Pin kernels<0.13.0 in [cuda] extra to prevent huggingface-hub conflic…

[gps23]

Fixes #91

Root cause: kernels>=0.13.0 bumped its huggingface-hub requirement to >=1.10.0, which conflicts with transformers>=4.57.6,<5.0's requirement of huggingface-hub<1.0. When kernels is installed/upgraded independently (e.g. in a separate pip install step after transformers is already resolved), pip doesn't re-validate transformers' constraint, so you end up with a newer huggingface-hub installed alongside old transformers — which fails transformers' own version guard at import time.

Verification: Bisected the exact breaking version using Docker:

• kernels==0.12.3 + transformers==4.57.6 → imports fine
• kernels==0.13.0 + transformers==4.57.6 → ImportError: huggingface-hub>=0.34.0,<1.0 is required for a normal functioning of this module, but found huggingface-hub==1.24.0.

Note: a single-command pip install kernels transformers (no version drift) naturally resolves to a compatible kernels version, since pip's resolver considers both constraints together. The bug only surfaces when kernels is installed/upgraded in a separate step from transformers, which is training_hub's own install pattern (pip install training-hub[lora] then pip install training-hub[cuda]).

Fix: Capped kernels<0.13.0 in the [cuda] extra.

Summary by CodeRabbit

• Chores
  • Updated the CUDA installation dependency range for improved compatibility and stability.
  • CUDA setups now use kernels versions from 0.9.0 up to, but not including, 0.13.0.

[coderabbitai]

📝 Walkthrough

Walkthrough

The CUDA optional dependency changes its kernels requirement from an unbounded minimum to the range >=0.9.0,<0.13.0.

Changes

CUDA dependency constraints

Layer / File(s)	Summary
CUDA kernels version range pyproject.toml	The CUDA extra now constrains kernels to versions from 0.9.0 through 0.12.x.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Possibly related PRs

• Red-Hat-AI-Innovation-Team/training_hub#9: Also updates CUDA optional dependency constraints in pyproject.toml.

Suggested reviewers: maxusmusti

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the dependency pin in the cuda extra and the compatibility issue it addresses.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[gps23]

Hi @Maxusmusti, could you please take a look at this PR when you have a chance? It pins kernels to <0.13.0 in the CUDA extra to prevent the huggingface-hub incompatibility with transformers. All current checks are passing. Thank