpu reports zero total memory", "gpu", gpuID)
 			continue
 		}
-		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(id), GPUUsedMemoryFileGlob)
+		usedGlob := filepath.Join(AMDNodesSysfsDir, strconv.Itoa(nodeID), GPUUsedMemoryFileGlob)
 		usedFiles, err := filepath.Glob(usedGlob)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("error looking up used GPU memory: %s %s", usedGlob, err))
+			slog.Warn("error looking up used GPU memory", "glob", usedGlob, "error", err)
 			continue
 		}
 		for _, usedFile := range usedFiles {
 			fp, err := os.Open(usedFile)
 			if err != nil {
-				slog.Warn(fmt.Sprintf("failed to open sysfs node file %s: %s", usedFile, err))
+				slog.Warn("failed to open sysfs node", "file", usedFile, "error", err)
 				continue
 			}
 			defer fp.Close()
 			data, err := io.ReadAll(fp)
 			if err != nil {
-				slog.Warn(fmt.Sprintf("failed to read sysfs node file %s: %s", usedFile, err))
+				slog.Warn("failed to read sysfs node", "file", usedFile, "error", err)
 				continue
 			}
 			used, err := strconv.ParseUint(strings.TrimSpace(string(data)), 10, 64)
 			if err != nil {
-				slog.Warn(fmt.Sprintf("malformed used memory %s: %s", string(data), err))
+				slog.Warn("malformed used memory", "data", string(data), "error", err)
 				continue
 			}
 			usedMemory += used
 		}
-		slog.Info(fmt.Sprintf("[%d] amdgpu totalMemory %d", id, totalMemory))
-		slog.Info(fmt.Sprintf("[%d] amdgpu freeMemory  %d", id, (totalMemory - usedMemory)))
-		resp.memInfo.DeviceCount++
-		resp.memInfo.TotalMemory += totalMemory
-		resp.memInfo.FreeMemory += (totalMemory - usedMemory)
+
+		// 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", gpuID, "total", format.HumanBytes2(totalMemory))
+			continue
+		}
+
+		slog.Info("amdgpu memory", "gpu", gpuID, "total", format.HumanBytes2(totalMemory))
+		slog.Info("amdgpu memory", "gpu", gpuID, "available", format.HumanBytes2(totalMemory-usedMemory))
+		gpuInfo := GpuInfo{
+			Library: "rocm",
+			memInfo: memInfo{
+				TotalMemory: totalMemory,
+				FreeMemory:  (totalMemory - usedMemory),
+			},
+			ID: fmt.Sprintf("%d", gpuID),
+			// Name: not exposed in sysfs directly, would require pci device id lookup
+			Major:         int(major),
+			Minor:         int(minor),
+			Patch:         int(patch),
+			MinimumMemory: rocmMinimumMemory,
+		}
+
+		// 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
+			for _, visible := range visibleDevices {
+				if visible == gpuInfo.ID {
+					include = true
+					break
+				}
+			}
+			if !include {
+				slog.Info("filtering out device per user request", "id", gpuInfo.ID, "visible_devices", visibleDevices)
+				continue
+			}
+		}
+
+		// Final validation is gfx compatibility - load the library if we haven't already loaded it
+		// even if the user overrides, we still need to validate the library
+		if libDir == "" {
+			libDir, err = AMDValidateLibDir()
+			if err != nil {
+				slog.Warn("unable to verify rocm library, will use cpu", "error", err)
+				return []GpuInfo{}
+			}
+		}
+		gpuInfo.DependencyPath = libDir
+
+		if gfxOverride == "" {
+			// Only load supported list once
+			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 []GpuInfo{}
+				}
+				slog.Debug("rocm supported GPUs", "types", supported)
+			}
+			gfx := fmt.Sprintf("gfx%d%d%x", gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch)
+			if !slices.Contains[[]string, string](supported, gfx) {
+				slog.Warn("amdgpu is not supported", "gpu", gpuInfo.ID, "gpu_type", gfx, "library", libDir, "supported_types", supported)
+				// TODO - consider discrete markdown just for ROCM troubleshooting?
+				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/gpu.md#overrides for HSA_OVERRIDE_GFX_VERSION usage")
+				continue
+			} else {
+				slog.Info("amdgpu is supported", "gpu", gpuInfo.ID, "gpu_type", gfx)
+			}
+		} else {
+			slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
+		}
+
+		// The GPU has passed all the verification steps and is supported
+		resp = append(resp, gpuInfo)
 	}
-	if resp.memInfo.DeviceCount > 0 {
-		resp.Library = "rocm"
+	if len(resp) == 0 {
+		slog.Info("no compatible amdgpu devices detected")
 	}
+	return resp
 }

 // Quick check for AMD driver so we can skip amdgpu discovery if not present
@ -252,87 +297,24 @@ func AMDDetected() bool {
 		slog.Debug("amdgpu driver not detected " + sysfsDir)
 		return false
 	} else if err != nil {
-		slog.Debug(fmt.Sprintf("error looking up amd driver %s %s", sysfsDir, err))
+		slog.Debug("error looking up amd driver", "path", sysfsDir, "error", err)
 		return false
 	}
 	return true
 }

-func setupLink(source, target string) error {
-	if err := os.RemoveAll(target); err != nil {
-		return fmt.Errorf("failed to remove old rocm directory %s %w", target, err)
-	}
-	if err := os.Symlink(source, target); err != nil {
-		return fmt.Errorf("failed to create link %s => %s %w", source, target, err)
-	}
-	slog.Debug(fmt.Sprintf("host rocm linked %s => %s", source, target))
-	return nil
-}
-
-// Ensure the AMD rocm lib dir is wired up
 // Prefer to use host installed ROCm, as long as it meets our minimum requirements
 // failing that, tell the user how to download it on their own
 func AMDValidateLibDir() (string, error) {
-	// We rely on the rpath compiled into our library to find rocm
-	// so we establish a symlink to wherever we find it on the system
-	// to <payloads>/rocm
-	payloadsDir, err := PayloadsDir()
-	if err != nil {
-		return "", err
-	}
-
-	// If we already have a rocm dependency wired, nothing more to do
-	rocmTargetDir := filepath.Clean(filepath.Join(payloadsDir, "..", "rocm"))
-	if rocmLibUsable(rocmTargetDir) {
-		return rocmTargetDir, nil
-	}
-
-	// next to the running binary
-	exe, err := os.Executable()
+	libDir, err := commonAMDValidateLibDir()
 	if err == nil {
-		peerDir := filepath.Dir(exe)
-		if rocmLibUsable(peerDir) {
-			slog.Debug("detected ROCM next to ollama executable " + peerDir)
-			return rocmTargetDir, setupLink(peerDir, rocmTargetDir)
-		}
-		peerDir = filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(peerDir) {
-			slog.Debug("detected ROCM next to ollama executable " + peerDir)
-			return rocmTargetDir, setupLink(peerDir, rocmTargetDir)
-		}
+		return libDir, nil
 	}

 	// Well known ollama installer path
 	installedRocmDir := "/usr/share/ollama/lib/rocm"
 	if rocmLibUsable(installedRocmDir) {
-		return rocmTargetDir, setupLink(installedRocmDir, rocmTargetDir)
-	}
-
-	// Prefer explicit HIP env var
-	hipPath := os.Getenv("HIP_PATH")
-	if hipPath != "" {
-		hipLibDir := filepath.Join(hipPath, "lib")
-		if rocmLibUsable(hipLibDir) {
-			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
-			return rocmTargetDir, setupLink(hipLibDir, rocmTargetDir)
-		}
-	}
-
-	// Scan the library path for potential matches
-	ldPaths := strings.Split(os.Getenv("LD_LIBRARY_PATH"), ":")
-	for _, ldPath := range ldPaths {
-		d, err := filepath.Abs(ldPath)
-		if err != nil {
-			continue
-		}
-		if rocmLibUsable(d) {
-			return rocmTargetDir, setupLink(d, rocmTargetDir)
-		}
-	}
-
-	// Well known location(s)
-	if rocmLibUsable("/opt/rocm/lib") {
-		return rocmTargetDir, setupLink("/opt/rocm/lib", rocmTargetDir)
+		return installedRocmDir, nil
 	}

 	// If we still haven't found a usable rocm, the user will have to install it on their own
@ -356,63 +338,3 @@ func AMDDriverVersion() (string, error) {
 	}
 	return strings.TrimSpace(string(verString)), nil
 }
-
-func AMDGFXVersions() map[int]Version {
-	res := map[int]Version{}
-	matches, _ := filepath.Glob(GPUPropertiesFileGlob)
-	for _, match := range matches {
-		fp, err := os.Open(match)
-		if err != nil {
-			slog.Debug(fmt.Sprintf("failed to open sysfs node file %s: %s", match, err))
-			continue
-		}
-		defer fp.Close()
-		i, err := strconv.Atoi(filepath.Base(filepath.Dir(match)))
-		if err != nil {
-			slog.Debug(fmt.Sprintf("failed to parse node ID %s", err))
-			continue
-		}
-
-		scanner := bufio.NewScanner(fp)
-		for scanner.Scan() {
-			line := strings.TrimSpace(scanner.Text())
-			if strings.HasPrefix(line, "gfx_target_version") {
-				ver := strings.Fields(line)
-				if len(ver) != 2 || len(ver[1]) < 5 {
-
-					if ver[1] == "0" {
-						// Silently skip the CPU
-						continue
-					} else {
-						slog.Debug("malformed " + line)
-					}
-					res[i] = Version{
-						Major: 0,
-						Minor: 0,
-						Patch: 0,
-					}
-					continue
-				}
-				l := len(ver[1])
-				patch, err1 := strconv.ParseUint(ver[1][l-2:l], 10, 32)
-				minor, err2 := strconv.ParseUint(ver[1][l-4:l-2], 10, 32)
-				major, err3 := strconv.ParseUint(ver[1][:l-4], 10, 32)
-				if err1 != nil || err2 != nil || err3 != nil {
-					slog.Debug("malformed int " + line)
-					continue
-				}
-
-				res[i] = Version{
-					Major: uint(major),
-					Minor: uint(minor),
-					Patch: uint(patch),
-				}
-			}
-		}
-	}
-	return res
-}
-
-func (v Version) ToGFXString() string {
-	return fmt.Sprintf("gfx%d%d%d", v.Major, v.Minor, v.Patch)
-}
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@ -7,7 +7,10 @@ import (
 	"os"
 	"path/filepath"
 	"slices"
+	"strconv"
 	"strings"
+
+	"github.com/ollama/ollama/format"
 )

 const (
@ -22,36 +25,32 @@ var (
 	ROCmLibGlobs = []string{"hipblas.dll", "rocblas"} // TODO - probably include more coverage of files here...
 )

-func AMDGetGPUInfo(resp *GpuInfo) {
+func AMDGetGPUInfo() []GpuInfo {
+	resp := []GpuInfo{}
 	hl, err := NewHipLib()
 	if err != nil {
 		slog.Debug(err.Error())
-		return
+		return nil
 	}
 	defer hl.Release()
-	skip := map[int]interface{}{}
-	ids := []int{}
-	resp.memInfo.DeviceCount = 0
-	resp.memInfo.TotalMemory = 0
-	resp.memInfo.FreeMemory = 0

 	ver, err := hl.AMDDriverVersion()
 	if err == nil {
 		slog.Info("AMD Driver: " + ver)
 	} else {
 		// For now this is benign, but we may eventually need to fail compatibility checks
-		slog.Debug(fmt.Sprintf("error looking up amd driver version: %s", err))
+		slog.Debug("error looking up amd driver version", "error", err)
 	}

-	// Note: the HIP library automatically handles HIP_VISIBLE_DEVICES
+	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
 	if count == 0 {
-		return
+		return nil
 	}
 	libDir, err := AMDValidateLibDir()
 	if err != nil {
-		slog.Warn(fmt.Sprintf("unable to verify rocm library, will use cpu: %s", err))
-		return
+		slog.Warn("unable to verify rocm library, will use cpu", "error", err)
+		return nil
 	}

 	var supported []string
@ -59,95 +58,120 @@ func AMDGetGPUInfo(resp *GpuInfo) {
 	if gfxOverride == "" {
 		supported, err = GetSupportedGFX(libDir)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("failed to lookup supported GFX types, falling back to CPU mode: %s", err))
-			return
+			slog.Warn("failed to lookup supported GFX types, falling back to CPU mode", "error", err)
+			return nil
 		}
 	} else {
 		slog.Debug("skipping rocm gfx compatibility check with HSA_OVERRIDE_GFX_VERSION=" + gfxOverride)
 	}

-	slog.Info(fmt.Sprintf("detected %d hip devices", count))
+	slog.Info("detected hip devices", "count", count)
+	// TODO how to determine the underlying device ID when visible devices is causing this to subset?
 	for i := 0; i < count; i++ {
-		ids = append(ids, i)
 		err = hl.HipSetDevice(i)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
-			skip[i] = struct{}{}
+			slog.Warn("set device", "id", i, "error", err)
 			continue
 		}

 		props, err := hl.HipGetDeviceProperties(i)
 		if err != nil {
-			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
-			skip[i] = struct{}{}
+			slog.Warn("get properties", "id", i, "error", err)
 			continue
 		}
 		n := bytes.IndexByte(props.Name[:], 0)
 		name := string(props.Name[:n])
-		slog.Info(fmt.Sprintf("[%d] Name: %s", i, name))
+		// TODO is UUID actually populated on windows?
+		// Can luid be used on windows for setting visible devices (and is it actually set?)
 		n = bytes.IndexByte(props.GcnArchName[:], 0)
 		gfx := string(props.GcnArchName[:n])
-		slog.Info(fmt.Sprintf("[%d] GcnArchName: %s", i, gfx))
+		slog.Info("hip device", "id", i, "name", name, "gfx", gfx)
+		var major, minor, patch string
+		switch len(gfx) {
+		case 6:
+			major, minor, patch = gfx[3:4], gfx[4:5], gfx[5:]
+		case 7:
+			major, minor, patch = gfx[3:5], gfx[5:6], gfx[6:]
+		}
 		//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(fmt.Sprintf("iGPU detected [%d] skipping", i))
-			skip[i] = struct{}{}
+			slog.Info("iGPU detected skipping", "id", i)
 			continue
 		}
 		if gfxOverride == "" {
 			if !slices.Contains[[]string, string](supported, gfx) {
-				slog.Warn(fmt.Sprintf("amdgpu [%d] %s is not supported by %s %v", i, gfx, libDir, supported))
+				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")
-				skip[i] = struct{}{}
 				continue
 			} else {
-				slog.Info(fmt.Sprintf("amdgpu [%d] %s is supported", i, gfx))
+				slog.Info("amdgpu is supported", "gpu", i, "gpu_type", gfx)
 			}
 		}

-		totalMemory, freeMemory, err := hl.HipMemGetInfo()
+		freeMemory, totalMemory, err := hl.HipMemGetInfo()
 		if err != nil {
-			slog.Warn(fmt.Sprintf("[%d] %s", i, err))
+			slog.Warn("get mem info", "id", i, "error", err)
 			continue
 		}

-		// TODO according to docs, freeMem may lie on windows!
-		slog.Info(fmt.Sprintf("[%d] Total Mem: %d", i, totalMemory))
-		slog.Info(fmt.Sprintf("[%d] Free Mem:  %d", i, freeMemory))
-		resp.memInfo.DeviceCount++
-		resp.memInfo.TotalMemory += totalMemory
-		resp.memInfo.FreeMemory += freeMemory
+		// 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
+		}
+
+		// TODO revisit this once ROCm v6 is available on windows.
+		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
+		slog.Info("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
+		slog.Info("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
+		gpuInfo := GpuInfo{
+			Library: "rocm",
+			memInfo: memInfo{
+				TotalMemory: totalMemory,
+				FreeMemory:  freeMemory,
+			},
+			ID:             fmt.Sprintf("%d", i), // TODO this is probably wrong if we specify visible devices
+			DependencyPath: libDir,
+			MinimumMemory:  rocmMinimumMemory,
+		}
+		if major != "" {
+			gpuInfo.Major, err = strconv.Atoi(major)
+			if err != nil {
+				slog.Info("failed to parse version", "version", gfx, "error", err)
+			}
+		}
+		if minor != "" {
+			gpuInfo.Minor, err = strconv.Atoi(minor)
+			if err != nil {
+				slog.Info("failed to parse version", "version", gfx, "error", err)
+			}
+		}
+		if patch != "" {
+			// Patch rev is hex; e.g. gfx90a
+			p, err := strconv.ParseInt(patch, 16, 0)
+			if err != nil {
+				slog.Info("failed to parse version", "version", gfx, "error", err)
+			} else {
+				gpuInfo.Patch = int(p)
+			}
+		}
+		if gpuInfo.Major < RocmComputeMin {
+			slog.Warn(fmt.Sprintf("amdgpu [%s] too old gfx%d%d%x", gpuInfo.ID, gpuInfo.Major, gpuInfo.Minor, gpuInfo.Patch))
+			continue
+		}
+
+		resp = append(resp, gpuInfo)
 	}
-	if resp.memInfo.DeviceCount > 0 {
-		resp.Library = "rocm"
-	}
-	// Abort if all GPUs are skipped
-	if len(skip) >= count {
-		slog.Info("all detected amdgpus are skipped, falling back to CPU")
-		return
-	}
-	if len(skip) > 0 {
-		amdSetVisibleDevices(ids, skip)
-	}
-	UpdatePath(libDir)
+
+	return resp
 }

 func AMDValidateLibDir() (string, error) {
-	// On windows non-admins typically can't create links
-	// so instead of trying to rely on rpath and a link in
-	// $LibDir/rocm, we instead rely on setting PATH to point
-	// to the location of the ROCm library
-
-	// Installer payload location if we're running the installed binary
-	exe, err := os.Executable()
+	libDir, err := commonAMDValidateLibDir()
 	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
+		return libDir, nil
 	}

