I tried your patch, but unfortunately the build did not go through. Here is the tail of the build:

[ 79%] Building CXX object caffe2/CMakeFiles/caffe2_gpu.dir/queue/queue_ops_gpu.cc.o
[ 80%] Building CXX object caffe2/CMakeFiles/caffe2_gpu.dir/sgd/iter_op_gpu.cc.o
[ 80%] Building CXX object caffe2/CMakeFiles/caffe2_gpu.dir/sgd/learning_rate_op_gpu.cc.o
[ 80%] Linking CXX shared library ../lib/libcaffe2_gpu.so
impure path `/usr/local/cuda/lib/libcudnn.so.7' used in link
collect2: error: ld returned 1 exit status
make[2]: *** [caffe2/CMakeFiles/caffe2_gpu.dir/build.make:4687: lib/libcaffe2_gpu.so] Error 1
make[1]: *** [CMakeFiles/Makefile2:5422: caffe2/CMakeFiles/caffe2_gpu.dir/all] Error 2
make: *** [Makefile:141: all] Error 2
setup.py::build_deps::run()
Failed to run 'bash ../tools/build_pytorch_libs.sh --use-cuda --use-nnpack --use-mkldnn --use-qnnpack caffe2'
builder for '/nix/store/i5n4iqqk9kzkhhn4a8v4bnm3b00f3k64-python3.7-pytorch-1.0.0.drv' failed with exit code 1
cannot build derivation '/nix/store/5mhc9w2bi01f0p1jpb0j8fc9hp76pziy-python3-3.7.3-env.drv': 1 dependencies couldn't be built
error: build of '/nix/store/5mhc9w2bi01f0p1jpb0j8fc9hp76pziy-python3-3.7.3-env.drv' failed

@jyp It sounds like you’ve run into a different problem: the Nix package expects to use the packaged CUDA and cuDNN in /nix/store, but you have a local copy installed in /usr/local/cuda. Can you provide some more context? Are you using NixOS or something else? Why do you have a /usr/local/cuda, and does it work if you remove that? What happens without the patch? What happens with only the first hunk of the patch (i.e. add cudatoolkit_joined to nativeBuildInputs without removing it from buildInputs)?

Clearly sandboxing was disabled in @jyp's build.

Indeed it works using sandbox. (I was confused by #51671)

I'm here after reading this and a few related issues (eg. #51671). Here is my shell.nix -- on first dropping into the this shell, PyTorch was built form source:

with import <nixpkgs> {};

let
  py = pkgs.python37;
in
stdenv.mkDerivation rec {
  name = "python-environment";

  buildInputs = [
    py
    py.pkgs.matplotlib
    py.pkgs.tkinter
    py.pkgs.numpy
    py.pkgs.pytorchWithCuda
  ];
}

However, during the build, there were several messages like "CUDA not available, skipping tests". Maybe relatedly, there were errors like "Error in cpuinfo: failed to parse the list of present procesors in /sys/devices/system/cpu/present". I can provide a full log if need be.

Additional outputs:

$LD_LIBRARY_PATH=/run/opengl-driver/lib

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 415.27       Driver Version: 415.27       CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+

In a python session:

>>> import torch
>>> torch.cuda.is_available()
False
>>>

I'm still pretty new to nix, so I might be doing something wrong. Should this have worked/what can I do to fix it?

@mtn These errors were fixed for me when using a sandboxed build.

How can I do this, and can I still use nix-shell? I'm finding this issue hard to parse: #903. Or maybe there's a easier way I should be building and using this? I just want the result to be isolated.

Edit: @jyp I'm running nixos 19.03, so shouldn't sandboxing be happening by default?

I have the same problem as @mtn and I have no idea how to resolve on 19.03

@mtn @fkstef I just ran into this problem and spent a long time trying to solve it. The answer is actually quite simple. This fix has not been backported to 19.03. You can either pin to this fix using pinning or use an overrideAttr directive. I did the second, but the first might be more useful if you're planning on a long term thing.

For instance, here is an example shell.nix

let
  pkgs = import <nixpkgs> {};
  pytorch-cuda = pkgs.python37Packages.pytorchWithCuda.overrideAttrs (oldAttrs: {
    nativeBuildInputs = oldAttrs.nativeBuildInputs ++ [ pkgs.cudatoolkit ];
  });
  python3 = pkgs.python3.withPackages (ps: with ps; [opencv4 numpy pytorch-cuda]);

in pkgs.stdenv.mkDerivation (with pkgs; {
  name = "env";

  buildInputs = [
    python3
  ];

})