Open netw0rkf10w opened 1 year ago
Thanks! Is there some documentation on what's required to make things compatible with torch.compile?
@tridao That's a great question! Could you please ask the PyTorch developers directly? They are very actively looking for feedback. This issue on their repo could be useful: https://github.com/pytorch/pytorch/issues/90550.
PyTorch 2.0 has introduced
torch.compile
for accelerating training and inference. I have tried it on top of flash attention but unfortunatelytorch
seems to unable to compile flash attention:Hopefully you could make flash attention compatible with PyTorch 2.0 in the near future. (FYI flash attention is still faster than non-flash attention +
torch.compile
for ViT).