cudaPackages.cuda_nvcc: patch math_functions.h signatures#484031
Conversation
|
|
This indeed fixes |
|
Remind me, does this affect all versions of CUDA? If that includes CUDA 13.0, do we need to package cuda_crt instead of cuda_nvcc (I can't remember if cuda_crt is the one that provides the CRT headers, if we copy them into cuda_nvcc, or something else). |
|
As a followup, do we need to worry about backporting these changes? I remember GLIBC was bumped to fix several vulnerabilities -- not sure if patches or the version bump would be backported. |
Seems like |
|
Ah this was a simpler diff than I thought, shouldn't have blocked. That said, |
Fix compatibility with glibc 2.42
0fcee74 to
eb6a294
Compare
Done! |
| # The cospi|sinpi|rsqrt function signatures in include/common/math_functions.h do not match | ||
| # glibc 2.42's. | ||
| # Indeed, there they are declared with noexcept(true) which is not the case in cuda_nvcc. | ||
| + lib.optionals (lib.versionAtLeast glibc.version "2.42") '' |
There was a problem hiding this comment.
Ugh kill me
Can you make a hotfix and I’ll merge it
Things done
Fix compatibility with glibc 2.42
The latest
glibc(2.42) brought by the lateststaging-nextiteration does not play well with cuda 12.8.This at least breaks
bitsandbytes(#483507).The
cospi|sinpi|rsqrtfunction signatures ininclude/common/math_functions.hdo not match glibc 2.42's.Indeed, there they are declared with
noexcept(true)which is not the case incuda_nvcc.This PR patches
math_functions.hincuda_nvcc's output to fix the issue.cc @NixOS/cuda-maintainers @kirillrdy
passthru.tests.nixpkgs-reviewon this PR. See nixpkgs-review usage../result/bin/.