happyme531/open_rknpu2_lm_infer is an open-source inference engine that runs LLMs on the RK3588 NPU. The commit merged on November 3, 2026 (84db49f) added an ARM64 container image publishing path and a Compose deployment template to this project.
The Dockerfile downloads the v0.11.0 release tarball from debian:bookworm-slim and builds a scratch image containing only the executables and the required shared libraries. The default command is serve --host 0.0.0.0 --port 8080, and, exactly as written in the README, it exposes an OpenAI-compatible LLM API on port 8080. The compose.yaml pins the image ghcr.io/happyme531/open_rknpu2_lm_infer:${RKNPU_VERSION:-0.11.0} to linux/arm64 and passes the NPU DRM device (default /dev/dri/renderD129) through to the container. The model directory is mounted read-only at /models, and the tuning cache is kept on a separate volume. The bind address defaults to 127.0.0.1, and the API key is passed via the OPEN_RKNPU_API_KEY environment variable. The image label lists the license as AGPL-3.0-only.
Image publishing is handled by .github/workflows/container.yml. After release.yml uploads a release, the container is built automatically, and the workflow_dispatch in ci.yml can also select only an existing release by entering a specific version and publishing it. After the build, the --version output and the version and model_families in the JSON produced by the models subcommand are checked, and the image is then pushed to GHCR; the latest tag is attached only when it is the newest release.
However, what this commit verifies is limited to the version string and the model list output. There are no performance or accuracy figures in the commit history, and the contents of the newly added llm_generated_docs/DOCKER.md and DOCKER_VALIDATION_20261104.md are not included in this record. The image's default tag is 0.11.0, and since what the change targets is not a new feature but the deployment method for existing releases, the release that is actually the deployment target must be confirmed by checking the release tags directly.