Boris Fomitchev

> The failing tests seem to be those I did not touch - like test_packed_sequence. I can fix them as well though. Sure those failures turned out to be from...

> @borisfom The changes seem to breaks three CI tests. Any idea? Fixed all issues. Should be good to go now.

> If the PR is approved, I think it makes sense that the other two RNN modes should also be modified in this PR? Yes RNN/GRU are also modified.

> Would you update their tests in onnx as well? I think they are parametrized but in different functions Done.

@kit1980 : how do we get this merged ?

@kit1980 @titaiwangms : Can we get this merged ?

I sure will! I have noticed, however, that when I install pytorch_scatter, I end up calling Torch's scatter_add instead anyway. Is current scatter code here obsolete ? In any case,...

Thanks for the confirmation! I guess there is no point of doing PR then.

If you have an idea ho to implement same scatter semantics with __nv_bfloat162 type, that may be something that PyT folks can use! As it stands now, the best thing...

Is it still the case?