NVJPEG debug tests are fine, but fails in the actual session

[Rush-iam]

Appreciate any help or advices on what I should try.

Env

• Xpra 4.4.4
• Ubuntu 22.04 inside Docker container (it gets Nvidia libs passed to container from the host system, not from apt packages)
• Tesla T4
• tried different Pycuda pip packages

Previously I worked with Xpra 4.3.4 + Ubuntu 18.04 without problems with NVJPEG. Now, I upgraded to 4.4.4 + 22.04.

Xpra cmd line

xpra start
    --bind-ws=0.0.0.0:8998
    --exec-wrapper="vglrun -d egl"
    --daemon=no
    --mdns=no
    --pulseaudio=no
    --speaker=disabled
    --microphone=disabled
    --webcam=no
    --printing=no
    --notifications=no
    --bandwidth-detection=no
    --quality=85
    --encodings=all,-h264
    -d nvenc,nvjpeg

Logs

• Startup log with NVJPEG/NVENC debug lines and successful NVJPEG tests: xpra.log
• During a session, when fails to encode:

2023-04-19 13:25:13,882 Warning: failed to setup video pipeline (68, (1, 1), Non
e, 0, 0, None, 'BGRX', (1, 1), 1872, 1018, nvjpeg(BGRX to jpeg))
Traceback (most recent call last):                                              
  File "/usr/lib/python3/dist-packages/xpra/server/window/window_video_source.py
", line 1692, in setup_pipeline                                                 
    if self.setup_pipeline_option(width, height, src_format, *option):
  File "/usr/lib/python3/dist-packages/xpra/server/window/window_video_source.py
", line 1778, in setup_pipeline_option                                          
    ve.init_context(encoder_spec.encoding, enc_width, enc_height, enc_in_format,
 options)                                                                       
  File "xpra/codecs/nvjpeg/encoder.pyx", line 194, in xpra.codecs.nvjpeg.encoder
.Encoder.init_context
AssertionError: no cuda device context  

• nvidia-smi:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.161.03   Driver Version: 470.161.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            On   | 00000000:00:1E.0 Off |                    0 |
| N/A   47C    P0    34W /  70W |    104MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

[totaam]

Please add -d video,cuda to your command line and post the output.

[Rush-iam]

here it is: xpra-video-cuda.log

[totaam]

I don't see any failures in that log sample.

[Rush-iam]

But why it says device Tesla T4 is not supported: cannot open encoding session? It worked fine in Ubuntu 18.04 🤔

[totaam]

device Tesla T4 is not supported: cannot open encoding session:
This indicates that an invalid struct version was used by the client., 0 contexts are in use

Means that your drivers and / or libraries are not compatible with the CUDA SDK version used for building the codec.

[Rush-iam]

Thank you, @totaam!
Downgrading from v4.4.4 to v4.4.3 solved the problem!

[totaam]

That's not a good solution as you won't be able to get updates.
I'm pretty sure that there are newer drivers available for Ubuntu 22.04, using that is a better solution.