Sean Silva
We added support for this in #1169
The current e2e testing is enough to close this issue, I think. I don't think per-op non-e2e tests add much value here.
I commented in the upstream issue https://github.com/pytorch/pytorch/issues/80738 - I hope they can change the return type to `Tensor?`. You can temporarily hack around it by putting some special case logic...
@navahgar @LucQueen -- do you have steps for reproducing? Our CI is green at head, so we will need a bit more information for reproducing. This seems like a weird...
@Yancey1989 -- thanks for raising this. If you are using the JIT IR importer (and not LTC), I have had good success with serializing the JIT IR module (TorchScript module)...
It looks like the shape function is not correct (failing PyTorch tests), see https://github.com/pytorch/pytorch/pull/80860
Oh, sorry, it is an "incorrect shape compute mapping function schema name" not incorrect shape function. Yes, we in general wait for these change to land upstream. Is there an...
I'm currently revamping InlineGlobalSlots -- will address this as part of the rewrite.
This should be handleable identical to how we handle broadcasting for elementwise ops. See here: https://github.com/llvm/torch-mlir/blob/9d6ee48661cefa677950bcd262fe2824a4fa9d17/lib/Conversion/TorchToLinalg/Utils.cpp#L157 In particular, dynamic size-1 broadcast is an error.
Yeah, I guess that would be ok.