RepoJournal
Hugging Face @huggingface
89 wires and counting

$ follow Hugging Face

Keep up with Hugging Face in about 3 minutes a day: 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 morning, this isn't your newsletter.

One email a day. Unsubscribe in one click.

$ status

wire 2026-08-28
stories 96
GitHub RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

SAGE ATTENTION GETS ITS SECOND WIND

By RepoJournal · Filed · About Hugging Face

Sayak Paul albertvillanova Shadow_Lu Quentin Gallouédec tarekziade hf-security-analysis[bot] kaixuanliu
13 people shipped this
Sayak Paul @sayakpaul 2 cited albertvillanova @albertvillanova 2 cited Shadow_Lu @LuShadowX 1 cited Quentin Gallouédec @qgallouedec 1 cited tarekziade @tarekziade 1 cited hf-security-analysis[bot] @hf-security-analysis[bot] 1 cited kaixuanliu @kaixuanliu 1 cited drbh @drbh 1 cited akshan-main @akshan-main Yih-Dar @ydshieh Zhiliang Peng @pengzhiliang sergiopaniego @sergiopaniego verma8076 @verma8076

Sayak Paul re-synced SageAttention with upstream, and suddenly the consumer-GPU video crowd has a reason to care again.

The kernels-community desk leads with sage-attention: Sage and Sage2 are now synced with upstream after a period of staleness, and Sage3 is out of the tree entirely, since it's Blackwell-only and its FP4 mainloop relies on `mma.sync` instructions that CUTLASS only enables for consumer Blackwell (sm120/sm121), not datacenter sm100a [1][2]. The repo also picked up a nested build-toml feature to support the incoming kernel dirs [3], plus an XPU conv1D kernel for causal-conv1d [4]. On the security front, a workflow vulnerability in pr-comment-build.yaml got remediated, so if you run kernels-community in CI, that one is worth a look [5]. Over in diffusers, a callback fix closes a fatal KeyError when users request prompt_embeds or masked_image_latents in pipelines like LEditsPP and SDXL ControlNet Union [6]. TRL is quietly locking down quantization coverage: QLoRA tests now exercise the documented quantization_config argument across SFT, DPO, KTO, and newly GRPO, RLOO, Reward, and Distillation trainers, closing a real coverage gap [7][8]. Transformers is silencing the noise: continuous batching no longer spews five WARNING lines and a progress bar at default verbosity, which is a small win for anyone reading logs in production [9]. And the mlinter repo is prepping 0.1.5, adding the missing TRF058 public-API assertions and fixing two ty diagnostics in trf009 [10].

Action items

References

  1. [1] sage-attention: sync sage and sage2 ↗ huggingface/kernels-community
  2. [2] sage-attention: sync sage and sage2 (#1095) huggingface/kernels-community
  3. [3] feat: Nested build toml support ↗ huggingface/kernels-community
  4. [4] causal-conv1d: add conv1D kernel for XPU (#1079) huggingface/kernels-community
  5. [5] fix(security): remediate workflow vulnerability in .github/workflows/pr-comment-build.yaml (#1116) huggingface/kernels-community
  6. [6] Fix callback tensor inputs that are never bound in the denoising loop (#14416) huggingface/diffusers
  7. [7] Convert the QLoRA tests to the quantization_config trainer argument ↗ huggingface/trl
  8. [8] Add a QLoRA test to the trainers that support quantization_config ↗ huggingface/trl
  9. [9] Quiet continuous batching at default verbosity (#48314) huggingface/transformers
  10. [10] preparing for 0.1.5 huggingface/transformers-mlinter
Topics: Python AI / ML Full archive →

Quick answers

What shipped in Hugging Face on August 28, 2026?
Sayak Paul re-synced SageAttention with upstream, and suddenly the consumer-GPU video crowd has a reason to care again. In total, 48 commits and 48 pull requests landed.
Who contributed to Hugging Face on August 28, 2026?
13 developers shipped this update, including Sayak Paul, Shadow_Lu, akshan-main, Quentin Gallouédec, Yih-Dar, Zhiliang Peng, albertvillanova, and sergiopaniego, and 5 more.
What were the notable Hugging Face updates?
sage-attention: sync sage and sage2, sage-attention: sync sage and sage2 (#1095), and feat: Nested build toml support.

$ log @huggingface --recent

TRANSFORMERS V5.16.0 SHIPS WITH QWEN4-EXP 131 shipped TRL UNBLOCKS MOE LORA AS PEFT PATCHES AUTOCAST 82 shipped Transformers MoE loss cut 99.7% as TRL gains tool calling 62 shipped Transformers linter hunts nn.Buffer misuse as kernels ships hotfix 7 shipped

$ log --all # elsewhere on the wire

AI Agents AGENT TOOL SIGNATURES NOW ENFORCED BEFORE EXECUTION Local LLMs LLAMA.CPP UNLEASHES QWEN3.8-FLASH-NEXT SUPPORT

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.