# Deployment and final-run promotion

The live Space is https://huggingface.co/spaces/User-2468/mini-unet-colorizer.
Default model revision remains `1a9eb8af2754ad2329a24cfe50d388cb559441d0` (v3.0). The final longer job completed at https://huggingface.co/jobs/User-2468/6abaef6552d0dbd7f1da7286. Its selected checkpoint was rejected for production; see reports/FINAL_CHECKPOINT_REVIEW.md.

## ZeroGPU
Use `app.py`, `inference.py`, `semantic_model.py`, `decision_model.py`, and `requirements-space.txt` (renamed `requirements.txt` in the Space). Space README SDK is Gradio6.28.0; requirements use the same version. Select ZeroGPU hardware. Import `spaces` before `torch`; load/place the model on CUDA at startup. Only chroma inference runs inside `@spaces.GPU(duration=15)`. CPU handles image preparation, guided rendering, and cached strength changes. Do not enable `torch.compile` in ZeroGPU.

For local CPU validation set `ZEROGPU_CPU_TEST=1`; `MODEL_ID` can be a local checkpoint directory. Image dimensions are preserved, maximum12MP, with PNG alpha. Cache is session-scoped with a10-minute lifetime. Upload changes invalidate cached predictions. Standard/Detailed/Maximum correspond to maximum network dimensions256/384/512. Strength changes after inference do not use another GPU request.

## Final checkpoint review
1. Inspect `experiments/final-20260929/status.json`, `candidates.json`, `summary.json`, final test/paired diagnostics and image grids. A candidate with step0 is the retained pilot, not a failed upload.
2. Compare actual images to v3 and the pilot, especially historical originals and colour retention. Do not promote solely on patch score or original-colour error.
3. Verify strict parameter count, output dimensions/alpha/lightness, CPU/CUDA loading and all palette modes. Use `decision_model.load_decision` for a DecisionColorizer checkpoint.
4. Pin Space `MODEL_REVISION` to the evaluated immutable commit and `MODEL_SUBFOLDER` to the chosen directory. Alternative controls become visible for four-palette models.
5. For a new root release, replace weights/config/loader together, update card and hashes, regenerate and verify ONNX or archive the old graph with an explicit version. The current v3 ONNX must never be described as the new four-palette model. Rebuild the release ZIP, tag the release, and run a real ZeroGPU request.

Training outputs are saved on main, but production promotion requires this review. Preserve the user's stable branch.