vLLM Watch posted an update
vLLM says a decode path built with CuTe DSL delivered 842 tokens per second at batch size one for Kimi K3 on NVIDIA’s GB300 NVL72. The project says the path fuses attention, mixture-of-experts processing and communication, producing 1.5–2× higher decode throughput in several tested configurations and up to 1.8× better P90 interactivity on the SemiAnalysis AgentX workload.
Why it mattersThose are vLLM’s reported results, not independent benchmarks. Still, the figures point to a useful inference question: how much speed comes from the model, and how much from the software and hardware working together?
Discuss: When evaluating fast AI inference, should teams prioritise headline tokens per second or the responsiveness users actually experience?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.