ci: improvements from #15982

2026-05-13 06:21:28 +00:00 · 2026-05-08 14:05:11 -07:00 · 2026-05-08 14:05:11 -07:00 · 4d142c0724
commit 4d142c0724
parent 88935c21b5
10 changed files with 250 additions and 38 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -16,7 +16,7 @@ jobs:
    outputs:
      GOFLAGS: ${{ steps.goflags.outputs.GOFLAGS }}
      VERSION: ${{ steps.goflags.outputs.VERSION }}
-      vendorsha: ${{ steps.changes.outputs.vendorsha }}
+      vendorsha: ${{ steps.goflags.outputs.vendorsha }}
    steps:
      - uses: actions/checkout@v4
      - name: Set environment
@ -142,6 +142,13 @@ jobs:
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
    steps:
+      - if: startsWith(matrix.preset, 'MLX ')
+        name: Increase pagefile to 200 GB
+        uses: al-cheb/configure-pagefile-action@v1.5
+        with:
+          minimum-size: 16GB
+          maximum-size: 200GB
+          disk-root: "D:"
      - name: Install system dependencies
        run: |
          choco install -y --no-progress ccache ninja
@ -381,17 +388,33 @@ jobs:
            dist/*.ps1
            dist/OllamaSetup.exe

-  linux-build:
+  linux-depends:
    strategy:
      matrix:
        include:
-          - os: linux
-            arch: amd64
-            target: archive
-          - os: linux
-            arch: arm64
-            target: archive
-    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+          - arch: amd64
+            target: llama-server-cpu
+          - arch: amd64
+            target: llama-server-cuda-v12
+          - arch: amd64
+            target: llama-server-cuda-v13
+          - arch: amd64
+            target: mlx
+          - arch: amd64
+            target: llama-server-rocm
+          - arch: amd64
+            target: llama-server-vulkan
+          - arch: arm64
+            target: llama-server-cpu
+          - arch: arm64
+            target: llama-server-cuda-v12
+          - arch: arm64
+            target: llama-server-cuda-v13
+          - arch: arm64
+            target: jetpack-5
+          - arch: arm64
+            target: jetpack-6
+    runs-on: ${{ matrix.arch == 'arm64' && 'linux-arm64' || 'linux' }}
    environment: release
    needs: setup-environment
    env:
@ -399,17 +422,101 @@ jobs:
    steps:
      - uses: actions/checkout@v4
      - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
+      - if: matrix.target == 'mlx'
+        name: Increase Linux swap to 200 GB
+        shell: bash
+        run: |
+          set -e
+          SWAP_PATH=/swapfile-mlx
+          SWAP_SIZE_GB=200
+          if [ -f "$SWAP_PATH" ]; then
+              sudo swapoff "$SWAP_PATH" 2>/dev/null || true
+              sudo rm -f "$SWAP_PATH"
+          fi
+          if ! sudo fallocate -l ${SWAP_SIZE_GB}G "$SWAP_PATH" 2>/dev/null; then
+              echo "fallocate unsupported, falling back to dd"
+              sudo dd if=/dev/zero of="$SWAP_PATH" bs=1M count=$((SWAP_SIZE_GB * 1024))
+          fi
+          sudo chmod 600 "$SWAP_PATH"
+          sudo mkswap "$SWAP_PATH"
+          sudo swapon "$SWAP_PATH"
+          swapon --show
+          free -h
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/${{ matrix.arch }}
+          target: ${{ matrix.target }}
+          provenance: false
+          sbom: false
+          build-args: |
+            GOFLAGS=${{ env.GOFLAGS }}
+            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
+            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
+            OLLAMA_MLX_BUILD_JOBS=16
+            OLLAMA_MLX_NVCC_THREADS=6
+            APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+          cache-from: |
+            type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }}
+            type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-to: type=registry,ref=ollama/release:cache-${{ matrix.arch }}-${{ matrix.target }},mode=max
+
+  linux-build:
+    strategy:
+      matrix:
+        include:
+          - os: linux
+            arch: amd64
+            target: archive
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda-v12
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda-v13
+              type=registry,ref=ollama/release:cache-amd64-mlx
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm
+              type=registry,ref=ollama/release:cache-amd64-llama-server-vulkan
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          - os: linux
+            arch: arm64
+            target: archive
+            cache-from: |
