Install/Setup build does not work with CUDA 11.8 #127
Comments
|
Hi, it works with CUDA 11.8. My current setup:
gcc version 9.4.0 (Ubuntu 9.4.0-1ubuntu1~20.04.1) |
|
This is strange then, because your setup seems to be exactly similar to mine, and yet I'm not able to install/compile the package.. |
|
have you set the CUDA_HOME environment variable? can you please type the following and tell me what is the output? |
|
Hi, I just added the following lines to my .bashrc file: These are the post-installation actions specified at NVIDIA docs. About CUDA_HOME variable, this one is imported by torch here, so if your CUDA_HOME variable is empty is a matter of torch-CUDA communication. To sum up, by adding aforementioned lines to your .bashrc file (and rebooting your computer or reloading terminal) it should work. If not, check torch-CUDA interaction. PS: find attached a screenshot of my CUDA_HOME env var: |
|
Hi @dankresio, I'm accessing the PC through SSH and its a remote linux server. I was also exploring the code where CUDA_HOME variable is being set, and the value of that variable is nvcc seems to be working just fine for me, so I'm not sure if its any issue with CUDA. Further, I always use GPU in pytorch code and that also works fine. PyTorch was also installed properly, its just the issues with compilation of this specific project. Attached screenshot of the code output which sets CUDA_HOME I will try to check if I'm able to find the path for CUDA as obtained by you |
|
Hi! I think that the problem is in your torch versions. Last, try the following:
That should work. PS: check if your |
|
Hello, I tried the steps which you have mentioned with python 3.10, and 3.11, but it did not work unfortunately. I had also added the flag After doing that, I got this new error saying it is not able to find 'thrust/complex.h' file I think maybe some issues with the CUDA installation in my machine? But it's strange that PyTorch works just fine and I'm able to train models on GPU on these machines. Perhaps its only issues with the compilation... It would be nice if authors could provide pre-compiled version of this work just like other popular packages and libraries. I might then be able to use it easily. For now I think I may have to use older CUDA (11.6) and older PyTorch on my machine to get this to work... |
Sorry not this reason, I meet this for “subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.”. But after I modify by ↑, still not working. My root cause is "nvcc fatal : Could not open output file {tmpdir}", which {tmpdir} is from $TMPDIR, and it's not exists. After some mkdir it works.... And your root cause is "ERROR: No supported gcc/g++ host compiler found.". |
|
I noticed a similar issue with cuda 11.6 as well, I received the following issue |
|
@aradhyamathur I think it has to do with CUDA/compiler installation like gcc/g++ in your machine. Since I don't have sudo access to my machine I just went with pip installation of this library which is already pre-compiled. |
|
I use WSL2 with windows 10. My cuda version is 11.8, and my g++ version is 11.4. It works for me |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I have created a brand new Conda environment with
python=3.10When using CUDA 11.8:
and installing the latest PyTorch version
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidiaThe setup process does not work and i get the following error:
ERROR: No supported gcc/g++ host compiler found.
snippet:
My GCC version is:
gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0and it seems to work properly.
I modified extra_compile_args in setup.py
extra_compile_args = {"cxx": [], "nvcc": ["-ccbin", "/usr/bin/gcc"]}But still it did not help.
However using CUDA 11.6 and older PyTorch version:
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidiaseems to work and the setup build is successful.
Requesting you to please look into the error and fix it.
I know there is a Google Colab notebook for inference and demo, but I have specific use cases to run this in my own machine where I currently have access and not through Google Colab.
The text was updated successfully, but these errors were encountered: