Hugging Face contributor tests report an H100 SDPA step falling from 2.06 to 1.02 seconds; on Trainium2, the revised code produced step timings where the previous path crashed.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Hugging Face, the open-source AI company co-founded by Clement Delangue, Julien Chaumond and Thomas Wolf, merged a TRL fix on November 13th, 2026 that halves a reported Qwen3 training step on an H100 and produced step timings on AWS Trainium2 where the previous layout crashed. The merged pull request keeps best-fit-decreasing packed examples as separate batch rows for standard attention models, instead of flattening each packed batch into one long row by default.…
source & further reading
runtimewire.com — original article
Hugging Face makes exported models portable across ONNX, OpenVINO, ExecuTorch and Dynamo
Vercel patches WorkflowAgent's ignored approval policy in AI SDK code
Claude Code on desktop now lets users build custom themes with mods