NVIDIA Triton Inference Server 2.73.0

Apply these rules to Triton server 2.73.0, its 26.09 containers, clients, model repositories, backends, cloud frontends, and embedded APIs. Triton serves inference; put application identity, tenant authorization, and data policy at the application boundary.

Release and deployment contract

• Select server and SDK image digests from the official release catalog, verify the running server and backend artifacts against the reviewed release, and build custom server, core, common, client, and backend components from matching release branches. Review NVIDIA bulletins against deployed and rollback images; move to a newer monthly release when a bulletin names the deployed release as affected. Keep the complete model/configuration artifact available for rollback.
• Qualify the container's Ubuntu 24.04 and Python 3.12 environment, its CUDA 13.4.1 and TensorRT 11.3 libraries, and the chosen backend against the host GPU and driver. The 26.09 GPU container requires compute capability 7.5 or later; verify CUDA driver compatibility for actual features and PTX, not just a nominal toolkit version. Qualify CPU, SBSA, Jetson, and wheel artifacts separately.
• Match custom Python environments to triton_python_backend_stub and review their dependencies. Do not assume the TensorRT-LLM backend container is available in this release. Lock and scan client dependencies and retain upstream security fixes in custom builds.
• Record the digest and branch of every backend shipped beside the server; a server version string alone does not prove that a custom backend contains the reviewed code.