2 ms·but unlike cuda there's no custom kernels for inference in vllm repo... I thinkby tough 1y agobut unlike cuda there's no custom kernels for inference in vllm repo... I think