docker.yml

name: docker

# Build the parakeet-cli container images and publish them to GitHub Container
# Registry (ghcr.io/<owner>/parakeet.cpp-cli).
#
# Each variant (cpu, cuda) is a multi-arch image (linux/amd64 + linux/arm64).
# Every arch is built natively on its own runner (no QEMU): amd64 on
# ubuntu-24.04, arm64 on ubuntu-24.04-arm. The per-arch images are pushed by
# digest, then a merge job assembles one multi-arch manifest per variant.
#
# The CUDA images use the CUDA 13 base so ggml compiles the Blackwell
# architectures (sm_120 + sm_121); that is what makes the arm64 CUDA image run
# on GB10 / Grace-Blackwell (DGX Spark). CUDA 12.6 tops out at sm_90.
#
# pull_request builds the CPU variant only, as a fast Dockerfile gate. The CUDA
# build takes tens of minutes (it compiles many GPU architectures), so it runs
# only on push to the default branch, tags, and manual dispatch, all of which
# also push the image. Use workflow_dispatch to exercise CUDA before merging.
on:
  push:
    branches: [master]
    tags: ['v*']
  pull_request:
  workflow_dispatch:

env:
  REGISTRY: ghcr.io
  # Named for the binary it ships, so a future server image can live alongside
  # it (e.g. parakeet.cpp-server). Resolves to <owner>/parakeet.cpp-cli.
  IMAGE_NAME: ${{ github.repository }}-cli

jobs:
  # -------------------------------------------------------------------------
  # setup: choose the build matrix for this event. PRs get CPU only (fast
  # gate); everything else gets CPU + CUDA.
  # -------------------------------------------------------------------------
  setup:
    runs-on: ubuntu-latest
    outputs:
      matrix: ${{ steps.set.outputs.matrix }}
    steps:
      - name: Select build matrix
        id: set
        run: |
          CPU='{"variant":"cpu","arch":"amd64","runner":"ubuntu-24.04","build_base":"ubuntu:24.04","runtime_base":"ubuntu:24.04","cmake_args":"","cuda_archs":""},{"variant":"cpu","arch":"arm64","runner":"ubuntu-24.04-arm","build_base":"ubuntu:24.04","runtime_base":"ubuntu:24.04","cmake_args":"","cuda_archs":""}'
          # CUDA: drop the libcuda driver-lib dependency (GGML_CUDA_NO_VMM) since
          # the build container has no GPU driver. amd64 takes ggml's default
          # (broad) arch list; arm64 only targets Grace GPUs (Hopper + GB10).
          CUDA='{"variant":"cuda","arch":"amd64","runner":"ubuntu-24.04","build_base":"nvidia/cuda:13.0.1-devel-ubuntu24.04","runtime_base":"nvidia/cuda:13.0.1-runtime-ubuntu24.04","cmake_args":"-DPARAKEET_GGML_CUDA=ON -DGGML_CUDA_NO_VMM=ON","cuda_archs":""},{"variant":"cuda","arch":"arm64","runner":"ubuntu-24.04-arm","build_base":"nvidia/cuda:13.0.1-devel-ubuntu24.04","runtime_base":"nvidia/cuda:13.0.1-runtime-ubuntu24.04","cmake_args":"-DPARAKEET_GGML_CUDA=ON -DGGML_CUDA_NO_VMM=ON","cuda_archs":"90;121-real"}'
          if [ "${{ github.event_name }}" = "pull_request" ]; then
            echo "matrix={\"include\":[${CPU}]}" >> "$GITHUB_OUTPUT"
          else
            echo "matrix={\"include\":[${CPU},${CUDA}]}" >> "$GITHUB_OUTPUT"
          fi

  # -------------------------------------------------------------------------
  # build: one job per (variant, arch). Builds natively on the matching runner
  # and pushes the image by digest (untagged). PRs build only (cache-only).
  # -------------------------------------------------------------------------
  build:
    needs: setup
    runs-on: ${{ matrix.runner }}
    permissions:
      contents: read
      packages: write
    strategy:
      fail-fast: false
      matrix: ${{ fromJSON(needs.setup.outputs.matrix) }}
    steps:
      - name: Checkout (with submodules)
        uses: actions/checkout@v4
        with:
          submodules: recursive

