Introduction

Triton Inference Server is an open-source, high-performance inference serving software that facilitates the deployment of machine learning models in production environments. Developed by NVIDIA, Triton provides a flexible and scalable solution for serving deep learning models.

Triton Inference Server supports a wide range of deep learning and machine learning frameworks, including TensorFlow, PyTorch, Python, ONNX, NVIDIA® TensorRT™, RAPIDS™ cuML, XGBoost, scikit-learn RandomForest, OpenVINO, custom C++, and more. Its primary use cases are:

  • Serving multiple models from a single server instance.
  • Dynamic model loading and unloading without server restart.
  • Ensemble inference, allowing multiple models to be used together to achieve results.
  • Model versioning for A/B testing and rolling updates.

Resources:

Prerequisites

  1. Pick the Official NVIDIA Triton Server Docker Image:
  1. Create a Model Repository. The model repository is the directory where you place the models that you want Triton to serve. Ensure your models are organized in the following folder structure:
    model-name/
    ├── 1
    │ └── model_file
    └── config.pbtxt
    
  • model-name It can be anything, but keep in mind you need this name to send requests to the correct model. Example: yolo
  • 1/: This subdirectory suggests different model versions might be stored here, with “1” representing the first version.
  • config.pbtxt: This file stores the model configuration in a text-based format, specifying details like input/output tensors, metadata, and backend used.
  • model_file: The model itself.

You can upload multiple models and versions

Deploy on SaladCloud

To deploy a Triton Server on SaladCloud, create a Docker image using the official Triton Server image as a base. Include your models in the image and route IPv6 requests to the Triton Server HTTP port.

Step 1: Create a Dockerfile and add a Base Image

Create a new file called Dockerfile and open it in your preferred text editor. Add the base image to your Dockerfile. Example:

FROM nvcr.io/nvidia/tritonserver:23.09-py3

Step 2: Copy Your Models to the Image

Make sure the folder containing your model is structured as mentioned in the prerequisites, then copy it to your image by updating Dockerfile. Example:

COPY tmp/triton_repo /models

Step 3: Enable IPv6

Refer to the official SaladCloud documentation on enabling IPv6: Enabling IPv6

Create a shell script that will run the Triton server with your model and route traffic from container port 80 to triton http-port 8000 using socat. That shell script will be an entrypoint in the image.

start_services.sh

#!/bin/bash

# Start tritonserver
tritonserver --model-repository=/models --http-port=8000 &

# Start socat
socat TCP6-LISTEN:80,fork TCP4:127.0.0.1:8000 &

# Keep the container running
wait

Step 4: Complete the Dockerfile

Add the script to the Docker image and configure it to run on container startup. Here is the full Dockerfile example:

# Use the Triton server image as the base
FROM nvcr.io/nvidia/tritonserver:23.09-py3

RUN apt-get update && apt-get install -y socat

# Copy the local tmp/triton_repo directory to /models in the image
COPY tmp/triton_repo /models

# Copy the start_services script into the image
COPY start_services.sh /start_services.sh

# Make the script executable
RUN chmod +x /start_services.sh

EXPOSE 80

CMD ["/start_services.sh"]

Step 5: Build your image and push it to docker hub (or the container registry of your choice.)

Step 6: Deploy your image on Salad, using either the Portal or the SaladCloud Public API

Follow this easy steps to deploy your container on salad: Quickstart - SaladCloud

Usage Example

Accessing the Application: Copy the Access Domain Name from the Container Group created above. Detailed instructions on how to find it can be found here: Setup Container Gateway

To test the solution, we deployed a Triton server running a YOLO model. To send requests to it, run the following Python script:

import numpy as np
import gevent.ssl
import tritonclient.http as httpclient
from PIL import Image

# Define a callable that returns an SSLContext with custom settings
def custom_ssl_context_factory():
    context = gevent.ssl.create_default_context()
    return context

# Load and preprocess the image
image_path = "test_pic.jpg"
image = Image.open(image_path).resize((640, 640))
image = np.array(image).astype(np.float32)
image = np.transpose(image, (2, 0, 1))  # Convert to CHW format
image = np.expand_dims(image, axis=0)  # Add batch dimension

# Create Triton client
url = "romaine-fennel.salad.cloud"
model_name = "yolo"

client = httpclient.InferenceServerClient(
    url=url,
    ssl=True,
    ssl_context_factory=custom_ssl_context_factory
)
# Prepare inputs and outputs
inputs = [httpclient.InferInput("images", image.shape, "FP32")]
inputs[0].set_data_from_numpy(image)

outputs = [httpclient.InferRequestedOutput("output0")]

# Perform inference
response = client.infer(model_name, inputs, outputs=outputs)

# Get the output
output_data = response.as_numpy("output0")
print("Inference result:", output_data)

Explanation:

  1. Image Preprocessing: The image is loaded, resized to 640x640, and converted to the required format (CHW).
  2. Triton Client Setup: A Triton client is created to communicate with the server.
  3. Inputs and Outputs: The input tensor is prepared and set with the image data. The output tensor is specified to retrieve the inference results.
  4. Inference: The infer method is called to perform inference, and the results are printed.

Replace romaine-fennel.salad.cloud with your Access Domain Name, and yolo with the name of your model.

By following these steps, you can successfully deploy, manage, and test the NVIDIA Triton Inference Server on SaladCloud, enabling high-performance serving of your machine learning models.