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda-v12
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda-v13
+              type=registry,ref=ollama/release:cache-arm64-jetpack-5
+              type=registry,ref=ollama/release:cache-arm64-jetpack-6
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
+    environment: release
+    needs: [setup-environment, linux-depends]
+    env:
+      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ vars.DOCKER_USER }}
+          password: ${{ secrets.DOCKER_ACCESS_TOKEN }}
      - uses: docker/build-push-action@v6
        with:
          context: .
          platforms: ${{ matrix.os }}/${{ matrix.arch }}
          target: ${{ matrix.target }}
+          provenance: false
+          sbom: false
          build-args: |
            GOFLAGS=${{ env.GOFLAGS }}
            CGO_CFLAGS=${{ env.CGO_CFLAGS }}
            CGO_CXXFLAGS=${{ env.CGO_CXXFLAGS }}
+            OLLAMA_MLX_BUILD_JOBS=16
+            OLLAMA_MLX_NVCC_THREADS=6
          outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-from: ${{ matrix.cache-from }}
          cache-to: type=inline
      - name: Deduplicate CUDA libraries
        run: |
@ -458,6 +565,15 @@ jobs:
              CGO_CXXFLAGS
              GOFLAGS
              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda-v12
+              type=registry,ref=ollama/release:cache-arm64-llama-server-cuda-v13
+              type=registry,ref=ollama/release:cache-arm64-jetpack-5
+              type=registry,ref=ollama/release:cache-arm64-jetpack-6
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
          - os: linux
            arch: amd64
            build-args: |
@ -465,6 +581,15 @@ jobs:
              CGO_CXXFLAGS
              GOFLAGS
              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda-v12
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cuda-v13
+              type=registry,ref=ollama/release:cache-amd64-mlx
+              type=registry,ref=ollama/release:cache-amd64-llama-server-vulkan
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
          - os: linux
            arch: amd64
            suffix: '-rocm'
@ -474,9 +599,15 @@ jobs:
              GOFLAGS
              FLAVOR=rocm
              APT_MIRROR=http://azure.archive.ubuntu.com/ubuntu
+              OLLAMA_MLX_BUILD_JOBS=16
+              OLLAMA_MLX_NVCC_THREADS=6
+            cache-from: |
+              type=registry,ref=ollama/release:cache-amd64-llama-server-cpu
+              type=registry,ref=ollama/release:cache-amd64-llama-server-rocm
+              type=registry,ref=${{ vars.DOCKER_REPO }}:latest
    runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
    environment: release
-    needs: setup-environment
+    needs: [setup-environment, linux-depends]
    env:
      GOFLAGS: ${{ needs.setup-environment.outputs.GOFLAGS }}
    steps:
@ -491,9 +622,11 @@ jobs:
        with:
          context: .
          platforms: ${{ matrix.os }}/${{ matrix.arch }}
+          provenance: false
+          sbom: false
          build-args: ${{ matrix.build-args }}
          outputs: type=image,name=${{ vars.DOCKER_REPO }},push-by-digest=true,name-canonical=true,push=true
-          cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
+          cache-from: ${{ matrix.cache-from }}
          cache-to: type=inline
      - run: |
          mkdir -p ${{ matrix.os }}-${{ matrix.arch }}
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@ -71,7 +71,7 @@ jobs:
          - preset: 'MLX CUDA 13'
            container: nvidia/cuda:13.0.0-devel-ubuntu22.04
            extra-packages: libcudnn9-dev-cuda-13 libopenblas-dev liblapack-dev liblapacke-dev git curl
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87 -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
+            flags: '-DCMAKE_CUDA_ARCHITECTURES=87 -DMLX_CUDA_ARCHITECTURES=80-virtual -DBLAS_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu -DLAPACK_INCLUDE_DIRS=/usr/include/x86_64-linux-gnu'
            install-go: true
            mlx_preset: 'MLX CUDA 13'
    runs-on: linux
@ -155,7 +155,7 @@ jobs:
          - 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
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
+            flags: '-DCMAKE_CUDA_ARCHITECTURES=80 -DMLX_CUDA_ARCHITECTURES=80-virtual'
            mlx_preset: 'MLX CUDA 13'
            cuda-components:
              - '"cudart"'
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -92,15 +92,21 @@ if(MLX_ENGINE)
        list(APPEND MLX_INCLUDE_REGEXES "^dl\\.dll$")
    endif()