 	// Installer payload (if we're running from some other location)
@ -159,21 +183,6 @@ func AMDValidateLibDir() (string, error) {
 		return rocmTargetDir, nil
 	}

-	// Prefer explicit HIP env var
-	hipPath := os.Getenv("HIP_PATH")
-	if hipPath != "" {
-		hipLibDir := filepath.Join(hipPath, "bin")
-		if rocmLibUsable(hipLibDir) {
-			slog.Debug("detected ROCM via HIP_PATH=" + hipPath)
-			return hipLibDir, nil
-		}
-	}
-
-	// Well known location(s)
-	if rocmLibUsable(RocmStandardLocation) {
-		return RocmStandardLocation, nil
-	}
-
 	// Should not happen on windows since we include it in the installer, but stand-alone binary might hit this
 	slog.Warn("amdgpu detected, but no compatible rocm library found.  Please install ROCm")
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
--- a/gpu/assets.go
+++ b/gpu/assets.go
@ -1,13 +1,17 @@
 package gpu

 import (
+	"errors"
 	"fmt"
 	"log/slog"
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"sync"
+	"syscall"
+	"time"
 )

 var (
@ -18,11 +22,77 @@ var (
 func PayloadsDir() (string, error) {
 	lock.Lock()
 	defer lock.Unlock()
+	var err error
 	if payloadsDir == "" {
-		tmpDir, err := os.MkdirTemp("", "ollama")
-		if err != nil {
-			return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+		runnersDir := os.Getenv("OLLAMA_RUNNERS_DIR")
+		// On Windows we do not carry the payloads inside the main executable
+		if runtime.GOOS == "windows" && runnersDir == "" {
+			appExe, err := os.Executable()
+			if err != nil {
+				slog.Error("failed to lookup executable path", "error", err)
+				return "", err
+			}
+
+			cwd, err := os.Getwd()
+			if err != nil {
+				slog.Error("failed to lookup working directory", "error", err)
+				return "", err
+			}
+
+			var paths []string
+			for _, root := range []string{filepath.Dir(appExe), cwd} {
+				paths = append(paths,
+					filepath.Join(root),
+					filepath.Join(root, "windows-"+runtime.GOARCH),
+					filepath.Join(root, "dist", "windows-"+runtime.GOARCH),
+				)
+			}
+
+			// Try a few variations to improve developer experience when building from source in the local tree
+			for _, p := range paths {
+				candidate := filepath.Join(p, "ollama_runners")
+				_, err := os.Stat(candidate)
+				if err == nil {
+					runnersDir = candidate
+					break
+				}
+			}
+			if runnersDir == "" {
+				err = fmt.Errorf("unable to locate llm runner directory.  Set OLLAMA_RUNNERS_DIR to the location of 'ollama_runners'")
+				slog.Error("incomplete distribution", "error", err)
+				return "", err
+			}
 		}
+		if runnersDir != "" {
+			payloadsDir = runnersDir
+			return payloadsDir, nil
+		}
+
+		// The remainder only applies on non-windows where we still carry payloads in the main executable
+		cleanupTmpDirs()
+		tmpDir := os.Getenv("OLLAMA_TMPDIR")
+		if tmpDir == "" {
+			tmpDir, err = os.MkdirTemp("", "ollama")
+			if err != nil {
+				return "", fmt.Errorf("failed to generate tmp dir: %w", err)
+			}
+		} else {
+			err = os.MkdirAll(tmpDir, 0755)
+			if err != nil {
+				return "", fmt.Errorf("failed to generate tmp dir %s: %w", tmpDir, err)
+			}
+		}
+
+		// Track our pid so we can clean up orphaned tmpdirs
+		pidFilePath := filepath.Join(tmpDir, "ollama.pid")
+		pidFile, err := os.OpenFile(pidFilePath, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.ModePerm)
+		if err != nil {
+			return "", err
+		}
+		if _, err := pidFile.Write([]byte(fmt.Sprint(os.Getpid()))); err != nil {
+			return "", err
+		}
+
 		// We create a distinct subdirectory for payloads within the tmpdir
 		// This will typically look like /tmp/ollama3208993108/runners on linux
 		payloadsDir = filepath.Join(tmpDir, "runners")
@ -30,16 +100,52 @@ func PayloadsDir() (string, error) {
 	return payloadsDir, nil
 }

+// Best effort to clean up prior tmpdirs
+func cleanupTmpDirs() {
+	dirs, err := filepath.Glob(filepath.Join(os.TempDir(), "ollama*"))
+	if err != nil {
+		return
+	}
+	for _, d := range dirs {
+		info, err := os.Stat(d)
+		if err != nil || !info.IsDir() {
+			continue
+		}
+		raw, err := os.ReadFile(filepath.Join(d, "ollama.pid"))
+		if err == nil {
+			pid, err := strconv.Atoi(string(raw))
+			if err == nil {
+				if proc, err := os.FindProcess(int(pid)); err == nil && !errors.Is(proc.Signal(syscall.Signal(0)), os.ErrProcessDone) {
+					// Another running ollama, ignore this tmpdir
+					continue
+				}
+			}
+		} else {
+			slog.Debug("failed to open ollama.pid", "path", d, "error", err)
+		}
+		err = os.RemoveAll(d)
+		if err != nil {
+			slog.Debug("unable to cleanup stale tmpdir", "path", d, "error", err)
+		}
+	}
+}
+
 func Cleanup() {
 	lock.Lock()
 	defer lock.Unlock()
-	if payloadsDir != "" {
+	runnersDir := os.Getenv("OLLAMA_RUNNERS_DIR")
+	if payloadsDir != "" && runnersDir == "" && runtime.GOOS != "windows" {
 		// We want to fully clean up the tmpdir parent of the payloads dir
 		tmpDir := filepath.Clean(filepath.Join(payloadsDir, ".."))
 		slog.Debug("cleaning up", "dir", tmpDir)
 		err := os.RemoveAll(tmpDir)
 		if err != nil {
-			slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
+			// On windows, if we remove too quickly the llama.dll may still be in-use and fail to remove
+			time.Sleep(1000 * time.Millisecond)
+			err = os.RemoveAll(tmpDir)
+			if err != nil {
+				slog.Warn("failed to clean up", "dir", tmpDir, "err", err)
+			}
 		}
 	}
 }
@ -60,7 +166,7 @@ func UpdatePath(dir string) {
 			}
 		}
 		newPath := strings.Join(append([]string{dir}, pathComponents...), ";")
-		slog.Info(fmt.Sprintf("Updating PATH to %s", newPath))
+		slog.Info("updating", "PATH", newPath)
 		os.Setenv("PATH", newPath)
 	}
 	// linux and darwin rely on rpath
--- a/gpu/cuda_common.go
+++ b/gpu/cuda_common.go
@ -0,0 +1,22 @@
+//go:build linux || windows
+
+package gpu
+
+import (
+	"log/slog"
+	"strings"
+)
+
+func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "cuda" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("cudaGetVisibleDevicesEnv skipping over non-cuda device", "library", info.Library)
+			continue
+		}
+		ids = append(ids, info.ID)
+	}
+	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
+
+}
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@ -16,84 +16,109 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
-	"strconv"
 	"strings"
 	"sync"
 	"unsafe"
+
+	"github.com/ollama/ollama/format"
 )

 type handles struct {
-	cuda *C.cuda_handle_t
+	deviceCount int
+	cudart      *C.cudart_handle_t
 }

+const (
+	cudaMinimumMemory = 457 * format.MebiByte
+	rocmMinimumMemory = 457 * format.MebiByte
+)
+
 var gpuMutex sync.Mutex
-var gpuHandles *handles = nil

 // With our current CUDA compile flags, older than 5.0 will not work properly
 var CudaComputeMin = [2]C.int{5, 0}

-// Possible locations for the nvidia-ml library
-var CudaLinuxGlobs = []string{
-	"/usr/local/cuda/lib64/libnvidia-ml.so*",
-	"/usr/lib/x86_64-linux-gnu/nvidia/current/libnvidia-ml.so*",
-	"/usr/lib/x86_64-linux-gnu/libnvidia-ml.so*",
-	"/usr/lib/wsl/lib/libnvidia-ml.so*",
-	"/usr/lib/wsl/drivers/*/libnvidia-ml.so*",
-	"/opt/cuda/lib64/libnvidia-ml.so*",
-	"/usr/lib*/libnvidia-ml.so*",
-	"/usr/local/lib*/libnvidia-ml.so*",
-	"/usr/lib/aarch64-linux-gnu/nvidia/current/libnvidia-ml.so*",
-	"/usr/lib/aarch64-linux-gnu/libnvidia-ml.so*",
+var RocmComputeMin = 9

-	// TODO: are these stubs ever valid?
-	"/opt/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so*",
+// TODO find a better way to detect iGPU instead of minimum memory
+const IGPUMemLimit = 1 * format.GibiByte // 512G is what they typically report, so anything less than 1G must be iGPU
+
+var CudartLinuxGlobs = []string{
+	"/usr/local/cuda/lib64/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/x86_64-linux-gnu/libcudart.so*",
+	"/usr/lib/wsl/lib/libcudart.so*",
+	"/usr/lib/wsl/drivers/*/libcudart.so*",
+	"/opt/cuda/lib64/libcudart.so*",
+	"/usr/local/cuda*/targets/aarch64-linux/lib/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/nvidia/current/libcudart.so*",
+	"/usr/lib/aarch64-linux-gnu/libcudart.so*",
+	"/usr/local/cuda/lib*/libcudart.so*",
+	"/usr/lib*/libcudart.so*",
+	"/usr/local/lib*/libcudart.so*",
 }

-var CudaWindowsGlobs = []string{
-	"c:\\Windows\\System32\\nvml.dll",
+var CudartWindowsGlobs = []string{
+	"c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\bin\\cudart64_*.dll",
 }

+// Jetson devices have JETSON_JETPACK="x.y.z" factory set to the Jetpack version installed.
+// Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
+var CudaTegra string = os.Getenv("JETSON_JETPACK")
+
 // Note: gpuMutex must already be held
-func initGPUHandles() {
+func initGPUHandles() *handles {

 	// TODO - if the ollama build is CPU only, don't do these checks as they're irrelevant and confusing

-	gpuHandles = &handles{nil}
-	var cudaMgmtName string
-	var cudaMgmtPatterns []string
+	gpuHandles := &handles{}
+	var cudartMgmtName string
+	var cudartMgmtPatterns []string
+
+	tmpDir, _ := PayloadsDir()
 	switch runtime.GOOS {
 	case "windows":
-		cudaMgmtName = "nvml.dll"
-		cudaMgmtPatterns = make([]string, len(CudaWindowsGlobs))
-		copy(cudaMgmtPatterns, CudaWindowsGlobs)
+		cudartMgmtName = "cudart64_*.dll"
+		localAppData := os.Getenv("LOCALAPPDATA")
+		cudartMgmtPatterns = []string{filepath.Join(localAppData, "Programs", "Ollama", cudartMgmtName)}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartWindowsGlobs...)
 	case "linux":
-		cudaMgmtName = "libnvidia-ml.so"
-		cudaMgmtPatterns = make([]string, len(CudaLinuxGlobs))
-		copy(cudaMgmtPatterns, CudaLinuxGlobs)
+		cudartMgmtName = "libcudart.so*"
+		if tmpDir != "" {
+			// TODO - add "payloads" for subprocess
+			cudartMgmtPatterns = []string{filepath.Join(tmpDir, "cuda*", cudartMgmtName)}
+		}
+		cudartMgmtPatterns = append(cudartMgmtPatterns, CudartLinuxGlobs...)
 	default:
-		return
+		return gpuHandles
 	}

-	slog.Info("Detecting GPU type")
-	cudaLibPaths := FindGPULibs(cudaMgmtName, cudaMgmtPatterns)
-	if len(cudaLibPaths) > 0 {
-		cuda := LoadCUDAMgmt(cudaLibPaths)
-		if cuda != nil {
-			slog.Info("Nvidia GPU detected")
-			gpuHandles.cuda = cuda
-			return
+	slog.Info("Detecting GPUs")
+	cudartLibPaths := FindGPULibs(cudartMgmtName, cudartMgmtPatterns)
+	if len(cudartLibPaths) > 0 {
+		deviceCount, cudart, libPath := LoadCUDARTMgmt(cudartLibPaths)
+		if cudart != nil {
+			slog.Info("detected GPUs", "library", libPath, "count", deviceCount)
+			gpuHandles.cudart = cudart
+			gpuHandles.deviceCount = deviceCount
+			return gpuHandles
 		}
 	}
+	return gpuHandles
 }

-func GetGPUInfo() GpuInfo {
+func GetGPUInfo() GpuInfoList {
 	// TODO - consider exploring lspci (and equivalent on windows) to check for
 	// GPUs so we can report warnings if we see Nvidia/AMD but fail to load the libraries
 	gpuMutex.Lock()
 	defer gpuMutex.Unlock()
-	if gpuHandles == nil {
-		initGPUHandles()
-	}
+
+	gpuHandles := initGPUHandles()
+	defer func() {
+		if gpuHandles.cudart != nil {
+			C.cudart_release(*gpuHandles.cudart)
+		}
+	}()

 	// All our GPU builds on x86 have AVX enabled, so fallback to CPU if we don't detect at least AVX
 	cpuVariant := GetCPUVariant()
@ -102,50 +127,63 @@ func GetGPUInfo() GpuInfo {
 	}

 	var memInfo C.mem_info_t
-	resp := GpuInfo{}
-	if gpuHandles.cuda != nil && (cpuVariant != "" || runtime.GOARCH != "amd64") {
-		C.cuda_check_vram(*gpuHandles.cuda, &memInfo)
+	resp := []GpuInfo{}
+
+	// NVIDIA first
+	for i := 0; i < gpuHandles.deviceCount; i++ {
+		// TODO once we support CPU compilation variants of GPU libraries refine this...
+		if cpuVariant == "" && runtime.GOARCH == "amd64" {
+			continue
+		}
+		gpuInfo := GpuInfo{
+			Library: "cuda",
+		}
+		C.cudart_check_vram(*gpuHandles.cudart, C.int(i), &memInfo)
 		if memInfo.err != nil {
-			slog.Info(fmt.Sprintf("error looking up CUDA GPU memory: %s", C.GoString(memInfo.err)))
+			slog.Info("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
 			C.free(unsafe.Pointer(memInfo.err))
-		} else if memInfo.count > 0 {
-			// Verify minimum compute capability
-			var cc C.cuda_compute_capability_t
-			C.cuda_compute_capability(*gpuHandles.cuda, &cc)
-			if cc.err != nil {
-				slog.Info(fmt.Sprintf("error looking up CUDA GPU compute capability: %s", C.GoString(cc.err)))
-				C.free(unsafe.Pointer(cc.err))
-			} else if cc.major > CudaComputeMin[0] || (cc.major == CudaComputeMin[0] && cc.minor >= CudaComputeMin[1]) {
-				slog.Info(fmt.Sprintf("CUDA Compute Capability detected: %d.%d", cc.major, cc.minor))
-				resp.Library = "cuda"
-			} else {
-				slog.Info(fmt.Sprintf("CUDA GPU is too old. Falling back to CPU mode. Compute Capability detected: %d.%d", cc.major, cc.minor))
-			}
+			continue
 		}
-	} else {
-		AMDGetGPUInfo(&resp)
-		if resp.Library != "" {
-			return resp
+		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
 		}
-	}
-	if resp.Library == "" {
-		C.cpu_check_ram(&memInfo)
-		resp.Library = "cpu"
-		resp.Variant = cpuVariant
-	}
-	if memInfo.err != nil {
-		slog.Info(fmt.Sprintf("error looking up CPU memory: %s", C.GoString(memInfo.err)))
-		C.free(unsafe.Pointer(memInfo.err))
-		return resp
+		gpuInfo.TotalMemory = uint64(memInfo.total)
+		gpuInfo.FreeMemory = uint64(memInfo.free)
+		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+		gpuInfo.Major = int(memInfo.major)
+		gpuInfo.Minor = int(memInfo.minor)
+		gpuInfo.MinimumMemory = cudaMinimumMemory
+
+		// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
+		resp = append(resp, gpuInfo)
+	}
+
+	// Then AMD
+	resp = append(resp, AMDGetGPUInfo()...)
+
+	if len(resp) == 0 {
+		C.cpu_check_ram(&memInfo)
+		if memInfo.err != nil {
+			slog.Info("error looking up CPU memory", "error", C.GoString(memInfo.err))
+			C.free(unsafe.Pointer(memInfo.err))
+			return resp
+		}
+		gpuInfo := GpuInfo{
+			Library: "cpu",
+			Variant: cpuVariant,
+		}
+		gpuInfo.TotalMemory = uint64(memInfo.total)
+		gpuInfo.FreeMemory = uint64(memInfo.free)
+		gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
+
+		resp = append(resp, gpuInfo)
 	}

