omnidreams-worldlens-canary.yml

name: OmniDreams WorldLens Canary

on:
  push:
    branches:
      - main
      - "pull-request/[0-9]+"
    paths:
      - ".github/workflows/omnidreams-worldlens-canary.yml"
      - "pyproject.toml"
      - "uv.lock"
      - "flashdreams/pyproject.toml"
      - "integrations_v2/omnidreams/**"
      - "flashdreams/flashdreams/core/**"
      - "flashdreams/flashdreams/infra/**"
      - "flashdreams/flashdreams/recipes/taehv/**"
      - "flashdreams/flashdreams/recipes/wan/**"
  merge_group:
    branches: [main]
  workflow_dispatch:

permissions:
  contents: read

jobs:
  canary:
    name: canary3 WorldLens baseline
    runs-on: linux-amd64-gpu-rtxpro6000-latest-2
    timeout-minutes: 120
    continue-on-error: true
    defaults:
      run:
        shell: bash
    container:
      image: nvidia/cuda:13.2.1-cudnn-devel-ubuntu24.04
      options: --gpus all
    env:
      UV_PROJECT_ENVIRONMENT: /tmp/flashdreams-venv
      UV_LINK_MODE: copy
      # imgui-bundle has Linux ARM64 wheels for Python 3.11+, but not 3.10.
      UV_PYTHON: "3.11"
      UV_NO_BUILD_PACKAGE: imgui-bundle
      MAX_JOBS: 8
      RUN_ROOT: artifacts/omnidreams_worldlens_canary/run
      SCRATCH_ROOT: /tmp/omnidreams-worldlens-scratch
      SPLIT: od_26_01_canary3_tb4
      TOTAL_BLOCKS: "4"
      BASELINE: integrations_v2/omnidreams/impl/eval_baselines/od-26.01-canary3-tb4-v1.json
      CANARY_UUID_REGEX: '"uuid": "(01d503d4-449b-46fc-8d78-9085e70d3554|02e075b9-fd24-426b-971b-7cfcb2074cc9|0499fb41-122d-4180-83af-f954a9974d3b)"'
    steps:
      - name: Detect GPU architecture
        id: gpu-arch
        run: |
          nvidia-smi
          compute_cap=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | head -1 | tr -d '[:space:]')
          arch=$(echo "${compute_cap}" | tr -d '.')
          echo "arch=${arch}" >> "$GITHUB_OUTPUT"
          echo "Detected GPU compute capability: ${compute_cap} -> sm_${arch}"

      - name: Checkout
        uses: actions/checkout@v4

      - name: Install system dependencies
        run: |
          apt-get update -qq
          DEBIAN_FRONTEND=noninteractive apt-get install -y -qq --no-install-recommends \
            python3 python3-dev python3-venv \
            ffmpeg \
            gcc g++ ninja-build \
            libgl1-mesa-dev \
            libx11-dev libxcursor-dev libxext-dev libxi-dev libxinerama-dev libxrandr-dev \
            libnccl-dev \
            curl wget git ca-certificates unzip
          rm -rf /var/lib/apt/lists/*

      - name: Setup proxy cache
        uses: nv-gha-runners/setup-proxy-cache@main

      - name: Setup uv
        uses: astral-sh/setup-uv@v6
        with:
          enable-cache: true
          cache-suffix: "omnidreams-worldlens-canary-sm${{ steps.gpu-arch.outputs.arch }}"
          prune-cache: false

      - name: Install dependencies
        env:
          NVTE_CUDA_ARCHS: ${{ steps.gpu-arch.outputs.arch }}
        run: |
          uv venv --clear
          uv sync --locked --extra dev

      - name: Run OmniDreams WorldLens canary
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          set -euo pipefail

          mkdir -p "${RUN_ROOT}" "${SCRATCH_ROOT}"

          oe() {
            uv run --no-sync --package flashdreams-omnidreams omnidreams-eval "$@"
          }

          oe discover \
            --output "${RUN_ROOT}/manifest-full.jsonl"

