torch_xla nightly installation instruction is causing missing symbols

[tengyifei]

Someone followed the nightly installation instruction and got:

import torch_xla
ImportError: .../python3.10/site-packages/_XLAC.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN5torch4lazy13MetricFnValueEd

Usually, this happens because the C++ parts of torch and torch_xla are built with different C++ ABI flags.

[bhavya01]

I'll take this issue. It's also causing our nightly tests to fail

[bhavya01]

I tried installing the nightly builds triggered by torch_xla

pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.7.0.dev-cp310-cp310-linux_x86_64.whl
pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.7.0.dev-cp310-cp310-linux_x86_64.whl

This doesn't give me a missing symbol error.

This indicates that upstream pytorch might have changed the way they build their wheels. Going to look at that next.

[bhavya01]

Torch wheels are compiled with C++11 ABI by default now. For the latest nightly wheel installed using pip3 install --user --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cpu

Python 3.10.16 (main, Dec 11 2024, 16:24:50) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.compiled_with_cxx11_abi()
True

I will update torch_xla scripts to just generate one nightly wheel built with c++11 ABI