name: Build & Publish AMD Strix Halo Toolboxes

on:
  workflow_dispatch:
    inputs:
      backends:
        description: >
          Comma-separated backends to build (e.g. "rocm-7beta,rocm-7rc").
          Use "all" to build everything.
        required: false
        default: all

env:
  DOCKERHUB_REPO: gitea.wefers.page/julian/amd-strix-halo-toolboxes
  LOCAL_PREFIX: llama

jobs:
  # 1) Prepare a clean JSON array for the matrix
  prepare:
    runs-on: ubuntu-latest
    outputs:
      matrix_json: ${{ steps.mk.outputs.matrix_json }}
    steps:
      - id: mk
        shell: bash
        run: |
          # Input from the Run workflow form
          IN='${{ inputs.backends }}'

          if [[ "$IN" == "all" || -z "$IN" ]]; then
            JSON='["rocm-6.4.4","rocm-7.2.3","rocm7-nightlies","vulkan-amdvlk","vulkan-radv"]'
          else
            # Remove spaces and build JSON array from comma list
            IN_CLEAN=$(echo "$IN" | tr -d '[:space:]')
            JSON='["'${IN_CLEAN//,/\",\"}'"]'
          fi

          echo "matrix_json=${JSON}" >> "$GITHUB_OUTPUT"
          echo "Using matrix: ${JSON}"

  # 2) Build each backend in parallel using the prepared matrix
  build-and-push:
    needs: prepare
    runs-on: ubuntu-latest
    strategy:
      fail-fast: false
      matrix:
        backend: ${{ fromJson(needs.prepare.outputs.matrix_json) }}

    steps:
      - name: Free up runner disk space
        run: |
          echo "Before cleanup:" && df -h /
          sudo rm -rf \
            /usr/share/dotnet \
            /usr/local/lib/android \
            /opt/ghc \
            /opt/hostedtoolcache/CodeQL
          docker system prune --all --force
          docker builder prune --all --force
          echo "After cleanup:" && df -h /

      - name: Check out repository
        uses: actions/checkout@v3

      - name: Cache podman storage for ${{ matrix.backend }}
        uses: actions/cache@v5
        with:
          key: podman-storage-${{ matrix.backend }}
          restore-keys: |
            podman-storage-${{ matrix.backend }}
            podman-storage
          path: ~/.local/share/containers/storage

      - name: Log in to Docker Hub
        run: |
          podman login -u ${{ secrets.DOCKERHUB_USERNAME}} -p ${{ secrets.DOCKERHUB_TOKEN }}

      - name: Set build timestamp
        run: echo "BUILD_TS=$(date +%Y%m%dT%H%M%S)" >> $GITHUB_ENV

      - name: Build & push ${{ matrix.backend }}
        working-directory: toolboxes
        shell: bash
        run: |
          set -euo pipefail
          B="${{ matrix.backend }}"
          DF="Dockerfile.$B"
          NAME="${B}"
          LI="${LOCAL_PREFIX}-${NAME}"
          TAG="${NAME}_${BUILD_TS}"
          IMM="${DOCKERHUB_REPO}:${TAG}"
          CHN="${DOCKERHUB_REPO}:${NAME}"

          echo "→ Building ${DF}"

          # we use buildah to eventually make use of pushing with
          # zstd:chunked compression, which is much more efficient
          # than dockers gzip format.
          # --pull: ensure we use the latest version of the base image
          # --squash: flatten the final result image into one single layer.
          #           Avoids large image sizes due to intermediate files
          #            that are irrelevant for the user
          # --format oci: use the OCI image format, which allows for pushing with zstd:chunked
          # --no-cache: Recompute every step in the dockerfile, even if the previous layer
          #             has not ben invalidated. Needed since we pull from ze internet.
          # --cache-(to|from): pull/push the intermedia cache layers resulting from
          #                    --mount options in the Dockerfile
          # NOTE: we are mounting cache layers for dnf and pushing them. This cache
          #       layer is shared amongst all Dockerfiles, since they have the identical
          #       mount parameter. When parallel building with buildah, those cache layers
          #       compete. In parallel, they all pull the latest fitting cache, then maybe
          #       add some packages relevant to their specific variant, then afterwards push
          #       the cache again. When multiple buildahs push the dnf cache, they could invalidate
          #       the just-pushed cache of another builder instance, so some packages might
          #       always be missing. SOLUTION: we give each containers dnf cache an individual
          #       id, thus cache per variant.
          buildah bud \
            --pull \
            --squash \
            --format oci \
            --no-cache \
            -t "${LI}" \
            -f "${DF}" \
            .

          echo "→ Running smoke test..."
          podman run --rm "${LI}" llama version
          podman run --rm "${LI}" llama-cli --help || { status=$?; echo "llama-cli exited with status $status"; [[ $status -eq 0 || $status -eq 1 || $status -eq 134 ]]; }
          podman run --rm "${LI}" llama-server --help || { status=$?; echo "llama-server exited with status $status"; [[ $status -eq 0 || $status -eq 1 || $status -eq 134 ]]; }

          # push with zstd:chunked compression, see https://github.com/containers/storage/blob/main/docs/containers-storage-zstd-chunked.md

          echo "→ Tag & push immutable → ${IMM}"
          buildah tag "${LI}" "${IMM}"
          buildah push --compression-format zstd:chunked "${IMM}"

          echo "→ Tag & push channel → ${CHN}"
          buildah tag "${IMM}" "${CHN}"
          buildah push --compression-format zstd:chunked "${CHN}"