add cuda 13.2 to jetson arm64 builds (main)#239
Conversation
There was a problem hiding this comment.
Pull request overview
Updates the self-hosted GitHub Actions CMake workflows to expand CUDA coverage (notably adding CUDA 13.2 to Jetson/ARM64 builds) and to make runner paths less user-specific.
Changes:
- Extend the ARM64 matrix to build/test with CUDA 12.9 and 13.2 for both GCC and Clang host compilers.
- Add an ARM64 workflow LD_LIBRARY_PATH override intended to support CUDA 13.2 compatibility libraries.
- Replace hard-coded
/home/<user>/...CMake path usage with$HOME/...in CI workflows.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.
| File | Description |
|---|---|
| .github/workflows/cmake-linux-arm64.yml | Adds CUDA 13.2 to the ARM64 matrix and introduces CUDA-dependent LD_LIBRARY_PATH / PATH environment setup. |
| .github/workflows/cmake-linux-amd64.yml | Replaces a hard-coded home directory path with $HOME for CMake on x86_64 self-hosted runners. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| echo "CXX=${{matrix.host_compiler}}" >> "$GITHUB_ENV" | ||
|
|
||
| echo "CXX=${{matrix.host_compiler}}" >> "$GITHUB_ENV" | ||
| echo "LD_LIBRARY_PATH=/usr/local/${{matrix.cuda_ldpath}}:$LD_LIBRARY_PATH" >> "$GITHUB_ENV" |
There was a problem hiding this comment.
Setting LD_LIBRARY_PATH as "/usr/local/...:$LD_LIBRARY_PATH" can leave a trailing ':' when LD_LIBRARY_PATH is empty/unset, which causes the dynamic linker to also search the current working directory. Please adjust the append logic so an empty LD_LIBRARY_PATH does not introduce an empty path element (for safety and determinism).
| echo "LD_LIBRARY_PATH=/usr/local/${{matrix.cuda_ldpath}}:$LD_LIBRARY_PATH" >> "$GITHUB_ENV" | |
| echo "LD_LIBRARY_PATH=/usr/local/${{matrix.cuda_ldpath}}${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" >> "$GITHUB_ENV" |
| - host_compiler: "g++-11" | ||
| cuda_toolkit: "12.9" | ||
| cuda_ldpath: "cuda-12.9" |
There was a problem hiding this comment.
The matrix value cuda_ldpath: "cuda-12.9" expands to /usr/local/cuda-12.9 in LD_LIBRARY_PATH, but CUDA shared libraries are typically under a lib directory (e.g., targets/aarch64-linux/lib / lib64) rather than the toolkit root. If this variable is meant to make runtime linking work, it should point at the directory that actually contains the .so files.
albertandaluz
changed the title
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| - host_compiler: "g++-11" | ||
| cuda_toolkit: "13.2" | ||
| cuda_ldpath: "cuda-13.2/compat_orin" | ||
| - host_compiler: "clang++-21" | ||
| cuda_toolkit: "13.2" | ||
| cuda_ldpath: "cuda-13.2/compat_orin" |
There was a problem hiding this comment.
cuda_ldpath values point at CUDA install roots (e.g. /usr/local/cuda-12.9 and /usr/local/cuda-13.2/compat_orin). LD_LIBRARY_PATH needs directories that directly contain the shared libraries (commonly .../lib64 or .../targets/aarch64-linux/lib, and for the compatibility package likely a .../lib subdir). As-is, the dynamic loader won’t search subdirectories, so runtime/test steps may fail to locate CUDA libs. Consider changing cuda_ldpath to the actual library directory paths for each toolkit/compat combo.
morousg
deleted the
238-add-cuda-132-arm64-visa-cuda-compat-oin-main-branch
branch
same as LTS, we add cuda 13.2 to jerson arm64 builds using the compatiblity package until jetpack 7.x is released.