nikitaved
nikitaved
The issue with distributed seems unrelated - it fails for me either way
It operates over a callable. Does it really make sense to have it as a transform?
I would like to keep it to be able to see whether there is benefit to it after all. It also restores the parity with thunder.compile we had before.
No additional safety, I got it wrong. What I mean is that we can indeed make it a transform to potentially decide which parts are safe to capture.
@tfogal , the executor is in good shape, just not complete. This means, there is no "advanced" logic on handling data-dep operations and fusion regions between graph breaks (with dynamic...
OK, looks like `max_pool_with_indices` comes from https://github.com/Lightning-AI/lightning-thunder/pull/163. max_pool without indices has a well-tested meta-function, and it could be re-used here.
As of now, we cannot support data-dependent ops, alas...
@carmocca , looking at the code I think the solution could be modifying the model in the package. The result of `topk` can be sorted, and then we do not...
@IvanYashchuk , looks like we should update the meta function for `where`. To be frank, I did not even know about this overload... Might be a very nice issue for...
Hey, @MYT677 , could you please provide your environment specifics with the following? I am getting different numbers. ``` wget https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py # For security purposes, please check the contents of...