Jason Ansel
Jason Ansel
Can we just update the meta function of torch__dynamo_create_parameter_op_tracable_create_parameter to return an alias? If I recall I made it lie and say it didn't alias to workaround functionalization bugs.
> The obvious solution is to keep set_() in the graph, and require that compiler backends know how to implement it. This seems... probably ok? Another option might be to...
Getting the C++ (CPU) backend for `torch.compile` working on Windows should be straightforward. I suspect it is mostly updating the build commands in https://github.com/pytorch/pytorch/blob/main/torch/_inductor/codecache.py -- and might be a good...
cc @bertmaher
https://github.com/pytorch/pytorch/pull/123318 might fix the issue with EasyDict
#123318 hasn't been merged yet, so you will need to build PyTorch from source to get it and cherry-pick that PR. Assuming it passes the tests I plan to merge...
Yes a few days after it is merged it will be in the nightly build
Not sure, you can do `python -c "import torch; print(torch.version.git_version)"` then looking that version on github to see if it is before or after #123318.
Should we delete the special einops handling from `torch._dynamo`?
Can you check if this fixes it: https://github.com/pytorch/pytorch/pull/124084