          {
            head -n 1 "${RUN_ROOT}/manifest-full.jsonl"
            grep -E "${CANARY_UUID_REGEX}" "${RUN_ROOT}/manifest-full.jsonl" || true
          } > "${RUN_ROOT}/manifest.jsonl"

          line_count=$(wc -l < "${RUN_ROOT}/manifest.jsonl")
          if [ "${line_count}" -ne 4 ]; then
            echo "expected one manifest header plus 3 canary cases, got ${line_count} lines" >&2
            exit 1
          fi

          oe plan-batches \
            --manifest "${RUN_ROOT}/manifest.jsonl" \
            --output "${RUN_ROOT}/batches.json" \
            --batch-size 3

          oe stage-batch \
            --manifest "${RUN_ROOT}/manifest.jsonl" \
            --batch-plan "${RUN_ROOT}/batches.json" \
            --batch-id batch-00000 \
            --scratch-root "${SCRATCH_ROOT}" \
            --output "${RUN_ROOT}/staged/canary3.jsonl"

          oe generate \
            --staged-manifest "${RUN_ROOT}/staged/canary3.jsonl" \
            --run-root "${RUN_ROOT}" \
            --total-blocks "${TOTAL_BLOCKS}"

          oe validate-generation \
            --run-root "${RUN_ROOT}" \
            --output "${RUN_ROOT}/validation.json"

          oe setup-worldlens \
            --cache-dir "${RUN_ROOT}/evaluators" \
            --output "${RUN_ROOT}/worldlens-checkout.json"

          oe prepare-worldlens \
            --staged-manifest "${RUN_ROOT}/staged/canary3.jsonl" \
            --run-root "${RUN_ROOT}" \
            --worldlens-root "${RUN_ROOT}/evaluators/WorldLens" \
            --split "${SPLIT}" \
            --method-name omnidreams \
            --write-config \
            --force

          uv run --locked \
            --package flashdreams-omnidreams \
            --with hydra-core \
            --with omegaconf \
            --with openai-clip \
            --with loguru \
            --with decord \
            --with tqdm \
            omnidreams-eval worldlens-evaluate \
              --worldlens-root "${RUN_ROOT}/evaluators/WorldLens" \
              --split "${SPLIT}" \
              --method-name omnidreams \
              --python python

          oe summarize-run \
            --run-root "${RUN_ROOT}"

          oe check-baseline \
            --summary "${RUN_ROOT}/evaluation-summary.json" \
            --baseline "${BASELINE}" \
            --output-json "${RUN_ROOT}/baseline-check.json"

      - name: Trim uv cache for upload
        if: always()
        run: |
          cache_dir="${UV_CACHE_DIR:-/github/home/.cache/uv}"
          echo "=== Cache size before trim ==="
          du -sh "${cache_dir}" || true
          du -sh "${cache_dir}"/*/ 2>/dev/null || true

          # Remove pre-built wheels; the proxy cache can re-download them
          # faster than GitHub can upload and restore a multi-GB uv cache.
          rm -rf "${cache_dir}/wheels-v6"

          # Remove unzipped wheel archives; uv can re-extract these on demand.
          rm -rf "${cache_dir}/archive-v0"

          # Remove build artifacts from git checkouts that bloat the cache but
          # are not needed for reuse.
          find "${cache_dir}/git-v0/checkouts" \
            \( -name "build" -o -name "*.egg-info" -o -name "__pycache__" \) \
            -type d -exec rm -rf {} + 2>/dev/null || true

          # Remove editable installs; workspace packages rebuild quickly.
          rm -rf "${cache_dir}/sdists-v9/editable"

          echo ""
          echo "=== Cache size after trim ==="
          du -sh "${cache_dir}" || true
          du -sh "${cache_dir}"/*/ 2>/dev/null || true
          echo ""
          echo "=== Cached built wheels (sdists-v9) ==="
          find "${cache_dir}/sdists-v9" -name "*.whl" -exec ls -lh {} \; 2>/dev/null || true

      - name: Upload canary artifacts
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: omnidreams-worldlens-canary
          path: |
            ${{ env.RUN_ROOT }}/manifest.jsonl
            ${{ env.RUN_ROOT }}/batches.json
            ${{ env.RUN_ROOT }}/staged/canary3.jsonl
            ${{ env.RUN_ROOT }}/validation.json
            ${{ env.RUN_ROOT }}/evaluation-summary.json
            ${{ env.RUN_ROOT }}/evaluation-summary.md
            ${{ env.RUN_ROOT }}/baseline-check.json
            ${{ env.RUN_ROOT }}/generated/*/generated.mp4
            ${{ env.RUN_ROOT }}/generated/*/generation.json
            ${{ env.RUN_ROOT }}/generated/*/flashdreams-run.log
            ${{ env.RUN_ROOT }}/evaluators/WorldLens/cache/eval_logs/**
            ${{ env.RUN_ROOT }}/evaluators/WorldLens/generated_results/omnidreams/stage_manifest.json
            ${{ env.RUN_ROOT }}/evaluators/WorldLens/generated_results/omnidreams/*/*.json
          if-no-files-found: ignore