elizaos--eliza/.github/workflows/gpu-bench-nightly.yml

# Nightly per-GPU voice-bench smoke run.
#
# STATUS: stub. Gated by workflow_dispatch input until a self-hosted CUDA GPU
# runner is wired. L4 (the cheapest GitHub-hosted GPU runner) is a reasonable
# starting target; for the full matrix (3090 / 4090 / 5090 / H200) we
# need a self-hosted pool.
#
# Per docs/inference/gpu-tier.md the per-GPU autotune profiles live in
# packages/inference/configs/gpu/, and `voice-bench bench gpu` prints
# the resolved plan for the host card. Once a real cuda PipelineDriver
# is wired (see packages/inference/voice-bench/README.md "Wiring the
# real pipeline"), this workflow runs the actual benchmark.

name: gpu-bench-nightly

on:
  schedule:
    # Nightly at 06:00 UTC. Adjust once we know what window the
    # self-hosted runners are available.
    - cron: "0 6 * * *"
  workflow_dispatch:
    inputs:
      bundle:
        description: "Bundle id (e.g. eliza-1-9b)"
        required: false
        default: "eliza-1-9b"
      enable_gpu_bench:
        description: "Run the scaffold GPU bench jobs"
        required: false
        type: boolean
        default: false

# Default to least privilege. Override per-job where needed.
permissions:
  contents: read

jobs:
  detect:
    name: detect host GPU
    if: ${{ github.event_name == 'workflow_dispatch' && inputs.enable_gpu_bench == true }}
    runs-on: [self-hosted, gpu-cuda-12.6]
    steps:
      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
      - uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6
        with:
          bun-version: canary
      - name: install
        run: bun install --frozen-lockfile || bun install --no-frozen-lockfile
      - name: probe GPU + print autotune plan
        run: |
          bun run --cwd packages/inference/voice-bench bench gpu \
            --bundle "${{ github.event.inputs.bundle || 'eliza-1-9b' }}" \
            | tee gpu-plan.json
      - name: upload plan
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a
        with:
          name: gpu-plan
          path: gpu-plan.json
          if-no-files-found: error
          retention-days: 7

  bench:
    name: voice-bench (cuda)
    needs: detect
    if: ${{ github.event_name == 'workflow_dispatch' && inputs.enable_gpu_bench == true }}
    runs-on: [self-hosted, gpu-cuda-12.6]
    timeout-minutes: 60
    steps:
      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
      - uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6
        with:
          bun-version: canary
      - name: install
        run: bun install --frozen-lockfile || bun install --no-frozen-lockfile
      - name: run voice-bench
        run: |
          bun run --cwd packages/inference/voice-bench bench \
            --bundle "${{ github.event.inputs.bundle || 'eliza-1-9b' }}" \
            --backend cuda \
            --scenario all \
            --runs 5 \
            --output bench.json
      - name: upload bench result
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a
        with:
          name: bench-${{ github.run_id }}
          path: bench.json
          if-no-files-found: error
          retention-days: 7