Hardcoded lib path in build script causes error

[musl]

Description

Library path on setup_build is hardcoded. Some systems, such as Debian, do not install nvidia libraries to /usr/lib64/ when installing nvidia-cuda-dev and nvidia-cuda-toolkit through apt:

$ find /usr/ -name libcudart_static*
/usr/lib/x86_64-linux-gnu/libcudart_static.a

I solved this issue by doing a symlink:

ln -s /usr/lib/x86_64-linux-gnu/libcuda* /usr/lib64/

However, the correct solution for this issue would be to not hardcode paths or add conditions to it.

To Reproduce

poetry add cupy

On Debian:

$ uname -a
Linux anon 6.6.9-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.6.9-1 (2024-01-01) x86_64 GNU/Linux

Installation

Source (pip install cupy)

Environment

OS                           : Linux-6.6.9-amd64-x86_64-with-glibc2.37
Python Version               : 3.11.7
CuPy Version                 : 13.0.0
CuPy Platform                : NVIDIA CUDA
NumPy Version                : 1.26.3
SciPy Version                : 1.11.4
Cython Build Version         : 0.29.36
Cython Runtime Version       : None
CUDA Root                    : /usr
nvcc PATH                    : /usr/bin/nvcc
CUDA Build Version           : 12020
CUDA Driver Version          : 12000
CUDA Runtime Version         : 12020 (linked to CuPy) / 12000 (locally installed)
cuBLAS Version               : (available)
cuFFT Version                : 11001
cuRAND Version               : 10301
cuSOLVER Version             : (11, 4, 3)
cuSPARSE Version             : (available)
NVRTC Version                : (12, 0)
Thrust Version               : 200200
CUB Build Version            : 200200
Jitify Build Version         : <unknown>
cuDNN Build Version          : (not loaded; try `import cupy.cuda.cudnn` first)
cuDNN Version                : (not loaded; try `import cupy.cuda.cudnn` first)
NCCL Build Version           : None
NCCL Runtime Version         : None
cuTENSOR Version             : None
cuSPARSELt Build Version     : None
Device 0 Name                : NVIDIA GeForce GTX 1050
Device 0 Compute Capability  : 61
Device 0 PCI Bus ID          : 0000:08:00.0

Additional Information

No response

[Leo Fang]

Just curious, was building against distro packages ever work for CuPy? i.e. did CuPy v12 and earlier work with this setup, but v13 does not?

[Kenichi Maehashi]

Maybe I should have let the linker find static libraries by specifying -lcudart_static, instead of discovering a file path on our own?

[Leo Fang]

Not sure, would that work on Windows too?

[Leo Fang]

My question is more about the distro package layout. If we have libs and headers scattered (e.g. /usr/lib/x86_64-linux-gnu/, /usr/include, ...) I suppose the old CUDA_PATH based logic never worked?

[Leo Fang]

Maybe I should have let the linker find static libraries by specifying -lcudart_static

I guess it'd work!

[Kenichi Maehashi]

@muslll, we've just merged the fix #8134. Could you try with the latest main branch?

@leofang

My question is more about the distro package layout. If we have libs and headers scattered (e.g. /usr/lib/x86_64-linux-gnu/, /usr/include, ...) I suppose the old CUDA_PATH based logic never worked?

I guess these distro-default paths are on the compiler/linker search path by default?

[musl]

Thanks @kmaehashi 👍