Inference Protocols and APIs#

Clients can communicate with Triton using either an HTTP/REST protocol, a GRPC protocol, or by an in-process C API.

HTTP/REST and GRPC Protocols#

Triton exposes both HTTP/REST and GRPC endpoints based on standard inference protocols that have been proposed by the KServe project. To fully enable all capabilities Triton also implements HTTP/REST and GRPC extensions to the KServe inference protocol. GRPC protocol also provides a bi-directional streaming version of the inference RPC to allow a sequence of inference requests/responses to be sent over a GRPC stream. We typically recommend using the unary version for inference requests. The streaming version should be used only if the situation demands it. Some of such use cases can be:

  • Assume a system with multiple Triton server instances running behind a Load Balancer. If a sequence of inference requests is needed to hit the same Triton server instance, a GRPC stream will hold a single connection throughout the lifetime and hence ensure the requests are delivered to the same Triton instance.

  • If the order of requests/responses needs to be preserved over the network, a GRPC stream will ensure that the server receives the requests in the same order as they were sent from the client.

The HTTP/REST and GRPC protocols also provide endpoints to check server and model health, metadata and statistics. Additional endpoints allow model loading and unloading, and inferencing. See the KServe and extension documentation for details.

HTTP Options#

Triton provides the following configuration options for server-client network transactions over HTTP protocol.


Triton allows the on-wire compression of request/response on HTTP through its clients. See HTTP Compression for more details.

GRPC Options#

Triton exposes various GRPC parameters for configuring the server-client network transactions. For usage of these options, refer to the output from tritonserver --help.


These options can be used to configure a secured channel for communication. The server-side options include:

  • --grpc-use-ssl

  • --grpc-use-ssl-mutual

  • --grpc-server-cert

  • --grpc-server-key

  • --grpc-root-cert

For client-side documentation, see Client-Side GRPC SSL/TLS

For more details on overview of authentication in gRPC, refer here.


Triton allows the on-wire compression of request/response messages by exposing following option on server-side:

  • --grpc-infer-response-compression-level

For client-side documentation, see Client-Side GRPC Compression

Compression can be used to reduce the amount of bandwidth used in server-client communication. For more details, see gRPC Compression.

GRPC KeepAlive#

Triton exposes GRPC KeepAlive parameters with the default values for both client and server described here.

These options can be used to configure the KeepAlive settings:

  • --grpc-keepalive-time

  • --grpc-keepalive-timeout

  • --grpc-keepalive-permit-without-calls

  • --grpc-http2-max-pings-without-data

  • --grpc-http2-min-recv-ping-interval-without-data

  • --grpc-http2-max-ping-strikes

For client-side documentation, see Client-Side GRPC KeepAlive.

In-Process Triton Server API#

The Triton Inference Server provides a backwards-compatible C API that allows Triton to be linked directly into a C/C++ application. This API is called the “Triton Server API” or just “Server API” for short. The API is implemented in the Triton shared library which is built from source contained in the core repository. On Linux this library is and on Windows it is tritonserver.dll. In the Triton Docker image the shared library is found in /opt/tritonserver/lib. The header file that defines and documents the Server API is tritonserver.h. Java bindings for In-Process Triton Server API are built on top of tritonserver.h and can be used for Java applications that need to use Tritonserver in-process.

All capabilities of Triton server are encapsulated in the shared library and are exposed via the Server API. The tritonserver executable implements HTTP/REST and GRPC endpoints and uses the Server API to communicate with core Triton logic. The primary source files for the endpoints are and In these source files you can see the Server API being used.

You can use the Server API in your own application as well. A simple example using the Server API can be found in

API Description#

Triton server functionality is encapsulated in a shared library which is built from source contained in the core repository. You can include the full capabilities of Triton by linking the shared library into your application and by using the C API defined in tritonserver.h.

When you link the Triton shared library into your application you are not spawning a separate Triton process, instead, you are including the Triton core logic directly in your application. The Triton HTTP/REST or GRPC protocols are not used to communicate with this Triton core logic, instead all communication between your appliation and the Triton core logic must take place via the Server API.

The top-level abstraction used by Server API is TRITONSERVER_Server, which represents the Triton core logic that is capable of implementing all of the features and capabilities of Triton. A TRITONSERVER_Server object is created by calling TRITONSERVER_ServerNew with a set of options that indicate how the object should be initialized. Use of TRITONSERVER_ServerNew is demonstrated in Once you have created a TRITONSERVER_Server object, you can begin using the rest of the Server API as described below.

