NVIDIA TensorRT Inference Server

The NVIDIA TensorRT Inference Server provides a cloud inferencing solution optimized for NVIDIA GPUs. The server provides an inference service via an HTTP or GRPC endpoint, allowing remote clients to request inferencing for any model being managed by the server.

What’s New In 1.4.0

  • Added libtorch as a new backend. PyTorch models manually decorated or automatically traced to produce TorchScript can now be run directly by the inference server.

  • Build system converted from bazel to CMake. The new CMake-based build system is more transparent, portable and modular.

  • To simplify the creation of custom backends, a Custom Backend SDK and improved documentation is now available.

  • Improved AsyncRun API in C++ and Python client libraries.

  • perf_client can now use user-supplied input data (previously perf_client could only use random or zero input data).

  • perf_client now reports latency at multiple confidence percentiles (p50, p90, p95, p99) as well as a user-supplied percentile that is also used to stabilize latency results.

  • Improvements to automatic model configuration creation (--strict-model-config=false).

  • C++ and Python client libraries now allow additional HTTP headers to be specified when using the HTTP protocol.

Features

  • Multiple framework support. The server can manage any number and mix of models (limited by system disk and memory resources). Supports TensorRT, TensorFlow GraphDef, TensorFlow SavedModel, ONNX, PyTorch, and Caffe2 NetDef model formats. Also supports TensorFlow-TensorRT integrated models. Variable-size input and output tensors are allowed if supported by the framework. See Capabilities for detailed support information for each framework.

  • Concurrent model execution support. Multiple models (or multiple instances of the same model) can run simultaneously on the same GPU.

  • Batching support. For models that support batching, the server can accept requests for a batch of inputs and respond with the corresponding batch of outputs. The inference server also supports multiple scheduling and batching algorithms that combine individual inference requests together to improve inference throughput. These scheduling and batching decisions are transparent to the client requesting inference.

  • Custom backend support. The inference server allows individual models to be implemented with custom backends instead of by a deep-learning framework. With a custom backend a model can implement any logic desired, while still benefiting from the GPU support, concurrent execution, dynamic batching and other features provided by the server.

  • Ensemble support. An ensemble represents a pipeline of one or more models and the connection of input and output tensors between those models. A single inference request to an ensemble will trigger the execution of the entire pipeline.

  • Multi-GPU support. The server can distribute inferencing across all system GPUs.

  • The inference server monitors the model repository for any change and dynamically reloads the model(s) when necessary, without requiring a server restart. Models and model versions can be added and removed, and model configurations can be modified while the server is running.

  • Model repositories may reside on a locally accessible file system (e.g. NFS) or in Google Cloud Storage.

  • Readiness and liveness health endpoints suitable for any orchestration or deployment framework, such as Kubernetes.

  • Metrics indicating GPU utilization, server throughput, and server latency.

Indices and tables