CUDA_ERROR_INVALID_CONTEXT on NVIDIA RTX PRO 4000 Blackwell (sm_120)

[afrancke]

GPU: NVIDIA RTX PRO 4000 Blackwell SFF
Driver: 590.48.01
CUDA runtime: 13.1
NVEncC: 9.10 (built against CUDA 12.9)
OS: Ubuntu 25.10 (Resolute), kernel 6.19.0-5-generic

nvencc --check-hw returns:
gpuinfo: cuCtxPopCurrent error:0xc9 (CUDA_ERROR_INVALID_CONTEXT)
No GPU found suitable for NVEnc Encoding.
Failed to initialize CUDA.

Notes:

• NVENC works correctly via Plex and ffmpeg hevc_nvenc
• cuInit(0) succeeds from Python/ctypes
• Reproduced both on host and inside Docker container with --gpus all
• Built with CUDA 12.6 and 12.9, same result both times

[rigaya]

Unfortunately, I’m not being able to reproduce the issue, perhaps the environment is completely different.
It might be caused of the driver handling of user CUDA context code of NVEnc, which ffmpeg won’t use, if so, will be difficult to avoid.
Would you run
nvencc --log-level debug --check-hw
and provide the full log to check where it has failed?