`cuda.h` missing error with `torch.compile`

1

Hello, I am using PyTorch 2.0 and having issues with torch.compile

I have installed the nightly build of torch using the instructions on the pytorch website with CUDA 11.7

# pytorch install command
❯ conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch-nightly -c nvidia

# confirming versions
❯ python -V
Python 3.10.8
❯ python -c 'import torch;print(torch.__version__)'
2.0.0.dev20230105
❯ conda list | grep cuda
cuda                      11.7.1                        0    nvidia
...
pytorch                   2.0.0.dev20230105 py3.10_cuda11.7_cudnn8.5.0_0    pytorch-nightly
pytorch-cuda              11.7                 h67b0de4_2    pytorch-nightly

When I try torch.compile I get no issues with the first example provided in the docs

In [1]: import torch
   ...:
   ...: def foo(x, y):
   ...:     a = torch.sin(x)
   ...:     b = torch.cos(x)
   ...:     return a + b
   ...: opt_foo1 = torch.compile(foo)
   ...: print(opt_foo1(torch.randn(10, 10), torch.randn(10, 10)))
# works !

However, trying the next example that uses a resnet18 model on GPU I get many errors of the following form

/tmp/tmpsov1pory/main.c:2:10: fatal error: cuda.h: No such file or directory
 #include "cuda.h"
          ^~~~~~~~
compilation terminated.

This message prints ~ 30 times listing different tmp directories.

Is there a certain environment variable that needs to be set to allow torch to find cuda.h. I tried running find /path/to/my/conda/env -name cuda.h and adding the directories to my LD_LIBRARY_PATH but still got the same error.

Please let me know if anyone can suggest what to troubleshoot next.

Thank you

2

Thanks for reporting as this issue should not happen. Could you try to use CUDA_HOME=/path/to/cuda python script.py args as described here, please?

3

I encounter the same problem, and I solved it by setting the environment variable $C_INCLUDE_PATH

The reason for this problem is that I am using HPC cluster of my institution, and they install cuda in a weird place. Therefore in order for the GCC compiler to find cuda.h, I need to manually include the path to $C_INCLUDE_PATH

You can locate where you have your cuda installed by which nvcc. Mine is /sw/cuda/11.8.0/bin/nvcc, therefore the path I am looking for is /sw/cuda/11.8.0/include

Then type in your terminal export C_INCLUDE_PATH=$C_INCLUDE_PATH:/the/path/to/library/include. Note that the path should end with include.

This solves my problem.