Describe the bug
CUDA initialization fails inside containers only, while the same host can use NVENC successfully outside containers.
The key symptom is:
nvidia-smi works inside the container- FFmpeg inside the container shows
h264_nvenc and cuda
- but actual CUDA initialization fails with:
cuInit(0) failed -> CUDA_ERROR_UNKNOWN: unknown error
I first noticed this with Plex in Docker, but I can reproduce the same failure with a plain FFmpeg container, so this does not appear to be Plex-specific.
To Reproduce
- Verify the host GPU works:
- Verify host FFmpeg NVENC works:
ffmpeg -hide_banner -encoders | grep nvenc
ffmpeg -hide_banner -decoders | grep cuvid
ffmpeg -hide_banner -hwaccels
ffmpeg -hide_banner -f lavfi -i testsrc=size=1280x720:rate=30 -t 5 -c:v h264_nvenc -f null -
On the host, this completes successfully.
- Verify basic GPU visibility in containers:
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
- Reproduce the failure in a plain FFmpeg container:
docker run --rm \
--runtime=nvidia \
--gpus all \
--entrypoint /bin/bash \
-e NVIDIA_VISIBLE_DEVICES=all \
-e NVIDIA_DRIVER_CAPABILITIES=compute,video,utility \
linuxserver/ffmpeg:latest \
-lc '
echo "=== NVIDIA-SMI ==="
nvidia-smi || true
echo
echo "=== ENCODERS (nvenc) ==="
ffmpeg -hide_banner -encoders | grep nvenc || true
echo
echo "=== HWACCELS ==="
ffmpeg -hide_banner -hwaccels || true
echo
echo "=== NVENC TEST ==="
ffmpeg -hide_banner -f lavfi -i testsrc=size=1280x720:rate=30 -t 5 -c:v h264_nvenc -f null - || true
'
- Observe the failure:
[h264_nvenc @ ...] dl_fn->cuda_dl->cuInit(0) failed -> CUDA_ERROR_UNKNOWN: unknown error
Expected behavior
I expect CUDA initialization to succeed in the container, just like it does on the host.
Specifically, I expect the FFmpeg NVENC test inside the container to complete successfully instead of failing at cuInit(0).
Environment (please provide the following information):
nvidia-container-toolkit version: 1.19.0- NVIDIA Driver Version:
590.48.01
- Host OS:
Arch Linux
- Kernel Version:
6.14.0-arch1-1
- Container Runtime Version:
Docker with nvidia-container-runtime configured
- CPU Architecture:
x86_64
- GPU Model(s):
NVIDIA RTX 2000 Ada Generation (AD107GL)
- CUDA Version:
13.1 (as reported by nvidia-smi)
Not applicable:
- Kubernetes Distro and Version
- NVIDIA GPU Operator version
Additional context
I originally saw this with Plex in Docker. Plex first failed with Cannot load libcuda.so.1, but after simplifying the setup and using the NVIDIA libraries already present in /usr/lib inside the container, Plex can now load libcuda.so.1.
However, Plex then fails at the same point as the plain FFmpeg container:
DEBUG - [GPU] Got device: NVIDIA AD107GL [RTX 2000 / 2000E Ada Generation], nvidia@unknown, default true, best true, ID 10de:28b0:10de:1870@0000:15:00.0, DevID [10de:28b0:10de:1870], flags 0xe8
ERROR - [GPU] Cuda function failed with error unknown error
DEBUG - [Req#e0/Transcode] [FFMPEG] - Loaded lib: libcuda.so.1
ERROR - [Req#e0/Transcode] [FFMPEG] - -> CUDA_ERROR_UNKNOWN: unknown error
So the important point is that this is reproducible outside Plex as well.