Error Handling#

Most Server API functions return an error object indicating success or failure. Success is indicated by return nullptr (NULL). Failure is indicated by returning a TRITONSERVER_Error object. The error code and message can be retrieved from a TRITONSERVER_Error object with TRITONSERVER_ErrorCode and TRITONSERVER_ErrorMessage.

The lifecycle and ownership of all Server API objects is documented in tritonserver.h. For TRITONSERVER_Error, ownership of the object passes to the caller of the Server API function. As a result, your application is responsible for managing the lifecycle of the returned TRITONSERVER_Error object. You must delete the error object using TRITONSERVER_ErrorDelete when you are done using it. Macros such as FAIL_IF_ERR shown in common.h are useful for managing error object lifetimes.

Versioning and Backwards Compatibility#

A typical pattern, demonstrated in and shown below, shows how you can compare the Server API version provided by the shared library against the Server API version that you compiled your application against. The Server API is backwards compatible, so as long as the major version provided by the shared library matches the major version that you compiled against, and the minor version provided by the shared library is greater-than-or-equal to the minor version that you compiled against, then your application can use the Server API.

#include "tritonserver.h"
// Error checking removed for clarity...
uint32_t api_version_major, api_version_minor;
TRITONSERVER_ApiVersion(&api_version_major, &api_version_minor);
if ((TRITONSERVER_API_VERSION_MAJOR != api_version_major) ||
    (TRITONSERVER_API_VERSION_MINOR > api_version_minor)) {
  // Error, the shared library implementing the Server API is older than
  // the version of the Server API that you compiled against.

Non-Inference APIs#

The Server API contains functions for checking health and readiness, getting model information, getting model statistics and metrics, loading and unloading models, etc. The use of these functions is straightforward and some of these functions are demonstrated in and all are documented in tritonserver.h.

Inference APIs#

Performing an inference request requires the use of many Server API functions and objects, as demonstrated in The general usage requires the following steps.

  • Create a TRITONSERVER_ResponseAllocator using TRITONSERVER_ResponseAllocatorNew. You can use the same response allocator for all of your inference requests, or you can create multiple response allocators. When Triton produces an output tensor, it needs a memory buffer into which it can store the contents of that tensor. Triton defers the allocation of these output buffers by invoking callback functions in your application. You communicate these callback functions to Triton with the TRITONSERVER_ResponseAllocator object. You must implement two callback functions, one for buffer allocation and one for buffer free. The signatures for these functions are TRITONSERVER_ResponseAllocatorAllocFn_t and TRITONSERVER_ResponseAllocatorReleaseFn_t as defined in tritonserver.h. In, these callback functions are implemented as ResponseAlloc and ResponseRelease.

  • Create an inference request as a TRITONSERVER_InferenceRequest object. The inference request is where you specify what model you want to use, the input tensors and their values, the output tensors that you want returned, and other request parameters. You create an inference request using TRITONSERVER_InferenceRequestNew. You create each input tensor in the request using TRITONSERVER_InferenceRequestAddInput and set the data for the input tensor using TRITONSERVER_InferenceRequestAppendInputData (or one of the TRITONSERVER_InferenceRequestAppendInputData* variants defined in tritonserver.h). By default, Triton will return all output tensors, but you can limit Triton to only return some outputs by using TRITONSERVER_InferenceRequestAddRequestedOutput.

    To correctly manage the lifecycle of the inference request, you must use TRITONSERVER_InferenceRequestSetReleaseCallback to set a callback into a function in your application. This callback will be invoke by Triton to return ownership of the TRITONSERVER_InferenceRequest object. Typically, in this callback you will just delete the TRITONSERVER_InferenceRequest object by using TRITONSERVER_InferenceRequestDelete. But you may also implement a different lifecycle management; for example, if you are reusing inference request objects you would want to make the object available for reuse.

    You can optionally use TRITONSERVER_InferenceRequestSetId to set a user-defined ID on the request. This ID is not used by Triton but will be returned in the response.

    You can reuse an existing TRITONSERVER_InferenceRequest object for a new inference request. A couple of examples of how this is done and why it is useful are shown in

  • Ask Triton to execute the inference request using TRITONSERVER_ServerInferAsync. TRITONSERVER_ServerInferAsync is a asynchronous call that returns immediately. The inference response is returned via a callback into your application. You register this callback using TRITONSERVER_InferenceRequestSetResponseCallback before you invoke TRITONSERVER_ServerInferAsync. In this callback is InferResponseComplete.

    When you invoke TRITONSERVER_ServerInferAsync and it returns without error, you are passing ownership of the TRITONSERVER_InferenceRequest object to Triton, and so you must not access that object in any way until Triton returns ownership to you via the callback you registered with TRITONSERVER_InferenceRequestSetReleaseCallback.

  • Process the inference response. The inference response is returned to the callback function you registered with TRITONSERVER_InferenceRequestSetResponseCallback. Your callback receives the response as a TRITONSERVER_InferenceResponse object. Your callback takes ownership of the TRITONSERVER_InferenceResponse object and so must free it with TRITONSERVER_InferenceResponseDelete when it is no longer needed.

    The first step in processing a response is to use TRITONSERVER_InferenceResponseError to check if the response is returning an error or if it is returning valid results. If the response is valid you can use TRITONSERVER_InferenceResponseOutputCount to iterate over the output tensors, and TRITONSERVER_InferenceResponseOutput to get information about each output tensor.

    Note that the example uses a std::promise to simply wait for the response, but synchronizing response handling in this way is not required. You can have multiple inference requests in flight at the same time and can issue inference requests from the same thread or from multiple different threads. allows Triton to be linked directly to a C/C++ application. The API is documented in tritonserver.h.

A simple example using the C API can be found in A more complicated example can be found in the source that implements the HTTP/REST and GRPC endpoints for Triton. These endpoints use the C API to communicate with the core of Triton. The primary source files for the endpoints are and

Java bindings for In-Process Triton Server API#

The Triton Inference Server uses Java CPP to create bindings around Tritonserver to create Java API.

The API is documented in Alternatively, the user can refer to the web version API docs generated from A simple example using the Java API can be found in Samples folder which includes which is similar to Please refer to sample usage documentation to learn about how to build and run

In the QA folder, folders starting with L0_java include Java API tests. These can be useful references for getting started, such as the ResNet50 test.

Java API setup instructions#

To use the Tritonserver Java API, you will need to have the Tritonserver library and dependencies installed in your environment. There are two ways to do this:

  1. Use a Tritonserver docker container with

    1. .jar Java bindings to C API (recommended)

    2. maven and build bindings yourself

  2. Build Triton from your environment without Docker (not recommended)

Run Tritonserver container and install dependencies#

To set up your environment with Triton Java API, please follow the following steps:

  1. First run Docker container:

 $ docker run -it --gpus=all -v ${pwd}:/workspace<your container version>-py3 bash
  1. Install jdk:

 $ apt update && apt install -y openjdk-11-jdk
  1. Install maven (only if you want to build the bindings yourself):

$ cd /opt/tritonserver
 $ wget
 $ tar zxvf apache-maven-3.8.4-bin.tar.gz
 $ export PATH=/opt/tritonserver/apache-maven-3.8.4/bin:$PATH

Run Java program with Java bindings Jar#

After ensuring that Tritonserver and dependencies are installed, you can run your Java program with the Java bindings with the following steps:

  1. Place Java bindings into your environment. You can do this by either:

    a. Building Java API bindings with provided build script:

    # Clone Triton client repo. Recommended client repo tag is: main
    $ git clone --single-branch --depth=1 -b <client repo tag>
    # Run build script
    $ source clientrepo/src/java-api-bindings/scripts/

    This will install the Java bindings to /workspace/install/java-api-bindings/tritonserver-java-bindings.jar


    b. Copying “Uber Jar” from Triton SDK container to your environment

    $ id=$(docker run -dit<triton container version>-py3-sdk bash)
    $ docker cp ${id}:/workspace/install/java-api-bindings/tritonserver-java-bindings.jar <Uber Jar directory>/tritonserver-java-bindings.jar
    $ docker stop ${id}
  2. Use the built “Uber Jar” that contains the Java bindings

    $ java -cp <Uber Jar directory>/tritonserver-java-bindings.jar <your Java program>

Build Java bindings and run Java program with Maven#

If you want to make changes to the Java bindings, then you can use Maven to build yourself. You can refer to part 1.a of Run Java program with Java bindings Jar to also build the jar yourself without any modifications to the Tritonserver bindings in JavaCPP-presets. You can do this using the following steps:

  1. Create the JNI binaries in your local repository (/root/.m2/repository) with javacpp-presets/tritonserver

 $ git clone
 $ cd javacpp-presets
 $ mvn clean install --projects .,tritonserver
 $ mvn clean install -f platform --projects ../tritonserver/platform -Djavacpp.platform=linux-x86_64
  1. Create your custom *.pom file for Maven. Please refer to samples/simple/pom.xml as reference for how to create your pom file.

  2. After creating your pom.xml file you can build your application with:

 $ mvn compile exec:java -Djavacpp.platform=linux-x86_64 -Dexec.args="<your input args>"