sm90 efficient attention test fixes (#105978)
Fixes the following two test cases involving efficient attention on sm90:
Explanations:
functorch/test_ops.py: test_vjp_nn_functional_scaled_dot_product_attention_cuda_float32
* originally the test had xfail for all sm
* in https://github.com/pytorch/pytorch/issues/102029, we found that it was unexpectedly passing on sm90
* I made https://github.com/pytorch/pytorch/pull/102131 to update the test to let it pass
* @drisspg seems to have made changes to the behavior such that the original xfail was getting triggered (https://github.com/pytorch/pytorch/issues/102029#issuecomment-1560071148)
* the CI began complaining about the failure again: https://github.com/pytorch/pytorch/issues/102663
* I'm now reverting https://github.com/pytorch/pytorch/pull/102131 to bring back the original xfail now that the behavior has been fixed by @drisspg to trigger the xfail in sm90 similar to all other sm
test_transformers.py: test_mem_efficient_fail_sm90_cuda
* the test as it's currently written seems to expect the sdp dispatcher to fail for mem efficient attention on sm90; however, testing this on H100, it actually succeeds, so I'm disabling the test for now as the current expected result may be outdated
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105978
Approved by: https://github.com/eqy, https://github.com/kshitij12345, https://github.com/zou3519
