cuda_error_invalid_source: device kernel image on RTX 4090 graphic cards

[Richard Meng]

Hi everyone,

We recently released a desktop app that uses the CuPy library for accelerated computation on GPUs. However, we have received reports from several users with RTX 4090 graphic cards that they are encountering an error during `cupy.RawKernel` execution. The error message they see is "cuda_error_invalid_source: device kernel image."  So far, we did not observe the same error in another hardware setup. 

Our program uses CUDA 11.7 (which came with pytorch installation) and python 3.9, running on Windows 11 Pro. 

If anyone has encountered this issue before or has any suggestions on resolving it, we would be very grateful for your help. Please let me know if providing other info will help with the diagnosis. 

Thank you in advance for any assistance you can provide.

[Kenichi Maehashi]

Hi,

RTX 4090 (Ada Lovelace) requires CUDA 11.8 or later.

https://docs.nvidia.com/cuda/archive/11.8.0/cuda-toolkit-release-notes/index.html#cuda-general-new-features

Kenichi

2023年3月4日土曜日 17:34:46 UTC+9 sharm...@gmail.com:

[Richard Meng]

Thank you... will try out the 11.8 version and see how it goes! 

Kenichi Maehashi 在 2023年3月5日 星期日上午11:33:56 [UTC+8] 的信中寫道：