AI Training - Tutorial - Get started with NVIDIA Triton Inference Server and AI Training

Wissensdatenbanken

KB0060355

AI Training - Tutorial - Get started with NVIDIA Triton Inference Server and AI Training


Icons/System/eye-open Created with Sketch. 4603 Ansichten 27.06.2025 AI Training

Objective

Triton Inference Server from NVIDIA allows you to streamline and standardize AI inference.

Triton enables users to deploy, run and scale trained AI models from any framework and on any type of resources: GPU or CPU. It also allows developers to deliver high-performance inference in the Cloud.

The goal of this tutorial is to see how it is possible to deploy Triton Inference Server easily thanks to the OVHcloud AI Training tool.

Overview

Requirements

Instructions

Create the model repository

The model repository is the directory in which you place the AI models you want Triton Inference Server to serve.

Custom models

Triton Inference Sever allows you to deploy several models from multiple Deep Learning and Machine Learning frameworks. For example, Triton Inference Server can serve the following model formats: TensorRT, TensorFlow, PyTorch, ONNX, ...

Follow the NVIDIA Triton Inference Server user guide to know more about the capabilities.

To use custom AI models, the directories and files that compose a model repository must follow a required layout.

├── <model-repository-path>/
    └── <model-name>/
        └── [config.pbtxt]
        └── [<output-labels-file> ...]
        └── <version>/
            └── <model-definition-file>
        └── <version>/
            └── <model-definition-file>
        ...
    └── <model-name>/
        └── [config.pbtxt]
        └── [<output-labels-file> ...]
        └── <version>/
            └── <model-definition-file>
        └── <version>/
            └── <model-definition-file>
        ...
    ...

Find more information on how to use custom models with Triton in this documentation.

An example model repository is included in the docs/examples/model_repository.

We will use these models already trained in this tutorial.

Sample models

First of all, you have to clone locally the Triton Inference Server GitHub repository.

In this command, specify also the version of Triton you are interested in. For this example, we will use the r23.03.

Clone the GitHub repository:

git clone -b r23.03 https://github.com/triton-inference-server/server.git

Before using the repository, you have to fetch any missing model definition files from their public model zoos via the provided script: fetch_models.sh.

The following models will be downloaded:

  • densenet_onnx
  • inception_graphdef
  • simple
  • simple_dyna_sequence
  • simple_identity
  • simple_int8
  • simple_sequence
  • simple_string
cd server/docs/examples
./fetch_models.sh

Check the content and architecture of your model repository with the tree command:

tree model_repository/

You should get:

model_repository/
├── densenet_onnx
│   ├── 1
│   │   └── model.onnx
│   ├── config.pbtxt
│   └── densenet_labels.txt
├── inception_graphdef
│   ├── 1
│   │   └── model.graphdef
│   ├── config.pbtxt
│   └── inception_labels.txt
├── simple
│   ├── 1
│   │   └── model.graphdef
│   └── config.pbtxt
├── simple_dyna_sequence
│   ├── 1
│   │   └── model.graphdef
│   └── config.pbtxt
├── simple_identity
│   ├── 1
│   │   └── model.savedmodel
│   │       └── saved_model.pb
│   └── config.pbtxt
├── simple_int8
│   ├── 1
│   │   └── model.graphdef
│   └── config.pbtxt
├── simple_sequence
│   ├── 1
│   │   └── model.graphdef
│   └── config.pbtxt
└── simple_string
    ├── 1
    │   └── model.graphdef
    └── config.pbtxt

17 directories, 18 files

Upload the models to an Object Storage container

To upload the models to an OVHcloud Object Storage container, you can use the ovhai CLI.

