RepoJournal
Hugging Face

@huggingface

Transformers, Datasets, and the open AI-model layer

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

One email a day. Unsubscribe in one click.

Pick a date

  1. Sat 11 jul
  2. Sun 12 jul
  3. Mon 13 jul
  4. Tue 14 jul
  5. Wed 15 jul
  6. Thu 16 jul
  7. Fri 17 jul
  8. Sat 18 jul
  9. Sun 19 jul
  10. Mon 20 jul
  11. Tue 21 jul
  12. Wed 22 jul
  13. Thu 23 jul
  14. Fri 24 jul
  15. Sat 25 jul
  16. Sun 26 jul
  17. Mon 27 jul
  18. Tue 28 jul
  19. Wed 29 jul
  20. Thu 30 jul
  21. Fri 31 jul
  22. Sat 1 aug
  23. Sun 2 aug
  24. Mon 3 aug
  25. Tue 4 aug
  26. Wed 5 aug
  27. Thu 6 aug
  28. Fri 7 aug
  29. Sat 8 aug
  30. Today 9 aug
Topics: Python AI / ML Full archive →

The Wire · Showcase

SPEECH-TO-SPEECH PROTOCOL OVERHAUL SHIPS, DIFFUSERS FIXES LORA STATE CORRUPTION

By RepoJournal · Filed · About Hugging Face

andimarafioti danieldk Aloys Jehwin sayakpaul YaseenBashaT behroozazarkhalili Mayankpratapsingh022
7 people shipped this
andimarafioti @andimarafioti 4 cited danieldk @danieldk 3 cited Aloys Jehwin @AloysJehwin 1 cited sayakpaul @sayakpaul 1 cited YaseenBashaT @YaseenBashaT 1 cited behroozazarkhalili @behroozazarkhalili 1 cited Mayankpratapsingh022 @Mayankpratapsingh022

The speech-to-speech realtime pipeline just rebuilt its entire response lifecycle to keep audio, transcripts, and tool calls synchronized across WebSocket and WebRTC, while diffusers patched a critical bug where partially unfused LoRA adapters were marked as clean but remained fused in silent components.

The speech-to-speech team landed four interconnected protocol fixes that address the root cause of output lifecycle fragmentation [1][2][3][4]. The changes reserve stable assistant output identities at audio start, align transcript event semantics to emit full completion signals, introduce an explicit backend registry to eliminate conditional duplication, and populate `response.done` with the actual generated output items so clients get a coherent view of what the assistant produced. This is the protocol stabilization that makes tool-calling reliable at scale. On the diffusers side, Aloys Jehwin fixed a state corruption bug where `unfuse_lora(components=[...])` with a component subset would drop adapters from `_merged_adapters` as soon as they unfused from the first component, even though they remained physically baked into the others [5]. The fix recomputes `_merged_adapters` by scanning all loadable components after unmerge. Sayak Paul also fixed torchao test failures [6]. The kernels team opened kernel-kernel dependency support: danieldk added `kernel-depends` to `build.toml` metadata across kernels-data and kernels repos, laying groundwork for kernel composition [7][8], and started testing einops as the first kernel with inter-kernel dependencies [9]. TRL squashed a GRPO loss aggregation bug where the `luspo` branch never applied per-token completion masks elementwise [10] and excluded a buggy TruffleHog detector that was triggering false positives on ordinary 35-character test names [11].

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] Harden Realtime response output lifecycle ↗ huggingface/speech-to-speech
  2. [2] Align Realtime transcript event semantics ↗ huggingface/speech-to-speech
  3. [3] Introduce backend registry for pipeline construction ↗ huggingface/speech-to-speech
  4. [4] Populate response.done with generated output items ↗ huggingface/speech-to-speech
  5. [5] fix(lora): only drop adapter from _merged_adapters when unfused from all components (#14385) huggingface/diffusers
  6. [6] [tests] fix torchao tests (#14258) huggingface/diffusers
  7. [7] Add `kernel-depends` option to `general` options in `build.toml` ↗ huggingface/kernels
  8. [8] Add `kernel-depends` option to `general` options in `build.toml` (#749) huggingface/kernels
  9. [9] einops: add ↗ huggingface/kernels-community
  10. [10] [GRPO] Apply the completion mask elementwise in the luspo loss aggregation ↗ huggingface/trl
  11. [11] ci: exclude the buggy TruffleHog lob detector ↗ huggingface/trl

