Ramiro Leal-Cavazos comments

Results 77 comments of


                                            Ramiro Leal-Cavazos

'unsupported by backend contract: non-value tensor type' with torch.zeros

Hi @JBloodless, do you have an example of calling how you're calling `pad_packed_squence` and compiling it with torch-mlir? I can take a look. You're right that `torch.zeros` should work.

'unsupported by backend contract: non-value tensor type' with torch.zeros

Yeah, the mutation case here is not currently supported by maximize value semantics. Adding support for it might require a bit of work. A way to circumvent this is to...

torch-mlir install requirements has conflict

What command are you using to install torch-mlir?

torch-mlir install requirements has conflict

The `requirements.txt` file should not be installing torch-mlir, AFAICT. The command to pip install should be the one from the README: ``` pip install --pre torch-mlir torchvision \ -f https://llvm.github.io/torch-mlir/package-index/...

Add torch macos universal whl files to release assets and package-index

@powderluv, PTAL

[fx]add all_reduce test.

I don't think I've ever seen this particular issue. We do have a place where we check the PyTorch version because of differences in ops supported: https://github.com/llvm/torch-mlir/blob/77ae56337dbf95eb809f4a7d218a9fb3dc1f41b0/projects/pt1/python/torch_mlir/dynamo.py#L69 but the issue...

[fx]add all_reduce test.

> It seems that torch.export always generate op with signature c10d_functional::all_reduce : (Tensor, str, str, int[], int) -> (Tensor) even on nightly-build torch. Interesting. Yeah, your proposed solution seems fine...