Triton Inference Server

The Triton Inference Server provides a cloud inferencing solution optimized for both CPUs and 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. For edge deployments, Triton Server is also available as a shared library with an API that allows the full functionality of the server to be included directly in an application.

What's New in 2.24.0

• Auto-Complete is enabled by default. The --strict-model-config option has been soft deprecated, use the new --disable-auto-complete-config CLI option instead.

• New example backend demonstrating Business Logic Scripting in C++.

• Users can provide values for "init_ops" in Tensorflow TF1.x GraphDef models through json file.

• New asyncio compatible API to the Python GRPC/HTTP APIs.

• Added thread pool to reduce service downtime for concurrently loading models. The thread pool size is configurable with the new --model-load-thread-count tritonserver option. You can find more information here.

• Model Analyzer now doesn't require config.pbtxt file for models that can be auto-completed in Triton.

• Refer to the 22.07 column of the Frameworks Support Matrix for container image versions on which the 22.07 inference server container is based.

Known Issues

• JetPack release will be published later in the month in order to align with JetPack SDK public availability.

• Auto-complete could cause an increase in server start time. To avoid a start time increase, users should provide the full model configuration.

• When auto-completing some model configs, backends may generate a model config even though there is not enough metadata (ex. Graphdef models for Tensorflow Backend). The user will see the model successfully load but fail to inference. In this case the user should provide the full model configuration for these models or use the --disable-auto-complete-config CLI option to show which models fail to load.

• Can't do autocomplete for PyTorch models, not enough metadata. Can only verify that the number of inputs is correct and the input names match what is specified in the model configuration. No info about number of outputs and datatypes. Related pytorch bug: pytorch/pytorch#38273.

• Running inference on multiple TensorRT model instances in Triton may fail with signal(6). The issue is expected to be fixed in a future release. Details can be found at #4566.

• Perf Analyzer stability criteria has been changed which may result in reporting instability for scenarios that were previously considered stable. This change has been made to improve the accuracy of Perf Analyzer results. If you observe this message, it can be resolved by increasing the --measurement-interval in the time windows mode or --measurement-request-count in the count windows mode.

• Unlike previously noted, 22.07 is the last release that defaults to TensorFlow version 1. From 22.08 onwards Triton will change the default TensorFlow version to 2.X.

• Triton PIP wheels for ARM SBSA are not available from PyPI and pip will install an incorrect Jetson version of Triton for Arm SBSA.

  The correct wheel file can be pulled directly from the Arm SBSA SDK image and manually installed.

• Traced models in PyTorch seem to create overflows when int8 tensor values are transformed to int32 on the GPU.

  Refer to issue pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices (A100 and A30).

• Triton metrics might not work if the host machine is running a separate DCGM agent on bare-metal or in a container.

• Starting from 22.02, the Triton container, which uses the 22.02 or above PyTorch container, will report an error during model loading in the PyTorch backend when using scripted models that were exported in the legacy format (using our 19.09 or previous PyTorch NGC containers corresponding to PyTorch 1.2.0 or previous releases).

  To load the model successfully in Triton, you need to export the model again by using a recent version of PyTorch.

• Model Analyzer reported values for GPU utilization and GPU power are known to be inaccurate and generally lower than reality

Client Libraries and Examples

Ubuntu 20.04 builds of the client libraries and examples are included in this release in the attached v2.24.0_ubuntu2004.clients.tar.gz file. The SDK is also available for as an Ubuntu 20.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

For windows, the client libraries and some examples are available in the attached tritonserver2.24.0-sdk-win.zip file.

Windows Support

A beta release of Triton for Windows is provided in the attached file:tritonserver2.24.0-win.zip. This is a beta release so functionality is limited and performance is not optimized. Additional features and improved performance will be provided in future releases. Specifically in this release:

• HTTP/REST and GRPC endpoints are supported.

• ONNX models are supported by the ONNX Runtime backend. The ONNX Runtime version is 1.12.0. The CPU, CUDA, and TensorRT execution providers are supported. The OpenVINO execution provider is not supported.

• OpenVINO models are supported. The OpenVINO version is 2021.4.

• Prometheus metrics endpoint is not supported.

• System and CUDA shared memory are not supported.

To use the Windows version of Triton, you must install all the necessary dependencies on your Windows system. These dependencies are available in the Dockerfile.win10.min. The Dockerfile includes the following CUDA-related components:

• CUDA 11.7.1

• cuDNN 8.4.1.5

• TensorRT 8.4.1.5

Jetson Jetpack Support

A release of Triton for JetPack is provided in the attached tar file: A release of Triton for JetPack is provided in the attached tar file: tritonserver2.24.0-jetpack5.0.2.tgz..

• This release supports TensorFlow 2.9.1, TensorFlow 1.15.5, TensorRT 8.4.1.5, Onnx Runtime 1.12.0, PyTorch 1.13.0, Python 3.8 and as well as ensembles.
• Onnx Runtime backend does not support the OpenVino and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.24.0-py3-none-manylinux2014_aarch64.whl[all]