build-embedded.yml

name: Build Embedded ARM64 Binaries

on:
  workflow_dispatch:
    inputs:
      version:
        description: "llama.cpp tag to build (e.g. b8502)"
        required: true
        default: "b8502"
      profiles:
        description: "Comma-separated profiles to build (or 'all')"
        required: true
        default: "all"

  push:
    tags:
      - "llama-embedded-*"

# ──────────────────────────────────────────────────────────────────────────────
# Strategy:
#
#  Cross-compiled on ubuntu-24.04-arm runners (no device needed):
#    rpi5-cpu, rpi5-vulkan, rpi4-cpu, rk3588-cpu, rk3588-vulkan,
#    a76-cpu, a76-vulkan, a72-cpu, armv8-cpu
#
#  Native build required (CUDA — must run on actual Jetson):
#    jetson-orin-cuda   → self-hosted runner labeled: [self-hosted, jetson-orin]
#    jetson-xavier-cuda → self-hosted runner labeled: [self-hosted, jetson-xavier]
#
# To register a Jetson as a self-hosted runner:
#   https://docs.github.com/en/actions/hosting-your-own-runners
# ──────────────────────────────────────────────────────────────────────────────

jobs:

  # ── Determine which profiles to build ──────────────────────────────────────
  setup:
    runs-on: ubuntu-latest
    outputs:
      version: ${{ steps.parse.outputs.version }}
      cpu_profiles: ${{ steps.parse.outputs.cpu_profiles }}
      cuda_profiles: ${{ steps.parse.outputs.cuda_profiles }}

    steps:
      - name: Parse inputs
        id: parse
        run: |
          VERSION="${{ github.event.inputs.version || 'b8502' }}"
          # Strip 'llama-embedded-' prefix if triggered by tag
          VERSION="${VERSION#llama-embedded-}"
          echo "version=${VERSION}" >> "$GITHUB_OUTPUT"

          PROFILES="${{ github.event.inputs.profiles || 'all' }}"

          ALL_CPU="rpi5-cpu rpi5-vulkan rpi4-cpu rk3588-cpu rk3588-vulkan a76-cpu a76-vulkan a72-cpu armv8-cpu"
          ALL_CUDA="jetson-orin-cuda jetson-xavier-cuda"

          if [ "$PROFILES" = "all" ]; then
            CPU_LIST="$ALL_CPU"
            CUDA_LIST="$ALL_CUDA"
          else
            CPU_LIST=""
            CUDA_LIST=""
            IFS=',' read -ra REQ <<< "$PROFILES"
            for p in "${REQ[@]}"; do
              p="${p// /}"
              if [[ " $ALL_CPU " == *" $p "* ]]; then
                CPU_LIST="$CPU_LIST $p"
              elif [[ " $ALL_CUDA " == *" $p "* ]]; then
                CUDA_LIST="$CUDA_LIST $p"
              fi
            done
          fi

          # Convert to JSON arrays for matrix
          CPU_JSON=$(echo "$CPU_LIST" | tr ' ' '\n' | grep -v '^$' | jq -Rc '[.]' | jq -sc 'add')
          CUDA_JSON=$(echo "$CUDA_LIST" | tr ' ' '\n' | grep -v '^$' | jq -Rc '[.]' | jq -sc 'add')

          echo "cpu_profiles=${CPU_JSON}" >> "$GITHUB_OUTPUT"
          echo "cuda_profiles=${CUDA_JSON}" >> "$GITHUB_OUTPUT"

          echo "Building CPU profiles: $CPU_LIST"
          echo "Building CUDA profiles: $CUDA_LIST"

  # ── CPU / Vulkan builds (cross-compiled on GitHub-hosted ARM runners) ───────
  build-cpu:
    needs: setup
    if: ${{ needs.setup.outputs.cpu_profiles != '[]' }}
    runs-on: ubuntu-24.04-arm   # GitHub-hosted ARM64 runner

    strategy:
      fail-fast: false
      matrix:
        profile: ${{ fromJson(needs.setup.outputs.cpu_profiles) }}

    steps:
      - name: Checkout build scripts
        uses: actions/checkout@v4

      - name: Install dependencies
        run: |
          sudo apt-get update -qq
          sudo apt-get install -y --no-install-recommends \
            build-essential cmake git \
            libvulkan-dev vulkan-tools glslc \
            pkg-config

      - name: Build ${{ matrix.profile }}
        run: |
          chmod +x ./scripts/build-embedded.sh
          OUTPUT_DIR=./dist ./scripts/build-embedded.sh \
            "${{ needs.setup.outputs.version }}" \
            "${{ matrix.profile }}"

      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
          name: llama-server-${{ needs.setup.outputs.version }}-${{ matrix.profile }}
          path: dist/*.tar.gz
          retention-days: 30

