NVIDIA Triton Inference Server is an open-source model-serving platform designed to standardize production deployment across heterogeneous model formats and hardware backends. Its defining capability is multi-framework support: a single Triton instance can concurrently serve models trained in TensorFlow, PyTorch, ONNX Runtime, TensorRT, OpenVINO, and custom Python/C++ backends, exposing them through unified HTTP/REST and gRPC inference APIs. This eliminates the need for framework-specific serving stacks and lets teams standardize deployment infrastructure independent of how a given model was trained.

Option B is a common but incorrect assumption — Triton explicitly supports both GPU and CPU inference, which matters for cost-sensitive or edge deployments where GPU availability is limited. Option C confuses Triton with cuGraph, a separate RAPIDS library for GPU-accelerated graph analytics (unrelated to model serving), and option A describes a generative task (denoising diffusion), which is a *model capability*, not a Triton *server* function — Triton can host such a model, but "generating images from noise" is not what the server itself does.

Triton also provides dynamic batching, concurrent model execution, model ensembling (chaining pre/post-processing with inference), and metrics export — features tested elsewhere in the Software Development and Engineering and Performance Optimization domains.