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3

      # Only authenticate when we actually push (i.e. not on pull_request).
      - name: Log in to ghcr.io
        if: github.event_name != 'pull_request'
        uses: docker/login-action@v3
        with:
          registry: ${{ env.REGISTRY }}
          username: ${{ github.actor }}
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Build and push by digest (${{ matrix.variant }}/${{ matrix.arch }})
        id: build
        uses: docker/build-push-action@v6
        with:
          context: .
          file: ./Dockerfile
          platforms: linux/${{ matrix.arch }}
          build-args: |
            BUILD_BASE=${{ matrix.build_base }}
            RUNTIME_BASE=${{ matrix.runtime_base }}
            CMAKE_EXTRA_ARGS=${{ matrix.cmake_args }}
            CUDA_ARCHS=${{ matrix.cuda_archs }}
          # PRs: build only (cache-only, nothing pushed). Otherwise push the
          # image by digest so the merge job can stitch the arches together.
          outputs: ${{ github.event_name != 'pull_request' && format('type=image,name={0}/{1},push-by-digest=true,name-canonical=true,push=true', env.REGISTRY, env.IMAGE_NAME) || 'type=cacheonly' }}
          cache-from: type=gha,scope=${{ matrix.variant }}-${{ matrix.arch }}
          cache-to: type=gha,mode=max,scope=${{ matrix.variant }}-${{ matrix.arch }}

      - name: Export digest
        if: github.event_name != 'pull_request'
        run: |
          mkdir -p /tmp/digests
          digest="${{ steps.build.outputs.digest }}"
          touch "/tmp/digests/${digest#sha256:}"

      - name: Upload digest
        if: github.event_name != 'pull_request'
        uses: actions/upload-artifact@v4
        with:
          name: digests-${{ matrix.variant }}-${{ matrix.arch }}
          path: /tmp/digests/*
          if-no-files-found: error
          retention-days: 1

  # -------------------------------------------------------------------------
  # merge: combine the per-arch digests of each variant into one multi-arch
  # manifest and tag it. Skipped on pull_request (nothing was pushed).
  # -------------------------------------------------------------------------
  merge:
    if: github.event_name != 'pull_request'
    needs: build
    runs-on: ubuntu-latest
    permissions:
      contents: read
      packages: write
    strategy:
      fail-fast: false
      matrix:
        include:
          - variant: cpu
            suffix: ""
          - variant: cuda
            suffix: "-cuda"
    steps:
      - name: Download digests
        uses: actions/download-artifact@v4
        with:
          path: /tmp/digests
          pattern: digests-${{ matrix.variant }}-*
          merge-multiple: true

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3

      - name: Log in to ghcr.io
        uses: docker/login-action@v3
        with:
          registry: ${{ env.REGISTRY }}
          username: ${{ github.actor }}
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Compute image tags
        id: meta
        uses: docker/metadata-action@v5
        with:
          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
          # cpu  -> latest, sha-xxxx, vX.Y.Z
          # cuda -> latest-cuda, sha-xxxx-cuda, vX.Y.Z-cuda
          flavor: |
            suffix=${{ matrix.suffix }},onlatest=true
          tags: |
            type=raw,value=latest,enable={{is_default_branch}}
            type=ref,event=tag
            type=sha

      - name: Create multi-arch manifest and push
        working-directory: /tmp/digests
        run: |
          docker buildx imagetools create \
            $(jq -cr '.tags | map("-t " + .) | join(" ")' <<< "$DOCKER_METADATA_OUTPUT_JSON") \
            $(printf '${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}@sha256:%s ' *)

      - name: Inspect manifest
        run: |
          docker buildx imagetools inspect \
            ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest${{ matrix.suffix }}