What is retained
The NIM on KServe entry retains the KServe subtree of NVIDIA/nim-deploy at an exact commit: the serving runtimes and the model-by-GPU shapes, with a digest index over every file. The deployment scaffolding is Apache-2.0; the license read verified that against the actual terms before anything was built.
One shape you can read, instead of a count
examples/aicr/kserve-nim-inference/upstream/kserve/nim-models/llama-3.1-8b-instruct_1xgpu_1.1.0.yaml is a complete InferenceService: one nvidia.com/gpu, the named runtime nvidia-nim-llama-3.1-8b-instruct-1.1.0, and model storage on pvc://nvidia-nim-pvc/. Every other shape in the directory has the same anatomy at different sizes, up to llama3-70b-instruct_4xh100.
The licensing boundary
The runtime images and models behind these shapes are NGC-gated under NVIDIA AI Enterprise licensing. The catalog retains the configuration shapes and never redistributes the gated artifacts. NGC API keys enter a deployment only as target facts you supply; a literal credential value anywhere in the retained tree stops the compile.
What is not covered, plainly
- Triton, vLLM, and Ray Serve are not retained here. Searching this catalog for them finds nothing, and that is the true answer today.
- GPU node provisioning (device plugins, node selectors, taints) is covered for the charts that carry it - see karpenter and nvidia-device-plugin in the catalog - not as a general guide.
- No GPU workload has run to produce any receipt in this catalog. Every NIM claim here is config-plane: import, render, digest pinning, delivery wiring.
If one of the absences blocks you, say so with a problem chart report: demand is what orders the queue.