vLLM cuts Step 5 sparse prefill kernel latency: top-k down 84%, indexer down 43%
vllm_project · x · 2026-10-15
The vLLM team shared kernel-level optimizations for Step 5 Preview's sparse attention prefill path, benchmarked on B300 with DEP8, MTP3, FP8 KV cache, a 60K-token input, and the MegaMoE path:
- Indexer logits kernel: 43% lower latency
- Top-k kernel: 84% lower latency
- Overall prefill: 2.2% lower latency (stages are individually normalized and not additive)
The work also removes 7 runtime JIT warmups and caps chunks at 4,096 tokens. The team notes that large per-kernel gains translate to only a modest end-to-end improvement.
More from Infra
- Chip equipment beats bull case: 2026 WFE seen at $153B, $278B by 2030 — BenBajarin · 2026-10-16
- Steam now launches in 5.23s, beating Epic's 5.47s — and vLLM gets roasted for slow startup — AlpinDale · 2026-10-16
- Shadeform to Announce Six New Features on Oct 19: Lower GPU Costs, More Suppliers — _jaydeepkarale · 2026-10-16
- Fastino launches Training API for SFT/RL, claiming 25-85% cost cut vs Tinker — kalyan_kpl · 2026-10-16
- Dallas Fed: AI data center debt issuance equals ~1/8 of Treasury duration supply — _onionesque · 2026-10-16
- Running Wan2.1 14B FP16 on a 12GB laptop GPU: VRAM swap stress test finishes in 4m11s — RYuu-1120 · 2026-10-16