Warning! In order to use Triton Inference Server, you must ensure that your model repository is correctly architected. For example, keep the empty directories, they are necessary for the proper functioning. 

ovhai bucket object upload my-models@GRA model_repository/ --remove-prefix model_repository/

Once your models have been uploaded, check that all your models are available:

ovhai bucket object list my-models@GRA

You should obtain:

DATE                       BYTES    NAME                                              DESCRIPTION ETAG
2023-05-04T12:45:41.643960 31.2 MiB densenet_onnx/1/model.onnx                        Object      c037b5c86fdfc940900b309299657f92
2023-05-04T12:45:41.663860 387 B    densenet_onnx/config.pbtxt                        Object      d0479dc66442434320897957068e5d51
2023-05-04T12:45:41.668330 10.1 KiB densenet_onnx/densenet_labels.txt                 Object      8ea0e7b022e01f34c5cccecbdf90b14f
2023-05-04T12:45:41.664780 91.3 MiB inception_graphdef/1/model.graphdef               Object      4b0459bcf60c61cf2ec759ff89ebbbb7
2023-05-04T12:45:41.672190 340 B    inception_graphdef/config.pbtxt                   Object      7c62d703609e2db1145cbb709ed0bf35
2023-05-04T12:45:41.659140 10.1 KiB inception_graphdef/inception_labels.txt           Object      aafe29c3854eda34886973b95036e66d
2023-05-04T12:45:41.668390 310 B    simple/1/model.graphdef                           Object      de0bafd39f604ea09ca677fc79c70bb5
2023-05-04T12:45:41.663640 370 B    simple/config.pbtxt                               Object      c805202e04542e47a8473181f9dcd0b9
2023-05-04T12:45:41.778020 879 B    simple_dyna_sequence/1/model.graphdef             Object      bce7d0674c80275ccb865a14eaa0c29a
2023-05-04T12:45:41.786770 2.6 KiB  simple_dyna_sequence/config.pbtxt                 Object      ffe0ca21fb2fd3352f976d139869a87e
2023-05-04T12:45:41.783670 531 B    simple_identity/1/model.savedmodel/saved_model.pb Object      eb2ce10be33de0013d98b26348cc8818
2023-05-04T12:45:41.789540 242 B    simple_identity/config.pbtxt                      Object      5994e38fdf5f2789a0dc942dfdeff422
2023-05-04T12:45:41.822000 281 B    simple_int8/1/model.graphdef                      Object      ff193cfa1d711e1e0d647e7f0d09965d
2023-05-04T12:45:41.888750 371 B    simple_int8/config.pbtxt                          Object      fa24596d26d5d03b6c78c02dacabe039
2023-05-04T12:45:41.949530 621 B    simple_sequence/1/model.graphdef                  Object      9c3c5501f283e1a1c955b299b027811c
2023-05-04T12:45:41.971580 2.1 KiB  simple_sequence/config.pbtxt                      Object      fd5b8ae1375aef54998845c4690b7af9
2023-05-04T12:45:42.042580 696 B    simple_string/1/model.graphdef                    Object      4b8239460b9c647e71e0d79e3f42d1c1
2023-05-04T12:45:42.056610 382 B    simple_string/config.pbtxt                        Object      840206b87fb3259eabdc3bb6478ab687

Launch Triton Inference Server with AI Training

Triton is optimized to provide the best inferencing performance by using GPUs. In this tutorial, it will be running with one GPU thanks to the OVHcloud AI Training tool.

First of all, pull the Triton Inference Server Docker image from the NGC repository. In this tutorial, we will use the 23.03-py3 container version.

docker pull nvcr.io/nvidia/tritonserver:23.03-py3

Customize the Triton Inference Server Docker image for OVHcloud AI Tools

To make Triton Inference Server image compatible with AI Training, you have to rebuild the image by giving the OVHcloud user rights on the /workspace content.

For more information on how to build Docker images compatible with OVHcloud AI solutions, please refer to this documentation.

Create the Dockerfile

Your Dockerfile should start with the FROM instruction indicating the parent image to use. In our case we choose to start from the FROM nvcr.io/nvidia/tritonserver:23.03-py3 pulled image:

FROM nvcr.io/nvidia/tritonserver:23.03-py3

Create the home directory and copy your files to it:

WORKDIR /workspace
COPY . /workspace

Give correct access rights to the OVHcloud user (42420:42420):

Don't forget the --user=42420:42420 argument if you want to simulate the exact same behaviour that will occur on AI Training jobs. It executes the Docker container as the specific OVHcloud user (user 42420:42420). 

