Reigh-Worker

GPU worker for Reigh, for running locally or in the cloud via Reigh Worker Orchestrator — processes video generation tasks using Wan2GP.

Quick Start

# 1. Linux baseline: Ubuntu 22.04/24.04 with NVIDIA drivers, ffmpeg, git, curl
#    Ubuntu 24.04+ needs Python 3.10 from deadsnakes before the command below works:
#    sudo add-apt-repository ppa:deadsnakes/ppa && sudo apt-get update

# 2. Bootstrap uv once
curl -LsSf https://astral.sh/uv/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"

# 3. Sync the locked environment from the repo root
uv sync --locked --python 3.10 --extra cuda124

# 4. Run the worker
uv run --python 3.10 python run_worker.py \
  --reigh-access-token "your-worker-token" \
  --wgp-profile 4 \
  --idle-release-minutes 15

Get credentials from reigh.art.

The Reigh app now generates two command tabs:

• Install: bootstrap or force-resync at the configured Worker repo location
• Run: the normal day-to-day launch path, which still runs uv sync before starting

Both commands always cd into the configured repo path first so copy-pasting from a fresh terminal in your home directory still works.

Packaging Notes

• pyproject.toml is the canonical dependency definition for the worker runtime.
• requirements.txt remains committed as rollback ballast during the rollout.
• uv.lock is expected to be refreshed from the worker repo root with Python 3.10.
• Wan2GP/requirements.txt now comes from the pinned Wan2GP/ submodule. If upstream requirements change, bump the submodule pointer, mirror the required dependency changes into the root project metadata, and regenerate uv.lock.

Rollback

• There is no runtime pip fallback on the migrated branch. Failed uv sync runs should fail loudly.
• If the very first uv-based launch fails after migrating an existing machine, restore the newest venv.pre-uv-* or .venv.pre-uv-* backup back to venv/ or .venv/, remove .uv-migrated, and investigate from there.
• If the release itself must be rolled back, revert the uv rollout commits and return to the pre-uv revision that still bootstraps from the committed requirements.txt.

Standalone Usage

Use the generation engine without Reigh for local testing or custom pipelines:

# Join two video clips with AI-generated transition
python examples/join_clips_example.py \
    --clip1 scene1.mp4 --clip2 scene2.mp4 \
    --output transition.mp4 --prompt "smooth camera glide"

# Regenerate corrupted frames
python examples/inpaint_frames_example.py \
    --video my_video.mp4 --start-frame 45 --end-frame 61 \
    --output fixed.mp4 --prompt "smooth motion"

Using HeadlessTaskQueue Directly

from headless_model_management import HeadlessTaskQueue, GenerationTask
from pathlib import Path

queue = HeadlessTaskQueue(wan_dir=str(Path(__file__).parent / "Wan2GP"), max_workers=1)
queue.start()

task = GenerationTask(
    id="my_task",
    model="wan_2_2_vace_lightning_baseline_2_2_2",
    prompt="a cat walking through a garden",
    parameters={"video_length": 81, "resolution": "896x512", "seed": 42}
)

queue.submit_task(task)
result = queue.wait_for_completion(task.id, timeout=600)
print(f"Output: {result.get('output_path')}" if result.get("success") else f"Error: {result.get('error')}")

queue.stop()

Debugging

uv run --python 3.10 python -m debug task <task_id>          # Investigate a task
uv run --python 3.10 python -m debug tasks --status Failed   # List recent failures

VibeComfy Live Validation

The normal development fallback is still a fresh live-test pod:

python -m scripts.live_test.main --variant fresh --backend vibecomfy --case z_image_turbo

Use the prebuilt path when a runpod-lifecycle prebuilt volume exists for the target RunPod data center. It reuses the baked worker/VibeComfy environments, then validates the actual consumer pod before worker registration, launch, or matrix queueing.

# Cheap target-manifest handoff with no DB, RunPod, pod launch, or VibeComfy import.
python -m scripts.live_test.main --variant fresh --backend vibecomfy \
  --case z_image_turbo \
  --emit-targets-json /tmp/reigh-targets.json

# Preferred consumer smoke after the prebuilt health gate is green.
python -m scripts.live_test.main --variant prebuilt --backend vibecomfy \
  --case z_image_turbo \
  --prebuilt-volume-name reigh-livetest-prebuilt-portable-<data-center>

--variant auto is opt-in. It checks for a prebuilt volume by the configured prefix and falls back to fresh when no matching volume is available. The default variant remains fresh.

Prebuilt report evidence is written under scripts/live_test/runs/<timestamp>/:

• targets.json: selected Reigh cases/routes/templates only, including explicit target/template counts;
• targets.enriched.json: VibeComfy-owned source/schema/model-asset metadata;
• env.health.json: grouped environment/custom-node/source/schema/asset checks;
• report.json and report.md: matrix result plus prebuilt metadata.

The report.json metadata block records the volume id/name, data center, manifest/health/target/enriched paths and hashes, selected profile/templates, GPU type, and health-check summary. Fresh and update reports keep the old shape unless their callers pass metadata explicitly.

When adding or onboarding a VibeComfy workflow for app-active use:

1. Emit the selected Reigh target JSON with --emit-targets-json.
2. Enrich it in VibeComfy with python -m vibecomfy.cli workflows enrich-targets.
3. Run rl prebuilt check --enriched-targets-json ... for the portable RTX 4090 profile.
4. Reconcile explicit selected assets if the check reports missing models.
5. Run the cheap z_image_turbo smoke first, then broaden the matrix only after health and smoke pass.
6. Record the run directory and report evidence in the contract or migration notes.

Code Health

Project Structure

See STRUCTURE.md for detailed project layout.

Powered By

Wan2GP by deepbeepmeep — the Wan2GP/ directory is a git submodule pinned to banodoco/Wan2GP.