Hugging Face AI Updates: October 10, 2026

1. TRL v1.15.0 Adds a Fused LM Head for Up to 6.9x Longer Training Sequences

Hugging Face. TRL v1.15.0 makes SFT, DPO, KTO, GRPO, RLOO, and distillation trainers score tokens with a fused LM head, a Triton kernel that computes per-token log-probs and entropy in tiles without materializing the full logits tensor. On a gemma-3-1b benchmark, maximum trainable sequence length grew by up to 6.89x for KTO and 5.8x for DPO, and peak memory at 8,192 tokens dropped 52% to 82% with slightly faster steps. The change removes the full-logits and use_liger_kernel chunked paths from GRPO, RLOO, DPO, and KTO, so PEFT adapters on lm_head now raise an error and use_liger_kernel=True is deprecated ahead of v2.0.0. Source

2. Ai2 Details Fair-Share GPU Scheduling on the Hugging Face Blog

Hugging Face. The Allen Institute for AI (Ai2) published a post on the Hugging Face blog describing how its infrastructure team replaced priority-based GPU scheduling with GPU time budgets, hierarchical fair-share allocation over a 7-day lookback window, and time-slicing with an 8-hour cap on protected runtime. Over a 30-day test, teams received 98% of the GPU hours they were owed while cluster occupancy held at 98%, p90 queue wait for debug workloads fell from 2 hours to 30 seconds, and repairs needing human intervention dropped 74%. Source

3. huggingface_hub v2.2.0 Makes hf jobs stats Return Snapshots by Default

Hugging Face. The huggingface_hub v2.2.0 release changes hf jobs stats to print a single snapshot of job metrics and exit, with a new -f/--follow flag for the previous live view and --format json|agent|quiet output options. The Python fetch_job_metrics function gains a matching follow parameter, and the release fixes hanging log commands, ensures concurrent downloads reliably use hf_xet, and makes HfFileSystem caching more predictable. Source