RUN chown -R 42420:42420 /workspace
ENV HOME=/workspace

Find the complete Dockerfile below:

FROM nvcr.io/nvidia/tritonserver:23.03-py3

WORKDIR /workspace
COPY . /workspace

RUN chown -R 42420:42420 /workspace
ENV HOME=/workspace
Build the Docker image from the Dockerfile

From the directory containing your Dockerfile, run one of the following commands to build your application image:

# Build the image using your machine's default architecture
docker build . -t triton-inference-server:23.03-py3

# Build image targeting the linux/amd64 architecture
docker buildx build --platform linux/amd64 -t triton-inference-server:23.03-py3 .

  • The first command builds the image using your system’s default architecture. This may work if your machine already uses the linux/amd64 architecture, which is required to run containers with our AI products. However, on systems with a different architecture (e.g. ARM64 on Apple Silicon), the resulting image will not be compatible and cannot be deployed.

  • The second command explicitly targets the linux/AMD64 architecture to ensure compatibility with our AI services. This requires buildx, which is not installed by default. If you haven’t used buildx before, you can install it by running: docker buildx install

The dot . argument indicates that your build context (place of the Dockerfile and other needed files) is the current directory.

The -t argument allows you to choose the identifier to give to your image. Usually image identifiers are composed of a name and a version tag <name>:<version>. For this example we chose triton-inference-server:23.03-py3.

Push the server image into the shared registry

The shared registry of AI Training should only be used for testing purposes. Please consider attaching your own registry. More information about this can be found here. The images pushed to this registry are for AI Tools workloads only, and will not be accessible for external uses.

Find the address of your shared registry by launching this command:

ovhai registry list

Log in on the shared registry with your usual AI Platform user credentials:

docker login -u <user> -p <password> <shared-registry-address>

Push the compiled image into the shared registry:

docker tag triton-inference-server:23.03-py3 <shared-registry-address>/triton-inference-server:23.03-py3
docker push <shared-registry-address>/triton-inference-server:23.03-py3

Launch Triton Inference Server in a dedicated job

Triton Inference Server exposes both HTTP/REST and gRPC endpoints. Therefore, clients can communicate with Triton using either an HTTP/REST and GRPC protocols.

Launch Triton Inference Server with OVHcloud AI Training tool to benefit from the GPU power:

ovhai job run --default-http-port 8000 \
              --gpu 1 \
              --volume my-models@GRA:/workspace/models:RO \
              <shared-registry-address>/triton-inference-server:23.03-py3 \
              -- bash -c 'tritonserver --model-repository=/workspace/models'

So that the client can communicate with the Triton Inference Server, you must retrieve the server job IP and URL.

If necessary, install jq as follows:

sudo apt-get install jq

To get the server job url, run the following command:

ovhai job get <job_id> -o json | jq '.status.url' -r

You can verify that Triton Inference Server is running properly.

curl -v <job_url>/v2/health/ready

Then, you can get the server job ip:

ovhai job get <job_id> -o json | jq '.status.ip' -r

Send inference requests

First, you have to use a docker pull to get the client libraries and examples image from NGC.

docker pull nvcr.io/nvidia/tritonserver:23.03-py3-sdk

In the following example, we'll use densenet, an image classification model. We'll see the results obtained on the following image:

DogImage

Test it locally (optional)

docker run -it --rm --net=host nvcr.io/nvidia/tritonserver:23.03-py3-sdk

Then, test the densenet_onnx model on the example image.

To test it locally, you have to specify the server_job_url to be able to send http requests.You can get it by using the following command:

ovhai job get <job_id> -o json | jq '.status.url' -r 

/workspace/install/bin/image_client -u <server_job_url> -m densenet_onnx -c 3 -s INCEPTION /workspace/images/mug.jpg

You should obtain the following result:

Image '/workspace/images/dog.jpg':
12.465825 (264) = CARDIGAN
11.567661 (263) = PEMBROKE
9.468060 (151) = CHIHUAHUA

Customize the Triton client Docker image for OVHcloud AI Tools

Here, the method is exactly the same as the one used previously for the Triton server. Refer to the part concerning Triton server Docker image for more details.

