Merge pull request #7932 from ollama/mxyng/fix-merges

fix unmarshaling merges
llm: normalise kvct parameter handling (#7926 )
2026-08-07 05:23:57 -04:00 · 2024-12-04 10:04:52 -08:00 · 2024-12-04 09:21:56 -08:00 · 2024-12-03 16:30:40 -08:00 · 2024-12-03 15:57:19 -08:00 · 2024-12-03 15:00:05 -08:00
144 changed files with 9304 additions and 1820 deletions
@@ -1,5 +1,9 @@
 name: release

+env:
+  ROCM_WINDOWS_URL: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
+  MSYS2_URL: https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-x86_64-20240727.exe
+
 on:
  push:
    tags:
@@ -8,7 +12,7 @@ on:
 jobs:
  # Full build of the Mac assets
  build-darwin:
-    runs-on: macos-12
+    runs-on: macos-13
    environment: release
    steps:
      - uses: actions/checkout@v4
@@ -39,8 +43,8 @@ jobs:
          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
+          SDKROOT: /Applications/Xcode_14.1.0.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
+          DEVELOPER_DIR: /Applications/Xcode_14.1.0.app/Contents/Developer
        run: |
          ./scripts/build_darwin.sh

@@ -60,51 +64,34 @@ jobs:
      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
    steps:
      - uses: actions/checkout@v4
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
      - name: Set Version
        shell: bash
        run: echo "VERSION=${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
+      - name: Add msys paths
        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
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
        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"
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
      - 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
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$env:PATH"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make
        name: make
      - uses: actions/upload-artifact@v4
        with:
          name: generate-windows-cpu
          path: |
            build/**/*
-            build/**/*.a
            dist/windows-amd64/**

  # ROCm generation step
@@ -115,75 +102,55 @@ jobs:
      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
    steps:
      - uses: actions/checkout@v4
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
      - name: Set Version
        shell: bash
        run: echo "VERSION=${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
+      - name: Add msys paths
        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
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
        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"
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
          cache: true
-      - name: 'Install ROCm'
+      # ROCM installation steps
+      - name: 'Cache ROCm installer'
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: rocm-install.exe
+          key: ${{ env.ROCM_WINDOWS_URL }}
+      - name: 'Conditionally Download ROCm'
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
        run: |
          $ErrorActionPreference = "Stop"
-          write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
-          write-host "Completed AMD HIP"
+          Invoke-WebRequest -Uri "${env:ROCM_WINDOWS_URL}" -OutFile "rocm-install.exe"
+      - name: 'Install ROCm'
+        run: |
+          Start-Process "rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
      - name: 'Verify ROCm'
        run: |
          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
-      - run: go get ./...
-      - run: |
-          $gopath=(get-command go).source | split-path -parent
+          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path | select -first 1)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+      - name: make rocm runner
+        run: |
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$env:PATH"
-          $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
-        name: make
-      - name: 'gather rocm dependencies'
-        run: |
-          $HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          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\"
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make -C llama print-HIP_PATH print-HIP_LIB_DIR
+          make rocm
      - uses: actions/upload-artifact@v4
        with:
          name: generate-windows-rocm
          path: |
            build/**/*
            dist/windows-amd64/**
-      - uses: actions/upload-artifact@v4
-        with:
-          name: windows-rocm-deps
-          path: dist/deps/*

  # CUDA generation step
  generate-windows-cuda:
@@ -192,89 +159,80 @@ jobs:
    strategy:
      matrix:
        cuda:
-          - version: "11"
-            url: 'https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe'
-          - version: "12"
-            url: 'https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe'
+          - version: "11.3"
+            url: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+          - version: "12.4"
+            url: https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe
    env:
      KEY_CONTAINER: ${{ vars.KEY_CONTAINER }}
    steps:
      - uses: actions/checkout@v4
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
      - name: Set Version
        shell: bash
        run: echo "VERSION=${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
+      - name: Install msys2
        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
+          $msys2_url="https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-x86_64-20240727.exe"
+          write-host "Downloading msys2"
+          Invoke-WebRequest -Uri "${msys2_url}" -OutFile "${env:RUNNER_TEMP}\msys2.exe"
+          write-host "Installing msys2"
+          Start-Process "${env:RUNNER_TEMP}\msys2.exe" -ArgumentList @("in", "--confirm-command", "--accept-messages", "--root", "C:/msys64") -NoNewWindow -Wait
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
        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"
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang", "make") -NoNewWindow -Wait
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: verify tools
+        run: |
+          get-command gcc
+          gcc --version
+          get-command make
+          make --version
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
          cache: true
-      - name: 'Install CUDA ${{ matrix.cuda.version }}'
+      # CUDA installation steps
+      - name: 'Cache CUDA installer'
+        id: cache-cuda
+        uses: actions/cache@v4
+        with:
+          path: cuda-install.exe
+          key: ${{ matrix.cuda.url }}
+      - name: 'Conditionally Download CUDA'
+        if: steps.cache-cuda.outputs.cache-hit != 'true'
        run: |
          $ErrorActionPreference = "Stop"
-          write-host "downloading CUDA Installer"
-          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -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"
+          Invoke-WebRequest -Uri "${{ matrix.cuda.url }}" -OutFile "cuda-install.exe"
+      - name: 'Install CUDA'
+        run: |
+          $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | foreach-object {"${_}_${{ matrix.cuda.version }}"}
+          Start-Process "cuda-install.exe" -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
+      - name: 'Verify CUDA'
+        run: |
+          & (resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0] --version
          $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}" >> $env:GITHUB_ENV
-      - name: 'Verify CUDA'
-        run: nvcc -V
-      - run: go get ./...
-      - name: make
+          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      - name: make cuda runner
        run: |
-          $gopath=(get-command go).source | split-path -parent
-          $cudabin=(get-command nvcc).source | split-path
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$cudabin;$env:PATH"
-          $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores
-      - 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\"
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make cuda_v$(($env:CUDA_PATH | split-path -leaf) -replace 'v(\d+).*', '$1')
      - uses: actions/upload-artifact@v4
        with:
          name: generate-windows-cuda-${{ matrix.cuda.version }}
          path: |
            build/**/*
            dist/windows-amd64/**
-      - uses: actions/upload-artifact@v4
-        with:
-          name: windows-cuda-deps-${{ matrix.cuda.version }}
-          path: dist/deps/*
-

  # windows arm64 generate, go build, and zip file (no installer)
  # Output of this build is aggregated into the final x86 build
@@ -416,7 +374,7 @@ jobs:
          $env:PATH="$gopath;$gccpath;$env:PATH"
          echo $env:PATH
          $env:ARCH="arm64"
-          .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies distZip
+          .\scripts\build_windows.ps1 buildOllama buildApp gatherDependencies sign distZip
        name: 'Windows Build'
      - uses: actions/upload-artifact@v4
        with:
@@ -466,6 +424,24 @@ jobs:
          write-host "Installing plugin"
          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
          write-host "plugin installed"
+      - name: Install msys2
+        run: |
+          $msys2_url="https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-x86_64-20240727.exe"
+          write-host "Downloading msys2"
+          Invoke-WebRequest -Uri "${msys2_url}" -OutFile "${env:RUNNER_TEMP}\msys2.exe"
+          write-host "Installing msys2"
+          Start-Process "${env:RUNNER_TEMP}\msys2.exe" -ArgumentList @("in", "--confirm-command", "--accept-messages", "--root", "C:/msys64") -NoNewWindow -Wait
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang", "make") -NoNewWindow -Wait
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: verify tools
+        run: |
+          get-command gcc
+          gcc --version
+          get-command make
+          make --version
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
@@ -476,19 +452,10 @@ jobs:
          name: generate-windows-cpu
      - uses: actions/download-artifact@v4
        with:
-          name: generate-windows-cuda-11
+          name: generate-windows-cuda-11.3
      - uses: actions/download-artifact@v4
        with:
-          name: generate-windows-cuda-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-11
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-cuda-deps-12
-      - uses: actions/download-artifact@v4
-        with:
-          name: windows-rocm-deps
+          name: generate-windows-cuda-12.4
      - uses: actions/download-artifact@v4
        with:
          name: generate-windows-rocm
@@ -498,12 +465,11 @@ jobs:
          path: dist
      - run: dir build
      - run: |
-          $gopath=(get-command go).source | split-path -parent
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$env:PATH"
          $env:OLLAMA_SKIP_GENERATE="1"
+          $env:ARCH="amd64"
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
          & .\scripts\build_windows.ps1
      - uses: actions/upload-artifact@v4
        with:
@@ -1,5 +1,11 @@
 name: test

+env:
+  ROCM_WINDOWS_URL: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
+  MSYS2_URL: https://github.com/msys2/msys2-installer/releases/download/2024-07-27/msys2-x86_64-20240727.exe
+  CUDA_12_WINDOWS_URL: https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe
+  CUDA_12_WINDOWS_VER: 12.4
+
 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.
@@ -99,30 +105,45 @@ jobs:
        with:
          go-version-file: go.mod
          cache: true
-      - name: 'Install ROCm'
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      # ROCM installation steps
+      - name: 'Cache ROCm installer'
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: rocm-install.exe
+          key: ${{ env.ROCM_WINDOWS_URL }}
+      - name: 'Conditionally Download ROCm'
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
        run: |
          $ErrorActionPreference = "Stop"
-          write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
-          write-host "Installing AMD HIP"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
-          write-host "Completed AMD HIP"
+          Invoke-WebRequest -Uri "${env:ROCM_WINDOWS_URL}" -OutFile "rocm-install.exe"
+      - name: 'Install ROCm'
+        run: |
+          Start-Process "rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
      - name: 'Verify ROCm'
        run: |
          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
-      - run: go get ./...
-      - run: |
-          $gopath=(get-command go).source | split-path -parent
+          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path | select -first 1)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      - name: Add msys paths
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+
+      - name: make rocm runner
+        run: |
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:PATH="$gopath;$env:PATH"
-          $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          write-host $env:HIP_PATH
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
          make -C llama print-HIP_PATH print-HIP_LIB_DIR
-          make -j $cores rocm
-        name: make
+          make rocm

  # CUDA generation step
  runners-windows-cuda:
@@ -135,36 +156,49 @@ jobs:
        with:
          go-version-file: go.mod
          cache: true
-      - name: 'Install CUDA'
+      - name: Set make jobs default
+        run: |
+          echo "MAKEFLAGS=--jobs=$((Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores)" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      # CUDA installation steps
+      - name: 'Cache CUDA installer'
+        id: cache-cuda
+        uses: actions/cache@v4
+        with:
+          path: cuda-install.exe
+          key: ${{ env.CUDA_12_WINDOWS_URL }}
+      - name: 'Conditionally Download CUDA'
+        if: steps.cache-cuda.outputs.cache-hit != 'true'
        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 "${env:RUNNER_TEMP}\cuda-install.exe" -ArgumentList '-s' -NoNewWindow -Wait
-          write-host "Completed CUDA"
+          Invoke-WebRequest -Uri "${env:CUDA_12_WINDOWS_URL}" -OutFile "cuda-install.exe"
+      - name: 'Install CUDA'
+        run: |
+          $subpackages = @("cudart", "nvcc", "cublas", "cublas_dev") | foreach-object {"${_}_${{ env.CUDA_12_WINDOWS_VER }}"}
+          Start-Process "cuda-install.exe" -ArgumentList (@("-s") + $subpackages) -NoNewWindow -Wait
+      - name: 'Verify CUDA'
+        run: |
+          & (resolve-path "c:\Program Files\NVIDIA*\CUDA\v*\bin\nvcc.exe")[0] --version
          $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}" >> $env:GITHUB_ENV
-      - name: 'Verify CUDA'
-        run: nvcc -V
-      - run: go get ./...
-      - name: make
+          echo "$cudaPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_V${cudaVer}=$cudaPath" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          echo "CUDA_PATH_VX_Y=CUDA_PATH_V${cudaVer}" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+
+      - name: Add msys paths
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
+      - name: make cuda runner
        run: |
-          $gopath=(get-command go).source | split-path -parent
-          $cudabin=(get-command nvcc).source | split-path
          import-module 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\Common7\Tools\Microsoft.VisualStudio.DevShell.dll'
          Enter-VsDevShell -vsinstallpath 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise' -skipautomaticlocation -DevCmdArguments '-arch=x64 -no_logo'
-          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
-          $env:PATH="$gopath;$cudabin;$env:PATH"
-          $env:OLLAMA_SKIP_CPU_GENERATE="1"
-          $cores = (Get-ComputerInfo -Property CsProcessors).CsProcessors.NumberOfCores
-          make -j $cores cuda_v11
-        env:
-          OLLAMA_SKIP_CPU_GENERATE: '1'
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
+          make cuda_v$(($env:CUDA_PATH | split-path -leaf) -replace 'v(\d+).*', '$1')

  runners-cpu:
    needs: [changes]
@@ -189,7 +223,15 @@ jobs:
        with:
          go-version-file: go.mod
          cache: true
-      - run: go get ./...
+      - name: Add msys paths
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          echo "c:\msys64\usr\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\msys64\clang64\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+      - name: Install msys2 tools
+        if: ${{ startsWith(matrix.os, 'windows-') }}
+        run: |
+          Start-Process "c:\msys64\usr\bin\pacman.exe" -ArgumentList @("-S", "--noconfirm", "mingw-w64-clang-x86_64-gcc-compat", "mingw-w64-clang-x86_64-clang") -NoNewWindow -Wait
      - name: 'Build Windows Go Runners'
        if: ${{ startsWith(matrix.os, 'windows-') }}
        run: |
@@ -200,6 +242,7 @@ jobs:
          $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
          $env:PATH="$gopath;$gccpath;$env:PATH"
          echo $env:PATH
+          if (!(gcc --version | select-string -quiet clang)) { throw "wrong gcc compiler detected - must be clang" }
          make -j 4      
      - name: 'Build Unix Go Runners'
        if: ${{ ! startsWith(matrix.os, 'windows-') }}
@@ -238,7 +281,7 @@ jobs:
        shell: bash
      - uses: golangci/golangci-lint-action@v6
        with:
-          args: --timeout 8m0s -v
+          args: --timeout 10m0s -v
  test:
    strategy:
      matrix:
@@ -1,20 +1,21 @@
-# Note: once we have fully transitioned to the Go server, this will replace the old Dockerfile at the top of the tree
-ARG GOLANG_VERSION=1.22.5
+ARG GOLANG_VERSION=1.22.8
 ARG CMAKE_VERSION=3.22.1
 ARG CUDA_VERSION_11=11.3.1
 ARG CUDA_V11_ARCHITECTURES="50;52;53;60;61;62;70;72;75;80;86"
 ARG CUDA_VERSION_12=12.4.0
 ARG CUDA_V12_ARCHITECTURES="60;61;62;70;72;75;80;86;87;89;90;90a"
 ARG ROCM_VERSION=6.1.2
+ARG JETPACK_6=r36.2.0
+ARG JETPACK_5=r35.4.1

 ### To create a local image for building linux binaries on mac or windows with efficient incremental builds
 #
-# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile.new --target unified-builder-amd64 .
+# docker build --platform linux/amd64 -t builder-amd64 -f Dockerfile --target unified-builder-amd64 .
 # docker run --platform linux/amd64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-amd64
 #
 ### Then incremental builds will be much faster in this container
 #
-# make -C llama -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
+# make -j 10 && go build -trimpath -o dist/linux-amd64/ollama .
 #
 FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCM_VERSION}-complete AS unified-builder-amd64
 ARG CMAKE_VERSION
@@ -41,7 +42,7 @@ ENTRYPOINT [ "zsh" ]
 ### To create a local image for building linux binaries on mac or linux/arm64 with efficient incremental builds
 # Note: this does not contain jetson variants
 #
-# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile.new --target unified-builder-arm64 .
+# docker build --platform linux/arm64 -t builder-arm64 -f Dockerfile --target unified-builder-arm64 .
 # docker run --platform linux/arm64 --rm -it -v $(pwd):/go/src/github.com/ollama/ollama/ builder-arm64
 #
 FROM --platform=linux/arm64 rockylinux:8 AS unified-builder-arm64
@@ -77,9 +78,9 @@ ARG CUDA_V12_ARCHITECTURES
 ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
    if grep "^flags" /proc/cpuinfo|grep avx>/dev/null; then \
-        make -C llama -j $(expr $(nproc) / 2 ) ; \
+        make -j $(expr $(nproc) / 2 ) ; \
    else \
-        make -C llama -j 5 ; \
+        make -j 5 ; \
    fi

 FROM --platform=linux/arm64 unified-builder-arm64 AS runners-arm64
@@ -91,7 +92,46 @@ ARG CUDA_V11_ARCHITECTURES
 ARG CUDA_V12_ARCHITECTURES
 ARG OLLAMA_FAST_BUILD
 RUN --mount=type=cache,target=/root/.ccache \
-    make -C llama -j 8
+    make -j 5
+
+# Jetsons need to be built in discrete stages
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_5} AS runners-jetpack5-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 cuda_v11 \
+        CUDA_ARCHITECTURES="72;87" \
+        GPU_RUNNER_VARIANT=_jetpack5 \
+        CGO_EXTRA_LDFLAGS_LINUX=-L/usr/local/cuda/lib64/stubs \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ollama/cuda_jetpack5
+
+FROM --platform=linux/arm64 nvcr.io/nvidia/l4t-jetpack:${JETPACK_6} AS runners-jetpack6-arm64
+ARG GOLANG_VERSION
+RUN apt-get update && apt-get install -y git curl ccache && \
+    curl -s -L https://dl.google.com/go/go${GOLANG_VERSION}.linux-arm64.tar.gz | tar xz -C /usr/local && \
+    ln -s /usr/local/go/bin/go /usr/local/bin/go && \
+    ln -s /usr/local/go/bin/gofmt /usr/local/bin/gofmt && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+WORKDIR /go/src/github.com/ollama/ollama/
+COPY . .
+ARG CGO_CFLAGS
+ENV GOARCH arm64
+RUN --mount=type=cache,target=/root/.ccache \
+    make -j 5 cuda_v12 \
+        CUDA_ARCHITECTURES="87" \
+        GPU_RUNNER_VARIANT=_jetpack6 \
+        CGO_EXTRA_LDFLAGS_LINUX=-L/usr/local/cuda/lib64/stubs \
+        DIST_LIB_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama \
+        DIST_GPU_RUNNER_DEPS_DIR=/go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ollama/cuda_jetpack6


 # Intermediate stages used for ./scripts/build_linux.sh
@@ -135,12 +175,20 @@ FROM --platform=linux/arm64 builder-arm64 AS build-arm64
 COPY . .
 COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
 COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/build/ build/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/ dist/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/build/ build/
 ARG GOFLAGS
 ARG CGO_CFLAGS
 RUN --mount=type=cache,target=/root/.ccache \
    go build -trimpath -o dist/linux-arm64/bin/ollama .
 RUN cd dist/linux-$GOARCH && \
    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH.tgz
+RUN cd dist/linux-$GOARCH-jetpack5 && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack5.tgz
+RUN cd dist/linux-$GOARCH-jetpack6 && \
+    tar --exclude runners -cf - . | pigz --best > ../ollama-linux-$GOARCH-jetpack6.tgz

 FROM --platform=linux/amd64 scratch AS dist-amd64
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/ollama-linux-*.tgz /
@@ -181,16 +229,19 @@ RUN rm -rf \
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-amd64
 RUN apt-get update && \
    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
+    apt-get clean && rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
 COPY --from=runners-cuda-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/

 FROM --platform=linux/arm64 ubuntu:22.04 AS runtime-arm64
 RUN apt-get update && \
    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
+    apt-get clean && rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/bin/ /bin/
 COPY --from=runners-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64/lib/ /lib/
+COPY --from=runners-jetpack5-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack5/lib/ /lib/
+COPY --from=runners-jetpack6-arm64 /go/src/github.com/ollama/ollama/dist/linux-arm64-jetpack6/lib/ /lib/
+

 # ROCm libraries larger so we keep it distinct from the CPU/CUDA image
 FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
@@ -199,7 +250,7 @@ FROM --platform=linux/amd64 ubuntu:22.04 AS runtime-rocm
 COPY --from=build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64-rocm/lib/ /lib/
 RUN apt-get update && \
    apt-get install -y ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
+    apt-get clean && rm -rf /var/lib/apt/lists/*
 COPY --from=container-build-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/bin/ /bin/
 COPY --from=runners-rocm-amd64 /go/src/github.com/ollama/ollama/dist/linux-amd64/lib/ /lib/

@@ -12,7 +12,7 @@ Get up and running with large language models.

 [Download](https://ollama.com/download/Ollama-darwin.zip)

-### Windows preview
+### Windows

 [Download](https://ollama.com/download/OllamaSetup.exe)

@@ -47,26 +47,28 @@ Ollama supports a list of models available on [ollama.com/library](https://ollam

 Here are some example models that can be downloaded:

-| Model              | Parameters | Size  | Download                       |
-| ------------------ | ---------- | ----- | ------------------------------ |
-| Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`          |
-| Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`       |
-| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`          |
-| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`      |
-| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`     |
-| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`              |
-| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`       |
-| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`         |
-| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`            |
-| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`        |
-| Mistral            | 7B         | 4.1GB | `ollama run mistral`           |
-| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`         |
-| 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` |
-| LLaVA              | 7B         | 4.5GB | `ollama run llava`             |
-| Solar              | 10.7B      | 6.1GB | `ollama run solar`             |
+| Model              | Parameters | Size  | Download                         |
+| ------------------ | ---------- | ----- | -------------------------------- |
+| Llama 3.2          | 3B         | 2.0GB | `ollama run llama3.2`            |
+| Llama 3.2          | 1B         | 1.3GB | `ollama run llama3.2:1b`         |
+| Llama 3.2 Vision   | 11B        | 7.9GB | `ollama run llama3.2-vision`     |
+| Llama 3.2 Vision   | 90B        | 55GB  | `ollama run llama3.2-vision:90b` |
+| Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`            |
+| Llama 3.1          | 70B        | 40GB  | `ollama run llama3.1:70b`        |
+| Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
+| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`                |
+| Phi 3 Medium       | 14B        | 7.9GB | `ollama run phi3:medium`         |
+| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`           |
+| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`              |
+| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`          |
+| Mistral            | 7B         | 4.1GB | `ollama run mistral`             |
+| Moondream 2        | 1.4B       | 829MB | `ollama run moondream`           |
+| 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`   |
+| LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
+| 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.
@@ -296,7 +298,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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](https://github.com/reid41/QA-Pilot) (Chat with Code Repository)
+- [QA-Pilot](https://github.com/reid41/QA-Pilot) (Interactive chat tool that can leverage Ollama models for rapid understanding and navigation of GitHub code repositories)
 - [ChatOllama](https://github.com/sugarforever/chat-ollama) (Open Source Chatbot based on Ollama with Knowledge Bases)
 - [CRAG Ollama Chat](https://github.com/Nagi-ovo/CRAG-Ollama-Chat) (Simple Web Search with Corrective RAG)
 - [RAGFlow](https://github.com/infiniflow/ragflow) (Open-source Retrieval-Augmented Generation engine based on deep document understanding)
@@ -306,11 +308,17 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Ollama RAG Chatbot](https://github.com/datvodinh/rag-chatbot.git) (Local Chat with multiple PDFs using Ollama and RAG)
 - [BrainSoup](https://www.nurgo-software.com/products/brainsoup) (Flexible native client with RAG & multi-agent automation)
 - [macai](https://github.com/Renset/macai) (macOS client for Ollama, ChatGPT, and other compatible API back-ends)
+- [RWKV-Runner](https://github.com/josStorer/RWKV-Runner) (RWKV offline LLM deployment tool, also usable as a client for ChatGPT and Ollama)
+- [Ollama Grid Search](https://github.com/dezoito/ollama-grid-search) (app to evaluate and compare models)
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
+- [Shinkai Desktop](https://github.com/dcSpark/shinkai-apps) (Two click install Local AI using Ollama + Files + RAG)
 - [AiLama](https://github.com/zeyoyt/ailama) (A Discord User App that allows you to interact with Ollama anywhere in discord )
 - [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama)
+- [R2R](https://github.com/SciPhi-AI/R2R) (Open-source RAG engine)
+- [Ollama-Kis](https://github.com/elearningshow/ollama-kis) (A simple easy to use GUI with sample custom LLM for Drivers Education)
+- [OpenGPA](https://opengpa.org) (Open-source offline-first Enterprise Agentic Application)
 - [Painting Droid](https://github.com/mateuszmigas/painting-droid) (Painting app with AI integrations)
 - [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
 - [AI Studio](https://github.com/MindWorkAI/AI-Studio)
@@ -318,6 +326,8 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LLMStack](https://github.com/trypromptly/LLMStack) (No-code multi-agent framework to build LLM agents and workflows)
 - [BoltAI for Mac](https://boltai.com) (AI Chat Client for Mac)
 - [Harbor](https://github.com/av/harbor) (Containerized LLM Toolkit with Ollama as default backend)
+- [PyGPT](https://github.com/szczyglis-dev/py-gpt) (AI desktop assistant for Linux, Windows and Mac)
+- [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT/blob/master/docs/content/platform/ollama.md) (AutoGPT Ollama integration)
 - [Go-CREW](https://www.jonathanhecl.com/go-crew/) (Powerful Offline RAG in Golang)
 - [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
 - [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
@@ -327,10 +337,35 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
 - [crewAI with Mesop](https://github.com/rapidarchitect/ollama-crew-mesop) (Mesop Web Interface to run crewAI with Ollama)
+- [Tkinter-based client](https://github.com/chyok/ollama-gui) (Python tkinter-based Client for Ollama)
 - [LLMChat](https://github.com/trendy-design/llmchat) (Privacy focused, 100% local, intuitive all-in-one chat interface)
+- [Local Multimodal AI Chat](https://github.com/Leon-Sander/Local-Multimodal-AI-Chat) (Ollama-based LLM Chat with support for multiple features, including PDF RAG, voice chat, image-based interactions, and integration with OpenAI.)
 - [ARGO](https://github.com/xark-argo/argo) (Locally download and run Ollama and Huggingface models with RAG on Mac/Windows/Linux)
+- [OrionChat](https://github.com/EliasPereirah/OrionChat) - OrionChat is a web interface for chatting with different AI providers
 - [G1](https://github.com/bklieger-groq/g1) (Prototype of using prompting strategies to improve the LLM's reasoning through o1-like reasoning chains.)
+- [Web management](https://github.com/lemonit-eric-mao/ollama-web-management) (Web management page)
+- [Promptery](https://github.com/promptery/promptery) (desktop client for Ollama.)
 - [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
+- [SpaceLlama](https://github.com/tcsenpai/spacellama) (Firefox and Chrome extension to quickly summarize web pages with ollama in a sidebar)
+- [YouLama](https://github.com/tcsenpai/youlama) (Webapp to quickly summarize any YouTube video, supporting Invidious as well)
+- [DualMind](https://github.com/tcsenpai/dualmind) (Experimental app allowing two models to talk to each other in the terminal or in a web interface)
+- [ollamarama-matrix](https://github.com/h1ddenpr0cess20/ollamarama-matrix) (Ollama chatbot for the Matrix chat protocol)
+- [ollama-chat-app](https://github.com/anan1213095357/ollama-chat-app) (Flutter-based chat app)
+- [Perfect Memory AI](https://www.perfectmemory.ai/) (Productivity AI assists personalized by what you have seen on your screen, heard and said in the meetings)
+- [Hexabot](https://github.com/hexastack/hexabot) (A conversational AI builder)
+- [Reddit Rate](https://github.com/rapidarchitect/reddit_analyzer) (Search and Rate Reddit topics with a weighted summation)
+- [OpenTalkGpt](https://github.com/adarshM84/OpenTalkGpt) (Chrome Extension to manage open-source models supported by Ollama, create custom models, and chat with models from a user-friendly UI)
+- [VT](https://github.com/vinhnx/vt.ai) (A minimal multimodal AI chat app, with dynamic conversation routing. Supports local models via Ollama)
+- [Nosia](https://github.com/nosia-ai/nosia) (Easy to install and use RAG platform based on Ollama)
+- [Witsy](https://github.com/nbonamy/witsy) (An AI Desktop application avaiable for Mac/Windows/Linux)
+- [Abbey](https://github.com/US-Artificial-Intelligence/abbey) (A configurable AI interface server with notebooks, document storage, and YouTube support)
+- [Minima](https://github.com/dmayboroda/minima) (RAG with on-premises or fully local workflow)
+
+### Cloud
+
+- [Google Cloud](https://cloud.google.com/run/docs/tutorials/gpu-gemma2-with-ollama)
+- [Fly.io](https://fly.io/docs/python/do-more/add-ollama/)
+- [Koyeb](https://www.koyeb.com/deploy/ollama)

 ### Terminal

@@ -346,7 +381,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [shell-pilot](https://github.com/reid41/shell-pilot)(Interact with models via pure shell scripts on Linux or macOS)
 - [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)
@@ -354,11 +389,19 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [tlm](https://github.com/yusufcanb/tlm)
 - [podman-ollama](https://github.com/ericcurtin/podman-ollama)
 - [gollama](https://github.com/sammcj/gollama)
+- [ParLlama](https://github.com/paulrobello/parllama)
 - [Ollama eBook Summary](https://github.com/cognitivetech/ollama-ebook-summary/)
 - [Ollama Mixture of Experts (MOE) in 50 lines of code](https://github.com/rapidarchitect/ollama_moe)
 - [vim-intelligence-bridge](https://github.com/pepo-ec/vim-intelligence-bridge) Simple interaction of "Ollama" with the Vim editor
+- [x-cmd ollama](https://x-cmd.com/mod/ollama)
+- [bb7](https://github.com/drunkwcodes/bb7)
+- [SwollamaCLI](https://github.com/marcusziade/Swollama) bundled with the Swollama Swift package. [Demo](https://github.com/marcusziade/Swollama?tab=readme-ov-file#cli-usage)
+- [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
+- [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
+- [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.

 ### Apple Vision Pro
+
 - [Enchanted](https://github.com/AugustDev/enchanted)

 ### Database
@@ -380,9 +423,11 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LangChain](https://python.langchain.com/docs/integrations/llms/ollama) and [LangChain.js](https://js.langchain.com/docs/integrations/chat/ollama/) with [example](https://js.langchain.com/docs/tutorials/local_rag/)
 - [Firebase Genkit](https://firebase.google.com/docs/genkit/plugins/ollama)
 - [crewAI](https://github.com/crewAIInc/crewAI)
+- [Spring AI](https://github.com/spring-projects/spring-ai) with [reference](https://docs.spring.io/spring-ai/reference/api/chat/ollama-chat.html) and [example](https://github.com/tzolov/ollama-tools)
 - [LangChainGo](https://github.com/tmc/langchaingo/) with [example](https://github.com/tmc/langchaingo/tree/main/examples/ollama-completion-example)
 - [LangChain4j](https://github.com/langchain4j/langchain4j) with [example](https://github.com/langchain4j/langchain4j-examples/tree/main/ollama-examples/src/main/java)
 - [LangChainRust](https://github.com/Abraxas-365/langchain-rust) with [example](https://github.com/Abraxas-365/langchain-rust/blob/main/examples/llm_ollama.rs)
+- [LLPhant](https://github.com/theodo-group/LLPhant?tab=readme-ov-file#ollama)
 - [LlamaIndex](https://docs.llamaindex.ai/en/stable/examples/llm/ollama/) and [LlamaIndexTS](https://ts.llamaindex.ai/modules/llms/available_llms/ollama)
 - [LiteLLM](https://github.com/BerriAI/litellm)
 - [OllamaFarm for Go](https://github.com/presbrey/ollamafarm)
@@ -407,11 +452,20 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Portkey](https://portkey.ai/docs/welcome/integration-guides/ollama)
 - [PromptingTools.jl](https://github.com/svilupp/PromptingTools.jl) with an [example](https://svilupp.github.io/PromptingTools.jl/dev/examples/working_with_ollama)
 - [LlamaScript](https://github.com/Project-Llama/llamascript)
+- [llm-axe](https://github.com/emirsahin1/llm-axe) (Python Toolkit for Building LLM Powered Apps)
 - [Gollm](https://docs.gollm.co/examples/ollama-example)
+- [Gollama for Golang](https://github.com/jonathanhecl/gollama)
 - [Ollamaclient for Golang](https://github.com/xyproto/ollamaclient)
 - [High-level function abstraction in Go](https://gitlab.com/tozd/go/fun)
 - [Ollama PHP](https://github.com/ArdaGnsrn/ollama-php)
 - [Agents-Flex for Java](https://github.com/agents-flex/agents-flex) with [example](https://github.com/agents-flex/agents-flex/tree/main/agents-flex-llm/agents-flex-llm-ollama/src/test/java/com/agentsflex/llm/ollama)
+- [Parakeet](https://github.com/parakeet-nest/parakeet) is a GoLang library, made to simplify the development of small generative AI applications with Ollama.
+- [Haverscript](https://github.com/andygill/haverscript) with [examples](https://github.com/andygill/haverscript/tree/main/examples)
+- [Ollama for Swift](https://github.com/mattt/ollama-swift)
+- [Swollama for Swift](https://github.com/marcusziade/Swollama) with [DocC](https://marcusziade.github.io/Swollama/documentation/swollama/)
+- [GoLamify](https://github.com/prasad89/golamify)
+- [Ollama for Haskell](https://github.com/tusharad/ollama-haskell)
+- [multi-llm-ts](https://github.com/nbonamy/multi-llm-ts) (A Typescript/JavaScript library allowing access to different LLM in unified API)

 ### Mobile

@@ -425,6 +479,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Raycast extension](https://github.com/MassimilianoPasquini97/raycast_ollama)
 - [Discollama](https://github.com/mxyng/discollama) (Discord bot inside the Ollama discord channel)
 - [Continue](https://github.com/continuedev/continue)
+- [Vibe](https://github.com/thewh1teagle/vibe) (Transcribe and analyze meetings with Ollama)
 - [Obsidian Ollama plugin](https://github.com/hinterdupfinger/obsidian-ollama)
 - [Logseq Ollama plugin](https://github.com/omagdy7/ollama-logseq)
 - [NotesOllama](https://github.com/andersrex/notesollama) (Apple Notes Ollama plugin)
@@ -447,14 +502,24 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [ChatGPTBox: All in one browser extension](https://github.com/josStorer/chatGPTBox) with [Integrating Tutorial](https://github.com/josStorer/chatGPTBox/issues/616#issuecomment-1975186467)
 - [Discord AI chat/moderation bot](https://github.com/rapmd73/Companion) Chat/moderation bot written in python. Uses Ollama to create personalities.
 - [Headless Ollama](https://github.com/nischalj10/headless-ollama) (Scripts to automatically install ollama client & models on any OS for apps that depends on ollama server)
- [vnc-lm](https://github.com/jk011ru/vnc-lm) (A containerized Discord bot with support for attachments and web links)
+- [Terraform AWS Ollama & Open WebUI](https://github.com/xuyangbocn/terraform-aws-self-host-llm) (A Terraform module to deploy on AWS a ready-to-use Ollama service, together with its front end Open WebUI service.)
+- [node-red-contrib-ollama](https://github.com/jakubburkiewicz/node-red-contrib-ollama)
+- [Local AI Helper](https://github.com/ivostoykov/localAI) (Chrome and Firefox extensions that enable interactions with the active tab and customisable API endpoints. Includes secure storage for user prompts.)
+- [vnc-lm](https://github.com/jake83741/vnc-lm) (Discord bot for messaging with LLMs through Ollama and LiteLLM. Seamlessly move between local and flagship models.)
 - [LSP-AI](https://github.com/SilasMarvin/lsp-ai) (Open-source language server for AI-powered functionality)
 - [QodeAssist](https://github.com/Palm1r/QodeAssist) (AI-powered coding assistant plugin for Qt Creator)
 - [Obsidian Quiz Generator plugin](https://github.com/ECuiDev/obsidian-quiz-generator)
+- [TextCraft](https://github.com/suncloudsmoon/TextCraft) (Copilot in Word alternative using Ollama)
+- [Alfred Ollama](https://github.com/zeitlings/alfred-ollama) (Alfred Workflow)

 ### Supported backends

 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.

