Steven S. Lyubomirsky

The original proposal has been implemented in #16129, though focusing mainly on dataflow blocks instead. I am also now working on an addition to have special handling for `split` and...

Discussion from the Feb. 7, 2023 community meeting after this work was presented: * This seems to be focused on the single-device case, so the natural comparison is against PyTorch...

I would be curious to hear about the plans for further developing AD in Relax. We should be able to support all of the language's features by building up the...

If you mean the branches of `If` nodes, that would be a mistake if we did it. We require those to be `SeqExpr`s. Edit: If you mean that `return` statements...

Okay, thanks for the update. I think the workaround is fine for my purposes. It would be nice for the parser to have syntactic sugar for it and at minimum...

Yeah, `return` is not part of Relax's grammar; it's from Python's grammar and it is up to us to determine how to parse it. This is actually opening a pretty...

The Relax AST does not permit omitting the false branch, so how would we parse it? (Even without a return)

I think we should parse that to `If(y, R.add(x, x), R.mul(x, x))`. The tricky thing is we're parsing Python _statements_ into Relax _expressions_.

> Edit: The motivation here is that the behavior of `return` in the script should match the behavior of it in python code. Yeah, this is a bit of the...

Results of discussion in the Feb. 14, 2023, community meeting: There was no opposition voiced to this proposal. I will implement this change and we can determine if any modifications...