Sean Silva
I suspect it is a missing case here: https://github.com/llvm/torch-mlir/blob/e06ee0850658b304bd5003147d7403f132b03302/lib/Dialect/Torch/Transforms/ReifyShapeCalculations.cpp#L72 Can you try to fix it?
I think the existing logic should be able to be generalized -- basically prim.TupleIndex'es of prim.TupleConstruct ops just create additional aliases.
Correct. I suspect that the OPT model needs to be wrapped in Python to just return the logits or something. Usually these models return a bunch of information (various "scores"...
Can you just do `return self.forward()[0]` on the result of the model? Or `self.forward()[0][0]` or something.
I don't think it is nested that deep. I only see 3 levels of nesting in the IR snippet you presented. And the model should have a stable API in...
What is self.forward()[0] -- logits? token ids?
We've changed a lot about the LTC build. I'm closing this issue for now. Please file a new issue if you run into a new issue.
I would recommend that we decompose the op -- @vivekkhandelwal1 do you know if there is a decomposition in functorch for this?
@vivekkhandelwal1 -- can you point MissyDu how to add the decomposition?
It looks like this was resolved to use a decomposition.