+    # Keep mlx/mlxc targets separate from runtime dependencies so
+    # --strip only applies to the binaries we build, not vendor DLLs/libs.
    install(TARGETS mlx mlxc
-        RUNTIME_DEPENDENCIES
-            DIRECTORIES ${MLX_RUNTIME_DIRS}
-            PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
-            PRE_EXCLUDE_REGEXES ".*"
+        RUNTIME_DEPENDENCY_SET mlx_runtime_deps
        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
    )
+    install(RUNTIME_DEPENDENCY_SET mlx_runtime_deps
+        DIRECTORIES ${MLX_RUNTIME_DIRS}
+        PRE_INCLUDE_REGEXES ${MLX_INCLUDE_REGEXES}
+        PRE_EXCLUDE_REGEXES ".*"
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX_VENDOR
+    )

    if(TARGET jaccl)
        install(TARGETS jaccl
@ -230,8 +236,7 @@ if(MLX_ENGINE)
        if(MLX_CUDA_LIBS)
            install(FILES ${MLX_CUDA_LIBS}
                DESTINATION ${OLLAMA_INSTALL_DIR}
-                COMPONENT MLX)
+                COMPONENT MLX_VENDOR)
        endif()
    endif()
 endif()
-
--- a/CMakePresets.json
+++ b/CMakePresets.json
@ -29,7 +29,8 @@
      "name": "MLX CUDA 13",
      "inherits": [ "MLX" ],
      "cacheVariables": {
-        "MLX_CUDA_ARCHITECTURES": "86;89;90;90a;100;103;75-virtual;80-virtual;110-virtual;120-virtual;121-virtual",
+        "CMAKE_CUDA_FLAGS": "-t 2",
+        "MLX_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
      }
    }
--- a/9
+++ b/9
@ -177,6 +177,8 @@ RUN --mount=type=cache,target=/root/.ccache \

 FROM base AS mlx
 ARG CUDA13VERSION=13.0
+ARG OLLAMA_MLX_BUILD_JOBS=
+ARG OLLAMA_MLX_NVCC_THREADS=2
 RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
    && dnf install -y openblas-devel lapack-devel \
    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
@ -202,9 +204,10 @@ RUN --mount=type=cache,target=/root/.ccache \
    && if [ -f /tmp/local-mlx-c/CMakeLists.txt ]; then \
        export OLLAMA_MLX_C_SOURCE=/tmp/local-mlx-c; \
    fi \
-    && cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
-        && cmake --build --preset 'MLX CUDA 13' -- -l $(nproc) \
-        && cmake --install build --component MLX --strip
+    && cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas -DCMAKE_CUDA_FLAGS="-t ${OLLAMA_MLX_NVCC_THREADS}" \
+        && cmake --build --preset 'MLX CUDA 13' -- -l $(nproc) ${OLLAMA_MLX_BUILD_JOBS:+-j ${OLLAMA_MLX_BUILD_JOBS}} \
+        && cmake --install build --component MLX --strip \
+        && cmake --install build --component MLX_VENDOR

 #
 # Go build
--- a/scripts/build_darwin.sh
+++ b/scripts/build_darwin.sh
@ -71,6 +71,7 @@ _build_darwin() {
                -DOLLAMA_RUNNER_DIR=./
            cmake --build $BUILD_DIR --target mlx mlxc -j
            cmake --install $BUILD_DIR --component MLX
+            cmake --install $BUILD_DIR --component MLX_VENDOR
            # Build llama-server statically (no Metal on x86, CPU-only)
            status "Building darwin $ARCH llama-server (static)"
            cmake -S llama/server --preset cpu \
@ -97,6 +98,7 @@ _build_darwin() {
                -DCMAKE_INSTALL_PREFIX=$INSTALL_PREFIX
            cmake --build $BUILD_DIR --target mlx mlxc --parallel
            cmake --install $BUILD_DIR --component MLX
+            cmake --install $BUILD_DIR --component MLX_VENDOR

