elizaos--eliza/.github/workflows/build-llama-ffi-linux.yml

name: Build libelizainference (Linux, FFI)

# Builds the in-process bun:ffi shared library `libelizainference.so` for
# linux-x64-cpu and (opt-in) linux-x64-cuda. Used by the desktop FFI
# path on Linux servers and the Cloud-hosted local-inference fallback.
#
# STATUS: SCAFFOLD. The build steps mirror the omnivoice workflow but
# the target name `libelizainference.so` has not been wired through the
# llama.cpp CMakeLists install rules yet. The `run_real_build` input keeps
# the workflow out of the default fan-out until verified.

on:
  workflow_dispatch:
    inputs:
      run_real_build:
        description: "Run the scaffold build once it has been verified"
        required: false
        type: boolean
        default: false
      run_cuda:
        description: "Also build the Linux x64 CUDA artifact (self-hosted runner)"
        required: false
        type: boolean
        default: false
  push:
    tags:
      - "llama-ffi-*"

concurrency:
  group: build-llama-ffi-linux-${{ github.ref }}
  cancel-in-progress: true

permissions:
  contents: read

env:
  NODE_VERSION: "24"
  BUN_VERSION: "canary"

jobs:
  build-cpu:
    name: ubuntu-22.04 cpu
    if: ${{ inputs.run_real_build == true }}
    runs-on: ubuntu-22.04
    timeout-minutes: 90
    steps:
      - name: Checkout
        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
        with:
          submodules: false
          show-progress: false

      - name: Setup Node.js
        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e
        with:
          node-version: ${{ env.NODE_VERSION }}

      - name: Setup Bun
        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6
        with:
          bun-version: ${{ env.BUN_VERSION }}

      - name: Install Linux build deps
        run: |
          set -euo pipefail
          APT_ARGS=(-o Acquire::ForceIPv4=true -o Acquire::Retries=5 -o Acquire::http::Timeout=30)
          for attempt in 1 2 3; do
            if sudo apt-get "${APT_ARGS[@]}" update && \
               sudo apt-get "${APT_ARGS[@]}" install -y --fix-missing \
                 cmake ninja-build build-essential ccache git python3; then
              break
            fi
            if [ "$attempt" -eq 3 ]; then
              echo "apt install failed after ${attempt} attempts"
              exit 1
            fi
            sleep $((attempt * 10))
          done
          cmake --version

      - name: Cache llama.cpp cmake build dir
        uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9
        with:
          path: plugins/plugin-local-inference/native/llama.cpp/build
          key: llama-ffi-linux-cpu-${{ hashFiles('plugins/plugin-local-inference/native/llama.cpp/CMakeLists.txt') }}
          restore-keys: |
            llama-ffi-linux-cpu-

      - name: Configure llama.cpp (CPU, shared, no server)
        run: |
          set -euo pipefail
          cmake -S plugins/plugin-local-inference/native/llama.cpp \
                -B plugins/plugin-local-inference/native/llama.cpp/build \
                -G Ninja \
                -DCMAKE_BUILD_TYPE=Release \
                -DBUILD_SHARED_LIBS=ON \
                -DLLAMA_BUILD_EXAMPLES=OFF \
                -DLLAMA_BUILD_TOOLS=OFF \
                -DLLAMA_BUILD_TESTS=OFF \
                -DLLAMA_BUILD_SERVER=OFF \
                -DGGML_NATIVE=OFF \
                -DGGML_OPENMP=ON

      - name: Build llama.cpp shared library
        run: |
          set -euo pipefail
          cmake --build plugins/plugin-local-inference/native/llama.cpp/build --target llama -j 4

