Thanks for the contribution.

I think this would be better implemented in the C++ code. Specifically embedding/embedding_backward in aten/src/ATen/native/Embedding.cpp.

The new functionality also deserves a test. (test_embedding_padding_idx in test_nn.py)

Pushing it down to the C++ level would be harder, because it's hard to choose the "no padding" value if it's just an int. You either need an extra flag, or choose INT_MIN or INT_MAX, but these are annoying to get from Python

@colesbury @apaszke thanks for the feedback.

I kept the changes in Python because the C++ code uses -1 to denote no padding, and changing that would require jumping through many more hoops.

I have added unit tests to test_nn.py to account for the new functionality.

Hi @colesbury @apaszke sorry to bother you, but do the changes look fine to you? Can this be merged in that case?

Thanks @dasguptar!