fix(build): update CUDA linking strategy for Windows. #757
Conversation
This build script is modified to properly handle platform-specific CUDA linking: On Windows: NVIDIA’s CUDA SDK does not ship culibos.lib, and static libraries like cublas_static.lib are typically unavailable. Attempting to statically link cudart_static on Windows will fail due to missing symbols provided by culibos. Therefore, on Windows, dynamic linking is preferred. The .lib files (like cudart.lib) act as import libraries for corresponding .dll files (e.g., cudart64_*.dll).
|
Hi! I just encountered this exact issue while trying to build with CUDA on Windows. The culibos library is indeed Linux-specific and causes the build to fail on Windows. I've tested the proposed changes locally and can confirm they resolve the issue perfectly. This fix would be really valuable for Windows users wanting to use CUDA acceleration. Is there anything blocking this PR from being merged? Happy to help with testing or provide additional feedback if needed. Thanks for all the great work on this project! |
|
I'd like to see it rebased on top of main because there was a resonably sized version bump of upstream that just landed. Otherwise this looks good to me |
|
it would be awesome if you could get the windows check passing - that would make it a lot eaiser to ensure we don't break it in the future (or at least know what broke it :) |
This build script is modified to properly handle platform-specific CUDA linking:
On Windows:
NVIDIA’s CUDA SDK does not ship culibos.lib, and static libraries like cublas_static.lib are typically unavailable.
Attempting to statically link cudart_static on Windows will fail due to missing symbols provided by culibos.
Therefore, on Windows, dynamic linking is preferred. The .lib files (like cudart.lib) act as import libraries for corresponding .dll files (e.g., cudart64_*.dll).