+### Observability
+
+- [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
+- [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production. 
@@ -55,7 +55,7 @@ func checkError(resp *http.Response, body []byte) error {

 // 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
+// port on which the ollama service is listening. The format of this variable
 // is:
 //
 //	<scheme>://<host>:<port>
@@ -12,7 +12,7 @@ import (
 	"time"
 )

-// StatusError is an error with and HTTP status code.
+// StatusError is an error with an HTTP status code and message.
 type StatusError struct {
 	StatusCode   int
 	Status       string
@@ -57,7 +57,7 @@ type GenerateRequest struct {
 	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.
+	// [Client.Generate]. 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.
@@ -90,14 +90,14 @@ type ChatRequest struct {
 	// Messages is the messages of the chat - can be used to keep a chat memory.
 	Messages []Message `json:"messages"`

-	// Stream enable streaming of returned response; true by default.
+	// Stream enables streaming of returned responses; true by default.
 	Stream *bool `json:"stream,omitempty"`

 	// Format is the format to return the response in (e.g. "json").
 	Format string `json:"format"`

 	// KeepAlive controls how long the model will stay loaded into memory
-	// followin the request.
+	// following the request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`

 	// Tools is an optional list of tools the model has access to.
@@ -146,6 +146,7 @@ type ToolCall struct {
 }

 type ToolCallFunction struct {
+	Index     int                       `json:"index,omitempty"`
 	Name      string                    `json:"name"`
 	Arguments ToolCallFunctionArguments `json:"arguments"`
 }
@@ -203,8 +204,8 @@ type Metrics struct {
 	EvalDuration       time.Duration `json:"eval_duration,omitempty"`
 }

-// Options specified in [GenerateRequest], if you add a new option here add it
-// to the API docs also.
+// Options specified in [GenerateRequest].  If you add a new option here, also
+// add it to the API docs.
 type Options struct {
 	Runner

@@ -236,7 +237,7 @@ type Runner struct {
 	NumGPU    int   `json:"num_gpu,omitempty"`
 	MainGPU   int   `json:"main_gpu,omitempty"`
 	LowVRAM   bool  `json:"low_vram,omitempty"`
-	F16KV     bool  `json:"f16_kv,omitempty"`
+	F16KV     bool  `json:"f16_kv,omitempty"` // Deprecated: This option is ignored
 	LogitsAll bool  `json:"logits_all,omitempty"`
 	VocabOnly bool  `json:"vocab_only,omitempty"`
 	UseMMap   *bool `json:"use_mmap,omitempty"`
@@ -613,7 +614,6 @@ func DefaultOptions() Options {
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
 			NumThread: 0,  // let the runtime decide
 			LowVRAM:   false,
-			F16KV:     true,
 			UseMLock:  false,
 			UseMMap:   nil,
 		},
@@ -11,10 +11,12 @@ import (

 	"github.com/ollama/ollama/app/store"
 	"github.com/ollama/ollama/app/tray"
+	"github.com/ollama/ollama/envconfig"
 )

 func Run() {
 	InitLogging()
+	slog.Info("app config", "env", envconfig.Values())

 	ctx, cancel := context.WithCancel(context.Background())
 	var done chan int
@@ -36,8 +36,13 @@ func init() {
 		ServerLogFile = filepath.Join(AppDataDir, "server.log")
 		UpgradeLogFile = filepath.Join(AppDataDir, "upgrade.log")

-		// Executables are stored in APPDATA
-		AppDir = filepath.Join(localAppData, "Programs", "Ollama")
+		exe, err := os.Executable()
+		if err != nil {
+			slog.Warn("error discovering executable directory", "error", err)
+			AppDir = filepath.Join(localAppData, "Programs", "Ollama")
+		} else {
+			AppDir = filepath.Dir(exe)
+		}

 		// Make sure we have PATH set correctly for any spawned children
 		paths := strings.Split(os.Getenv("PATH"), ";")
@@ -64,7 +69,7 @@ func init() {
 		}

 		// Make sure our logging dir exists
-		_, err := os.Stat(AppDataDir)
+		_, err = os.Stat(AppDataDir)
 		if errors.Is(err, os.ErrNotExist) {
 			if err := os.MkdirAll(AppDataDir, 0o755); err != nil {
 				slog.Error(fmt.Sprintf("create ollama dir %s: %v", AppDataDir, err))
@@ -18,11 +18,17 @@ func getCLIFullPath(command string) string {
 	var cmdPath string
 	appExe, err := os.Executable()
 	if err == nil {
+		// Check both the same location as the tray app, as well as ./bin
 		cmdPath = filepath.Join(filepath.Dir(appExe), command)
 		_, err := os.Stat(cmdPath)
 		if err == nil {
 			return cmdPath
 		}
+		cmdPath = filepath.Join(filepath.Dir(appExe), "bin", command)
+		_, err = os.Stat(cmdPath)
+		if err == nil {
+			return cmdPath
+		}
 	}
 	cmdPath, err = exec.LookPath(command)
 	if err == nil {
@@ -26,19 +26,15 @@ func DoUpgrade(cancel context.CancelFunc, done chan int) error {
 	slog.Info("starting upgrade with " + installerExe)
 	slog.Info("upgrade log file " + UpgradeLogFile)

-	// When running in debug mode, we'll be "verbose" and let the installer pop up and prompt
+	// make the upgrade show progress, but non interactive
 	installArgs := []string{
 		"/CLOSEAPPLICATIONS",                    // Quit the tray app if it's still running
 		"/LOG=" + filepath.Base(UpgradeLogFile), // Only relative seems reliable, so set pwd
 		"/FORCECLOSEAPPLICATIONS",               // Force close the tray app - might be needed
-	}
-	// make the upgrade as quiet as possible (no GUI, no prompts)
-	installArgs = append(installArgs,
-		"/SP", // Skip the "This will install... Do you wish to continue" prompt
-		"/SUPPRESSMSGBOXES",
+		"/SP",                                   // Skip the "This will install... Do you wish to continue" prompt
+		"/NOCANCEL",                             // Disable the ability to cancel upgrade mid-flight to avoid partially installed upgrades
 		"/SILENT",
-		"/VERYSILENT",
-	)
+	}

 	// Safeguard in case we have requests in flight that need to drain...
 	slog.Info("Waiting for server to shutdown")
@@ -53,8 +53,8 @@ RestartIfNeededByRun=no
 ; https://jrsoftware.org/ishelp/index.php?topic=setup_wizardimagefile
 WizardSmallImageFile=.\assets\setup.bmp

-; TODO verifty actual min windows version...
-; OG Win 10
+; Ollama requires Windows 10 22H2 or newer for proper unicode rendering
+; TODO: consider setting this to 10.0.19045
 MinVersion=10.0.10240

 ; First release that supports WinRT UI Composition for win32 apps
@@ -136,7 +136,7 @@ Type: filesandordirs; Name: "{%TEMP}\ollama*"
 Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"

 [Messages]
-WizardReady=Ollama Windows Preview
+WizardReady=Ollama
 ReadyLabel1=%nLet's get you up and running with your own large language models.
 SetupAppRunningError=Another Ollama installer is running.%n%nPlease cancel or finish the other installer, then click OK to continue with this install, or Cancel to exit.

@@ -64,7 +64,7 @@ func initStore() {
 		slog.Debug(fmt.Sprintf("unexpected error searching for store: %s", err))
 	}
 	slog.Debug("initializing new store")
-	store.ID = uuid.New().String()
+	store.ID = uuid.NewString()
 	writeStore(getStorePath())
 }

@@ -11,12 +11,13 @@ import (
 )

 const (
-	updateAvailableMenuID = 1
-	updateMenuID          = updateAvailableMenuID + 1
-	separatorMenuID       = updateMenuID + 1
-	diagLogsMenuID        = separatorMenuID + 1
-	diagSeparatorMenuID   = diagLogsMenuID + 1
-	quitMenuID            = diagSeparatorMenuID + 1
+	_ = iota
+	updateAvailableMenuID
+	updateMenuID
+	separatorMenuID
+	diagLogsMenuID
+	diagSeparatorMenuID
+	quitMenuID
 )

 func (t *winTray) initMenus() error {
@@ -38,7 +39,7 @@ func (t *winTray) UpdateAvailable(ver string) error {
 		if err := t.addOrUpdateMenuItem(updateAvailableMenuID, 0, updateAvailableMenuTitle, true); err != nil {
 			return fmt.Errorf("unable to create menu entries %w", err)
 		}
-		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenutTitle, false); err != nil {
+		if err := t.addOrUpdateMenuItem(updateMenuID, 0, updateMenuTitle, false); err != nil {
 			return fmt.Errorf("unable to create menu entries %w", err)
 		}
 		if err := t.addSeparatorMenuItem(separatorMenuID, 0); err != nil {
@@ -10,6 +10,6 @@ const (

 	quitMenuTitle            = "Quit Ollama"
 	updateAvailableMenuTitle = "An update is available"
-	updateMenutTitle         = "Restart to update"
+	updateMenuTitle          = "Restart to update"
 	diagLogsMenuTitle        = "View logs"
 )
@@ -361,7 +361,7 @@ func (t *winTray) showMenu() error {

 	boolRet, _, err = pTrackPopupMenu.Call(
 		uintptr(t.menus[0]),
-		TPM_BOTTOMALIGN|TPM_LEFTALIGN,
+		TPM_BOTTOMALIGN|TPM_LEFTALIGN|TPM_RIGHTBUTTON,
 		uintptr(p.X),
 		uintptr(p.Y),
 		0,
@@ -67,6 +67,7 @@ const (
 	SW_HIDE             = 0
 	TPM_BOTTOMALIGN     = 0x0020
 	TPM_LEFTALIGN       = 0x0000
+	TPM_RIGHTBUTTON     = 0x0002
 	WM_CLOSE            = 0x0010
 	WM_USER             = 0x0400
 	WS_CAPTION          = 0x00C00000
@@ -19,9 +19,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
-	"regexp"
 	"runtime"
-	"slices"
 	"strconv"
 	"strings"
 	"sync/atomic"
@@ -36,39 +34,67 @@ import (
 	"golang.org/x/term"

 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/auth"
 	"github.com/ollama/ollama/envconfig"
 	"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"
 )

+var (
+	errModelNotFound     = errors.New("no Modelfile or safetensors files found")
+	errModelfileNotFound = errors.New("specified Modelfile wasn't found")
+)
+
+func getModelfileName(cmd *cobra.Command) (string, error) {
+	fn, _ := cmd.Flags().GetString("file")
+
+	filename := fn
+	if filename == "" {
+		filename = "Modelfile"
+	}
+
+	absName, err := filepath.Abs(filename)
+	if err != nil {
+		return "", err
+	}
+
+	_, err = os.Stat(absName)
+	if err != nil {
+		return fn, err
+	}
+
+	return absName, nil
+}
+
 func CreateHandler(cmd *cobra.Command, args []string) error {
-	filename, _ := cmd.Flags().GetString("file")
-	filename, err := filepath.Abs(filename)
-	if err != nil {
-		return err
-	}
-
-	client, err := api.ClientFromEnvironment()
-	if err != nil {
-		return err
-	}
-
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()

-	f, err := os.Open(filename)
-	if err != nil {
-		return err
-	}
-	defer f.Close()
+	var reader io.Reader

-	modelfile, err := parser.ParseFile(f)
+	filename, err := getModelfileName(cmd)
+	if os.IsNotExist(err) {
+		if filename == "" {
+			reader = strings.NewReader("FROM .\n")
+		} else {
+			return errModelfileNotFound
+		}
+	} else if err != nil {
+		return err
+	} else {
+		f, err := os.Open(filename)
+		if err != nil {
+			return err
+		}
+
+		reader = f
+		defer f.Close()
+	}
+
+	modelfile, err := parser.ParseFile(reader)
 	if err != nil {
 		return err
 	}
@@ -83,6 +109,11 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p.Add(status, spinner)
 	defer p.Stop()

+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
 	for i := range modelfile.Commands {
 		switch modelfile.Commands[i].Name {
 		case "model", "adapter":
@@ -221,7 +252,7 @@ func tempZipFiles(path string) (string, error) {
 		// covers consolidated.x.pth, consolidated.pth
 		files = append(files, pt...)
 	} else {
-		return "", errors.New("no safetensors or torch files found")
+		return "", errModelNotFound
 	}

 	// add configuration files, json files are detected as text/plain
@@ -422,6 +453,10 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	if len(prompts) > 0 {
 		interactive = false
 	}
+	// Be quiet if we're redirecting to a pipe or file
+	if !term.IsTerminal(int(os.Stdout.Fd())) {
+		interactive = false
+	}

 	nowrap, err := cmd.Flags().GetBool("nowordwrap")
 	if err != nil {
@@ -453,7 +488,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

-	opts.MultiModal = slices.Contains(info.Details.Families, "clip")
+	opts.MultiModal = len(info.ProjectorInfo) != 0
 	opts.ParentModel = info.Details.ParentModel

 	if interactive {
@@ -478,47 +513,6 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	return generate(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 {
@@ -565,6 +559,8 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 	}

 	request := api.PushRequest{Name: args[0], Insecure: insecure}
+
+	n := model.ParseName(args[0])
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
 		if spinner != nil {
 			spinner.Stop()
@@ -572,18 +568,19 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		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
 	}

+	p.Stop()
 	spinner.Stop()
+
+	destination := n.String()
+	if strings.HasSuffix(n.Host, ".ollama.ai") || strings.HasSuffix(n.Host, ".ollama.com") {
+		destination = "https://ollama.com/" + strings.TrimSuffix(n.DisplayShortest(), ":latest")
+	}
+	fmt.Printf("\nYou can find your model at:\n\n")
+	fmt.Printf("\t%s\n", destination)
+
 	return nil
 }

@@ -766,9 +763,9 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		case "parameters":
 			fmt.Println(resp.Parameters)
 		case "system":
-			fmt.Println(resp.System)
+			fmt.Print(resp.System)
 		case "template":
-			fmt.Println(resp.Template)
+			fmt.Print(resp.Template)
 		}

 		return nil
@@ -1284,7 +1281,7 @@ func NewCLI() *cobra.Command {
 	log.SetFlags(log.LstdFlags | log.Lshortfile)
 	cobra.EnableCommandSorting = false

-	if runtime.GOOS == "windows" {
+	if runtime.GOOS == "windows" && term.IsTerminal(int(os.Stdout.Fd())) {
 		console.ConsoleFromFile(os.Stdin) //nolint:errcheck
 	}

@@ -1316,7 +1313,7 @@ func NewCLI() *cobra.Command {
 		RunE:    CreateHandler,
 	}

-	createCmd.Flags().StringP("file", "f", "Modelfile", "Name of the Modelfile")
+	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\"")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_0)")

 	showCmd := &cobra.Command{
@@ -1448,6 +1445,7 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_SCHED_SPREAD"],
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
+				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
@@ -4,10 +4,10 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
-	"path/filepath"
 	"strings"
 	"testing"

@@ -179,18 +179,14 @@ Weigh anchor!

 	t.Run("license", func(t *testing.T) {
 		var b bytes.Buffer
-		license, err := os.ReadFile(filepath.Join("..", "LICENSE"))
-		if err != nil {
-			t.Fatal(err)
-		}
-
+		license := "MIT License\nCopyright (c) Ollama\n"
 		if err := showInfo(&api.ShowResponse{
 			Details: api.ModelDetails{
 				Family:            "test",
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-			License: string(license),
+			License: license,
 		}, &b); err != nil {
 			t.Fatal(err)
 		}
@@ -270,3 +266,226 @@ func TestDeleteHandler(t *testing.T) {
 		t.Fatalf("DeleteHandler failed: expected error about stopping non-existent model, got %v", err)
 	}
 }
+
+func TestGetModelfileName(t *testing.T) {
+	tests := []struct {
+		name          string
+		modelfileName string
+		fileExists    bool
+		expectedName  string
+		expectedErr   error
+	}{
+		{
+			name:          "no modelfile specified, no modelfile exists",
+			modelfileName: "",
+			fileExists:    false,
+			expectedName:  "",
+			expectedErr:   os.ErrNotExist,
+		},
+		{
+			name:          "no modelfile specified, modelfile exists",
+			modelfileName: "",
+			fileExists:    true,
+			expectedName:  "Modelfile",
+			expectedErr:   nil,
+		},
+		{
+			name:          "modelfile specified, no modelfile exists",
+			modelfileName: "crazyfile",
+			fileExists:    false,
+			expectedName:  "crazyfile",
+			expectedErr:   os.ErrNotExist,
+		},
+		{
+			name:          "modelfile specified, modelfile exists",
+			modelfileName: "anotherfile",
+			fileExists:    true,
+			expectedName:  "anotherfile",
+			expectedErr:   nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cmd := &cobra.Command{
+				Use: "fakecmd",
+			}
+			cmd.Flags().String("file", "", "path to modelfile")
+
+			var expectedFilename string
+
+			if tt.fileExists {
+				tempDir, err := os.MkdirTemp("", "modelfiledir")
+				defer os.RemoveAll(tempDir)
+				if err != nil {
+					t.Fatalf("temp modelfile dir creation failed: %v", err)
+				}
+				var fn string
+				if tt.modelfileName != "" {
+					fn = tt.modelfileName
+				} else {
+					fn = "Modelfile"
+				}
+
+				tempFile, err := os.CreateTemp(tempDir, fn)
+				if err != nil {
+					t.Fatalf("temp modelfile creation failed: %v", err)
+				}
+
+				expectedFilename = tempFile.Name()
+				err = cmd.Flags().Set("file", expectedFilename)
+				if err != nil {
+					t.Fatalf("couldn't set file flag: %v", err)
+				}
+			} else {
+				if tt.modelfileName != "" {
+					expectedFilename = tt.modelfileName
+					err := cmd.Flags().Set("file", tt.modelfileName)
+					if err != nil {
+						t.Fatalf("couldn't set file flag: %v", err)
+					}
+				}
+			}
+
+			actualFilename, actualErr := getModelfileName(cmd)
+
+			if actualFilename != expectedFilename {
+				t.Errorf("expected filename: '%s' actual filename: '%s'", expectedFilename, actualFilename)
+			}
+
+			if tt.expectedErr != os.ErrNotExist {
+				if actualErr != tt.expectedErr {
+					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
+				}
+			} else {
+				if !os.IsNotExist(actualErr) {
+					t.Errorf("expected err: %v actual err: %v", tt.expectedErr, actualErr)
+				}
+			}
+		})
+	}
+}
+
+func TestPushHandler(t *testing.T) {
+	tests := []struct {
+		name           string
+		modelName      string
+		serverResponse map[string]func(w http.ResponseWriter, r *http.Request)
+		expectedError  string
+		expectedOutput string
+	}{
+		{
+			name:      "successful push",
+			modelName: "test-model",
+			serverResponse: map[string]func(w http.ResponseWriter, r *http.Request){
+				"/api/push": func(w http.ResponseWriter, r *http.Request) {
+					if r.Method != http.MethodPost {
+						t.Errorf("expected POST request, got %s", r.Method)
+					}
+
+					var req api.PushRequest
+					if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+						http.Error(w, err.Error(), http.StatusBadRequest)
+						return
+					}
+
+					if req.Name != "test-model" {
+						t.Errorf("expected model name 'test-model', got %s", req.Name)
+					}
+
+					// Simulate progress updates
+					responses := []api.ProgressResponse{
+						{Status: "preparing manifest"},
+						{Digest: "sha256:abc123456789", Total: 100, Completed: 50},
+						{Digest: "sha256:abc123456789", Total: 100, Completed: 100},
+					}
+
+					for _, resp := range responses {
+						if err := json.NewEncoder(w).Encode(resp); err != nil {
+							http.Error(w, err.Error(), http.StatusInternalServerError)
+							return
+						}
+						w.(http.Flusher).Flush()
+					}
+				},
+			},
+			expectedOutput: "\nYou can find your model at:\n\n\thttps://ollama.com/test-model\n",
+		},
+		{
+			name:      "unauthorized push",
+			modelName: "unauthorized-model",
+			serverResponse: map[string]func(w http.ResponseWriter, r *http.Request){
+				"/api/push": func(w http.ResponseWriter, r *http.Request) {
+					w.Header().Set("Content-Type", "application/json")
+					w.WriteHeader(http.StatusUnauthorized)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error": "access denied",
+					})
+					if err != nil {
+						t.Fatal(err)
+					}
+				},
+			},
+			expectedError: "you are not authorized to push to this namespace, create the model under a namespace you own",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mockServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if handler, ok := tt.serverResponse[r.URL.Path]; ok {
+					handler(w, r)
+					return
+				}
+				http.Error(w, "not found", http.StatusNotFound)
+			}))
+			defer mockServer.Close()
+
+			t.Setenv("OLLAMA_HOST", mockServer.URL)
+
+			cmd := &cobra.Command{}
+			cmd.Flags().Bool("insecure", false, "")
+			cmd.SetContext(context.TODO())
+
+			// Redirect stderr to capture progress output
+			oldStderr := os.Stderr
+			r, w, _ := os.Pipe()
+			os.Stderr = w
+
+			// Capture stdout for the "Model pushed" message
+			oldStdout := os.Stdout
+			outR, outW, _ := os.Pipe()
+			os.Stdout = outW
+
+			err := PushHandler(cmd, []string{tt.modelName})
+
+			// Restore stderr
+			w.Close()
+			os.Stderr = oldStderr
+			// drain the pipe
+			if _, err := io.ReadAll(r); err != nil {
+				t.Fatal(err)
+			}
+
+			// Restore stdout and get output
+			outW.Close()
+			os.Stdout = oldStdout
+			stdout, _ := io.ReadAll(outR)
+
+			if tt.expectedError == "" {
+				if err != nil {
+					t.Errorf("expected no error, got %v", err)
+				}
+				if tt.expectedOutput != "" {
+					if got := string(stdout); got != tt.expectedOutput {
+						t.Errorf("expected output %q, got %q", tt.expectedOutput, got)
+					}
+				}
+			} else {
+				if err == nil || !strings.Contains(err.Error(), tt.expectedError) {
+					t.Errorf("expected error containing %q, got %v", tt.expectedError, err)
+				}
+			}
+		})
+	}
+}
@@ -319,8 +319,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 						opts.Messages = append(opts.Messages, newMessage)
 					}
 					fmt.Println("Set system message.")
-					sb.Reset()
-
 					sb.Reset()
 					continue
 				default:
@@ -494,35 +492,29 @@ func buildModelfile(opts runOptions) string {
 }

 func normalizeFilePath(fp string) string {
-	// Define a map of escaped characters and their replacements
-	replacements := map[string]string{
-		"\\ ":  " ",  // Escaped space
-		"\\(":  "(",  // Escaped left parenthesis
-		"\\)":  ")",  // Escaped right parenthesis
-		"\\[":  "[",  // Escaped left square bracket
-		"\\]":  "]",  // Escaped right square bracket
-		"\\{":  "{",  // Escaped left curly brace
-		"\\}":  "}",  // Escaped right curly brace
-		"\\$":  "$",  // Escaped dollar sign
-		"\\&":  "&",  // Escaped ampersand
-		"\\;":  ";",  // Escaped semicolon
-		"\\'":  "'",  // Escaped single quote
-		"\\\\": "\\", // Escaped backslash
-		"\\*":  "*",  // Escaped asterisk
-		"\\?":  "?",  // Escaped question mark
-	}
-
-	for escaped, actual := range replacements {
-		fp = strings.ReplaceAll(fp, escaped, actual)
-	}
-	return fp
+	return strings.NewReplacer(
+		"\\ ", " ", // Escaped space
+		"\\(", "(", // Escaped left parenthesis
+		"\\)", ")", // Escaped right parenthesis
+		"\\[", "[", // Escaped left square bracket
+		"\\]", "]", // Escaped right square bracket
+		"\\{", "{", // Escaped left curly brace
+		"\\}", "}", // Escaped right curly brace
+		"\\$", "$", // Escaped dollar sign
+		"\\&", "&", // Escaped ampersand
+		"\\;", ";", // Escaped semicolon
+		"\\'", "'", // Escaped single quote
+		"\\\\", "\\", // Escaped backslash
+		"\\*", "*", // Escaped asterisk
+		"\\?", "?", // Escaped question mark
+	).Replace(fp)
 }

 func extractFileNames(input string) []string {
 	// Regex to match file paths starting with optional drive letter, / ./ \ or .\ and include escaped or unescaped spaces (\ or %20)
 	// and followed by more characters and a file extension
 	// This will capture non filename strings, but we'll check for file existence to remove mismatches
-	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png|svg)\b`
+	regexPattern := `(?:[a-zA-Z]:)?(?:\./|/|\\)[\S\\ ]+?\.(?i:jpg|jpeg|png)\b`
 	re := regexp.MustCompile(regexPattern)

 	return re.FindAllString(input, -1)
@@ -535,10 +527,9 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 	for _, fp := range filePaths {
 		nfp := normalizeFilePath(fp)
 		data, err := getImageData(nfp)
-		if err != nil {
-			if os.IsNotExist(err) {
-				continue
-			}
+		if errors.Is(err, os.ErrNotExist) {
+			continue
+		} else if err != nil {
 			fmt.Fprintf(os.Stderr, "Couldn't process image: %q\n", err)
 			return "", imgs, err
 		}
@@ -546,7 +537,7 @@ func extractFileData(input string) (string, []api.ImageData, error) {
 		input = strings.ReplaceAll(input, fp, "")
 		imgs = append(imgs, data)
 	}
-	return input, imgs, nil
+	return strings.TrimSpace(input), imgs, nil
 }

 func getImageData(filePath string) ([]byte, error) {
@@ -12,44 +12,45 @@ import (
 func TestExtractFilenames(t *testing.T) {
 	// Unix style paths
 	input := ` some preamble 
- ./relative\ path/one.png inbetween1 ./not a valid two.jpg inbetween2
-/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.svg`
+ ./relative\ path/one.png inbetween1 ./not a valid two.jpg inbetween2 ./1.svg
+/unescaped space /three.jpeg inbetween3 /valid\ path/dir/four.png "./quoted with spaces/five.JPG`
 	res := extractFileNames(input)
 	assert.Len(t, res, 5)
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[1], "two.jpg")
 	assert.Contains(t, res[2], "three.jpeg")
 	assert.Contains(t, res[3], "four.png")
-	assert.Contains(t, res[4], "five.svg")
+	assert.Contains(t, res[4], "five.JPG")
 	assert.NotContains(t, res[4], '"')
-	assert.NotContains(t, res, "inbtween")
+	assert.NotContains(t, res, "inbetween1")
+	assert.NotContains(t, res, "./1.svg")

 	// Windows style paths
 	input = ` some preamble
 c:/users/jdoe/one.png inbetween1 c:/program files/someplace/two.jpg inbetween2 
 /absolute/nospace/three.jpeg inbetween3 /absolute/with space/four.png inbetween4
-./relative\ path/five.svg inbetween5 "./relative with/spaces/six.png inbetween6
-d:\path with\spaces\seven.svg inbetween7 c:\users\jdoe\eight.png inbetween8 
- d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.svg some ending
+./relative\ path/five.JPG inbetween5 "./relative with/spaces/six.png inbetween6
+d:\path with\spaces\seven.JPEG inbetween7 c:\users\jdoe\eight.png inbetween8 
+ d:\program files\someplace\nine.png inbetween9 "E:\program files\someplace\ten.PNG some ending
 `
 	res = extractFileNames(input)
 	assert.Len(t, res, 10)
-	assert.NotContains(t, res, "inbtween")
+	assert.NotContains(t, res, "inbetween2")
 	assert.Contains(t, res[0], "one.png")
 	assert.Contains(t, res[0], "c:")
 	assert.Contains(t, res[1], "two.jpg")
 	assert.Contains(t, res[1], "c:")
 	assert.Contains(t, res[2], "three.jpeg")
 	assert.Contains(t, res[3], "four.png")
-	assert.Contains(t, res[4], "five.svg")
+	assert.Contains(t, res[4], "five.JPG")
 	assert.Contains(t, res[5], "six.png")
-	assert.Contains(t, res[6], "seven.svg")
+	assert.Contains(t, res[6], "seven.JPEG")
 	assert.Contains(t, res[6], "d:")
 	assert.Contains(t, res[7], "eight.png")
 	assert.Contains(t, res[7], "c:")
 	assert.Contains(t, res[8], "nine.png")
 	assert.Contains(t, res[8], "d:")
-	assert.Contains(t, res[9], "ten.svg")
+	assert.Contains(t, res[9], "ten.PNG")
 	assert.Contains(t, res[9], "E:")
 }

@@ -9,7 +9,7 @@ import (
 	"log/slog"
 	"strings"

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

 type ModelParameters struct {
@@ -27,8 +27,8 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }

-func (ModelParameters) KV(t *Tokenizer) fileutils.KV {
-	kv := fileutils.KV{
+func (ModelParameters) KV(t *Tokenizer) llm.KV {
+	kv := llm.KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -54,7 +54,7 @@ func (ModelParameters) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p AdapterParameters) KV() fileutils.KV {
+func (p AdapterParameters) KV() llm.KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -62,7 +62,7 @@ func (p AdapterParameters) KV() fileutils.KV {
 		alpha = p.LoraParameters.Alpha
 	}

-	kv := fileutils.KV{
+	kv := llm.KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -79,19 +79,19 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }

-func (ModelParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
-	return fileutils.WriteGGUF(ws, kv, ts)
+func (ModelParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }

-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv fileutils.KV, ts []fileutils.Tensor) error {
-	return fileutils.WriteGGUF(ws, kv, ts)
+func (AdapterParameters) writeFile(ws io.WriteSeeker, kv llm.KV, ts []llm.Tensor) error {
+	return llm.WriteGGUF(ws, kv, ts)
 }

 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) fileutils.KV
+	KV(*Tokenizer) llm.KV
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
-	Tensors([]Tensor) []fileutils.Tensor
+	Tensors([]Tensor) []llm.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
@@ -99,7 +99,7 @@ type ModelConverter interface {
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
 	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, fileutils.KV, []fileutils.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }

 type moreParser interface {
@@ -108,17 +108,17 @@ type moreParser interface {

 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(fileutils.KV) fileutils.KV
+	KV(llm.KV) llm.KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
-	Tensors([]Tensor) []fileutils.Tensor
+	Tensors([]Tensor) []llm.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string

-	writeFile(io.WriteSeeker, fileutils.KV, []fileutils.Tensor) error
+	writeFile(io.WriteSeeker, llm.KV, []llm.Tensor) error
 }

-func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV fileutils.KV) error {
+func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV llm.KV) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -8,7 +8,7 @@ import (
 	"slices"
 	"strings"

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

 type bertModel struct {
@@ -85,7 +85,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }

-func (p *bertModel) KV(t *Tokenizer) fileutils.KV {
+func (p *bertModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false
@@ -132,8 +132,8 @@ func (p *bertModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p *bertModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *bertModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		if slices.Contains([]string{
 			"embeddings.position_ids",
@@ -143,7 +143,7 @@ func (p *bertModel) Tensors(ts []Tensor) []fileutils.Tensor {
 			continue
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"

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

 type gemmaModel struct {
@@ -23,7 +23,7 @@ type gemmaModel struct {

 var _ ModelConverter = (*gemmaModel)(nil)

-func (p *gemmaModel) KV(t *Tokenizer) fileutils.KV {
+func (p *gemmaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings
@@ -42,14 +42,14 @@ func (p *gemmaModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p *gemmaModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *gemmaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		if strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -1,7 +1,7 @@
 package convert

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

 type gemma2Model struct {
@@ -11,7 +11,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }

-func (p *gemma2Model) KV(t *Tokenizer) fileutils.KV {
+func (p *gemma2Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings
@@ -6,7 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"

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

 type gemma2Adapter struct {
@@ -15,14 +15,14 @@ type gemma2Adapter struct {

 var _ AdapterConverter = (*gemma2Adapter)(nil)

-func (p *gemma2Adapter) KV(baseKV fileutils.KV) fileutils.KV {
+func (p *gemma2Adapter) KV(baseKV llm.KV) llm.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv
 }

-func (p *gemma2Adapter) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *gemma2Adapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -31,7 +31,7 @@ func (p *gemma2Adapter) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -9,7 +9,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"

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

 type llamaModel struct {
@@ -46,7 +46,7 @@ type llamaModel struct {

 var _ ModelConverter = (*llamaModel)(nil)

-func (p *llamaModel) KV(t *Tokenizer) fileutils.KV {
+func (p *llamaModel) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize
@@ -120,11 +120,11 @@ func (p *llamaModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p *llamaModel) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *llamaModel) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor

 	if p.RopeScaling.factors != nil {
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     "rope_freqs.weight",
 			Kind:     0,
 			Shape:    []uint64{uint64(len(p.RopeScaling.factors))},
@@ -138,7 +138,7 @@ func (p *llamaModel) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -7,7 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"

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

 type llamaAdapter struct {
@@ -18,7 +18,7 @@ type llamaAdapter struct {

 var _ AdapterConverter = (*llamaAdapter)(nil)

-func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
+func (p *llamaAdapter) KV(baseKV llm.KV) llm.KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
 	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
@@ -29,8 +29,8 @@ func (p *llamaAdapter) KV(baseKV fileutils.KV) fileutils.KV {
 	return kv
 }

-func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
-	var out []fileutils.Tensor
+func (p *llamaAdapter) Tensors(ts []Tensor) []llm.Tensor {
+	var out []llm.Tensor
 	for _, t := range ts {
 		shape := t.Shape()
 		if (strings.HasSuffix(t.Name(), "weight.lora_a") && shape[0] > shape[1]) ||
@@ -41,7 +41,7 @@ func (p *llamaAdapter) Tensors(ts []Tensor) []fileutils.Tensor {
 			t.SetRepacker(p.repack)
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    shape,
@@ -6,7 +6,7 @@ import (
 	"slices"
 	"strings"

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

 type mixtralModel struct {
@@ -15,7 +15,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }

-func (p *mixtralModel) KV(t *Tokenizer) fileutils.KV {
+func (p *mixtralModel) KV(t *Tokenizer) llm.KV {
 	kv := p.llamaModel.KV(t)

 	if p.NumLocalExperts > 0 {
@@ -29,7 +29,7 @@ func (p *mixtralModel) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p *mixtralModel) Tensors(ts []Tensor) []fileutils.Tensor {
+func (p *mixtralModel) Tensors(ts []Tensor) []llm.Tensor {
 	oldnew := []string{
 		"model.layers", "blk",
 		"w1", "ffn_gate_exps",
@@ -56,10 +56,10 @@ func (p *mixtralModel) Tensors(ts []Tensor) []fileutils.Tensor {
 		return true
 	})

-	var out []fileutils.Tensor
+	var out []llm.Tensor
 	for n, e := range experts {
 		// TODO(mxyng): sanity check experts
-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     n,
 			Kind:     e[0].Kind(),
 			Shape:    append([]uint64{uint64(len(e))}, e[0].Shape()...),
@@ -8,7 +8,7 @@ import (
 	"strings"
 	"sync"

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

 type phi3Model struct {
@@ -37,7 +37,7 @@ type phi3Model struct {

 var _ ModelConverter = (*phi3Model)(nil)

-func (p *phi3Model) KV(t *Tokenizer) fileutils.KV {
+func (p *phi3Model) KV(t *Tokenizer) llm.KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings
@@ -68,19 +68,19 @@ func (p *phi3Model) KV(t *Tokenizer) fileutils.KV {
 	return kv
 }

-func (p *phi3Model) Tensors(ts []Tensor) []fileutils.Tensor {
+func (p *phi3Model) Tensors(ts []Tensor) []llm.Tensor {
 	var addRopeFactors sync.Once

-	out := make([]fileutils.Tensor, 0, len(ts)+2)
+	out := make([]llm.Tensor, 0, len(ts)+2)
 	for _, t := range ts {
 		if strings.HasPrefix(t.Name(), "blk.0.") {
 			addRopeFactors.Do(func() {
-				out = append(out, fileutils.Tensor{
+				out = append(out, llm.Tensor{
 					Name:     "rope_factors_long.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.LongFactor))},
 					WriterTo: p.RopeScaling.LongFactor,
-				}, fileutils.Tensor{
+				}, llm.Tensor{
 					Name:     "rope_factors_short.weight",
 					Kind:     0,
 					Shape:    []uint64{uint64(len(p.RopeScaling.ShortFactor))},
@@ -89,7 +89,7 @@ func (p *phi3Model) Tensors(ts []Tensor) []fileutils.Tensor {
 			})
 		}

-		out = append(out, fileutils.Tensor{
+		out = append(out, llm.Tensor{
 			Name:     t.Name(),
 			Kind:     t.Kind(),
 			Shape:    t.Shape(),
@@ -20,7 +20,7 @@ import (

