Phi 3.5 MoE fp8 ?

#2
by nfunctor - opened

Hello,

Thanks for the wonderful models!

Do you think you might make a w8a8 version of Phi-MoE? It was added to VLLM a few days ago, even if the 0.6.0 version has a weight_name one-liner bug (PR accepted) it seems to work nicely so far.

Sign up or log in to comment