RepoJournal
Hugging Face @huggingface
90 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 2026-08-29
stories 76
GitHub RSS Archive Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

Diffusers pins revisions once, TRL fixes vLLM weight-sync hangs

By RepoJournal · Filed · About Hugging Face

osolmaz Wauplin AmineDiro albertvillanova Muhammad Saeed Quentin Gallouédec sayakpaul
11 people shipped this
osolmaz @osolmaz 4 cited Wauplin @Wauplin 2 cited AmineDiro @AmineDiro 2 cited albertvillanova @albertvillanova 2 cited Muhammad Saeed @Muhammadsaeed707 2 cited Quentin Gallouédec @qgallouedec 1 cited sayakpaul @sayakpaul Tarek Ziade @tarekziade Rémi Ouazan @remi-or Kakulukian @Kakulukian Daniel van Strien @davanstrien

Diffusers now resolves the git revision a single time, and TRL turns silent vLLM weight-sync hangs into explicit errors.

Diffusers' from_pretrained now calls huggingface_hub.resolve_revision once at the start, so every fetched file is pinned to the same commit and served from cache without re-resolving on each download [1][2]. This mirrors a change already running in production in vLLM and is a follow-up to huggingface_hub v1.26.0, which introduced resolve_revision and ResolvedRevision. TRL fixed a GRPO weight-sync path where vLLM responses were discarded and the trainer blocked in an uninterruptible ncclBroadcast, hanging the run; the fix checks each request's status and raises with the server's error text, and moves the NCCL collective to a daemon thread [3][4]. Separately, the CI rerun filter now retries cuBLAS allocation errors, which pytest-rerunfailures previously missed because it only matches the outermost exception [5][6]. Harbor pinned all active deployment profiles to updated FX trial-worker images for the system-message ordering, provider-tool filter, and gateway repairs, regenerating every content-derived profile ID without changing any model or provider setting [7][8][9][10]. Transformers now raises when a paged attention forward is called with no cache, rather than silently computing non-causal attention [11]. Hub-docs documents a Network Security exemption for Enterprise Plus service account tokens, allowing CI automation outside corporate IP ranges to bypass IP restrictions and the Content Access Policy [12][13].

Action items

References

  1. [1] [Download] Resolve the revision once at the beginning of from_pretrained ↗ huggingface/diffusers
  2. [2] [Download] Resolve the revision once at the beginning of from_pretrained (#14340) ↗ huggingface/diffusers
  3. [3] Fix: vLLM weight sync fails instead of hanging ↗ huggingface/trl
  4. [4] Fix: vLLM weight sync fails instead of hanging (#6913) ↗ huggingface/trl
  5. [5] Retry tests failing on cuBLAS allocation errors the rerun filter missed ↗ huggingface/trl
  6. [6] Retry tests failing on cuBLAS allocation errors the rerun filter missed (#6916) ↗ huggingface/trl
  7. [7] fix(profiles): pin merged-system FX worker ↗ huggingface/harbor-hf
  8. [8] fix(profiles): pin FX system-order worker ↗ huggingface/harbor-hf
  9. [9] fix(profiles): pin FX provider-filter worker ↗ huggingface/harbor-hf
  10. [10] fix(profiles): pin FX gateway worker image ↗ huggingface/harbor-hf
  11. [11] Raise when a paged attention forward is called with no cache (#48297) ↗ huggingface/transformers
  12. [12] Document Network Security exemption for service account tokens (#2737) ↗ huggingface/hub-docs
  13. [13] Document Network Security exemption for service account tokens ↗ huggingface/hub-docs
Topics: Python AI / ML Full archive →

Quick answers

What shipped in Hugging Face on August 29, 2026?
Diffusers now resolves the git revision a single time, and TRL turns silent vLLM weight-sync hangs into explicit errors. In total, 38 commits and 38 pull requests landed.
Who contributed to Hugging Face on August 29, 2026?
11 developers shipped this update, including Wauplin, sayakpaul, Quentin Gallouédec, Tarek Ziade, Rémi Ouazan, osolmaz, AmineDiro, and albertvillanova, and 3 more.
What were the notable Hugging Face updates?
[Download] Resolve the revision once at the beginning of from_pretrained, [Download] Resolve the revision once at the beginning of from_pretrained (#14340), and Fix: vLLM weight sync fails instead of hanging.

$ log @huggingface --recent

SAGE ATTENTION GETS ITS SECOND WIND 96 shipped 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

$ log --all # elsewhere on the wire

AI Agents Langchain ships MCP adapter, Cline fixes telemetry and CVEs Local LLMs OpenVINO backend lands Qwen3.5 on NPU, Whisper.cpp 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.