Quick answers

What shipped in Hugging Face on August 8, 2026?
The speech-to-speech realtime pipeline just rebuilt its entire response lifecycle to keep audio, transcripts, and tool calls synchronized across WebSocket and WebRTC, while diffusers patched a critical bug where partially unfused LoRA adapters were marked as clean but remained fused in silent components. In total, 32 commits and 15 pull requests landed.
Who contributed to Hugging Face on August 8, 2026?
7 developers shipped this update, including Aloys Jehwin, sayakpaul, andimarafioti, Mayankpratapsingh022, YaseenBashaT, behroozazarkhalili, and danieldk.
What were the notable Hugging Face updates?
Harden Realtime response output lifecycle, Align Realtime transcript event semantics, and Introduce backend registry for pipeline construction.

More from @huggingface

Hugging Face @huggingface

MOON-IDE SHIPS RELIABILITY OVERHAUL ACROSS MESSAGING AND SESSIONS

Hugging Face's IDE tooling just got significantly more resilient, with five commits landing stale response handling, markdown rendering fixes, and a pending message queue for network hiccups.

David Swanson HuggingFaceInfra ydshieh coyotte508
python 11 shipped
Hugging Face @huggingface

TRANSFORMERS SHIPS AUTOMATED CODE QUALITY REVIEWS, MLCLAW FIXES BROKER COMPATIBILITY

Transformers now posts mlinter findings directly to PRs as inline comments, while MLClaw v0.10.2 ships a fix for Codex broker metadata rejection.

ydshieh sywangyi Albert Villanova del Moral sayakpaul +1
python 29 shipped
Hugging Face @huggingface

SPEECH-TO-SPEECH 0.2.12 SHIPS WITH SMARTER TURN-TAKING AND WEBRTC SUPPORT

The final 0.2.x release lands Smart Turn v3.2 endpointing, WebRTC support, and direct audio input for audio-capable LLMs, while diffusers onboards MiniMax-H3 for joint video and audio generation.

andimarafioti PigeonDan1 akshan-main apolinario +5
python 57 shipped
Hugging Face @huggingface

KERNELS REFACTOR CLEANS UP LOADING PATH, TOKENIZERS KILLS DEPENDABOT NOISE

Hugging Face is untangling kernel loading internals across multiple repos while simultaneously eliminating 50% of its Dependabot burden through surgical dependency cleanup.

danieldk IlyasMoutawwakil vasqu Arthur
python 26 shipped

Daily updates, in your inbox

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

One email a day. Unsubscribe in one click. Read a past issue →

Elsewhere on the wire

AI Agents AI Agents

CLINE FIXES ABORT-WINDOW PROMPT LOSS, CLIPBOARD IMAGES LAND ON DESKTOP

Cline's core agent now preserves queued prompts when you hit stop, and desktop users can finally paste screenshots directly into the composer.

dependabot malhotra5 Saoud Rizwan
ai-agents 40 shipped
Local LLMs Local LLMs

OLLAMA CUTS MODEL CONFIG BLOAT 100KB+ PER MODEL, LLAMA.CPP SHIPS TOOL ISOLATION AND CUDA FUSION GAINS

Ollama eliminated redundant manifest data ballooning config blobs on safetensors models, while llama.cpp landed docker-backed tool isolation and fused three GPU kernels into one.

Daniel Hiltgen github-actions[bot] grafail Isotr0py +7
llms 102 shipped
CachyOS @CachyOS

WIKI TRANSLATION PUSH HITS SPANISH AND PORTUGUESE, NVIDIA DRIVERS GET 7.2 SUPPORT

CachyOS documentation expands globally with AI-assisted Spanish and Portuguese updates while the core team hardens nvidia-470xx for kernel 7.2.

witchlliee github-actions Vasiliy Stelmachenok Vladislav Nepogodin +3
infra 73 shipped
Elixir & Phoenix Elixir & Phoenix

MACRO.TO_STRING BUG BREAKS FOR COMPREHENSIONS WITH TRAILING KEYWORDS

Elixir's Macro.to_string/1 silently corrupts for-comprehensions that use do-end blocks with non-block trailing keywords, rendering valid code that no longer compiles.

makoto-developer
elixir 2 shipped
Want every project, not just this one?