CI Build Failing due to PyTorch `fmtlib`.

[ysiraichi]

🐛 Bug

CI build failure found in: https://github.com/pytorch/xla/actions/runs/17946995075/job/51035845884?pr=9652

    In file included from bazel-out/k8-opt/bin/external/torch/_virtual_includes/headers/torch/csrc/jit/python/pybind_utils.h:27,
                     from bazel-out/k8-opt/bin/external/torch/_virtual_includes/headers/torch/csrc/jit/python/pybind.h:12,
                     from ./torch_xla/csrc/xla_sharding_util.h:4,
                     from torch_xla/csrc/aten_xla_type.cpp:56:
    bazel-out/k8-opt/bin/external/torch/_virtual_includes/headers/torch/csrc/utils/python_arg_parser.h:42:10: fatal error: fmt/format.h: No such file or directory
       42 | #include <fmt/format.h>
          |          ^~~~~~~~~~~~~~
    compilation terminated.

Maybe it's time to pin PyTorch.

[ysiraichi]

@bhavya01 I'm currently working on this. Here's my plan:

1. Create a PyTorch commit pin, and have it point to a somewhere before test_aten_xla_tensor_1 SegFaulting #9651 started occurring
   • This should also take care of this specific build issue
2. Bump the PyTorch pin to the main branch whenever those issues are solved

Let me know if you have any suggestions.