MLX: add header vendoring and remove go build tag (#14642)

* prefer rocm v6 on windows

Avoid building with v7 - more changes are needed

* MLX: add header vendoring and remove go build tag

This switches to using a vendoring approach for the mlx-c headers so that Go
can build without requiring a cmake first.  This enables building the new MLX
based code by default.  Every time cmake runs, the headers are refreshed, so we
can easily keep them in sync when we bump mlx versions.  Basic Windows
and Linux support are verified.

* ci: harden for flaky choco repo servers

CI sometimes fails due to choco not actually installing cache.  Since it just speeds up the build, we can proceed without.

* review comments

.github/workflows/release.yaml

@@ -117,6 +117,25 @@ jobs:
             install: https://sdk.lunarg.com/sdk/download/1.4.321.1/windows/vulkansdk-windows-X64-1.4.321.1.exe
             flags: ''
             runner_dir: 'vulkan'
+          - os: windows
+            arch: amd64
+            preset: 'MLX CUDA 13'
+            install: https://developer.download.nvidia.com/compute/cuda/13.0.0/local_installers/cuda_13.0.0_windows.exe
+            cudnn-install: https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.18.1.3_cuda13-archive.zip
+            cuda-components:
+              - '"cudart"'
+              - '"nvcc"'
+              - '"cublas"'
+              - '"cublas_dev"'
+              - '"cufft"'
+              - '"cufft_dev"'
+              - '"nvrtc"'
+              - '"nvrtc_dev"'
+              - '"crt"'
+              - '"nvvm"'
+              - '"nvptxcompiler"'
+            cuda-version: '13.0'
+            flags: ''
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -125,8 +144,10 @@ jobs:
       - name: Install system dependencies
         run: |
           choco install -y --no-progress ccache ninja
-          ccache -o cache_dir=${{ github.workspace }}\.ccache
-      - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'ROCm ') || startsWith(matrix.preset, 'Vulkan')
+          if (Get-Command ccache -ErrorAction SilentlyContinue) {
+            ccache -o cache_dir=${{ github.workspace }}\.ccache
+          }
+      - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'ROCm ') || startsWith(matrix.preset, 'Vulkan') || startsWith(matrix.preset, 'MLX ')
         id: cache-install
         uses: actions/cache/restore@v4
         with:
@@ -134,8 +155,9 @@ jobs:
             C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
             C:\Program Files\AMD\ROCm
             C:\VulkanSDK
-          key: ${{ matrix.install }}
-      - if: startsWith(matrix.preset, 'CUDA ')
+            C:\Program Files\NVIDIA\CUDNN
+          key: ${{ matrix.install }}-${{ matrix.cudnn-install }}
+      - if: startsWith(matrix.preset, 'CUDA ') || startsWith(matrix.preset, 'MLX ')
         name: Install CUDA ${{ matrix.cuda-version }}
         run: |
           $ErrorActionPreference = "Stop"
@@ -179,6 +201,23 @@ jobs:
         run: |
           echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+      - if: startsWith(matrix.preset, 'MLX ')
+        name: Install cuDNN for MLX
+        run: |
+          $ErrorActionPreference = "Stop"
+          $cudnnRoot = "C:\Program Files\NVIDIA\CUDNN"
+          if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
+            Invoke-WebRequest -Uri "${{ matrix.cudnn-install }}" -OutFile "cudnn.zip"
+            Expand-Archive -Path cudnn.zip -DestinationPath cudnn-extracted
+            $cudnnDir = (Get-ChildItem -Path cudnn-extracted -Directory)[0].FullName
+            New-Item -ItemType Directory -Force -Path $cudnnRoot
+            Copy-Item -Path "$cudnnDir\*" -Destination "$cudnnRoot\" -Recurse
+          }
+
+          echo "CUDNN_ROOT_DIR=$cudnnRoot" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CUDNN_INCLUDE_PATH=$cudnnRoot\include" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CUDNN_LIBRARY_PATH=$cudnnRoot\lib\x64" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "$cudnnRoot\bin\x64" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@v4
         with:
@@ -186,7 +225,8 @@ jobs:
             C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
             C:\Program Files\AMD\ROCm
             C:\VulkanSDK
-          key: ${{ matrix.install }}
+            C:\Program Files\NVIDIA\CUDNN
+          key: ${{ matrix.install }}-${{ matrix.cudnn-install }}
       - uses: actions/checkout@v4
       - uses: actions/cache@v4
         with:
@@ -198,7 +238,7 @@ jobs:
           Enter-VsDevShell -VsInstallPath 'C:\Program Files\Microsoft Visual Studio\2022\Enterprise' -SkipAutomaticLocation  -DevCmdArguments '-arch=x64 -no_logo'
           cmake --preset "${{ matrix.preset }}" ${{ matrix.flags }} --install-prefix "$((pwd).Path)\dist\${{ matrix.os }}-${{ matrix.arch }}"
           cmake --build --parallel ([Environment]::ProcessorCount) --preset "${{ matrix.preset }}"
-          cmake --install build --component "${{ startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || startsWith(matrix.preset, 'Vulkan') && 'Vulkan' || 'CPU' }}" --strip
+          cmake --install build --component "${{ startsWith(matrix.preset, 'MLX ') && 'MLX' || startsWith(matrix.preset, 'CUDA ') && 'CUDA' || startsWith(matrix.preset, 'ROCm ') && 'HIP' || startsWith(matrix.preset, 'Vulkan') && 'Vulkan' || 'CPU' }}" --strip
           Remove-Item -Path dist\lib\ollama\rocm\rocblas\library\*gfx906* -ErrorAction SilentlyContinue
         env:
           CMAKE_GENERATOR: Ninja