hf-dynacell

name

hf-dynacell

description

Develop, deploy, and maintain the DynaCell virtual-staining HuggingFace demo hosted at biohub/dynacell (ZeroGPU). Covers hosting in the biohub Dynacell resource group, the private-artifacts + HF_TOKEN access model, ZeroGPU configuration, and the deploy/smoke-test workflow. Use when the user asks to "push the dynacell space", "update the dynacell demo", "upload checkpoints", "deploy the hf space", mentions biohub/dynacell, or works in applications/dynacell/examples/hf_demo/.

DynaCell HuggingFace Demo — maintenance

Operational reference for the demo. Content + file layout live in applications/dynacell/examples/hf_demo/AGENT.md; this skill holds hosting, access, ZeroGPU, and deploy knowledge.

Hosting — biohub org, Dynacell resource group

Three repos under biohub, all in the Dynacell resource group (id 6a234bb4507cbbbb04456767):

Repo	Type	Visibility
biohub/dynacell	Space (Gradio, ZeroGPU)	private (→ public at launch)
biohub/dynacell-checkpoints	model	private
biohub/dynacell-demo-data	dataset	private

RG members who can push: edyoshikun, shalinmehta (admin); alxndrkalinin, dihan-zheng (write). Migrated from dihan-zheng/* (byte-for-byte mirror).

Moving an existing repo into a resource group needs org-admin. As org-write members, create repos directly in the group with create_repo(..., resource_group_id="6a234bb4507cbbbb04456767") — the upload scripts do this.

Access model — public demo, private artifacts

Decision (2026-06): checkpoints and demo data stay private even when the Space is public. The Space downloads them server-side via an HF_TOKEN secret (Space → Settings → Variables and secrets) — a fine-grained, read-only token scoped to exactly those two repos. Public visitors get only outputs (plots, PCC); raw artifacts stay gated. If the token is revoked or loses RG read access, runtime downloads break (rotate via "Replace").

ZeroGPU configuration (validated 2026-06)

• Hardware zero-a10g (RTX Pro 6000 Blackwell, 48 GB). Enterprise quota 60 GPU-min/day, shared.
• python_version: "3.12" required in the Space README — the default base image is 3.10, but VisCy needs >=3.12 (else BUILD_ERROR: viscy-data requires ... not in '>=3.12'). ZeroGPU supports Python 3.12.12 / 3.10.13, PyTorch 2.8→latest, Gradio SDK only.
• requirements.txt includes spaces; @spaces.GPU(duration=120) decorates run_prediction and compute_trajectory (predict_runner.py).
• A subprocess inside @spaces.GPU inherits the GPU, so the regression dynacell predict subprocess works on ZeroGPU. gr.Progress args work under the decorator. Keep import torch lazy (no CUDA at import).
• Tuning: CELL-Diff at 50–100 ODE steps can approach the 120 s ceiling on a cold call; raise duration or use a dynamic-duration callable if timeouts appear.

Code pointers

CHECKPOINT_REPO (predict_runner.py), _DEMO_REPO (app.py), and the README models:/datasets: fields all point at biohub/*.

Deploy

1. hf auth login (token with write to the Dynacell RG).
2. Edit hf_space/.
3. Push: uv run python applications/dynacell/examples/hf_demo/upload_hf_space.py.
4. Watch: HfApi().get_space_runtime("biohub/dynacell").stage. Build/run logs at GET /api/spaces/biohub/dynacell/logs/{build,run} are SSE streams — read with stream=True + a line cap; a plain GET hangs.

Re-publish checkpoints from HPC: edit the path map in upload_checkpoints.py and run it where /hpc/projects/... is mounted.

Smoke test

Drive endpoints with gradio_client (Client("biohub/dynacell", token=get_token()); token optional now the Space is public). /load_demo_data renders the data view. The run endpoints /run_regression and /run_generative read data from gr.State, which the client resets per call — they can't be driven headlessly, so exercise them in the live UI. Last validated GPU compute: regression VSCyto3D → Spectral PCC ~0.95 (~35 s); CELL-Diff trajectory (10 steps) ~74 s.

State

• Space biohub/dynacell is public (set 2026-06); checkpoints + dataset repos stay private, so the HF_TOKEN secret must remain.

Launch checklist

• Fill citation/DOI TODOs in cards/*.md and the uploaded READMEs.
• Re-pin requirements.txt VisCy refs off @dynacell-models once merged/renamed.
• Only if the artifact repos are ever made public: the HF_TOKEN secret can then be removed.