Chiron Maintainer

Chiron / Tabula maintainer skill

Use this skill when working on the Chiron platform internals: setting up a local Tabula backend, building or redeploying a BioEngine app (chiron-manager, chiron-orchestrator, tabula-trainer), or running a federated training session. For agent or end-user workflows on the public platform, use public/skills/chiron-platform/SKILL.md instead.

Tabula backend setup

All commands assume the tabula repo is checked out at ../tabula/ (sibling of this repo). If not present, clone first:

git clone https://github.com/aicell-lab/tabula ../tabula

conda create -n tabula python=3.11 -y && conda activate tabula
pip install torch==1.13.1+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
MAX_JOBS=4 pip install flash-attn==2.3.5 --no-build-isolation
pip install anndata==0.12.6
pip install "git+https://github.com/aicell-lab/bioengine.git@375dadf#egg=bioengine[datasets,worker]"
pip install -r ../tabula/requirements.txt && pip install -e ../tabula/

Running a local worker

# Dataset server (standalone)
python -m tabula.datasets --data-dir /path/to/data

# BioEngine Worker that auto-loads chiron-manager
python -m bioengine.worker \
  --mode single-machine \
  --head-num-cpus 3 --head-num-gpus 1 --head-memory-in-gb 30 \
  --startup-applications '{"artifact_id":"chiron-platform/chiron-manager","application_id":"chiron-manager"}'

Docker is the preferred local path. The .env file in ../tabula/ must define HYPHA_TOKEN, DATA_DIR, BIOENGINE_HOME, UID, GID. Important: unset HYPHA_TOKEN from the shell before running docker compose, otherwise the exported shell variable overrides the value in .env.

cd ../tabula/
unset HYPHA_TOKEN && docker compose up -d worker-tabula

# To restart with a refreshed token:
unset HYPHA_TOKEN && docker compose down worker-tabula && docker compose up -d worker-tabula

The current trainer image is ghcr.io/aicell-lab/tabula:0.6.0.

Uploading and deploying BioEngine apps

Always use the local BioEngine worker to upload apps. npx hypha-cli art cp bypasses the worker's upload pipeline and may not stage or commit correctly.

The token in ../tabula/.env must be valid for the chiron-platform workspace. Check expiry before running:

import base64, json, time
payload = HYPHA_TOKEN.split('.')[1] + '=='
data = json.loads(base64.urlsafe_b64decode(payload))
remaining = data['exp'] - time.time()
print(f"Token valid for {remaining/3600:.1f}h, expires {time.strftime('%Y-%m-%d %H:%M UTC', time.gmtime(data['exp']))}")

import asyncio
from hypha_rpc import connect_to_server

HYPHA_TOKEN = "<chiron-platform token from ../tabula/.env>"
APP_DIR     = "../tabula/apps/chiron_manager"  # or whichever app

async def main():
    server = await connect_to_server({'server_url': 'https://hypha.aicell.io', 'token': HYPHA_TOKEN})

    svcs = await server.list_services()
    worker_svc = next(s for s in svcs if 'bioengine-worker' in s['id'] and 'rtc' not in s['id'])
    worker = await server.get_service(worker_svc['id'])

    files = []
    for fname in ['manifest.yaml', 'manager.py']:  # adjust per app
        with open(f"{APP_DIR}/{fname}") as f:
            files.append({'name': fname, 'content': f.read(), 'type': 'text'})
    artifact_id = await worker.upload_app(files=files)
    print('Uploaded:', artifact_id)

    result = await worker.deploy_app(artifact_id=artifact_id, application_id='chiron-manager')
    print('Deployed:', result)

asyncio.run(main())

Notes:

• worker.upload_app(files=) uploads to the artifact store and returns the artifact id.
• worker.deploy_app(artifact_id=, application_id=) deploys or redeploys the app on Ray Serve. Reuse the same application_id to replace in place.
• Pass only manifest.yaml and the Python source files. Skip tutorial and docs files.
• The worker must be in the chiron-platform workspace. Verify with server.list_services().
• Do not pass _rkwargs=True to worker.upload_app or worker.deploy_app. The BioEngine worker's schema validator rejects it.

Federated training session

Once at least one orchestrator and one or more trainer workers are running, drive the session from the Chiron UI at https://chiron.aicell.io/#/training:

1. Create an Orchestrator application.
2. Create one or more Tabula Trainer applications.
3. Register trainers to the orchestrator.
4. Start federated training rounds.
5. Monitor progress and publish trained weights to the artifact hub.

Resource baseline per site:

The same flow is available via Hypha RPC. See public/skills/chiron-platform/apps/chiron-orchestrator.md for the underlying start_training contract.