Kai Londenberg

[BUG] cellpose causing numerical issues in unrelated code in same process.

3

**Describe the bug** the file [vit_sam.py](https://github.com/MouseLand/cellpose/blob/v4.0.4/cellpose/vit_sam.py) contains the line ```python torch.backends.cuda.matmul.allow_tf32 = True ``` right within the imports at the top. This is globally changing the pytorch configuration, which can...

Fix Flash3 CUDA capability check

4

## What does this PR do? Fixes #1361 and #1356 flash3.py incorrectly checks the CUDA COMPUTE CAPABILITY reporting that the GPU is "too new" when it clearly is not.

Incorrect CUDA_MINIMUM_COMPUTE_CAPABILITY check in fmha/flash3.py

1

Inside xformers flash3.py, there's an incorrect check of CUDA_MINIMUM_COMPUTE_CAPABILITY See https://github.com/facebookresearch/xformers/blob/97b1c754c2d1c80f4106eb2c17693e1ce2f32971/xformers/ops/fmha/flash3.py#L648-L658 It should be a "smaller-than" comparison, but it's greater-than.