Running the Server

For best performance the Triton Inference Server should be run on a system that contains Docker, nvidia-docker, CUDA and one or more supported GPUs, as explained in Running The Inference Server. The inference server can also be run on non-CUDA, non-GPU systems as described in Running The Inference Server On A System Without A GPU.

If you build the inference server outside of Docker, you can then run the inference server without Docker, as explained in Running The Inference Server Without Docker.

Example Model Repository

Before running the Triton Inference Server, you must first set up a model repository containing the models that the server will make available for inferencing.

An example model repository containing a Caffe2 ResNet50, a TensorFlow Inception model, an ONNX densenet model, a simple TensorFlow GraphDef model (used by the simple_client example), and a simple TensorFlow GraphDef model using String tensors (used by the simple_string_client example) are provided in the docs/examples/model_repository directory. Before using the example model repository you must fetch any missing model definition files from their public model zoos. Be sure to checkout the release version of the branch that corresponds to the server you are using (or the master branch if you are using a server build from master):

$ git checkout r20.02
$ cd docs/examples
$ ./

An example ensemble model repository is also provided in the docs/examples/ensemble_model_repository directory. It contains a custom image preprocess model, Caffe2 ResNet50, and an ensemble (used by the ensemble_image_client example).

Before using the example ensemble model repository, in addition to fetching public model definition files as mentioned above, you must build the model definition file for the custom image preprocess model (see Building A Custom Backend for instructions on how to build it). Also note that although ensemble models are fully specified in their model configuration, empty version directories are required for them to be recognized as valid model directories:

$ cd docs/examples
$ mkdir -p ensemble_model_repository/preprocess_resnet50_ensemble/1

Running The Inference Server

Before running the inference server, you must first set up a model repository containing the models that the server will make available for inferencing. Section Model Repository describes how to create your own model repository. You can also use Example Model Repository to set up an example model repository.

Assuming the sample model repository is available in /path/to/model/repository, the following command runs the container you pulled from NGC or built locally:

$ nvidia-docker run --rm --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -p8000:8000 -p8001:8001 -p8002:8002 -v/path/to/model/repository:/models <tritonserver image name> trtserver --model-repository=/models

Where <tritonserver image name> will be something like if you pulled the container from the NGC registry, or tritonserver if you built it from source.

The nvidia-docker -v option maps /path/to/model/repository on the host into the container at /models, and the --model-repository option to the server is used to point to /models as the model repository.

The -p flags expose the container ports where the inference server listens for HTTP requests (port 8000), listens for GRPC requests (port 8001), and reports Prometheus metrics (port 8002).

The --shm-size and --ulimit flags are recommended to improve the server’s performance. For --shm-size the minimum recommended size is 1g but smaller or larger sizes may be used depending on the number and size of models being served.

For more information on the Prometheus metrics provided by the inference server see Metrics.

Running The Inference Server On A System Without A GPU

On a system without GPUs, the inference server should be run using docker instead of nvidia-docker, but is otherwise identical to what is described in Running The Inference Server:

$ docker run --rm --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -p8000:8000 -p8001:8001 -p8002:8002 -v/path/to/model/repository:/models <tritonserver image name> trtserver --model-repository=/models

Because a GPU is not available, the inference server will be unable to load any model configuration that requires a GPU or that specifies a GPU instance by an instance-group configuration.

Running The Inference Server Without Docker

After building the inference server outside of Docker, the trtserver executable will be in builddir/trtis/install/bin and the required shared libraries will be in builddir/trtis/install/lib. The trtserver executable and libraries are configured to be installed and executed from the /opt/tensorrtserver directory, so copy builddir/trtis/install/* to /opt/tensorrtserver/. . Then execute trtserver with the desired arguments:

$ /opt/tensorrtserver/bin/trtserver --model-repository=/models

Checking Inference Server Status

The simplest way to verify that the inference server is running correctly is to use the Status API to query the server’s status. From the host system use curl to access the HTTP endpoint to request server status. The response is protobuf text showing the status for the server and for each model being served, for example:

$ curl localhost:8000/api/status
id: "inference:0"
version: "0.6.0"
uptime_ns: 23322988571
model_status {
  key: "resnet50_netdef"
  value {
    config {
      name: "resnet50_netdef"
      platform: "caffe2_netdef"
    version_status {
      key: 1
      value {
        ready_state: MODEL_READY
ready_state: SERVER_READY

This status shows configuration information as well as indicating that version 1 of the resnet50_netdef model is MODEL_READY. This means that the server is ready to accept inferencing requests for version 1 of that model. A model version ready_state will show up as MODEL_UNAVAILABLE if the model failed to load for some reason.