ollama-rocm

Version	2026.229.1218
Repo	box/github.com/opencharly/layer-ollama-rocm:v2026.243.1755

AMD ROCm inference backend for the ollama LLM server Adds the ROCm/HIP ggml backend so a composing box runs inference on an AMD card instead of falling back to CPU. This is an ADD-ON to the ollama candy, never a replacement: it contributes only the backend libraries that drop into ollama’s own runner directory (/usr/lib/ollama/rocm_v*/libggml-hip.so plus the HIP/HSA runtime it links), which is why the base ollama candy can stay CPU-only and small on a packaged distro. Unlike CUDA, ROCm is never in the upstream tarball at all, so there is no path on which the base candy carries it. It is by far the largest of the backends (~2.9 GiB installed), so keeping it in its own candy is what stops every ollama image from paying for it — an image opts in only when it targets AMD hardware.

On Arch (CachyOS included) the split ollama-rocm package supplies the backend and depends on the ollama package, so the two compose rather than conflict. A distro without an ollama package is NOT covered here: the base upstream tarball ships no ROCm backend (upstream publishes ROCm as a separate ollama-linux-amd64-rocm archive), so composing this candy there would install nothing and the check below would correctly fail rather than pretend. Extending it to the tarball distros means teaching the ollama candy to select the ROCm archive, which is a change to that candy, not a second copy of this one.

keep-groups preserves the host video/render groups so a passed-through AMD GPU (/dev/kfd + /dev/dri/renderD*) is usable from the container. The backend library is verifiable at build scope; live GPU enumeration is verifiable only on a host with a real AMD GPU passed through.

Acceptance plan

This candy’s plan: — the runnable spec charly check executes against a live deployment. check: steps are idempotent probes; run: steps change state.

Intent	Step
check	the ROCm/HIP ggml backend library is installed in ollama’s runner directory