Are there any plans to support flash-attn 1?

[mateiradu88]

Are there any plans to support flash-attn 1? This is a serious limitation at the moment for anyone still running Turing architecture, as flash-attn 2 does not support Turing.

[mateiradu88]

To anyone struggling with this issue and running on Turing, simply disabling the flash-attn 2 checks in code seems to work just fine! You will then be able to run on flash-attn 1 without errors, and results look similar if not the exact same as hugging-face demo.

[thomashay]

@mateiradu88 How exactly did you do this? I tried config=self.config, use_flash_attention_2 = False in line 115 of meshanything.py, but that results in errors later from shape_opt.py.

[mateiradu88]

I deleted the files, but I found an if condition checking for the flash attention version 2, deleting the condition and running the branch anyway worked like a charm, no other modifications as far as I remember.

[thomashay]

Also worked for me, thanks! For anyone else: I just removed the ifelse in line 347 in shape_opt.py (only the if and else statement, not the body of the loop, so just lines 347, 356 and 357) and could run the point cloud command line inference test.