 	"golang.org/x/exp/maps"

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

 type tensorData struct {
@@ -29,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }

-func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, fileutils.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, llm.KV, *llm.Tensors) {
 	t.Helper()

 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -48,7 +48,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, fileutils.Te
 	}
 	t.Cleanup(func() { r.Close() })

-	m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
+	m, _, err := llm.DecodeGGML(r, math.MaxInt)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -60,7 +60,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, fileutils.KV, fileutils.Te
 	return r, m.KV(), m.Tensors()
 }

-func generateResultsJSON(t *testing.T, f *os.File, kv fileutils.KV, tensors fileutils.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv llm.KV, tensors *llm.Tensors) map[string]string {
 	actual := make(map[string]string)
 	for k, v := range kv {
 		if s, ok := v.(json.Marshaler); !ok {
@@ -330,7 +330,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()

-			m, _, err := fileutils.DecodeGGML(r, math.MaxInt)
+			m, _, err := llm.DecodeGGML(r, math.MaxInt)
 			if err != nil {
 				t.Fatal(err)
 			}
@@ -10,6 +10,7 @@ import (
 	"log/slog"
 	"os"
 	"slices"
+	"strings"

 	"golang.org/x/exp/maps"
 )
@@ -60,7 +61,25 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 			addedTokens[t.Content] = t
 		}

-		t.Merges = tt.Model.Merges
+		if len(tt.Model.Merges) == 0 {
+			// noop; merges is empty
+		} else if err := json.Unmarshal(tt.Model.Merges, &t.Merges); err == nil {
+			// noop; merges is []string
+		} else if merges, err := func() ([][]string, error) {
+			var merges [][]string
+			if err := json.Unmarshal(tt.Model.Merges, &merges); err != nil {
+				return nil, err
+			}
+
+			return merges, nil
+		}(); err == nil {
+			t.Merges = make([]string, len(merges))
+			for i := range merges {
+				t.Merges[i] = strings.Join(merges[i], " ")
+			}
+		} else {
+			return nil, fmt.Errorf("could not parse tokenizer merges. expected []string or [][]string: %w", err)
+		}

 		sha256sum := sha256.New()
 		for _, pt := range tt.PreTokenizer.PreTokenizers {
@@ -156,9 +175,9 @@ func parseTokenizer(fsys fs.FS, specialTokenTypes []string) (*Tokenizer, error)
 type tokenizer struct {
 	AddedTokens []token `json:"added_tokens"`
 	Model       struct {
-		Type   string         `json:"type"`
-		Vocab  map[string]int `json:"vocab"`
-		Merges []string       `json:"merges"`
+		Type   string          `json:"type"`
+		Vocab  map[string]int  `json:"vocab"`
+		Merges json.RawMessage `json:"merges"`
 	} `json:"model"`

 	PreTokenizer struct {
@@ -191,6 +191,62 @@ func TestParseTokenizer(t *testing.T) {
 				Pre: "default",
 			},
 		},
+		{
+			name: "list string merges",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"model": {
+						"merges": [
+							"a b",
+							"c d",
+							"e f"
+						]
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model: "gpt2",
+				},
+				Merges: []string{
+					"a b",
+					"c d",
+					"e f",
+				},
+				Pre: "default",
+			},
+		},
+		{
+			name: "list list string merges",
+			fsys: createTokenizerFS(t, t.TempDir(), map[string]io.Reader{
+				"tokenizer.json": strings.NewReader(`{
+					"model": {
+						"merges": [
+							[
+								"a", "b"
+							],
+							[
+								"c", "d"
+							],
+							[
+								"e", "f"
+							]
+						]
+					}
+				}`),
+			}),
+			want: &Tokenizer{
+				Vocabulary: &Vocabulary{
+					Model: "gpt2",
+				},
+				Merges: []string{
+					"a b",
+					"c d",
+					"e f",
+				},
+				Pre: "default",
+			},
+		},
 	}

 	for _, tt := range cases {
@@ -1,3 +0,0 @@
-# `discover`
-
-This package is responsible for discovering information about the system and the capabilities to run LLM.  This includes GPU and CPU discovery so the optimal runner can be chosen for a given model.  The ollama scheduler relies on up-to-date available memory information, so this package provides the ability to refresh free memory as efficiently as possible.
@@ -37,19 +37,6 @@ func GetSupportedGFX(libDir string) ([]string, error) {
 	return ret, nil
 }

-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
-		}
-		ids = append(ids, info.ID)
-	}
-	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
-}
-
 func commonAMDValidateLibDir() (string, error) {
 	// Favor our bundled version

@@ -64,7 +64,7 @@ func NewHipLib() (*HipLib, error) {
 	return hl, nil
 }

-// The hip library only evaluates the HIP_VISIBLE_DEVICES variable at startup
+// The hip library only evaluates the ROCR_VISIBLE_DEVICES variable at startup
 // so we have to unload/reset the library after we do our initial discovery
 // to make sure our updates to that variable are processed by llama.cpp
 func (hl *HipLib) Release() {
@@ -64,16 +64,13 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 	// Determine if the user has already pre-selected which GPUs to look at, then ignore the others
 	var visibleDevices []string
 	hipVD := envconfig.HipVisibleDevices()   // zero based index only
-	rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID, but consumer cards seem to not support UUID
+	rocrVD := envconfig.RocrVisibleDevices() // zero based index or UUID
 	gpuDO := envconfig.GpuDeviceOrdinal()    // 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 hipVD != "":
+		visibleDevices = strings.Split(hipVD, ",")
 	case gpuDO != "":
 		visibleDevices = strings.Split(gpuDO, ",")
 	}
@@ -99,7 +96,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		}
 		return a < b
 	})
-	cpuCount := 0
+	gpuCount := 0
 	for _, match := range matches {
 		slog.Debug("evaluating amdgpu node " + match)
 		fp, err := os.Open(match)
@@ -108,11 +105,6 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			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
@@ -186,20 +178,19 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		// do reliably report VRAM usage.

 		if isCPU {
-			cpuCount++
 			continue
 		}

-		// CPUs are always first in the list
-		gpuID := nodeID - cpuCount
-
-		// Shouldn't happen, but just in case...
-		if gpuID < 0 {
-			err := fmt.Errorf("unexpected amdgpu sysfs data resulted in negative GPU ID, please set OLLAMA_DEBUG=1 and report an issue")
-			slog.Error(err.Error())
-			return nil, err
+		// Skip over any GPUs that are masked
+		if major == 0 && minor == 0 && patch == 0 {
+			slog.Debug("skipping gpu with gfx000")
+			continue
 		}

+		// Keep track of numeric IDs based on valid GPUs
+		gpuID := gpuCount
+		gpuCount += 1
+
 		// Look up the memory for the current node
 		totalMemory := uint64(0)
 		usedMemory := uint64(0)
@@ -273,6 +264,14 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 			name = fmt.Sprintf("%04x:%04x", vendor, device)
 		}

+		// Favor UUIDs if available to reduce possibility of getting the numeric IDs wrong
+		var ID string
+		if uniqueID != 0 {
+			ID = fmt.Sprintf("GPU-%016x", uniqueID)
+		} else {
+			ID = strconv.Itoa(gpuID)
+		}
+
 		gpuInfo := RocmGPUInfo{
 			GpuInfo: GpuInfo{
 				Library: "rocm",
@@ -280,7 +279,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 					TotalMemory: totalMemory,
 					FreeMemory:  (totalMemory - usedMemory),
 				},
-				ID:            strconv.Itoa(gpuID),
+				ID:            ID,
 				Name:          name,
 				Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
 				MinimumMemory: rocmMinimumMemory,
@@ -288,6 +287,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				DriverMinor:   driverMinor,
 			},
 			usedFilepath: usedFile,
+			index:        gpuID,
 		}

 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
@@ -319,7 +319,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		if len(visibleDevices) > 0 {
 			include := false
 			for _, visible := range visibleDevices {
-				if visible == gpuInfo.ID {
+				if visible == gpuInfo.ID || visible == strconv.Itoa(gpuInfo.index) {
 					include = true
 					break
 				}
@@ -350,7 +350,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				return nil, err
 			}
 		}
-		gpuInfo.DependencyPath = libDir
+		gpuInfo.DependencyPath = []string{libDir}

 		if gfxOverride == "" {
 			// Only load supported list once
@@ -516,3 +516,20 @@ func verifyKFDDriverAccess() error {
 	fd.Close()
 	return nil
 }
+
+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
+		}
+		ids = append(ids, info.ID)
+	}
+	// There are 3 potential env vars to use to select GPUs.
+	// ROCR_VISIBLE_DEVICES supports UUID or numeric so is our preferred on linux
+	// GPU_DEVICE_ORDINAL supports numeric IDs only
+	// HIP_VISIBLE_DEVICES supports numeric IDs only
+	return "ROCR_VISIBLE_DEVICES", strings.Join(ids, ",")
+}
@@ -43,7 +43,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 		slog.Debug("error looking up amd driver version", "error", err)
 	}

-	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
+	// Note: the HIP library automatically handles subsetting to any *_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
 	if count == 0 {
 		err := fmt.Errorf("no compatible amdgpu devices detected")
@@ -111,7 +111,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				UnreliableFreeMemory: true,

 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				DependencyPath: libDir,
+				DependencyPath: []string{libDir},
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
@@ -201,3 +201,20 @@ func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
 	}
 	return nil
 }
+
+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
+		}
+		ids = append(ids, info.ID)
+	}
+	// There are 3 potential env vars to use to select GPUs.
+	// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
+	// HIP_VISIBLE_DEVICES supports numeric IDs only
+	// GPU_DEVICE_ORDINAL supports numeric IDs only
+	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
+}
@@ -240,7 +240,7 @@ func GetGPUInfo() GpuInfoList {
 					Library:        "cpu",
 					Variant:        cpuCapability.String(),
 					ID:             "0",
-					DependencyPath: depPath,
+					DependencyPath: []string{depPath},
 				},
 				CPUs: details,
 			},
@@ -293,11 +293,11 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMinor = driverMinor
 				variant := cudaVariant(gpuInfo)
 				if depPath != "" {
-					gpuInfo.DependencyPath = depPath
+					gpuInfo.DependencyPath = []string{depPath}
 					// Check for variant specific directory
 					if variant != "" {
 						if _, err := os.Stat(filepath.Join(depPath, "cuda_"+variant)); err == nil {
-							gpuInfo.DependencyPath = filepath.Join(depPath, "cuda_"+variant)
+							gpuInfo.DependencyPath = []string{filepath.Join(depPath, "cuda_"+variant), depPath}
 						}
 					}
 				}
@@ -316,7 +316,9 @@ func GetGPUInfo() GpuInfoList {
 				// query the management library as well so we can record any skew between the two
 				// which represents overhead on the GPU we must set aside on subsequent updates
 				if cHandles.nvml != nil {
-					C.nvml_get_free(*cHandles.nvml, C.int(gpuInfo.index), &memInfo.free, &memInfo.total, &memInfo.used)
+					uuid := C.CString(gpuInfo.ID)
+					defer C.free(unsafe.Pointer(uuid))
+					C.nvml_get_free(*cHandles.nvml, uuid, &memInfo.free, &memInfo.total, &memInfo.used)
 					if memInfo.err != nil {
 						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
 						C.free(unsafe.Pointer(memInfo.err))
@@ -368,7 +370,7 @@ func GetGPUInfo() GpuInfoList {
 						gpuInfo.FreeMemory = uint64(memInfo.free)
 						gpuInfo.ID = C.GoString(&memInfo.gpu_id[0])
 						gpuInfo.Name = C.GoString(&memInfo.gpu_name[0])
-						gpuInfo.DependencyPath = depPath
+						gpuInfo.DependencyPath = []string{depPath}
 						oneapiGPUs = append(oneapiGPUs, gpuInfo)
 					}
 				}
@@ -417,7 +419,9 @@ func GetGPUInfo() GpuInfoList {
 		}
 		for i, gpu := range cudaGPUs {
 			if cHandles.nvml != nil {
-				C.nvml_get_free(*cHandles.nvml, C.int(gpu.index), &memInfo.free, &memInfo.total, &memInfo.used)
+				uuid := C.CString(gpu.ID)
+				defer C.free(unsafe.Pointer(uuid))
+				C.nvml_get_free(*cHandles.nvml, uuid, &memInfo.free, &memInfo.total, &memInfo.used)
 			} else if cHandles.cudart != nil {
 				C.cudart_bootstrap(*cHandles.cudart, C.int(gpu.index), &memInfo)
 			} else if cHandles.nvcuda != nil {
@@ -4,6 +4,7 @@
 #include "gpu_info_nvcuda.h"

 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
+  LOG(resp->ch.verbose, "initializing %s\n", nvcuda_lib_path);
  CUresult ret;
  resp->err = NULL;
  resp->num_devices = 0;
@@ -57,8 +58,10 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
      resp->cudaErr = -1;
      return;
    }
+    LOG(resp->ch.verbose, "dlsym: %s - %p\n", l[i].s, *l[i].p);
  }

+  LOG(resp->ch.verbose, "calling cuInit\n");
  ret = (*resp->ch.cuInit)(0);
  if (ret != CUDA_SUCCESS) {
    LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
@@ -75,15 +78,18 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
  resp->ch.driver_minor = 0;

  // Report driver version if we're in verbose mode, ignore errors
+  LOG(resp->ch.verbose, "calling cuDriverGetVersion\n");
  ret = (*resp->ch.cuDriverGetVersion)(&version);
  if (ret != CUDA_SUCCESS) {
    LOG(resp->ch.verbose, "cuDriverGetVersion failed: %d\n", ret);
  } else {
+    LOG(resp->ch.verbose, "raw version 0x%x\n", version);
    resp->ch.driver_major = version / 1000;
    resp->ch.driver_minor = (version - (resp->ch.driver_major * 1000)) / 10;
    LOG(resp->ch.verbose, "CUDA driver version: %d.%d\n", resp->ch.driver_major, resp->ch.driver_minor);
  }

+  LOG(resp->ch.verbose, "calling cuDeviceGetCount\n");
  ret = (*resp->ch.cuDeviceGetCount)(&resp->num_devices);
  if (ret != CUDA_SUCCESS) {
    LOG(resp->ch.verbose, "cuDeviceGetCount err: %d\n", ret);
@@ -94,6 +100,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
    resp->cudaErr = ret;
    return;
  }
+  LOG(resp->ch.verbose, "device count %d\n", resp->num_devices);
 }

 const int buflen = 256;
@@ -17,7 +17,7 @@ void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
  } l[] = {
      {"nvmlInit_v2", (void *)&resp->ch.nvmlInit_v2},
      {"nvmlShutdown", (void *)&resp->ch.nvmlShutdown},
-      {"nvmlDeviceGetHandleByIndex", (void *)&resp->ch.nvmlDeviceGetHandleByIndex},
+      {"nvmlDeviceGetHandleByUUID", (void *)&resp->ch.nvmlDeviceGetHandleByUUID},
      {"nvmlDeviceGetMemoryInfo", (void *)&resp->ch.nvmlDeviceGetMemoryInfo},
      {NULL, NULL},
  };
@@ -67,20 +67,20 @@ void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp) {
 }


-void nvml_get_free(nvml_handle_t h, int device_id, uint64_t *free, uint64_t *total, uint64_t *used) {
+void nvml_get_free(nvml_handle_t h, char *uuid, uint64_t *free, uint64_t *total, uint64_t *used) {
    nvmlDevice_t device;
    nvmlMemory_t memInfo = {0};
    nvmlReturn_t ret;
-    ret = (*h.nvmlDeviceGetHandleByIndex)(device_id, &device);
+    ret = (*h.nvmlDeviceGetHandleByUUID)((const char *)(uuid), &device);
    if (ret != NVML_SUCCESS) {
-        LOG(1, "unable to get device handle %d: %d", device_id, ret);
+        LOG(1, "unable to get device handle %s: %d", uuid, ret);
        *free = 0;
        return;
    }

    ret = (*h.nvmlDeviceGetMemoryInfo)(device, &memInfo);
    if (ret != NVML_SUCCESS) {
-        LOG(1, "device memory info lookup failure %d: %d", device_id, ret);
+        LOG(1, "device memory info lookup failure %s: %d", uuid, ret);
        *free = 0;
        return;
    }
@@ -25,7 +25,7 @@ typedef struct nvml_handle {
  uint16_t verbose;
  nvmlReturn_t (*nvmlInit_v2)(void);
  nvmlReturn_t (*nvmlShutdown)(void);
-  nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(unsigned int, nvmlDevice_t *);
+  nvmlReturn_t (*nvmlDeviceGetHandleByUUID)(const char *, nvmlDevice_t *);
  nvmlReturn_t (*nvmlDeviceGetMemoryInfo)(nvmlDevice_t, nvmlMemory_t *);
 } nvml_handle_t;

@@ -41,7 +41,7 @@ typedef struct nvml_compute_capability {
 } nvml_compute_capability_t;

 void nvml_init(char *nvml_lib_path, nvml_init_resp_t *resp);
-void nvml_get_free(nvml_handle_t ch,  int device_id, uint64_t *free, uint64_t *total, uint64_t *used);
+void nvml_get_free(nvml_handle_t ch, char *uuid, uint64_t *free, uint64_t *total, uint64_t *used);
 void nvml_release(nvml_handle_t ch);

 #endif  // __GPU_INFO_NVML_H__
@@ -3,9 +3,11 @@ package discover
 import (
 	"bufio"
 	"fmt"
+	"io"
 	"os"
 	"reflect"
 	"regexp"
+	"sort"
 	"strings"

 	"github.com/ollama/ollama/format"
@@ -109,6 +111,10 @@ func GetCPUDetails() ([]CPU, error) {
 	if err != nil {
 		return nil, err
 	}
+	return linuxCPUDetails(file)
+}
+
+func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	reColumns := regexp.MustCompile("\t+: ")
 	scanner := bufio.NewScanner(file)
 	cpuInfos := []linuxCpuInfo{}
@@ -131,6 +137,9 @@ func GetCPUDetails() ([]CPU, error) {
 			cpu = &linuxCpuInfo{}
 		}
 	}
+	if cpu.ID != "" {
+		cpuInfos = append(cpuInfos, *cpu)
+	}

 	// Process the sockets/cores/threads
 	socketByID := map[string]*CPU{}
@@ -177,10 +186,14 @@ func GetCPUDetails() ([]CPU, error) {
 			s.EfficiencyCoreCount = efficiencyCoreCount
 		}
 	}
-
-	result := []CPU{}
-	for _, c := range socketByID {
-		result = append(result, *c)
+	keys := make([]string, 0, len(socketByID))
+	result := make([]CPU, 0, len(socketByID))
+	for k := range socketByID {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	for _, k := range keys {
+		result = append(result, *socketByID[k])
 	}
 	return result, nil
 }
@@ -25,7 +25,7 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	MinimumMemory uint64 `json:"-"`

 	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
-	DependencyPath string `json:"lib_path,omitempty"`
+	DependencyPath []string `json:"lib_path,omitempty"`

 	// Extra environment variables specific to the GPU as list of [key,value]
 	EnvWorkarounds [][2]string `json:"envs,omitempty"`
@@ -175,6 +175,25 @@ func (si SystemInfo) GetOptimalThreadCount() int {
 	if len(si.System.CPUs) == 0 {
 		return 0
 	}
-	// Allocate thread count matching the performance cores on a single socket
-	return si.System.CPUs[0].CoreCount - si.System.CPUs[0].EfficiencyCoreCount
+
+	coreCount := 0
+	for _, c := range si.System.CPUs {
+		coreCount += c.CoreCount - c.EfficiencyCoreCount
+	}
+
+	return coreCount
+}
+
+// For each GPU, check if it does NOT support flash attention
+func (l GpuInfoList) FlashAttentionSupported() bool {
+	for _, gpu := range l {
+		supportsFA := gpu.Library == "metal" ||
+			(gpu.Library == "cuda" && gpu.DriverMajor >= 7) ||
+			gpu.Library == "rocm"
+
+		if !supportsFA {
+			return false
+		}
+	}
+	return true
 }
@@ -49,10 +49,10 @@ Advanced parameters (optional):
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `system`: system message to (overrides what is defined in the `Modelfile`)
 - `template`: the prompt template to use (overrides what is defined in the `Modelfile`)
- `context`: the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 - `stream`: if `false` the response will be returned as a single response object, rather than a stream of objects
 - `raw`: if `true` no formatting will be applied to the prompt. You may choose to use the `raw` parameter if you are specifying a full templated prompt in your request to the API
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+- `context` (deprecated): the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory

 #### JSON mode

@@ -355,7 +355,6 @@ curl http://localhost:11434/api/generate -d '{
    "num_gpu": 1,
    "main_gpu": 0,
    "low_vram": false,
-    "f16_kv": true,
    "vocab_only": false,
    "use_mmap": true,
    "use_mlock": false,
@@ -831,10 +830,30 @@ Create a model from a [`Modelfile`](./modelfile.md). It is recommended to set `m

 ### Parameters

- `name`: name of the model to create
+- `model`: name of the model to create
 - `modelfile` (optional): contents of the Modelfile
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects
 - `path` (optional): path to the Modelfile
+- `quantize` (optional): quantize a non-quantized (e.g. float16) model
+
+#### Quantization types
+
+| Type | Recommended |
+| --- | :-: |
+| q2_K | |
+| q3_K_L | |
+| q3_K_M | |
+| q3_K_S | |
+| q4_0 | |
+| q4_1 | |
+| q4_K_M | * |
+| q4_K_S | |
+| q5_0 | |
+| q5_1 | |
+| q5_K_M | |
+| q5_K_S | |
+| q6_K | |
+| q8_0 | * |

 ### Examples

@@ -846,14 +865,14 @@ Create a new model from a `Modelfile`.

 ```shell
 curl http://localhost:11434/api/create -d '{
-  "name": "mario",
+  "model": "mario",
  "modelfile": "FROM llama3\nSYSTEM You are mario from Super Mario Bros."
 }'
 ```

 ##### Response

-A stream of JSON objects. Notice that the final JSON object shows a `"status": "success"`.
+A stream of JSON objects is returned:

 ```json
 {"status":"reading model metadata"}
@@ -869,13 +888,43 @@ A stream of JSON objects. Notice that the final JSON object shows a `"status": "
 {"status":"success"}
 ```

+#### Quantize a model
+
+Quantize a non-quantized model.
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/create -d '{
+  "model": "llama3.1:quantized",
+  "modelfile": "FROM llama3.1:8b-instruct-fp16",
+  "quantize": "q4_K_M"
+}'
+```
+
+##### Response
+
+A stream of JSON objects is returned:
+
+```
+{"status":"quantizing F16 model to Q4_K_M"}
+{"status":"creating new layer sha256:667b0c1932bc6ffc593ed1d03f895bf2dc8dc6df21db3042284a6f4416b06a29"}
+{"status":"using existing layer sha256:11ce4ee3e170f6adebac9a991c22e22ab3f8530e154ee669954c4bc73061c258"}
+{"status":"using existing layer sha256:0ba8f0e314b4264dfd19df045cde9d4c394a52474bf92ed6a3de22a4ca31a177"}
+{"status":"using existing layer sha256:56bb8bd477a519ffa694fc449c2413c6f0e1d3b1c88fa7e3c9d88d3ae49d4dcb"}
+{"status":"creating new layer sha256:455f34728c9b5dd3376378bfb809ee166c145b0b4c1f1a6feca069055066ef9a"}
+{"status":"writing manifest"}
+{"status":"success"}
+```
+
+
 ### Check if a Blob Exists

 ```shell
 HEAD /api/blobs/:digest
 ```

-Ensures that the file blob used for a FROM or ADAPTER field exists on the server. This is checking your Ollama server and not Ollama.ai.
+Ensures that the file blob used for a FROM or ADAPTER field exists on the server. This is checking your Ollama server and not ollama.com.

 #### Query Parameters

@@ -980,7 +1029,7 @@ Show information about a model including details, modelfile, template, parameter

 ### Parameters

- `name`: name of the model to show
+- `model`: name of the model to show
 - `verbose`: (optional) if set to `true`, returns full data for verbose response fields

 ### Examples
@@ -989,7 +1038,7 @@ Show information about a model including details, modelfile, template, parameter

 ```shell
 curl http://localhost:11434/api/show -d '{
-  "name": "llama3.2"
+  "model": "llama3.2"
 }'
 ```

@@ -1069,7 +1118,7 @@ Delete a model and its data.

 ### Parameters

- `name`: model name to delete
+- `model`: model name to delete

 ### Examples

@@ -1077,7 +1126,7 @@ Delete a model and its data.

 ```shell
 curl -X DELETE http://localhost:11434/api/delete -d '{
-  "name": "llama3:13b"
+  "model": "llama3:13b"
 }'
 ```

@@ -1095,7 +1144,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where

 ### Parameters

- `name`: name of the model to pull
+- `model`: name of the model to pull
 - `insecure`: (optional) allow insecure connections to the library. Only use this if you are pulling from your own library during development.
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects

@@ -1105,7 +1154,7 @@ Download a model from the ollama library. Cancelled pulls are resumed from where

 ```shell
 curl http://localhost:11434/api/pull -d '{
-  "name": "llama3.2"
+  "model": "llama3.2"
 }'
 ```

@@ -1167,7 +1216,7 @@ Upload a model to a model library. Requires registering for ollama.ai and adding

 ### Parameters

- `name`: name of the model to push in the form of `<namespace>/<model>:<tag>`
+- `model`: name of the model to push in the form of `<namespace>/<model>:<tag>`
 - `insecure`: (optional) allow insecure connections to the library. Only use this if you are pushing to your library during development.
 - `stream`: (optional) if `false` the response will be returned as a single response object, rather than a stream of objects

@@ -1177,7 +1226,7 @@ Upload a model to a model library. Requires registering for ollama.ai and adding

 ```shell
 curl http://localhost:11434/api/push -d '{
-  "name": "mattw/pygmalion:latest"
+  "model": "mattw/pygmalion:latest"
 }'
 ```

@@ -108,7 +108,7 @@ Custom CPU settings are not currently supported in the new Go server build but w

 #### Containerized Linux Build

-If you have Docker available, you can build linux binaries with `OLLAMA_NEW_RUNNERS=1 ./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`
+If you have Docker available, you can build linux binaries with `./scripts/build_linux.sh` which has the CUDA and ROCm dependencies included. The resulting binary is placed in `./dist`

 ### Windows

@@ -118,10 +118,13 @@ The following tools are required as a minimal development environment to build C
  - https://go.dev/dl/
 - Git
  - https://git-scm.com/download/win
- GCC and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
+- clang with gcc compat and Make.  There are multiple options on how to go about installing these tools on Windows.  We have verified the following, but others may work as well:  
  - [MSYS2](https://www.msys2.org/)
-    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-ucrt-x86_64-gcc make` to install the required tools
-  - Assuming you used the default install prefix for msys2 above, add `c:\msys64\ucrt64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+    - After installing, from an MSYS2 terminal, run `pacman -S mingw-w64-clang-x86_64-gcc-compat mingw-w64-clang-x86_64-clang make` to install the required tools
+  - Assuming you used the default install prefix for msys2 above, add `C:\msys64\clang64\bin` and `c:\msys64\usr\bin` to your environment variable `PATH` where you will perform the build steps below (e.g. system-wide, account-level, powershell, cmd, etc.)
+
+> [!NOTE]  
+> Due to bugs in the GCC C++ library for unicode support, Ollama should be built with clang on windows.

 Then, build the `ollama` binary:

@@ -50,6 +50,9 @@ sudo systemctl restart docker
 docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
 ```

+> [!NOTE]  
+> If you're running on an NVIDIA JetPack system, Ollama can't automatically discover the correct JetPack version. Pass the environment variable JETSON_JETPACK=5 or JETSON_JETPACK=6 to the container to select version 5 or 6.
+
 ### AMD GPU

 To run Ollama using Docker with AMD GPUs, use the `rocm` tag and the following command:
@@ -151,7 +151,7 @@ Refer to the section [above](#how-do-i-configure-ollama-server) for how to set e

 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:

-```
+```nginx
 server {
    listen 80;
    server_name example.com;  # Replace with your domain or IP
@@ -285,4 +285,28 @@ Note: Windows with Radeon GPUs currently default to 1 model maximum due to limit

 ## How does Ollama load models on multiple GPUs?

-Installing multiple GPUs of the same brand can be a great way to increase your available VRAM to load larger models.  When you load a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transfering across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
+When loading a new model, Ollama evaluates the required VRAM for the model against what is currently available.  If the model will entirely fit on any single GPU, Ollama will load the model on that GPU.  This typically provides the best performance as it reduces the amount of data transferring across the PCI bus during inference.  If the model does not fit entirely on one GPU, then it will be spread across all the available GPUs.
+
+## How can I enable Flash Attention?
+
+Flash Attention is a feature of most modern models that can significantly reduce memory usage as the context size grows.  To enable Flash Attention, set the `OLLAMA_FLASH_ATTENTION` environment variable to `1` when starting the Ollama server.
+
+## How can I set the quantization type for the K/V cache?
+
+The K/V context cache can be quantized to significantly reduce memory usage when Flash Attention is enabled.
+
+To use quantized K/V cache with Ollama you can set the following environment variable:
+
+- `OLLAMA_KV_CACHE_TYPE` - The quantization type for the K/V cache.  Default is `f16`.
+
+> Note: Currently this is a global option - meaning all models will run with the specified quantization type.
+
+The currently available K/V cache quantization types are:
+
+- `f16` - high precision and memory usage (default).
+- `q8_0` - 8-bit quantization, uses approximately 1/2 the memory of `f16` with a very small loss in precision, this usually has no noticeable impact on the model's quality (recommended if not using f16).
+- `q4_0` - 4-bit quantization, uses approximately 1/4 the memory of `f16` with a small-medium loss in precision that may be more noticeable at higher context sizes.
+
+How much the cache quantization impacts the model's response quality will depend on the model and the task.  Models that have a high GQA count (e.g. Qwen2) may see a larger impact on precision from quantization than models with a low GQA count.
+
+You may need to experiment with different quantization types to find the best balance between memory usage and quality.
@@ -74,6 +74,10 @@ 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.

+If you have multiple GPUs with different GFX versions, append the numeric device
+number to the environment variable to set them individually.  For example,
+`HSA_OVERRIDE_GFX_VERSION_0=10.3.0` and  `HSA_OVERRIDE_GFX_VERSION_1=11.0.0`
+
 At this time, the known supported GPU types on linux are the following LLVM Targets.
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
@@ -99,9 +103,10 @@ Reach out on [Discord](https://discord.gg/ollama) or file an
 ### 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.
+subset, you can set `ROCR_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")
+and force CPU usage, use an invalid GPU ID (e.g., "-1").  When available, use the
+`Uuid` to uniquely identify the device instead of numeric value.

 ### Container Permission

@@ -32,7 +32,7 @@ ollama run my-model

 Ollama supports importing adapters based on several different model architectures including:

-  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Llama (including Llama 2, Llama 3, Llama 3.1, and Llama 3.2);
  * Mistral (including Mistral 1, Mistral 2, and Mixtral); and
  * Gemma (including Gemma 1 and Gemma 2)

@@ -67,14 +67,12 @@ ollama run my-model

 Ollama supports importing models for several different architectures including:

-  * Llama (including Llama 2, Llama 3, and Llama 3.1);
+  * Llama (including Llama 2, Llama 3, Llama 3.1, and Llama 3.2);
  * Mistral (including Mistral 1, Mistral 2, and Mixtral);
  * Gemma (including Gemma 1 and Gemma 2); and
  * Phi3

-This includes importing foundation models as well as any fine tuned models which which have been _fused_ with a foundation model.
-
-
+This includes importing foundation models as well as any fine tuned models which have been _fused_ with a foundation model.
 ## Importing a GGUF based model or adapter

 If you have a GGUF based model or adapter it is possible to import it into Ollama. You can obtain a GGUF model or adapter by:
@@ -83,7 +81,7 @@ If you have a GGUF based model or adapter it is possible to import it into Ollam
  * converting a Safetensors adapter with the `convert_lora_to_gguf.py` from Llama.cpp; or
  * downloading a model or adapter from a place such as HuggingFace

-To import a GGUF model, create a `Modelfile` containg:
+To import a GGUF model, create a `Modelfile` containing:

 ```dockerfile
 FROM /path/to/file.gguf
@@ -112,6 +112,21 @@ sudo systemctl status ollama
 > https://www.amd.com/en/support/linux-drivers for best support of your Radeon
 > GPU.

+## Customizing
+
+To customize the installation of Ollama, you can edit the systemd service file or the environment variables by running:
+
+```
+sudo systemctl edit ollama
+```
+
+Alternatively, create an override file manually in `/etc/systemd/system/ollama.service.d/override.conf`:
+
+```ini
+[Service]
+Environment="OLLAMA_DEBUG=1"
+```
+
 ## Updating

 Update Ollama by running the install script again:
@@ -129,7 +144,7 @@ sudo tar -C /usr -xzf ollama-linux-amd64.tgz

 ## Installing specific versions

-Use `OLLAMA_VERSION` environment variable with the install script to install a specific version of Ollama, including pre-releases. You can find the version numbers in the [releases page](https://github.com/ollama/ollama/releases). 
+Use `OLLAMA_VERSION` environment variable with the install script to install a specific version of Ollama, including pre-releases. You can find the version numbers in the [releases page](https://github.com/ollama/ollama/releases).

 For example:

@@ -63,7 +63,7 @@ SYSTEM You are Mario from super mario bros, acting as an assistant.
 To use this:

 1. Save it as a file (e.g. `Modelfile`)
-2. `ollama create choose-a-model-name -f <location of the file e.g. ./Modelfile>'`
+2. `ollama create choose-a-model-name -f <location of the file e.g. ./Modelfile>`
 3. `ollama run choose-a-model-name`
 4. Start using the model!

@@ -120,7 +120,7 @@ FROM <model directory>
 The model directory should contain the Safetensors weights for a supported architecture.

 Currently supported model architectures:
-  * Llama (including Llama 2, Llama 3, and Llama 3.1)
+  * Llama (including Llama 2, Llama 3, Llama 3.1, and Llama 3.2)
  * Mistral (including Mistral 1, Mistral 2, and Mixtral)
  * Gemma (including Gemma 1 and Gemma 2)
  * Phi3
@@ -156,7 +156,7 @@ PARAMETER <parameter> <parametervalue>
 | seed           | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0)                                                                                       | int        | seed 42              |
 | stop           | Sets the stop sequences to use. When this pattern is encountered the LLM will stop generating text and return. Multiple stop patterns may be set by specifying multiple separate `stop` parameters in a modelfile.                                      | string     | stop "AI assistant:" |
 | tfs_z          | Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting. (default: 1)                                               | float      | tfs_z 1              |
-| num_predict    | Maximum number of tokens to predict when generating text. (Default: 128, -1 = infinite generation, -2 = fill context)                                                                                                                                   | int        | num_predict 42       |
+| num_predict    | Maximum number of tokens to predict when generating text. (Default: -1, infinite generation)                                                                                                                                   | int        | num_predict 42       |
 | top_k          | Reduces the probability of generating nonsense. A higher value (e.g. 100) will give more diverse answers, while a lower value (e.g. 10) will be more conservative. (Default: 40)                                                                        | int        | top_k 40             |
 | top_p          | Works together with top-k. A higher value (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. (Default: 0.9)                                                                 | float      | top_p 0.9            |
 | min_p          | Alternative to the top_p, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. (Default: 0.0) | float      | min_p 0.05            |
@@ -37,7 +37,7 @@ response = client.chat.completions.create(
                {"type": "text", "text": "What's in this image?"},
                {
                    "type": "image_url",
-                    "image_url": "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",
+                    "image_url": "data:image/png;base64,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",
                },
            ],
        }
@@ -86,7 +86,7 @@ const response = await openai.chat.completions.create({
            { type: "text", text: "What's in this image?" },
            {
            type: "image_url",
-            image_url: "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",
+            image_url: "data:image/png;base64,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",
            },
        ],
        },
@@ -142,7 +142,7 @@ curl http://localhost:11434/v1/chat/completions \
          {
            "type": "image_url",
            "image_url": {
-               "url": "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"
+               "url": "data:image/png;base64,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"
            }
          }
        ]
@@ -95,13 +95,21 @@ If none of those resolve the problem, gather additional information and file an

 On linux, AMD GPU access typically requires `video` and/or `render` group membership to access the `/dev/kfd` device.  If permissions are not set up correctly, Ollama will detect this and report an error in the server log.

