Ramiro Leal-Cavazos

Performance issue with `torch.linear` op

3

### Request description I am currently benchmarking a Bert model, and I noticed that the time it takes for the model to run the PyTorch op [`torch.linear`](https://pytorch.org/docs/stable/generated/torch.nn.functional.linear.html?highlight=linear#torch.nn.functional.linear) with IREE+CUDA is...

Replace `mean_dim` with `sum_mean_dim` in shape lib + fix `arange`

2

PyTorch recently renamed the shape function `mean_dim` with the name `sum_mean_dim`, and they added the right handling of the optional `dim` argument (https://github.com/pytorch/pytorch/commit/c8eae2de5200b95d59ab090a37eed96528c877ff). This commit updates the shape functions that...

Add support for returning more than one copy of the same tensor

3

One of the simplifications made by the pass `RefinePublicReturn` currently only happens if the tensor in question only has one user. However, the current method of checking this does not...

If the [input to the e2e test for `gelu`](https://github.com/llvm/torch-mlir/blob/bdb5083d3338b73a9d5c5ce2266c6fd0fe822c31/python/torch_mlir_e2e_test/test_suite/elementwise.py#L409) is changed to use a distribution with values in range `[-1, 0]`, the test fails when using the TOSA backend. ```...

Below is an example of a test that would pass in torch-mlir despite having an annotation that does not match the input tensors. Namely, the size and dtype of the...

The op `IndexPutHackedTwinOp` takes a [list of index tensors](https://github.com/llvm/torch-mlir/blob/acd57a352033f3c8315847e6d60b1fbb4188c44c/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td#L5843) as an argument, not a list of optional tensors. This means that all the logic in the lowering checking if indices...