Two builds are failing because of test_variant_consistency_jit_addr_cuda_bfloat16 with

RuntimeError: at::cuda::blas::gemv: not implemented for N3c108BFloat16E

@zasdfgbnm @ngimel we have a sip to recommend for this, right? Maybe we should think about giving that skip a clearer name so it's easier to read and discover.

Update: actually, nevermind. We should fix this in that test and not bother with a more precise skip. @anjali411, would you just add this dtype to the skip list?

Just need to skip the one failing test.

Edit: oops, one other thing. The existing method_test entries should be incorporated into addr's sample inputs and removed from method_tests:

pytorch/torch/testing/_internal/common_methods_invocations.py

Line 2065 in 7e05d07

('addr', (S, M), ((S,), (M,)),),

are you suggesting to add the sample inputs to test broadcasting in the torch.addr OpInfo based test and skip est_inplace_grad and test_variant_consistency_eager for all dtypes?

I didn't delete the method_test entry because I didn't add broadcasting tests in the new OpInfo based test. I also saw that torch.addmm method_tests haven't been removed yet even though a new OpInfo based test (which does not test broadcasting, as tested in method_tests) has been added for it.

are you suggesting to add the sample inputs to test broadcasting in the torch.addr OpInfo based test and skip test_inplace_grad and test_variant_consistency_eager for all dtypes?

Yes.

I didn't delete the method_test entry because I didn't add broadcasting tests in the new OpInfo based test. I also saw that torch.addmm method_tests haven't been removed yet even though a new OpInfo based test (which does not test broadcasting, as tested in method_tests) has been added for it.

That makes sense, but I think it's OK to move and skip in this case. We don't need to update addmm here.

@anjali411 merged this pull request in 1cc8f8a.