-	resp.DeviceCount = uint32(memInfo.count)
-	resp.FreeMemory = uint64(memInfo.free)
-	resp.TotalMemory = uint64(memInfo.total)
 	return resp
 }

-func getCPUMem() (memInfo, error) {
+func GetCPUMem() (memInfo, error) {
 	var ret memInfo
 	var info C.mem_info_t
 	C.cpu_check_ram(&info)
@ -158,37 +196,11 @@ func getCPUMem() (memInfo, error) {
 	return ret, nil
 }

-func CheckVRAM() (int64, error) {
-	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseInt(userLimit, 10, 64)
-		if err != nil {
-			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
-		}
-		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
-		return avail, nil
-	}
-	gpuInfo := GetGPUInfo()
-	if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
-		// leave 10% or 1024MiB of VRAM free per GPU to handle unaccounted for overhead
-		overhead := gpuInfo.FreeMemory / 10
-		gpus := uint64(gpuInfo.DeviceCount)
-		if overhead < gpus*1024*1024*1024 {
-			overhead = gpus * 1024 * 1024 * 1024
-		}
-		avail := int64(gpuInfo.FreeMemory - overhead)
-		slog.Debug(fmt.Sprintf("%s detected %d devices with %dM available memory", gpuInfo.Library, gpuInfo.DeviceCount, avail/1024/1024))
-		return avail, nil
-	}
-
-	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation
-}
-
 func FindGPULibs(baseLibName string, patterns []string) []string {
 	// Multiple GPU libraries may exist, and some may not work, so keep trying until we exhaust them
 	var ldPaths []string
 	gpuLibPaths := []string{}
-	slog.Info(fmt.Sprintf("Searching for GPU management library %s", baseLibName))
+	slog.Debug("Searching for GPU library", "name", baseLibName)

 	switch runtime.GOOS {
 	case "windows":
@ -206,7 +218,7 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 		}
 		patterns = append(patterns, filepath.Join(d, baseLibName+"*"))
 	}
-	slog.Debug(fmt.Sprintf("gpu management search paths: %v", patterns))
+	slog.Debug("gpu library search", "globs", patterns)
 	for _, pattern := range patterns {
 		// Ignore glob discovery errors
 		matches, _ := filepath.Glob(pattern)
@ -234,25 +246,25 @@ func FindGPULibs(baseLibName string, patterns []string) []string {
 			}
 		}
 	}
-	slog.Info(fmt.Sprintf("Discovered GPU libraries: %v", gpuLibPaths))
+	slog.Debug("discovered GPU libraries", "paths", gpuLibPaths)
 	return gpuLibPaths
 }

-func LoadCUDAMgmt(cudaLibPaths []string) *C.cuda_handle_t {
-	var resp C.cuda_init_resp_t
+func LoadCUDARTMgmt(cudartLibPaths []string) (int, *C.cudart_handle_t, string) {
+	var resp C.cudart_init_resp_t
 	resp.ch.verbose = getVerboseState()
-	for _, libPath := range cudaLibPaths {
+	for _, libPath := range cudartLibPaths {
 		lib := C.CString(libPath)
 		defer C.free(unsafe.Pointer(lib))
-		C.cuda_init(lib, &resp)
+		C.cudart_init(lib, &resp)
 		if resp.err != nil {
-			slog.Info(fmt.Sprintf("Unable to load CUDA management library %s: %s", libPath, C.GoString(resp.err)))
+			slog.Debug("Unable to load cudart", "library", libPath, "error", C.GoString(resp.err))
 			C.free(unsafe.Pointer(resp.err))
 		} else {
-			return &resp.ch
+			return int(resp.num_devices), &resp.ch, libPath
 		}
 	}
-	return nil
+	return 0, nil, ""
 }

 func getVerboseState() C.uint16_t {
@ -261,3 +273,22 @@ func getVerboseState() C.uint16_t {
 	}
 	return C.uint16_t(0)
 }
+
+// Given the list of GPUs this instantiation is targeted for,
+// figure out the visible devices environment variable
+//
+// If different libraries are detected, the first one is what we use
+func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
+	if len(l) == 0 {
+		return "", ""
+	}
+	switch l[0].Library {
+	case "cuda":
+		return cudaGetVisibleDevicesEnv(l)
+	case "rocm":
+		return rocmGetVisibleDevicesEnv(l)
+	default:
+		slog.Debug("no filter required for library " + l[0].Library)
+		return "", ""
+	}
+}
--- a/gpu/gpu_darwin.go
+++ b/gpu/gpu_darwin.go
@ -9,52 +9,47 @@ package gpu
 */
 import "C"
 import (
-	"fmt"
-	"log/slog"
-	"os"
 	"runtime"
-	"strconv"
+
+	"github.com/ollama/ollama/format"
 )

-// CheckVRAM returns the free VRAM in bytes on Linux machines with NVIDIA GPUs
-func CheckVRAM() (int64, error) {
-	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseInt(userLimit, 10, 64)
-		if err != nil {
-			return 0, fmt.Errorf("Invalid OLLAMA_MAX_VRAM setting %s: %s", userLimit, err)
-		}
-		slog.Info(fmt.Sprintf("user override OLLAMA_MAX_VRAM=%d", avail))
-		return avail, nil
-	}
+const (
+	metalMinimumMemory = 512 * format.MebiByte
+)

+func GetGPUInfo() GpuInfoList {
+	mem, _ := GetCPUMem()
 	if runtime.GOARCH == "amd64" {
-		// gpu not supported, this may not be metal
-		return 0, nil
-	}
-	recommendedMaxVRAM := int64(C.getRecommendedMaxVRAM())
-	return recommendedMaxVRAM, nil
-}
-
-func GetGPUInfo() GpuInfo {
-	mem, _ := getCPUMem()
-	if runtime.GOARCH == "amd64" {
-		return GpuInfo{
-			Library: "cpu",
-			Variant: GetCPUVariant(),
-			memInfo: mem,
+		return []GpuInfo{
+			{
+				Library: "cpu",
+				Variant: GetCPUVariant(),
+				memInfo: mem,
+			},
 		}
 	}
-	return GpuInfo{
+	info := GpuInfo{
 		Library: "metal",
-		memInfo: mem,
+		ID:      "0",
 	}
+	info.TotalMemory = uint64(C.getRecommendedMaxVRAM())
+
+	// TODO is there a way to gather actual allocated video memory? (currentAllocatedSize doesn't work)
+	info.FreeMemory = info.TotalMemory
+
+	info.MinimumMemory = metalMinimumMemory
+	return []GpuInfo{info}
 }

-func getCPUMem() (memInfo, error) {
+func GetCPUMem() (memInfo, error) {
 	return memInfo{
-		TotalMemory: 0,
+		TotalMemory: uint64(C.getPhysicalMemory()),
 		FreeMemory:  0,
-		DeviceCount: 0,
 	}, nil
 }
+
+func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
+	// No-op on darwin
+	return "", ""
+}
--- a/gpu/gpu_info.h
+++ b/gpu/gpu_info.h
@ -38,12 +38,17 @@
 extern "C" {
 #endif

+#define GPU_ID_LEN 64
+
 typedef struct mem_info {
+  char *err;  // If non-nill, caller responsible for freeing
+  char gpu_id[GPU_ID_LEN];
  uint64_t total;
  uint64_t free;
-  unsigned int count;
-  int igpu_index; // If >= 0, we detected an integrated GPU to ignore
-  char *err;  // If non-nill, caller responsible for freeing
+
+  // Compute Capability
+  int major; 
+  int minor;
 } mem_info_t;

 void cpu_check_ram(mem_info_t *resp);
@ -52,7 +57,7 @@ void cpu_check_ram(mem_info_t *resp);
 }
 #endif

-#include "gpu_info_cuda.h"
+#include "gpu_info_cudart.h"