-When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -ld /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the group assignments on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.
+When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -lnd /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the **numeric** group IDs on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.   For example, in the following output `crw-rw---- 1 0  44 226,   0 Sep 16 16:55 /dev/dri/card0` the group ID column is `44` 
+
+If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.

 If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
 - `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
 - `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported
 - Check dmesg for any errors from amdgpu or kfd drivers `sudo dmesg | grep -i amdgpu` and `sudo dmesg | grep -i kfd`

+## Multiple AMD GPUs
+
+If you experience gibberish responses when models load across multiple AMD GPUs on Linux, see the following guide.
+
+- https://rocm.docs.amd.com/projects/radeon/en/latest/docs/install/native_linux/mgpu.html#mgpu-known-issues-and-limitations
+
 ## Windows Terminal Errors

 Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
@@ -1,83 +0,0 @@
-# Running Ollama on Fly.io GPU Instances
-
-Ollama runs with little to no configuration on [Fly.io GPU instances](https://fly.io/docs/gpus/gpu-quickstart/). If you don't have access to GPUs yet, you'll need to [apply for access](https://fly.io/gpu/) on the waitlist. Once you're accepted, you'll get an email with instructions on how to get started.
-
-Create a new app with `fly apps create`:
-
-```bash
-fly apps create
-```
-
-Then create a `fly.toml` file in a new folder that looks like this:
-
-```toml
-app = "sparkling-violet-709"
-primary_region = "ord"
-vm.size = "a100-40gb" # see https://fly.io/docs/gpus/gpu-quickstart/ for more info
-
-[build]
-  image = "ollama/ollama"
-
-[http_service]
-  internal_port = 11434
-  force_https = false
-  auto_stop_machines = true
-  auto_start_machines = true
-  min_machines_running = 0
-  processes = ["app"]
-
-[mounts]
-  source = "models"
-  destination = "/root/.ollama"
-  initial_size = "100gb"
-```
-
-Then create a [new private IPv6 address](https://fly.io/docs/reference/private-networking/#flycast-private-load-balancing) for your app:
-
-```bash
-fly ips allocate-v6 --private
-```
-
-Then deploy your app:
-
-```bash
-fly deploy
-```
-
-And finally you can access it interactively with a new Fly.io Machine:
-
-```
-fly machine run -e OLLAMA_HOST=http://your-app-name.flycast --shell ollama/ollama
-```
-
-```bash
-$ ollama run openchat:7b-v3.5-fp16
->>> How do I bake chocolate chip cookies?
- To bake chocolate chip cookies, follow these steps:
-
-1. Preheat the oven to 375°F (190°C) and line a baking sheet with parchment paper or silicone baking mat.
-
-2. In a large bowl, mix together 1 cup of unsalted butter (softened), 3/4 cup granulated sugar, and 3/4
-cup packed brown sugar until light and fluffy.
-
-3. Add 2 large eggs, one at a time, to the butter mixture, beating well after each addition. Stir in 1
-teaspoon of pure vanilla extract.
-
-4. In a separate bowl, whisk together 2 cups all-purpose flour, 1/2 teaspoon baking soda, and 1/2 teaspoon
-salt. Gradually add the dry ingredients to the wet ingredients, stirring until just combined.
-
-5. Fold in 2 cups of chocolate chips (or chunks) into the dough.
-
-6. Drop rounded tablespoons of dough onto the prepared baking sheet, spacing them about 2 inches apart.
-
-7. Bake for 10-12 minutes, or until the edges are golden brown. The centers should still be slightly soft.
-
-8. Allow the cookies to cool on the baking sheet for a few minutes before transferring them to a wire rack
-to cool completely.
-
-Enjoy your homemade chocolate chip cookies!
-```
-
-When you set it up like this, it will automatically turn off when you're done using it. Then when you access it again, it will automatically turn back on. This is a great way to save money on GPU instances when you're not using them. If you want a persistent wake-on-use connection to your Ollama instance, you can set up a [connection to your Fly network using WireGuard](https://fly.io/docs/reference/private-networking/#discovering-apps-through-dns-on-a-wireguard-connection). Then you can access your Ollama instance at `http://your-app-name.flycast`.
-
-And that's it!
@@ -1,77 +0,0 @@
-# Using LangChain with Ollama using JavaScript
-
-In this tutorial, we are going to use JavaScript with LangChain and Ollama to learn about something just a touch more recent. In August 2023, there was a series of wildfires on Maui. There is no way an LLM trained before that time can know about this, since their training data would not include anything as recent as that. So we can find the [Wikipedia article about the fires](https://en.wikipedia.org/wiki/2023_Hawaii_wildfires) and ask questions about the contents.
-
-To get started, let's just use **LangChain** to ask a simple question to a model. To do this with JavaScript, we need to install **LangChain**:
-
-```bash
-npm install @langchain/community
-```
-
-Now we can start building out our JavaScript:
-
-```javascript
-import { Ollama } from "@langchain/community/llms/ollama";
-
-const ollama = new Ollama({
-  baseUrl: "http://localhost:11434",
-  model: "llama3.2",
-});
-
-const answer = await ollama.invoke(`why is the sky blue?`);
-
-console.log(answer);
-```
-
-That will get us the same thing as if we ran `ollama run llama3.2 "why is the sky blue"` in the terminal. But we want to load a document from the web to ask a question against. **Cheerio** is a great library for ingesting a webpage, and **LangChain** uses it in their **CheerioWebBaseLoader**. So let's install **Cheerio** and build that part of the app.
-
-```bash
-npm install cheerio
-```
-
-```javascript
-import { CheerioWebBaseLoader } from "langchain/document_loaders/web/cheerio";
-
-const loader = new CheerioWebBaseLoader("https://en.wikipedia.org/wiki/2023_Hawaii_wildfires");
-const data = await loader.load();
-```
-
-That will load the document. Although this page is smaller than the Odyssey, it is certainly bigger than the context size for most LLMs. So we are going to need to split into smaller pieces, and then select just the pieces relevant to our question. This is a great use for a vector datastore. In this example, we will use the **MemoryVectorStore** that is part of **LangChain**. But there is one more thing we need to get the content into the datastore. We have to run an embeddings process that converts the tokens in the text into a series of vectors. And for that, we are going to use **Tensorflow**. There is a lot of stuff going on in this one. First, install the **Tensorflow** components that we need.
-
-```javascript
-npm install @tensorflow/tfjs-core@3.6.0 @tensorflow/tfjs-converter@3.6.0 @tensorflow-models/universal-sentence-encoder@1.3.3 @tensorflow/tfjs-node@4.10.0
-```
-
-If you just install those components without the version numbers, it will install the latest versions, but there are conflicts within **Tensorflow**, so you need to install the compatible versions.
-
-```javascript
-import { RecursiveCharacterTextSplitter } from "langchain/text_splitter"
-import { MemoryVectorStore } from "langchain/vectorstores/memory";
-import "@tensorflow/tfjs-node";
-import { TensorFlowEmbeddings } from "langchain/embeddings/tensorflow";
-
-// Split the text into 500 character chunks. And overlap each chunk by 20 characters
-const textSplitter = new RecursiveCharacterTextSplitter({
- chunkSize: 500,
- chunkOverlap: 20
-});
-const splitDocs = await textSplitter.splitDocuments(data);
-
-// Then use the TensorFlow Embedding to store these chunks in the datastore
-const vectorStore = await MemoryVectorStore.fromDocuments(splitDocs, new TensorFlowEmbeddings());
-```
-
-To connect the datastore to a question asked to a LLM, we need to use the concept at the heart of **LangChain**: the chain. Chains are a way to connect a number of activities together to accomplish a particular tasks. There are a number of chain types available, but for this tutorial we are using the **RetrievalQAChain**.
-
-```javascript
-import { RetrievalQAChain } from "langchain/chains";
-
-const retriever = vectorStore.asRetriever();
-const chain = RetrievalQAChain.fromLLM(ollama, retriever);
-const result = await chain.call({query: "When was Hawaii's request for a major disaster declaration approved?"});
-console.log(result.text)
-```
-
-So we created a retriever, which is a way to return the chunks that match a query from a datastore. And then connect the retriever and the model via a chain. Finally, we send a query to the chain, which results in an answer using our document as a source. The answer it returned was correct, August 10, 2023.
-
-And that is a simple introduction to what you can do with **LangChain** and **Ollama.**
@@ -1,85 +0,0 @@
-# Using LangChain with Ollama in Python
-
-Let's imagine we are studying the classics, such as **the Odyssey** by **Homer**. We might have a question about Neleus and his family. If you ask llama2 for that info, you may get something like:
-
-> I apologize, but I'm a large language model, I cannot provide information on individuals or families that do not exist in reality. Neleus is not a real person or character, and therefore does not have a family or any other personal details. My apologies for any confusion. Is there anything else I can help you with?
-
-This sounds like a typical censored response, but even llama2-uncensored gives a mediocre answer:
-
-> Neleus was a legendary king of Pylos and the father of Nestor, one of the Argonauts. His mother was Clymene, a sea nymph, while his father was Neptune, the god of the sea.
-
-So let's figure out how we can use **LangChain** with Ollama to ask our question to the actual document, the Odyssey by Homer, using Python.
-
-Let's start by asking a simple question that we can get an answer to from the **Llama2** model using **Ollama**. First, we need to install the **LangChain** package:
-
-`pip install langchain_community`
-
-Then we can create a model and ask the question:
-
-```python
-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.
-
-Now let's load a document to ask questions against. I'll load up the Odyssey by Homer, which you can find at Project Gutenberg. We will need **WebBaseLoader** which is part of **LangChain** and loads text from any webpage. On my machine, I also needed to install **bs4** to get that to work, so run `pip install bs4`.
-
-```python
-from langchain.document_loaders import WebBaseLoader
-loader = WebBaseLoader("https://www.gutenberg.org/files/1727/1727-h/1727-h.htm")
-data = loader.load()
-```
-
-This file is pretty big. Just the preface is 3000 tokens. Which means the full document won't fit into the context for the model. So we need to split it up into smaller pieces.
-
-```python
-from langchain.text_splitter import RecursiveCharacterTextSplitter
-
-text_splitter=RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
-all_splits = text_splitter.split_documents(data)
-```
-
-It's split up, but we have to find the relevant splits and then submit those to the model. We can do this by creating embeddings and storing them in a vector database. We can use Ollama directly to instantiate an embedding model. We will use ChromaDB in this example for a vector database. `pip install chromadb`
-We also need to pull embedding model: `ollama pull nomic-embed-text`
-```python
-from langchain.embeddings import OllamaEmbeddings
-from langchain.vectorstores import Chroma
-oembed = OllamaEmbeddings(base_url="http://localhost:11434", model="nomic-embed-text")
-vectorstore = Chroma.from_documents(documents=all_splits, embedding=oembed)
-```
-
-Now let's ask a question from the document. **Who was Neleus, and who is in his family?** Neleus is a character in the Odyssey, and the answer can be found in our text.
-
-```python
-question="Who is Neleus and who is in Neleus' family?"
-docs = vectorstore.similarity_search(question)
-len(docs)
-```
-
-This will output the number of matches for chunks of data similar to the search.
-
-The next thing is to send the question and the relevant parts of the docs to the model to see if we can get a good answer. But we are stitching two parts of the process together, and that is called a chain. This means we need to define a chain:
-
-```python
-from langchain.chains import RetrievalQA
-qachain=RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
-res = qachain.invoke({"query": question})
-print(res['result'])
-```
-
-The answer received from this chain was:
-
-> Neleus is a character in Homer's "Odyssey" and is mentioned in the context of Penelope's suitors. Neleus is the father of Chloris, who is married to Neleus and bears him several children, including Nestor, Chromius, Periclymenus, and Pero. Amphinomus, the son of Nisus, is also mentioned as a suitor of Penelope and is known for his good natural disposition and agreeable conversation.
-
-It's not a perfect answer, as it implies Neleus married his daughter when actually Chloris "was the youngest daughter to Amphion son of Iasus and king of Minyan Orchomenus, and was Queen in Pylos".
-
-I updated the chunk_overlap for the text splitter to 20 and tried again and got a much better answer:
-
-> Neleus is a character in Homer's epic poem "The Odyssey." He is the husband of Chloris, who is the youngest daughter of Amphion son of Iasus and king of Minyan Orchomenus. Neleus has several children with Chloris, including Nestor, Chromius, Periclymenus, and Pero.
-
-And that is a much better answer.
@@ -1,15 +0,0 @@
-# Running Ollama on NVIDIA Jetson Devices
-
-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. 
-
-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`
- Pull the model you want to use (e.g. mistral): `ollama pull mistral`
- 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).
@@ -1,22 +1,15 @@
-# Ollama Windows Preview
+# Ollama Windows

-Welcome to the Ollama Windows preview.
+Welcome to Ollama for Windows.

 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
+After installing Ollama for Windows, 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 22H2 or newer, Home or Pro
@@ -25,6 +18,32 @@ Logs will often be helpful in diagnosing the problem (see

 Ollama uses unicode characters for progress indication, which may render as unknown squares in some older terminal fonts in Windows 10. If you see this, try changing your terminal font settings.

+## Filesystem Requirements
+
+The Ollama install does not require Administrator, and installs in your home directory by default.  You'll need at least 4GB of space for the binary install.  Once you've installed Ollama, you'll need additional space for storing the Large Language models, which can be tens to hundreds of GB in size.  If your home directory doesn't have enough space, you can change where the binaries are installed, and where the models are stored.
+
+### Changing Install Location
+
+To install the Ollama application in a location different than your home directory, start the installer with the following flag
+
+```powershell
+OllamaSetup.exe /DIR="d:\some\location"
+```
+
+### Changing Model Location
+
+To change where Ollama stores the downloaded models instead of using your home directory, set the environment variable `OLLAMA_MODELS` in your user account.
+
+1. Start the Settings (Windows 11) or Control Panel (Windows 10) application and search for _environment variables_.
+
+2. Click on _Edit environment variables for your account_.
+
+3. Edit or create a new variable for your user account for `OLLAMA_MODELS` where you want the models stored
+
+4. Click OK/Apply to save.
+
+If Ollama is already running, Quit the tray application and relaunch it from the Start menu, or a new terminal started after you saved the environment variables.
+
 ## API Access

 Here's a quick example showing API access from `powershell`
@@ -34,10 +53,6 @@ Here's a quick example showing API access from `powershell`

 ## Troubleshooting

-While we're in preview, `OLLAMA_DEBUG` is always enabled, which adds
-a "view logs" menu item to the app, and increases 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
@@ -52,6 +67,10 @@ the explorer window by hitting `<cmd>+R` and type in:

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

+> [!NOTE]
+> If you have [changed the OLLAMA_MODELS location](#changing-model-location), the installer will not remove your downloaded models
+
+
 ## Standalone CLI

 The easiest way to install Ollama on Windows is to use the `OllamaSetup.exe`
@@ -72,6 +72,7 @@ func Origins() (origins []string) {
 		"app://*",
 		"file://*",
 		"tauri://*",
+		"vscode-webview://*",
 	)

 	return origins
@@ -152,6 +153,8 @@ var (
 	Debug = Bool("OLLAMA_DEBUG")
 	// FlashAttention enables the experimental flash attention feature.
 	FlashAttention = Bool("OLLAMA_FLASH_ATTENTION")
+	// KvCacheType is the quantization type for the K/V cache.
+	KvCacheType = String("OLLAMA_KV_CACHE_TYPE")
 	// NoHistory disables readline history.
 	NoHistory = Bool("OLLAMA_NOHISTORY")
 	// NoPrune disables pruning of model blobs on startup.
@@ -233,6 +236,7 @@ func AsMap() map[string]EnvVar {
 	ret := map[string]EnvVar{
 		"OLLAMA_DEBUG":             {"OLLAMA_DEBUG", Debug(), "Show additional debug information (e.g. OLLAMA_DEBUG=1)"},
 		"OLLAMA_FLASH_ATTENTION":   {"OLLAMA_FLASH_ATTENTION", FlashAttention(), "Enabled flash attention"},
+		"OLLAMA_KV_CACHE_TYPE":     {"OLLAMA_KV_CACHE_TYPE", KvCacheType(), "Quantization type for the K/V cache (default: f16)"},
 		"OLLAMA_GPU_OVERHEAD":      {"OLLAMA_GPU_OVERHEAD", GpuOverhead(), "Reserve a portion of VRAM per GPU (bytes)"},
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host(), "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive(), "The duration that models stay loaded in memory (default \"5m\")"},
@@ -264,9 +268,9 @@ func AsMap() map[string]EnvVar {

 	if runtime.GOOS != "darwin" {
 		ret["CUDA_VISIBLE_DEVICES"] = EnvVar{"CUDA_VISIBLE_DEVICES", CudaVisibleDevices(), "Set which NVIDIA devices are visible"}
-		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible"}
-		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices(), "Set which AMD devices are visible"}
-		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal(), "Set which AMD devices are visible"}
+		ret["HIP_VISIBLE_DEVICES"] = EnvVar{"HIP_VISIBLE_DEVICES", HipVisibleDevices(), "Set which AMD devices are visible by numeric ID"}
+		ret["ROCR_VISIBLE_DEVICES"] = EnvVar{"ROCR_VISIBLE_DEVICES", RocrVisibleDevices(), "Set which AMD devices are visible by UUID or numeric ID"}
+		ret["GPU_DEVICE_ORDINAL"] = EnvVar{"GPU_DEVICE_ORDINAL", GpuDeviceOrdinal(), "Set which AMD devices are visible by numeric ID"}
 		ret["HSA_OVERRIDE_GFX_VERSION"] = EnvVar{"HSA_OVERRIDE_GFX_VERSION", HsaOverrideGfxVersion(), "Override the gfx used for all detected AMD GPUs"}
 		ret["OLLAMA_INTEL_GPU"] = EnvVar{"OLLAMA_INTEL_GPU", IntelGPU(), "Enable experimental Intel GPU detection"}
 	}
@@ -68,6 +68,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://10.0.0.1", []string{
 			"http://10.0.0.1",
@@ -86,6 +87,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://172.16.0.1,https://192.168.0.1", []string{
 			"http://172.16.0.1",
@@ -105,6 +107,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 		{"http://totally.safe,http://definitely.legit", []string{
 			"http://totally.safe",
@@ -124,6 +127,7 @@ func TestOrigins(t *testing.T) {
 			"app://*",
 			"file://*",
 			"tauri://*",
+			"vscode-webview://*",
 		}},
 	}
 	for _, tt := range cases {
@@ -1,6 +1,6 @@
 from langchain.llms import Ollama

-input = input("What is your question?")
+input = input("What is your question?\n> ")
 llm = Ollama(model="llama3.2")
-res = llm.predict(input)
+res = llm.invoke(input)
 print (res)
@@ -1,3 +0,0 @@
-# `modelfile`
-
-This package provides utilities for loading and inspecting model files
@@ -1 +0,0 @@
-package fileutils
@@ -1,18 +1,18 @@
 module github.com/ollama/ollama

-go 1.22.5
+go 1.22.8

 require (
 	github.com/containerd/console v1.0.3
 	github.com/emirpasic/gods v1.18.1
 	github.com/gin-gonic/gin v1.10.0
 	github.com/golang/protobuf v1.5.4 // indirect
-	github.com/google/uuid v1.1.2
+	github.com/google/uuid v1.6.0
 	github.com/olekukonko/tablewriter v0.0.5
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.3.0
+	golang.org/x/sync v0.9.0
 )

 require (
@@ -22,13 +22,14 @@ require (
 	github.com/mattn/go-runewidth v0.0.14
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
+	golang.org/x/image v0.22.0
 )

 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 // indirect
 	github.com/bytedance/sonic/loader v0.1.1 // indirect
 	github.com/chewxy/hm v1.0.0 // indirect
-	github.com/chewxy/math32 v1.10.1 // indirect
+	github.com/chewxy/math32 v1.11.0 // indirect
 	github.com/cloudwego/base64x v0.1.4 // indirect
 	github.com/cloudwego/iasm v0.2.0 // indirect
 	github.com/davecgh/go-spew v1.1.1 // indirect
@@ -72,7 +73,7 @@ require (
 	golang.org/x/net v0.25.0 // indirect
 	golang.org/x/sys v0.20.0
 	golang.org/x/term v0.20.0
-	golang.org/x/text v0.15.0
+	golang.org/x/text v0.20.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )
@@ -21,8 +21,8 @@ github.com/census-instrumentation/opencensus-proto v0.2.1/go.mod h1:f6KPmirojxKA
 github.com/chewxy/hm v1.0.0 h1:zy/TSv3LV2nD3dwUEQL2VhXeoXbb9QkpmdRAVUFiA6k=
 github.com/chewxy/hm v1.0.0/go.mod h1:qg9YI4q6Fkj/whwHR1D+bOGeF7SniIP40VweVepLjg0=
 github.com/chewxy/math32 v1.0.0/go.mod h1:Miac6hA1ohdDUTagnvJy/q+aNnEk16qWUdb8ZVhvCN0=
-github.com/chewxy/math32 v1.10.1 h1:LFpeY0SLJXeaiej/eIp2L40VYfscTvKh/FSEZ68uMkU=
-github.com/chewxy/math32 v1.10.1/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs=
+github.com/chewxy/math32 v1.11.0 h1:8sek2JWqeaKkVnHa7bPVqCEOUPbARo4SGxs6toKyAOo=
+github.com/chewxy/math32 v1.11.0/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs=
 github.com/client9/misspell v0.3.4/go.mod h1:qj6jICC3Q7zFZvVWo7KLAzC3yx5G7kyvSDkc90ppPyw=
 github.com/cloudwego/base64x v0.1.4 h1:jwCgWpFanWmN8xoIUHa2rtzmkd5J2plF/dnLS6Xd/0Y=
 github.com/cloudwego/base64x v0.1.4/go.mod h1:0zlkT4Wn5C6NdauXdJRhSKRlJvmclQ1hhJgA0rcu/8w=
@@ -113,8 +113,9 @@ github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
 github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
-github.com/google/uuid v1.1.2 h1:EVhdT+1Kseyi1/pUmXKaFxYsDNy9RQYkMWRH68J/W7Y=
 github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
+github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/grpc-ecosystem/grpc-gateway v1.16.0/go.mod h1:BDjrQk3hbvj6Nolgz8mAMFbcEtjT1g+wF4CSlocrBnw=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
@@ -230,6 +231,8 @@ golang.org/x/image v0.0.0-20200430140353-33d19683fad8/go.mod h1:FeLwcggjj3mMvU+o
 golang.org/x/image v0.0.0-20200618115811-c13761719519/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20201208152932-35266b937fa6/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
 golang.org/x/image v0.0.0-20210216034530-4410531fe030/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/image v0.22.0 h1:UtK5yLUzilVrkjMAZAZ34DXGpASN8i8pj8g+O+yd10g=
+golang.org/x/image v0.22.0/go.mod h1:9hPFhljd4zZ1GNSIZJ49sqbp45GKK9t6w+iXvGqZUz4=
 golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
 golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
 golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=
@@ -263,8 +266,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E=
-golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=
+golang.org/x/sync v0.9.0 h1:fEo0HyrW1GIgZdpbhCRO0PkJajUS5H9IFUztCgEo2jQ=
+golang.org/x/sync v0.9.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -289,8 +292,8 @@ 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 v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.15.0 h1:h1V/4gjBv8v9cjcR6+AR5+/cIYK5N/WAgiv4xlsEtAk=
-golang.org/x/text v0.15.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+golang.org/x/text v0.20.0 h1:gK/Kv2otX8gz+wn7Rmb3vT96ZwuoxnQlY+HlJVj7Qug=
+golang.org/x/text v0.20.0/go.mod h1:D4IsuqiFMhST5bX19pQ9ikHC2GsaKyk/oF+pn3ducp4=
 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=
@@ -30,6 +30,48 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 	GenerateTestHelper(ctx, t, req, []string{"rayleigh", "scattering"})
 }

+func TestUnicode(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		// DeepSeek has a Unicode tokenizer regex, making it a unicode torture test
+		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
+		Prompt: "天空为什么是蓝色的?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+			// Workaround deepseek context shifting bug
+			"num_ctx":     8192,
+			"num_predict": 2048,
+		},
+	}
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	DoGenerate(ctx, t, client, req, []string{"散射", "频率"}, 120*time.Second, 120*time.Second)
+}
+
+func TestExtendedUnicodeOutput(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "gemma2:2b",
+		Prompt: "Output some smily face emoji",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	DoGenerate(ctx, t, client, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"}, 120*time.Second, 120*time.Second)
+}
+
 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" {
@@ -60,7 +60,8 @@ func TestMultiModelConcurrency(t *testing.T) {
 	for i := 0; i < len(req); i++ {
 		go func(i int) {
 			defer wg.Done()
-			DoGenerate(ctx, t, client, req[i], resp[i], 60*time.Second, 10*time.Second)
+			// Note: CPU based inference can crawl so don't give up too quickly
+			DoGenerate(ctx, t, client, req[i], resp[i], 90*time.Second, 30*time.Second)
 		}(i)
 	}
 	wg.Wait()
@@ -10,7 +10,38 @@ import (
 	"github.com/ollama/ollama/api"
 )

+func TestLongInputContext(t *testing.T) {
+	// Setting NUM_PARALLEL to 1 ensures the allocated context is exactly what
+	// we asked for and there is nothing extra that we could spill over into
+	t.Setenv("OLLAMA_NUM_PARALLEL", "1")
+
+	// Longer needed for small footprint GPUs
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  "llama2",
+		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
+		Stream: &stream,
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+			"num_ctx":     128,
+		},
+	}
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("PullIfMissing failed: %v", err)
+	}
+	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia"}, 120*time.Second, 10*time.Second)
+}
+
 func TestContextExhaustion(t *testing.T) {
+	// Setting NUM_PARALLEL to 1 ensures the allocated context is exactly what
+	// we asked for and there is nothing extra that we could spill over into
+	t.Setenv("OLLAMA_NUM_PARALLEL", "1")
+
 	// Longer needed for small footprint GPUs
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
@@ -12,7 +12,7 @@ import (
 	"github.com/stretchr/testify/require"
 )

-func TestIntegrationMultimodal(t *testing.T) {
+func TestIntegrationLlava(t *testing.T) {
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{
@@ -39,6 +39,33 @@ func TestIntegrationMultimodal(t *testing.T) {
 	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
 }

+func TestIntegrationMllama(t *testing.T) {
+	image, err := base64.StdEncoding.DecodeString(imageEncoding)
+	require.NoError(t, err)
+	req := api.GenerateRequest{
+		// TODO fix up once we publish the final image
+		Model:  "x/llama3.2-vision",
+		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"
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	// mllama models on CPU can be quite slow to start,
+	DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
+}
+
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb
 AAUAAAABAAAAUgEoAAMAAAABAAIAAIdpAAQAAAABAAAAWgAAAAAAAABIAAAAAQAAAEgAAAABAAOgAQADAAAAAQABAACgAgAEAAAAAQAAANKgAwAEAAAAAQAA
 AHgAAAAAXdsepgAAAAlwSFlzAAALEwAACxMBAJqcGAAAAVlpVFh0WE1MOmNvbS5hZG9iZS54bXAAAAAAADx4OnhtcG1ldGEgeG1sbnM6eD0iYWRvYmU6bnM6
@@ -16,7 +16,6 @@ import (
 	"github.com/stretchr/testify/require"

 	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/envconfig"
 )

 func TestMaxQueue(t *testing.T) {
@@ -27,12 +26,8 @@ func TestMaxQueue(t *testing.T) {

 	// Note: This test can be quite slow when running in CPU mode, so keep the threadCount low unless your on GPU
 	// Also note that by default Darwin can't sustain > ~128 connections without adjusting limits
-	threadCount := 32
-	if maxQueue := envconfig.MaxQueue(); maxQueue != 0 {
-		threadCount = int(maxQueue)
-	} else {
-		t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
-	}
+	threadCount := 16
+	t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))

 	req := api.GenerateRequest{
 		Model:  "orca-mini",
@@ -55,7 +55,7 @@ go build -tags avx,cuda .

 ### ROCm

-Install the [CUDA toolkit v11.3.1](https://developer.nvidia.com/cuda-11-3-1-download-archive):
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).

 ```shell
 make ggml_hipblas.so
@@ -77,7 +77,7 @@ go build -tags avx,cuda .

 ### ROCm

-Install [ROCm 5.7.1](https://rocm.docs.amd.com/en/docs-5.7.1/).
+Install [ROCm](https://rocm.docs.amd.com/en/latest/).

 ```shell
 make ggml_hipblas.dll
@@ -93,7 +93,7 @@ make -j

 ## Vendoring

-Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model.  While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit.  A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.
+Ollama currently vendors [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [ggml](https://github.com/ggerganov/ggml) through a vendoring model. While we generally strive to contribute changes back upstream to avoid drift, we cary a small set of patches which are applied to the tracking commit. A set of make targets are available to aid developers in updating to a newer tracking commit, or to work on changes.

 If you update the vendoring code, start by running the following command to establish the tracking llama.cpp repo in the `./vendor/` directory.

@@ -105,35 +105,35 @@ make apply-patches

 **Pin to new base commit**

-To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring.env`
+To update to a newer base commit, select the upstream git tag or commit and update `llama/vendoring`

 #### Applying patches

 When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.

-Start by applying the patches.  If any of the patches have conflicts, the `git am` will stop at the first failure.
+Start by applying the patches. If any of the patches have conflicts, the `git am` will stop at the first failure.

 ```
 make apply-patches
 ```

-If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed.  Save the file(s) and continue the patch series with `git am --continue` .  If any additional patches fail, follow the same pattern until the full patch series is applied.  Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.
+If you see an error message about a conflict, go into the `./vendor/` directory, and perform merge resolution using your preferred tool to the patch commit which failed. Save the file(s) and continue the patch series with `git am --continue` . If any additional patches fail, follow the same pattern until the full patch series is applied. Once finished, run a final `create-patches` and `sync` target to ensure everything is updated.

 ```
 make create-patches sync
 ```

-Build and test Ollama, and make any necessary changes to the Go code based on the new base commit.  Submit your PR to the Ollama repo.
+Build and test Ollama, and make any necessary changes to the Go code based on the new base commit. Submit your PR to the Ollama repo.

 ### Generating Patches

-When working on new fixes or features that impact vendored code, use the following model.  First get a clean tracking repo with all current patches applied:
+When working on new fixes or features that impact vendored code, use the following model. First get a clean tracking repo with all current patches applied:

 ```
 make apply-patches
 ```

-Now edit the upstream native code in the `./vendor/` directory.  You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing.  Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:
+Now edit the upstream native code in the `./vendor/` directory. You do not need to commit every change in order to build, a dirty working tree in the tracking repo is OK while developing. Simply save in your editor, and run the following to refresh the vendored code with your changes, build the backend(s) and build ollama:

 ```
 make sync
@@ -142,9 +142,9 @@ go build .
 ```

 > [!IMPORTANT]
-> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo.  It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).
+> Do **NOT** run `apply-patches` while you're iterating as that will reset the tracking repo. It will detect a dirty tree and abort, but if your tree is clean and you accidentally ran this target, use `git reflog` to recover your commit(s).

