Closed Jasonlee1995 closed 2 years ago
I used the ffcv imagenet codes and use it to my dataset.
What I experienced DDP with pytorch dataset and dataloader, they use mostly the same vram (balanced)
However when I use with FFCV dataset and dataloader, they are not balanced (0 gpu is getting way more vram than others)
Is this normal when I use DDP with FFCV dataset and dataloader or it's bug?
Hello,
This is a well known bug and has been fixed, we haven't made a formal release yet but you can install directly from source and it should be fixed (branch 0.0.4 or 0.0.1 should work)
I used the ffcv imagenet codes and use it to my dataset.
What I experienced DDP with pytorch dataset and dataloader, they use mostly the same vram (balanced)
However when I use with FFCV dataset and dataloader, they are not balanced (0 gpu is getting way more vram than others)
Is this normal when I use DDP with FFCV dataset and dataloader or it's bug?