FFmpeg inside a Docker container can't see the GPU

Graphics / Linux Linux
, , , , , ,
1

I’m using FFmpeg to apply a GLSL .frag shader to a video. I do it with this command

docker run --rm \	--gpus all \	--device /dev/dri \	-v $(pwd):/config \	lscr.io/linuxserver/ffmpeg \	-init_hw_device vulkan=vk:0 -v verbose \	-i /config/input.mp4 \	-vf "libplacebo=custom_shader_path=/config/shader.frag" \	-c:v h264_nvenc \	/config/output.mp4 \	2>&1 | less -F

but the extremely low speed made me suspicious

frame=16 fps=0.3 q=45.0 size=0KiB time=00:00:00.43 bitrate=0.9kbits/s speed=0.00767x elapsed=0:00:56.52

The CPU activity was at 99.3% and the GPU at 0%. So I searched through the verbose output and found this:

[Vulkan @ 0x63691fd82b40] Using device: llvmpipe (LLVM 18.1.3, 256 bits)

For context:

I’m using an EC2 instance (g6f.xlarge) with ubuntu 24.04.
I’ve installed the NVIDIA GRID drivers following the official AWS guide, and the NVIDIA Container Toolkit following this other guide.
Vulkan can see the GPU outside of the container

ubuntu@ip-172-31-41-83:~/liquid-glass$ vulkaninfo | grep -A2 "deviceName" 'DISPLAY' environment variable not set... skipping surface info deviceName = NVIDIA L4-3Q pipelineCacheUUID = 178e3b81-98ac-43d3-f544-6258d2c33ef5

More info that might be useful

ubuntu@ip-172-31-41-83:~$ docker run --rm --gpus all --device /dev/dri nvidia/vulkan:1.3-470 vulkaninfo --summary Unable to find image 'nvidia/vulkan:1.3-470' locally 1.3-470: Pulling from nvidia/vulkan ea362f368469: Pull complete 33ad2f6e0261: Pull complete 91ce2d96e937: Pull complete 9b499af792c6: Pull complete c04906891ebc: Pull complete ebf96a6776b2: Pull complete dfd3c8943767: Pull complete 49eaa5311171: Pull complete Digest: sha256:de1866cacf9cf0edad969a07504b0aaf24da57fc069b166a913b33188fe60231 Status: Downloaded newer image for nvidia/vulkan:1.3-470 Cannot create Vulkan instance. This problem is often caused by a faulty installation of the Vulkan driver or attempting to use a GPU that does not support Vulkan. ERROR at /vulkan-sdk/1.3.204.1/source/Vulkan-Tools/vulkaninfo/vulkaninfo.h:649:vkCreateInstance failed with ERROR_INCOMPATIBLE_DRIVER 
ubuntu@ip-172-31-41-83:~$ docker info | grep -i nvidia Runtimes: nvidia runc io.containerd.runc.v2 
ubuntu@ip-172-31-41-83:~$ ls -la /usr/share/vulkan/icd.d/ total 36 drwxr-xr-x 2 root root 4096 Sep 17 18:13 . drwxr-xr-x 6 root root 4096 Sep 17 18:13 .. -rw-r--r-- 1 root root 167 Aug 18 08:26 gfxstream_vk_icd.x86_64.json -rw-r--r-- 1 root root 169 Aug 18 08:26 intel_hasvk_icd.x86_64.json -rw-r--r-- 1 root root 163 Aug 18 08:26 intel_icd.x86_64.json -rw-r--r-- 1 root root 161 Aug 18 08:26 lvp_icd.x86_64.json -rw-r--r-- 1 root root 165 Aug 18 08:26 nouveau_icd.x86_64.json -rw-r--r-- 1 root root 164 Aug 18 08:26 radeon_icd.x86_64.json -rw-r--r-- 1 root root 164 Aug 18 08:26 virtio_icd.x86_64.json 
ubuntu@ip-172-31-41-83:~$ ls -la /dev/dri/ total 0 drwxr-xr-x 3 root root 120 Sep 17 18:38 . drwxr-xr-x 17 root root 3520 Sep 17 19:19 .. drwxr-xr-x 2 root root 100 Sep 17 18:38 by-path crw-rw---- 1 root video 226, 0 Sep 17 18:38 card0 crw-rw---- 1 root video 226, 1 Sep 17 18:38 card1 crw-rw---- 1 root render 226, 128 Sep 17 18:38 renderD128 
ubuntu@ip-172-31-41-83:~$ dpkg -l | grep nvidia ii libnvidia-container-tools 1.17.8-1 amd64 NVIDIA container runtime library (command-line tools) ii libnvidia-container1:amd64 1.17.8-1 amd64 NVIDIA container runtime library ii nvidia-container-toolkit 1.17.8-1 amd64 NVIDIA Container toolkit ii nvidia-container-toolkit-base 1.17.8-1 amd64 NVIDIA Container Toolkit Base 
ubuntu@ip-172-31-41-83:~/liquid-glass$ sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi Thu Sep 18 10:41:16 2025 +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 570.172.08 Driver Version: 570.172.08 CUDA Version: 12.8 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |

What I tried

  1. I tried locating the nvidia_icd.json file and passing it manually in two different ways
docker run --rm \ --gpus all \ --device /dev/dri \ -v $(pwd):/config \ -v /etc/vulkan/icd.d:/etc/vulkan/icd.d \ -v /usr/share/vulkan/icd.d:/usr/share/vulkan/icd.d \ lscr.io/linuxserver/ffmpeg \ -init_hw_device vulkan=vk:0 -v verbose \ -i /config/input.mp4 \ -vf "libplacebo=custom_shader_path=/config/shader.frag" \ -c:v h264_nvenc \ /config/output.mp4 \ 2>&1 | less -F 
docker run --rm \ --gpus all \ --device /dev/dri \ -v $(pwd):/config \ -v /etc/vulkan/icd.d:/etc/vulkan/icd.d \ -e VULKAN_ICD_FILENAMES=/etc/vulkan/icd.d/nvidia_icd.json \ -e NVIDIA_VISIBLE_DEVICES=all \ -e NVIDIA_DRIVER_CAPABILITIES=all \ lscr.io/linuxserver/ffmpeg \ -init_hw_device vulkan=vk:0 -v verbose \ -i /config/input.mp4 \ -vf "libplacebo=custom_shader_path=/config/shader.frag" \ -c:v h264_nvenc \ /config/output.mp4 \ 2>&1 | less -F
  1. I tried installing other packages that ended up breaking the NVIDIA driver (then I wiped everything out and started again from scratch with new working drivers)
sudo apt install nvidia-driver-570 nvidia-utils-570 
ubuntu@ip-172-31-41-83:\~$ nvidia-smi NVIDIA-SMI couldn't find libnvidia-ml.so library in your system. Please make sure that the NVIDIA Display Driver is properly installed and present in your system. Please also try adding directory that contains libnvidia-ml.so to your system PATH.
  1. I tried setting vk:1 instead of vk:0
[Vulkan @ 0x5febdd1e7b40\] Supported layers: [Vulkan @ 0x5febdd1e7b40\] GPU listing: [Vulkan @ 0x5febdd1e7b40\] 0: llvmpipe (LLVM 18.1.3, 256 bits) (software) [Vulkan @ 0x5febdd1e7b40\] Unable to find device with index 1!

  1. Tried both adding --volume /dev/dri:/dev/dri and putting it instead of --device /dev/dri. Also removing --device /dev/dri.

  2. Adding --runtime=nvidia

  3. Giving full permissions

sudo chmod 777 /dev/dri/*
2

Did you tried to “kick start” CUDA / compute capabilities form outside the container as describe here: Unrelaibale cuInit with sandboxes and containers (CUDA cuInit: Unknown error)