-Iterate until you're ready to submit PRs.  Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with
+Iterate until you're ready to submit PRs. Once your code is ready, commit a change in the `./vendor/` directory, then generate the patches for ollama with

 ```
 make create-patches
@@ -157,4 +157,4 @@ In your `./vendor/` directory, create a branch, and cherry-pick the new commit t

 Commit the changes in the ollama repo and submit a PR to Ollama, which will include the vendored code update with your change, along with the patches.

-After your PR upstream is merged, follow the **Updating Base Commit** instructions above, however first remove your patch before running `apply-patches` since the new base commit contains your change already.
+After your PR upstream is merged, follow the **Updating Base Commit** instructions above, however first remove your patch before running `apply-patches` since the new base commit contains your change already.
@@ -2296,6 +2296,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
        case GGML_OP_PAD:
            ggml_cuda_op_pad(ctx, dst);
            break;
+        case GGML_OP_UNPAD:
+            ggml_cuda_op_unpad(ctx, dst);
+            break;
        case GGML_OP_ARANGE:
            ggml_cuda_op_arange(ctx, dst);
            break;
@@ -3018,6 +3021,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
        case GGML_OP_GROUP_NORM:
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_LEAKY_RELU:
@@ -73,3 +73,49 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
 }
+
+static __global__ void unpad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    }
+}
+
+static void unpad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    unpad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    unpad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
@@ -29,3 +29,4 @@
 #define CUDA_PAD_BLOCK_SIZE 256

 void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@@ -2055,6 +2055,51 @@ kernel void kernel_pad_f32(
    }
 }

+kernel void kernel_unpad_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            }
+        }
+
+        return;
+    }
+}
+
 kernel void kernel_arange_f32(
    device        char * dst,
    constant   int64_t & ne0,
@@ -219,6 +219,7 @@ enum ggml_metal_kernel_type {
    GGML_METAL_KERNEL_TYPE_IM2COL_F32,
    GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
    GGML_METAL_KERNEL_TYPE_PAD_F32,
+    GGML_METAL_KERNEL_TYPE_UNPAD_F32,
    GGML_METAL_KERNEL_TYPE_ARANGE_F32,
    GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
    GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
@@ -715,6 +716,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) {
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32,                     unpad_f32,                        true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
@@ -872,6 +874,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
            return false;
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_ARGSORT:
@@ -2681,6 +2684,36 @@ static void ggml_metal_encode_node(

                const int nth = MIN(1024, ne0);

+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_UNPAD:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UNPAD_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN(1024, ne0);
+
                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
            } break;
        case GGML_OP_ARANGE:
@@ -3023,6 +3023,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
    "POOL_2D_BACK",
    "UPSCALE",
    "PAD",
+    "UNPAD",
    "ARANGE",
    "TIMESTEP_EMBEDDING",
    "ARGSORT",
@@ -3056,7 +3057,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
    "OPT_STEP_ADAMW",
 };

-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");

 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
    "none",
@@ -3117,6 +3118,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
    "pool_2d_back(x)",
    "upscale(x)",
    "pad(x)",
+    "unpad(x)",
    "arange(start, stop, step)",
    "timestep_embedding(timesteps, dim, max_period)",
    "argsort(x)",
@@ -3150,7 +3152,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
    "adamw(x)",
 };

-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");

 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");

@@ -6981,6 +6983,32 @@ struct ggml_tensor * ggml_pad(
    return result;
 }

+// ggml_unpad
+
+struct ggml_tensor * ggml_unpad(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+    bool is_node = false;
+
+    if (a->grad) {
+        GGML_ABORT("fatal error"); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] - p0,
+            a->ne[1] - p1,
+            a->ne[2] - p2,
+            a->ne[3] - p3);
+
+    result->op = GGML_OP_UNPAD;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange

 struct ggml_tensor * ggml_arange(
@@ -15338,6 +15366,58 @@ static void ggml_compute_forward_pad(
    }
 }

+static void ggml_compute_forward_unpad_f32(
+    const struct ggml_compute_params *params,
+    struct ggml_tensor *dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_unpad(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_unpad_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}

 // ggml_compute_forward_arange

@@ -17320,6 +17400,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_pad(params, tensor);
            } break;
+        case GGML_OP_UNPAD:
+            {
+                ggml_compute_forward_unpad(params, tensor);
+            } break;
        case GGML_OP_ARANGE:
            {
                ggml_compute_forward_arange(params, tensor);
@@ -18395,6 +18479,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
            {
                GGML_ABORT("fatal error"); // TODO: not implemented
            }
+        case GGML_OP_UNPAD:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
        case GGML_OP_ARANGE:
            {
                GGML_ABORT("fatal error"); // TODO: not implemented
@@ -19191,6 +19279,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
            } break;
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_ARGSORT:
@@ -532,6 +532,7 @@ extern "C" {
        GGML_OP_POOL_2D_BACK,
        GGML_OP_UPSCALE, // nearest interpolate
        GGML_OP_PAD,
+        GGML_OP_UNPAD,
        GGML_OP_ARANGE,
        GGML_OP_TIMESTEP_EMBEDDING,
        GGML_OP_ARGSORT,
@@ -1790,6 +1791,15 @@ extern "C" {
            int                  p2,
            int                  p3);

+    // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
+    GGML_API struct ggml_tensor * ggml_unpad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
    // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
    // timesteps: [N,]
    // return: [N, dim]
@@ -415,7 +415,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
            case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
                regex_exprs = {
                    "[\r\n]",
-                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿǄ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿﬀ-ﬆﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿǄ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿﬀ-ﬆﬓ-ﬗＡ-Ｚａ-ｚ\U00010400-\U0001044f𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
                    "\\s?[!-/:-~！-／：-～‘-‟　-。]+",
                    "\\s+$",
                    "[一-龥ࠀ-一가-퟿]+",
@@ -195,6 +195,7 @@ static std::string format(const char * fmt, ...) {

 enum llm_arch {
    LLM_ARCH_LLAMA,
+    LLM_ARCH_MLLAMA,
    LLM_ARCH_FALCON,
    LLM_ARCH_BAICHUAN,
    LLM_ARCH_GROK,
@@ -249,6 +250,7 @@ enum llm_arch {

 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_LLAMA,           "llama"        },
+    { LLM_ARCH_MLLAMA,          "mllama"       },
    { LLM_ARCH_FALCON,          "falcon"       },
    { LLM_ARCH_GROK,            "grok"         },
    { LLM_ARCH_GPT2,            "gpt2"         },
@@ -356,6 +358,7 @@ enum llm_kv {
    LLM_KV_ATTENTION_SLIDING_WINDOW,
    LLM_KV_ATTENTION_SCALE,
    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,

    LLM_KV_ROPE_DIMENSION_COUNT,
    LLM_KV_ROPE_FREQ_BASE,
@@ -465,6 +468,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                    },
    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },

    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -639,6 +643,14 @@ enum llm_tensor {
    LLM_TENSOR_CLS,
    LLM_TENSOR_CLS_OUT,
    LLM_TENSOR_BSKCN_TV,
+    LLM_TENSOR_CROSS_ATTN_K_NORM,
+    LLM_TENSOR_CROSS_ATTN_K_PROJ,
+    LLM_TENSOR_CROSS_ATTN_O_PROJ,
+    LLM_TENSOR_CROSS_ATTN_Q_NORM,
+    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
+    LLM_TENSOR_CROSS_ATTN_V_PROJ,
+    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
+    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
 };

 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -668,6 +680,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
        },
    },
+    {
+        LLM_ARCH_MLLAMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
+            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
+            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
+            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
+            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
+            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
+            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
+            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
+        },
+    },
    {
        LLM_ARCH_BAICHUAN,
        {
@@ -2416,6 +2462,7 @@ enum e_model {
    MODEL_40B,
    MODEL_65B,
    MODEL_70B,
+    MODEL_90B,
    MODEL_236B,
    MODEL_314B,
    MODEL_SMALL,
@@ -2460,6 +2507,7 @@ struct llama_hparams {
    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;

    std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;

    uint32_t n_layer_dense_lead = 0;
    uint32_t n_lora_q = 0;
@@ -2528,10 +2576,11 @@ struct llama_hparams {
        if (this->n_expert      != other.n_expert)      return true;
        if (this->n_expert_used != other.n_expert_used) return true;

-        if (this->n_head_arr    != other.n_head_arr)    return true;
-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+        if (this->n_head_arr        != other.n_head_arr)        return true;
+        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
+        if (this->n_ff_arr          != other.n_ff_arr)          return true;
+        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
+        if (this->cross_attn_layers != other.cross_attn_layers) return true;

        if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
        if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
@@ -2649,6 +2698,10 @@ struct llama_hparams {

        GGML_ABORT("fatal error");
    }
+
+    bool cross_attention_layers(uint32_t il) const {
+        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+    }
 };

 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
@@ -2678,6 +2731,9 @@ struct llama_cparams {
    bool offload_kqv;
    bool flash_attn;
    bool no_perf;
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    bool cross_attn = false;

    enum llama_pooling_type pooling_type;

@@ -2832,6 +2888,16 @@ struct llama_layer {
    struct ggml_tensor * ffn_down_scale;

    struct ggml_tensor * bskcn_tv;
+
+    // cross attention
+    struct ggml_tensor * cross_attn_k_norm;
+    struct ggml_tensor * cross_attn_k_proj;
+    struct ggml_tensor * cross_attn_o_proj;
+    struct ggml_tensor * cross_attn_q_norm;
+    struct ggml_tensor * cross_attn_q_proj;
+    struct ggml_tensor * cross_attn_v_proj;
+    struct ggml_tensor * cross_attn_attn_gate;
+    struct ggml_tensor * cross_attn_mlp_gate;
 };

 // very similar to llama_batch,
@@ -3478,6 +3544,8 @@ struct llama_context {
    struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
    struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
    struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+
+    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
 };

 struct llama_lora_weight {
@@ -3712,6 +3780,18 @@ static bool llama_kv_cache_init(
    cache.v_l.reserve(n_layer);

    for (int i = 0; i < (int) n_layer; i++) {
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
+            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
+            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+            ggml_format_name(k, "cache_k_l%d", i);
+            ggml_format_name(v, "cache_v_l%d", i);
+            cache.k_l.push_back(k);
+            cache.v_l.push_back(v);
+            continue;
+        }
+
        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();

@@ -5486,12 +5566,14 @@ static void llm_load_hparams(
    }

    // zero-out the per-layer hparams
-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(hparams.n_head_arr.begin(),             hparams.n_head_arr.end(),        0);
+    std::fill(hparams.n_head_kv_arr.begin(),          hparams.n_head_kv_arr.end(),     0);
+    std::fill(hparams.n_ff_arr.begin(),               hparams.n_ff_arr.end(),          0);
+    std::fill(hparams.cross_attn_layers.begin(),      hparams.cross_attn_layers.end(), -1);

-    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer);
-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
+    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);

    // n_head_kv is optional, default to n_head
    hparams.n_head_kv_arr = hparams.n_head_arr;
@@ -5540,7 +5622,7 @@ static void llm_load_hparams(

        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);

-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.arch == LLM_ARCH_FALCON) {
            if (hparams.n_rot != hparams.n_embd_head_k) {
                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
            }
@@ -5580,6 +5662,16 @@ static void llm_load_hparams(
                    }
                }
            } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_11B; break;
+                    case 100: model.type = e_model::MODEL_90B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
        case LLM_ARCH_MINICPM:
            {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -7275,6 +7367,55 @@ static bool llm_load_tensors(
                        layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                    }
                } break;
+            case LLM_ARCH_MLLAMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        if (hparams.cross_attention_layers(i)) {
+                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
+                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
+                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
+                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
+                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
+                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        } else {
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                            layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        }
+                    }
+                } break;
            case LLM_ARCH_GROK:
                {
                    if (n_expert == 0) {
@@ -9119,7 +9260,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam

        if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
            model.hparams.n_vocab != model.vocab.id_to_token.size()) {
-            throw std::runtime_error("vocab size mismatch");
+            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
        }

        if (params.vocab_only) {
@@ -9219,6 +9360,21 @@ static struct ggml_tensor * llm_build_inp_embd(
    return inpL;
 }

+static struct ggml_tensor * llm_build_inp_cross_attn_state(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+        const llama_hparams & hparams,
+         const llm_build_cb & cb) {
+    const int64_t n_embd = hparams.n_embd;
+
+    struct ggml_tensor * inpCAS = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
+    cb(inpCAS, "inp_cross_attn_state", -1);
+    ggml_set_input(inpCAS);
+    lctx.inp_cross_attn_state = inpCAS;
+
+    return inpCAS;
+}
+
 static void llm_build_kv_store(
        struct ggml_context * ctx,
        const llama_hparams & hparams,
@@ -10193,6 +10349,7 @@ struct llm_build_context {
        lctx.inp_pos_bucket    = nullptr;
        lctx.inp_embd_enc      = nullptr;
        lctx.inp_KQ_mask_cross = nullptr;
+        lctx.inp_cross_attn_state = nullptr;
    }

    void free() {
@@ -10780,6 +10937,239 @@ struct llm_build_context {
                LLM_NORM_RMS, cb, -1);
        cb(cur, "result_norm", -1);

+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mllama() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        struct ggml_tensor * inpCAS;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.cross_attention_layers(il)) {
+                if (!batch.embd && !cparams.cross_attn) {
+                    continue;
+                }
+
+                // cross attention layer
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 2, 1, 3));
+                cb(Qcur, "Qcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur, * Vcur;
+                if (batch.embd) {
+                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
+                    cb(Kcur, "Kcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
+
+                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
+                    cb(Vcur, "Vcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il]));
+                } else {
+                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
+                    cb(Kcur, "Kcur (view)", il);
+
+                    Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]);
+                    cb(Vcur, "Vcur (view)", il);
+                }
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
+                cb(kq, "kq", il);
+
+                // TODO: apply causal masks
+                struct ggml_tensor * kq_soft_max = ggml_soft_max_ext(ctx0, kq, nullptr, 1.f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+                cb(kq_soft_max, "kq_soft_max", il);
+
+                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
+                cb(Vcur, "Vcur", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
+                cb(cur, "cur", il);
+
+                // TODO: do this in place once?
+                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                // TODO: do this inplace once?
+                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            } else {
+                // self attention layer
+
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
        // lm_head
        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
        cb(cur, "result_output", -1);
@@ -16527,6 +16917,10 @@ static struct ggml_cgraph * llama_build_graph(
            {
                result = llm.build_llama();
            } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                result = llm.build_mllama();
+            } break;
        case LLM_ARCH_BAICHUAN:
            {
                result = llm.build_baichuan();
@@ -16787,10 +17181,19 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
    }

    if (batch.embd) {
-        const int64_t n_embd   = hparams.n_embd;
-        const int64_t n_tokens = batch.n_tokens;
+        if (lctx.inp_cross_attn_state && lctx.inp_cross_attn_state->buffer) {
+            ggml_backend_tensor_set(lctx.inp_cross_attn_state, batch.embd, 0, ggml_nbytes(lctx.inp_cross_attn_state));
+            // zero out inp_embd since it's not used
+            float * inp_embd_data = (float *)lctx.inp_embd->data;
+            for (int i = 0; i < ggml_nelements(lctx.inp_embd); ++i) {
+                inp_embd_data[i] = 0.0f;
+            }
+        } else {
+            const int64_t n_embd   = hparams.n_embd;
+            const int64_t n_tokens = batch.n_tokens;

-        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+            ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        }
    }

    if (batch.pos && lctx.inp_pos) {
@@ -17371,7 +17774,7 @@ static int llama_decode_internal(
        n_outputs = 1;
    }

-    lctx.sbatch.from_batch(batch_all, n_embd,
+    lctx.sbatch.from_batch(batch_all, batch_all.n_embd,
        /* simple_split */ !kv_self.recurrent,
        /* logits_all   */ n_outputs == n_tokens_all);

@@ -17664,7 +18067,7 @@ static int llama_encode_internal(

    const int64_t n_embd = hparams.n_embd;

-    lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    lctx.sbatch.from_batch(batch, batch.n_embd, /* simple_split */ true, /* logits_all */ true);

    const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens);

@@ -18674,7 +19077,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
        if (llama_model_has_encoder(&model)) {
            n_attn_layer *= 3;
        }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        if (qs.n_attention_wv != n_attn_layer) {
+            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
+        }
    }

    size_t total_size_org = 0;
@@ -19840,6 +20245,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {

        // use what we call a normal RoPE, operating on pairs of consecutive head values
        case LLM_ARCH_LLAMA:
+        case LLM_ARCH_MLLAMA:
        case LLM_ARCH_BAICHUAN:
        case LLM_ARCH_STARCODER:
        case LLM_ARCH_PLAMO:
@@ -21256,6 +21662,10 @@ void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) {
    ctx->cparams.causal_attn = causal_attn;
 }

+void llama_set_cross_attention(struct llama_context * ctx, bool cross_attention) {
+    ctx->cparams.cross_attn = cross_attention;
+}
+
 struct llama_batch llama_batch_get_one(
             llama_token * tokens,
                 int32_t   n_tokens,
@@ -21265,6 +21675,7 @@ struct llama_batch llama_batch_get_one(
        /*n_tokens       =*/ n_tokens,
        /*tokens         =*/ tokens,
        /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
        /*pos            =*/ nullptr,
        /*n_seq_id       =*/ nullptr,
        /*seq_id         =*/ nullptr,
@@ -21280,6 +21691,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
        /*n_tokens       =*/ 0,
        /*tokens         =*/ nullptr,
        /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
        /*pos            =*/ nullptr,
        /*n_seq_id       =*/ nullptr,
        /*seq_id         =*/ nullptr,
@@ -21291,6 +21703,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_

    if (embd) {
        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+        batch.n_embd = embd;
    } else {
        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
    }
@@ -21,6 +21,8 @@ package llama
 #cgo cuda CFLAGS: -fPIE -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo cuda CXXFLAGS: -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo cuda CXXFLAGS: -DGGML_USE_CUDA -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
+#cgo cuda_jetpack5 LDFLAGS: -lggml_cuda_jetpack5 -L/usr/local/cuda-11/lib64
+#cgo cuda_jetpack6 LDFLAGS: -lggml_cuda_jetpack6 -L/usr/local/cuda-12/lib64
 #cgo cuda_v11 LDFLAGS: -lggml_cuda_v11 -L/usr/local/cuda-11/lib64
 #cgo cuda_v12 LDFLAGS: -lggml_cuda_v12 -L/usr/local/cuda-12/lib64
 #cgo darwin,amd64 CFLAGS: -Wno-incompatible-pointer-types-discards-qualifiers
@@ -36,8 +38,8 @@ package llama
 #cgo linux CXXFLAGS: -D_GNU_SOURCE
 #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/Linux/amd64
 #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/Linux/amd64
-#cgo linux,arm64 CFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA -D__ARM_FEATURE_MATMUL_INT8
-#cgo linux,arm64 CXXFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA -D__ARM_FEATURE_MATMUL_INT8
+#cgo linux,arm64 CFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA
+#cgo linux,arm64 CXXFLAGS: -D__aarch64__ -D__ARM_NEON -D__ARM_FEATURE_FMA
 #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/Linux/arm64
 #cgo linux,arm64,sve CFLAGS: -march=armv8.6-a+sve
 #cgo linux,arm64,sve CXXFLAGS: -march=armv8.6-a+sve
@@ -46,8 +48,8 @@ package llama
 #cgo rocm CFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm CXXFLAGS: -DGGML_USE_CUDA -DGGML_USE_HIPBLAS -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_MMV_Y=1 -DGGML_BUILD=1
 #cgo rocm LDFLAGS: -L${SRCDIR} -lggml_rocm -lhipblas -lamdhip64 -lrocblas
-#cgo windows CFLAGS: -Wno-discarded-qualifiers
-#cgo windows CFLAGS: -Wno-discarded-qualifiers
+#cgo windows CFLAGS: -Wno-discarded-qualifiers -D_WIN32_WINNT=0x602
+#cgo windows CXXFLAGS: -D_WIN32_WINNT=0x602
 #cgo windows LDFLAGS: -lmsvcrt
 #cgo windows LDFLAGS: -lmsvcrt -static-libstdc++ -static-libgcc -static
 #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/Windows/amd64
@@ -62,10 +64,23 @@ package llama
 #include <stdlib.h>
 #include "llama.h"
 #include "clip.h"
+#include "ggml.h"
 #include "llava.h"
+#include "mllama.h"
 #include "sampling_ext.h"

 bool llamaProgressCallback(float progress, void *user_data);
+
+typedef enum {COMP_UNKNOWN,COMP_GCC,COMP_CLANG} COMPILER;
+COMPILER inline get_compiler() {
+#if defined(__clang__)
+	return COMP_CLANG;
+#elif defined(__GNUC__)
+	return COMP_GCC;
+#else
+	return UNKNOWN_COMPILER;
+#endif
+}
 */
 import "C"

@@ -75,6 +90,7 @@ import (
 	"fmt"
 	"runtime"
 	"runtime/cgo"
+	"slices"
 	"strings"
 	"unsafe"
 )
@@ -86,14 +102,45 @@ func BackendInit() {
 }

 func PrintSystemInfo() string {
-	return C.GoString(C.llama_print_system_info())
+	var compiler string
+	switch C.get_compiler() {
+	case C.COMP_UNKNOWN:
+		compiler = "cgo(unknown_compiler)"
+	case C.COMP_GCC:
+		compiler = "cgo(gcc)"
+	case C.COMP_CLANG:
+		compiler = "cgo(clang)"
+	}
+	return C.GoString(C.llama_print_system_info()) + compiler
+}
+
+func GetModelArch(modelPath string) (string, error) {
+	mp := C.CString(modelPath)
+	defer C.free(unsafe.Pointer(mp))
+
+	gguf_ctx := C.gguf_init_from_file(mp, C.struct_gguf_init_params{no_alloc: true, ctx: (**C.struct_ggml_context)(C.NULL)})
+	if gguf_ctx == nil {
+		return "", errors.New("unable to load model file")
+	}
+	defer C.gguf_free(gguf_ctx)
+
+	key := C.CString("general.architecture")
+	defer C.free(unsafe.Pointer(key))
+	arch_index := C.gguf_find_key(gguf_ctx, key)
+	if int(arch_index) < 0 {
+		return "", errors.New("unknown model architecture")
+	}
+
+	arch := C.gguf_get_val_str(gguf_ctx, arch_index)
+
+	return C.GoString(arch), nil
 }

 type ContextParams struct {
 	c C.struct_llama_context_params
 }

-func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, flashAttention bool) ContextParams {
+func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, flashAttention bool, kvCacheType string) ContextParams {
 	params := C.llama_context_default_params()
 	params.n_ctx = C.uint(numCtx)
 	params.n_batch = C.uint(batchSize)
@@ -102,17 +149,34 @@ func NewContextParams(numCtx int, batchSize int, numSeqMax int, threads int, fla
 	params.n_threads_batch = params.n_threads
 	params.embeddings = C.bool(true)
 	params.flash_attn = C.bool(flashAttention)
+	params.type_k = kvCacheTypeFromStr(strings.ToLower(kvCacheType))
+	params.type_v = kvCacheTypeFromStr(strings.ToLower(kvCacheType))
+
 	return ContextParams{c: params}
 }

+// kvCacheTypeFromStr converts a string cache type to the corresponding GGML type value
+func kvCacheTypeFromStr(s string) C.enum_ggml_type {
+	if s == "" {
+		return C.GGML_TYPE_F16
+	}
+
+	switch s {
+	case "q8_0":
+		return C.GGML_TYPE_Q8_0
+	case "q4_0":
+		return C.GGML_TYPE_Q4_0
+	default:
+		return C.GGML_TYPE_F16
+	}
+}
+
 type Context struct {
 	c          *C.struct_llama_context
 	numThreads int
 }

-func (c *Context) KvCacheClear() {
-	C.llama_kv_cache_clear(c.c)
-}
+var ErrKvCacheFull = errors.New("could not find a kv cache slot")

 func (c *Context) Decode(batch *Batch) error {
 	// Positive return values does not mean a fatal error, but rather a warning.
@@ -126,7 +190,7 @@ func (c *Context) Decode(batch *Batch) error {
 	}

 	if code > 0 {
-		return fmt.Errorf("could not find a KV slot for the batch - try reducing the size of the batch or increase the context. code: %d", code)
+		return ErrKvCacheFull
 	}

 	return nil
@@ -136,10 +200,6 @@ func (c *Context) Model() *Model {
 	return &Model{c: C.llama_get_model(c.c)}
 }

-func (c *Context) GetLogitsIth(i int) []float32 {
-	return unsafe.Slice((*float32)(unsafe.Pointer(C.llama_get_logits_ith(c.c, C.int(i)))), c.Model().NumVocab())
-}
-
 func (c *Context) KvCacheSeqAdd(seqId int, p0 int, p1 int, delta int) {
 	C.llama_kv_cache_seq_add(c.c, C.int(seqId), C.int(p0), C.int(p1), C.int(delta))
 }
@@ -152,6 +212,14 @@ func (c *Context) KvCacheSeqCp(srcSeqId int, dstSeqId int, p0 int, p1 int) {
 	C.llama_kv_cache_seq_cp(c.c, C.int(srcSeqId), C.int(dstSeqId), C.int(p0), C.int(p1))
 }

+func (c *Context) KvCacheClear() {
+	C.llama_kv_cache_clear(c.c)
+}
+
+func (c *Context) KvCacheDefrag() {
+	C.llama_kv_cache_defrag(c.c)
+}
+
 // Get the embeddings for a sequence id
 func (c *Context) GetEmbeddingsSeq(seqId int) []float32 {
 	embeddings := unsafe.Pointer(C.llama_get_embeddings_seq(c.c, C.int(seqId)))
@@ -163,7 +231,12 @@ func (c *Context) GetEmbeddingsSeq(seqId int) []float32 {
 }

 func (c *Context) GetEmbeddingsIth(i int) []float32 {
-	return unsafe.Slice((*float32)(unsafe.Pointer(C.llama_get_embeddings_ith(c.c, C.int32_t(i)))), c.Model().NEmbd())
+	embeddings := unsafe.Pointer(C.llama_get_embeddings_ith(c.c, C.int32_t(i)))
+	if embeddings == nil {
+		return nil
+	}
+
+	return unsafe.Slice((*float32)(embeddings), c.Model().NEmbd())
 }

 type ModelParams struct {
@@ -184,7 +257,7 @@ func llamaProgressCallback(progress C.float, userData unsafe.Pointer) C.bool {
 	return true
 }

-func LoadModelFromFile(modelPath string, params ModelParams) *Model {
+func LoadModelFromFile(modelPath string, params ModelParams) (*Model, error) {
 	cparams := C.llama_model_default_params()
 	cparams.n_gpu_layers = C.int(params.NumGpuLayers)
 	cparams.main_gpu = C.int32_t(params.MainGpu)
@@ -214,18 +287,28 @@ func LoadModelFromFile(modelPath string, params ModelParams) *Model {
 		cparams.progress_callback_user_data = unsafe.Pointer(&handle)
 	}

-	return &Model{c: C.llama_load_model_from_file(C.CString(modelPath), cparams)}
+	m := Model{c: C.llama_load_model_from_file(C.CString(modelPath), cparams)}
+	if m.c == nil {
+		return nil, fmt.Errorf("unable to load model: %s", modelPath)
+	}
+
+	return &m, nil
 }

 func FreeModel(model *Model) {
 	C.llama_free_model(model.c)
 }

-func NewContextWithModel(model *Model, params ContextParams) *Context {
-	return &Context{
+func NewContextWithModel(model *Model, params ContextParams) (*Context, error) {
+	c := Context{
 		c:          C.llama_new_context_with_model(model.c, params.c),
 		numThreads: int(params.c.n_threads),
 	}
+	if c.c == nil {
+		return nil, errors.New("unable to create llama context")
+	}
+
+	return &c, nil
 }

 func (m *Model) NumVocab() int {
@@ -245,6 +328,9 @@ func (m *Model) ApplyLoraFromFile(context *Context, loraPath string, scale float
 	defer C.free(unsafe.Pointer(cLoraPath))

 	loraAdapter := C.llama_lora_adapter_init(m.c, cLoraPath)
+	if loraAdapter == nil {
+		return errors.New("unable to load lora")
+	}

 	err := -1
 	if loraAdapter != nil {
@@ -260,18 +346,40 @@ func (m *Model) ApplyLoraFromFile(context *Context, loraPath string, scale float
 type Batch struct {
 	c         C.struct_llama_batch
 	batchSize int
+	maxSeq    int
 	embedSize int
 }

-// Creates a new batch for either word tokens if embed is 0 or
-// image embeddings if embed is specified. Batches cannot contain
-// both types at the same time
-func NewBatch(nTokens int, embed int, maxSeq int) *Batch {
-	return &Batch{
-		c:         C.llama_batch_init(C.int(nTokens), C.int(embed), C.int(maxSeq)),
-		batchSize: nTokens,
-		embedSize: embed,
+// Creates a new batch for either word tokens or image embeddings (if embedSize is non-zero).
+// Batches cannot contain both types at the same time. batchSize is the maximum number of entries
+// that can be added per sequence
+func NewBatch(batchSize int, maxSeq int, embedSize int) (*Batch, error) {
+	b := Batch{
+		c:         C.llama_batch_init(C.int(batchSize*maxSeq), C.int(embedSize), C.int(maxSeq)),
+		batchSize: batchSize,
+		maxSeq:    maxSeq,
+		embedSize: embedSize,
 	}
+
+	// Check to see if any of the allocations in llama_batch_init() failed
+	nilPointer := (embedSize == 0 && b.c.token == nil) || (embedSize != 0 && b.c.embd == nil) ||
+		b.c.pos == nil || b.c.n_seq_id == nil || b.c.seq_id == nil || b.c.logits == nil ||
+		slices.Contains(unsafe.Slice(b.c.seq_id, b.allocSize()), nil)
+
+	if nilPointer {
+		C.llama_batch_free(b.c)
+		return nil, fmt.Errorf("unable to allocate batch (batchSize=%v maxSeq=%v embedSize=%v)", batchSize, maxSeq, embedSize)
+	}
+
+	return &b, nil
+}
+
+func (b *Batch) Size() int {
+	return b.batchSize
+}
+
+func (b *Batch) allocSize() int {
+	return b.batchSize * b.maxSeq
 }

 func (b *Batch) NumTokens() int {
@@ -286,21 +394,23 @@ func (b *Batch) IsEmbedding() bool {
 // when the batch was initialized. The other argument will be ignored. Adds to the
 // batch with the given position for the given sequence ids, and optionally instructs
 // to include logits.
-func (b *Batch) Add(token int, embed []float32, pos int, seqIds []int, logits bool) {
+func (b *Batch) Add(token int, embed []float32, pos int, logits bool, seqIds ...int) {
 	if !b.IsEmbedding() {
-		unsafe.Slice(b.c.token, b.batchSize)[b.c.n_tokens] = C.llama_token(token)
+		unsafe.Slice(b.c.token, b.allocSize())[b.c.n_tokens] = C.llama_token(token)
 	} else {
-		copy(unsafe.Slice((*float32)(b.c.embd), b.batchSize*b.embedSize)[int(b.c.n_tokens)*b.embedSize:], embed)
+		copy(unsafe.Slice((*float32)(b.c.embd), b.allocSize()*b.embedSize)[int(b.c.n_tokens)*b.embedSize:], embed)
 	}
-	unsafe.Slice(b.c.pos, b.batchSize)[b.c.n_tokens] = C.llama_pos(pos)
-	unsafe.Slice(b.c.n_seq_id, b.batchSize)[b.c.n_tokens] = C.int(len(seqIds))
+	unsafe.Slice(b.c.pos, b.allocSize())[b.c.n_tokens] = C.llama_pos(pos)
+	unsafe.Slice(b.c.n_seq_id, b.allocSize())[b.c.n_tokens] = C.int(len(seqIds))

 	for i, s := range seqIds {
-		unsafe.Slice((unsafe.Slice(b.c.seq_id, b.batchSize)[b.c.n_tokens]), C.int(len(seqIds)))[i] = C.int32_t(s)
+		unsafe.Slice((unsafe.Slice(b.c.seq_id, b.allocSize())[b.c.n_tokens]), C.int(len(seqIds)))[i] = C.int32_t(s)
 	}

 	if logits {
-		unsafe.Slice(b.c.logits, b.batchSize)[b.c.n_tokens] = 1
+		unsafe.Slice(b.c.logits, b.allocSize())[b.c.n_tokens] = 1
+	} else {
+		unsafe.Slice(b.c.logits, b.allocSize())[b.c.n_tokens] = 0
 	}

 	b.c.n_tokens += 1
@@ -410,29 +520,42 @@ func Quantize(infile, outfile string, ftype uint32) error {
 	return nil
 }

-// llava
+// vision processing
 type ClipContext struct {
 	c *C.struct_clip_ctx
 }

-func NewClipContext(modelPath string) *ClipContext {
+func NewClipContext(llamaContext *Context, modelPath string) (*ClipContext, error) {
 	mp := C.CString(modelPath)
 	defer C.free(unsafe.Pointer(mp))
-	cc := C.clip_model_load(mp, 1)
-	return &ClipContext{c: cc}
+	c := C.clip_model_load(mp, 1)
+	if c == nil {
+		return nil, fmt.Errorf("unable to load clip model: %v", modelPath)
+	}
+
+	projEmbedSize := int(C.clip_n_mmproj_embd(c))
+	modelEmbedSize := llamaContext.Model().NEmbd()
+	if projEmbedSize != modelEmbedSize {
+		return nil, fmt.Errorf("projector embedding size (%d) does not match model (%d)", projEmbedSize, modelEmbedSize)
+	}
+
+	return &ClipContext{c: c}, nil
 }

 func (c *ClipContext) Free() {
 	C.clip_free(c.c)
 }

-func NewLlavaImageEmbed(llamaContext *Context, clipContext *ClipContext, data []byte) [][]float32 {
-	c := C.llava_image_embed_make_with_bytes(clipContext.c, C.int(llamaContext.numThreads), (*C.uchar)(unsafe.Pointer(&data[0])), C.int(len(data)))
+func (c *ClipContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32, error) {
+	l := C.llava_image_embed_make_with_bytes(c.c, C.int(llamaContext.numThreads), (*C.uchar)(unsafe.Pointer(&data[0])), C.int(len(data)))
+	if l == nil {
+		return nil, errors.New("unable to make llava embedding from image")
+	}

-	numTokens := int(c.n_image_pos)
+	numTokens := int(l.n_image_pos)
 	numEmbed := llamaContext.Model().NEmbd()

-	s := unsafe.Slice((*float32)(c.embed), numEmbed*numTokens)
+	s := unsafe.Slice((*float32)(l.embed), numEmbed*numTokens)

 	embed := make([][]float32, numTokens)
 	rows := make([]float32, len(s))
@@ -442,9 +565,70 @@ func NewLlavaImageEmbed(llamaContext *Context, clipContext *ClipContext, data []
 		embed[i] = rows[i*numEmbed : (i+1)*numEmbed]
 	}

-	C.llava_image_embed_free(c)
+	C.llava_image_embed_free(l)

-	return embed
+	return embed, nil
+}
+
+type MllamaContext struct {
+	c *C.struct_mllama_ctx
+}
+
+func NewMllamaContext(llamaContext *Context, modelPath string) (*MllamaContext, error) {
+	mp := C.CString(modelPath)
+	defer C.free(unsafe.Pointer(mp))
+	c := C.mllama_model_load(mp, 1)
+	if c == nil {
+		return nil, fmt.Errorf("unable to load mllama model: %v", modelPath)
+	}
+
+	projEmbedSize := int(C.mllama_n_embd(c))
+	modelEmbedSize := llamaContext.Model().NEmbd()
+	if projEmbedSize != modelEmbedSize {
+		return nil, fmt.Errorf("projector embedding size (%d) does not match model (%d)", projEmbedSize, modelEmbedSize)
