$ cat huggingface/week/2026-10-12.log
the week in review · Oct 12 – Oct 18, 2026
Transformers 5.20 ships Muse Spark Open
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The same release pulls the deprecated GLUE/SQuAD/XNLI data helpers, while TRL v1.16.0 adds AWS Trainium training.
Release v5.20.0 huggingface/transformers
The release adds Muse Spark Open, a multimodal mixture-of-experts model from Meta that takes text and images. For anyone on 5.19, this is the release that also carries the removal of the legacy data helpers, so read item two before bumping.
Remove deprecated data helpers (#49360) huggingface/transformers
GlueDataset, SquadDataset, the GLUE/SQuAD/XNLI processors and metrics are gone from the library. The PR says they have had no users since the legacy example scripts were retired, so the breakage cost is presumably low, but any pinned import of those names now fails.
v1.16.0 huggingface/trl
TRL adds a second non-CUDA training path after the fused LM head from v1.15 made Neuron tractable; Trainium is the remaining target. Combined with the Neuron padding work this week, TRL is now tuning for accelerator-specific compile behavior rather than treating non-NVIDIA hardware as an edge case.
Pass the tools to `parse_response` so tool-call arguments are typed by the tool's JSON schema huggingface/trl
Tool-call arguments are now typed by the tool's JSON schema instead of a lenient JSON guess. On Qwen3.5, 3.6, 3.8, GLM-4-MoE and Nemotron 3, chat templates render `"75001"` and `75001` identically, so zip codes came back as ints and bools came back wrong; passing the tools into `parse_response` closes that.
Tree attention huggingface/trl
AsyncGRPO packs every rollout row as a prefix tree, with no `packing` option to set: a tree row with nothing shared is the old sequence packing at the same speed. Multi-turn and multi-generation rollouts repeat enough prefix to make the tree the only sane layout.
ITF: dedupe serge PRs by test, retire stale and superseded ones huggingface/transformers-ci
Nightly triage now dedupes serge PRs by test and retires superseded ones, because 19 of the 47 open serge PRs on the transformers repo on 2026-10-11 were stale or duplicates. If you watch that queue, expect it to shrink on its own instead of growing.
ITF: parse the CI dataset's "(line N)" traces huggingface/transformers-ci
The CI dataset's trace parser now handles `(line N)` traces, which every one of the 376 integration failures on 2026-10-10 used, so the terminal exception is actually extracted. That makes failure grouping and the related-issues box on the test page trustworthy rather than routed on partial text.
Default `pad_to_multiple_of` to `"auto"`: pad to a multiple of 128 on Neuron huggingface/trl
Alongside: TRL defaults `pad_to_multiple_of` to `"auto"`, padding to 128 on Neuron and `None` everywhere else, which avoids the roughly one-minute stall per new sequence length that Neuron's per-shape compile causes in SFT, DPO, KTO, Reward, GRPO, RLOO and Distillation.
$ ls huggingface/week/ # the briefings behind this review
Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
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.