NVIDIA's vLLM Optimization Boosts Kimi K3 Decoding on GB300

vLLM merged an NVIDIA-developed decoding implementation for Kimi K3 (PR #61867), using a unified CuTe DSL path that fuses attention, MoE and communication. It delivers 1.5-2x faster decoding on GB300 NVL72, reaching 842 tokens/s at batch 1.

2026-10-18 ~ 2026-10-18 · 2 related posts

• vLLM and NVIDIA push Kimi K3 decode to 842 tokens/s on GB300 with fused CuTe DSL path — vllm_project · 2026-10-18
• vLLM merges NVIDIA's GB300-specialized Kimi K3 decode: 1.5-2x speedup — marksaroufim · 2026-10-18