PyTorch CUDA extensions break during compilation with CUDA10.2.89

[ptrblck]

🐛 Bug

Currently CUDA extensions in PyTorch break using nvcc 10.2.89 with the following error:

/workspace/src/pytorch/torch/include/c10/util/variant.h: In function ‘constexpr decltype(auto) c10::visit(Visitor&&, Vs&& ...)’:
/workspace/src/pytorch/torch/include/c10/util/variant.h:2718:97: error: parameter packs not expanded with ‘...’:
     return (detail_::all({!vs.valueless_by_exception()...})

After debugging this issue, it appears to be related to an nvcc issue in CUDA version 10.2.89. As we are working on fixing this issue, it is recommended for those who compile CUDA extensions to use CUDA 10.1.168.

To Reproduce

Steps to reproduce the behavior:

# PyTorch compiled with CUDA10.2.89
git clone https://github.com/pytorch/extension-cpp
cd extension-cpp/cuda
python setup.py install

CC @csarofeen @ngimel

cc @ezyang @gchanan @zou3519 @yf225 @ngimel

[csverma610]

Hello,

I am confused. Should we install only Cuda10.1 until the issues mentioned are resolved. Somehow
installing pytorch from the source is non-trivial.

[ptrblck]

@csverma610
If you are planning on building custom CUDA extensions, then we would recommend to stick to CUDA10.1 until the PR gets merged.
Otherwise you could build PyTorch from source using CUDA10.2.

If you are stuck in the process of building, I would recommend to create a topic in our discussion board and tag me there.

[zhangguanheng66]

Thanks for the post. Will keep monitoring.