Kimi K3 Performance Optimizations in vLLM: The Road to 2.8× Throughput
Kimi K3 Performance Optimizations in vLLM: The Road to 2.8× Throughput
摘要
Kimi K3 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU kernels.
Kimi K3 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU kernels.
译自 vLLM · 官方博客 · 录于 二〇二六年九月十五日