No longer able to build in Docker

[iarspider]

Hi!

We were able to build torchvision 0.3.0 w/cuda in docker container, which doesn't have access to NVidia GPU. However when building 0.4.1 we get the following error:

AssertionError: 
Found no NVIDIA driver on your system. Please check that you
have an NVIDIA GPU and installed a driver from
http://www.nvidia.com/Download/index.aspx

Is there a way to build torchvision without "NVidia driver"?

[iarspider]

This is how we build Torch:

export USE_NINJA=1
export USE_MKLDNN=0
export CUDA_HOME=/path/to/cuda
python setup.py install --prefix=/install/torch

This is how we build Torchvision:

export LD_LIBRARY_PATH=/install/torch/lib/pythonX.Y/site-packages/torch/lib
export FORCE_CUDA=1
export CUDA_HOME=/path/to/cuda
python setup.py install --prefix=/install/torchvision

[fmassa]

Hi,

I don't have much experience with docker, and I'm not what has changed that could break things for you compared to 0.3.0.

Can you try building torchvision from master branch?

[iarspider]

Hi,

same problem with "master" (commit 681c6c1) version of torchvision.

[ahirner]

Solved by supplying TORCH_CUDA_ARCH_LIST, otherwise cpp_extension.py tries to dynamically determine ARCH. See trace if TORCH_CUDA_ARCH_LIST is not supplied for 0.4.2 and pytorch 1.3.1.

You may want to check out our custom builds here.

[fmassa]

@ahirner can you send a PR fixing the docker in torchvision?

Also, cc @seemethere for awareness

[iarspider]

Worked for me, thanks @ahirner !

[ahirner]

I think the error only pertains custom builds not current wheels? We don't use them either so it would be good to repro the case with official builds and open a new issue if it is @fmassa