Public bug reported:
Note that I'm not entirely sure where to report this. To my mind
upgrading podman makes most sense, but I don't know.
On AWS at least, Ubuntu upgraded kernel to 7.0. Since then, the kernel loads
`nvidia_drm` module:
```
$ uname -a
Linux ip-172-22-29-16 7.0.0-1011-aws #11~24.04.1-Ubuntu SMP PREEMPT Mon Aug 10
15:20:57 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
$ lsmod | grep nvidia_drm
nvidia_drm 159744 0
nvidia_modeset 1544192 1 nvidia_drm
drm_ttm_helper 20480 1 nvidia_drm
```
which was not present with the previous 6.17 kernel:
```
$ uname -a
Linux ip-172-22-29-13 6.17.0-1019-aws #19~24.04.1-Ubuntu SMP Tue Jun 23
18:53:06 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
$ lsmod | grep nvidia_drm
```
That creates new devices:
```
$ ls -al /dev/dri/
total 0
drwxr-xr-x 3 root root 120 Sep 3 13:07 .
drwxr-xr-x 17 root root 3600 Sep 3 08:04 ..
drwxr-xr-x 2 root root 100 Sep 3 13:07 by-path
crw-rw---- 1 root video 226, 0 Sep 3 08:04 card0
crw-rw---- 1 root video 226, 1 Sep 3 13:07 card1
crw-rw---- 1 root render 226, 128 Sep 3 13:07 renderD128
```
vs the old kernel:
```
$ ls -al /dev/dri/
total 0
drwxr-xr-x 3 root root 80 Sep 3 09:26 .
drwxr-xr-x 17 root root 3520 Sep 3 09:26 ..
drwxr-xr-x 2 root root 60 Sep 3 09:26 by-path
crw-rw---- 1 root video 226, 0 Sep 3 09:26 card0
```
and that causes nvidia-ctk to generate a different version of the spec, one
that podman 4.9.3 doesn't understand:
```
$ nvidia-ctk --quiet cdi generate | head -2
---
cdiVersion: 0.7.0
```
vs the old kernel (or new kernel with `rmmod nvidia_drm`):
```
$ nvidia-ctk --quiet cdi generate | head -2
---
cdiVersion: 0.5.0
```
and that makes podman unable to work with the GPU:
```
$ sudo podman --log-level=debug run --rm --device nvidia.com/gpu=0
docker.io/library/busybox true 2>&1 | grep 'CDI'
time="2026-09-03T13:27:30Z" level=debug msg="Identified CDI device
nvidia.com/gpu=0"
time="2026-09-03T13:27:30Z" level=debug msg="The following error was triggered
when refreshing the CDI registry: failed to load CDI Spec failed to parse CDI
Spec \"/var/run/cdi/nvidia.yaml\": failed to unmarshal CDI Spec: error
unmarshaling JSON: while decoding JSON: json: unknown field \"additionalGids\""
Error: setting up CDI devices: unresolvable CDI devices nvidia.com/gpu=0
```
because it doesn't understand the 0.7.0 spec.
I won't pretend to fully understand how all this works, and I do realize
upgrading podman will likely raise different demons. It did fall behind
quite a bit though.
** Affects: libpod (Ubuntu)
Importance: Undecided
Status: New
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2166357
Title:
podman 4.9.3 in 24.04 can't parse Nvidia's CDI 0.7 on new linux-aws
kernel v7
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/libpod/+bug/2166357/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