Create the Dockerfile
FROM nvcr.io/nvidia/tritonserver:23.03-py3-sdk

WORKDIR /workspace
COPY . /workspace

RUN chown -R 42420:42420 /workspace
ENV HOME=/workspace
Build the Docker image from the Dockerfile

Launch the following command from the Dockerfile directory to build your application image:

docker build . -t triton-inference-server:23.03-py3-sdk
Push the client image into the shared registry

Push the compiled image into the shared registry:

docker tag triton-inference-server:23.03-py3-sdk <shared-registry-address>/triton-inference-server:23.03-py3-sdk
docker push <shared-registry-address>/triton-inference-server:23.03-py3-sdk

Launch a Triton Client job for inference

Run the example image client inside an AI Training job to perform image classification using the example densenet_onnx model.

To be able to classify your own images, you can create an Object Container in the OVHcloud Object Storage and add your images there:

In this example, we add the image dog.jpg into the bucket. 

ovhai bucket object upload my-images@GRA image_repository/ --remove-prefix image_repository/dog.jpg

To send a request for the densenet_onnx model, use an image from the /workspace/images directory. In this case we ask for the top 3 classifications.

HTTP/REST protocol

You just have to use the -u flag to point at the gRPC endpoint on Triton. The inference server URL will be your server job URL.

You should have obtained this server job IP when launching the Triton Inference Server with AI Training in the step Launch Triton Server in a dedicated job.

You have to specify the port 8000 if you want to use the HTTP/REST protocol.

Launch the inference job for HTTP protocol:

ovhai job run \
     --cpu 1 \
     --volume my-images@GRA:/workspace/images:RO \
     <shared-registry-address>/triton-inference-server:23.03-py3-sdk \
     -- bash -c '/workspace/install/bin/image_client -u <server_job_ip>:8000 -m densenet_onnx -c 3 -s INCEPTION /workspace/images/'

Once your job is RUNNING, check the logs to obtain the results of the classification.

ovhai job logs <job_id> --from job

You should obtain something like that:

2023-08-21T14:47:03Z [job] Image '/workspace/images//dog.jpg':
2023-08-21T14:47:03Z [job] 12.465825 (264) = CARDIGAN
2023-08-21T14:47:03Z [job] 11.567661 (263) = PEMBROKE
2023-08-21T14:47:03Z [job] 9.468060 (151) = CHIHUAHUA
gRPC protocol

If you want to use the gRPC protocol by providing the -i flag, you must also use the -u flag to point at the gRPC endpoint on Triton. The inference server URL will be your server job IP.

You should have obtained this server job IP when launching the Triton Inference Server with AI Training in the step Launch Triton Server in a dedicated job.

You have to specify the port 8001 if you want to use the gRPC protocol.

Launch the inference job for gRPC protocol:

ovhai job run \
     --cpu 1 \
     --volume my-images@GRA:/workspace/images:RO \
     <shared-registry-address>/triton-inference-server:23.03-py3-sdk \
     -- bash -c '/workspace/install/bin/image_client -i grpc -u <server_job_ip>:8001 -m densenet_onnx -c 3 -s INCEPTION /workspace/images/'

Once your job is RUNNING, check the logs to obtain the results of the classification.

ovhai job logs <job_id> --from job

You should get the following result:

2023-08-21T14:47:03Z [job] Image '/workspace/images/dog.jpg':
2023-08-21T14:47:03Z [job] 12.465825 (264) = CARDIGAN
2023-08-21T14:47:03Z [job] 11.567661 (263) = PEMBROKE
2023-08-21T14:47:03Z [job] 9.468060 (151) = CHIHUAHUA

Go further

  • To go further with Triton Inference Server and OVHcloud AI Tools, you can follow these tutorials.

Feedback

Please send us your questions, feedback and suggestions to improve the service: