A6000 support for FlashAttentionV2
zhangliang-04 opened this issue · 3 comments
Is your feature request related to a problem? Please describe.
It seems that this codebase only supports the use of FlashAttentionV2 on A100 or A40. However, in theory all Turing or Ampere GPUs should work with it. Is there any plan for supporting A6000?
Describe the solution you'd like
Add FlashAttention support for A6000 gpu.
Thanks for pointing that out! Yeah I think it is totally possible, and all the required change is this line in install.sh (https://github.com/OptimalScale/LMFlow/blob/main/install.sh#L6). I will modify it soon when I have some time. Also, if you are interested in contributing, we welcome all kinds of contributions via PR. Thanks 😄
We have updated install.sh accordingly (#669). Could you please let us know if the problem is solved? Thanks 😄
Yes, it works! Thanks for your contribution😄