NCA-GENL - Software Development - Section 2.4

Deploy machine learning models on Triton Inference Server.

Deploy trained models using NVIDIA Triton Inference Server, configuring the model repository, backend (TensorRT, ONNX, PyTorch, TensorFlow), and instance groups for concurrent execution. Understand how Triton's dynamic batching and ensemble pipelines improve GPU utilisation and reduce inference latency in production.

Triton Inference Server

Practice question for this objective

Free sampleSoftware Developmentmedium

A developer needs to run a PyTorch model and an ONNX model side by side on the same Triton Inference Server instance, serving both through a single HTTP endpoint prefix. What aspect of Triton makes this directly possible?

  • ATriton's concurrent model execution, which lets multiple model instances share GPU memory and run simultaneously.
  • BTriton's model ensemble feature, which links the PyTorch and ONNX models into a single composite pipeline.
  • CTriton's model repository layout, which stores each model in its own versioned subdirectory and serves all of them under one server.
  • DTriton's multiple framework backend support, which allows models from different frameworks to be loaded and served concurrently within one server process. Correct
Recognise that Triton supports multiple framework backends simultaneously, enabling PyTorch, ONNX, TensorRT, and other models to be served from a single server instance. Triton Inference Server uses a pluggable backend architecture. Each backend (PyTorch, ONNX Runtime, TensorFlow, TensorRT, Python, and others) is implemented as a shared library and loaded independently. When Triton starts and reads the model repository, it inspects the backend field in each model's config.pbtxt and loads the appropriate library. Multiple backends can be active concurrently in the same process, so a PyTorch model and an ONNX model can both be served through the unified REST and gRPC API without conversion or separate server instances.

Why A is wrong: Concurrent model execution relates to running multiple instances of the same or different models at the same time for throughput, which is tempting here, but it is not the feature that enables different framework backends to coexist. Concurrent execution relies on the multi-backend architecture but is a distinct capability.

Why B is wrong: Ensembles are plausible because they connect multiple models, but an ensemble requires explicitly defining the data flow between models. Simply running two unrelated models side by side does not require an ensemble, and an ensemble would not be needed just to serve them independently.

Why C is wrong: The model repository structure is a prerequisite for loading models, and it is tempting because it is what you configure to add models. However, the repository layout alone does not explain why different framework types can coexist; that is determined by the backend architecture.

Why D is correct: Triton ships with separate backends for PyTorch (libtorch), ONNX Runtime, TensorFlow, TensorRT, and others. Each backend is a shared library loaded on demand, so models from different frameworks coexist in the same server process and are accessible through the same API endpoint without any conversion.

See more NCA-GENL practice questions, answers explained.

More in this domain

Back to all Software Development objectives, or the NCA-GENL cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.