Suppress C++ stacktrace on XLA_CHECK*() calls.

[ysiraichi]

This PR improves error messages in PyTorch/XLA by suppressing the display of C++ stack traces during XLA check failures, making them more user-friendly. Currently, when XLA_CHECK*() fails, the resulting error output includes a lengthy and verbose C++ stacktrace. While these can be useful for deep-dive debugging by developers, they often add noise for end-users.

Key Changes:

• Suppression of C++ stacktraces for all XLA_CHECK*() failures
• Add source location information if XLA_SHOW_CPP_ERROR_CONTEXT=1
  • Following the changes in Error Handling: refactor XlaCoordinator to use status types. #9386

Before:

Traceback (most recent call last):
  File "dot.py", line 6, in <module>
    torch.dot(a, b)
RuntimeError: torch_xla/csrc/aten_xla_bridge.cpp:110 : Check failed: xtensor
*** Begin stack trace ***
        tsl::CurrentStackTrace[abi:cxx11]()
        torch_xla::bridge::GetXlaTensor(at::Tensor const&)
        torch_xla::XLANativeFunctions::dot(at::Tensor const&, at::Tensor const&)

        c10::BoxedKernel::callBoxed(c10::OperatorHandle const&, c10::DispatchKeySet, std::vector<c10::IValue, std::allocator<c10::IValue> >*) const
        c10::KernelFunction::callBoxed(c10::OperatorHandle const&, c10::DispatchKeySet, std::vector<c10::IValue, std::allocator<c10::IValue> >*) const
        c10::Dispatcher::callBoxed(c10::OperatorHandle const&, std::vector<c10::IValue, std::allocator<c10::IValue> >*) const



        c10::BoxedKernel::callBoxed(c10::OperatorHandle const&, c10::DispatchKeySet, std::vector<c10::IValue, std::allocator<c10::IValue> >*) const


        at::_ops::dot::redispatch(c10::DispatchKeySet, at::Tensor const&, at::Tensor const&)





        at::_ops::dot::call(at::Tensor const&, at::Tensor const&)
        at::Tensor::dot(at::Tensor const&) const



        _PyObject_MakeTpCall
        _PyEval_EvalFrameDefault

        PyEval_EvalCode



        _PyRun_SimpleFileObject
        _PyRun_AnyFileObject
        Py_RunMain
        Py_BytesMain
        __libc_start_main
        _start
*** End stack trace ***
Input tensor is not an XLA tensor: torch.FloatTensor

After:

Traceback (most recent call last):
  File "dot.py", line 6, in <module>
    torch.dot(a, b)
RuntimeError: Check failed: xtensor: Input tensor is not an XLA tensor: torch.FloatTensor

[zhanyong-wan]

Thanks!

[zhanyong-wan]

Can we include the C++ stack trace when this is true?

[zhanyong-wan]

Style: [[nodiscard]]

[zhanyong-wan]

Document how these macros react to XLA_SHOW_CPP_ERROR_CONTEXT?

[zhanyong-wan]

Add tests to verify the new behavior?