 #endif  // __GPU_INFO_H__
 #endif  // __APPLE__
--- a/gpu/gpu_info_cpu.c
+++ b/gpu/gpu_info_cpu.c
@ -8,9 +8,11 @@ void cpu_check_ram(mem_info_t *resp) {
  MEMORYSTATUSEX info;
  info.dwLength = sizeof(info);
  if (GlobalMemoryStatusEx(&info) != 0) {
-    resp->count = 1;
    resp->total = info.ullTotalPhys;
    resp->free = info.ullAvailPhys;
+    resp->major = 0;
+    resp->minor = 0;
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
  } else {
    resp->err = LOAD_ERR();
  }
@ -27,9 +29,11 @@ void cpu_check_ram(mem_info_t *resp) {
  if (sysinfo(&info) != 0) {
    resp->err = strdup(strerror(errno));
  } else {
-    resp->count = 1;
    resp->total = info.totalram * info.mem_unit;
    resp->free = info.freeram * info.mem_unit;
+    resp->major = 0;
+    resp->minor = 0;
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "0");
  }
  return;
 }
--- a/gpu/gpu_info_cuda.c
+++ b/gpu/gpu_info_cuda.c
@ -1,214 +0,0 @@
-#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
-
-#include "gpu_info_cuda.h"
-
-#include <string.h>
-
-void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp) {
-  nvmlReturn_t ret;
-  resp->err = NULL;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
-      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
-      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
-      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
-      {"nvmlDeviceGetCount_v2", (void *)&resp->ch.nvmlDeviceGetCount_v2},
-      {"nvmlDeviceGetCudaComputeCapability", (void *)&resp->ch.nvmlDeviceGetCudaComputeCapability},
-      {"nvmlSystemGetDriverVersion", (void *)&resp->ch.nvmlSystemGetDriverVersion},
-      {"nvmlDeviceGetName", (void *)&resp->ch.nvmlDeviceGetName},
-      {"nvmlDeviceGetSerial", (void *)&resp->ch.nvmlDeviceGetSerial},
-      {"nvmlDeviceGetVbiosVersion", (void *)&resp->ch.nvmlDeviceGetVbiosVersion},
-      {"nvmlDeviceGetBoardPartNumber", (void *)&resp->ch.nvmlDeviceGetBoardPartNumber},
-      {"nvmlDeviceGetBrand", (void *)&resp->ch.nvmlDeviceGetBrand},
-      {NULL, NULL},
-  };
-
-  resp->ch.handle = LOAD_LIBRARY(cuda_lib_path, RTLD_LAZY);
-  if (!resp->ch.handle) {
-    char *msg = LOAD_ERR();
-    LOG(resp->ch.verbose, "library %s load err: %s\n", cuda_lib_path, msg);
-    snprintf(buf, buflen,
-             "Unable to load %s library to query for Nvidia GPUs: %s",
-             cuda_lib_path, msg);
-    free(msg);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  // TODO once we've squashed the remaining corner cases remove this log
-  LOG(resp->ch.verbose, "wiring nvidia management library functions in %s\n", cuda_lib_path);
-  
-  for (i = 0; l[i].s != NULL; i++) {
-    // TODO once we've squashed the remaining corner cases remove this log
-    LOG(resp->ch.verbose, "dlsym: %s\n", l[i].s);
-
-    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
-    if (!l[i].p) {
-      resp->ch.handle = NULL;
-      char *msg = LOAD_ERR();
-      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
-      UNLOAD_LIBRARY(resp->ch.handle);
-      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
-               msg);
-      free(msg);
-      resp->err = strdup(buf);
-      return;
-    }
-  }
-
-  ret = (*resp->ch.nvmlInit_v2)();
-  if (ret != NVML_SUCCESS) {
-    LOG(resp->ch.verbose, "nvmlInit_v2 err: %d\n", ret);
-    UNLOAD_LIBRARY(resp->ch.handle);
-    resp->ch.handle = NULL;
-    snprintf(buf, buflen, "nvml vram init failure: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  // Report driver version if we're in verbose mode, ignore errors
-  ret = (*resp->ch.nvmlSystemGetDriverVersion)(buf, buflen);
-  if (ret != NVML_SUCCESS) {
-    LOG(resp->ch.verbose, "nvmlSystemGetDriverVersion failed: %d\n", ret);
-  } else {
-    LOG(resp->ch.verbose, "CUDA driver version: %s\n", buf);
-  }
-}
-
-void cuda_check_vram(cuda_handle_t h, mem_info_t *resp) {
-  resp->err = NULL;
-  nvmlDevice_t device;
-  nvmlMemory_t memInfo = {0};
-  nvmlReturn_t ret;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  if (h.handle == NULL) {
-    resp->err = strdup("nvml handle sn't initialized");
-    return;
-  }
-
-  ret = (*h.nvmlDeviceGetCount_v2)(&resp->count);
-  if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "unable to get device count: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  resp->total = 0;
-  resp->free = 0;
-  for (i = 0; i < resp->count; i++) {
-    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
-    if (ret != NVML_SUCCESS) {
-      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
-      resp->err = strdup(buf);
-      return;
-    }
-
-    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
-    if (ret != NVML_SUCCESS) {
-      snprintf(buf, buflen, "device memory info lookup failure %d: %d", i, ret);
-      resp->err = strdup(buf);
-      return;
-    }
-    if (h.verbose) {
-      nvmlBrandType_t brand = 0;
-      // When in verbose mode, report more information about
-      // the card we discover, but don't fail on error
-      ret = (*h.nvmlDeviceGetName)(device, buf, buflen);
-      if (ret != NVML_SUCCESS) {
-        LOG(h.verbose, "nvmlDeviceGetName failed: %d\n", ret);
-      } else {
-        LOG(h.verbose, "[%d] CUDA device name: %s\n", i, buf);
-      }
-      ret = (*h.nvmlDeviceGetBoardPartNumber)(device, buf, buflen);
-      if (ret != NVML_SUCCESS) {
-        LOG(h.verbose, "nvmlDeviceGetBoardPartNumber failed: %d\n", ret);
-      } else {
-        LOG(h.verbose, "[%d] CUDA part number: %s\n", i, buf);
-      }
-      ret = (*h.nvmlDeviceGetSerial)(device, buf, buflen);
-      if (ret != NVML_SUCCESS) {
-        LOG(h.verbose, "nvmlDeviceGetSerial failed: %d\n", ret);
-      } else {
-        LOG(h.verbose, "[%d] CUDA S/N: %s\n", i, buf);
-      }
-      ret = (*h.nvmlDeviceGetVbiosVersion)(device, buf, buflen);
-      if (ret != NVML_SUCCESS) {
-        LOG(h.verbose, "nvmlDeviceGetVbiosVersion failed: %d\n", ret);
-      } else {
-        LOG(h.verbose, "[%d] CUDA vbios version: %s\n", i, buf);
-      }
-      ret = (*h.nvmlDeviceGetBrand)(device, &brand);
-      if (ret != NVML_SUCCESS) {
-        LOG(h.verbose, "nvmlDeviceGetBrand failed: %d\n", ret);
-      } else {
-        LOG(h.verbose, "[%d] CUDA brand: %d\n", i, brand);
-      }
-    }
-
-    LOG(h.verbose, "[%d] CUDA totalMem %ld\n", i, memInfo.total);
-    LOG(h.verbose, "[%d] CUDA usedMem %ld\n", i, memInfo.used);
-
-    resp->total += memInfo.total;
-    resp->free += memInfo.free;
-  }
-}
-
-void cuda_compute_capability(cuda_handle_t h, cuda_compute_capability_t *resp) {
-  resp->err = NULL;
-  resp->major = 0;
-  resp->minor = 0;
-  nvmlDevice_t device;
-  int major = 0;
-  int minor = 0;
-  nvmlReturn_t ret;
-  const int buflen = 256;
-  char buf[buflen + 1];
-  int i;
-
-  if (h.handle == NULL) {
-    resp->err = strdup("nvml handle not initialized");
-    return;
-  }
-
-  unsigned int devices;
-  ret = (*h.nvmlDeviceGetCount_v2)(&devices);
-  if (ret != NVML_SUCCESS) {
-    snprintf(buf, buflen, "unable to get device count: %d", ret);
-    resp->err = strdup(buf);
-    return;
-  }
-
-  for (i = 0; i < devices; i++) {
-    ret = (*h.nvmlDeviceGetHandleByIndex)(i, &device);
-    if (ret != NVML_SUCCESS) {
-      snprintf(buf, buflen, "unable to get device handle %d: %d", i, ret);
-      resp->err = strdup(buf);
-      return;
-    }
-
-    ret = (*h.nvmlDeviceGetCudaComputeCapability)(device, &major, &minor);
-    if (ret != NVML_SUCCESS) {
-      snprintf(buf, buflen, "device compute capability lookup failure %d: %d", i, ret);
-      resp->err = strdup(buf);
-      return;
-    }
-    // Report the lowest major.minor we detect as that limits our compatibility
-    if (resp->major == 0 || resp->major > major ) {
-      resp->major = major;
-      resp->minor = minor;
-    } else if ( resp->major == major && resp->minor > minor ) {
-      resp->minor = minor;
-    }
-  }
-}
-#endif  // __APPLE__
--- a/gpu/gpu_info_cuda.h
+++ b/gpu/gpu_info_cuda.h
@ -1,56 +0,0 @@
-#ifndef __APPLE__
-#ifndef __GPU_INFO_CUDA_H__
-#define __GPU_INFO_CUDA_H__
-#include "gpu_info.h"
-
-// Just enough typedef's to dlopen/dlsym for memory information
-typedef enum nvmlReturn_enum {
-  NVML_SUCCESS = 0,
-  // Other values omitted for now...
-} nvmlReturn_t;
-typedef void *nvmlDevice_t;  // Opaque is sufficient
-typedef struct nvmlMemory_st {
-  unsigned long long total;
-  unsigned long long free;
-  unsigned long long used;
-} nvmlMemory_t;
-
-typedef enum nvmlBrandType_enum
-{
-    NVML_BRAND_UNKNOWN          = 0,
-} nvmlBrandType_t;
-
-typedef struct cuda_handle {
-  void *handle;
-  uint16_t verbose;
-  nvmlReturn_t (*nvmlInit_v2)(void);
-  nvmlReturn_t (*nvmlShutdown)(void);
-  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
-  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
-  nvmlReturn_t (*nvmlDeviceGetCount_v2)(unsigned int *);
-  nvmlReturn_t (*nvmlDeviceGetCudaComputeCapability)(nvmlDevice_t, int* major, int* minor);
-  nvmlReturn_t (*nvmlSystemGetDriverVersion) (char* version, unsigned int  length);
-  nvmlReturn_t (*nvmlDeviceGetName) (nvmlDevice_t device, char* name, unsigned int  length);
-  nvmlReturn_t (*nvmlDeviceGetSerial) (nvmlDevice_t device, char* serial, unsigned int  length);
-  nvmlReturn_t (*nvmlDeviceGetVbiosVersion) (nvmlDevice_t device, char* version, unsigned int  length);
-  nvmlReturn_t (*nvmlDeviceGetBoardPartNumber) (nvmlDevice_t device, char* partNumber, unsigned int  length);
-  nvmlReturn_t (*nvmlDeviceGetBrand) (nvmlDevice_t device, nvmlBrandType_t* type);
-} cuda_handle_t;
-
-typedef struct cuda_init_resp {
-  char *err;  // If err is non-null handle is invalid
-  cuda_handle_t ch;
-} cuda_init_resp_t;
-
-typedef struct cuda_compute_capability {
-  char *err;
-  int major;
-  int minor;
-} cuda_compute_capability_t;
-
-void cuda_init(char *cuda_lib_path, cuda_init_resp_t *resp);
-void cuda_check_vram(cuda_handle_t ch, mem_info_t *resp);
-void cuda_compute_capability(cuda_handle_t ch, cuda_compute_capability_t *cc);
-
-#endif  // __GPU_INFO_CUDA_H__
-#endif  // __APPLE__
--- a/gpu/gpu_info_cudart.c
+++ b/gpu/gpu_info_cudart.c
@ -0,0 +1,181 @@
+#ifndef __APPLE__  // TODO - maybe consider nvidia support on intel macs?
+
+#include <string.h>
+#include "gpu_info_cudart.h"
+
+void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp) {
+  cudartReturn_t ret;
+  resp->err = NULL;
+  resp->num_devices = 0;
+  const int buflen = 256;
+  char buf[buflen + 1];
+  int i;
+
+  struct lookup {
+    char *s;
+    void **p;
+  } l[] = {
+      {"cudaSetDevice", (void *)&resp->ch.cudaSetDevice},
+      {"cudaDeviceSynchronize", (void *)&resp->ch.cudaDeviceSynchronize},
+      {"cudaDeviceReset", (void *)&resp->ch.cudaDeviceReset},
+      {"cudaMemGetInfo", (void *)&resp->ch.cudaMemGetInfo},
+      {"cudaGetDeviceCount", (void *)&resp->ch.cudaGetDeviceCount},
+      {"cudaDeviceGetAttribute", (void *)&resp->ch.cudaDeviceGetAttribute},
+      {"cudaDriverGetVersion", (void *)&resp->ch.cudaDriverGetVersion},
+      {"cudaGetDeviceProperties", (void *)&resp->ch.cudaGetDeviceProperties},
+      {NULL, NULL},
+  };
+
+  resp->ch.handle = LOAD_LIBRARY(cudart_lib_path, RTLD_LAZY);
+  if (!resp->ch.handle) {
+    char *msg = LOAD_ERR();
+    LOG(resp->ch.verbose, "library %s load err: %s\n", cudart_lib_path, msg);
+    snprintf(buf, buflen,
+            "Unable to load %s library to query for Nvidia GPUs: %s",
+            cudart_lib_path, msg);
+    free(msg);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  for (i = 0; l[i].s != NULL; i++) {
+    *l[i].p = LOAD_SYMBOL(resp->ch.handle, l[i].s);
+    if (!l[i].p) {
+      char *msg = LOAD_ERR();
+      LOG(resp->ch.verbose, "dlerr: %s\n", msg);
+      UNLOAD_LIBRARY(resp->ch.handle);
+      resp->ch.handle = NULL;
+      snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
+              msg);
+      free(msg);
+      resp->err = strdup(buf);
+      return;
+    }
+  }
+
+  ret = (*resp->ch.cudaSetDevice)(0);
+  if (ret != CUDART_SUCCESS) {
+    LOG(resp->ch.verbose, "cudaSetDevice err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
+      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
+      return;
+    }
+    snprintf(buf, buflen, "cudart init failure: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  int version = 0;
+  cudartDriverVersion_t driverVersion;
+  driverVersion.major = 0;
+  driverVersion.minor = 0;
+
+  // Report driver version if we're in verbose mode, ignore errors
+  ret = (*resp->ch.cudaDriverGetVersion)(&version);
+  if (ret != CUDART_SUCCESS) {
+    LOG(resp->ch.verbose, "cudaDriverGetVersion failed: %d\n", ret);
+  } else {
+    driverVersion.major = version / 1000;
+    driverVersion.minor = (version - (driverVersion.major * 1000)) / 10;
+    LOG(resp->ch.verbose, "CUDA driver version: %d-%d\n", driverVersion.major, driverVersion.minor);
+  }
+
+  ret = (*resp->ch.cudaGetDeviceCount)(&resp->num_devices);
+  if (ret != CUDART_SUCCESS) {
+    LOG(resp->ch.verbose, "cudaGetDeviceCount err: %d\n", ret);
+    UNLOAD_LIBRARY(resp->ch.handle);
+    resp->ch.handle = NULL;
+    snprintf(buf, buflen, "unable to get device count: %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+}
+
+
+void cudart_check_vram(cudart_handle_t h, int i, mem_info_t *resp) {
+  resp->err = NULL;
+  cudartMemory_t memInfo = {0,0,0};
+  cudartReturn_t ret;
+  const int buflen = 256;
+  char buf[buflen + 1];
+
+  if (h.handle == NULL) {
+    resp->err = strdup("cudart handle isn't initialized");
+    return;
+  }
+
+  ret = (*h.cudaSetDevice)(i);
+  if (ret != CUDART_SUCCESS) {
+    snprintf(buf, buflen, "cudart device failed to initialize");
+    resp->err = strdup(buf);
+    return;
+  }
+
+  cudaDeviceProp_t props;
+  ret = (*h.cudaGetDeviceProperties)(&props, i);
+  if (ret != CUDART_SUCCESS) {
+    LOG(h.verbose, "[%d] device properties lookup failure: %d\n", i, ret);
+    snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
+    resp->major = 0;
+    resp->minor = 0;
+  } else {
+    int allNull = 1;
+    for (int j = 0; j < 16; j++) {
+      if (props.uuid.bytes[j] != 0) {
+        allNull = 0;
+        break;
+      }
+    }
+    if (allNull != 0) {
+      snprintf(&resp->gpu_id[0], GPU_ID_LEN, "%d", i);
+    } else {
+      // GPU-d110a105-ac29-1d54-7b49-9c90440f215b
+      snprintf(&resp->gpu_id[0], GPU_ID_LEN,
+          "GPU-%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+          props.uuid.bytes[0],
+          props.uuid.bytes[1],
+          props.uuid.bytes[2],
+          props.uuid.bytes[3],
+          props.uuid.bytes[4],
+          props.uuid.bytes[5],
+          props.uuid.bytes[6],
+          props.uuid.bytes[7],
+          props.uuid.bytes[8],
+          props.uuid.bytes[9],
+          props.uuid.bytes[10],
+          props.uuid.bytes[11],
+          props.uuid.bytes[12],
+          props.uuid.bytes[13],
+          props.uuid.bytes[14],
+          props.uuid.bytes[15]
+        );
+    }
+    resp->major = props.major;
+    resp->minor = props.minor;
+
+    // TODO add other useful properties from props
+  }
+  ret = (*h.cudaMemGetInfo)(&memInfo.free, &memInfo.total);
+  if (ret != CUDART_SUCCESS) {
+    snprintf(buf, buflen, "cudart device memory info lookup failure %d", ret);
+    resp->err = strdup(buf);
+    return;
+  }
+
+  resp->total = memInfo.total;
+  resp->free = memInfo.free;
+
+  LOG(h.verbose, "[%s] CUDA totalMem %lu\n", resp->gpu_id, resp->total);
+  LOG(h.verbose, "[%s] CUDA freeMem %lu\n", resp->gpu_id, resp->free);
+  LOG(h.verbose, "[%s] Compute Capability %d.%d\n", resp->gpu_id, resp->major, resp->minor);
+}
+
+void cudart_release(cudart_handle_t h) {
+  LOG(h.verbose, "releasing cudart library\n");
+  UNLOAD_LIBRARY(h.handle);
+  h.handle = NULL;
+}
+
+#endif  // __APPLE__
--- a/gpu/gpu_info_cudart.h
+++ b/gpu/gpu_info_cudart.h
@ -0,0 +1,147 @@
+#ifndef __APPLE__
+#ifndef __GPU_INFO_CUDART_H__
+#define __GPU_INFO_CUDART_H__
+#include "gpu_info.h"
+
+// Just enough typedef's to dlopen/dlsym for memory information
+typedef enum cudartReturn_enum {
+  CUDART_SUCCESS = 0,
+  CUDA_ERROR_INVALID_VALUE = 1,
+  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  // Other values omitted for now...
+} cudartReturn_t;
+
+typedef enum cudartDeviceAttr_enum {
+  cudartDevAttrComputeCapabilityMajor = 75,
+  cudartDevAttrComputeCapabilityMinor = 76,
+
+  // TODO - not yet wired up but may be useful for Jetson or other
+  // integrated GPU scenarios with shared memory
+  cudaDevAttrIntegrated = 18
+
+} cudartDeviceAttr_t;
+
+typedef void *cudartDevice_t;  // Opaque is sufficient
+typedef struct cudartMemory_st {
+  size_t total;
+  size_t free;
+  size_t used;
+} cudartMemory_t;
+
+typedef struct cudartDriverVersion {
+  int major;
+  int minor;
+} cudartDriverVersion_t;
+
+typedef struct cudaUUID {
+    unsigned char bytes[16];
+} cudaUUID_t;
+typedef struct cudaDeviceProp {
+    char         name[256];                  /**< ASCII string identifying device */
+    cudaUUID_t   uuid;                       /**< 16-byte unique identifier */
+    char         luid[8];                    /**< 8-byte locally unique identifier. Value is undefined on TCC and non-Windows platforms */
+    unsigned int luidDeviceNodeMask;         /**< LUID device node mask. Value is undefined on TCC and non-Windows platforms */
+    size_t       totalGlobalMem;             /**< Global memory available on device in bytes */
+    size_t       sharedMemPerBlock;          /**< Shared memory available per block in bytes */
+    int          regsPerBlock;               /**< 32-bit registers available per block */
+    int          warpSize;                   /**< Warp size in threads */
+    size_t       memPitch;                   /**< Maximum pitch in bytes allowed by memory copies */
+    int          maxThreadsPerBlock;         /**< Maximum number of threads per block */
+    int          maxThreadsDim[3];           /**< Maximum size of each dimension of a block */
+    int          maxGridSize[3];             /**< Maximum size of each dimension of a grid */
+    int          clockRate;                  /**< Clock frequency in kilohertz */
+    size_t       totalConstMem;              /**< Constant memory available on device in bytes */
+    int          major;                      /**< Major compute capability */
+    int          minor;                      /**< Minor compute capability */
+    size_t       textureAlignment;           /**< Alignment requirement for textures */
+    size_t       texturePitchAlignment;      /**< Pitch alignment requirement for texture references bound to pitched memory */
+    int          deviceOverlap;              /**< Device can concurrently copy memory and execute a kernel. Deprecated. Use instead asyncEngineCount. */
+    int          multiProcessorCount;        /**< Number of multiprocessors on device */
+    int          kernelExecTimeoutEnabled;   /**< Specified whether there is a run time limit on kernels */
+    int          integrated;                 /**< Device is integrated as opposed to discrete */
+    int          canMapHostMemory;           /**< Device can map host memory with cudaHostAlloc/cudaHostGetDevicePointer */
+    int          computeMode;                /**< Compute mode (See ::cudaComputeMode) */
+    int          maxTexture1D;               /**< Maximum 1D texture size */
+    int          maxTexture1DMipmap;         /**< Maximum 1D mipmapped texture size */
+    int          maxTexture1DLinear;         /**< Deprecated, do not use. Use cudaDeviceGetTexture1DLinearMaxWidth() or cuDeviceGetTexture1DLinearMaxWidth() instead. */
+    int          maxTexture2D[2];            /**< Maximum 2D texture dimensions */
+    int          maxTexture2DMipmap[2];      /**< Maximum 2D mipmapped texture dimensions */
+    int          maxTexture2DLinear[3];      /**< Maximum dimensions (width, height, pitch) for 2D textures bound to pitched memory */
+    int          maxTexture2DGather[2];      /**< Maximum 2D texture dimensions if texture gather operations have to be performed */
+    int          maxTexture3D[3];            /**< Maximum 3D texture dimensions */
+    int          maxTexture3DAlt[3];         /**< Maximum alternate 3D texture dimensions */
+    int          maxTextureCubemap;          /**< Maximum Cubemap texture dimensions */
+    int          maxTexture1DLayered[2];     /**< Maximum 1D layered texture dimensions */
+    int          maxTexture2DLayered[3];     /**< Maximum 2D layered texture dimensions */
+    int          maxTextureCubemapLayered[2];/**< Maximum Cubemap layered texture dimensions */
+    int          maxSurface1D;               /**< Maximum 1D surface size */
+    int          maxSurface2D[2];            /**< Maximum 2D surface dimensions */
+    int          maxSurface3D[3];            /**< Maximum 3D surface dimensions */
+    int          maxSurface1DLayered[2];     /**< Maximum 1D layered surface dimensions */
+    int          maxSurface2DLayered[3];     /**< Maximum 2D layered surface dimensions */
+    int          maxSurfaceCubemap;          /**< Maximum Cubemap surface dimensions */
+    int          maxSurfaceCubemapLayered[2];/**< Maximum Cubemap layered surface dimensions */
+    size_t       surfaceAlignment;           /**< Alignment requirements for surfaces */
+    int          concurrentKernels;          /**< Device can possibly execute multiple kernels concurrently */
+    int          ECCEnabled;                 /**< Device has ECC support enabled */
+    int          pciBusID;                   /**< PCI bus ID of the device */
+    int          pciDeviceID;                /**< PCI device ID of the device */
+    int          pciDomainID;                /**< PCI domain ID of the device */
+    int          tccDriver;                  /**< 1 if device is a Tesla device using TCC driver, 0 otherwise */
+    int          asyncEngineCount;           /**< Number of asynchronous engines */
+    int          unifiedAddressing;          /**< Device shares a unified address space with the host */
+    int          memoryClockRate;            /**< Peak memory clock frequency in kilohertz */
+    int          memoryBusWidth;             /**< Global memory bus width in bits */
+    int          l2CacheSize;                /**< Size of L2 cache in bytes */
+    int          persistingL2CacheMaxSize;   /**< Device's maximum l2 persisting lines capacity setting in bytes */
+    int          maxThreadsPerMultiProcessor;/**< Maximum resident threads per multiprocessor */
+    int          streamPrioritiesSupported;  /**< Device supports stream priorities */
+    int          globalL1CacheSupported;     /**< Device supports caching globals in L1 */
+    int          localL1CacheSupported;      /**< Device supports caching locals in L1 */
+    size_t       sharedMemPerMultiprocessor; /**< Shared memory available per multiprocessor in bytes */
+    int          regsPerMultiprocessor;      /**< 32-bit registers available per multiprocessor */
+    int          managedMemory;              /**< Device supports allocating managed memory on this system */
+    int          isMultiGpuBoard;            /**< Device is on a multi-GPU board */
+    int          multiGpuBoardGroupID;       /**< Unique identifier for a group of devices on the same multi-GPU board */
+    int          hostNativeAtomicSupported;  /**< Link between the device and the host supports native atomic operations */
+    int          singleToDoublePrecisionPerfRatio; /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */
+    int          pageableMemoryAccess;       /**< Device supports coherently accessing pageable memory without calling cudaHostRegister on it */
+    int          concurrentManagedAccess;    /**< Device can coherently access managed memory concurrently with the CPU */
+    int          computePreemptionSupported; /**< Device supports Compute Preemption */
+    int          canUseHostPointerForRegisteredMem; /**< Device can access host registered memory at the same virtual address as the CPU */
+    int          cooperativeLaunch;          /**< Device supports launching cooperative kernels via ::cudaLaunchCooperativeKernel */
+    int          cooperativeMultiDeviceLaunch; /**< Deprecated, cudaLaunchCooperativeKernelMultiDevice is deprecated. */
+    size_t       sharedMemPerBlockOptin;     /**< Per device maximum shared memory per block usable by special opt in */
+    int          pageableMemoryAccessUsesHostPageTables; /**< Device accesses pageable memory via the host's page tables */
+    int          directManagedMemAccessFromHost; /**< Host can directly access managed memory on the device without migration. */
+    int          maxBlocksPerMultiProcessor; /**< Maximum number of resident blocks per multiprocessor */
+    int          accessPolicyMaxWindowSize;  /**< The maximum value of ::cudaAccessPolicyWindow::num_bytes. */
+    size_t       reservedSharedMemPerBlock;  /**< Shared memory reserved by CUDA driver per block in bytes */
+  } cudaDeviceProp_t;
+
+typedef struct cudart_handle {
+  void *handle;
+  uint16_t verbose;
+  cudartReturn_t (*cudaSetDevice)(int device);
+  cudartReturn_t (*cudaDeviceSynchronize)(void);
+  cudartReturn_t (*cudaDeviceReset)(void);
+  cudartReturn_t (*cudaMemGetInfo)(size_t *, size_t *);
+  cudartReturn_t (*cudaGetDeviceCount)(int *);
+  cudartReturn_t (*cudaDeviceGetAttribute)(int* value, cudartDeviceAttr_t attr, int device);
+  cudartReturn_t (*cudaDriverGetVersion) (int *driverVersion);
+  cudartReturn_t (*cudaGetDeviceProperties) (cudaDeviceProp_t* prop, int device);
+} cudart_handle_t;
+
+typedef struct cudart_init_resp {
+  char *err;  // If err is non-null handle is invalid
+  cudart_handle_t ch;
+  int num_devices;
+} cudart_init_resp_t;
+
+void cudart_init(char *cudart_lib_path, cudart_init_resp_t *resp);
+void cudart_check_vram(cudart_handle_t ch, int device_id, mem_info_t *resp);
+void cudart_release(cudart_handle_t ch);
+
+#endif  // __GPU_INFO_CUDART_H__
+#endif  // __APPLE__
--- a/gpu/gpu_info_darwin.h
+++ b/gpu/gpu_info_darwin.h
@ -1,3 +1,4 @@
 #import <Metal/Metal.h>
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
+uint64_t getPhysicalMemory();
--- a/gpu/gpu_info_darwin.m
+++ b/gpu/gpu_info_darwin.m
@ -1,11 +1,13 @@
-//go:build darwin
+// go:build darwin
 #include "gpu_info_darwin.h"

