🐛 Bug

Hi~
Thanks for your excellent work.
However, I am very confused that the minimum architecture required in flash3 is SM_80 when selecting the version of flash_attn:
https://github.com/facebookresearch/xformers/blob/main/xformers/ops/fmha/flash3.py#L597
Actually, flash_attn3 should run on a GPU with Hopper architecture, and the required architecture is SM_90. I am not sure if this is a little bug?
In my practical use, I found that the flash_attn version used in xformers is flash3, which works well on H20 GPU, while there is a CUDA illegal memory access issue on L20 GPU. Further investigation using cuda-gdb revealed that the exception occurred in the flash:: prepare_varlen_num-blocks_kernel, which is exactly one of the kernels in _C.so in xformers. To solve this problem, I changed CUDA_MINIMUM_COMPUTE_CAPABILITY to (9,0) when running on L20, and the program can run normally.
I'm not sure if it's my improper use or if there's a bug in the selection logic here.

To Reproduce

When I change CUDA_MINIMUM_COMPUTE_CAPABILITY to (8, 0), it can be stably reproduced on L20
While it works well if changing CUDA_MINIMUM_COMPUTE_CAPABILITY to (9, 0)

Environment

Python version: 3.10.12
PyTorch version: 2.7.1+cu126

CUDA runtime version: 12.2.91
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA L20
GPU 1: NVIDIA L20
GPU 2: NVIDIA L20
GPU 3: NVIDIA L20

torch==2.7.1
transformers==4.54.1
xformers==0.0.31
vllm==0.10.0

Addition notes

I find a PR #1178 on Dec 18, 2024, which enables the default version of flash_attn to 3. However, there is no code executed to check if it is compatible with flash3.
https://github.com/facebookresearch/xformers/blob/v0.0.31/xformers/ops/fmha/dispatch.py#L22