+	}
+
+	return &MllamaContext{c: c}, nil
+}
+
+func (m *MllamaContext) Free() {
+	C.mllama_free(m.c)
+}
+
+func (m *MllamaContext) NewEmbed(llamaContext *Context, data []byte, aspectRatioId int) ([][]float32, error) {
+	img := C.mllama_image_init()
+	defer C.mllama_image_free(img)
+
+	ok := bool(C.mllama_image_load_from_data(unsafe.Pointer(&data[0]), C.int(len(data)), 560, 560, 3, 4, C.int(aspectRatioId), img))
+	if !ok {
+		return nil, errors.New("unable to load mllama image data")
+	}
+
+	rows := make([]float32, m.EmbedSize(llamaContext))
+	ok = bool(C.mllama_image_encode(m.c, C.int(llamaContext.numThreads), img, (*C.float)(unsafe.Pointer(&rows[0]))))
+	if !ok {
+		return nil, errors.New("unable to make mllama embedding from image")
+	}
+
+	embed := make([][]float32, 1)
+	embed[0] = rows
+
+	return embed, nil
+}
+
+func (m *MllamaContext) EmbedSize(llamaContext *Context) int {
+	numTokens := int(C.mllama_n_positions(m.c) * C.mllama_n_tiles(m.c))
+	numEmbed := llamaContext.Model().NEmbd()
+
+	return numTokens * numEmbed
+}
+
+func (c *Context) SetCrossAttention(state bool) {
+	C.llama_set_cross_attention(c.c, C.bool(state))
+}
+
+func (c *Context) Synchronize() {
+	C.llama_synchronize(c.c)
 }

 // sampling
@@ -472,7 +656,7 @@ type SamplingParams struct {
 	Grammar        string
 }

-func NewSamplingContext(model *Model, params SamplingParams) *SamplingContext {
+func NewSamplingContext(model *Model, params SamplingParams) (*SamplingContext, error) {
 	var cparams C.struct_gpt_sampler_cparams
 	cparams.top_k = C.int32_t(params.TopK)
 	cparams.top_p = C.float(params.TopP)
@@ -495,9 +679,13 @@ func NewSamplingContext(model *Model, params SamplingParams) *SamplingContext {

 	cparams.grammar = grammar
 	context := &SamplingContext{c: C.gpt_sampler_cinit(model.c, &cparams)}
+	if context.c == nil {
+		return nil, errors.New("unable to create sampling context")
+	}
+
 	runtime.SetFinalizer(context, func(s *SamplingContext) { C.gpt_sampler_cfree(s.c) })

-	return context
+	return context, nil
 }

 func (s *SamplingContext) Reset() {
@@ -266,6 +266,7 @@ extern "C" {

        llama_token  *  token;
        float        *  embd;
+        int32_t         n_embd;
        llama_pos    *  pos;
        int32_t      *  n_seq_id;
        llama_seq_id ** seq_id;
@@ -449,6 +450,10 @@ extern "C" {
                     struct llama_model * model,
            struct llama_context_params   params);

+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state);
+
    // Frees all allocated memory
    LLAMA_API void llama_free(struct llama_context * ctx);

@@ -435,7 +435,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
        if (n_eval > n_batch) {
            n_eval = n_batch;
        }
-        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), n_embd, nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
        if (llama_decode(ctx_llama, batch)) {
            LOG_ERR("%s : failed to eval\n", __func__);
            return false;
@@ -17,7 +17,8 @@ else ifeq ($(OS),linux)
 	GPU_LIB_DIR_LINUX := $(HIP_PATH)/lib
 	GPU_COMPILER_LINUX := $(shell X=$$(which hipcc 2>/dev/null) && echo $$X)
 	GPU_COMPILER:=$(GPU_COMPILER_LINUX)
-	ROCM_TRANSITIVE_LIBS = $(shell ldd $(ROCM_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf | sort -u )
+	ROCM_TRANSITIVE_LIBS_INITIAL = $(sort $(shell ldd $(GPU_LIBS) | grep "=>" | cut -f2 -d= | cut -f2 -d' '  | grep -e rocm -e amdgpu -e libtinfo -e libnuma -e libelf))
+	GPU_TRANSITIVE_LIBS = $(sort $(shell readlink -f $(ROCM_TRANSITIVE_LIBS_INITIAL)) $(ROCM_TRANSITIVE_LIBS_INITIAL))
 endif

 # TODO future multi-variant support for ROCm
@@ -32,14 +33,18 @@ GPU_RUNNER_DRIVER_LIB_LINK := -lamdhip64
 GPU_RUNNER_LIBS_SHORT := hipblas rocblas
 GPU_PATH_ROOT_WIN=$(dir $(GPU_LIB_DIR_WIN))
 GPU_PATH_ROOT_LINUX=$(dir $(GPU_LIB_DIR_LINUX))
-GPU_COMPILER_CFLAGS_WIN = $(CFLAGS)
+GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -fPIC -D_GNU_SOURCE
-GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS)
+GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -fPIC -D_GNU_SOURCE

-ROCM_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
-ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
-ROCM_DIST_DEPS_LIBS = $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(ROCM_LIBS)) $(notdir $(ROCM_TRANSITIVE_LIBS)))
+GPU_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT),$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
+ifeq ($(OS),windows)
+	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH))/lib/ollama
+else ifeq ($(OS),linux)
+	ROCM_DIST_DEPS_DIR = $(abspath $(SRC_DIR)/../dist/$(OS)-$(ARCH)-rocm)/lib/ollama
+endif
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(ROCM_DIST_DEPS_DIR)/,$(notdir $(GPU_LIBS)) $(notdir $(GPU_TRANSITIVE_LIBS))))
 ROCBLAS_DIST_DEP_MANIFEST = $(ROCM_DIST_DEPS_DIR)/rocblas/library/TensileManifest.txt

 ifeq ($(OS),linux)
@@ -53,6 +58,8 @@ endif
 GPU_COMPILER_CUFLAGS = \
 	$(GPU_COMPILER_FPIC) \
 	$(addprefix -m,$(GPU_RUNNER_CPU_FLAGS)) \
+	-mf16c \
+	-mfma \
 	-parallel-jobs=2 \
 	-c \
 	-O3 \
@@ -72,6 +79,9 @@ GPU_COMPILER_CUFLAGS = \
 	-D_CRT_SECURE_NO_WARNINGS \
 	-D_GNU_SOURCE \
 	-D_XOPEN_SOURCE=600 \
+	-DUSE_PROF_API=1 \
+	-std=gnu++14 \
+	-x hip \
 	-mllvm=-amdgpu-early-inline-all=true \
 	-mllvm=-amdgpu-function-calls=false \
 	-Wno-expansion-to-defined \
@@ -80,18 +90,20 @@ GPU_COMPILER_CUFLAGS = \
 	-Wno-pass-failed \
 	-Wno-deprecated-declarations \
 	-Wno-unused-result \
-	-I. \
-	$(foreach arch, $(HIP_ARCHS_COMMON), --offload-arch=$(arch))
+	-I.
+
+# Workaround buggy P2P copy on some windows multi-GPU setups
+# This workaround breaks linux systems with small system RAM, so only enable on windows
+ifeq ($(OS),windows)
+	GPU_COMPILER_CUFLAGS += -DGGML_CUDA_NO_PEER_COPY=1
+endif

 include make/gpu.make

 # Adjust the rules from gpu.make to handle the ROCm dependencies properly
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST) $(ROCM_DIST_DEPS_LIBS)
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(ROCBLAS_DIST_DEP_MANIFEST)
 $(ROCBLAS_DIST_DEP_MANIFEST):
 	@-mkdir -p $(dir $@)
 	@echo "Copying rocblas library..."
 	cd $(GPU_LIB_DIR)/rocblas/library/ && tar cf - . | (cd $(dir $@) && tar xf - )
 	@echo "rocblas library copy complete"
-$(ROCM_DIST_DEPS_LIBS): 
-	@-mkdir -p $(dir $@)
-	$(CP) $(dir $(filter %$(notdir $@),$(ROCM_LIBS) $(ROCM_TRANSITIVE_LIBS)))/$(notdir $@)* $(dir $@)
@@ -3,7 +3,7 @@
 REPO_ROOT:=$(dir $(patsubst %/,%,$(dir $(patsubst %/,%,$(dir $(abspath $(lastword $(MAKEFILE_LIST))))))))
 DST_DIR:=$(dir $(patsubst %/,%,$(dir $(abspath $(lastword $(MAKEFILE_LIST))))))

-include $(REPO_ROOT)llama/vendoring.env
+include $(REPO_ROOT)llama/vendoring

 LLAMACPP_REPO := $(REPO_ROOT)llama/vendor/

@@ -76,3 +76,9 @@ else
 	CP := cp -af
 endif

+COMMON_SRCS := \
+	$(wildcard *.c) \
+	$(wildcard *.cpp)
+COMMON_HDRS := \
+	$(wildcard *.h) \
+	$(wildcard *.hpp)
@@ -19,6 +19,9 @@ GPU_COMPILER_CFLAGS_WIN = $(CFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CFLAGS_LINUX = $(CFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
 GPU_COMPILER_CXXFLAGS_WIN = $(CXXFLAGS) -D_WIN32_WINNT=0x602
 GPU_COMPILER_CXXFLAGS_LINUX = $(CXXFLAGS) -Xcompiler -fPIC -D_GNU_SOURCE
+GPU_LIBS = $(sort $(wildcard $(addsuffix *.$(SHARED_EXT)*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT)))))
+GPU_DIST_DEPS_LIBS= $(sort $(addprefix $(DIST_GPU_RUNNER_DEPS_DIR)/,$(notdir $(GPU_LIBS))))
+
 ifeq ($(OS),linux)
 	CUDA_PATH?=/usr/local/cuda
 	GPU_COMPILER_FPIC = -fPIC -Wno-unused-function -std=c++11
@@ -34,13 +34,6 @@ endif
 GPU_RUNNER_LIBS = $(wildcard $(addsuffix .$(SHARED_EXT).*,$(addprefix $(GPU_LIB_DIR)/$(SHARED_PREFIX),$(GPU_RUNNER_LIBS_SHORT))))
 DIST_GPU_RUNNER_LIB_DEPS = $(addprefix $(DIST_GPU_RUNNER_DEPS_DIR)/,$(notdir $(GPU_RUNNER_LIBS)))

-COMMON_SRCS := \
-	$(wildcard *.c) \
-	$(wildcard *.cpp)
-COMMON_HDRS := \
-	$(wildcard *.h) \
-	$(wildcard *.hpp)
-
 GPU_RUNNER_SRCS := \
 	ggml-cuda.cu \
 	$(filter-out $(wildcard ggml-cuda/fattn*.cu),$(wildcard ggml-cuda/*.cu)) \
@@ -79,7 +72,7 @@ $(GPU_RUNNER_NAME): $(BUILD_RUNNERS) $(DIST_RUNNERS) $(PAYLOAD_RUNNERS)
 # Build targets
 $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.cu
 	@-mkdir -p $(dir $@)
-	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CUFLAGS) $(GPU_RUNNER_ARCH_FLAGS) -o $@ $<
+	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) $(GPU_COMPILER_CUFLAGS) $(GPU_RUNNER_ARCH_FLAGS) -o $@ $<
 $(BUILD_DIR)/%.$(GPU_RUNNER_NAME).$(OBJ_EXT): %.c
 	@-mkdir -p $(dir $@)
 	$(CCACHE) $(GPU_COMPILER) -c $(GPU_COMPILER_CFLAGS) -o $@ $<
@@ -92,19 +85,22 @@ $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(RUNNERS
 	GOARCH=$(ARCH) CGO_LDFLAGS="$(TARGET_CGO_LDFLAGS)" go build -buildmode=pie  $(GPU_GOFLAGS) -trimpath -tags $(subst $(space),$(comma),$(GPU_RUNNER_CPU_FLAGS) $(GPU_RUNNER_GO_TAGS)) -o $@ ./runner
 $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(GPU_RUNNER_OBJS) $(DIST_GPU_RUNNER_LIB_DEPS) $(COMMON_HDRS) $(GPU_RUNNER_HDRS)
 	@-mkdir -p $(dir $@)
-	$(CCACHE) $(GPU_COMPILER) --shared $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@
+	$(CCACHE) $(GPU_COMPILER) --shared -L$(GPU_LIB_DIR) $(GPU_RUNNER_DRIVER_LIB_LINK) -L${DIST_GPU_RUNNER_DEPS_DIR} $(foreach lib, $(GPU_RUNNER_LIBS_SHORT), -l$(lib)) $(GPU_RUNNER_OBJS) -o $@

 # Distribution targets
 $(RUNNERS_DIST_DIR)/%: $(RUNNERS_BUILD_DIR)/%
 	@-mkdir -p $(dir $@)
-	cp $< $@
-$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
+	$(CP) $< $@
+$(RUNNERS_DIST_DIR)/$(GPU_RUNNER_NAME)/ollama_llama_server$(EXE_EXT): $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT) $(GPU_DIST_DEPS_LIBS)
 $(DIST_LIB_DIR)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT): $(RUNNERS_BUILD_DIR)/$(GPU_RUNNER_NAME)/$(SHARED_PREFIX)ggml_$(GPU_RUNNER_NAME).$(SHARED_EXT)
 	@-mkdir -p $(dir $@)
-	cp $< $@
+	$(CP) $< $@
 $(DIST_GPU_RUNNER_LIB_DEPS): 
 	@-mkdir -p $(dir $@)
-	$(CP) $(GPU_LIB_DIR)/$(notdir $@)* $(dir $@)
+	$(CP) $(GPU_LIB_DIR)/$(notdir $@) $(dir $@)
+$(GPU_DIST_DEPS_LIBS): 
+	@-mkdir -p $(dir $@)
+	$(CP) $(dir $(filter %$(notdir $@),$(GPU_LIBS) $(GPU_TRANSITIVE_LIBS)))/$(notdir $@) $(dir $@)

 # Payload targets
 $(RUNNERS_PAYLOAD_DIR)/%/ollama_llama_server.gz: $(RUNNERS_BUILD_DIR)/%/ollama_llama_server 
@@ -0,0 +1,900 @@
+// NOTE: This is modified from clip.cpp for Mllama only
+#include "mllama.h"
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#include <algorithm>
+#include <cmath>
+#include <cstdarg>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <stdexcept>
+#include <vector>
+
+#define REQUIRE(x)                                           \
+    do {                                                     \
+        if (!(x)) {                                          \
+            throw std::runtime_error("REQUIRE failed: " #x); \
+        }                                                    \
+    } while (0)
+
+#define LOG(fmt, ...) fprintf(stderr, "%s: " fmt "\n", __func__, ##__VA_ARGS__)
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#if __GLIBCXX__
+#include <cstdio>
+#include <ext/stdio_filebuf.h>
+#include <fcntl.h>
+#endif
+#endif
+
+struct mllama_image {
+    int width;
+    int height;
+
+    int num_channels = 3;
+    int num_tiles = 4;
+
+    int aspect_ratio_id;
+
+    std::vector<float> data;
+};
+
+static std::string format(const char *fmt, ...) {
+    va_list args;
+    va_start(args, fmt);
+    std::vector<char> b(128);
+    int n = vsnprintf(b.data(), b.size(), fmt, args);
+    REQUIRE(n >= 0 && n < b.size());
+    va_end(args);
+    return std::string(b.data(), b.size());
+}
+
+//
+// utilities to get data from a gguf file
+//
+
+static int get_key_index(const gguf_context *ctx, const char *key) {
+    int key_index = gguf_find_key(ctx, key);
+    REQUIRE(key_index != -1);
+    return key_index;
+}
+
+static std::vector<uint32_t> get_u32_array(const gguf_context *ctx, const std::string &key) {
+    const int i = get_key_index(ctx, key.c_str());
+    const int n = gguf_get_arr_n(ctx, i);
+    const uint32_t *data = (uint32_t *)gguf_get_arr_data(ctx, i);
+
+    std::vector<uint32_t> s(n);
+    for (size_t j = 0; j < s.size(); j++) {
+        s[j] = data[j];
+    }
+
+    return s;
+}
+
+static uint32_t get_u32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_u32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static float get_f32(const gguf_context *ctx, const std::string &key) {
+    return gguf_get_val_f32(ctx, get_key_index(ctx, key.c_str()));
+}
+
+static std::string get_ftype(int ftype) {
+    return ggml_type_name(static_cast<ggml_type>(ftype));
+}
+
+//
+// mllama layers
+//
+
+struct mllama_hparams {
+    uint32_t image_size;
+    uint32_t patch_size;
+    uint32_t hidden_size;
+    uint32_t n_intermediate;
+    uint32_t projection_dim;
+    uint32_t n_head;
+    uint32_t n_layer;
+    uint32_t n_global_layer;
+    uint32_t n_tiles;
+
+    float eps;
+
+    std::vector<bool> intermediate_layers;
+};
+
+struct mllama_layer {
+    // attention
+    struct ggml_tensor *k_w;
+    struct ggml_tensor *k_b;
+    struct ggml_tensor *q_w;
+    struct ggml_tensor *q_b;
+    struct ggml_tensor *v_w;
+    struct ggml_tensor *v_b;
+
+    struct ggml_tensor *o_w;
+    struct ggml_tensor *o_b;
+
+    struct ggml_tensor *attn_gate;
+
+    // layernorm 1
+    struct ggml_tensor *ln_1_w;
+    struct ggml_tensor *ln_1_b;
+
+    // ff
+    struct ggml_tensor *ff_i_w;
+    struct ggml_tensor *ff_i_b;
+
+    struct ggml_tensor *ff_o_w;
+    struct ggml_tensor *ff_o_b;
+
+    struct ggml_tensor *ff_gate;
+
+    // layernorm 2
+    struct ggml_tensor *ln_2_w;
+    struct ggml_tensor *ln_2_b;
+};
+
+struct mllama_vision_model {
+    struct mllama_hparams hparams;
+
+    // embeddings
+    struct ggml_tensor *class_embedding;
+    struct ggml_tensor *patch_embeddings;
+    struct ggml_tensor *position_embeddings;
+    struct ggml_tensor *position_embeddings_gate;
+    struct ggml_tensor *tile_position_embeddings;
+    struct ggml_tensor *tile_position_embeddings_gate;
+    struct ggml_tensor *pre_tile_position_embeddings;
+    struct ggml_tensor *pre_tile_position_embeddings_gate;
+    struct ggml_tensor *post_tile_position_embeddings;
+    struct ggml_tensor *post_tile_position_embeddings_gate;
+
+    struct ggml_tensor *pre_ln_w;
+    struct ggml_tensor *pre_ln_b;
+
+    std::vector<mllama_layer> layers;
+    std::vector<mllama_layer> global_layers;
+
+    struct ggml_tensor *post_ln_w;
+    struct ggml_tensor *post_ln_b;
+
+    struct ggml_tensor *mm_0_w;
+    struct ggml_tensor *mm_0_b;
+};
+
+struct mllama_ctx {
+    struct mllama_vision_model vision_model;
+
+    uint32_t ftype = 1;
+
+    struct gguf_context *ctx_gguf;
+    struct ggml_context *ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
+
+    // memory buffers to evaluate the model
+    ggml_backend_buffer_t params_buffer = nullptr;
+
+    ggml_backend_t backend = nullptr;
+    ggml_gallocr_t compute_alloc = nullptr;
+};
+
+static ggml_tensor *mllama_image_build_encoder_layer(
+    struct ggml_context *ctx0, const size_t il, const struct mllama_layer &layer, struct ggml_tensor *embeddings,
+    const float eps, const int hidden_size, const int batch_size, const int n_head, const int d_head) {
+    struct ggml_tensor *cur = embeddings;
+
+    {
+        // layernorm1
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_1_w), layer.ln_1_b);
+        ggml_set_name(cur, format("%d pre layernorm", il).c_str());
+    }
+
+    {
+        // self-attention
+        struct ggml_tensor *Q = ggml_mul_mat(ctx0, layer.q_w, cur);
+        if (layer.q_b != nullptr) {
+            Q = ggml_add(ctx0, Q, layer.q_b);
+        }
+
+        Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, Q->ne[1], batch_size);
+        Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+        ggml_set_name(Q, format("%d query", il).c_str());
+
+        struct ggml_tensor *K = ggml_mul_mat(ctx0, layer.k_w, cur);
+        if (layer.k_b != nullptr) {
+            K = ggml_add(ctx0, K, layer.k_b);
+        }
+
+        K = ggml_reshape_4d(ctx0, K, d_head, n_head, K->ne[1], batch_size);
+        K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+        ggml_set_name(K, format("%d key", il).c_str());
+
+        struct ggml_tensor *V = ggml_mul_mat(ctx0, layer.v_w, cur);
+        if (layer.v_b != nullptr) {
+            V = ggml_add(ctx0, V, layer.v_b);
+        }
+
+        V = ggml_reshape_4d(ctx0, V, d_head, n_head, V->ne[1], batch_size);
+        V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+        ggml_set_name(V, format("%d value", il).c_str());
+
+        struct ggml_tensor *KQ = ggml_mul_mat(ctx0, K, Q);
+        KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head));
+        KQ = ggml_soft_max_inplace(ctx0, KQ);
+        ggml_set_name(KQ, format("%d KQ", il).c_str());
+
+        struct ggml_tensor *KQV = ggml_mul_mat(ctx0, V, KQ);
+        KQV = ggml_reshape_4d(ctx0, KQV, d_head, KQV->ne[1], n_head, batch_size);
+        KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+        KQV = ggml_cont_3d(ctx0, KQV, hidden_size, KQV->ne[2], batch_size);
+        ggml_set_name(KQV, format("%d KQV", il).c_str());
+
+        cur = ggml_mul_mat(ctx0, layer.o_w, KQV);
+        if (layer.o_b != nullptr) {
+            cur = ggml_add(ctx0, cur, layer.o_b);
+        }
+        ggml_set_name(cur, format("%d self attention", il).c_str());
+
+        if (layer.attn_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.attn_gate);
+            ggml_set_name(cur, format("%d self attention gate", il).c_str());
+        }
+    }
+
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    {
+        // layernorm2
+        cur = ggml_norm(ctx0, cur, eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_2_w), layer.ln_2_b);
+        ggml_set_name(cur, format("%d post layernorm", il).c_str());
+    }
+
+    {
+        // feed forward
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_i_w, cur), layer.ff_i_b);
+        cur = ggml_gelu_inplace(ctx0, cur);
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, layer.ff_o_w, cur), layer.ff_o_b);
+        ggml_set_name(cur, format("%d feed forward", il).c_str());
+
+        if (layer.ff_gate != nullptr) {
+            cur = ggml_mul_inplace(ctx0, cur, layer.ff_gate);
+            ggml_set_name(cur, format("%d feed forward gate", il).c_str());
+        }
+    }
+
+    // residual 2
+    cur = ggml_add(ctx0, cur, embeddings);
+    ggml_set_name(cur, format("%d residual", il).c_str());
+
+    embeddings = cur;
+
+    return embeddings;
+}
+
+static ggml_cgraph *mllama_image_build_graph(mllama_ctx *ctx, const mllama_image_batch *imgs) {
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    const int image_size_width = image_size;
+    const int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+    const int hidden_size = hparams.hidden_size;
+    const int n_head = hparams.n_head;
+    const int d_head = hidden_size / n_head;
+
+    const int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    int num_tiles = 4;
+    int num_channels = 3;
+    if (imgs->data != nullptr) {
+        num_tiles = imgs->data[0].num_tiles > 0 ? imgs->data[0].num_tiles : num_tiles;
+        num_channels = imgs->data[0].num_channels > 0 ? imgs->data[0].num_channels : num_channels;
+    }
+
+    struct ggml_init_params params = {
+        ctx->buf_compute_meta.size(), // mem_size
+        ctx->buf_compute_meta.data(), // mem_buffer
+        true,                         // no_alloc
+    };
+
+    struct ggml_context *ctx0 = ggml_init(params);
+    struct ggml_cgraph *gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor *inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, num_channels, num_tiles);
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
+
+    struct ggml_tensor *inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
+
+    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, num_tiles);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+
+    struct ggml_tensor *aspect_ratios = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, imgs->size);
+    ggml_set_name(aspect_ratios, "aspect_ratios");
+    ggml_set_input(aspect_ratios);
+
+    if (model.pre_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *pre_tile_position_embeddings = ggml_get_rows(ctx0, model.pre_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(pre_tile_position_embeddings, "pre_tile_position_embeddings");
+
+        pre_tile_position_embeddings = ggml_reshape_3d(ctx0, pre_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.pre_tile_position_embeddings_gate != nullptr) {
+            pre_tile_position_embeddings = ggml_mul_inplace(ctx0, pre_tile_position_embeddings, model.pre_tile_position_embeddings_gate);
+        }
+
+        inp = ggml_add(ctx0, inp, pre_tile_position_embeddings);
+    }
+
+    struct ggml_tensor *embeddings = inp;
+
+    if (model.class_embedding != nullptr) {
+        // concat class_embeddings and patch_embeddings
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, num_tiles);
+        ggml_set_name(embeddings, "embeddings");
+        ggml_set_input(embeddings);
+        for (int i = 0; i < num_tiles; ++i) {
+            // repeat class embeddings for each tile
+            embeddings = ggml_acc_inplace(ctx0, embeddings, model.class_embedding, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], i * embeddings->nb[2]);
+        }
+
+        embeddings = ggml_acc_inplace(ctx0, embeddings, inp, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
+
+    struct ggml_tensor *positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    struct ggml_tensor *position_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
+    if (model.position_embeddings_gate != nullptr) {
+        position_embd = ggml_mul_inplace(ctx0, position_embd, model.position_embeddings_gate);
+    }
+
+    embeddings = ggml_add(ctx0, embeddings, position_embd);
+
+    if (model.tile_position_embeddings != nullptr) {
+        struct ggml_tensor *tile_position_embeddings = ggml_get_rows(ctx0, model.tile_position_embeddings, aspect_ratios);
+        ggml_set_name(tile_position_embeddings, "tile_position_embeddings");
+
+        tile_position_embeddings = ggml_reshape_3d(ctx0, tile_position_embeddings, hidden_size, num_positions, num_tiles);
+        if (model.tile_position_embeddings_gate != nullptr) {
+            tile_position_embeddings = ggml_mul_inplace(ctx0, tile_position_embeddings, model.tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, tile_position_embeddings);
+    }
+
+    // pre-layernorm
+    if (model.pre_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.pre_ln_w);
+        if (model.pre_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.pre_ln_b);
+        }
+
+        ggml_set_name(embeddings, "pre layernorm");
+    }
+
+    const int num_padding_patches = 8 - (embeddings->ne[1] % 8) % 8;
+
+    embeddings = ggml_pad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_view_3d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1] * embeddings->ne[2], batch_size, embeddings->nb[1], embeddings->nb[2] * embeddings->ne[3], 0);
+
+    std::vector<struct ggml_tensor *> intermediate_embeddings;
+
+    // encoder
+    for (size_t il = 0; il < model.layers.size(); il++) {
+        if (hparams.intermediate_layers[il]) {
+            intermediate_embeddings.push_back(embeddings);
+        }
+
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    // post-layernorm
+    if (model.post_ln_w != nullptr) {
+        embeddings = ggml_mul(ctx0, ggml_norm(ctx0, embeddings, hparams.eps), model.post_ln_w);
+        if (model.post_ln_b != nullptr) {
+            embeddings = ggml_add(ctx0, embeddings, model.post_ln_b);
+        }
+
+        ggml_set_name(embeddings, "post layernorm");
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+
+    if (model.post_tile_position_embeddings != nullptr) {
+        struct ggml_tensor *post_tile_position_embeddings = ggml_get_rows(ctx0, model.post_tile_position_embeddings, aspect_ratios);
+        ggml_set_name(post_tile_position_embeddings, "post_tile_position_embeddings");
+
+        post_tile_position_embeddings = ggml_reshape_3d(ctx0, post_tile_position_embeddings, hidden_size, 1, num_tiles);
+        if (model.post_tile_position_embeddings_gate != nullptr) {
+            post_tile_position_embeddings = ggml_mul(ctx0, post_tile_position_embeddings, model.post_tile_position_embeddings_gate);
+        }
+
+        embeddings = ggml_add(ctx0, embeddings, post_tile_position_embeddings);
+    }
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_tiles * (num_positions + num_padding_patches), 1);
+
+    // global encoder
+    for (size_t il = 0; il < model.global_layers.size(); il++) {
+        embeddings = mllama_image_build_encoder_layer(
+            ctx0, il, model.global_layers[il], embeddings,
+            hparams.eps, hidden_size, batch_size, n_head, d_head);
+    }
+
+    struct ggml_tensor *stacked_embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 0, hidden_size, (num_positions + num_padding_patches) * num_tiles);
+    for (size_t i = 0; i < intermediate_embeddings.size(); ++i) {
+        stacked_embeddings = ggml_concat(ctx0, stacked_embeddings, ggml_reshape_3d(ctx0, intermediate_embeddings[i], 1, intermediate_embeddings[i]->ne[0], intermediate_embeddings[i]->ne[1]), 0);
+    }
+
+    stacked_embeddings = ggml_reshape_4d(ctx0, stacked_embeddings, intermediate_embeddings.size() * hidden_size, num_positions + num_padding_patches, num_tiles, batch_size);
+    stacked_embeddings = ggml_unpad(ctx0, stacked_embeddings, 0, num_padding_patches, 0, 0);
+
+    embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size, num_positions + num_padding_patches, num_tiles);
+    embeddings = ggml_unpad(ctx0, embeddings, 0, num_padding_patches, 0, 0);
+    embeddings = ggml_concat(ctx0, embeddings, stacked_embeddings, 0);
+
+    // mllama projector
+    embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_0_w, embeddings), model.mm_0_b);
+    ggml_set_name(embeddings, "multi modal projector");
+
+    // build the graph
+    ggml_build_forward_expand(gf, embeddings);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_tensor *mllama_tensor_load(struct ggml_context *ctx, const char *name, const bool optional) {
+    struct ggml_tensor *cur = ggml_get_tensor(ctx, name);
+    REQUIRE(cur != nullptr || optional);
+    return cur;
+}
+
+static std::vector<struct mllama_layer> mllama_layers_load(struct ggml_context *ctx, const char *prefix, const int n) {
+    std::vector<struct mllama_layer> layers(n);
+    for (size_t i = 0; i < layers.size(); i++) {
+        auto &layer = layers[i];
+        layer.ln_1_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.weight", prefix, i).c_str(), false);
+        layer.ln_1_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln1.bias", prefix, i).c_str(), false);
+        layer.ln_2_w = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.weight", prefix, i).c_str(), false);
+        layer.ln_2_b = mllama_tensor_load(ctx, format("%s.blk.%d.ln2.bias", prefix, i).c_str(), false);
+
+        layer.k_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.weight", prefix, i).c_str(), false);
+        layer.k_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_k.bias", prefix, i).c_str(), true);
+        layer.q_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.weight", prefix, i).c_str(), false);
+        layer.q_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_q.bias", prefix, i).c_str(), true);
+        layer.v_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.weight", prefix, i).c_str(), false);
+        layer.v_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_v.bias", prefix, i).c_str(), true);
+        layer.o_w = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.weight", prefix, i).c_str(), false);
+        layer.o_b = mllama_tensor_load(ctx, format("%s.blk.%d.attn_out.bias", prefix, i).c_str(), true);
+
+        layer.ff_i_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.weight", prefix, i).c_str(), false);
+        layer.ff_i_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_down.bias", prefix, i).c_str(), false);
+        layer.ff_o_w = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.weight", prefix, i).c_str(), false);
+        layer.ff_o_b = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_up.bias", prefix, i).c_str(), false);
+
+        layer.attn_gate = mllama_tensor_load(ctx, format("%s.blk.%d.attn_gate", prefix, i).c_str(), true);
+        layer.ff_gate = mllama_tensor_load(ctx, format("%s.blk.%d.ffn_gate", prefix, i).c_str(), true);
+    }
+
+    return layers;
+}
+
+// read and create ggml_context containing the tensors and their data
+struct mllama_ctx *mllama_model_load(const char *fname, const int verbosity = 1) {
+    struct ggml_context *meta = nullptr;
+
+    struct gguf_init_params params = {
+        true,  // no_alloc
+        &meta, // ctx
+    };
+
+    struct gguf_context *ctx = gguf_init_from_file(fname, params);
+    REQUIRE(ctx != nullptr);
+
+    if (verbosity >= 1) {
+        const int n_tensors = gguf_get_n_tensors(ctx);
+        const int n_kv = gguf_get_n_kv(ctx);
+        const std::string ftype = get_ftype(get_u32(ctx, "general.file_type"));
+        const int idx_desc = get_key_index(ctx, "general.description");
+        const std::string description = gguf_get_val_str(ctx, idx_desc);
+        const int idx_name = gguf_find_key(ctx, "general.name");
+        if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
+            const std::string name = gguf_get_val_str(ctx, idx_name);
+            LOG("model name:   %s", name.c_str());
+        }
+        LOG("description:  %s", description.c_str());
+        LOG("GGUF version: %d", gguf_get_version(ctx));
+        LOG("alignment:    %zu", gguf_get_alignment(ctx));
+        LOG("n_tensors:    %d", n_tensors);
+        LOG("n_kv:         %d", n_kv);
+        LOG("ftype:        %s", ftype.c_str());
+        LOG("");
+    }
+    const int n_tensors = gguf_get_n_tensors(ctx);
+
+    mllama_ctx *new_mllama = new mllama_ctx{};
+
+#ifdef GGML_USE_CUDA
+    new_mllama->backend = ggml_backend_cuda_init(0);
+    LOG("vision using CUDA backend");
+#endif
+
+#ifdef GGML_USE_METAL
+    new_mllama->backend = ggml_backend_metal_init();
+    LOG("vision using Metal backend");
+#endif
+
+#ifdef GGML_USE_CANN
+    new_mllama->backend = ggml_backend_cann_init(0);
+    LOG("vision using CANN backend");
+#endif
+
+#ifdef GGML_USE_VULKAN
+    new_mllama->backend = ggml_backend_vk_init(0);
+    LOG("vision using Vulkan backend");
+#endif
+
+    if (!new_mllama->backend) {
+        new_mllama->backend = ggml_backend_cpu_init();
+        LOG("vision using CPU backend");
+    }
+
+    // load tensors
+    {
+        std::vector<uint8_t> read_buf;
+        struct ggml_init_params params = {
+            (n_tensors + 1) * ggml_tensor_overhead(), // mem_size
+            nullptr,                                  // mem_buffer
+            true,                                     // no_alloc
+        };
+
+        new_mllama->ctx_data = ggml_init(params);
+        if (!new_mllama->ctx_data) {
+            LOG("ggml_init() failed");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+#ifdef _WIN32
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+        if (!wlen) {
+            return NULL;
+        }
+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
+        if (!wlen) {
+            free(wbuf);
+            return NULL;
+        }
+#if __GLIBCXX__
+        int fd = _wopen(wbuf, _O_RDONLY | _O_BINARY);
+        __gnu_cxx::stdio_filebuf<char> buffer(fd, std::ios_base::in);
+        std::istream fin(&buffer);
+#else // MSVC
+        // unused in our current build
+        auto fin = std::ifstream(wbuf, std::ios::binary);
+#endif
+        free(wbuf);
+#else
+        auto fin = std::ifstream(fname, std::ios::binary);
+#endif
+        if (!fin) {
+            LOG("cannot open model file for loading tensors\n");
+            mllama_free(new_mllama);
+            gguf_free(ctx);
+            return nullptr;
+        }
+
+        // add tensors to context
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *t = ggml_get_tensor(meta, name);
+            struct ggml_tensor *cur = ggml_dup_tensor(new_mllama->ctx_data, t);
+            ggml_set_name(cur, name);
+        }
+
+        // alloc memory and offload data
+        new_mllama->params_buffer = ggml_backend_alloc_ctx_tensors(new_mllama->ctx_data, new_mllama->backend);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char *name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor *cur = ggml_get_tensor(new_mllama->ctx_data, name);
+            const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
+            fin.seekg(offset, std::ios::beg);
+            if (!fin) {
+                LOG("failed to seek for tensor %s\n", name);
+                mllama_free(new_mllama);
+                gguf_free(ctx);
+                return nullptr;
+            }
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_mllama->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
+        }
+
+#if defined(_WIN32) && defined(__GLIBCXX__)
+        close(fd);
+#else
+        fin.close();
+#endif
+    }
+
+    // vision model
+    // load vision model
+    auto &vision_model = new_mllama->vision_model;
+    auto &hparams = vision_model.hparams;
+    hparams.hidden_size = get_u32(ctx, "mllama.vision.embedding_length");
+    hparams.n_head = get_u32(ctx, "mllama.vision.attention.head_count");
+    hparams.n_intermediate = get_u32(ctx, "mllama.vision.feed_forward_length");
+    hparams.n_layer = get_u32(ctx, "mllama.vision.block_count");
+    hparams.n_global_layer = get_u32(ctx, "mllama.vision.global.block_count");
+    hparams.n_tiles = get_u32(ctx, "mllama.vision.max_num_tiles");
+    hparams.image_size = get_u32(ctx, "mllama.vision.image_size");
+    hparams.patch_size = get_u32(ctx, "mllama.vision.patch_size");
+    hparams.projection_dim = get_u32(ctx, "mllama.vision.projection_dim");
+    hparams.eps = get_f32(ctx, "mllama.vision.attention.layer_norm_epsilon");
+
+    std::vector<uint32_t> intermediate_layers_indices = get_u32_array(ctx, "mllama.vision.intermediate_layers_indices");
+    hparams.intermediate_layers.resize(hparams.n_layer);
+    for (size_t i = 0; i < intermediate_layers_indices.size(); i++) {
+        hparams.intermediate_layers[intermediate_layers_indices[i]] = true;
+    }
+
+    if (verbosity >= 2) {
+        LOG("");
+        LOG("vision model hparams");
+        LOG("image_size         %d", hparams.image_size);
+        LOG("patch_size         %d", hparams.patch_size);
+        LOG("v_hidden_size      %d", hparams.hidden_size);
+        LOG("v_n_intermediate   %d", hparams.n_intermediate);
+        LOG("v_projection_dim   %d", hparams.projection_dim);
+        LOG("v_n_head           %d", hparams.n_head);
+        LOG("v_n_layer          %d", hparams.n_layer);
+        LOG("v_n_global_layer   %d", hparams.n_global_layer);
+        LOG("v_eps              %f", hparams.eps);
+    }
+
+    vision_model.class_embedding = mllama_tensor_load(new_mllama->ctx_data, "v.class_embd", true);
+    vision_model.patch_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.patch_embd.weight", true);
+
+    vision_model.position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.weight", true);
+    vision_model.position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.position_embd.gate", true);
+
+    vision_model.pre_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.weight", true);
+    vision_model.pre_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.pre_ln.bias", true);
+    vision_model.post_ln_w = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.weight", true);
+    vision_model.post_ln_b = mllama_tensor_load(new_mllama->ctx_data, "v.post_ln.bias", true);
+
+    vision_model.tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.weight", true);
+    vision_model.tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.tile_position_embd.gate", true);
+
+    vision_model.pre_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.weight", true);
+    vision_model.pre_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.pre_tile_position_embd.gate", true);
+
+    vision_model.post_tile_position_embeddings = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.weight", true);
+    vision_model.post_tile_position_embeddings_gate = mllama_tensor_load(new_mllama->ctx_data, "v.post_tile_position_embd.gate", true);
+
+    vision_model.mm_0_w = mllama_tensor_load(new_mllama->ctx_data, "mm.0.weight", false);
+    vision_model.mm_0_b = mllama_tensor_load(new_mllama->ctx_data, "mm.0.bias", false);
+
+    vision_model.layers = mllama_layers_load(new_mllama->ctx_data, "v", hparams.n_layer);
+    vision_model.global_layers = mllama_layers_load(new_mllama->ctx_data, "v.global", hparams.n_global_layer);
+
+    ggml_free(meta);
+
+    new_mllama->ctx_gguf = ctx;
+
+    {
+        // measure mem requirement and allocate
+        new_mllama->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_mllama->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_mllama->backend));
