RepoJournal
Hugging Face

@huggingface

Transformers, Datasets, and the open AI-model layer

Pick a date

The Wire · Showcase

FLASH-ATTN3 TARGETS TORCH 2.9 AS UPLOAD TOOLING GETS OVERHAUL

By RepoJournal · Filed · About Hugging Face

Flash-attn3 is moving to Torch 2.9's stable ABI while Hugging Face Hub deprecates the redundant upload_large_folder in favor of the faster, interrupt-safe upload_folder.

The kernels team is locking flash-attn3 onto Torch 2.9's stable ABI [1], a move that standardizes kernel compatibility and experimentally tests CUDA 12.6 builds. Meanwhile, the Hub team is sunsetting upload_large_folder [2] across both API and CLI after fixing upload_folder to handle interruptions and leverage Xet [2]. Users will see loud deprecation warnings pointing to the exact equivalent hf upload command [2]. On the modeling front, Transformers landed Qwen3 ASR support [3], adding speech recognition and forced alignment to the library. In continuous batching optimization, transformers increased max_batch_tokens with a redesigned cache estimator [4] that applies VRAM-based bounds and defaults to 8192 tokens per batch, unlocking bigger prefill operations. Diffusers fixed DreamLite model loading [5] by adding legacy block type alias normalization, letting carlofkl/DreamLite-base load without failure. The diffusers team also bumped safetensors to 0.8.0 [6], bringing the latest serialization improvements.

Action items

References

  1. [1] flash-attn3: target Torch 2.9 stable ABI ↗ huggingface/kernels-community
  2. [2] [Upload] Deprecate upload_large_folder (API + CLI) ↗ huggingface/huggingface_hub
  3. [3] Qwen3 ASR and Forced Aligner ↗ huggingface/transformers
  4. [4] [CB] Changes to increase max_batch_tokens ↗ huggingface/transformers
  5. [5] Fix DreamLite legacy block type aliases ↗ huggingface/diffusers
  6. [6] feat: bump safetensors to 0.8.0 ↗ huggingface/diffusers

FAQ

What changed in Hugging Face on June 27, 2026?
Flash-attn3 is moving to Torch 2.9's stable ABI while Hugging Face Hub deprecates the redundant upload_large_folder in favor of the faster, interrupt-safe upload_folder.
What should Hugging Face teams do about it?
Test flash-attn3 Torch 2.9 builds in your kernel CI pipeline • Migrate from upload_large_folder to hf upload or upload_folder before deprecation window closes • Verify DreamLite model loading works in your diffusers workflows
Which Hugging Face repositories shipped on June 27, 2026?
huggingface/kernels-community, huggingface/huggingface_hub, huggingface/transformers, huggingface/diffusers

Related across the cluster

For your repos

The showcase is a teaser.
Your wire is the product.

Same engine. Different stack. Below: what changes when the wire is yours.

Showcase wire

  • 14 famous open source orgs
  • One wire per day
  • Public, generic
  • Read on the web, when you remember

Your wire

  • Up to 1,500 of your repos - orgs, deps, vendors
  • Morning and evening briefs
  • Action items routed to your team
  • Slack delivery, email, breaking-news CVE alerts

Want a hands-on demo first? Ask a current user for an invite link.