-uint64_t getRecommendedMaxVRAM()
-{
-	id<MTLDevice> device = MTLCreateSystemDefaultDevice();
-	uint64_t result = device.recommendedMaxWorkingSetSize;
-	CFRelease(device);
-	return result;
+uint64_t getRecommendedMaxVRAM() {
+  id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+  uint64_t result = device.recommendedMaxWorkingSetSize;
+  CFRelease(device);
+  return result;
 }

+uint64_t getPhysicalMemory() {
+  return [[NSProcessInfo processInfo] physicalMemory];
+}
--- a/gpu/gpu_test.go
+++ b/gpu/gpu_test.go
@ -9,23 +9,16 @@ import (

 func TestBasicGetGPUInfo(t *testing.T) {
 	info := GetGPUInfo()
-	assert.Contains(t, "cuda rocm cpu metal", info.Library)
-
-	switch runtime.GOOS {
-	case "darwin":
-		// TODO - remove this once MacOS returns some size for CPU
-		return
-	case "linux", "windows":
-		assert.Greater(t, info.TotalMemory, uint64(0))
-		assert.Greater(t, info.FreeMemory, uint64(0))
-		assert.Greater(t, info.DeviceCount, uint32(0))
-	default:
-		return
+	assert.Greater(t, len(info), 0)
+	assert.Contains(t, "cuda rocm cpu metal", info[0].Library)
+	if info[0].Library != "cpu" {
+		assert.Greater(t, info[0].TotalMemory, uint64(0))
+		assert.Greater(t, info[0].FreeMemory, uint64(0))
 	}
 }

 func TestCPUMemInfo(t *testing.T) {
-	info, err := getCPUMem()
+	info, err := GetCPUMem()
 	assert.NoError(t, err)
 	switch runtime.GOOS {
 	case "darwin":
--- a/gpu/types.go
+++ b/gpu/types.go
@ -3,7 +3,6 @@ package gpu
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
-	DeviceCount uint32 `json:"device_count,omitempty"`
 }

 // Beginning of an `ollama info` command
@ -14,11 +13,52 @@ type GpuInfo struct {
 	// Optional variant to select (e.g. versions, cpu feature flags)
 	Variant string `json:"variant,omitempty"`

-	// TODO add other useful attributes about the card here for discovery information
+	// MinimumMemory represents the minimum memory required to use the GPU
+	MinimumMemory uint64 `json:"-"`
+
+	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
+	DependencyPath string `json:"lib_path,omitempty"`
+
+	// GPU information
+	ID    string `json:"gpu_id"`          // string to use for selection of this specific GPU
+	Name  string `json:"name"`            // user friendly name if available
+	Major int    `json:"major,omitempty"` // Major compatibility version (CC or gfx)
+	Minor int    `json:"minor,omitempty"` // Minor compatibility version (CC or gfx)
+	Patch int    `json:"patch,omitempty"` // Patch compatibility only matters on AMD
+
+	// TODO other performance capability info to help in scheduling decisions
 }

-type Version struct {
-	Major uint
-	Minor uint
-	Patch uint
+type GpuInfoList []GpuInfo
+
+// Split up the set of gpu info's by Library and variant
+func (l GpuInfoList) ByLibrary() []GpuInfoList {
+	resp := []GpuInfoList{}
+	libs := []string{}
+	for _, info := range l {
+		found := false
+		requested := info.Library
+		if info.Variant != "" {
+			requested += "_" + info.Variant
+		}
+		for i, lib := range libs {
+			if lib == requested {
+				resp[i] = append(resp[i], info)
+				found = true
+				break
+			}
+		}
+		if !found {
+			libs = append(libs, info.Library)
+			resp = append(resp, []GpuInfo{info})
+		}
+	}
+	return resp
 }
+
+// Sort by Free Space
+type ByFreeMemory []GpuInfo
+
+func (a ByFreeMemory) Len() int           { return len(a) }
+func (a ByFreeMemory) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
+func (a ByFreeMemory) Less(i, j int) bool { return a[i].FreeMemory < a[j].FreeMemory }
--- a/integration/README.md
+++ b/integration/README.md
@ -0,0 +1,11 @@
+# Integration Tests
+
+This directory contains integration tests to exercise Ollama end-to-end to verify behavior
+
+By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...`
+
+
+The integration tests have 2 modes of operating.
+
+1. By default, they will start the server on a random port, run the tests, and then shutdown the server.
+2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote
--- a/integration/basic_test.go
+++ b/integration/basic_test.go
@ -0,0 +1,70 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/require"
+)
+
+func TestOrcaMiniBlueSky(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "orca-mini",
+		Prompt: "why is the sky blue?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+}
+
+func TestUnicodeModelDir(t *testing.T) {
+	// This is only useful for Windows with utf-16 characters, so skip this test for other platforms
+	if runtime.GOOS != "windows" {
+		t.Skip("Unicode test only applicable to windows")
+	}
+	// Only works for local testing
+	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
+		t.Skip("TestUnicodeModelDir only works for local testing, skipping")
+	}
+
+	modelDir, err := os.MkdirTemp("", "ollama_埃")
+	require.NoError(t, err)
+	defer os.RemoveAll(modelDir)
+	slog.Info("unicode", "OLLAMA_MODELS", modelDir)
+
+	oldModelsDir := os.Getenv("OLLAMA_MODELS")
+	if oldModelsDir == "" {
+		defer os.Unsetenv("OLLAMA_MODELS")
+	} else {
+		defer os.Setenv("OLLAMA_MODELS", oldModelsDir)
+	}
+	err = os.Setenv("OLLAMA_MODELS", modelDir)
+	require.NoError(t, err)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	req := api.GenerateRequest{
+		Model:  "orca-mini",
+		Prompt: "why is the sky blue?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
+}
--- a/integration/concurrency_test.go
+++ b/integration/concurrency_test.go
@ -0,0 +1,225 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"strconv"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/require"
+)
+
+func TestMultiModelConcurrency(t *testing.T) {
+	var (
+		req = [2]api.GenerateRequest{
+			{
+				Model:  "orca-mini",
+				Prompt: "why is the ocean blue?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:  "tinydolphin",
+				Prompt: "what is the origin of the us thanksgiving holiday?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			},
+		}
+		resp = [2][]string{
+			[]string{"sunlight"},
+			[]string{"england", "english", "massachusetts", "pilgrims"},
+		}
+	)
+	var wg sync.WaitGroup
+	wg.Add(len(req))
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	for i := 0; i < len(req); i++ {
+		go func(i int) {
+			defer wg.Done()
+			GenerateTestHelper(ctx, t, req[i], resp[i])
+		}(i)
+	}
+	wg.Wait()
+}
+
+func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute) // GTX 750 2G card takes ~9 minutes
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	req, resp := GenerateRequests()
+	// Get the server running (if applicable) warm the model up with a single initial request
+	DoGenerate(ctx, t, client, req[0], resp[0], 60*time.Second, 5*time.Second)
+
+	var wg sync.WaitGroup
+	wg.Add(len(req))
+	for i := 0; i < len(req); i++ {
+		go func(i int) {
+			defer wg.Done()
+			for j := 0; j < 5; j++ {
+				slog.Info("Starting", "req", i, "iter", j)
+				// On slower GPUs it can take a while to process the 4 concurrent requests
+				// so we allow a much longer initial timeout
+				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
+			}
+		}(i)
+	}
+	wg.Wait()
+}
+
+// Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
+func TestMultiModelStress(t *testing.T) {
+	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	if vram == "" {
+		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
+	}
+	max, err := strconv.ParseUint(vram, 10, 64)
+	require.NoError(t, err)
+	const MB = uint64(1024 * 1024)
+	type model struct {
+		name string
+		size uint64 // Approximate amount of VRAM they typically use when fully loaded in VRAM
+	}
+
+	smallModels := []model{
+		{
+			name: "orca-mini",
+			size: 2992 * MB,
+		},
+		{
+			name: "phi",
+			size: 2616 * MB,
+		},
+		{
+			name: "gemma:2b",
+			size: 2364 * MB,
+		},
+		{
+			name: "stable-code:3b",
+			size: 2608 * MB,
+		},
+		{
+			name: "starcoder2:3b",
+			size: 2166 * MB,
+		},
+	}
+	mediumModels := []model{
+		{
+			name: "llama2",
+			size: 5118 * MB,
+		},
+		{
+			name: "mistral",
+			size: 4620 * MB,
+		},
+		{
+			name: "orca-mini:7b",
+			size: 5118 * MB,
+		},
+		{
+			name: "dolphin-mistral",
+			size: 4620 * MB,
+		},
+		{
+			name: "gemma:7b",
+			size: 5000 * MB,
+		},
+		// TODO - uncomment this once #3565 is merged and this is rebased on it
+		// {
+		// 	name: "codellama:7b",
+		// 	size: 5118 * MB,
+		// },
+	}
+
+	// These seem to be too slow to be useful...
+	// largeModels := []model{
+	// 	{
+	// 		name: "llama2:13b",
+	// 		size: 7400 * MB,
+	// 	},
+	// 	{
+	// 		name: "codellama:13b",
+	// 		size: 7400 * MB,
+	// 	},
+	// 	{
+	// 		name: "orca-mini:13b",
+	// 		size: 7400 * MB,
+	// 	},
+	// 	{
+	// 		name: "gemma:7b",
+	// 		size: 5000 * MB,
+	// 	},
+	// 	{
+	// 		name: "starcoder2:15b",
+	// 		size: 9100 * MB,
+	// 	},
+	// }
+
+	var chosenModels []model
+	switch {
+	case max < 10000*MB:
+		slog.Info("selecting small models")
+		chosenModels = smallModels
+	// case max < 30000*MB:
+	default:
+		slog.Info("selecting medium models")
+		chosenModels = mediumModels
+		// default:
+		// 	slog.Info("selecting large models")
+		// 	chosenModels = largModels
+	}
+
+	req, resp := GenerateRequests()
+
+	for i := range req {
+		if i > len(chosenModels) {
+			break
+		}
+		req[i].Model = chosenModels[i].name
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute) // TODO baseline -- 10m too short
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Make sure all the models are pulled before we get started
+	for _, r := range req {
+		require.NoError(t, PullIfMissing(ctx, client, r.Model))
+	}
+
+	var wg sync.WaitGroup
+	consumed := uint64(256 * MB) // Assume some baseline usage
+	for i := 0; i < len(req); i++ {
+		// Always get at least 2 models, but dont' overshoot VRAM too much or we'll take too long
+		if i > 1 && consumed > max {
+			slog.Info("achieved target vram exhaustion", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+			break
+		}
+		consumed += chosenModels[i].size
+		slog.Info("target vram", "count", i, "vramMB", max/1024/1024, "modelsMB", consumed/1024/1024)
+
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			for j := 0; j < 3; j++ {
+				slog.Info("Starting", "req", i, "iter", j, "model", req[i].Model)
+				DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 5*time.Second)
+			}
+		}(i)
+	}
+	wg.Wait()
+}
--- a/integration/context_test.go
+++ b/integration/context_test.go
@ -0,0 +1,28 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestContextExhaustion(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute) // TODO maybe shorter?
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "llama2",
+		Prompt: "Write me a story with a ton of emojis?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+			"num_ctx":     128,
+		},
+	}
+	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+}
--- a/integration/llm_image_test.go
+++ b/integration/llm_image_test.go
@ -1,49 +1,38 @@
 //go:build integration

