139 wires and counting

$ follow Hugging Face

Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-14
stories 80

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

AsyncGRPO packs rollouts as prefix trees, TRL types tool arguments by schema

By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology

TRL's async training stack gets the day's two substantive changes: prefix-tree packing for repeated multi-turn rollouts, and schema-typed tool-call arguments now that transformers 5.15 can supply the tool definition to `parse_response`.

Tree attention huggingface/trl

by AmineDiro

AsyncGRPO now packs every row as a prefix tree instead of forwarding repeated prompt and completion prefixes once per row. There is no `packing` option: with nothing shared a tree row is exactly the old sequence packing at the same speed, so existing runs keep their behavior.

Pass the tools to `parse_response` so tool-call arguments are typed by the tool's JSON schema huggingface/trl

by qgallouedec

Qwen3.5, GLM-4-MoE and Nemotron 3 chat templates render a string argument as raw text, so the response templates fall back to a lenient JSON parse: a zip code comes back as an int and a bool as the string "True" on Qwen3.5 and Nemotron 3. Passing `tools=` through to `parse_response`, the path transformers 5.15 added, types arguments from the tool's JSON schema instead.

Store async GRPO and distillation rollout sample tokens as numpy arrays huggingface/trl

by AmineDiro

The rollout worker handed samples to the trainer as three Python lists per token, and on a deep backlog the worker's gen-2 GC walked every queued cell while holding the GIL, freezing the queue feeder thread. On a run with 50k-token samples and roughly 67k samples backlogged, the trainer blocked on `get()` for tens of seconds with a full queue.

Raise a clear error for prefix tuning on peft versions that cannot support it huggingface/trl

by albertvillanova

Prefix tuning on peft 0.17.1 and below died with `AttributeError: 'DynamicCache' object has no attribute 'key_cache'` inside `map_cache_to_layer_device_map`. TRL now raises a clear error up front when the peft version cannot support prefix tuning.

Test page: flat layout, one stats row, related issues box (#239) huggingface/transformers-ci

by Tarek Ziade

Tarek Ziade's test page rework flattens the layout to one stats row covering duration, memory leak and Serge Actions, drops the Run/Job/PR buttons in favor of the Grafana link bar, and moves related issues into a sticky card box. The GPU trend panels also gained model-added and major-dependency-change markers, so a failure wave can be lined up against the merge that caused it.

Quick answers

What shipped in Hugging Face on October 14, 2026?
TRL's async training stack gets the day's two substantive changes: prefix-tree packing for repeated multi-turn rollouts, and schema-typed tool-call arguments now that transformers 5.15 can supply the tool definition to `parse_response`. In total, 47 commits and 33 pull requests landed.
Who contributed to Hugging Face on October 14, 2026?
8 developers shipped this update, including Tarek Ziade, Yih-Dar, qgallouedec, AmineDiro, albertvillanova, dacorvo, Matheus-F-Scatolin, and Sergio Paniego Blanco.
What were the notable Hugging Face updates?
Tree attention, Pass the tools to `parse_response` so tool-call arguments are typed by the tool's JSON schema, and Store async GRPO and distillation rollout sample tokens as numpy arrays.