+        struct mllama_image_batch batch;
+        batch.size = 1;
+        ggml_cgraph *gf = mllama_image_build_graph(new_mllama, &batch);
+        ggml_gallocr_reserve(new_mllama->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_mllama->compute_alloc, 0);
+        LOG("compute allocated memory: %.2f MB", compute_memory_buffer_size / 1024.0 / 1024.0);
+    }
+
+    return new_mllama;
+}
+
+struct mllama_image *mllama_image_init() {
+    return new mllama_image();
+}
+
+void mllama_image_free(struct mllama_image *img) { delete img; }
+void mllama_image_batch_free(struct mllama_image_batch *batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
+    }
+}
+
+bool mllama_image_load_from_data(const void *data, const int n, const int width, const int height, const int num_channels, const int num_tiles, const int aspect_ratio_id, struct mllama_image *img) {
+    img->width = width;
+    img->height = height;
+    img->num_channels = num_channels;
+    img->num_tiles = num_tiles;
+    img->aspect_ratio_id = aspect_ratio_id;
+    img->data.resize(n);
+
+    memcpy(img->data.data(), data, n);
+    return true;
+}
+
+inline int mllama(int x, int lower, int upper) {
+    return std::max(lower, std::min(x, upper));
+}
+
+void mllama_free(mllama_ctx *ctx) {
+    ggml_free(ctx->ctx_data);
+    gguf_free(ctx->ctx_gguf);
+
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
+    delete ctx;
+}
+
+bool mllama_image_encode(struct mllama_ctx *ctx, const int n_threads, mllama_image *img, float *vec) {
+    mllama_image_batch imgs{};
+    imgs.size = 1;
+    imgs.data = img;
+    return mllama_image_batch_encode(ctx, n_threads, &imgs, vec);
+}
+
+bool mllama_image_batch_encode(mllama_ctx *ctx, const int n_threads, const mllama_image_batch *imgs, float *vec) {
+    int batch_size = imgs->size;
+    REQUIRE(batch_size == 1);
+
+    // build the inference graph
+    ggml_cgraph *gf = mllama_image_build_graph(ctx, imgs);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto &model = ctx->vision_model;
+    const auto &hparams = model.hparams;
+
+    const int image_size = hparams.image_size;
+    int image_size_width = image_size;
+    int image_size_height = image_size;
+
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
+    const int num_positions = num_patches + (model.class_embedding == nullptr ? 0 : 1);
+
+    {
+        struct ggml_tensor *inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        ggml_backend_tensor_set(inp_raw, imgs->data[0].data.data(), 0, ggml_nbytes(inp_raw));
+    }
+
+    {
+        struct ggml_tensor *embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        if (embeddings != nullptr) {
+            void *zeros = malloc(ggml_nbytes(embeddings));
+            memset(zeros, 0, ggml_nbytes(embeddings));
+            ggml_backend_tensor_set(embeddings, zeros, 0, ggml_nbytes(embeddings));
+            free(zeros);
+        }
+    }
+
+    {
+        struct ggml_tensor *positions = ggml_graph_get_tensor(gf, "positions");
+        if (positions != nullptr) {
+            int *positions_data = (int *)malloc(ggml_nbytes(positions));
+            for (int i = 0; i < num_positions; i++) {
+                positions_data[i] = i;
+            }
+            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+            free(positions_data);
+        }
+    }
+
+    {
+        struct ggml_tensor *aspect_ratios = ggml_graph_get_tensor(gf, "aspect_ratios");
+        if (aspect_ratios != nullptr) {
+            int *aspect_ratios_data = (int *)malloc(ggml_nbytes(aspect_ratios));
+            aspect_ratios_data[0] = imgs->data[0].aspect_ratio_id;
+            ggml_backend_tensor_set(aspect_ratios, aspect_ratios_data, 0, ggml_nbytes(aspect_ratios));
+            free(aspect_ratios_data);
+        }
+    }
+
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
+    }
+
+    ggml_backend_graph_compute(ctx->backend, gf);
+
+    // the last node is the embedding tensor
+    struct ggml_tensor *embeddings = ggml_graph_node(gf, ggml_graph_n_nodes(gf) - 1);
+
+    // copy the embeddings to the location passed by the user
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
+
+    return true;
+}
+
+int32_t mllama_image_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.image_size;
+}
+
+int32_t mllama_patch_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.patch_size;
+}
+
+int32_t mllama_hidden_size(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.hidden_size;
+}
+
+int mllama_n_patches(const struct mllama_ctx *ctx) {
+    const auto &hparams = ctx->vision_model.hparams;
+    return (hparams.image_size / hparams.patch_size) * (hparams.image_size / hparams.patch_size);
+}
+
+int mllama_n_positions(const struct mllama_ctx *ctx) {
+    return mllama_n_patches(ctx) + (ctx->vision_model.class_embedding == nullptr ? 0 : 1);
+}
+
+int mllama_n_tiles(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.n_tiles;
+}
+
+int mllama_n_embd(const struct mllama_ctx *ctx) {
+    return ctx->vision_model.hparams.projection_dim;
+}
+
+size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx) {
+    return mllama_n_positions(ctx) * mllama_n_embd(ctx) * mllama_n_tiles(ctx) * sizeof(float);
+}
@@ -0,0 +1,61 @@
+#ifndef MLLAMA_H
+#define MLLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#if defined(_WIN32) && !defined(__MINGW32__)
+#ifdef LLAMA_BUILD
+#define MLLAMA_API __declspec(dllexport)
+#else
+#define MLLAMA_API __declspec(dllimport)
+#endif
+#else
+#define MLLAMA_API __attribute__((visibility("default")))
+#endif
+#else
+#define MLLAMA_API
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct mllama_ctx;
+
+struct mllama_image_batch {
+    struct mllama_image *data;
+    size_t size;
+};
+
+MLLAMA_API struct mllama_ctx *mllama_model_load(const char *fname, int verbosity);
+MLLAMA_API struct mllama_ctx *mllama_model_load_cpu(const char *fname, int verbosity);
+
+MLLAMA_API void mllama_free(struct mllama_ctx *ctx);
+
+MLLAMA_API int32_t mllama_image_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_patch_size(const struct mllama_ctx *ctx);
+MLLAMA_API int32_t mllama_hidden_size(const struct mllama_ctx *ctx);
+
+MLLAMA_API int mllama_n_patches(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_positions(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_tiles(const struct mllama_ctx *ctx);
+MLLAMA_API int mllama_n_embd(const struct mllama_ctx *ctx);
+MLLAMA_API size_t mllama_n_embd_bytes(const struct mllama_ctx *ctx);
+
+MLLAMA_API struct mllama_image *mllama_image_init();
+
+MLLAMA_API void mllama_image_free(struct mllama_image *img);
+MLLAMA_API void mllama_image_batch_free(struct mllama_image_batch *batch);
+
+MLLAMA_API bool mllama_image_load_from_data(const void *data, const int n, const int nx, const int ny, const int nc, const int nt, const int aspect_ratio_id, struct mllama_image *img);
+
+MLLAMA_API bool mllama_image_encode(struct mllama_ctx *ctx, int n_threads, struct mllama_image *img, float *vec);
+MLLAMA_API bool mllama_image_batch_encode(struct mllama_ctx *ctx, int n_threads, const struct mllama_image_batch *imgs, float *vec);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // MLLAMA_H
@@ -0,0 +1,732 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: jmorganca <jmorganca@gmail.com>
+Date: Thu, 17 Oct 2024 15:18:22 -0700
+Subject: [PATCH] add mllama support
+
+mllama adds cross-attention layers to the standard llama architecture
+it also requires a way to input a new tensor: cross_attention_state
+once per generation
+
+cross-attention layers don't change and so they are cached in the
+kv cache once per run
+
+remaining is to implement the cross attention mask
+---
+ examples/llava/llava.cpp |   2 +-
+ include/llama.h          |   5 +
+ src/llama.cpp            | 447 +++++++++++++++++++++++++++++++++++++--
+ 3 files changed, 436 insertions(+), 18 deletions(-)
+
+diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp
+index 8558c6bd..37b2f2e2 100644
+--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
+@@ -409,7 +409,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
+         if (n_eval > n_batch) {
+             n_eval = n_batch;
+         }
+-        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), n_embd, nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+         if (llama_decode(ctx_llama, batch)) {
+             LOG_ERR("%s : failed to eval\n", __func__);
+             return false;
+diff --git a/include/llama.h b/include/llama.h
+index 7cae1bbe..aca09310 100644
+--- a/include/llama.h
+++ b/include/llama.h
+@@ -240,6 +240,7 @@ extern "C" {
+ 
+         llama_token  *  token;
+         float        *  embd;
+        int32_t         n_embd;
+         llama_pos    *  pos;
+         int32_t      *  n_seq_id;
+         llama_seq_id ** seq_id;
+@@ -423,6 +424,10 @@ extern "C" {
+                      struct llama_model * model,
+             struct llama_context_params   params);
+ 
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state);
+
+     // Frees all allocated memory
+     LLAMA_API void llama_free(struct llama_context * ctx);
+ 
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 83b80b59..35748488 100644
+--- a/src/llama.cpp
+++ b/src/llama.cpp
+@@ -169,6 +169,7 @@ static std::string format(const char * fmt, ...) {
+ 
+ enum llm_arch {
+     LLM_ARCH_LLAMA,
+    LLM_ARCH_MLLAMA,
+     LLM_ARCH_FALCON,
+     LLM_ARCH_BAICHUAN,
+     LLM_ARCH_GROK,
+@@ -223,6 +224,7 @@ enum llm_arch {
+ 
+ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_LLAMA,           "llama"        },
+    { LLM_ARCH_MLLAMA,          "mllama"       },
+     { LLM_ARCH_FALCON,          "falcon"       },
+     { LLM_ARCH_GROK,            "grok"         },
+     { LLM_ARCH_GPT2,            "gpt2"         },
+@@ -330,6 +332,7 @@ enum llm_kv {
+     LLM_KV_ATTENTION_SLIDING_WINDOW,
+     LLM_KV_ATTENTION_SCALE,
+     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
+    LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+ 
+     LLM_KV_ROPE_DIMENSION_COUNT,
+     LLM_KV_ROPE_FREQ_BASE,
+@@ -439,6 +442,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"           },
+     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                    },
+     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,  "%s.attention.block_skip_connection.%d" },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers"   },
+ 
+     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
+     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+@@ -613,6 +617,14 @@ enum llm_tensor {
+     LLM_TENSOR_CLS,
+     LLM_TENSOR_CLS_OUT,
+     LLM_TENSOR_BSKCN_TV,
+    LLM_TENSOR_CROSS_ATTN_K_NORM,
+    LLM_TENSOR_CROSS_ATTN_K_PROJ,
+    LLM_TENSOR_CROSS_ATTN_O_PROJ,
+    LLM_TENSOR_CROSS_ATTN_Q_NORM,
+    LLM_TENSOR_CROSS_ATTN_Q_PROJ,
+    LLM_TENSOR_CROSS_ATTN_V_PROJ,
+    LLM_TENSOR_CROSS_ATTN_ATTN_GATE,
+    LLM_TENSOR_CROSS_ATTN_MLP_GATE,
+ };
+ 
+ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+@@ -642,6 +654,40 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
+             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+         },
+     },
+    {
+        LLM_ARCH_MLLAMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_CROSS_ATTN_K_NORM,    "blk.%d.cross_attn_k_norm" },
+            { LLM_TENSOR_CROSS_ATTN_K_PROJ,    "blk.%d.cross_attn_k_proj" },
+            { LLM_TENSOR_CROSS_ATTN_O_PROJ,    "blk.%d.cross_attn_o_proj" },
+            { LLM_TENSOR_CROSS_ATTN_Q_NORM,    "blk.%d.cross_attn_q_norm" },
+            { LLM_TENSOR_CROSS_ATTN_Q_PROJ,    "blk.%d.cross_attn_q_proj" },
+            { LLM_TENSOR_CROSS_ATTN_V_PROJ,    "blk.%d.cross_attn_v_proj" },
+            { LLM_TENSOR_CROSS_ATTN_ATTN_GATE, "blk.%d.cross_attn_attn_gate" },
+            { LLM_TENSOR_CROSS_ATTN_MLP_GATE,  "blk.%d.cross_attn_mlp_gate" },
+        },
+    },
+     {
+         LLM_ARCH_BAICHUAN,
+         {
+@@ -2390,6 +2436,7 @@ enum e_model {
+     MODEL_40B,
+     MODEL_65B,
+     MODEL_70B,
+    MODEL_90B,
+     MODEL_236B,
+     MODEL_314B,
+     MODEL_SMALL,
+@@ -2434,6 +2481,7 @@ struct llama_hparams {
+     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+ 
+     std::array<std::array<uint32_t, LLAMA_MAX_LAYERS>, 4> n_bskcn_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> cross_attn_layers;
+ 
+     uint32_t n_layer_dense_lead = 0;
+     uint32_t n_lora_q = 0;
+@@ -2502,10 +2550,11 @@ struct llama_hparams {
+         if (this->n_expert      != other.n_expert)      return true;
+         if (this->n_expert_used != other.n_expert_used) return true;
+ 
+-        if (this->n_head_arr    != other.n_head_arr)    return true;
+-        if (this->n_head_kv_arr != other.n_head_kv_arr) return true;
+-        if (this->n_ff_arr      != other.n_ff_arr)      return true;
+-        if (this->n_bskcn_arr   != other.n_bskcn_arr)   return true;
+        if (this->n_head_arr        != other.n_head_arr)        return true;
+        if (this->n_head_kv_arr     != other.n_head_kv_arr)     return true;
+        if (this->n_ff_arr          != other.n_ff_arr)          return true;
+        if (this->n_bskcn_arr       != other.n_bskcn_arr)       return true;
+        if (this->cross_attn_layers != other.cross_attn_layers) return true;
+ 
+         if (this->n_rel_attn_bkts    != other.n_rel_attn_bkts)    return true;
+         if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+@@ -2623,6 +2672,10 @@ struct llama_hparams {
+ 
+         GGML_ABORT("fatal error");
+     }
+
+    bool cross_attention_layers(uint32_t il) const {
+        return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+    }
+ };
+ 
+ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+@@ -2652,6 +2705,9 @@ struct llama_cparams {
+     bool offload_kqv;
+     bool flash_attn;
+     bool no_perf;
+    // TODO (jmorganca): this should most likely be passed in as part of a batch
+    // and not set on the context for all batches.
+    bool cross_attn = false;
+ 
+     enum llama_pooling_type pooling_type;
+ 
+@@ -2806,6 +2862,16 @@ struct llama_layer {
+     struct ggml_tensor * ffn_down_scale;
+ 
+     struct ggml_tensor * bskcn_tv;
+
+    // cross attention
+    struct ggml_tensor * cross_attn_k_norm;
+    struct ggml_tensor * cross_attn_k_proj;
+    struct ggml_tensor * cross_attn_o_proj;
+    struct ggml_tensor * cross_attn_q_norm;
+    struct ggml_tensor * cross_attn_q_proj;
+    struct ggml_tensor * cross_attn_v_proj;
+    struct ggml_tensor * cross_attn_attn_gate;
+    struct ggml_tensor * cross_attn_mlp_gate;
+ };
+ 
+ // very similar to llama_batch,
+@@ -3452,6 +3518,8 @@ struct llama_context {
+     struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
+     struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
+     struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+
+    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
+ };
+ 
+ struct llama_lora_weight {
+@@ -3686,6 +3754,18 @@ static bool llama_kv_cache_init(
+     cache.v_l.reserve(n_layer);
+ 
+     for (int i = 0; i < (int) n_layer; i++) {
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
+            struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
+            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+            ggml_format_name(k, "cache_k_l%d", i);
+            ggml_format_name(v, "cache_v_l%d", i);
+            cache.k_l.push_back(k);
+            cache.v_l.push_back(v);
+            continue;
+        }
+
+         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
+         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
+ 
+@@ -5460,12 +5540,14 @@ static void llm_load_hparams(
+     }
+ 
+     // zero-out the per-layer hparams
+-    std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
+-    std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
+-    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(hparams.n_head_arr.begin(),             hparams.n_head_arr.end(),        0);
+    std::fill(hparams.n_head_kv_arr.begin(),          hparams.n_head_kv_arr.end(),     0);
+    std::fill(hparams.n_ff_arr.begin(),               hparams.n_ff_arr.end(),          0);
+    std::fill(hparams.cross_attn_layers.begin(),      hparams.cross_attn_layers.end(), -1);
+ 
+-    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff_arr,   hparams.n_layer);
+-    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH,       hparams.n_ff_arr,          hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT,      hparams.n_head_arr,        hparams.n_layer);
+    ml.get_arr(LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, hparams.cross_attn_layers, false);
+ 
+     // n_head_kv is optional, default to n_head
+     hparams.n_head_kv_arr = hparams.n_head_arr;
+@@ -5514,7 +5596,7 @@ static void llm_load_hparams(
+ 
+         ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
+ 
+-        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_MLLAMA || model.arch == LLM_ARCH_FALCON) {
+             if (hparams.n_rot != hparams.n_embd_head_k) {
+                 throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
+             }
+@@ -5554,6 +5636,16 @@ static void llm_load_hparams(
+                     }
+                 }
+             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_11B; break;
+                    case 100: model.type = e_model::MODEL_90B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+         case LLM_ARCH_MINICPM:
+             {
+                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+@@ -7249,6 +7341,55 @@ static bool llm_load_tensors(
+                         layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                     }
+                 } break;
+            case LLM_ARCH_MLLAMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab+8});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        if (hparams.cross_attention_layers(i)) {
+                            layer.cross_attn_k_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_NORM,   "weight", i), {128});
+                            layer.cross_attn_k_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_K_PROJ,   "weight", i), {n_embd, 1024});
+                            layer.cross_attn_o_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_O_PROJ,   "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_q_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_NORM, "weight", i), {128});
+                            layer.cross_attn_q_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_Q_PROJ, "weight", i), {n_embd, n_embd});
+                            layer.cross_attn_v_proj = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_V_PROJ, "weight", i), {n_embd, 1024});
+                            layer.cross_attn_attn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_ATTN_GATE, i), {1});
+                            layer.cross_attn_mlp_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_CROSS_ATTN_MLP_GATE, i), {1});
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        } else {
+                            layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                            layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+                            layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                            layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        }
+                    }
+                } break;
+             case LLM_ARCH_GROK:
+                 {
+                     if (n_expert == 0) {
+@@ -9093,7 +9234,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+ 
+         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
+             model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+-            throw std::runtime_error("vocab size mismatch");
+            LLAMA_LOG_WARN("%s: vocab mismatch %u !- %zu ...\n", __func__, model.hparams.n_vocab, model.vocab.id_to_token.size());
+         }
+ 
+         if (params.vocab_only) {
+@@ -9193,6 +9334,21 @@ static struct ggml_tensor * llm_build_inp_embd(
+     return inpL;
+ }
+ 
+static struct ggml_tensor * llm_build_inp_cross_attn_state(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+        const llama_hparams & hparams,
+         const llm_build_cb & cb) {
+    const int64_t n_embd = hparams.n_embd;
+
+    struct ggml_tensor * inpCAS = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, 1601, 4);
+    cb(inpCAS, "inp_cross_attn_state", -1);
+    ggml_set_input(inpCAS);
+    lctx.inp_cross_attn_state = inpCAS;
+
+    return inpCAS;
+}
+
+ static void llm_build_kv_store(
+         struct ggml_context * ctx,
+         const llama_hparams & hparams,
+@@ -10167,6 +10323,7 @@ struct llm_build_context {
+         lctx.inp_pos_bucket    = nullptr;
+         lctx.inp_embd_enc      = nullptr;
+         lctx.inp_KQ_mask_cross = nullptr;
+        lctx.inp_cross_attn_state = nullptr;
+     }
+ 
+     void free() {
+@@ -10754,6 +10911,239 @@ struct llm_build_context {
+                 LLM_NORM_RMS, cb, -1);
+         cb(cur, "result_norm", -1);
+ 
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mllama() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        struct ggml_tensor * inpCAS;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+        inpCAS = llm_build_inp_cross_attn_state(ctx0, lctx, hparams, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            if (hparams.cross_attention_layers(il)) {
+                if (!batch.embd && !cparams.cross_attn) {
+                    continue;
+                }
+
+                // cross attention layer
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_q_proj, cur);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 2, 1, 3));
+                cb(Qcur, "Qcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].cross_attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur, * Vcur;
+                if (batch.embd) {
+                    Kcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_k_proj, inpCAS);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 6404);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].cross_attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
+                    cb(Kcur, "Kcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, kv_self.k_l[il]));
+
+                    Vcur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_v_proj, inpCAS);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, 6404);
+                    cb(Vcur, "Vcur", il);
+
+                    Vcur = ggml_permute(ctx0, Vcur, 0, 2, 1, 3);
+                    cb(Vcur, "Vcur", il);
+
+                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, kv_self.v_l[il]));
+                } else {
+                    Kcur = ggml_view_tensor(ctx0, kv_self.k_l[il]);
+                    cb(Kcur, "Kcur (view)", il);
+
+                    Vcur = ggml_view_tensor(ctx0, kv_self.v_l[il]);
+                    cb(Vcur, "Vcur (view)", il);
+                }
+
+                struct ggml_tensor * kq = ggml_mul_mat(ctx0, Kcur, Qcur);
+                cb(kq, "kq", il);
+
+                // TODO: apply causal masks
+                struct ggml_tensor * kq_soft_max = ggml_soft_max_ext(ctx0, kq, nullptr, 1.f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+                cb(kq_soft_max, "kq_soft_max", il);
+
+                Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, Vcur));
+                cb(Vcur, "Vcur", il);
+
+                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, Vcur, kq_soft_max);
+                cb(kqv, "kqv", il);
+
+                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                cb(kqv_merged, "kqv_merged", il);
+
+                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+                cb(cur, "kqv_merged_cont", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].cross_attn_o_proj, cur);
+                cb(cur, "cur", il);
+
+                // TODO: do this in place once?
+                cur = ggml_mul(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_attn_gate));
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                // TODO: do this inplace once?
+                cur = ggml_add_inplace(ctx0, ggml_mul_inplace(ctx0, cur, ggml_tanh(ctx0, model.layers[il].cross_attn_mlp_gate)), ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            } else {
+                // self attention layer
+
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+         // lm_head
+         cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+         cb(cur, "result_output", -1);
+@@ -16501,6 +16891,10 @@ static struct ggml_cgraph * llama_build_graph(
+             {
+                 result = llm.build_llama();
+             } break;
+        case LLM_ARCH_MLLAMA:
+            {
+                result = llm.build_mllama();
+            } break;
+         case LLM_ARCH_BAICHUAN:
+             {
+                 result = llm.build_baichuan();
+@@ -16761,10 +17155,19 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) {
+     }
+ 
+     if (batch.embd) {
+-        const int64_t n_embd   = hparams.n_embd;
+-        const int64_t n_tokens = batch.n_tokens;
+        if (lctx.inp_cross_attn_state && lctx.inp_cross_attn_state->buffer) {
+            ggml_backend_tensor_set(lctx.inp_cross_attn_state, batch.embd, 0, ggml_nbytes(lctx.inp_cross_attn_state));
+            // zero out inp_embd since it's not used
+            float * inp_embd_data = (float *)lctx.inp_embd->data;
+            for (int i = 0; i < ggml_nelements(lctx.inp_embd); ++i) {
+                inp_embd_data[i] = 0.0f;
+            }
+        } else {
+            const int64_t n_embd   = hparams.n_embd;
+            const int64_t n_tokens = batch.n_tokens;
+ 
+-        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+            ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        }
+     }
+ 
+     if (batch.pos && lctx.inp_pos) {
+@@ -17345,7 +17748,7 @@ static int llama_decode_internal(
+         n_outputs = 1;
+     }
+ 
+-    lctx.sbatch.from_batch(batch_all, n_embd,
+    lctx.sbatch.from_batch(batch_all, batch_all.n_embd,
+         /* simple_split */ !kv_self.recurrent,
+         /* logits_all   */ n_outputs == n_tokens_all);
+ 
+@@ -17638,7 +18041,7 @@ static int llama_encode_internal(
+ 
+     const int64_t n_embd = hparams.n_embd;
+ 
+-    lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    lctx.sbatch.from_batch(batch, batch.n_embd, /* simple_split */ true, /* logits_all */ true);
+ 
+     const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens);
+ 
+@@ -18648,7 +19051,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
+         if (llama_model_has_encoder(&model)) {
+             n_attn_layer *= 3;
+         }
+-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        if (qs.n_attention_wv != n_attn_layer) {
+            LLAMA_LOG_WARN("%s: n_attention_wv is unexpected, expected: %d, found: %d\n", __func__, n_attn_layer, qs.n_attention_wv);
+        }
+     }
+ 
+     size_t total_size_org = 0;
+@@ -19814,6 +20219,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+ 
+         // use what we call a normal RoPE, operating on pairs of consecutive head values
+         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_MLLAMA:
+         case LLM_ARCH_BAICHUAN:
+         case LLM_ARCH_STARCODER:
+         case LLM_ARCH_PLAMO:
+@@ -21230,6 +21636,10 @@ void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) {
+     ctx->cparams.causal_attn = causal_attn;
+ }
+ 
+void llama_set_cross_attention(struct llama_context * ctx, bool cross_attention) {
+    ctx->cparams.cross_attn = cross_attention;
+}
+
+ struct llama_batch llama_batch_get_one(
+              llama_token * tokens,
+                  int32_t   n_tokens,
+@@ -21239,6 +21649,7 @@ struct llama_batch llama_batch_get_one(
+         /*n_tokens       =*/ n_tokens,
+         /*tokens         =*/ tokens,
+         /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
+         /*pos            =*/ nullptr,
+         /*n_seq_id       =*/ nullptr,
+         /*seq_id         =*/ nullptr,
+@@ -21254,6 +21665,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
+         /*n_tokens       =*/ 0,
+         /*tokens         =*/ nullptr,
+         /*embd           =*/ nullptr,
+        /*n_embd         =*/ 0,
+         /*pos            =*/ nullptr,
+         /*n_seq_id       =*/ nullptr,
+         /*seq_id         =*/ nullptr,
+@@ -21265,6 +21677,7 @@ struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_
+ 
+     if (embd) {
+         batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+        batch.n_embd = embd;
+     } else {
+         batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
+     }
@@ -0,0 +1,409 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Thu, 17 Oct 2024 17:19:25 -0700
+Subject: [PATCH] add unpad operator
+
+---
+ ggml/include/ggml.h        | 10 ++++
+ ggml/src/ggml-cuda.cu      |  4 ++
+ ggml/src/ggml-cuda/pad.cu  | 46 +++++++++++++++++++
+ ggml/src/ggml-cuda/pad.cuh |  1 +
+ ggml/src/ggml-metal.m      | 33 ++++++++++++++
+ ggml/src/ggml-metal.metal  | 45 ++++++++++++++++++
+ ggml/src/ggml.c            | 93 +++++++++++++++++++++++++++++++++++++-
+ 7 files changed, 230 insertions(+), 2 deletions(-)
+
+diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
+index ce3d92cb..962cb5f7 100644
+--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
+@@ -506,6 +506,7 @@ extern "C" {
+         GGML_OP_POOL_2D_BACK,
+         GGML_OP_UPSCALE, // nearest interpolate
+         GGML_OP_PAD,
+        GGML_OP_UNPAD,
+         GGML_OP_ARANGE,
+         GGML_OP_TIMESTEP_EMBEDDING,
+         GGML_OP_ARGSORT,
+@@ -1764,6 +1765,15 @@ extern "C" {
+             int                  p2,
+             int                  p3);
+ 
+    // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
+    GGML_API struct ggml_tensor * ggml_unpad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
+     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+     // timesteps: [N,]
+     // return: [N, dim]
+diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
+index fe77b81c..6e84af56 100644
+--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
+@@ -2270,6 +2270,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+         case GGML_OP_PAD:
+             ggml_cuda_op_pad(ctx, dst);
+             break;
+        case GGML_OP_UNPAD:
+            ggml_cuda_op_unpad(ctx, dst);
+            break;
+         case GGML_OP_ARANGE:
+             ggml_cuda_op_arange(ctx, dst);
+             break;
+@@ -2992,6 +2995,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
+         case GGML_OP_GROUP_NORM:
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_LEAKY_RELU:
+diff --git a/ggml/src/ggml-cuda/pad.cu b/ggml/src/ggml-cuda/pad.cu
+index aba539e8..39fd4b16 100644
+--- a/ggml/src/ggml-cuda/pad.cu
+++ b/ggml/src/ggml-cuda/pad.cu
+@@ -47,3 +47,49 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+ }
+
+static __global__ void unpad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    }
+}
+
+static void unpad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    unpad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    unpad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
+diff --git a/ggml/src/ggml-cuda/pad.cuh b/ggml/src/ggml-cuda/pad.cuh
+index 8fd386b0..e2ededc3 100644
+--- a/ggml/src/ggml-cuda/pad.cuh
+++ b/ggml/src/ggml-cuda/pad.cuh
+@@ -3,3 +3,4 @@
+ #define CUDA_PAD_BLOCK_SIZE 256
+ 
+ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
+index 829c5e39..25702d85 100644
+--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
+@@ -193,6 +193,7 @@
+     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
+     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
+     GGML_METAL_KERNEL_TYPE_PAD_F32,
+    GGML_METAL_KERNEL_TYPE_UNPAD_F32,
+     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
+     GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
+     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
+@@ -689,6 +690,7 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32,                     unpad_f32,                        true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
+         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
+@@ -846,6 +848,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
+             return false;
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:
+@@ -2655,6 +2658,36 @@ static void ggml_metal_encode_node(
+ 
+                 const int nth = MIN(1024, ne0);
+ 
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_UNPAD:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UNPAD_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN(1024, ne0);
+
+                 [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+             } break;
+         case GGML_OP_ARANGE:
+diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
+index 2b200032..09887511 100644
+--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal.metal
+@@ -2029,6 +2029,51 @@ kernel void kernel_pad_f32(
+     }
+ }
+ 
+kernel void kernel_unpad_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            }
+        }
+
+        return;
+    }
+}
+
+ kernel void kernel_arange_f32(
+     device        char * dst,
+     constant   int64_t & ne0,
+diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
+index bcbc32d9..f4864ac8 100644
+--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
+@@ -2997,6 +2997,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "POOL_2D_BACK",
+     "UPSCALE",
+     "PAD",
+    "UNPAD",
+     "ARANGE",
+     "TIMESTEP_EMBEDDING",
+     "ARGSORT",
+@@ -3030,7 +3031,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+     "OPT_STEP_ADAMW",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+ 
+ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "none",
+@@ -3091,6 +3092,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "pool_2d_back(x)",
+     "upscale(x)",
+     "pad(x)",
+    "unpad(x)",
+     "arange(start, stop, step)",
+     "timestep_embedding(timesteps, dim, max_period)",
+     "argsort(x)",
+@@ -3124,7 +3126,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+     "adamw(x)",
+ };
+ 
+-static_assert(GGML_OP_COUNT == 80, "GGML_OP_COUNT != 80");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+ 
+ static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
+ 
+@@ -6955,6 +6957,32 @@ struct ggml_tensor * ggml_pad(
+     return result;
+ }
+ 
+// ggml_unpad
+
+struct ggml_tensor * ggml_unpad(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+    bool is_node = false;
+
+    if (a->grad) {
+        GGML_ABORT("fatal error"); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] - p0,
+            a->ne[1] - p1,
+            a->ne[2] - p2,
+            a->ne[3] - p3);
+
+    result->op = GGML_OP_UNPAD;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+ // ggml_arange
+ 
+ struct ggml_tensor * ggml_arange(
+@@ -15312,6 +15340,58 @@ static void ggml_compute_forward_pad(
+     }
+ }
+ 
+static void ggml_compute_forward_unpad_f32(
+    const struct ggml_compute_params *params,
+    struct ggml_tensor *dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_unpad(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_unpad_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+ 
+ // ggml_compute_forward_arange
+ 
+@@ -17294,6 +17374,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
+             {
+                 ggml_compute_forward_pad(params, tensor);
+             } break;
+        case GGML_OP_UNPAD:
+            {
+                ggml_compute_forward_unpad(params, tensor);
+            } break;
+         case GGML_OP_ARANGE:
+             {
+                 ggml_compute_forward_arange(params, tensor);
+@@ -18369,6 +18453,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
+             {
+                 GGML_ABORT("fatal error"); // TODO: not implemented
+             }
+        case GGML_OP_UNPAD:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
+         case GGML_OP_ARANGE:
+             {
+                 GGML_ABORT("fatal error"); // TODO: not implemented
+@@ -19165,6 +19253,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+             } break;
+         case GGML_OP_UPSCALE:
+         case GGML_OP_PAD:
+        case GGML_OP_UNPAD:
+         case GGML_OP_ARANGE:
+         case GGML_OP_TIMESTEP_EMBEDDING:
+         case GGML_OP_ARGSORT:
@@ -0,0 +1,66 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Daniel Hiltgen <daniel@ollama.com>
+Date: Fri, 25 Oct 2024 16:25:18 -0700
+Subject: [PATCH] fix deepseek deseret regex
+
+On windows compiled with gcc the c++ regex library failed to handle
+the characters
+---
+ src/llama-vocab.cpp |  2 +-
+ src/unicode.cpp     | 21 +++++++++++++++++++++
+ 2 files changed, 22 insertions(+), 1 deletion(-)
+
+diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
+index d2f34ddd..3ef6af19 100644
+--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
+@@ -389,7 +389,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
+             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
+                 regex_exprs = {