-package server
+package integration

 import (
 	"context"
 	"encoding/base64"
-	"log"
-	"os"
-	"strings"
 	"testing"
 	"time"

-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/llm"
-	"github.com/stretchr/testify/assert"
+	"github.com/ollama/ollama/api"
 	"github.com/stretchr/testify/require"
 )

 func TestIntegrationMultimodal(t *testing.T) {
-	SkipIFNoTestData(t)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{
-		Model:   "llava:7b",
-		Prompt:  "what does the text in this image say?",
-		Options: map[string]interface{}{},
+		Model:  "llava:7b",
+		Prompt: "what does the text in this image say?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"seed":        42,
+			"temperature": 0.0,
+		},
 		Images: []api.ImageData{
 			image,
 		},
 	}
-	resp := "the ollamas"
-	workDir, err := os.MkdirTemp("", "ollama")
-	require.NoError(t, err)
-	defer os.RemoveAll(workDir)
-	require.NoError(t, llm.Init(workDir))
-	ctx, cancel := context.WithTimeout(context.Background(), time.Second*60)
+
+	// Note: sometimes it returns "the ollamas" sometimes "the ollams"
+	resp := "the ollam"
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
-	opts := api.DefaultOptions()
-	opts.Seed = 42
-	opts.Temperature = 0.0
-	model, llmRunner := PrepareModelForPrompts(t, req.Model, opts)
-	defer llmRunner.Close()
-	response := OneShotPromptResponse(t, ctx, req, model, llmRunner)
-	log.Print(response)
-	assert.Contains(t, strings.ToLower(response), resp)
+	GenerateTestHelper(ctx, t, req, []string{resp})
 }

