pytorch / FBGEMM

FB (Facebook) + GEMM (General Matrix-Matrix Multiplication) - https://code.fb.com/ml-applications/fbgemm/

`.to("meta")` is leaked to the public main branch in the tests.

liligwu opened this issue a year ago · comments

Li Li commented a year ago

Please check the two tests below.
https://github.com/pytorch/FBGEMM/blob/b42ff6af1ded425e01ce098071aa304983cd9320/fbgemm_gpu/test/permute_pooled_embedding_test.py#L173C49-L173C49

FBGEMM/fbgemm_gpu/test/sparse_ops_test.py

Line 599 in b42ff6a

mx = torch.randint(low=0, high=100, size=(n,)).type(index_dtype).to("meta")

Benson Ma commented a year ago

Not sure if I understand the question correctly, but the meta here refers to meta tensors as opposed to Meta (company) -internal code (see https://discuss.pytorch.org/t/a-new-way-to-run-and-support-autograd-with-meta-tensors/160225 for example), so it should not be leaked internal code. @brad-mengchi, @shintaro-iwasaki, or @sryap could you confirm?

Shintaro Iwasaki commented a year ago

Thanks @q10. Yes it looks quite confusing, but this is a meta tensor.