            # Metal 4.x build (NAX-enabled, macOS 26+)
            # Only possible with Xcode 26+ SDK; skip on older toolchains.
@ -118,6 +120,7 @@ _build_darwin() {
                    -DFETCHCONTENT_SOURCE_DIR_METAL_CPP=$V3_DEPS/metal_cpp-src
                cmake --build $BUILD_DIR_V4 --target mlx mlxc --parallel
                cmake --install $BUILD_DIR_V4 --component MLX
+                cmake --install $BUILD_DIR_V4 --component MLX_VENDOR
            else
                status "Skipping MLX Metal v4 (SDK $SDK_MAJOR < 26, need Xcode 26+)"
            fi
--- a/scripts/build_linux.sh
+++ b/scripts/build_linux.sh
@ -1,9 +1,14 @@
 #!/bin/sh
 #
-# Mac ARM users, rosetta can be flaky, so to use a remote x86 builder
+# Mac ARM users, rosetta can be flaky, so to use a remote x86 builder.
+# Use the docker-container driver with the bundled buildkit GC config
+# for improved cache behavior.
 #
 # docker context create amd64 --docker host=ssh://mybuildhost
-# docker buildx create --name mybuilder amd64 --platform linux/amd64
+# docker buildx create --name mybuilder \
+#     --driver docker-container \
+#     --config ./scripts/buildkitd.toml.example \
+#     --bootstrap amd64 --platform linux/amd64
 # docker buildx create --name mybuilder --append desktop-linux --platform linux/arm64
 # docker buildx use mybuilder

@ -44,18 +49,27 @@ fi
 # buildx behavior changes for single vs. multiplatform
 echo "Compressing linux tar bundles..."
 if echo $PLATFORM | grep "," > /dev/null ; then
-        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
-        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
-        tar c -C ./dist/linux_amd64 --exclude rocm --exclude 'mlx*' --exclude include . | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
-        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
-        tar c -C ./dist/linux_amd64 ./lib/ollama/mlx_cuda_v13 ./lib/ollama/include  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-mlx.tar.zst
+        tar c -C ./dist/linux_arm64 --exclude cuda_jetpack5 --exclude cuda_jetpack6 . | zstd -9 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack5  | zstd -9 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/linux_arm64 ./lib/ollama/cuda_jetpack6  | zstd -9 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
+        tar c -C ./dist/linux_amd64 --exclude rocm --exclude 'mlx*' . | zstd -9 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/linux_amd64 ./lib/ollama/rocm  | zstd -9 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
+        ( cd ./dist/linux_amd64 && tar c lib/ollama/mlx* ) | zstd -9 -T0 >./dist/ollama-linux-amd64-mlx.tar.zst
 elif echo $PLATFORM | grep "arm64" > /dev/null ; then
-        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64.tar.zst
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
-        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | zstd --ultra -22 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
+        tar c -C ./dist/ --exclude cuda_jetpack5 --exclude cuda_jetpack6 bin lib | zstd -9 -T0 >./dist/ollama-linux-arm64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack5  | zstd -9 -T0 >./dist/ollama-linux-arm64-jetpack5.tar.zst
+        tar c -C ./dist/ ./lib/ollama/cuda_jetpack6  | zstd -9 -T0 >./dist/ollama-linux-arm64-jetpack6.tar.zst
 elif echo $PLATFORM | grep "amd64" > /dev/null ; then
-        tar c -C ./dist/ --exclude rocm --exclude 'mlx*' --exclude include bin lib | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64.tar.zst
-        tar c -C ./dist/ ./lib/ollama/rocm  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
-        tar c -C ./dist/ ./lib/ollama/mlx_cuda_v13 ./lib/ollama/include  | zstd --ultra -22 -T0 >./dist/ollama-linux-amd64-mlx.tar.zst
+        tar c -C ./dist/ --exclude rocm --exclude 'mlx*' bin lib | zstd -9 -T0 >./dist/ollama-linux-amd64.tar.zst
+        tar c -C ./dist/ ./lib/ollama/rocm  | zstd -9 -T0 >./dist/ollama-linux-amd64-rocm.tar.zst
+        ( cd ./dist/ && tar c lib/ollama/mlx* ) | zstd -9 -T0 >./dist/ollama-linux-amd64-mlx.tar.zst
 fi
+
+LIMIT=2147483648
+for f in ./dist/ollama-linux-*.tar.zst; do
+    [ -f "$f" ] || continue
+    size=$(stat -f%z "$f" 2>/dev/null || stat -c%s "$f")
+    if [ "$size" -gt "$LIMIT" ]; then
+        echo "WARNING: $f is $size bytes ($((size - LIMIT)) over the 2 GiB GitHub release-asset limit)" >&2
+    fi
+done
--- a/scripts/build_windows.ps1
+++ b/scripts/build_windows.ps1
@ -368,6 +368,8 @@ function mlxCuda13 {
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
            & cmake --install build\mlx_cuda_v$cudaMajorVer --component "MLX" --strip
            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+            & cmake --install build\mlx_cuda_v$cudaMajorVer --component "MLX_VENDOR"
+            if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
        } else {
            Write-Output "CUDA v$cudaMajorVer not detected, skipping MLX build"
        }
--- a/scripts/buildkitd.toml.example
+++ b/scripts/buildkitd.toml.example
@ -0,0 +1,21 @@
+# Suggested BuildKit GC config for Ollama local development.
+
+[worker.oci]
+  gc = true
+  gckeepstorage = "150GB"
+
+[[worker.oci.gcpolicy]]
+  filters = ["type==source.local", "type==source.git.checkout"]
+  keepDuration = "48h"
+  maxUsedSpace = "5GB"
+
+[[worker.oci.gcpolicy]]
+  filters = ["type==exec.cachemount"]
+  keepDuration = "168h"
+  maxUsedSpace = "20GB"
+
+[[worker.oci.gcpolicy]]
+  keepDuration = "720h"
+  reservedSpace = "20GB"
+  maxUsedSpace = "150GB"
+  minFreeSpace = "50GB"
--- a/x/imagegen/mlx/CMakeLists.txt
+++ b/x/imagegen/mlx/CMakeLists.txt
@ -94,6 +94,36 @@ FetchContent_Declare(
 )
 FetchContent_MakeAvailable(mlx-c)

