still need to test it in concurrency + try it in nvfp4 hopefully will be able to compare perplexity in bf16 / fp8 / nvfp4 as well as tp performance jumps from 2 -> 4 -> 8 nodes across all 3

→ View original post on X — @theahmadosman