pytorch - 💡(How to fix) Fix OSS CUDA13 CUDSS missing libs, not properly compiled on releases [1 comments, 1 participants]

Fix Plan

To fix the issue of OSS builds not linking against CUDSS in CUDA13, we need to install CUDSS and configure the build to link against it. Here are the steps:

• Install CUDSS:
  • Download the CUDSS library from the official NVIDIA website.
  • Follow the installation instructions to install the library on your system.
• Configure the build to link against CUDSS:
  • Set the CUDSS_LIBRARY_PATH and CUDSS_INCLUDE_PATH environment variables to point to the installed CUDSS library and include files.
  • Update the CMakeLists.txt file to include the CUDSS library in the build.

Example code snippet to update CMakeLists.txt:

find_package(CUDSS REQUIRED)

add_library(${PROJECT_NAME} ${SOURCES})
target_link_libraries(${PROJECT_NAME} ${CUDSS_LIBRARY})
target_include_directories(${PROJECT_NAME} PUBLIC ${CUDSS_INCLUDE_DIR})

• To dynamically link against CUDSS using pypi wheels, you can use the following code snippet:

import torch

# Install the CUDSS wheel
torch.utils.install_cudss()

# Verify that CUDSS is installed and linked correctly
print(torch.cuda.get_device_name(0))

Verification

To verify that the fix worked, you can check the build logs to ensure that the CUDSS library is being linked correctly. You can also run the following command to verify that the CUDSS library is installed and linked correctly:

cmake --build .

If the build is successful, you should see the CUDSS library being linked against the OSS build.

Extra Tips

• Make sure to install the correct version of CUDSS that matches your CUDA version.
• If you are using a virtual environment, make sure to activate it before installing CUDSS.
• You can also use the torch.utils.install_cudss() function to install CUDSS programmatically.