Mismatch between detected runtime version and precompiled toolkit

[randsley]

Description
NVIDIA has officially released CUDA Toolkit 13.1.0 (Dec 2025), which introduces significant changes including the CUDA Tile programming model and support for newer Blackwell architectures. Currently, CUDA.jl appears to be precompiled for/supporting up to 13.0.x, leading to initialization errors when a system-wide 13.1.0 driver/toolkit is detected.

Motivation
Users with the latest NVIDIA drivers or those working on Blackwell-based systems (H200/B200) require the 13.1.0 runtime for full compatibility and to leverage new performance enhancements in libraries like cuBLAS and cuDNN.

Error Observed
Error: You are using CUDA 13.1.0, but CUDA.jl was precompiled for CUDA 13.0.0.

CUDA toolchain:

• runtime 13.1, artifact installation
• driver 591.44.0 for 13.1
• compiler 13.0

CUDA libraries:

• CUBLAS: 13.1.0
• CURAND: 10.4.0
• CUFFT: 12.0.0
• CUSOLVER: 12.0.4
• CUSPARSE: 12.6.3
• CUPTI: 2025.3.1 (API 13.0.1)
• NVML: 13.0.0+591.44

Julia packages:

• CUDA: 5.9.5
• CUDA_Driver_jll: 13.1.0+0
• CUDA_Compiler_jll: 0.3.0+0
• CUDA_Runtime_jll: 0.19.2+0

Toolchain:

• Julia: 1.12.3
• LLVM: 18.1.7

Preferences:

• CUDA_Runtime_jll.version: 13.0

1 device:
0: NVIDIA RTX 5090 GPU

[maleadt]

Currently, CUDA.jl appears to be precompiled for/supporting up to 13.0.x, leading to initialization errors when a system-wide 13.1.0 driver/toolkit is detected.

That is not true. CUDA.jl selects toolkit 13.0, which is perfectly compatible with your CUDA 13.1 driver.

Error: You are using CUDA 13.1.0, but CUDA.jl was precompiled for CUDA 13.0.0.

This code is comparing the toolkit version as returned by querying libcudart, as shipped by our artifacts, to the value it was precompiled for. This should always match. Just a guess: do you have anything on LD_LIBRARY_PATH, which CUDA.jl may have warned about during loading?