+# MLX CUDA builds have a long tail across many architecture variants. Using
+# higher NVCC thread counts helps, but every .cu can then spawn multiple
+# memory-heavy compiler processes. Keep those compiles in a Ninja pool sized
+# from host RAM so CI can use more threads without falling over.
+if(CMAKE_GENERATOR STREQUAL "Ninja")
+    file(GLOB_RECURSE _mlx_cu
+        "${mlx_SOURCE_DIR}/mlx/backend/cuda/*.cu"
+        "${mlx_BINARY_DIR}/mlx/backend/cuda/*.cu"
+    )
+    if(_mlx_cu)
+        set(MLX_CUDA_RAM_MB 22000 CACHE STRING
+            "Per-file memory budget in MB for the MLX CUDA compile job pool.")
+        cmake_host_system_information(RESULT _ram_mb QUERY TOTAL_PHYSICAL_MEMORY)
+        math(EXPR _cuda_pool "${_ram_mb} / ${MLX_CUDA_RAM_MB}")
+        if(_cuda_pool LESS 2)
+            set(_cuda_pool 2)
+        endif()
+        set_property(GLOBAL APPEND PROPERTY JOB_POOLS cuda_compile=${_cuda_pool})
+        list(LENGTH _mlx_cu _cu_count)
+        foreach(f ${_mlx_cu})
+            set_property(SOURCE "${f}"
+                TARGET_DIRECTORY mlx
+                PROPERTY JOB_POOL_COMPILE cuda_compile)
+        endforeach()
+        message(STATUS "MLX cuda_compile JOB_POOL: ${_cu_count} files, pool size ${_cuda_pool} (host RAM ${_ram_mb} MB / ${MLX_CUDA_RAM_MB} MB per file)")
+    else()
+        message(WARNING "MLX cuda_compile JOB_POOL: no CUDA sources found under mlx/backend/cuda")
+    endif()
+endif()
+
 # Sync vendored headers with fetched version
 file(GLOB _mlx_c_hdrs "${mlx-c_SOURCE_DIR}/mlx/c/*.h")
 file(COPY ${_mlx_c_hdrs} DESTINATION "${CMAKE_SOURCE_DIR}/x/mlxrunner/mlx/include/mlx/c/")