RepoJournal

$ cat huggingface/week/2026-08-31.log

Hugging Face

Hugging Face

the week in review · Aug 31 – Sep 6, 2026

Transformers deprecations cleaned, NeoMME and offload land

♥

TRL removes PPOTrainer, and Transformers drops deprecated code while adding NeoMME and gradient checkpointing offload.

142 commits 142 PRs merged 3 releases 7 briefings covered
← previous week huggingface wire →

What shipped: Transformers landed NeoMME and NeoMME-Retriever [1], added offload to gradient checkpointing [2], and dropped deprecated code [3]. A RoPE encoder fix removes the position-indexed token type lookup [4], and a new capture_outputs option lets models keep only necessary hidden states [5]. Diffusers added an LTX-2.5 DFR pipeline [6], while deprecating ONNX [7]. TRL removed PPOTrainer [3] and vendored Liger-Kernel losses as part of that cleanup.

Why it matters: The deprecated code removal in Transformers [3] and PPOTrainer removal in TRL [3] are breaking changes already in effect; check your code against them. Diffusers' ONNX deprecation [7] sets a timeline for pipeline migration. The offload addition [2] directly reduces memory pressure for long-context training, and the RoPE fix [4] corrects a subtle bug that could affect generation quality in models with certain configs.

Also: Integration tests now pin attention layer types in TRL [8], kernels adds minver checks and capability validation [9][10], a flash-attn2 backward pass dangling pointer fix [11], and a Jinja engine alignment with Jinja2 was noted this week.

References

  1. [1] Add NeoMME and NeoMME-Retriever (#47992) ↗ huggingface/transformers
  2. [2] Add offload to gradient checkpointing ↗ huggingface/transformers
  3. [3] Deprecated stuff gone (#48367) ↗ huggingface/transformers
  4. [4] fix(models): Drop the position-indexed token type lookup in RoPE encoders (#48407) ↗ huggingface/transformers
  5. [5] Allow capturing only necessary hidden_states with capture_outputs (#48081) ↗ huggingface/transformers
  6. [6] Add LTX-2.5 DFR pipeline (keyframe slots, spatial detailing, tiled temporal rounds) ↗ huggingface/diffusers
  7. [7] [core] deprecate onnx ↗ huggingface/diffusers
  8. [8] Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types ↗ huggingface/trl
  9. [9] feat: support and check minver ↗ huggingface/kernels
  10. [10] feat: implement capability check in main ↗ huggingface/kernels
  11. [11] flash-attn2: fix dangling pointers in bwd when `num_kv_heads != num_heads` (#1125) ↗ huggingface/kernels-community

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

Mon Aug 31 Transformers tightens integration tests, TRL pins hybrid attention coverage Tue Sep 1 Transformers drops deprecated code, adds NeoMME and gradient checkpointing offload Wed Sep 2 kernels gains minver checks, diffusers deprecates ONNX Thu Sep 3 Jinja engine aligns with Jinja2; hub-kernel fallbacks now warn once Fri Sep 4 TRL documents 1M-token training, diffusers fixes group offloading, kernels fix Intel heuristics Sat Sep 5 TRL removes PPOTrainer, vendors Liger-Kernel losses Sun Sep 6 Transformers adds Kimi linear support
← previous week huggingface wire →