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

The Wire · Showcase

TRANSFORMERS SPLITS VLLM CI ACROSS DEDICATED RUNNERS; SENTENCE-TRANSFORMERS FIXES FLOAT PRECISION ACROSS ALL RANKING

By RepoJournal · Filed · About Hugging Face

ydshieh IlyasMoutawwakil dg845 KisuYang Mayankpratapsingh022 Hotragn andimarafioti
12 people shipped this
ydshieh @ydshieh 2 cited IlyasMoutawwakil @IlyasMoutawwakil 2 cited dg845 @dg845 2 cited KisuYang @KisuYang 2 cited Mayankpratapsingh022 @Mayankpratapsingh022 1 cited Hotragn @Hotragn 1 cited andimarafioti @andimarafioti 1 cited Sayak Paul @sayakpaul 1 cited tomaarsen @tomaarsen 1 cited Samoed @Samoed 1 cited dsulot @dsulot yiyixuxu @yiyixuxu

Transformers restructured its vLLM test pipeline to eliminate OOM failures while sentence-transformers shipped three critical precision fixes in float16/bfloat16 compute paths.

The transformers team moved vLLM CI to dedicated runners [1] [2], splitting test_initialization into an 8-shard matrix job to fix persistent out-of-memory errors that plagued sequential runs. This is the isolation win the team needed to keep vLLM test coverage in the pipeline without blowing up resource usage. On the sentence-transformers side, three separate fixes hit v6 simultaneously: CrossEncoder.rank now returns native Python floats instead of numpy.float32 scalars [3], which makes results JSON-serializable and stops rank comparisons from breaking on device tensors. More critically, both similarity computation [4] and cross-encoder logits [5] now upcast to float32 before activation, preventing reduced-mantissa formats from coarsely bucketing scores when models run in float16 or bfloat16. The team also exposed device and chunking parameters in MaxSim similarity functions to speed CUDA computation by up to 39x on large embeddings [6]. In diffusers, LTX25AutoBlocks got refactored into its own file with a dedicated model name [12] [14], and keep_in_fp32 module tests were tightened [13]. Speech-to-speech fixed Whisper STT handlers that were incorrectly finalizing turns on progressive audio [9] and rewrote language detection to handle recent transformers releases that no longer echo decoder prefixes [10]. Muse Glimmer is now exportable [7] [8], and the realtime protocol now overlaps tool-result generation [11].

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners ↗ huggingface/transformers
  2. [2] CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners (#47934) huggingface/transformers
  3. [3] [`v6`] refactor: return Python floats from CrossEncoder.rank ↗ huggingface/sentence-transformers
  4. [4] [`v6`] fix: compute similarity scores in float32 to avoid low-precision ties (HPS) ↗ huggingface/sentence-transformers
  5. [5] [`v6`] fix: upcast cross-encoder logits to float32 before activation (HPS) (#3893) huggingface/sentence-transformers
  6. [6] [`v6`] Forward similarity kwargs to similarity functions, expose MaxSim device & chunking ↗ huggingface/sentence-transformers
  7. [7] Make muse glimmer exportable ↗ huggingface/transformers
  8. [8] Make muse glimmer exportable (#47871) huggingface/transformers
  9. [9] fix(stt): stop Whisper handlers finalizing turns on progressive audio ↗ huggingface/speech-to-speech
  10. [10] Fix Whisper STT language detection reading a text token ↗ huggingface/speech-to-speech
  11. [11] Overlap tool follow-up generation within the Realtime protocol ↗ huggingface/speech-to-speech
  12. [12] Move `LTX25AutoBlocks` to new file with new model name ↗ huggingface/diffusers
  13. [13] [tests] tighten keep_in_32 modules tests (#14399) huggingface/diffusers
  14. [14] Move `LTX25AutoBlocks` to new file with new model name (#14453) huggingface/diffusers

Quick answers

What shipped in Hugging Face on August 13, 2026?
Transformers restructured its vLLM test pipeline to eliminate OOM failures while sentence-transformers shipped three critical precision fixes in float16/bfloat16 compute paths. In total, 18 commits and 17 pull requests landed.
Who contributed to Hugging Face on August 13, 2026?
12 developers shipped this update, including ydshieh, IlyasMoutawwakil, dsulot, Mayankpratapsingh022, Hotragn, andimarafioti, dg845, and yiyixuxu, and 4 more.
What were the notable Hugging Face updates?
CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners, CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners (#47934), and [`v6`] refactor: return Python floats from CrossEncoder.rank.

More from @huggingface

Hugging Face @huggingface

DIFFUSERS CLOSES LORA GAPS WHILE MOON-IDE PINS CACHE REPLICAS

Diffusers shipped LoRA loading for MiniMax-H3 and fixed test failures across the framework, while moon-ide added session affinity to stabilize prompt cache distribution.

coyotte508 dependabot Quentin Gallouédec Albert Villanova del Moral +6
python 31 shipped
Hugging Face @huggingface

OPTIMUM-INTEL ADDS META MUSE GLIMMER SUPPORT VIA OPENVINO

OpenVINO now exports Meta's Muse Glimmer multimodal model with int4 quantization, bringing image-text-to-text inference to the optimization stack.

rkazants Anastasiia Pnevskaia pkooij coyotte508 +5
python 20 shipped
Hugging Face @huggingface

SPEECH-TO-SPEECH SHIPS ORDERED ASSISTANT FIX, HUB SDK PATCHES PROGRESS BAR BUG

Hugging Face speech-to-speech preserves assistant text and tool output in a single ordered sequence, eliminating split-queue complexity across parsing, Realtime events, and conversation history.

andimarafioti imstevenpmwork pkooij Nicholas Broad +4
python 30 shipped
Hugging Face @huggingface

PYTORCH-IMAGE-MODELS FIXES POSITIONAL EMBEDDING RESAMPLING BUG

A shortcut in positional embedding resampling was silently skipping necessary transformations when token counts matched but grid shapes differed.

aswanth-07
python 2 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

LANGCHAIN SHIPS ENCRYPTED REASONING FIX, CLINE ADDS WEB SEARCH AND VOICE

ChatOpenAI now preserves OpenAI's encrypted reasoning from streamed responses, while Cline releases three versions with provider-executed web search and microphone transcription.

langchain-oss-model-profiles[bot] johannes117 langchain-oss-model-profiles github-actions[bot] +8
ai-agents 55 shipped
Local LLMs Local LLMs

QWEN3.8 LANDS IN OLLAMA WITH DEVELOPER INSTRUCTION SUPPORT

Ollama shipped full Qwen3.8 support overnight, adding reasoning-effort semantics and a critical fix for coding agents that send developer instructions before user messages.

dhiltgen 0 fairydreaming Hemanth Battu +12
llms 208 shipped
CachyOS @CachyOS

FIRMWARE AND INIT SYNC LANDS AS AUTOS ROLL FORWARD

CachyOS synced linux-firmware with Arch and bumped mkinitcpio to 41.1-1, while the package automation pipeline keeps qview and qtscrcpy current across the AUR-derived repo.

witchlliee github-actions Axel Olausson Holtenäs Peter Jung
infra 41 shipped
Elixir & Phoenix Elixir & Phoenix

PHOENIX LIVE VIEW SHIPS UPLOAD OVERHAUL, ELIXIR TUNES CORE PERFORMANCE

SteffenDE landed five critical fixes for LiveView's upload handler in one sweep, while the Elixir core team optimized Version comparison and regex operations.

ericmj preciz maennchen tomciopp +3
elixir 34 shipped
Want every project, not just this one?