Image generation with deepspeed --fp16#394

Open

afiaka87 wants to merge 4 commits intolucidrains:mainfrom afiaka87:vqgan_precision_fixes

+14-19

Commits on Nov 30, 2021

(fp16/deepspeed w generations) Cast submodules to half except vqgan.
Clay Mullis
committed
Autocast VQGAN calls to fp32 in fp16 training
Clay Mullis
committed
enable pytorch's built-in automatic mixed precision by default
Clay Mullis
committed
remove duplicate code
afiaka87
authored