Hacker News new | ask | show | jobs
by embedding-shape 122 days ago
The MXFP4 variant I suppose? My setup (RTX Pro 6000) does around ~140 tok/s with llama.cpp, around 160 tok/s with vLLM.
1 comments
vkaufmann 122 days ago
yep MXFP4 really fast :D
link