+                     "[\r\n]",
+-                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿǄ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿﬀ-ﬆﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+                    "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿǄ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿﬀ-ﬆﬓ-ﬗＡ-Ｚａ-ｚ\U00010400-\U0001044f𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+",
+                     "\\s?[!-/:-~！-／：-～‘-‟　-。]+",
+                     "\\s+$",
+                     "[一-龥ࠀ-一가-퟿]+",
+diff --git a/src/unicode.cpp b/src/unicode.cpp
+index f4e941cd..9d78ff16 100644
+--- a/src/unicode.cpp
+++ b/src/unicode.cpp
+@@ -2,6 +2,11 @@
+ #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+ #endif
+ 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+
+ #include "unicode.h"
+ #include "unicode-data.h"
+ 
+@@ -201,8 +206,24 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
+ }
+ 
+ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, NULL, 0);
+    if (!wlen) {
+        throw std::invalid_argument("failed to convert regex");
+    }
+    wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+    wlen = MultiByteToWideChar(CP_UTF8, 0, s.c_str(), -1, wbuf, wlen);
+    if (!wlen) {
+        free(wbuf);
+        throw std::invalid_argument("failed to convert regex");
+    }
+    std::wstring ret = std::wstring(wbuf);
+    free(wbuf);
+    return ret;
+#else
+     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
+     return conv.from_bytes(s);
+#endif
+ }
+ 
+ static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
@@ -2,7 +2,7 @@ package main

 import (
 	"errors"
-	"hash/maphash"
+	"fmt"
 	"log/slog"
 	"reflect"
 	"time"
@@ -20,14 +20,14 @@ type InputCache struct {
 	// optimize cache eviction for multiple users
 	multiUserCache bool

-	// cache of images to embeddings
-	images    []imageCache
-	imageHash maphash.Hash
-
 	lc *llama.Context
 }

-func NewInputCache(lc *llama.Context, kvSize int, numSlots int, multiUserCache bool) *InputCache {
+func NewInputCache(lc *llama.Context, kvSize int, numSlots int, multiUserCache bool) (*InputCache, error) {
+	if kvSize/numSlots < 1 {
+		return nil, fmt.Errorf("must have at least one kv cache entry per parallel sequence (kv: %v parallel: %v)", kvSize, numSlots)
+	}
+
 	slots := make([]InputCacheSlot, numSlots)

 	for i := range slots {
@@ -41,9 +41,8 @@ func NewInputCache(lc *llama.Context, kvSize int, numSlots int, multiUserCache b
 		numCtx:         kvSize / numSlots,
 		slots:          slots,
 		multiUserCache: multiUserCache,
-		images:         make([]imageCache, numSlots),
 		lc:             lc,
-	}
+	}, nil
 }

 // Locking: Operations on InputCacheSlot (including finding one
@@ -64,7 +63,7 @@ type InputCacheSlot struct {
 	lastUsed time.Time
 }

-func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCacheSlot, []input, int, error) {
+func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCacheSlot, []input, error) {
 	var slot *InputCacheSlot
 	var numPast int
 	var err error
@@ -81,7 +80,7 @@ func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCach
 		slot, numPast, err = c.findBestCacheSlot(prompt)
 	}
 	if err != nil {
-		return nil, nil, 0, err
+		return nil, nil, err
 	}

 	if !cachePrompt {
@@ -108,7 +107,7 @@ func (c *InputCache) LoadCacheSlot(prompt []input, cachePrompt bool) (*InputCach
 	prompt = prompt[numPast:]
 	slot.Inputs = slot.Inputs[:numPast]

-	return slot, prompt, numPast, nil
+	return slot, prompt, nil
 }

 func (c *InputCache) findLongestCacheSlot(prompt []input) (*InputCacheSlot, int, error) {
@@ -200,66 +199,48 @@ func countCommonPrefix(a []input, b []input) int {
 	return count
 }

-func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int, numDiscard int, numPast int) {
+func (c *InputCache) ShiftDiscard(inputLen int, numKeep int) int {
+	targetFree := (c.numCtx - numKeep) / 2
+	targetFree = max(targetFree, 1)
+
+	currentFree := c.numCtx - inputLen
+	discard := targetFree - currentFree
+
+	if discard < 0 {
+		discard = 0
+	}
+
+	return discard
+}
+
+// Frees up space in the KV cache by deleting the oldest half of history and shifting
+// the newest half into that space (saving numKeep inputs at the beginning).
+//
+// Assumes that at least 1 entry can be freed up by shifting (i.e. numKeep < numCtx)
+func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int) error {
+	if numKeep >= c.numCtx {
+		return fmt.Errorf("unable to shift context - keep exceeds context (keep: %v context: %v)", numKeep, c.numCtx)
+	}
+
+	discard := c.ShiftDiscard(len(slot.Inputs), numKeep)
+
+	if discard <= 0 {
+		return nil
+	}
+
+	slog.Debug("context limit hit - shifting", "id", slot.Id, "limit", c.numCtx, "input", len(slot.Inputs),
+		"keep", numKeep, "discard", discard)
+
 	// TODO (jessegross): KV cache removal can fail for certain types of models
-	// server.cpp doesn't handle this, though we can be more graceful
-	c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+numDiscard)
-	c.lc.KvCacheSeqAdd(slot.Id, numKeep+numDiscard, numPast, -numDiscard)
-
-	for i := numKeep + numDiscard; i < len(slot.Inputs); i++ {
-		slot.Inputs[i-numDiscard] = slot.Inputs[i]
+	if !c.lc.KvCacheSeqRm(slot.Id, numKeep, numKeep+discard) {
+		return fmt.Errorf("unable to remove old kv cache entries (id: %v, keep: %v discard: %v)", slot.Id, numKeep, discard)
 	}
-	slot.Inputs = slot.Inputs[:len(slot.Inputs)-numDiscard]
-}
+	c.lc.KvCacheSeqAdd(slot.Id, numKeep+discard, len(slot.Inputs), -discard)

-// Locking: Lookup and store operations on imageCache require a lock
-// to be held that serializes these with each other. Hash does not
-// require a lock nor they need to be serialized with InputCacheSlot.
-
-type imageCache struct {
-	key      uint64
-	val      [][]float32
-	lastUsed time.Time
-}
-
-func (c *InputCache) HashImage(image []byte) uint64 {
-	c.imageHash.Reset()
-	_, _ = c.imageHash.Write(image)
-	return c.imageHash.Sum64()
-}
-
-var ErrImageNotFound = errors.New("image not found in cache")
-
-func (c *InputCache) FindImage(hash uint64) ([][]float32, error) {
-	for i := range c.images {
-		if c.images[i].key == hash {
-			slog.Debug("loading image embeddings from cache", "entry", i)
-			c.images[i].lastUsed = time.Now()
-			return c.images[i].val, nil
-		}
+	for i := numKeep + discard; i < len(slot.Inputs); i++ {
+		slot.Inputs[i-discard] = slot.Inputs[i]
 	}
+	slot.Inputs = slot.Inputs[:len(slot.Inputs)-discard]

-	return nil, ErrImageNotFound
-}
-
-func (c *InputCache) AddImage(hash uint64, embed [][]float32) {
-	best := time.Now()
-	var bestImage int
-
-	for i := range c.images {
-		if c.images[i].key == hash {
-			bestImage = i
-			break
-		}
-
-		if c.images[i].lastUsed.Compare(best) < 0 {
-			best = c.images[i].lastUsed
-			bestImage = i
-		}
-	}
-
-	slog.Debug("storing image embeddings in cache", "entry", bestImage, "used", c.images[bestImage].lastUsed)
-	c.images[bestImage].key = hash
-	c.images[bestImage].val = embed
-	c.images[bestImage].lastUsed = time.Now()
+	return nil
 }
@@ -1,7 +1,6 @@
 package main

 import (
-	"reflect"
 	"testing"
 	"time"
 )
@@ -229,76 +228,65 @@ func TestFindCacheSlot(t *testing.T) {
 	}
 }

-func TestImageCache(t *testing.T) {
-	cache := NewInputCache(nil, 2048, 4, false)
-
-	valA := [][]float32{{0.1, 0.2}, {0.3}}
-	valB := [][]float32{{0.4}, {0.5}, {0.6}}
-	valC := [][]float32{{0.7}}
-	valD := [][]float32{{0.8}}
-	valE := [][]float32{{0.9}}
-
-	// Empty cache
-	result, err := cache.FindImage(0x5adb61d31933a946)
-	if err != ErrImageNotFound {
-		t.Errorf("found result in empty cache: result %v, err %v", result, err)
+func TestShiftDiscard(t *testing.T) {
+	tests := []struct {
+		name     string
+		numCtx   int
+		numKeep  int
+		inputLen int
+		expected int
+	}{
+		{
+			name:     "Shift",
+			numCtx:   2048,
+			numKeep:  5,
+			inputLen: 2048,
+			expected: 1021,
+		},
+		{
+			name:     "Max Keep",
+			numCtx:   2048,
+			numKeep:  2047,
+			inputLen: 2048,
+			expected: 1,
+		},
+		{
+			name:     "No Keep",
+			numCtx:   2048,
+			numKeep:  0,
+			inputLen: 2048,
+			expected: 1024,
+		},
+		{
+			name:     "Truncate",
+			numCtx:   2048,
+			numKeep:  5,
+			inputLen: 5000,
+			expected: 3973,
+		},
+		{
+			name:     "Truncate Keep",
+			numCtx:   2048,
+			numKeep:  2047,
+			inputLen: 5000,
+			expected: 2953,
+		},
+		{
+			name:     "No Op",
+			numCtx:   2048,
+			numKeep:  5,
+			inputLen: 512,
+			expected: 0,
+		},
 	}

-	// Insert A
-	cache.AddImage(0x5adb61d31933a946, valA)
-
-	result, err = cache.FindImage(0x5adb61d31933a946)
-	if !reflect.DeepEqual(result, valA) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-
-	// Insert B
-	cache.AddImage(0x011551369a34a901, valB)
-
-	result, err = cache.FindImage(0x5adb61d31933a946)
-	if !reflect.DeepEqual(result, valA) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0x011551369a34a901)
-	if !reflect.DeepEqual(result, valB) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-
-	// Replace B with C
-	cache.AddImage(0x011551369a34a901, valC)
-
-	result, err = cache.FindImage(0x5adb61d31933a946)
-	if !reflect.DeepEqual(result, valA) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0x011551369a34a901)
-	if !reflect.DeepEqual(result, valC) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-
-	// Evict A
-	cache.AddImage(0x756b218a517e7353, valB)
-	cache.AddImage(0x75e5e8d35d7e3967, valD)
-	cache.AddImage(0xd96f7f268ca0646e, valE)
-
-	result, err = cache.FindImage(0x5adb61d31933a946)
-	if reflect.DeepEqual(result, valA) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0x756b218a517e7353)
-	if !reflect.DeepEqual(result, valB) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0x011551369a34a901)
-	if !reflect.DeepEqual(result, valC) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0x75e5e8d35d7e3967)
-	if !reflect.DeepEqual(result, valD) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
-	}
-	result, err = cache.FindImage(0xd96f7f268ca0646e)
-	if !reflect.DeepEqual(result, valE) {
-		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			c := InputCache{numCtx: tt.numCtx}
+			result := c.ShiftDiscard(tt.inputLen, tt.numKeep)
+			if result != tt.expected {
+				t.Errorf("shiftDiscard(ctx: %v, keep: %v input: %v): have %v; want %v", tt.numCtx, tt.numKeep, tt.inputLen, result, tt.expected)
+			}
+		})
 	}
 }
@@ -0,0 +1,183 @@
+package main
+
+import (
+	"errors"
+	"fmt"
+	"hash/maphash"
+	"log/slog"
+	"slices"
+	"sync"
+	"time"
+
+	"github.com/ollama/ollama/llama"
+)
+
+const imageCacheSize = 4
+
+type ImageContext struct {
+	// mu is required to be held when generating embeddings or accessing the cache
+	mu sync.Mutex
+
+	clip   *llama.ClipContext
+	mllama *llama.MllamaContext
+
+	// cache of images to embeddings
+	images    []imageCache
+	imageHash maphash.Hash
+}
+
+func NewImageContext(llamaContext *llama.Context, modelPath string) (*ImageContext, error) {
+	arch, err := llama.GetModelArch(modelPath)
+	if err != nil {
+		return nil, fmt.Errorf("unable to determine vision architecture: %w (%s)", err, modelPath)
+	}
+
+	var c ImageContext
+	if arch == "clip" {
+		c.clip, err = llama.NewClipContext(llamaContext, modelPath)
+	} else if arch == "mllama" {
+		c.mllama, err = llama.NewMllamaContext(llamaContext, modelPath)
+	} else {
+		return nil, fmt.Errorf("unknown vision model architecture: %s", arch)
+	}
+
+	if err != nil {
+		return nil, err
+	}
+
+	c.images = make([]imageCache, imageCacheSize)
+
+	return &c, nil
+}
+
+func (c *ImageContext) Free(modelPath string) {
+	if c == nil {
+		return
+	}
+
+	if c.clip != nil {
+		c.clip.Free()
+	}
+	if c.mllama != nil {
+		c.mllama.Free()
+	}
+}
+
+func (c *ImageContext) NewEmbed(llamaContext *llama.Context, data []byte, aspectRatioId int) ([][]float32, error) {
+	if c == nil {
+		return nil, nil
+	}
+
+	if len(data) <= 0 {
+		return nil, errors.New("received zero length image")
+	}
+
+	hash := c.hashImage(data)
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	embed, err := c.findImage(hash)
+	if err != nil {
+		if c.mllama != nil {
+			embed, err = c.mllama.NewEmbed(llamaContext, data, aspectRatioId)
+			if err != nil {
+				return nil, err
+			}
+		} else if c.clip != nil {
+			embed, err = c.clip.NewEmbed(llamaContext, data)
+			if err != nil {
+				return nil, err
+			}
+		} else {
+			return nil, errors.New("received image but vision model not loaded")
+		}
+
+		c.addImage(hash, embed)
+	}
+
+	return embed, nil
+}
+
+func (c *ImageContext) BatchSize(configuredBatchSize int) int {
+	// If images are not supported, we don't need to allocate embedding batches
+	if c == nil {
+		return 0
+	}
+
+	// Mllama maps an image to 1 embedding token (llava creates many tokens)
+	// and doesn't support more than a single image per request.
+	// The embeddings are large (100 MB), so allocating a big batch can fail
+	// on some systems
+	if c.mllama != nil {
+		return 1
+	}
+
+	return configuredBatchSize
+}
+
+func (c *ImageContext) EmbedSize(llamaContext *llama.Context) int {
+	if c != nil && c.mllama != nil {
+		return c.mllama.EmbedSize(llamaContext)
+	} else {
+		return llamaContext.Model().NEmbd()
+	}
+}
+
+func (c *ImageContext) NeedCrossAttention(inputs ...input) bool {
+	if c == nil || c.mllama == nil {
+		return false
+	}
+
+	return slices.ContainsFunc(inputs, func(input input) bool {
+		return input.embed != nil
+	})
+}
+
+type imageCache struct {
+	key      uint64
+	val      [][]float32
+	lastUsed time.Time
+}
+
+func (c *ImageContext) hashImage(image []byte) uint64 {
+	c.imageHash.Reset()
+	_, _ = c.imageHash.Write(image)
+	return c.imageHash.Sum64()
+}
+
+var errImageNotFound = errors.New("image not found in cache")
+
+func (c *ImageContext) findImage(hash uint64) ([][]float32, error) {
+	for i := range c.images {
+		if c.images[i].key == hash {
+			slog.Debug("loading image embeddings from cache", "entry", i)
+			c.images[i].lastUsed = time.Now()
+			return c.images[i].val, nil
+		}
+	}
+
+	return nil, errImageNotFound
+}
+
+func (c *ImageContext) addImage(hash uint64, embed [][]float32) {
+	best := time.Now()
+	var bestImage int
+
+	for i := range c.images {
+		if c.images[i].key == hash {
+			bestImage = i
+			break
+		}
+
+		if c.images[i].lastUsed.Compare(best) < 0 {
+			best = c.images[i].lastUsed
+			bestImage = i
+		}
+	}
+
+	slog.Debug("storing image embeddings in cache", "entry", bestImage, "used", c.images[bestImage].lastUsed)
+	c.images[bestImage].key = hash
+	c.images[bestImage].val = embed
+	c.images[bestImage].lastUsed = time.Now()
+}
@@ -0,0 +1,80 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestImageCache(t *testing.T) {
+	cache := ImageContext{images: make([]imageCache, 4)}
+
+	valA := [][]float32{{0.1, 0.2}, {0.3}}
+	valB := [][]float32{{0.4}, {0.5}, {0.6}}
+	valC := [][]float32{{0.7}}
+	valD := [][]float32{{0.8}}
+	valE := [][]float32{{0.9}}
+
+	// Empty cache
+	result, err := cache.findImage(0x5adb61d31933a946)
+	if err != errImageNotFound {
+		t.Errorf("found result in empty cache: result %v, err %v", result, err)
+	}
+
+	// Insert A
+	cache.addImage(0x5adb61d31933a946, valA)
+
+	result, err = cache.findImage(0x5adb61d31933a946)
+	if !reflect.DeepEqual(result, valA) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+
+	// Insert B
+	cache.addImage(0x011551369a34a901, valB)
+
+	result, err = cache.findImage(0x5adb61d31933a946)
+	if !reflect.DeepEqual(result, valA) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0x011551369a34a901)
+	if !reflect.DeepEqual(result, valB) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+
+	// Replace B with C
+	cache.addImage(0x011551369a34a901, valC)
+
+	result, err = cache.findImage(0x5adb61d31933a946)
+	if !reflect.DeepEqual(result, valA) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0x011551369a34a901)
+	if !reflect.DeepEqual(result, valC) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+
+	// Evict A
+	cache.addImage(0x756b218a517e7353, valB)
+	cache.addImage(0x75e5e8d35d7e3967, valD)
+	cache.addImage(0xd96f7f268ca0646e, valE)
+
+	result, err = cache.findImage(0x5adb61d31933a946)
+	if reflect.DeepEqual(result, valA) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0x756b218a517e7353)
+	if !reflect.DeepEqual(result, valB) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0x011551369a34a901)
+	if !reflect.DeepEqual(result, valC) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0x75e5e8d35d7e3967)
+	if !reflect.DeepEqual(result, valD) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+	result, err = cache.findImage(0xd96f7f268ca0646e)
+	if !reflect.DeepEqual(result, valE) {
+		t.Errorf("failed to find expected value: result %v, err %v", result, err)
+	}
+}
@@ -18,6 +18,9 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
+
+	"golang.org/x/sync/semaphore"

 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llama"
@@ -33,9 +36,6 @@ type input struct {
 }

 type Sequence struct {
-	// number of inputs evaluated
-	numPast int
-
 	// batch index
 	iBatch int

@@ -45,12 +45,19 @@ type Sequence struct {
 	// prompt inputs left to evaluate
 	inputs []input

+	// inputs that have been added to a batch but not yet submitted to Decode
+	pendingInputs []input
+
 	// tokens that have been generated but not returned yet (e.g. for stop sequences)
 	pendingResponses []string

 	// input cache being used by this sequence
 	cache *InputCacheSlot

+	// does this sequence require cross-attention layers to be processed? - if we have seen
+	// an image for certain multi-modal models
+	crossAttention bool
+
 	// channel to send responses over
 	responses chan string

@@ -107,26 +114,28 @@ func (s *Server) NewSequence(prompt string, images []ImageData, params NewSequen
 		params.numKeep = len(inputs)
 	}

-	if !params.embedding {
-		// Subtracting 4 ensures that at least 1 input can be discarded during shift
-		params.numKeep = min(params.numKeep, s.cache.numCtx-4)
-		params.numKeep += s.bosToken
-	} else {
-		// Embeddings are 1 shot - just truncate to the context window, without ever shifting
-		params.numKeep = min(params.numKeep, s.cache.numCtx)
+	if s.model.AddBOSToken() {
+		params.numKeep += 1
 	}

-	// truncate to fit in context window
+	// Ensure that at least 1 input can be discarded during shift
+	params.numKeep = min(params.numKeep, s.cache.numCtx-1)
+
 	if len(inputs) > s.cache.numCtx {
-		slog.Warn("truncating input prompt", "limit", s.cache.numCtx, "prompt", len(inputs), "numKeep", params.numKeep)
+		discard := len(inputs) - s.cache.numCtx
 		newInputs := inputs[:params.numKeep]
-		newInputs = append(newInputs, inputs[len(inputs)-s.cache.numCtx+params.numKeep:]...)
+		newInputs = append(newInputs, inputs[params.numKeep+discard:]...)
+
+		slog.Warn("truncating input prompt", "limit", s.cache.numCtx, "prompt", len(inputs), "keep", params.numKeep, "new", len(newInputs))
 		inputs = newInputs
 	}

 	var sc *llama.SamplingContext
 	if params.samplingParams != nil {
-		sc = llama.NewSamplingContext(s.model, *params.samplingParams)
+		sc, err = llama.NewSamplingContext(s.model, *params.samplingParams)
+		if err != nil {
+			return nil, err
+		}
 		for _, input := range inputs {
 			if input.embed == nil {
 				sc.Accept(input.token, false)
@@ -155,22 +164,26 @@ func (s *Server) NewSequence(prompt string, images []ImageData, params NewSequen
 // generating image embeddings for each image
 func (s *Server) inputs(prompt string, images []ImageData) ([]input, error) {
 	var inputs []input
+	var parts []string
+	var matches [][]string

-	re := regexp.MustCompile(`\[img-(\d+)\]`)
-	parts := re.Split(prompt, -1)
-	matches := re.FindAllStringSubmatch(prompt, -1)
+	if s.image != nil {
+		re := regexp.MustCompile(`\[img-(\d+)\]`)
+		parts = re.Split(prompt, -1)
+		matches = re.FindAllStringSubmatch(prompt, -1)
+	} else {
+		parts = []string{prompt}
+	}

 	for i, part := range parts {
 		// text - tokenize
-		if strings.TrimSpace(part) != "" {
-			tokens, err := s.lc.Model().Tokenize(part, i == 0, true)
-			if err != nil {
-				return nil, err
-			}
+		tokens, err := s.lc.Model().Tokenize(part, i == 0, true)
+		if err != nil {
+			return nil, err
+		}

-			for _, t := range tokens {
-				inputs = append(inputs, input{token: t})
-			}
+		for _, t := range tokens {
+			inputs = append(inputs, input{token: t})
 		}

 		// image - generate image embedding
@@ -189,16 +202,10 @@ func (s *Server) inputs(prompt string, images []ImageData) ([]input, error) {
 				return nil, fmt.Errorf("invalid image index: %d", n)
 			}

-			hash := s.cache.HashImage(images[imageIndex].Data)
-
-			// Vision models cannot be accessed concurrently
-			s.clip.mu.Lock()
-			embed, err := s.cache.FindImage(hash)
+			embed, err := s.image.NewEmbed(s.lc, images[imageIndex].Data, images[imageIndex].AspectRatioID)
 			if err != nil {
-				embed = llama.NewLlavaImageEmbed(s.lc, s.clip.cc, images[imageIndex].Data)
-				s.cache.AddImage(hash, embed)
+				return nil, err
 			}
-			s.clip.mu.Unlock()

 			for _, e := range embed {
 				inputs = append(inputs, input{embed: e})
@@ -209,46 +216,52 @@ func (s *Server) inputs(prompt string, images []ImageData) ([]input, error) {
 	return inputs, nil
 }

-type clip struct {
-	cc *llama.ClipContext
-	mu sync.Mutex
-}
-
 type Server struct {
+	// is the server ready to process requests?
+	// protects access to model and image
+	ready sync.WaitGroup
+
+	// loaded model
 	model *llama.Model
-	lc    *llama.Context

-	// required for image embeddings
-	clip clip
+	// image model context for multi-modal models
+	image *ImageContext

-	batchSize int
+	// status for external health reporting - loading, ready to serve, etc.
+	status ServerStatus

-	// parallel is the number of parallel requests to handle
+	// current progress on loading the model
+	progress float32
+
+	// number of simultaneous requests to handle
 	parallel int

-	// seqs is the list of parallel sequences being evaluated
-	// TODO (jmorganca): this can probably be moved into run()
+	// maximum number of elements in a batch (per sequence)
+	// TODO (jmorganca): make this n_batch
+	batchSize int
+
+	// protects access to everything below this line
+	// this is context state needed for decoding
+	mu sync.Mutex
+
+	// indicates that data is ready for processing
+	cond *sync.Cond
+
+	// decoding state
+	lc *llama.Context
+
+	// the list of simultaneous sequences being evaluated
 	seqs []*Sequence

+	// seqs can have a maximum of parallel entries, which
+	// is enfoced by seqSem
+	seqsSem *semaphore.Weighted
+
 	// KV cache
 	cache *InputCache

-	// does this model require a beginning of sequence token?
-	bosToken int
-
 	// next sequence for prompt processing to avoid starvation
 	nextSeq int
-
-	// is the server ready to process requests?
-	ready sync.WaitGroup
-
-	mu sync.Mutex
-
-	cond *sync.Cond
-
-	progress float32
-
-	status ServerStatus
 }

 func (s *Server) allNil() bool {
@@ -260,30 +273,30 @@ func (s *Server) allNil() bool {
 	return true
 }

-func (s *Server) shiftContext(seq *Sequence) {
-	numLeft := seq.numPast - seq.numKeep
-	numDiscard := numLeft / 2
-
-	slog.Debug("context limit hit - shifting", "limit", s.cache.numCtx, "numPast", seq.numPast,
-		"numKeep", seq.numKeep, "numLeft", numLeft, "numDiscard", numDiscard)
-
-	s.cache.ShiftCacheSlot(seq.cache, seq.numKeep, numDiscard, seq.numPast)
-
-	seq.numPast -= numDiscard
-}
-
 func flushPending(seq *Sequence) bool {
-	for _, p := range seq.pendingResponses {
-		select {
-		case seq.responses <- p:
-		case <-seq.quit:
-			seq.pendingResponses = []string{}
-			return false
-		}
+	joined := strings.Join(seq.pendingResponses, "")
+	seq.pendingResponses = []string{}
+
+	// Check if there are any partial UTF-8 characters remaining.
+	// We already check and queue as we are generating but some may
+	// still make it here:
+	// - Sequence is ending, e.g. generation limit has been hit
+	// - Invalid characters in the middle of a string
+	// This is a stricter check to ensure we never output invalid Unicode.
+	for !utf8.ValidString(joined) {
+		joined = joined[:len(joined)-1]
 	}

-	seq.pendingResponses = []string{}
-	return true
+	if len(joined) == 0 {
+		return true
+	}
+
+	select {
+	case seq.responses <- joined:
+		return true
+	case <-seq.quit:
+		return false
+	}
 }

 func (s *Server) removeSequence(seqIndex int, reason string) {
@@ -295,25 +308,42 @@ func (s *Server) removeSequence(seqIndex int, reason string) {
 	close(seq.embedding)
 	seq.cache.InUse = false
 	s.seqs[seqIndex] = nil
+	s.seqsSem.Release(1)
 }

 func (s *Server) run(ctx context.Context) {
 	s.ready.Wait()

-	// logically these batches are used only within the context of processBatch
+	// Logically these batches are used only within the context of processBatch
 	// but it is better for performance to allocate them once here
-	tokenBatch := llama.NewBatch(s.batchSize*len(s.seqs), 0, len(s.seqs))
+	tokenBatch, err := llama.NewBatch(s.batchSize, len(s.seqs), 0)
+	if err != nil {
+		panic(err)
+	}
 	defer tokenBatch.Free()

-	embedBatch := llama.NewBatch(s.batchSize*len(s.seqs), s.lc.Model().NEmbd(), len(s.seqs))
-	defer embedBatch.Free()
+	var embedBatch *llama.Batch
+	embedBatchSize := s.image.BatchSize(s.batchSize)
+	if embedBatchSize != 0 {
+		embedBatch, err = llama.NewBatch(embedBatchSize, len(s.seqs), s.image.EmbedSize(s.lc))
+		if err != nil {
+			panic(err)
+		}
+		defer embedBatch.Free()
+	} else {
+		embedBatch = &llama.Batch{}
+	}

 	for {
 		select {
 		case <-ctx.Done():
 			return
 		default:
-			s.processBatch(tokenBatch, embedBatch)
+			err := s.processBatch(tokenBatch, embedBatch)
+			if err != nil {
+				panic(err)
+			}
+
 			tokenBatch.Clear()
 			embedBatch.Clear()
 		}
@@ -327,7 +357,7 @@ func (s *Server) run(ctx context.Context) {
 // these should instead be handled by the handlers
 // it should only be responsible for accepting tokens or embeddings and
 // processing batches as fast as possible
-func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch) {
+func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch) error {
 	s.mu.Lock()
 	for s.allNil() {
 		s.cond.Wait() // Wait until an item is added
@@ -335,6 +365,7 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 	defer s.mu.Unlock()

 	var batch *llama.Batch
+	crossAttention := false

 	seqIdx := s.nextSeq - 1
 	for range s.seqs {
@@ -346,17 +377,23 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 		}

 		// if past the num predict limit
-		if seq.numPredict > 0 && seq.numPredicted > seq.numPredict {
+		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
 			s.removeSequence(seqIdx, "limit")
 			continue
 		}

-		if seq.numPast+len(seq.inputs) > s.cache.numCtx {
-			s.shiftContext(seq)
-		}
-
-		var numInputsProcessed int
 		for i, input := range seq.inputs {
+			if len(seq.cache.Inputs)+len(seq.pendingInputs)+1 > s.cache.numCtx {
+				if len(seq.pendingInputs) == 0 {
+					err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
+					if err != nil {
+						return err
+					}
+				} else {
+					break
+				}
+			}
+
 			embedding := input.embed != nil

 			// If we don't currently have a batch, use one of the correct type and
@@ -368,37 +405,49 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 					batch = tokenBatch
 				} else {
 					batch = embedBatch
+					seq.crossAttention = s.image.NeedCrossAttention(input)
 				}
-			} else if embedding != batch.IsEmbedding() {
+			} else if embedding != batch.IsEmbedding() || crossAttention != seq.crossAttention {
 				s.nextSeq = seqIdx
 				break
 			}

-			// todo: make this n_batch
-			if i >= s.batchSize {
+			if i >= batch.Size() {
 				break
 			}

-			batch.Add(input.token, input.embed, seq.numPast, []int{seq.cache.Id}, numInputsProcessed+1 == len(seq.inputs))
-			seq.numPast++
-			numInputsProcessed++
-		}
-
-		if numInputsProcessed > 0 {
-			seq.cache.Inputs = append(seq.cache.Inputs, seq.inputs[:numInputsProcessed]...)
-			seq.inputs = seq.inputs[numInputsProcessed:]
+			crossAttention = seq.crossAttention
+			batch.Add(input.token, input.embed, len(seq.cache.Inputs)+len(seq.pendingInputs), i+1 == len(seq.inputs), seq.cache.Id)
+			seq.pendingInputs = append(seq.pendingInputs, input)
 			seq.iBatch = batch.NumTokens() - 1
 		}
+
+		seq.inputs = seq.inputs[len(seq.pendingInputs):]
 	}

 	if batch == nil || batch.NumTokens() == 0 {
-		return
+		return nil
 	}

+	s.lc.SetCrossAttention(crossAttention)
+
 	err := s.lc.Decode(batch)
 	if err != nil {
-		slog.Error("failed to decode batch", "error", err)
-		return
+		if errors.Is(err, llama.ErrKvCacheFull) {
+			slog.Debug("defragmenting kv cache")
+			s.cache.lc.KvCacheDefrag()
+			err = s.lc.Decode(batch)
+		}
+		if err != nil {
+			return fmt.Errorf("failed to decode batch: %w", err)
+		}
+	}
+
+	if crossAttention {
+		// synchronize state to ensure the cross attention batch is complete.
+		// needed specifically for multi-GPU systems otherwise an inflight
+		// task may be incorrectly invalidated causing a crash
+		s.lc.Synchronize()
 	}

 	for i, seq := range s.seqs {
@@ -406,6 +455,12 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 			continue
 		}

+		// After calling Decode, pending inputs are now in the cache
+		if len(seq.pendingInputs) > 0 {
+			seq.cache.Inputs = append(seq.cache.Inputs, seq.pendingInputs...)
+			seq.pendingInputs = []input{}
+		}
+
 		// don't sample prompt processing
 		if len(seq.inputs) != 0 {
 			continue
@@ -418,7 +473,7 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)

 		// if done processing the prompt, generate an embedding and return
 		if seq.embeddingOnly {
-			embed := s.lc.GetEmbeddingsSeq(i)
+			embed := s.lc.GetEmbeddingsSeq(seq.cache.Id)
 			if embed == nil {
 				embed = s.lc.GetEmbeddingsIth(seq.iBatch)
 			}
@@ -488,6 +543,8 @@ func (s *Server) processBatch(tokenBatch *llama.Batch, embedBatch *llama.Batch)
 			s.removeSequence(i, "connection")
 		}
 	}
+
+	return nil
 }

 // TODO (jmorganca): use structs from the api package to avoid duplication
@@ -517,8 +574,9 @@ type Options struct {
 }

 type ImageData struct {
-	Data []byte `json:"data"`
-	ID   int    `json:"id"`
+	Data          []byte `json:"data"`
+	ID            int    `json:"id"`
+	AspectRatioID int    `json:"aspect_ratio_id"`
 }

 type CompletionRequest struct {
@@ -600,24 +658,42 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 		return
 	}

-	// TODO (jmorganca): add to sequence queue instead of
-	// failing if a slot isn't available
+	// Ensure there is a place to put the sequence, released when removed from s.seqs
+	if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
+		if errors.Is(err, context.Canceled) {
+			slog.Info("aborting completion request due to client closing the connection")
+		} else {
+			slog.Error("Failed to acquire semaphore", "error", err)
+		}
+		return
+	}
+
 	s.mu.Lock()
+	found := false
 	for i, sq := range s.seqs {
 		if sq == nil {
-			seq.cache, seq.inputs, seq.numPast, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
+			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
 			if err != nil {
 				s.mu.Unlock()
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
 				return
 			}
+
+			seq.crossAttention = s.image.NeedCrossAttention(seq.cache.Inputs...)
+
 			s.seqs[i] = seq
 			s.cond.Signal()
+			found = true
 			break
 		}
 	}
 	s.mu.Unlock()

+	if !found {
+		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
+		return
+	}
+
 	for {
 		select {
 		case <-r.Context().Done():
@@ -681,11 +757,21 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 		return
 	}

-	// TODO (jessegross): Wait for a free slot instead of failing and blocking forever
+	// Ensure there is a place to put the sequence, released when removed from s.seqs
+	if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
+		if errors.Is(err, context.Canceled) {
+			slog.Info("aborting embeddings request due to client closing the connection")
+		} else {
+			slog.Error("Failed to acquire semaphore", "error", err)
+		}
+		return
+	}
+
 	s.mu.Lock()
+	found := false
 	for i, sq := range s.seqs {
 		if sq == nil {
-			seq.cache, seq.inputs, seq.numPast, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
+			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
 			if err != nil {
 				s.mu.Unlock()
 				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
@@ -693,11 +779,17 @@ func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
 			}
 			s.seqs[i] = seq
 			s.cond.Signal()
+			found = true
 			break
 		}
 	}
 	s.mu.Unlock()

+	if !found {
+		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
+		return
+	}
+
 	embedding := <-seq.embedding

 	if err := json.NewEncoder(w).Encode(&EmbeddingResponse{
@@ -741,40 +833,64 @@ func (s *Server) health(w http.ResponseWriter, r *http.Request) {
 	}
 }

+type multiLPath []string
+
+func (m *multiLPath) Set(value string) error {
+	*m = append(*m, value)
+	return nil
+}
+
+func (m *multiLPath) String() string {
+	return strings.Join(*m, ", ")
+}
+
 func (s *Server) loadModel(
 	params llama.ModelParams,
 	mpath string,
-	lpath string,
+	lpath multiLPath,
 	ppath string,
 	kvSize int,
+	kvCacheType string,
 	flashAttention bool,
 	threads int,
 	multiUserCache bool,
 ) {
 	llama.BackendInit()

-	s.model = llama.LoadModelFromFile(mpath, params)
+	var err error
+	s.model, err = llama.LoadModelFromFile(mpath, params)
+	if err != nil {
+		panic(err)
+	}

-	ctxParams := llama.NewContextParams(kvSize, s.batchSize*s.parallel, s.parallel, threads, flashAttention)
-	s.lc = llama.NewContextWithModel(s.model, ctxParams)
+	ctxParams := llama.NewContextParams(kvSize, s.batchSize*s.parallel, s.parallel, threads, flashAttention, kvCacheType)
+	s.lc, err = llama.NewContextWithModel(s.model, ctxParams)
+	if err != nil {
+		panic(err)
+	}

-	if lpath != "" {
-		err := s.model.ApplyLoraFromFile(s.lc, lpath, 1.0, threads)
+	if lpath.String() != "" {
+		for _, path := range lpath {
+			err := s.model.ApplyLoraFromFile(s.lc, path, 1.0, threads)
+			if err != nil {
+				panic(err)
+			}
+		}
+	}
+
+	if ppath != "" {
+		var err error
+		s.image, err = NewImageContext(s.lc, ppath)
 		if err != nil {
 			panic(err)
 		}
 	}

-	if s.model.AddBOSToken() {
-		s.bosToken = 1
+	s.cache, err = NewInputCache(s.lc, kvSize, s.parallel, multiUserCache)
+	if err != nil {
+		panic(err)
 	}

-	if ppath != "" {
-		s.clip.cc = llama.NewClipContext(ppath)
-	}
-
-	s.cache = NewInputCache(s.lc, kvSize, s.parallel, multiUserCache)
-
 	s.status = ServerStatusReady
 	s.ready.Done()
 }
@@ -788,7 +904,7 @@ func main() {
 	mainGpu := flag.Int("main-gpu", 0, "Main GPU")
 	flashAttention := flag.Bool("flash-attn", false, "Enable flash attention")
 	kvSize := flag.Int("ctx-size", 2048, "Context (or KV cache) size")
-	lpath := flag.String("lora", "", "Path to lora layer file")
+	kvCacheType := flag.String("kv-cache-type", "", "quantization type for KV cache (default: f16)")
 	port := flag.Int("port", 8080, "Port to expose the server on")
 	threads := flag.Int("threads", runtime.NumCPU(), "Number of threads to use during generation")
 	verbose := flag.Bool("verbose", false, "verbose output (default: disabled)")
@@ -796,13 +912,10 @@ func main() {
 	mlock := flag.Bool("mlock", false, "force system to keep model in RAM rather than swapping or compressing")
 	tensorSplit := flag.String("tensor-split", "", "fraction of the model to offload to each GPU, comma-separated list of proportions")
 	multiUserCache := flag.Bool("multiuser-cache", false, "optimize input cache algorithm for multiple users")
-	// Expose requirements as a JSON output to stdout
 	requirements := flag.Bool("requirements", false, "print json requirement information")

-	// These are either ignored by llama.cpp or have no significance to us
-	_ = flag.Bool("embedding", false, "enable embedding vector output (default: disabled)")
-	_ = flag.Bool("log-disable", false, "disables logging to a file")
-	_ = flag.Bool("memory-f32", false, "use f32 instead of f16 for memory key+value (default: disabled) not recommended: doubles context memory required and no measurable increase in quality")
+	var lpaths multiLPath
+	flag.Var(&lpaths, "lora", "Path to lora layer file (can be specified multiple times)")

 	flag.Parse()
 	if *requirements {
@@ -826,12 +939,13 @@ func main() {
 	})
 	slog.SetDefault(slog.New(handler))
 	slog.Info("starting go runner")
-	slog.Debug("system info", "cpu", llama.PrintSystemInfo(), "threads", *threads)
+	slog.Info("system", "info", llama.PrintSystemInfo(), "threads", *threads)

 	server := &Server{
 		batchSize: *batchSize,
 		parallel:  *parallel,
 		seqs:      make([]*Sequence, *parallel),
+		seqsSem:   semaphore.NewWeighted(int64(*parallel)),
 		status:    ServerStatusLoadingModel,
 	}

@@ -849,7 +963,7 @@ func main() {
 	params := llama.ModelParams{
 		NumGpuLayers: *nGpuLayers,
 		MainGpu:      *mainGpu,
-		UseMmap:      !*noMmap && *lpath == "",
+		UseMmap:      !*noMmap && lpaths.String() == "",
 		UseMlock:     *mlock,
 		TensorSplit:  tensorSplitFloats,
 		Progress: func(progress float32) {
@@ -858,7 +972,7 @@ func main() {
 	}

 	server.ready.Add(1)
-	go server.loadModel(params, *mpath, *lpath, *ppath, *kvSize, *flashAttention, *threads, *multiUserCache)
+	go server.loadModel(params, *mpath, lpaths, *ppath, *kvSize, *kvCacheType, *flashAttention, *threads, *multiUserCache)

 	server.cond = sync.NewCond(&server.mu)