invalid discovery mode: cdi

[Undefined443]

After upgrading to nvidia-container-toolkit v1.18.0-1, I followed the instructions from Generating a CDI specification to generate the CDI specification, but encountered the following problem:

$ sudo systemctl enable --now nvidia-cdi-refresh.service
Job for nvidia-cdi-refresh.service failed because the control process exited with error code.
See "systemctl status nvidia-cdi-refresh.service" and "journalctl -xeu nvidia-cdi-refresh.service" for details.

After checking the log records, I found that the error message was invalid discovery mode: cdi:

$ sudo journalctl -u nvidia-cdi-refresh.service
Oct 23 10:22:21 ubuntu systemd[1]: Starting nvidia-cdi-refresh.service - Refresh NVIDIA CDI specification file...
Oct 23 10:22:21 ubuntu nvidia-ctk[64294]: time="2025-10-23T10:22:21+08:00" level=error msg="invalid discovery mode: cdi"
Oct 23 10:22:21 ubuntu systemd[1]: nvidia-cdi-refresh.service: Main process exited, code=exited, status=1/FAILURE
Oct 23 10:22:21 ubuntu systemd[1]: nvidia-cdi-refresh.service: Failed with result 'exit-code'.
Oct 23 10:22:21 ubuntu systemd[1]: Failed to start nvidia-cdi-refresh.service - Refresh NVIDIA CDI specification file.

[Undefined443]

Finally I found that I needed to specify the discovery mode in the service file to start nvidia-cdi-refresh.service normally:

sudo vim /etc/systemd/system/nvidia-cdi-refresh.service

-ExecStart=/usr/bin/nvidia-ctk cdi generate
+ExecStart=/usr/bin/nvidia-ctk cdi generate --mode=auto

[elezar]

@Undefined443 the issue is that we're (incorrectly) reading the MODE for CDI from the /etc/nvidia-container-runtime/config.toml file. I assume that you have made modifications to this file to switch from the default of auto to cdi.

These modes are NOT compatible (or would need to be mapped correctly).

As a note, instead of updating the systemd unit you could add the following line to your /etc/nvidia-container-toolkit/nvidia-cdi-refresh.env file (or created it if it doesn't exist):

NVIDIA_CTK_CDI_GENERATE_MODE=auto

[mjk-vastai]

Hi, @elezar,

We've noticed that if we don't have the mode set at all, it can work for our stuff. why might that be?