Running Dynamo, Nvidia's software for serving AI models across many GPUs, on Kubernetes needs the right drivers, scheduling, health checks and monitoring to all work together. Setting that up by hand takes a long time.

Nvidia AI Cluster Runtime ships tested "recipes" with fixed versions of every part. Teams snapshot the cluster, generate a recipe, render a bundle, deploy it and check that it works before adding their own apps.

Researchers use ready-made models from the MONAI Model Zoo to analyse medical scans, but wanted them to run fast on AMD GPUs without changing the models.

Each model gets an extra settings file that switches on a faster memory layout, 16-bit numbers and PyTorch's compiler. The weights and the inference steps stay the same, and every model still meets its accuracy targets.

GPUDirect Storage lets GPUs read data straight from storage, but setting it up correctly means reading a lot of documentation.

gds-diag is an open-source command-line tool that checks a machine, marks each finding as FAIL, WARN, INFO or PASS, and suggests a fix. An optional add-on lets AI coding assistants run the checks from plain-language requests.

Teams building AI agents need a model, software to run the agent, tools, and controls over what the agent may touch. Putting all of that together safely is hard.

NemoClaw blueprints combine an existing agent framework with Nvidia's Nemotron models, tuning tools and the OpenShell runtime, which limits the agent's access to files, networks and passwords.