PyTorch extension generation step fails with nvcc9.1, gcc-4.9.3 to 7.4.0, pytorch-1.0.0 to 1.12 about dain

[laomao0]

You need to check your cuda: compute version.
For my GPU: RTX 2080 Ti, we found my compute version is 7.5, in en.wikipedia.org/wiki/CUDA.

Then you must do the following steps:

1. Do cd DAIN/my_package/xxxx/setup.py
2. modify: nvcc_args = [ '-gencode', 'arch=compute_75,code=sm_75' ]
3. run DAIN/my_package/build.sh

Notice: xxx is the subfolder, you need to modify each one under my_package.

[x2eliah]

Thanks, but that did not solve the issue. (We run 1080Ti's, which are the already-enabled-in-build-scripts 6.1).

Ultimately, this looks like the bug caused by cuda 9.0 / 9.1 inability to compile pytorch extensions due to miscommunication between our nvcc 9.1 and gcc.

The only way we ended up solving it, was to bite the bullet and upgrade the workstation's cuda to 9.2 (including the necessary driver upgrade) and ensuring that nvcc v9.2 was used in the build process.
All I can conclude, from our efforts at resolving this, is that pytorch extension building simply does not function with cuda 9.0 / 9.1, it is mandatory to use 9.2.

I'm closing the issue - since there's nothing the authors can do do resolve this, ultimately - but it would be nice to clarify in the instructions that the dependency is 9.2, not 9.0...