I am using CUDA.jl on a Pascal (sm_61) hardware.
Till now having driver 570xx (575xx also) worked when restricting cuDNN to v1.4.4 (so nvidia cudnn is <v9.10.0, later Pascal is deprecated). This restricted also CUDA.jl to version 5.8.5. Tying runtime_version to v12.x downloaded the right artifact and it worked.
Now I tried to use the latest driver version to support Pascal, 580.xx, and did the same restrictions as above. This time it crashed already using a minimal example with CuArray. dmesg showed me a segfault in libcuda.so.595.58.03 which points to a driver/artifact-library mismatch.
So my question: is it possible to get an artifact having v580 so this mismatch cannot happen?
julia> CUDA.versioninfo()
CUDA toolchain:
- runtime 12.4, artifact installation
- driver 580.159.4 for 13.2
- compiler 12.9
CUDA libraries:
- CUBLAS: 12.4.5
- CURAND: 10.3.5
- CUFFT: 11.2.1
- CUSOLVER: 11.6.1
- CUSPARSE: 12.3.1
- CUPTI: 2024.1.1 (API 12.4.0)
- NVML: 13.0.0+580.159.4
Julia packages:
- CUDA: 5.8.5
- CUDA_Driver_jll: 13.2.1+0
- CUDA_Compiler_jll: 0.2.2+0
- CUDA_Runtime_jll: 0.19.2+0
Toolchain:
- Julia: 1.12.6
- LLVM: 18.1.7
Preferences:
- CUDA_Runtime_jll.version: 12.4
1 device:
0: Quadro P2000 (sm_61, 3.682 GiB / 5.000 GiB available)