You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
For Llama3, when using FSDP2 the model weights are automatically converted to bfloat16 for mixed precision training. However, I notice with Llama4, when using FSDP2 I have to manually cast the weights to bfloat16 before this line otherwise they stay in fp32 and I can't use float8 GEMMs.