cuvidGetDecoderCaps(&decodecaps) returned CUDA_ERROR_NO_DEVICE (code: 100)

[Yyong25]

System Information:

system: ubuntu:20.04

Nvidia card: four NVIDIA GeForce RTX 3060

Nvidia dirver : 580.82.09

Code project address and replication method: git@github.com:Yyong25/nvidia_decoder_test.git

Create 4 containers, map 4 cards each, only one card can output correctly

[Yyong25]

I have a 4-card RTX 3060 system with driver version 580.82.09. Using the nvidia/cuda:13.0.1-cud decoder-devel-ubuntu22.04 base image, I created 4 separate containers mapping cards 0-3 into each container respectively. Cards 0, 1, and 2 all return the error cuvidGetDecoderCaps(&decodecaps) returned CUDA_ERROR_NO_DEVICE (code: 100), while only card 3 returns correct results from cuvidGetDecoderCaps(&decodecaps).

[elezar]

@Yyong25 which version of the nvidia-container-toolkit are you running?

[Yyong25]

@elezar
driver:580.82.09
NVIDIA Container Runtime Hook version 1.18.0
commit: f8daa5e

Is this error related to the version of the Nvidia container toolkit? I also have suspected the driver version. I conducted another experiment using the base image of NVIDIA/CUDA: 12.8.1-devel-ubunt22.04 on another 4-card RTX 6000 Ada with driver version 535, NVIDIA Container Runtime Hook version 1.17.8， Using the same method, create 4 containers and map each card to the container separately. This function return no error, but this basic image will still error on driver version 580