TRL 1.15 fused LM head reaches six trainers

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

Diffusers moved to minor releases coordinated around model integrations and pushed 0.42.0 to Python 3.11, while transformers shipped Mistral Large 4.

1039 commits 922 PRs merged 21 releases 31 briefings covered

all huggingface reviews →

← previous month huggingface wire →

v1.15.0 huggingface/trl

by qgallouedec

SFT, DPO, KTO, GRPO, RLOO, and Distillation now score tokens through a single fused head, which the release notes put at "up to 6.9× longer sequences on the same GPU." Sequence length, not throughput, was the binding constraint on most long-context training runs.

v5.22.0: Mistral Large 4 (Le Chonk) huggingface/transformers

by LysandreJik

Mistral Large 4 (Le Chonk) is a 1 trillion-parameter natively multimodal model and the largest Mistral has shipped; it lands mid-period with 5.18's Nemotron 3 Diarization streaming speaker diarization on the other side. Both are loadable through the standard pipeline path without a release hop.

Diffusers 0.42.0: FLUX.3, Fibo 2.0, Kandinsky 6, and more huggingface/diffusers

by sayakpaul

FLUX.3 Image Klein, Bria Fibo 2.0, Kandinsky 6, SANA-WM, and RFD4-Proteína arrive alongside new adapter support and modular pipelines. The release also restarts the clock on your Docker images: "Python 3.11 or newer is now required." The release notes state Diffusers now coordinates minor releases around model integrations, the way Transformers does.

v1.17.0 huggingface/trl

by albertvillanova

Fine-tuning mistralai/Mistral-Large-4-1T-A52B is documented for a single node, which is the concrete outcome of the fused head plus the AWS Trainium work in 1.16.0. Budget accordingly before you assume this needs a cluster reservation.

[hf-xet v1.7.0]: Telemetry, Dynamic Download Buffers, and fixes huggingface/xet-core

by github-actions[bot]

hf-xet 1.7.0 adds client-side telemetry reports for each upload and download, documented under the Hub's xet telemetry page, and sizes download buffers dynamically from available RAM. Worth a look if your CI pulls large model repos, since the buffer sizing changes memory behaviour on small runners.

Deprecate opencode_env and pi_env in favour of harbor_env huggingface/OpenEnv

by sergiopaniego

OpenCode and Pi are both validated Harbor harnesses with the same token-level capture, so the per-harness environments were a second path to the same agents. Both are removed in OpenEnv 0.8.0; migrate to harbor_env before then.

[MFU] Use device peak flops instead of hardcoded H100's huggingface/trl

by dwarez

Async trainers computed MFU against hardcoded H100 peak flops, so reported efficiency on any other device was wrong by construction. A per-device peak-flops map now backs the calculation; if you compared MFU numbers on Trainium or TPU, those figures were not measuring what they claimed.

Add selective activation checkpointing (SAC) to SFT huggingface/trl

by kashif

SFT gains selective activation checkpointing: with gradient checkpointing on, the attention output is saved during forward rather than recomputed in backward, recovering most of the long-context slowdown for one extra hidden-state-sized tensor.

$ ls huggingface/month/ # the briefings behind this review

Thu Oct 1 Transformers 5.18 ships Nemotron 3 diarization, TRL drops vLLM 0.20.0 Fri Oct 2 transformers clears out deprecated mask, cache, and rotary paths Sat Oct 3 QLoRA keeps DoRA in float32, flash-attention kwargs stop leaking into vision encoders Sun Oct 4 AnyLanguageModel catches up to Foundation Models 26 and 27 Mon Oct 5 TRL widens template matching for LFM2, LFM2.5, and Gemma 4 Tue Oct 6 Diffusers 0.41.0 ships Qwen-Image 2.1, TRL drops Python 3.10 Wed Oct 7 Multi-eos completions in TRL, Kandinsky 6 lands in diffusers Thu Oct 8 TRL moves every trainer onto the fused LM head Fri Oct 9 TRL 1.15 lands a fused LM head, transformers finishes its RoPE cleanup Sat Oct 10 OpenRouter joins Inference Providers in huggingface.js Sun Oct 11 Repo2RLEnv parsers learn cargo-nextest and Django, Tasksmith ships to PyPI Mon Oct 12 serge PR triage tightens, TPU and multi-GPU test fixes land Tue Oct 13 Transformers drops legacy data helpers, TRL rebuilds GOLD tests on real trainers Wed Oct 14 AsyncGRPO packs rollouts as prefix trees, TRL types tool arguments by schema Thu Oct 15 TRL defaults padding to "auto" on Neuron, drops vLLM 0.21.0 Fri Oct 16 Transformers 5.20 ships Muse Spark Open, TRL adds AWS Trainium Sat Oct 17 MoE expert sharding gets a test, lerobot swaps vendor SDKs for rustypot Sun Oct 18 kernels stops reloading duplicate kernels, restores pre-2.10 Torch CPU/ROCm builds Mon Oct 19 Ernie4.5-VL-MoE experts shard under TP and EP Tue Oct 20 Tensor-parallel embeddings get their own styles, and LoRA silently skips MoE expert layers Wed Oct 21 TRL reads tokenizers for text-only models, chat-ui queues messages mid-run Thu Oct 22 Security bumps in kernels, a torchao import fix, and Grounding DINO heads that stay untied Fri Oct 23 Online DPO fixes Qwen-VL multimodal RoPE, and a CPU parity test for continuous batching Sat Oct 24 Diffusers 0.42.0 requires Python 3.11, lerobot cuts sidecar memory 91% Sun Oct 25 XPU flash-attention speed fix, FSDP2 dtype cast for async RL Mon Oct 26 Image processors composite tRNS transparency; kernel test harness moves to get_kernel Tue Oct 27 TRL fixes strided fused logprob backward, AsyncGRPO mitigates FP8 sampler collapse Wed Oct 28 TRL fixes chunked loss precision, drops liger-kernel from dev deps Thu Oct 29 AsyncGRPO drops dead rollout groups; PEFT fixes FSDP2 adapter sizing Fri Oct 30 Transformers 5.22.0 ships Mistral Large 4, TRL 1.17.0 fine-tunes it on one node Sat Oct 31 Transformers drops torch.jit.script and changes Mac serve attention; LeRobot moves to native EMA

all huggingface reviews →

← previous month huggingface wire →