Build the image
The image holds the dspk command line and dspk-cli-sorter, on Debian slim, run as a non-root
user. Three targets share everything but the system libraries each GPU needs:
| Target | For | Adds | Run with |
|---|---|---|---|
cpu (default) |
no GPU | nothing (DSPK_RUNTIME=cpu) |
|
nvidia |
NVIDIA: CUDA and Vulkan | NVRTC 13.0 and the Vulkan loader; the driver comes from the host | --device nvidia.com/gpu=all |
mesa |
AMD, Intel: Vulkan (wgpu) | the Vulkan loader and Mesa’s Vulkan drivers | --device /dev/dri |
Build from the repository root, because the build needs the whole workspace:
podman build -f deploy/Containerfile --ignorefile deploy/.containerignore \ --target nvidia -t dspk-serve:nvidia .The first build compiles everything and takes a while. Another target afterwards compiles nothing: they share the build stage and Podman’s caches.
Build options
Section titled “Build options”| Option | What it does |
|---|---|
--build-arg CUDA_NVRTC=13-5 |
A newer NVRTC, at most the CUDA version of the host’s driver (nvidia-smi shows it) |
--build-arg FEATURES=cpu,hub |
Compile the CPU runtime only |
Next: run the server.