99 wires and counting

$ follow Hugging Face

Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-09-01
stories 51

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

Transformers drops deprecated code, adds NeoMME and gradient checkpointing offload

By RepoJournal · Filed · About Hugging Face

Transformers, TRL, and Kernels-Community ship fixes and features that touch long-standing workflows, from deprecated API removal to VLM training and flash-attn2 pointer bugs.

Transformers removed deprecated code in a sweeping cleanup [1], meaning any code relying on those APIs will need to migrate before the next release. In the same repo, NeoMME and NeoMME-Retriever were added [2], bringing new model architectures to the library, though integration tests are temporarily removed pending public weights. Gradient checkpointing now supports offload [3], which can reduce memory usage during training by moving activations off the GPU. A behavior change: SmolVLM/Idefics2/Idefics3 previously nulled pixel_values when image_hidden_states was present; now generate() raises an error if both are set, so users must clear pixel_values manually [4]. RoPE encoders dropped a position-indexed token type lookup [5], which could affect positional encoding behavior in those models. In TRL, AsyncGRPOTrainer now supports VLMs [6], but "this DOESNT support multimodal training" per the PR, freezing everything outside the text tower and fixing two specific functions. Kernels-community fixed dangling pointers in flash-attn2 backward pass when num_kv_heads != num_heads [7], which can prevent memory corruption in attention backprop.

Action items

References

  1. [1] Deprecated stuff gone (#48367) ↗ huggingface/transformers
  2. [2] Add NeoMME and NeoMME-Retriever (#47992) ↗ huggingface/transformers
  3. [3] Add offload to gradient checkpointing ↗ huggingface/transformers
  4. [4] Document image_hidden_states/pixel_values mutual exclusivity for SmolVLM/Idefics2/Idefics3 (#47714) ↗ huggingface/transformers
  5. [5] fix(models): Drop the position-indexed token type lookup in RoPE encoders (#48407) ↗ huggingface/transformers
  6. [6] fix vlm support asyncgrpo ↗ huggingface/trl
  7. [7] flash-attn2: fix dangling pointers in bwd when `num_kv_heads != num_heads` (#1125) ↗ huggingface/kernels-community

Quick answers

What shipped in Hugging Face on September 1, 2026?
Transformers, TRL, and Kernels-Community ship fixes and features that touch long-standing workflows, from deprecated API removal to VLM training and flash-attn2 pointer bugs. In total, 25 commits and 26 pull requests landed.
Who contributed to Hugging Face on September 1, 2026?
9 developers shipped this update, including Raushan Turganbay, Tony Wu, qgallouedec, verma8076, Harshal Janjani, dependabot, sayakpaul, and Daniël de Kok, and 1 more.
What were the notable Hugging Face updates?
Deprecated stuff gone (#48367), Add NeoMME and NeoMME-Retriever (#47992), and Add offload to gradient checkpointing.