$ the-wire · showcase
vLLM CI runs again, OpenVINO exporter unblocks fused attention
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
10 people shipped this
nbarr07
@nbarr07
1 cited
cfahlgren1
@cfahlgren1
1 cited
Yih-Dar
@ydshieh
1 cited
pculliton
@pculliton
1 cited
Ilyas Moutawwakil
@IlyasMoutawwakil
1 cited
HuggingFaceInfra
@HuggingFaceInfra
Bryan Hundven
@bhundven
eightman
@eightman999
BlivionIaG
@BlivionIaG
hf-security-analysis[bot]
@hf-security-analysis[bot]
Transformers restored a vLLM nightly CI that had failed in every job since 2026-10-16, and an exporter fix lets OpenVINO fuse attention, RoPE, the KV cache and compressed matmuls instead of decomposing them.
vLLM CI: fix install and KV-cache memory, add failure report job huggingface/transformers
The vLLM nightly CI had failed in all jobs at "Install dependencies" since 2026-10-16; this lands the install fix, the KV-cache memory fixes that surfaced afterward, and a new failure report job. That job ships structured output: analyze_gh_wf_runs.py is now limited to log fetching and parsing, and a new analyze_vllm_ci.py emits five JSON reports per run, including failing tests with trace and ...
OpenVINO exporter: let OpenVINO fuse attention, RoPE, the KV cache and compressed matmuls (#50311) huggingface/transformers
The OpenVINO exporter swapped repeat_kv for repeat_interleave, which lowered to a Gather between the stateful KV cache and attention, so OpenVINO's CPU plugin could not fuse it and every attention ran decomposed into MatMul/Softmax while copying the cache. Reverting to repeat_kv, including in the SDPA patch, restores the fused KV-cache attention path on exported models.
[Gemma4] Stop prefill at last KV-owning layer when logits_to_keep=1 huggingface/transformers
In KV-shared Gemma 4 models (num_kv_shared_layers > 0, e.g. Gemma 4 E2B with 20/35 shared layers and E4B with 18/42), the trailing shared layers own no KV, so with logits_to_keep=1 prefill now stops at the last KV-owning layer instead of running through them.
[3d datasets docs]: document 3d previews, mesh columns and the simready tag huggingface/hub-docs
The 3D datasets docs now cover .ply, .splat, .obj and .stl previews, preview size limits, Dataset Viewer thumbnails and the simready tag, plus how to add a Mesh column to a Parquet dataset since raw 3D files and Parquet cannot share one repository.
Remove Nscale text-to-image support from providers table (#2962) huggingface/hub-docs
Nscale is retiring its last text-to-image model, so it comes out of the hand-maintained providers table in index.md. On the routine side, a bot regenerated the Inference Providers docs, and hardware profiles gained the RTX 5070 Mobile 12G, Radeon RX 6400, Radeon Pro WX 2100, Ryzen AI Max+ PRO 495, Raspberry Pi 4 (BCM2711) and older Kepler-era cards like the GT 730.