  # ── CUDA builds (native Jetson — requires self-hosted runners) ─────────────
  build-jetson-orin:
    needs: setup
    if: ${{ contains(fromJson(needs.setup.outputs.cuda_profiles), 'jetson-orin-cuda') }}
    # Self-hosted Jetson Orin runner with JetPack 6 / CUDA 12
    runs-on: [self-hosted, jetson-orin]

    steps:
      - name: Checkout build scripts
        uses: actions/checkout@v4

      - name: Install cmake (if needed)
        run: |
          cmake --version || sudo apt-get install -y cmake

      - name: Build jetson-orin-cuda
        run: |
          chmod +x ./scripts/build-embedded.sh
          OUTPUT_DIR=./dist ./scripts/build-embedded.sh \
            "${{ needs.setup.outputs.version }}" \
            jetson-orin-cuda

      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
          name: llama-server-${{ needs.setup.outputs.version }}-jetson-orin-cuda
          path: dist/*.tar.gz
          retention-days: 30

  build-jetson-xavier:
    needs: setup
    if: ${{ contains(fromJson(needs.setup.outputs.cuda_profiles), 'jetson-xavier-cuda') }}
    # Self-hosted Jetson Xavier runner with JetPack 5 / CUDA 11
    runs-on: [self-hosted, jetson-xavier]

    steps:
      - name: Checkout build scripts
        uses: actions/checkout@v4

      - name: Build jetson-xavier-cuda
        run: |
          chmod +x ./scripts/build-embedded.sh
          OUTPUT_DIR=./dist ./scripts/build-embedded.sh \
            "${{ needs.setup.outputs.version }}" \
            jetson-xavier-cuda

      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
          name: llama-server-${{ needs.setup.outputs.version }}-jetson-xavier-cuda
          path: dist/*.tar.gz
          retention-days: 30

  # ── Collect all artifacts and create a release ──────────────────────────────
  release:
    needs: [setup, build-cpu, build-jetson-orin, build-jetson-xavier]
    if: always() && (needs.build-cpu.result == 'success' || needs.build-jetson-orin.result == 'success' || needs.build-jetson-xavier.result == 'success')
    runs-on: ubuntu-latest

    steps:
      - name: Download all artifacts
        uses: actions/download-artifact@v4
        with:
          path: all-artifacts
          merge-multiple: true

      - name: List built artifacts
        run: |
          echo "Built artifacts:"
          find all-artifacts -name "*.tar.gz" | sort

      - name: Create GitHub Release (on tag push)
        if: startsWith(github.ref, 'refs/tags/')
        uses: softprops/action-gh-release@v2
        with:
          files: all-artifacts/**/*.tar.gz
          tag_name: ${{ github.ref_name }}
          name: "llama-server embedded ${{ needs.setup.outputs.version }}"
          body: |
            ## Embedded ARM64 llama-server builds

            Based on llama.cpp **${{ needs.setup.outputs.version }}**

            All builds use device-specific `-mcpu` flags to avoid SVE instructions
            (`cntb`, etc.) that cause SIGILL on Cortex-A72/A76/A78AE CPUs.

            | Artifact | Device | Accel | CPU Flags |
            |---|---|---|---|
            | `linux-arm64-jetson-orin-cuda-12` | Jetson Orin Nano/NX/AGX | CUDA 12 | `-mcpu=cortex-a78ae` |
            | `linux-arm64-jetson-xavier-cuda-11` | Jetson Xavier NX/AGX | CUDA 11 | `-mcpu=carmel` |
            | `linux-arm64-rpi5-vulkan` | Raspberry Pi 5 | Vulkan | `-mcpu=cortex-a76` |
            | `linux-arm64-rpi5-cpu` | Raspberry Pi 5 | CPU | `-mcpu=cortex-a76` |
            | `linux-arm64-rpi4-cpu` | Raspberry Pi 4 | CPU | `-mcpu=cortex-a72` |
            | `linux-arm64-rk3588-vulkan` | Rockchip RK3588 | Vulkan | `-mcpu=cortex-a76` |
            | `linux-arm64-rk3588-cpu` | Rockchare RK3588 | CPU | `-mcpu=cortex-a76` |
            | `linux-arm64-a76-vulkan` | Generic A76 (OPi5, Rock 5B) | Vulkan | `-mcpu=cortex-a76` |
            | `linux-arm64-a76-cpu` | Generic Cortex-A76 | CPU | `-mcpu=cortex-a76` |
            | `linux-arm64-a72-cpu` | Generic Cortex-A72 | CPU | `-mcpu=cortex-a72` |
            | `linux-arm64-armv8-cpu` | Generic ARMv8-A | CPU | `-march=armv8-a` |