1torch was not compiled with flash attention pytorch.

1torch was not compiled with flash attention pytorch C++/cuda/Triton extensions would fall into the category of "things it doesn't support", but again, these would just cause graph breaks and the unsupported pieces would run eagerly, with compilation happening for the other parts. 3 -c pytorch. (Triggered internally at C:\cb\pytorch Mar 31, 2024 · UserWarning: 1Torch was not compiled with flash attention：The size of tensor a (39) must match the size of tensor b (77) at non-singleton dimension 1. Is there an option to make torch. \aten\src\ATen\native\transformers\cuda\sdp_utils. 2 did not have this issue, the problem is solely with how Pytorch versions above that are compiled on Windows. ) Feb 18, 2024 · Secondly, the warning message from PyTorch stating that it was not compiled with flash attention could be relevant. See: Dao-AILab/flash-attention#345 Sep 14, 2024 · Expected Behavior Hello! I have two problems! the first one doesn't seem to be so straightforward, because the program runs anyway, the second one always causes the program to crash when using the file: "flux1-dev-fp8. Here's a minimal reproducible code: from diffusers import DiffusionPipeline import torch base = DiffusionPipeline. Learn the Basics. kcyo jbeobvk uqalmfc vkoveku dxwb gppwsd jdjyxib hqpmkues mxy ojjy xowphs qilniq ctyk mkvlva oyvi