NVIDIA Triton Inference Server 2.72.0

Use these rules for NVIDIA Triton Inference Server 2.72.0 and its 26.08 containers, clients, model repositories, backends, repository agents, caches, and embedded APIs. Enforce application identity and tenant policy outside Triton, and treat loaded model code as trusted server code.

Release and compatibility boundary

• Reference the server and SDK images by digest, verify that the server reports the release this deployment was reviewed against, and build custom server, core, common, client, and backend revisions from the release branch that matches the image. Preserve the complete model and configuration artifact for rollback.
• Take the NGC image tags and digests from the NGC catalog or the Triton release notes rather than memory, leaving a placeholder that fails the deployment when you can't check, and keep the digests in deployment configuration that an update bot or a scheduled review keeps current. NVIDIA's bulletins direct Triton fixes to newer monthly releases, so move to one when a bulletin lists yours as affected. The 2.72.0 in this prompt's title names the release these rules were checked against, not a version to install.
• Qualify the Ubuntu 24.04/Python 3.12 container environment and each backend's target platform. Match a Python execution environment's interpreter to triton_python_backend_stub; changing the interpreter requires a matching rebuilt stub and dependencies.
• For GPU deployments, require the 26.08 compute-capability baseline of 7.5 or later and verify the actual image's CUDA, TensorRT, and backend libraries against the host driver. CUDA