      - name: Verify artifact
        run: |
          set -euo pipefail
          ART_DIR="plugins/plugin-local-inference/native/llama.cpp/build"
          if [ -f "$ART_DIR/libllama.so" ]; then
            ART="$ART_DIR/libllama.so"
          elif [ -f "$ART_DIR/bin/libllama.so" ]; then
            ART="$ART_DIR/bin/libllama.so"
          else
            echo "no libllama.so found under $ART_DIR"
            ls -la "$ART_DIR"
            exit 1
          fi
          ls -lh "$ART"
          file "$ART"

      - name: Stage artifact (linux-x64-cpu)
        id: stage
        run: |
          set -euo pipefail
          STAGE_DIR="artifacts/libelizainference/linux-x64-cpu"
          mkdir -p "$STAGE_DIR"
          for cand in \
            plugins/plugin-local-inference/native/llama.cpp/build/libllama.so \
            plugins/plugin-local-inference/native/llama.cpp/build/bin/libllama.so; do
            if [ -f "$cand" ]; then
              cp "$cand" "$STAGE_DIR/libelizainference.so"
              break
            fi
          done
          {
            echo "target=linux-x64-cpu"
            echo "backend=cpu"
            echo "lib_name=libelizainference.so"
            echo "ref=${{ github.ref }}"
            echo "sha=${{ github.sha }}"
            echo "runner=ubuntu-22.04"
            echo "llama_sha=$(git -C plugins/plugin-local-inference/native/llama.cpp rev-parse HEAD || echo n/a)"
          } > "$STAGE_DIR/BUILD_INFO.txt"
          echo "stage_dir=$STAGE_DIR" >> "$GITHUB_OUTPUT"

      - name: Upload libelizainference artifact
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a
        with:
          name: libelizainference-linux-x64-cpu
          path: ${{ steps.stage.outputs.stage_dir }}/*
          if-no-files-found: error
          retention-days: 30

  build-cuda:
    name: ubuntu-22.04 cuda (self-hosted)
    if: ${{ inputs.run_real_build == true && inputs.run_cuda == true }}
    runs-on: [self-hosted, gpu-cuda-12.6]
    timeout-minutes: 120
    continue-on-error: true
    steps:
      - name: Checkout
        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
        with:
          submodules: false
          show-progress: false

      - name: Probe CUDA toolchain
        run: |
          set -euo pipefail
          nvcc --version
          nvidia-smi || true

      - name: Configure llama.cpp (CUDA, shared, no server)
        run: |
          set -euo pipefail
          cmake -S plugins/plugin-local-inference/native/llama.cpp \
                -B plugins/plugin-local-inference/native/llama.cpp/build \
                -G Ninja \
                -DCMAKE_BUILD_TYPE=Release \
                -DBUILD_SHARED_LIBS=ON \
                -DLLAMA_BUILD_EXAMPLES=OFF \
                -DLLAMA_BUILD_TOOLS=OFF \
                -DLLAMA_BUILD_TESTS=OFF \
                -DLLAMA_BUILD_SERVER=OFF \
                -DGGML_NATIVE=OFF \
                -DGGML_CUDA=ON

      - name: Build llama.cpp shared library
        run: |
          set -euo pipefail
          cmake --build plugins/plugin-local-inference/native/llama.cpp/build --target llama -j 4

      - name: Stage artifact (linux-x64-cuda)
        id: stage
        run: |
          set -euo pipefail
          STAGE_DIR="artifacts/libelizainference/linux-x64-cuda"
          mkdir -p "$STAGE_DIR"
          for cand in \
            plugins/plugin-local-inference/native/llama.cpp/build/libllama.so \
            plugins/plugin-local-inference/native/llama.cpp/build/bin/libllama.so; do
            if [ -f "$cand" ]; then
              cp "$cand" "$STAGE_DIR/libelizainference.so"
              break
            fi
          done
          echo "stage_dir=$STAGE_DIR" >> "$GITHUB_OUTPUT"

      - name: Upload libelizainference artifact
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a
        with:
          name: libelizainference-linux-x64-cuda
          path: ${{ steps.stage.outputs.stage_dir }}/*
          if-no-files-found: error
          retention-days: 30