[jbohnslav]

Pretrained models now have six individual Hugging Face repositories grouped in a collection, with model cards, configuration summaries, licenses and checksum manifests. The README and setup guide document downloads for DEG-f, DEG-m and DEG-s. Model binaries stay outside this repository.

Test setup downloads the same published checkpoints at immutable commits and verifies their checksums, including when video fixtures already exist. CPU CI and the Modal GPU runner share this setup. New tests check artifact identity, strict parameter loading, finite forward outputs and model discovery for all six models. Video fixtures retain their existing source.

Validation:

• Full CPU suite: 36 passed.
• Modal GPU integration suite: 5 passed, covering all presets, feature extraction, sequence training and the softmax workflow.
• Ruff, pre-commit checks for all changed files, uv lock --check and git diff --check passed.

[github-actions]

Claude finished @jbohnslav's task in 20s —— View job

---

Review of PR #184

The automated /code-review skill errored, so this is a manual review of the code changes: setup_tests.py, modal_gpu_test.py, tests/test_pretrained_models.py, pyproject.toml and the workflow files. I did not run the tests. I also did not read the model cards or docs, or the contents of models.json, SHA256SUMS and the per-model model_config.json. The git diff origin/master...HEAD command also failed with "no merge base", so I read the commit directly.

I found no blocking bugs. The pinned revisions plus the checksum comparison against the checked-in manifests is a sound design. Setup now runs checkpoint preparation even when the video fixtures already exist, which fixes a real staleness gap.

Minor suggestions

1. Duplicated checksum logic. The checksum verification loop appears in both setup_tests.py:setup_pretrained_models and tests/test_pretrained_models.py::test_published_checkpoint. Extract it into one helper, for example in setup_tests.py, and import it in the test. The two copies could drift apart.
2. Confusing flow_generator flag. In the test, release["flow_generator"] is truthy for the DEG-f/m/s feature extractors, which produce 700-class logits. model_type is then set to "feature_extractor" when the flag is true. The behavior is consistent, but the name reads inverted. Consider renaming the field in models.json to something like uses_flow_generator, or add a comment.
3. Truncated downloads. snapshot_download(local_dir=...) leaves a .cache/huggingface folder inside each run directory. This is harmless for discovery, but if a download is interrupted, a partial checkpoint.pt fails only at the checksum step with a generic ValueError. Passing force_download on mismatch, or including the expected and actual digests in the message, would make failures easier to diagnose.
4. Exit status and error handling. setup_tests() catches every exception and returns False, and __main__ now exits non-zero, which is good. Printing the traceback, for example with traceback.print_exc(), would help CI debugging. At present only str(error) is shown.
5. Modal volume copy. ignore=lambda directory, names: ... in modal_gpu_test.py compares Path(directory) to VOLUME_ARCHIVE_PATH / "models". This works as long as copytree passes the exact source path, which it does. A short comment noting that the old volume snapshots are intentionally skipped would help.
6. Dev dependency. huggingface-hub>=0.24 is added to the dev group, and setup_tests.py imports it directly. This is fine, but anyone running it without --dev will hit an import error. The docstring already notes this.

The README and docs changes look consistent with the PR description. I did not verify the Hugging Face repositories or commit hashes against models.json.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-19T14:20:00.983158Z	5411587	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.