CUDA role in Memory Efficient Attention

[BillyGun27]

I am wondering why do we need NVIDIA GPUs with compute capability above 6.0 (P100+) for Memory-efficient attention?
Does this function require a specific CUDA operation that is not exist in the one with lower capability?

[danthe3rd]

Hi,
Memory-efficient attention has been tested on 6.0+, but should work with 5.0+. This kernel relies on CUTLASS which might not work below 5.0.
If you have an older GPU, it might work if you build from source (and with some luck), but I don't recommend that, and we won't support it if you have questions or issues.