NVIDIA NIM 2.1

Apply these rules to NVIDIA NIM for LLM and VLM 2.1 model-specific and model-free containers. Each image bundles either vLLM or SGLang; check the actual image, model, profile, and backend before using a backend-specific rule. This is standalone serving infrastructure, including its proxy, model downloads, adapters, deployment, and inference APIs.

Release and deployment integrity

• Run the newest security-maintained patch of the selected NIM release line and use the official Support Matrix and NGC catalog to choose the model-specific or model-free image, backend, tag, and validated GPU profile. Look up the current release instead of relying on a remembered number; keep reviewed image digests and Helm chart references current through an update bot or scheduled review, and never loosen an existing deployment constraint merely to complete an install.
• Inventory and audit the complete image and its bundled vLLM or SGLang backend in CI. Check NVIDIA security bulletins for the actual image and separately check the host OS, GPU driver, Container Toolkit, and GPU Operator; an image scan does not establish host safety or prove an upstream backend fix was backported.
• Keep the model, GPU, precision, tensor and pipeline parallelism, LoRA support, and image backend together as a validated profile from list-model-profiles and the Support Matrix. Recheck startup, inference, readiness, and intended API behavior after changing an image, driver, model, profile, or backend.
• Record the selected image digest, backend name and version, profile ID, model artifact identity, chart, and