FindCUDA says CUDA version is 12.8, but the CUDA headers say the version is ERROR: ld.so: object '/usr/lib/x86_64-linux-gnu/libjemalloc.so.2' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.

TL;DR

Update the LD_PRELOAD path in .ci/pytorch/build.sh to use a portable lookup for the jemalloc library location.

Guidance

• Identify the correct path to the jemalloc library on non-Debian distributions (e.g., /usr/lib64/libjemalloc.so.2 on RHEL, CentOS, Fedora) and update the LD_PRELOAD path accordingly.
• Consider using a platform-agnostic approach to determine the jemalloc library location, such as using the ldconfig command or a similar mechanism.
• Verify that the updated LD_PRELOAD path resolves the build failures and prevents the LD_PRELOAD error message from corrupting cmake output parsing.
• Test the updated build.sh script on multiple Linux distributions to ensure compatibility.

Example

# Example of using ldconfig to find the jemalloc library location
export LD_PRELOAD=$(ldconfig -p | grep libjemalloc.so | awk '{print $4}')

Note: This example is hypothetical and may not work as-is, but it illustrates the idea of using a platform-agnostic approach to determine the jemalloc library location.

Notes

The proposed solution assumes that the jemalloc library is installed and available on the system. If the library is not installed, additional steps may be necessary to install it.

Recommendation

Apply a workaround by updating the LD_PRELOAD path to use a portable lookup, as this will allow the build script to work across multiple Linux distributions without requiring platform-specific checks.