$ the-wire · showcase
TRL fixes chunked loss precision, drops liger-kernel from dev deps
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
![hf-security-analysis[bot]](https://avatars.githubusercontent.com/u/25720743?v=4)
TRL corrected gradient precision in its chunked log-probability and KL/JSD heads and finished detaching Liger kernels from its own tooling, while a new lint rule landed in transformers-mlinter.
Fix chunked loss precision huggingface/trl
The two precision issues from "A Fix for Softmax and Cross-Entropy" are fixed in the chunked log-probability and KL/JSD heads: confident targets keep their small gradients, and matching teachers now produce zero gradients instead of spurious ones. The issue is the same one Appendix B.4 of "Entropy-preserving reinforcement learning" reports for high-confidence policy gradients, and the separate ...
Drop liger-kernel from the dev dependencies, the Docker images and the tests huggingface/trl
`use_liger_kernel=True` is deprecated in every trainer and due for removal in v2.0.0, and TRL no longer ships fused Liger losses, so `liger-kernel` is gone from `dev` and both Docker images. The `liger` extra survives until v2.0.0, and the `trl-training` skill now points at `model_init_kwargs={"use_kernels": True}`.
TRF060: Check for absolute imports huggingface/transformers-mlinter
New rule TRF060 requires imports inside `transformers/models` to be relative, which will flag absolute intra-package imports in any new model code. A companion change prints IDE-compatible line numbers so console output jumps straight to the offending line, and the linter itself moved to Python 3.11.
Recognize the LFM2.5-VL-450M chat template for tool calling huggingface/trl
`add_response_schema` raised `Unrecognized chat template` and tool-calling `GRPOTrainer` failed at init on `LiquidAI/LFM2.5-VL-450M`, which now matches a stored template. Its format double-quotes string arguments (`city="Paris, France"`) where other LFM2.5 templates do not.
Recognize the LFM2.5-1.2B-Instruct chat template for tool calling huggingface/trl
The same recognition now covers `LiquidAI/LFM2.5-1.2B-Instruct`, `LFM2.5-350M`, `LFM2.5-1.2B-Thinking` and `LFM2-24B-A2B`, none of which matched a stored template at any Hub revision. Both changes are part of the same effort to widen tool-calling support across LiquidAI checkpoints.
Bump the actions group across 1 directory with 4 updates huggingface/optimum
optimum's CI actions group picked up four dependency updates, including `actions/setup-node` 7.0.0 to 7.1.0, alongside workflow reference refreshes from the security analysis bot.
Action items
