[FMHA] add support for later CUDA (8.x)

Open jqueguiner opened this issue 4 years ago • 4 comments

add support for later minor dprops 8.0+ for CUDA DEVICES such as A40

Nov 25 '21 10:11 jqueguiner

@jqueguiner Thanks for the PR but I'm afraid we need to relax the condition in backward path as well in https://github.com/NVIDIA/apex/blob/f7421555c3d2ff01eed0e7c0c4321f3e4dd58fc6/apex/contrib/csrc/fmha/fmha_api.cpp#L182-L191.

Thank you for your patience in my delayed response.

Dec 04 '21 00:12 crcrpar

@crcrpar : thanks for the feedback; should be fixed now.

thanks

Dec 08 '21 09:12 jqueguiner

Hi @crcrpar any update regarding this matter ?

Dec 16 '21 03:12 jqueguiner

I found that this FMHA failed when I ran https://github.com/NVIDIA/apex/blob/master/apex/contrib/test/fmha/test_fmha.py with this PR on A40 because the kernel might violate some limitation unfortunately...

cc'ing @yjk21 for visibility

Dec 16 '21 04:12 crcrpar