Failed Dispatch of Custom CUDA OP in TorchAO

[leslie-fang-intel]

Install PyTorch 2.6 and TorchAO nightly release and run custom ops' UT in TorchAO on a A100

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126

clear && python -u -m pytest -s -v test/test_ops.py -k test_dequantize_tensor_core_tiled_layout_correctness_quant_dequant

I get the following error

 NotImplementedError: Could not run 'torchao::dequantize_tensor_core_tiled_layout' with arguments from the 'CUDA' backend. 

The same error met, when I build torchAO from src, where I believe csrc/cuda/tensor_core_tiled_layout/tensor_core_tiled_layout.cu has been built from the building log.

[leslie-fang-intel]

It turns out to be there are 2 C extension in my system _C.cpython-310-x86_64-linux-gnu.so and _C.abi3.so. And the initiation load the so without custom kernels. Feels like there might 2 improvement:

1. setup.py clean should remove all the existing so
2. When there are more than one so, we may throw the error instead of using logging.debug

   ao/torchao/__init__.py

   Line 39 in 42e1345

   logging.debug(f"Skipping import of cpp extensions: {e}")

[leslie-fang-intel]

Close it as it relates to my ENV