Bump torch-mlir to 98e08023bbf71e00ab81e980eac9f7c96f1f24b4

[jinchen62]

No description provided.

[ScottTodd]

llvm/torch-mlir@fa39d91 included here seems to have broken some downstream usage of the IREE Python bindings: https://github.com/iree-org/iree-turbine/actions/runs/10617974632/job/29432221240?pr=112#step:5:204

We may be missing some test coverage in this repo.

Can someone verify / repro, then decide if we should revert or fix forward?

[ScottTodd]

I can repro with just IREE:

1. Build from source: https://iree.dev/building-from-source/getting-started/#building-with-cmake
2. Setup venv and install some deps:

   python -m venv .venv
   .venv\Scripts\activate.bat
   D:\dev\projects\iree-build\.env.bat
   python -m pip install onnx torch sympy

3. Try running

   >>> from iree.compiler.extras.fx_importer import ContextCache
   
   Traceback (most recent call last):  File "<stdin>", line 1, in <module>
     File "D:\dev\projects\iree-build\compiler\bindings\python\iree\compiler\extras\fx_importer.py", line 38, in <module>
       import torch
     File "D:\dev\scratch\iree\.venv\Lib\site-packages\torch\__init__.py", line 2120, in <module>
       from torch._higher_order_ops import cond
     File "D:\dev\scratch\iree\.venv\Lib\site-packages\torch\_higher_order_ops\__init__.py", line 1, in <module>
       from .cond import cond
     File "D:\dev\scratch\iree\.venv\Lib\site-packages\torch\_higher_order_ops\cond.py", line 5, in <module>
       import torch._subclasses.functional_tensor
     File "D:\dev\scratch\iree\.venv\Lib\site-packages\torch\_subclasses\functional_tensor.py", line 42, in <module>
       class FunctionalTensor(torch.Tensor):
     File "D:\dev\scratch\iree\.venv\Lib\site-packages\torch\_subclasses\functional_tensor.py", line 258, in FunctionalTensor
       cpu = _conversion_method_template(device=torch.device("cpu"))
   D:\dev\scratch\iree\.venv\Lib\site-packages\torch\_subclasses\functional_tensor.py:258: UserWarning: Failed to initialize NumPy: _ARRAY_API not found (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\utils\tensor_numpy.cpp:84.)
     cpu = _conversion_method_template(device=torch.device("cpu"))
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
     File "D:\dev\projects\iree-build\compiler\bindings\python\iree\compiler\extras\fx_importer.py", line 138, in <module>
       from .._mlir_libs._torchMlir import get_int64_max, get_int64_min
   ModuleNotFoundError: No module named 'iree.compiler._mlir_libs._torchMlir'

[ScottTodd]

I can't repro from running https://github.com/iree-org/iree/blob/main/compiler/bindings/python/test/extras/fx_importer_test.py though.

[ScottTodd]

Ah. https://github.com/iree-org/iree/blob/main/compiler/bindings/python/test/extras/fx_importer_test.py does catch this, if torch is installed. Should get that test really running, together with #16624. Maybe a new workflow similar to https://github.com/iree-org/iree/blob/main/.github/workflows/pkgci_test_tensorflow_cpu.yml focused on torch.

[ScottTodd]

Adding / enabling test coverage that would have caught this in #18420

[MaheshRavishankar]

I'd say we revert this change.

[jinchen62]

@MaheshRavishankar @ScottTodd So should I revert this PR or revert that torch-mlir commit?

[ScottTodd]

@MaheshRavishankar @ScottTodd So should I revert this PR or revert that torch-mlir commit?

Added a comment to the upstream PR. Still seeing if we can work around it locally in the meantime.

[ScottTodd]

But while we wait on movement upstream (or contribute a patch ourselves), I vote we revert this since it breaks some usage of IREE's Python APIs that iree-turbine depends on.