SafeTensors serialization for PyTorch models

[kunigori]

Adds SafeTensors-based serialization for PyTorch models (addresses #2532) and
implements metadata-driven loading to integrate cleanly with the materializer
workflow (per @bcdurak's feedback).

Changes

• ✅ Add safetensors optional extra in pyproject.toml
• ✅ Save state_dict to .safetensors when available; fallback to .pt with warning
• ✅ Write minimal metadata.json (class_path, serialization_format)
• ✅ Use TemporaryDirectory + copy_dir() for remote stores
• ✅ load() always returns nn.Module
• ✅ Backward compat: supports weights.pt, checkpoint.pt, and legacy entire_model.pt

New artifact layout

artifact_uri/
├─ weights.safetensors   # or weights.pt on fallback
└─ metadata.json         # class_path + format

Metadata

{
  "class_path": "my_package.models.MyModel",
  "serialization_format": "safetensors",
  "init_args": [],
  "init_kwargs": {},
  "factory_path": null
}

Why SafeTensors?

• Security: Avoids pickle-based code execution risks
• Performance: Faster, memory-mapped weight loads
• Compatibility: Works with S3/GCS/Azure via artifact stores

Tests

Local run:

pytest tests/unit/integrations/pytorch/materializers/test_pytorch_module_materializer.py -v
# 4 passed in 1.88s

Coverage:

• Round-trip with safetensors
• Pickle fallback path
• Metadata-driven load
• Legacy formats (weights.pt, checkpoint.pt, entire_model.pt)
• Clear error when safetensors extra is missing at load

Known limitations (Phase 1)

• Zero-argument __init__() requirement: Models needing config should use
  a factory method (planned for Phase 2)

• Legacy artifacts without metadata (weights.pt / checkpoint.pt) require:

  model = materializer.load(data_type=MyModel)

• Legacy entire_model.pt is loaded and returned as a Module directly
  (no data_type needed)

Documentation

Happy to add a short guide covering why/how/limits/troubleshooting.
Which file should I update?

• docs/book/component-guide/materializers/pytorch.md (materializer behavior)?
• docs/book/integration-guide/pytorch.md (integration landing)?

Or would you prefer a new section?

Future work (separate PRs)

• Phase 2: Support init_args / init_kwargs / factory functions
• Phase 3: PyTorch Lightning materializer
• Phase 4: HuggingFace Transformers support

Checklist

• Tests pass locally
• Code formatted (ruff check --fix + ruff format)
• Also ran project scripts: bash scripts/format.sh and bash scripts/lint.sh
• Type hints added (mypy clean)
• Backward compatibility maintained
• Rebased on develop
• Documentation updated (pending guidance on location)
• CLA signed

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

yusuke kunimitsu seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[schustmi]

Hey @kunigori, thanks for the PR! Can you please base your changes on the develop branch and then also change the target of this PR.