Cannot build PyTorch 2.3.1 with CUDA 12.9 on WSL2 (Ubuntu 22.04) — Missing nvToolsExt despite cuda-nvtx-12-9 installed

1

Hi all,

I’m trying to build PyTorch v2.3.1 from source on a WSL2 Ubuntu 22.04 environment with the following setup:
Any way to build pytorch 2.3.1 with cuda 12.9?
Cause it seems latest pytorch incompatible with some modules(TI QAT wrapper module) I need to use in my project, I need pytorch 2.3.1.

:white_check_mark: System Configuration:

  • OS: WSL2 (Ubuntu 22.04)
  • GPU: NVIDIA RTX 5080
  • CUDA Driver Version: 12.9
  • CUDA Toolkit Version: 12.9 (installed via APT)
  • cuDNN: libcudnn9 installed via APT
  • MAGMA: magma-cuda121 installed via conda
  • Python: Conda environment, Python 3.10
  • GCC: 11.4

Environment variables:

export CUDA_HOME=/usr/local/cuda-12.9
export _GLIBCXX_USE_CXX11_ABI=1
export USE_CUDA=1
export USE_CUDNN=1
export USE_MKLDNN=1
export USE_MAGMA=1
export USE_NCCL=0
export MAX_JOBS=$(nproc)
export _GLIBCXX_USE_CXX11_ABI=1
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}

:warning: Problem:

While building PyTorch 2.3.1 using python setup.py develop, the process fails with the following error:

CMake Error at cmake/public/cuda.cmake:70 (message):
  Failed to find nvToolsExt
2

No, I would not recommend trying to build old PyTorch releases with newer CUDA toolkits as needed changes might not have landed as can be seen in your error message.

3

Ok, it is very clear.