 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb
--- a/integration/llm_test.go
+++ b/integration/llm_test.go
@ -0,0 +1,47 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// TODO - this would ideally be in the llm package, but that would require some refactoring of interfaces in the server
+//        package to avoid circular dependencies
+
+var (
+	stream = false
+	req    = [2]api.GenerateRequest{
+		{
+			Model:  "orca-mini",
+			Prompt: "why is the ocean blue?",
+			Stream: &stream,
+			Options: map[string]interface{}{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		}, {
+			Model:  "orca-mini",
+			Prompt: "what is the origin of the us thanksgiving holiday?",
+			Stream: &stream,
+			Options: map[string]interface{}{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		},
+	}
+	resp = [2][]string{
+		[]string{"sunlight"},
+		[]string{"england", "english", "massachusetts", "pilgrims"},
+	}
+)
+
+func TestIntegrationSimpleOrcaMini(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	GenerateTestHelper(ctx, t, req[0], resp[0])
+}
--- a/integration/utils_test.go
+++ b/integration/utils_test.go
@ -0,0 +1,338 @@
+//go:build integration
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"math/rand"
+	"net"
+	"net/http"
+	"net/url"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/app/lifecycle"
+	"github.com/stretchr/testify/require"
+)
+
+func Init() {
+	lifecycle.InitLogging()
+}
+
+func FindPort() string {
+	port := 0
+	if a, err := net.ResolveTCPAddr("tcp", "localhost:0"); err == nil {
+		var l *net.TCPListener
+		if l, err = net.ListenTCP("tcp", a); err == nil {
+			port = l.Addr().(*net.TCPAddr).Port
+			l.Close()
+		}
+	}
+	if port == 0 {
+		port = rand.Intn(65535-49152) + 49152 // get a random port in the ephemeral range
+	}
+	return strconv.Itoa(port)
+}
+
+func GetTestEndpoint() (*api.Client, string) {
+	defaultPort := "11434"
+	ollamaHost := os.Getenv("OLLAMA_HOST")
+
+	scheme, hostport, ok := strings.Cut(ollamaHost, "://")
+	if !ok {
+		scheme, hostport = "http", ollamaHost
+	}
+
+	// trim trailing slashes
+	hostport = strings.TrimRight(hostport, "/")
+
+	host, port, err := net.SplitHostPort(hostport)
+	if err != nil {
+		host, port = "127.0.0.1", defaultPort
+		if ip := net.ParseIP(strings.Trim(hostport, "[]")); ip != nil {
+			host = ip.String()
+		} else if hostport != "" {
+			host = hostport
+		}
+	}
+
+	if os.Getenv("OLLAMA_TEST_EXISTING") == "" && port == defaultPort {
+		port = FindPort()
+	}
+
+	slog.Info("server connection", "host", host, "port", port)
+
+	return api.NewClient(
+		&url.URL{
+			Scheme: scheme,
+			Host:   net.JoinHostPort(host, port),
+		},
+		http.DefaultClient), fmt.Sprintf("%s:%s", host, port)
+}
+
+var serverMutex sync.Mutex
+var serverReady bool
+
+func startServer(ctx context.Context, ollamaHost string) error {
+	// Make sure the server has been built
+	CLIName, err := filepath.Abs("../ollama")
+	if err != nil {
+		return err
+	}
+
+	if runtime.GOOS == "windows" {
+		CLIName += ".exe"
+	}
+	_, err = os.Stat(CLIName)
+	if err != nil {
+		return fmt.Errorf("CLI missing, did you forget to build first?  %w", err)
+	}
+	serverMutex.Lock()
+	defer serverMutex.Unlock()
+	if serverReady {
+		return nil
+	}
+
+	if tmp := os.Getenv("OLLAMA_HOST"); tmp != ollamaHost {
+		slog.Info("setting env", "OLLAMA_HOST", ollamaHost)
+		os.Setenv("OLLAMA_HOST", ollamaHost)
+	}
+
+	slog.Info("starting server", "url", ollamaHost)
+	done, err := lifecycle.SpawnServer(ctx, "../ollama")
+	if err != nil {
+		return fmt.Errorf("failed to start server: %w", err)
+	}
+
+	go func() {
+		<-ctx.Done()
+		serverMutex.Lock()
+		defer serverMutex.Unlock()
+		exitCode := <-done
+		if exitCode > 0 {
+			slog.Warn("server failure", "exit", exitCode)
+		}
+		serverReady = false
+	}()
+
+	// TODO wait only long enough for the server to be responsive...
+	time.Sleep(500 * time.Millisecond)
+
+	serverReady = true
+	return nil
+}
+
+func PullIfMissing(ctx context.Context, client *api.Client, modelName string) error {
+	slog.Info("checking status of model", "model", modelName)
+	showReq := &api.ShowRequest{Name: modelName}
+
+	showCtx, cancel := context.WithDeadlineCause(
+		ctx,
+		time.Now().Add(5*time.Second),
+		fmt.Errorf("show for existing model %s took too long", modelName),
+	)
+	defer cancel()
+	_, err := client.Show(showCtx, showReq)
+	var statusError api.StatusError
+	switch {
+	case errors.As(err, &statusError) && statusError.StatusCode == http.StatusNotFound:
+		break
+	case err != nil:
+		return err
+	default:
+		slog.Info("model already present", "model", modelName)
+		return nil
+	}
+	slog.Info("model missing", "model", modelName)
+
+	stallDuration := 30 * time.Second // This includes checksum verification, which can take a while on larger models
+	stallTimer := time.NewTimer(stallDuration)
+	fn := func(resp api.ProgressResponse) error {
+		// fmt.Print(".")
+		if !stallTimer.Reset(stallDuration) {
+			return fmt.Errorf("stall was detected, aborting status reporting")
+		}
+		return nil
+	}
+
+	stream := true
+	pullReq := &api.PullRequest{Name: modelName, Stream: &stream}
+
+	var pullError error
+
+	done := make(chan int)
+	go func() {
+		pullError = client.Pull(ctx, pullReq, fn)
+		done <- 0
+	}()
+
+	select {
+	case <-stallTimer.C:
+		return fmt.Errorf("download stalled")
+	case <-done:
+		return pullError
+	}
+}
+
+var serverProcMutex sync.Mutex
+
+// Returns an Client, the testEndpoint, and a cleanup function, fails the test on errors
+// Starts the server if needed
+func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, string, func()) {
+	client, testEndpoint := GetTestEndpoint()
+	if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
+		serverProcMutex.Lock()
+		fp, err := os.CreateTemp("", "ollama-server-*.log")
+		if err != nil {
+			t.Fatalf("failed to generate log file: %s", err)
+		}
+		lifecycle.ServerLogFile = fp.Name()
+		fp.Close()
+		require.NoError(t, startServer(ctx, testEndpoint))
+	}
+
+	return client, testEndpoint, func() {
+		if os.Getenv("OLLAMA_TEST_EXISTING") == "" {
+			defer serverProcMutex.Unlock()
+			if t.Failed() {
+				fp, err := os.Open(lifecycle.ServerLogFile)
+				if err != nil {
+					slog.Error("failed to open server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					return
+				}
+				data, err := io.ReadAll(fp)
+				if err != nil {
+					slog.Error("failed to read server log", "logfile", lifecycle.ServerLogFile, "error", err)
+					return
+				}
+				slog.Warn("SERVER LOG FOLLOWS")
+				os.Stderr.Write(data)
+				slog.Warn("END OF SERVER")
+			}
+			err := os.Remove(lifecycle.ServerLogFile)
+			if err != nil && !os.IsNotExist(err) {
+				slog.Warn("failed to cleanup", "logfile", lifecycle.ServerLogFile, "error", err)
+			}
+		}
+	}
+}
+
+func GenerateTestHelper(ctx context.Context, t *testing.T, genReq api.GenerateRequest, anyResp []string) {
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	require.NoError(t, PullIfMissing(ctx, client, genReq.Model))
+	DoGenerate(ctx, t, client, genReq, anyResp, 30*time.Second, 10*time.Second)
+}
+
+func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq api.GenerateRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) {
+	stallTimer := time.NewTimer(initialTimeout)
+	var buf bytes.Buffer
+	fn := func(response api.GenerateResponse) error {
+		// fmt.Print(".")
+		buf.Write([]byte(response.Response))
+		if !stallTimer.Reset(streamTimeout) {
+			return fmt.Errorf("stall was detected while streaming response, aborting")
+		}
+		return nil
+	}
+
+	stream := true
+	genReq.Stream = &stream
+	done := make(chan int)
+	var genErr error
+	go func() {
+		genErr = client.Generate(ctx, &genReq, fn)
+		done <- 0
+	}()
+
+	select {
+	case <-stallTimer.C:
+		if buf.Len() == 0 {
+			t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
+		} else {
+			t.Errorf("generate stalled.  Response so far:%s", buf.String())
+		}
+	case <-done:
+		require.NoError(t, genErr, "failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
+		// Verify the response contains the expected data
+		response := buf.String()
+		atLeastOne := false
+		for _, resp := range anyResp {
+			if strings.Contains(strings.ToLower(response), resp) {
+				atLeastOne = true
+				break
+			}
+		}
+		require.True(t, atLeastOne, "none of %v found in %s", anyResp, response)
+		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
+	case <-ctx.Done():
+		t.Error("outer test context done while waiting for generate")
+	}
+}
+
+// Generate a set of requests
+// By default each request uses orca-mini as the model
+func GenerateRequests() ([]api.GenerateRequest, [][]string) {
+	return []api.GenerateRequest{
+			{
+				Model:  "orca-mini",
+				Prompt: "why is the ocean blue?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:  "orca-mini",
+				Prompt: "why is the color of dirt brown?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:  "orca-mini",
+				Prompt: "what is the origin of the us thanksgiving holiday?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:  "orca-mini",
+				Prompt: "what is the origin of independence day?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}, {
+				Model:  "orca-mini",
+				Prompt: "what is the composition of air?",
+				Stream: &stream,
+				Options: map[string]interface{}{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			},
+		},
+		[][]string{
+			[]string{"sunlight"},
+			[]string{"soil", "organic", "earth", "black", "tan"},
+			[]string{"england", "english", "massachusetts", "pilgrims"},
+			[]string{"fourth", "july", "declaration", "independence"},
+			[]string{"nitrogen", "oxygen", "carbon", "dioxide"},
+		}
+}
--- a/llm/dyn_ext_server.c
+++ b/llm/dyn_ext_server.c
@ -1,142 +0,0 @@
-#include "dyn_ext_server.h"
-
-#include <stdio.h>
-#include <string.h>
-
-#ifdef __linux__
-#include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
-#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
-#define LOAD_ERR() strdup(dlerror())
-#define UNLOAD_LIBRARY(handle) dlclose(handle)
-#elif _WIN32
-#include <windows.h>
-#define LOAD_LIBRARY(lib, flags) LoadLibrary(lib)
-#define LOAD_SYMBOL(handle, sym) GetProcAddress(handle, sym)
-#define UNLOAD_LIBRARY(handle) FreeLibrary(handle)
-#define LOAD_ERR() ({\
-  LPSTR messageBuffer = NULL; \
-  size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, \
-                                 NULL, GetLastError(), MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&messageBuffer, 0, NULL); \
-  char *resp = strdup(messageBuffer); \
-  LocalFree(messageBuffer); \
-  resp; \
-})
-#else
-#include <dlfcn.h>
-#define LOAD_LIBRARY(lib, flags) dlopen(lib, flags)
-#define LOAD_SYMBOL(handle, sym) dlsym(handle, sym)
-#define LOAD_ERR() strdup(dlerror())
-#define UNLOAD_LIBRARY(handle) dlclose(handle)
-#endif
-
-void dyn_init(const char *libPath, struct dynamic_llama_server *s,
-                       ext_server_resp_t *err) {
-  int i = 0;
-  struct lookup {
-    char *s;
-    void **p;
-  } l[] = {
-      {"llama_server_init", (void *)&s->llama_server_init},
-      {"llama_server_start", (void *)&s->llama_server_start},
-      {"llama_server_stop", (void *)&s->llama_server_stop},
-      {"llama_server_completion", (void *)&s->llama_server_completion},
-      {"llama_server_completion_next_result",
-       (void *)&s->llama_server_completion_next_result},
-      {"llama_server_completion_cancel",
-       (void *)&s->llama_server_completion_cancel},
-      {"llama_server_release_task_result",
-       (void *)&s->llama_server_release_task_result},
-      {"llama_server_tokenize", (void *)&s->llama_server_tokenize},
-      {"llama_server_detokenize", (void *)&s->llama_server_detokenize},
-      {"llama_server_embedding", (void *)&s->llama_server_embedding},
-      {"llama_server_release_json_resp",
-       (void *)&s->llama_server_release_json_resp},
-      {"", NULL},
-  };
-
-  printf("loading library %s\n", libPath);
-  s->handle = LOAD_LIBRARY(libPath, RTLD_LOCAL|RTLD_NOW);
-  if (!s->handle) {
-    err->id = -1;
-    char *msg = LOAD_ERR();
-    snprintf(err->msg, err->msg_len,
-             "Unable to load dynamic server library: %s", msg);
-    free(msg);
-    return;
-  }
-
-  for (i = 0; l[i].p != NULL; i++) {
-    *l[i].p = LOAD_SYMBOL(s->handle, l[i].s);
-    if (!l[i].p) {
-      UNLOAD_LIBRARY(s->handle);
-      err->id = -1;
-      char *msg = LOAD_ERR();
-      snprintf(err->msg, err->msg_len, "symbol lookup for %s failed: %s",
-               l[i].s, msg);
-      free(msg);
-      return;
-    }
-  }
-}
-
-inline void dyn_llama_server_init(struct dynamic_llama_server s,
-                                           ext_server_params_t *sparams,
-                                           ext_server_resp_t *err) {
-  s.llama_server_init(sparams, err);
-}
-
-inline void dyn_llama_server_start(struct dynamic_llama_server s) {
-  s.llama_server_start();
-}
-
-inline void dyn_llama_server_stop(struct dynamic_llama_server s) {
-  s.llama_server_stop();
-}
-
-inline void dyn_llama_server_completion(struct dynamic_llama_server s,
-                                                 const char *json_req,
-                                                 ext_server_resp_t *resp) {
-  s.llama_server_completion(json_req, resp);
-}
-
-inline void dyn_llama_server_completion_next_result(
-    struct dynamic_llama_server s, const int task_id,
-    ext_server_task_result_t *result) {
-  s.llama_server_completion_next_result(task_id, result);
-}
-
-inline void dyn_llama_server_completion_cancel(
-    struct dynamic_llama_server s, const int task_id, ext_server_resp_t *err) {
-  s.llama_server_completion_cancel(task_id, err);
-}
-inline void dyn_llama_server_release_task_result(
-    struct dynamic_llama_server s, ext_server_task_result_t *result) {
-  s.llama_server_release_task_result(result);
-}
-
-inline void dyn_llama_server_tokenize(struct dynamic_llama_server s,
-                                               const char *json_req,
-                                               char **json_resp,
-                                               ext_server_resp_t *err) {
-  s.llama_server_tokenize(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_detokenize(struct dynamic_llama_server s,
-                                                 const char *json_req,
-                                                 char **json_resp,
-                                                 ext_server_resp_t *err) {
-  s.llama_server_detokenize(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_embedding(struct dynamic_llama_server s,
-                                                const char *json_req,
-                                                char **json_resp,
-                                                ext_server_resp_t *err) {
-  s.llama_server_embedding(json_req, json_resp, err);
-}
-
-inline void dyn_llama_server_release_json_resp(
-    struct dynamic_llama_server s, char **json_resp) {
-  s.llama_server_release_json_resp(json_resp);
-}
--- a/llm/dyn_ext_server.go
+++ b/llm/dyn_ext_server.go
@ -1,368 +0,0 @@
-package llm
-
-/*
-#cgo CFLAGS: -I${SRCDIR}/ext_server -I${SRCDIR}/llama.cpp -I${SRCDIR}/llama.cpp/common -I${SRCDIR}/llama.cpp/examples/server
-#cgo CFLAGS: -DNDEBUG -DLLAMA_SERVER_LIBRARY=1 -D_XOPEN_SOURCE=600 -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64
-#cgo CFLAGS: -Wmissing-noreturn -Wextra -Wcast-qual -Wno-unused-function -Wno-array-bounds
-#cgo CPPFLAGS: -Ofast -Wextra -Wno-unused-function -Wno-unused-variable -Wno-deprecated-declarations
-#cgo darwin CFLAGS: -D_DARWIN_C_SOURCE
-#cgo darwin CPPFLAGS:  -DGGML_USE_ACCELERATE
-#cgo darwin CPPFLAGS: -DGGML_USE_METAL -DGGML_METAL_NDEBUG
-#cgo darwin LDFLAGS: -lc++ -framework Accelerate
-#cgo darwin LDFLAGS: -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
-#cgo linux CFLAGS: -D_GNU_SOURCE
-#cgo linux LDFLAGS: -lrt -ldl -lstdc++ -lm
-#cgo linux windows LDFLAGS: -lpthread
-
-#include <stdlib.h>
-#include "dyn_ext_server.h"
-
-*/
-import "C"
-
-import (
-	"bytes"
-	"context"
-	"encoding/json"
-	"fmt"
-	"log/slog"
-	"os"
-	"path/filepath"
-	"strings"
-	"sync"
-	"time"
-	"unsafe"
-
-	"github.com/jmorganca/ollama/api"
-	"github.com/jmorganca/ollama/gpu"
-)
-
-type dynExtServer struct {
-	s       C.struct_dynamic_llama_server
-	options api.Options
-}
-
-// Note: current implementation does not support concurrent instantiations
-var mutex sync.Mutex
-
-func newExtServerResp(len C.size_t) C.ext_server_resp_t {
-	var resp C.ext_server_resp_t
-	resp.msg_len = len
-	bytes := make([]byte, len)
-	resp.msg = (*C.char)(C.CBytes(bytes))
-	return resp
-}
-
-func freeExtServerResp(resp C.ext_server_resp_t) {
-	if resp.msg_len == 0 {
-		return
-	}
-	C.free(unsafe.Pointer(resp.msg))
-}
-
-func extServerResponseToErr(resp C.ext_server_resp_t) error {
-	return fmt.Errorf(C.GoString(resp.msg))
-}
-
-// Note: current implementation does not support concurrent instantiations
-var llm *dynExtServer
-
-func newDynExtServer(library, model string, adapters, projectors []string, opts api.Options) (LLM, error) {
-	if !mutex.TryLock() {
-		slog.Info("concurrent llm servers not yet supported, waiting for prior server to complete")
-		mutex.Lock()
-	}
-	gpu.UpdatePath(filepath.Dir(library))
-	libPath := C.CString(library)
-	defer C.free(unsafe.Pointer(libPath))
-	resp := newExtServerResp(512)
-	defer freeExtServerResp(resp)
-	var srv C.struct_dynamic_llama_server
-	C.dyn_init(libPath, &srv, &resp)
-	if resp.id < 0 {
-		mutex.Unlock()
-		return nil, fmt.Errorf("Unable to load dynamic library: %s", C.GoString(resp.msg))
-	}
-	llm = &dynExtServer{
-		s:       srv,
-		options: opts,
-	}
-	slog.Info(fmt.Sprintf("Loading Dynamic llm server: %s", library))
-
-	var sparams C.ext_server_params_t
-	sparams.model = C.CString(model)
-	defer C.free(unsafe.Pointer(sparams.model))
-
-	sparams.embedding = true
-	sparams.n_ctx = C.uint(opts.NumCtx)
-	sparams.n_batch = C.uint(opts.NumBatch)
-	sparams.n_gpu_layers = C.int(opts.NumGPU)
-	sparams.main_gpu = C.int(opts.MainGPU)
-	sparams.n_parallel = 1 // TODO - wire up concurrency
-
-	// Always use the value encoded in the model
-	sparams.rope_freq_base = 0.0
-	sparams.rope_freq_scale = 0.0
-	sparams.memory_f16 = C.bool(opts.F16KV)
-	sparams.use_mlock = C.bool(opts.UseMLock)
-	sparams.use_mmap = C.bool(opts.UseMMap)
-
-	if opts.UseNUMA {
-		sparams.numa = C.int(1)
-	} else {
-		sparams.numa = C.int(0)
-	}
-
-	sparams.lora_adapters = nil
-	for i := 0; i < len(adapters); i++ {
-		la := (*C.ext_server_lora_adapter_t)(C.malloc(C.sizeof_ext_server_lora_adapter_t))
-		defer C.free(unsafe.Pointer(la))
-		la.adapter = C.CString(adapters[i])
-		defer C.free(unsafe.Pointer(la.adapter))
-		la.scale = C.float(1.0) // TODO expose scale/weights up through ollama UX
-		la.next = nil
-		if i == 0 {
-			sparams.lora_adapters = la
-		} else {
-			tmp := sparams.lora_adapters
-			for ; tmp.next != nil; tmp = tmp.next {
-			}
-			tmp.next = la
-		}
-	}
-
-	if len(projectors) > 0 {
-		// TODO: applying multiple projectors is not supported by the llama.cpp server yet
-		sparams.mmproj = C.CString(projectors[0])
-		defer C.free(unsafe.Pointer(sparams.mmproj))
-	} else {
-		sparams.mmproj = nil
-	}
-
-	sparams.n_threads = C.uint(opts.NumThread)
-
-	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
-		sparams.verbose_logging = C.bool(true)
-	} else {
-		sparams.verbose_logging = C.bool(false)
-	}
-
-	slog.Info("Initializing llama server")
-	slog.Debug(fmt.Sprintf("server params: %+v", sparams))
-	initResp := newExtServerResp(512)
-	defer freeExtServerResp(initResp)
-	C.dyn_llama_server_init(llm.s, &sparams, &initResp)
-	if initResp.id < 0 {
-		mutex.Unlock()
-		err := extServerResponseToErr(initResp)
-		slog.Debug(fmt.Sprintf("failure during initialization: %s", err))
-		return nil, err
-	}
-
-	slog.Info("Starting llama main loop")
-	C.dyn_llama_server_start(llm.s)
-	return llm, nil
-}
-
-func (llm *dynExtServer) Predict(ctx context.Context, predict PredictOpts, fn func(PredictResult)) error {
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-
-	if len(predict.Images) > 0 {
-		slog.Info(fmt.Sprintf("loaded %d images", len(predict.Images)))
-	}
-
-	request := map[string]any{
-		"prompt":            predict.Prompt,
-		"stream":            true,
-		"n_predict":         predict.Options.NumPredict,
-		"n_keep":            predict.Options.NumKeep,
-		"temperature":       predict.Options.Temperature,
-		"top_k":             predict.Options.TopK,
-		"top_p":             predict.Options.TopP,
-		"tfs_z":             predict.Options.TFSZ,
-		"typical_p":         predict.Options.TypicalP,
-		"repeat_last_n":     predict.Options.RepeatLastN,
-		"repeat_penalty":    predict.Options.RepeatPenalty,
-		"presence_penalty":  predict.Options.PresencePenalty,
-		"frequency_penalty": predict.Options.FrequencyPenalty,
-		"mirostat":          predict.Options.Mirostat,
-		"mirostat_tau":      predict.Options.MirostatTau,
-		"mirostat_eta":      predict.Options.MirostatEta,
-		"penalize_nl":       predict.Options.PenalizeNewline,
-		"seed":              predict.Options.Seed,
-		"stop":              predict.Options.Stop,
-		"image_data":        predict.Images,
-		"cache_prompt":      true,
-	}
-
-	if predict.Format == "json" {
-		request["grammar"] = jsonGrammar
-	}
-
-	retryDelay := 100 * time.Microsecond
-	for retries := 0; retries < maxRetries; retries++ {
-		if retries > 0 {
-			time.Sleep(retryDelay) // wait before retrying
-			retryDelay *= 2        // exponential backoff
-		}
-
-		// Handling JSON marshaling with special characters unescaped.
-		buffer := &bytes.Buffer{}
-		enc := json.NewEncoder(buffer)
-		enc.SetEscapeHTML(false)
-
-		if err := enc.Encode(request); err != nil {
-			return fmt.Errorf("failed to marshal data: %w", err)
-		}
-
-		req := C.CString(buffer.String())
-		defer C.free(unsafe.Pointer(req))
-
-		C.dyn_llama_server_completion(llm.s, req, &resp)
-		if resp.id < 0 {
-			return extServerResponseToErr(resp)
-		}
-
-		retryNeeded := false
-	out:
-		for {
-			select {
-			case <-ctx.Done():
-				// This handles the request cancellation
-				C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
-				if resp.id < 0 {
-					return extServerResponseToErr(resp)
-				} else {
-					return nil
-				}
-			default:
-				var result C.ext_server_task_result_t
-				C.dyn_llama_server_completion_next_result(llm.s, resp.id, &result)
-				json_resp := C.GoString(result.json_resp)
-				C.dyn_llama_server_release_task_result(llm.s, &result)
-
-				var p prediction
-				if err := json.Unmarshal([]byte(json_resp), &p); err != nil {
-					C.dyn_llama_server_completion_cancel(llm.s, resp.id, &resp)
-					if resp.id < 0 {
-						return fmt.Errorf("error unmarshaling llm prediction response: %w and cancel %s", err, C.GoString(resp.msg))
-					} else {
-						return fmt.Errorf("error unmarshaling llm prediction response: %w", err)
-					}
-				}
-
-				if bool(result.error) && strings.Contains(json_resp, "slot unavailable") {
-					retryNeeded = true
-					// task will already be canceled
-					break out
-				}
-
-				if p.Content != "" {
-					fn(PredictResult{
-						Content: p.Content,
-					})
-				}
-
-				if p.Stop || bool(result.stop) {
-					fn(PredictResult{
-						Done:               true,
-						PromptEvalCount:    p.Timings.PromptN,
-						PromptEvalDuration: parseDurationMs(p.Timings.PromptMS),
-						EvalCount:          p.Timings.PredictedN,
-						EvalDuration:       parseDurationMs(p.Timings.PredictedMS),
-					})
-					return nil
-				}
-			}
-		}
-		if !retryNeeded {
-			return nil // success
-		}
-	}
-
-	// should never reach here ideally
-	return fmt.Errorf("max retries exceeded")
-}
-
-func (llm *dynExtServer) Encode(ctx context.Context, prompt string) ([]int, error) {
-	data, err := json.Marshal(TokenizeRequest{Content: prompt})
-	if err != nil {
-		return nil, fmt.Errorf("marshaling encode data: %w", err)
-	}
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_tokenize(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return nil, extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var encoded TokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &encoded); err2 != nil {
-		return nil, fmt.Errorf("unmarshal encode response: %w", err2)
-	}
-
-	return encoded.Tokens, err
-}
-
-func (llm *dynExtServer) Decode(ctx context.Context, tokens []int) (string, error) {
-	if len(tokens) == 0 {
-		return "", nil
-	}
-	data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
-	if err != nil {
-		return "", fmt.Errorf("marshaling decode data: %w", err)
-	}
-
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_detokenize(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return "", extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var decoded DetokenizeResponse
-	if err2 := json.Unmarshal([]byte(C.GoString(json_resp)), &decoded); err2 != nil {
-		return "", fmt.Errorf("unmarshal encode response: %w", err2)
-	}
-
-	return decoded.Content, err
-}
-
-func (llm *dynExtServer) Embedding(ctx context.Context, input string) ([]float64, error) {
-	data, err := json.Marshal(TokenizeRequest{Content: input})
-	if err != nil {
-		return nil, fmt.Errorf("error marshaling embed data: %w", err)
-	}
-
-	req := C.CString(string(data))
-	defer C.free(unsafe.Pointer(req))
-	var json_resp *C.char
-	resp := newExtServerResp(128)
-	defer freeExtServerResp(resp)
-	C.dyn_llama_server_embedding(llm.s, req, &json_resp, &resp)
-	if resp.id < 0 {
-		return nil, extServerResponseToErr(resp)
-	}
-	defer C.dyn_llama_server_release_json_resp(llm.s, &json_resp)
-
-	var embedding EmbeddingResponse
-	if err := json.Unmarshal([]byte(C.GoString(json_resp)), &embedding); err != nil {
-		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
-	}
-
-	return embedding.Embedding, nil
-}
-
-func (llm *dynExtServer) Close() {
-	C.dyn_llama_server_stop(llm.s)
-	mutex.Unlock()
-}
--- a/llm/dyn_ext_server.h
+++ b/llm/dyn_ext_server.h
@ -1,74 +0,0 @@
-#include <stdlib.h>
-
-#include "ext_server.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-struct dynamic_llama_server {
-  void *handle;
-  void (*llama_server_init)(ext_server_params_t *sparams,
-                            ext_server_resp_t *err);
-  void (*llama_server_start)();
-  void (*llama_server_stop)();
-  void (*llama_server_completion)(const char *json_req,
-                                  ext_server_resp_t *resp);
-  void (*llama_server_completion_next_result)(const int task_id,
-                                              ext_server_task_result_t *result);
-  void (*llama_server_completion_cancel)(const int task_id,
-                                         ext_server_resp_t *err);
-  void (*llama_server_release_task_result)(ext_server_task_result_t *result);
-  void (*llama_server_tokenize)(const char *json_req, char **json_resp,
-                                ext_server_resp_t *err);
-  void (*llama_server_detokenize)(const char *json_req, char **json_resp,
-                                  ext_server_resp_t *err);
-  void (*llama_server_embedding)(const char *json_req, char **json_resp,
-                                 ext_server_resp_t *err);
-  void (*llama_server_release_json_resp)(char **json_resp);
-};
-
-void dyn_init(const char *libPath, struct dynamic_llama_server *s,
-                       ext_server_resp_t *err);
-
-// No good way to call C function pointers from Go so inline the indirection
-void dyn_llama_server_init(struct dynamic_llama_server s,
-                                    ext_server_params_t *sparams,
-                                    ext_server_resp_t *err);
-
-void dyn_llama_server_start(struct dynamic_llama_server s);
-
-void dyn_llama_server_stop(struct dynamic_llama_server s);
-
-void dyn_llama_server_completion(struct dynamic_llama_server s,
-                                          const char *json_req,
-                                          ext_server_resp_t *resp);
-
-void dyn_llama_server_completion_next_result(
-    struct dynamic_llama_server s, const int task_id,
-    ext_server_task_result_t *result);
-
-void dyn_llama_server_completion_cancel(struct dynamic_llama_server s,
-                                                 const int task_id,
-                                                 ext_server_resp_t *err);
-
-void dyn_llama_server_release_task_result(
-    struct dynamic_llama_server s, ext_server_task_result_t *result);
-
-void dyn_llama_server_tokenize(struct dynamic_llama_server s,
-                                        const char *json_req, char **json_resp,
-                                        ext_server_resp_t *err);
-
-void dyn_llama_server_detokenize(struct dynamic_llama_server s,
-                                          const char *json_req,
-                                          char **json_resp,
-                                          ext_server_resp_t *err);
-
-void dyn_llama_server_embedding(struct dynamic_llama_server s,
-                                         const char *json_req, char **json_resp,
-                                         ext_server_resp_t *err);
-void dyn_llama_server_release_json_resp(struct dynamic_llama_server s,
-                                                 char **json_resp);
-
-#ifdef __cplusplus
-}
-#endif
--- a/llm/ext_server/CMakeLists.txt
+++ b/llm/ext_server/CMakeLists.txt
@ -1,25 +1,14 @@
-# Ollama specific CMakefile to include in llama.cpp/examples/server

-set(TARGET ext_server)
+set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
-    add_library(${TARGET} SHARED ../../../ext_server/ext_server.cpp ../../llama.cpp)
-else()
-    add_library(${TARGET} STATIC ../../../ext_server/ext_server.cpp ../../llama.cpp)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_include_directories(${TARGET} PRIVATE ../..)
-target_include_directories(${TARGET} PRIVATE ../../..)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
-target_compile_definitions(${TARGET} PUBLIC LLAMA_SERVER_LIBRARY=1)
-target_link_libraries(${TARGET} PRIVATE ggml llava common )
-set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>)
-install(TARGETS ext_server LIBRARY)
-
-if (CUDAToolkit_FOUND)
-    target_include_directories(${TARGET} PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-    if (WIN32)
-        target_link_libraries(${TARGET} PRIVATE nvml)
-    endif()
-endif()
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/llm/ext_server/README.md
+++ b/llm/ext_server/README.md
@ -1,18 +0,0 @@
-# Extern C Server
-
-This directory contains a thin facade we layer on top of the Llama.cpp server to
-expose `extern C` interfaces to access the functionality through direct API
-calls in-process.  The llama.cpp code uses compile time macros to configure GPU
-type along with other settings.  During the `go generate ./...` execution, the
-build will generate one or more copies of the llama.cpp `extern C` server based
-on what GPU libraries are detected to support multiple GPU types as well as CPU
-only support. The Ollama go build then embeds these different servers to support
-different GPUs and settings at runtime.
-
-If you are making changes to the code in this directory, make sure to disable
-caching during your go build to ensure you pick up your changes.  A typical
-iteration cycle from the top of the source tree looks like:
-
-```
-go generate ./... && go build -a .
-```
--- a/llm/ext_server/ext_server.cpp
+++ b/llm/ext_server/ext_server.cpp
@ -1,377 +0,0 @@
-#include "ext_server.h"
-#include <atomic>
-
-// Necessary evil since the server types are not defined in a header
-#include "server.cpp"
-
-// Low level API access to verify GPU access
-#if defined(GGML_USE_CUBLAS)
-#if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define cudaGetDevice hipGetDevice
-#define cudaError_t hipError_t
-#define cudaSuccess hipSuccess
-#define cudaGetErrorString hipGetErrorString
-#else
-#include <cuda_runtime.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-#endif // defined(GGML_USE_HIPBLAS)
-#endif // GGML_USE_CUBLAS
-
-// Expose the llama server as a callable extern "C" API
-llama_server_context *llama = NULL;
-std::thread ext_server_thread;
-bool shutting_down = false;
-std::atomic_int recv_counter;
-
-// RAII wrapper for tracking in-flight recv calls
-class atomicRecv {
-  public:
-    atomicRecv(std::atomic<int> &atomic) : atomic(atomic) {
-      ++this->atomic;
-    }
-    ~atomicRecv() {
-      --this->atomic;
-    }
-  private:
-    std::atomic<int> &atomic;
-};
- 
-void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
-  recv_counter = 0;
-  assert(err != NULL && sparams != NULL);
-  log_set_target(stderr);
-  if (!sparams->verbose_logging) {
-    server_verbose = true;
-    log_disable();
-  }
-
-  LOG_TEE("system info: %s\n", llama_print_system_info());
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    llama = new llama_server_context;
-    gpt_params params;
-    params.n_ctx = sparams->n_ctx;
-    params.n_batch = sparams->n_batch;
-    if (sparams->n_threads > 0) {
-      params.n_threads = sparams->n_threads;
-    }
-    params.n_parallel = sparams->n_parallel;
-    params.rope_freq_base = sparams->rope_freq_base;
-    params.rope_freq_scale = sparams->rope_freq_scale;
-
-    if (sparams->memory_f16) {
-      params.cache_type_k = "f16";
-      params.cache_type_v = "f16";
-    } else {
-      params.cache_type_k = "f32";
-      params.cache_type_v = "f32";
-    }
-
-    params.n_gpu_layers = sparams->n_gpu_layers;
-    params.main_gpu = sparams->main_gpu;
-    params.use_mlock = sparams->use_mlock;
-    params.use_mmap = sparams->use_mmap;
-    params.numa = (ggml_numa_strategy)sparams->numa;
-    params.embedding = sparams->embedding;
-    if (sparams->model != NULL) {
-      params.model = sparams->model;
-    }
-
-    if (sparams->lora_adapters != NULL) {
-      for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL;
-          la = la->next) {
-        params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
-      }
-
-      params.use_mmap = false;
-    }
-
-    if (sparams->mmproj != NULL) {
-      params.mmproj = std::string(sparams->mmproj);
-    }
-
-#if defined(GGML_USE_CUBLAS)
-    // Before attempting to init the backend which will assert on error, verify the CUDA/ROCM GPU is accessible
-    LOG_TEE("Performing pre-initialization of GPU\n");
-    int id;
-    cudaError_t cudaErr = cudaGetDevice(&id);
-    if (cudaErr != cudaSuccess) {
-      err->id = -1;
-      snprintf(err->msg, err->msg_len, "Unable to init GPU: %s", cudaGetErrorString(cudaErr));
-      return;
-    }
-#endif
-
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-  if (!llama->load_model(params)) { 
-    // an error occured that was not thrown
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "error loading model %s", params.model.c_str());
-    return;
-  }
-
-    llama->initialize();
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len,
-             "Unknown exception initializing llama server");
-  }
-}
-
-void llama_server_start() {
-  assert(llama != NULL);
-  // TODO mutex to protect thread creation
-  ext_server_thread = std::thread([&]() {
-    try {
-      LOG_TEE("llama server main loop starting\n");
-      ggml_time_init();
-      llama->queue_tasks.on_new_task(std::bind(
-        &llama_server_context::process_single_task, llama, std::placeholders::_1));
-      llama->queue_tasks.on_finish_multitask(std::bind(
-        &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
-      llama->queue_tasks.on_run_slots(std::bind(
-        &llama_server_context::update_slots, llama));
-      llama->queue_results.on_multitask_update(std::bind(
-          &llama_server_queue::update_multitask,
-          &llama->queue_tasks,
-          std::placeholders::_1,
-          std::placeholders::_2,
-          std::placeholders::_3
-        ));
-      llama->queue_tasks.start_loop();
-    } catch (std::exception &e) {
-      LOG_TEE("caught exception in llama server main loop: %s\n", e.what());
-    } catch (...) {
-      LOG_TEE("caught unknown exception in llama server main loop\n");
-    }
-    LOG_TEE("\nllama server shutting down\n");
-    llama_backend_free();
-  });
-}
-
-void llama_server_stop() {
-  assert(llama != NULL);
-  // Shutdown any in-flight requests and block incoming requests.
-  LOG_TEE("\ninitiating shutdown - draining remaining tasks...\n");
-  shutting_down = true;
-
-  while (recv_counter.load() > 0) {
-    std::this_thread::sleep_for(std::chrono::milliseconds(50));
-  }
-
-  // This may take a while for any pending tasks to drain
-  // TODO - consider a timeout to cancel tasks if it's taking too long
-  llama->queue_tasks.terminate();
-  ext_server_thread.join();
-  delete llama;
-  llama = NULL;
-  LOG_TEE("llama server shutdown complete\n");
-  shutting_down = false;
-}
-
-void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
-  assert(llama != NULL && json_req != NULL && resp != NULL);
-  resp->id = -1;
-  resp->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    json data = json::parse(json_req);
-    resp->id = llama->queue_tasks.get_new_id();
-    llama->queue_results.add_waiting_task_id(resp->id);
-    llama->request_completion(resp->id, data, false, false, -1);
-  } catch (std::exception &e) {
-    snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
-  } catch (...) {
-    snprintf(resp->msg, resp->msg_len, "Unknown exception during completion");
-  }
-}
-
-void llama_server_completion_next_result(const int task_id,
-                                         ext_server_task_result_t *resp) {
-  assert(llama != NULL && resp != NULL);
-  resp->id = -1;
-  resp->stop = false;
-  resp->error = false;
-  resp->json_resp = NULL;
-  std::string result_json;
-  try {
-    atomicRecv ar(recv_counter);
-    task_result result = llama->queue_results.recv(task_id);
-    result_json =
-        result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
-    resp->id = result.id;
-    resp->stop = result.stop;
-    resp->error = result.error;
-    if (result.error) {
-      LOG_TEE("next result cancel on error\n");
-      llama->request_cancel(task_id);
-      LOG_TEE("next result removing waiting tak ID: %d\n", task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-    } else if (result.stop) {
-      LOG_TEE("next result cancel on stop\n");
-      llama->request_cancel(task_id);
-      LOG_TEE("next result removing waiting task ID: %d\n", task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-    } else if (shutting_down) {
-      LOG_TEE("aborting completion due to shutdown %d\n", task_id);
-      llama->request_cancel(task_id);
-      llama->queue_results.remove_waiting_task_id(task_id);
-      resp->stop = true;
-    }
-  } catch (std::exception &e) {
-    resp->error = true;
-    resp->id = -1;
-    result_json = "{\"error\":\"exception " + std::string(e.what()) + "\"}";
-    LOG_TEE("llama server completion exception %s\n", e.what());
-  } catch (...) {
-    resp->error = true;
-    resp->id = -1;
-    result_json = "{\"error\":\"Unknown exception during completion\"}";
-    LOG_TEE("llama server completion unknown exception\n");
-  }
-  const std::string::size_type size = result_json.size() + 1;
-  resp->json_resp = new char[size];
-  snprintf(resp->json_resp, size, "%s", result_json.c_str());
-}
-
-void llama_server_release_task_result(ext_server_task_result_t *result) {
-  if (result == NULL || result->json_resp == NULL) {
-    return;
-  }
-  delete[] result->json_resp;
-}
-
-void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err) {
-  assert(llama != NULL && err != NULL);
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    llama->request_cancel(task_id);
-    llama->queue_results.remove_waiting_task_id(task_id);
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len,
-             "Unknown exception completion cancel in llama server");
-  }
-}
-
-void llama_server_tokenize(const char *json_req, char **json_resp,
-                           ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    std::vector<llama_token> tokens;
-    if (body.count("content") != 0) {
-      tokens = llama->tokenize(body["content"], false);
-    }
-    const json data = format_tokenizer_response(tokens);
-    std::string result_json = data.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during tokenize");
-  }
-}
-
-void llama_server_release_json_resp(char **json_resp) {
-  if (json_resp == NULL || *json_resp == NULL) {
-    return;
-  }
-  delete[] *json_resp;
-}
-
-void llama_server_detokenize(const char *json_req, char **json_resp,
-                             ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    std::string content;
-    if (body.count("tokens") != 0) {
-      const std::vector<llama_token> tokens = body["tokens"];
-      content = tokens_to_str(llama->ctx, tokens.cbegin(), tokens.cend());
-    }
-    const json data = format_detokenized_response(content);
-    std::string result_json = data.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during detokenize");
-  }
-}
-
-void llama_server_embedding(const char *json_req, char **json_resp,
-                            ext_server_resp_t *err) {
-  assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
-  *json_resp = NULL;
-  err->id = 0;
-  err->msg[0] = '\0';
-  try {
-    if (shutting_down) {
-      throw std::runtime_error("server shutting down");
-    }
-    const json body = json::parse(json_req);
-    json prompt;
-    if (body.count("content") != 0) {
-      prompt = body["content"];
-    } else {
-      prompt = "";
-    }
-    const int task_id = llama->queue_tasks.get_new_id();
-    llama->queue_results.add_waiting_task_id(task_id);
-    llama->request_completion(task_id, {{"prompt", prompt}, {"n_predict", 0}}, false, true, -1);
-    atomicRecv ar(recv_counter);
-    task_result result = llama->queue_results.recv(task_id);
-    std::string result_json = result.result_json.dump();
-    const std::string::size_type size = result_json.size() + 1;
-    *json_resp = new char[size];
-    snprintf(*json_resp, size, "%s", result_json.c_str());
-    llama->queue_results.remove_waiting_task_id(task_id);
-  } catch (std::exception &e) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "exception %s", e.what());
-  } catch (...) {
-    err->id = -1;
-    snprintf(err->msg, err->msg_len, "Unknown exception during embedding");
-  }
-}
--- a/llm/ext_server/ext_server.h
+++ b/llm/ext_server/ext_server.h
@ -1,95 +0,0 @@
-#if defined(LLAMA_SERVER_LIBRARY)
-#ifndef LLAMA_SERVER_H
-#define LLAMA_SERVER_H
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-#include <stdio.h>
-
-int __main(int argc, char **argv);
-
-// This exposes extern C entrypoints into the llama_server
-// To enable the server compile with LLAMA_SERVER_LIBRARY
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-typedef struct ext_server_resp {
-  int id;          // < 0 on error
-  size_t msg_len;  // caller must allocate msg and set msg_len
-  char *msg;
-} ext_server_resp_t;
-
-// Allocated and freed by caller
-typedef struct ext_server_lora_adapter {
-  char *adapter;
-  float scale;
-  struct ext_server_lora_adapter *next;
-} ext_server_lora_adapter_t;
-
-// Allocated and freed by caller
-typedef struct ext_server_params {
-  char *model;
-  uint32_t n_ctx;         // token context window, 0 = from model
-  uint32_t n_batch;       // prompt processing maximum batch size
-  uint32_t n_threads;     // number of threads to use for generation
-  int32_t n_parallel;     // number of parallel sequences to decodewra
-  float rope_freq_base;   // RoPE base frequency, 0 = from model
-  float rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
-  bool memory_f16;        // use f16 instead of f32 for memory kv
-  int32_t n_gpu_layers;  // number of layers to store in VRAM (-1 - use default)
-  int32_t main_gpu;      // the GPU that is used for scratch and small tensors
-  bool use_mlock;        // force system to keep model in RAM
-  bool use_mmap;         // use mmap if possible
-  int numa;              // attempt optimizations that help on some NUMA systems
-  bool embedding;        // get only sentence embedding
-  ext_server_lora_adapter_t *lora_adapters;
-  char *mmproj;
-  bool verbose_logging;  // Enable verbose logging of the server
-} ext_server_params_t;
-
-typedef struct ext_server_task_result {
-  int id;
-  bool stop;
-  bool error;
-  char *json_resp;  // null terminated, memory managed by ext_server
-} ext_server_task_result_t;
-
-// Initialize the server once per process
-// err->id = 0 for success and err->msg[0] = NULL
-// err->id != 0 for failure, and err->msg contains error message
-void llama_server_init(ext_server_params_t *sparams, ext_server_resp_t *err);
-
-// Run the main loop, called once per init
-void llama_server_start();
-// Stop the main loop and free up resources allocated in init and start.  Init
-// must be called again to reuse
-void llama_server_stop();
-
-// json_req null terminated string, memory managed by caller
-// resp->id >= 0 on success (task ID)
-// resp->id < 0 on error, and resp->msg contains error message
-void llama_server_completion(const char *json_req, ext_server_resp_t *resp);
-
-// Caller must call llama_server_release_task_result to free resp->json_resp
-void llama_server_completion_next_result(const int task_id,
-                                         ext_server_task_result_t *result);
-void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err);
-void llama_server_release_task_result(ext_server_task_result_t *result);
-
-// Caller must call llama_server_releaes_json_resp to free json_resp if err.id <
-// 0
-void llama_server_tokenize(const char *json_req, char **json_resp,
-                           ext_server_resp_t *err);
-void llama_server_detokenize(const char *json_req, char **json_resp,
-                             ext_server_resp_t *err);
-void llama_server_embedding(const char *json_req, char **json_resp,
-                            ext_server_resp_t *err);
-void llama_server_release_json_resp(char **json_resp);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
-#endif  // LLAMA_SERVER_LIBRARY
--- a/Show More
+++ b/Show More