$ the-wire · showcase
timm fixes small-dataset sampling, TRL patches vLLM weight-sync crash
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
Two training-loop breakages got fixes today: timm's sampler on datasets below the selection rounding threshold, and TRL's server-mode weight sync against vLLM 0.20 through 0.25.
Pretrained source custom load huggingface/pytorch-image-models
rwightman reworked how pretrained weights are resolved so a custom source override actually takes effect, and the same commit fixes distillation teacher setup, which had been reading the wrong pretrained source. If you pass your own weights via a local path or alternate hub, this is the release that makes the override stick.
Fix sampler small dataset huggingface/pytorch-image-models
Ordered and repeated augmentation sampling, NaFlex sampling, and per-rank distributed sampling all used to break on datasets smaller than the selection rounding threshold; they now pad per-rank samples with modulo indexing. Sampler lengths are kept consistent with iteration, so a DataLoader built over a tiny set will no longer under- or over-run its epoch.
Fix the server-mode crash at the first weight sync on vLLM 0.20 to 0.25 huggingface/trl
vLLM's /reset_prefix_cache returns 200 with an empty body up to 0.25.0 and only started returning {"success": ...} in 0.26.0, so TRL's client JSON-decoded nothing and server-mode training died at the first weight sync. VLLMClient.reset_prefix_cache now accepts the empty body, which unbreaks vLLM 0.20.0 through 0.25.x.
Fix exporters import on torch < 2.9 (is_contiguous_or_false) (#49124) huggingface/transformers
is_contiguous_or_false only exists in torch._prims_common from 2.9 on, but a bare module-level import made every model test fail to collect on the FA CI image, which is still on torch 2.8. The import is now guarded with a fallback that calls a.is_contiguous() and returns False on data-dependent shape errors, matching upstream behavior.
Update Inference Providers documentation (automated) (#2818) huggingface/hub-docs
The Inference Providers docs were regenerated by HuggingFaceInfra, and rwightman moved timm's shared agent guidance into AGENTS.md with CLAUDE.md referencing it, documenting expectations for minimal diffs and backwards compatibility.
