Misc. bug: GPU detection bug: The Docker version of llama.cpp did not detect the GPU. Vulkan Backend, AMD RX7900XTX (ROCm is fine)

[arbv]

Name and Version

[root@cauldron:/etc/nixos]# podman run --rm --security-opt=label=disable --device /dev/dri -e VK_LOADER_DEBUG=all --group-add video ghcr.io/ggml-org/llama.cpp:full-vulkan --server --version
version: 9592 (ac4cddeb0)
built with GNU 15.2.0 for Linux x86_64

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

podman run --rm --security-opt=label=disable --device /dev/dri -e VK_LOADER_DEBUG=all ghcr.io/ggml-org/llama.cpp:full-vulkan --server --list-devices --verbose
Available devices:

Problem description & steps to reproduce

llama.cpp does not detect my GPUs at all when using Vulkan backend. ROCm works fine.

First Bad Commit

No response

Relevant log output

Logs when using the official docker image

podman run --rm --security-opt=label=disable --device /dev/dri -e VK_LOADER_DEBUG=all ghcr.io/ggml-org/llama.cpp:full-vulkan --server --list-devices --verbose
Available devices:

Logs when using vulkaninfo with Debian Trixie image (to demonstrate that GPU passthrough for Vulkan works)

[root@cauldron:/etc/nixos]# podman run --rm --device /dev/dri   docker.io/library/debian:trixie   sh -c 'apt-get update -qq && apt-get install -y -qq mesa-vulkan-drivers vulkan-tools >/dev/null 2>&1; vulkaninfo --summary 2>/dev/null | grep -iE "deviceName|driverName|GPU id|apiVersion"'
	apiVersion         = 1.4.305
	deviceName         = AMD Radeon RX 7900 XTX (RADV NAVI31)
	driverName         = radv
	apiVersion         = 1.4.305
	deviceName         = AMD Radeon Graphics (RADV RAPHAEL_MENDOCINO)
	driverName         = radv
	apiVersion         = 1.4.305
	deviceName         = llvmpipe (LLVM 19.1.7, 256 bits)
	driverName         = llvmpipe

[kstoykov]

I can also confirm it.

I noticed the issue when I upgraded from b9570 to b9592.

[cibernox]

I can confirm this and have bisected the regression.

Last good build: b9426
First broken build: b9438

Tested by pulling each image and running --list-devices in a fresh container with /dev/dri passed through:

docker run --rm --device /dev/dri/renderD129 --device /dev/dri/card1 \
  --entrypoint /app/llama-server \
  ghcr.io/ggml-org/llama.cpp:server-vulkan-b9426 --list-devices
# → Vulkan0: AMD Radeon RX 7900 XTX (RADV NAVI31) (24560 MiB, 24533 MiB free)

docker run --rm --device /dev/dri/renderD129 --device /dev/dri/card1 \
  --entrypoint /app/llama-server \
  ghcr.io/ggml-org/llama.cpp:server-vulkan-b9438 --list-devices
# → Available devices:  (empty)

b9438 was built on 2026-05-31T05:54Z. The regression has been in the floating server-vulkan tag since then.

Notably, vulkaninfo works fine inside the broken container — the GPU passthrough and Vulkan loader are not the issue. The ggml Vulkan backend specifically fails to enumerate physical devices.

Workaround: pin to ghcr.io/ggml-org/llama.cpp:server-vulkan-b9426.

[TomG736]

I find that ghcr.io/ggml-org/llama.cpp:server-vulkan-b9570 works but b9592 onwards doesn't. I havent tested back to b9438
This is with an Intel ARC GPU

[Rand-o]

Confirmed it also broke for me on Strix Halo trying to use vulkan. I had to roll back

[klimekop6]

I can confirmed it also. My GPU RX 6800 is not visible when i run server-vulkan-b9592 or b9603. Switching to b9570 fix the issue

[Tech-no-1]

I can also confirm this with an Intel ARC gpu, b9570 is the last vulkan docker version detecting the gpu. SYCL docker detects the gpu, it's much slower though.

[Kononnable]

It seems we're silently running out of memory on CI builds.

#14 337.5 Cannot allocate memory
#14 337.5 Error executing command for matmul_id_subgroup_q5_0_f32_aligned: Failed to fork process

Latest working version: b9426 https://github.com/ggml-org/llama.cpp/actions/runs/26675729593/job/78627009078#step:9:1373
First broken version: b9438 https://github.com/ggml-org/llama.cpp/actions/runs/27397273833/job/80967212036#step:9:2828

[prcoe1]

Same here with AMD R9700, last working version: ghcr.io/ggml-org/llama.cpp:server-vulkan-b9570, any newer one from here doesn't pick up any GPU and runs on CPU only.

[npodbielski]

I confirm it does not work on strix halo.