ModelsBuildersMulti-source00:44Chinese contentinferact 开源 Kimi K3 TPU megakernel,单核跑出 709 tokens/sKimi K3 在 TPU 上跑到 709 tokens/s,比 GB200 还快,92 个 MoE 层塞进一个 Pallas kernel,做法挺巧,做推理的可以看看。#Kimi K3#TPU#Pallas#MoE7 sourcesEventOfficial accountvLLM@vllm_projectSourceRead laterReadWorth followingUsefulFollow Kimi K3