What currently works:
- host
nvidia-smi
- host FFmpeg NVENC
- container
nvidia-smi
- container reports
h264_nvenc
- container reports
cuda in ffmpeg -hwaccels
What fails:
- container FFmpeg actual CUDA initialization (
cuInit(0))
- Plex hardware transcoding in the container for the same reason
Information to attach
Output of nvidia-smi
Host:
GPU 0: NVIDIA RTX 2000 Ada Generation (UUID: GPU-8de1bc21-ab8b-71ba-c6af-9f5895aaabaf)
Container:
Tue Mar 24 08:33:14 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.48.01 Driver Version: 590.48.01 CUDA Version: 13.1 |
+-----------------------------------------+------------------------+----------------------+
| 0 NVIDIA RTX 2000 Ada Gene... Off | 00000000:15:00.0 Off | Off |
+-----------------------------------------------------------------------------------------+
Container logs
FFmpeg container output:
=== NVIDIA-SMI ===
Tue Mar 24 08:33:14 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.48.01 Driver Version: 590.48.01 CUDA Version: 13.1 |
+-----------------------------------------+------------------------+----------------------+
| 0 NVIDIA RTX 2000 Ada Gene... Off | 00000000:15:00.0 Off | Off |
+-----------------------------------------------------------------------------------------+
=== ENCODERS (nvenc) ===
V....D av1_nvenc
V....D h264_nvenc
V....D hevc_nvenc
=== HWACCELS ===
Hardware acceleration methods:
vdpau
cuda
vaapi
qsv
drm
opencl
vulkan
=== NVENC TEST ===
[h264_nvenc @ 0x55ceb1999d40] dl_fn->cuda_dl->cuInit(0) failed -> CUDA_ERROR_UNKNOWN: unknown error
[vost#0:0/h264_nvenc @ 0x55ceb1999740] [enc:h264_nvenc @ 0x55ceb1999cc0] Error while opening encoder - maybe incorrect parameters such as bit_rate, rate, width or height.
[out#0/null @ 0x55ceb19993c0] Nothing was written into output file, because at least one of its streams received no packets.
Conversion failed!
Additional diagnostic information
nvidia-ctk / CDI:
NVIDIA Container Toolkit CLI version 1.19.0
INFO[0000] Found 3 CDI devices
nvidia.com/gpu=0
nvidia.com/gpu=GPU-8de1bc21-ab8b-71ba-c6af-9f5895aaabaf
nvidia.com/gpu=all
Docker runtime info:
Runtimes: io.containerd.runc.v2 nvidia runc
Default Runtime: runc
/etc/docker/daemon.json:
{
"data-root": "/srv/docker",
"runtimes": {
"nvidia": {
"args": [],
"path": "nvidia-container-runtime"
}
}
}/etc/nvidia-container-runtime/config.toml relevant parts:
disable-require = false
supported-driver-capabilities = "compat32,compute,display,graphics,ngx,utility,video"
[nvidia-container-cli]
environment = []
ldconfig = "@/sbin/ldconfig"
load-kmods = true
[nvidia-container-runtime]
log-level = "info"
mode = "auto"
runtimes = ["runc", "crun"]
[nvidia-container-runtime.modes.cdi]
annotation-prefixes = ["cdi.k8s.io/"]
default-kind = "nvidia.com/gpu"
spec-dirs = ["/etc/cdi", "/var/run/cdi"]
Describe the bug
CUDA initialization fails inside containers only, while the same host can use NVENC successfully outside containers.
The key symptom is:
nvidia-smiworks inside the containerh264_nvencandcudaI first noticed this with Plex in Docker, but I can reproduce the same failure with a plain FFmpeg container, so this does not appear to be Plex-specific.
To Reproduce
On the host, this completes successfully.
Expected behavior
I expect CUDA initialization to succeed in the container, just like it does on the host.
Specifically, I expect the FFmpeg NVENC test inside the container to complete successfully instead of failing at
cuInit(0).Environment (please provide the following information):
nvidia-container-toolkitversion:1.19.0590.48.01Arch Linux6.14.0-arch1-1Dockerwithnvidia-container-runtimeconfiguredx86_64NVIDIA RTX 2000 Ada Generation(AD107GL)13.1(as reported bynvidia-smi)Not applicable:
Additional context
I originally saw this with Plex in Docker. Plex first failed with
Cannot load libcuda.so.1, but after simplifying the setup and using the NVIDIA libraries already present in/usr/libinside the container, Plex can now loadlibcuda.so.1.However, Plex then fails at the same point as the plain FFmpeg container:
So the important point is that this is reproducible outside Plex as well.
What currently works:
nvidia-sminvidia-smih264_nvenccudainffmpeg -hwaccelsWhat fails:
cuInit(0))Information to attach
Output of
nvidia-smiHost:
Container:
Container logs
FFmpeg container output:
Additional diagnostic information
nvidia-ctk/ CDI:Docker runtime info:
/etc/docker/daemon.json:{ "data-root": "/srv/docker", "runtimes": { "nvidia": { "args": [], "path": "nvidia-container-runtime" } } }/etc/nvidia-container-runtime/config.tomlrelevant parts: