Deadlock while running on GPU

[pierre-fribourg-context]

Hi all,
first, thanks for this great work.

I encounter a weird issue while testing decord on GPU on my computer.
I am running the following benchmark command:

python3 bench_decord.py --file /my_video.mp4 for CPU
python3 bench_decord.py --file /my_video.mp4 --gpu 0 for GPU

It is working fine on CPU ending with the following line:

300 frames, elapsed time for random access(accurate): 25.084531545639038

But on GPU it deadlocks almost right away after this line:

[h264 @ 0x171d080] mmco: unref short failure

This appends while running localy or through docker (from decord/gpu.Dockerfile) with nvdia-runtime enabled

Here is my nvidia-smi output (its the same locally or from inside the docker):

NVIDIA-SMI 470.63.01 Driver Version: 470.63.01 CUDA Version: 11.4

From inside the docker I don't see any process lines in nvdia-smi, but locally I crealy see the process line jumping onto the GPU as here:

| 0 N/A N/A 5444 C python3 205MiB |

GPU usage keep very low, and it deadlocks for ever.

Do I have to switch to another version of CUDA maybe ?

Thanks a lot.

[stdKonjac]

I meet the same problem, have you solved it?