$ the-wire · showcase
OpenEnv typed clients silently dropped every reward
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
A client-side field lookup bug meant chess_env training loops never saw a reward or an episode end, and Transformers dropped peft 0.19.1 outright because it cannot load adapter weights.
fix(envs): read reward and done from the step envelope in typed clients huggingface/OpenEnv
serialize_observation() excludes reward and done from the nested observation dict and puts them on the response envelope, but three typed clients read the observation dict and got the field defaults instead. chess_env was the worst case: StepResult.reward was always 0.0 and StepResult.done always False, so a training loop could run forever without ever observing a reward or an episode ending.
Bump `peft` version requirement (#48716) huggingface/transformers
peft>=0.19.1 used to be accepted, but 0.19.1 cannot correctly load adapter weights, so this commit drops support for peft<0.20.0 and updates every reference to the minimum. If you pin an older peft alongside Transformers, the requirement will now fail instead of loading a broken adapter silently.
Restore legacy tensor-parallel initialization for compatibility (#48797) huggingface/transformers
Legacy tensor-parallel and FSDP initialization paths are back, along with their public exports, with deprecation warnings on the old helpers. Code written against the previous initialization entry points keeps working while the helpers are phased out.
Publish to PyPI on tag push and guard VERSION against the tag huggingface/trl
TRL's publish workflow fired on any push to main that touched VERSION and only skipped the upload when the version string contained "dev", which is how v1.12.0 landed on PyPI as an accidental duplicate of v1.11.0 ninety seconds after 1.11.0. Publishing now triggers on tag push and rejects a post-release bump that is not a dev version, so an untagged version can no longer reach PyPI.
Bump dependencies huggingface/transformers.js
sharp goes to 0.35.4, @huggingface/tokenizers to 0.2.0, @huggingface/jinja to 0.5.10, and onnxruntime-node to 1.30.0, which raises the required Node.js version to >=20. Bumping sharp is what forces the runtime floor.
Retire three tiny model generators with no test consumer huggingface/trl
The long tail: TRL retired three tiny model generators with no test consumer, one of them never referenced under tests in any commit, after cross-referencing every push_to_hub call against the model ids in the suite.
Action items
- → Upgrade peft to 0.20.0 or newer for Transformers, since 0.19.1 cannot load adapter weights huggingface/transformers [plan]
- → Move to Node.js 20 or newer before upgrading transformers.js, required by sharp huggingface/transformers.js [plan]
