but unlike cuda there's no custom kernels for inference in vllm repo...

I think