NVIDIA NIM 2.0

Apply these standalone rules to the model-specific and model-free NVIDIA NIM for Large Language Models 2.0 containers using vLLM. Confirm the image and backend before applying them; separately versioned SGLang images and NIM 3.x require their own review.

Release, runtime, and profile selection

• Pin the container by digest, pin the Helm chart version, and inventory the actual backend and dependencies. The 2.0.11 release notes identify vLLM 0.27.0; a NIM tag alone does not establish which security backports an image contains.
• Take the NIM image tag and digest and the Helm chart version from the NGC catalog or the NIM release notes rather than memory, leaving a placeholder that fails the deployment when you can't check, and keep them in a values or lock file that an update bot or a scheduled review keeps current, so newer NVIDIA-supported images reach review. The 2.0 in this prompt's title names the release series these rules cover, and 2.0.11 is the release they were checked against, not a version to install.
• Require an NVIDIA-supported fixed image or verified vendor backports before exposing affected functionality to untrusted input. Upstream advisories include vLLM 0.27.0 in affected ranges for model-code loading, token input, cache salts, structured output, and multimodal processing. Gateway controls and restarts do not replace those fixes; do not upgrade packages inside the production container ad hoc.
• Validate the published Docker 24+, NVIDIA driver 580+, CUDA 12.9+, CPU architecture, and GPU requirements; these are compatibility minimums, not security