torch.aten.as_strided produces incorrect views, should not be decomposed from torch.chunk

[raayandhar]

From discussion with @zjgarvey, we have the following findings (mostly Zach, huge thanks):

The torch.chunk operation currently relies on aten.as_strided which is overkill for what the op actually does, and it can use .slice instead. However, we also note that aten.as_strided relies on accurate stride input layout, but in torch-mlir, tensors are materialized as contiguous (torch can have non-contiguous tensors, we don't. For example, transpose the input.). Therefore the static accessings args from as_strided refer to data that is no longer in the same format.

aten.as_strided is only generated in decompositions of split.Tensor. We propose that

1. Remove aten.as_strided from all decomposition lists in torch-mlir (all paths reachable post-import) fix: remove split.Tensor and split_with_sizes from decomp tables #4340
2. On the IREE side, we can add the RecomposeComplexOps pass from torch-mlir that will decompose .chunk into .slice instead of as_strided [Codegen] Add RecomposeComplexOps pass in Torch/InputConversion/Passes iree-org/iree#22276
3. Remove split.Tensor and split_with_sizes from IREE-Turbine's decomposition table http://github.com/iree-org/iree-turbine/pull/1175
   Eventually we should probably deprecate the aten.as_strided support and throw an error on input when we see a aten.as_strided op, since we cannot guarantee the correctness.

Related issues:

• (TorchToLinalg) Support for lowering torch.aten.as_strided #4191
• YOLOv12 Compilation from PyTorch iree-org/iree#22022

[zjgarvey]

This looks good. Ping me when you open PR's for these tasks and I can review, etc.

If you run into any trouble, e.g. in torch-mlir if you experience test failures, etc. Let me know. We may need to add recompose to certain test-config-specific pass pipelines.

[zjgarvey]

It may be useful to note:

1. torch.ops.aten.chunk.default gets converted into torch.ops.aten.split.Tensor during exported_program.run_decompositions, even if chunk is not in our decomposition table. As long as the decomposition table being passed is non-empty, it will actually do the graph execution and tracing, and chunk is implemented in c++ as a split.
2. torch.ops.aten.split.Tensor is on many of our default decomposition lists, but the current decomposition for this op converts to as_strided, which is untenable for torch-mlir.