vLLM ships Day-0 support for Step 5 Preview with sparse attention, MoE and P/D on Blackwell
vllm_project · x · 2026-10-15
vLLM announced Day-0 support for Step 5 Preview, covering sparse attention, MoE, multi-token prediction and vision across Blackwell and Hopper. Reported results: MTP3 on B300 DEP8 lifted AIME25 from 0.9500 to 0.9667 with GPQA/OCRBench roughly flat and zero request errors; sparse prefill kernels cut indexer logits latency 43% and top-k latency 84% (overall prefill -2.2%). A prefill/decode prototype linked DEP4 prefill to DEP8 decode on 12× GB300 with FP8, NIXL and a 384 GiB cache store, recording 514,784 prefix-hit tokens (91.24%) and OCR scores of 90.4%/89.4% with zero errors. These are pre-release, pre-rebase figures.
Related event: vLLM Ships Day-0 Support for Step 5 Preview Across Blackwell and Hopper(7 posts)→
More from Infra
- Magnitude open-source inference engine beats Ollama by over 2x with auto-tuned kernels — nickbaumann_ · 2026-10-15
- Nvidia to announce $28B in AI investments in Europe this week, report says — georgemillo · 2026-10-15
- A $10,811 Cloudflare Bill From an Unused AI-Written Test Project: Refunds Aren't a Safety Net — sujingshen · 2026-10-15
- Retraining-Free Method Turns Transformers Into Elastic Models Adapting to Any Compute Budget — VISLab-Amsterdam · 2026-10-15
- vLLM's MTP3 validation: AIME25 up to 0.9667, prefill kernels up to 84% lower latency — vllm_project · 2026-10-15
- Open-weights s1-llm-auto-router picks the right LLM in 26ms on a CPU — airesearch12 · 2026-10-15