High-performance serving with Triton Inference Server

Before following the steps in this article, make sure you have the following prerequisites:

• The Azure CLI and the ml extension to the Azure CLI. For more information, see Install, set up, and use the CLI (v2).

  Important

  The CLI examples in this article assume that you are using the Bash (or compatible) shell. For example, from a Linux system or Windows Subsystem for Linux.

• An Azure Machine Learning workspace. If you don't have one, use the steps in the Install, set up, and use the CLI (v2) to create one.

• A working Python 3.8 (or higher) environment.

• You must have additional Python packages installed for scoring and may install them with the code below. They include:

  • NumPy - An array and numerical computing library
  • Triton Inference Server Client - Facilitates requests to the Triton Inference Server
  • Pillow - A library for image operations
  • Gevent - A networking library used when connecting to the Triton Server

pip install numpy
pip install tritonclient[http]
pip install pillow
pip install gevent

• Access to NCv3-series VMs for your Azure subscription.

  Important

  You might need to request a quota increase for your subscription before you can use this series of VMs. For more information, see NCv3-series.

NVIDIA Triton Inference Server requires a specific model repository structure, where there is a directory for each model and subdirectories for the model version. The contents of each model version subdirectory is determined by the type of the model and the requirements of the backend that supports the model. To see all the model repository structure https://github.com/triton-inference-server/server/blob/main/docs/user_guide/model_repository.md#model-files

The information in this document is based on using a model stored in ONNX format, so the directory structure of the model repository is <model-repository>/<model-name>/1/model.onnx. Specifically, this model performs image identification.

The information in this article is based on code samples contained in the azureml-examples repository. To run the commands locally without having to copy/paste YAML and other files, clone the repo and then change directories to the cli directory in the repo:

git clone https://github.com/Azure/azureml-examples --depth 1
cd azureml-examples
cd cli

If you haven't already set the defaults for the Azure CLI, save your default settings. To avoid passing in the values for your subscription, workspace, and resource group multiple times, use the following commands. Replace the following parameters with values for your specific configuration:

• Replace <subscription> with your Azure subscription ID.
• Replace <workspace> with your Azure Machine Learning workspace name.
• Replace <resource-group> with the Azure resource group that contains your workspace.
• Replace <location> with the Azure region that contains your workspace.

az account set --subscription <subscription>
az configure --defaults workspace=<workspace> group=<resource-group> location=<location>

APPLIES TO: Python SDK azure-ai-ml v2 (current)

Before following the steps in this article, make sure you have the following prerequisites:

• An Azure Machine Learning workspace. If you don't have one, use the steps in the Quickstart: Create workspace resources article to create one.

• To install the Python SDK v2, use the following command:

  pip install azure-ai-ml azure-identity
  

  To update an existing installation of the SDK to the latest version, use the following command:

  pip install --upgrade azure-ai-ml azure-identity
  

  For more information, see Install the Python SDK v2 for Azure Machine Learning.

• A working Python 3.8 (or higher) environment.

• You must have additional Python packages installed for scoring and may install them with the code below. They include:

  • NumPy - An array and numerical computing library
  • Triton Inference Server Client - Facilitates requests to the Triton Inference Server
  • Pillow - A library for image operations
  • Gevent - A networking library used when connecting to the Triton Server

  pip install numpy
  pip install tritonclient[http]
  pip install pillow
  pip install gevent
  

• Access to NCv3-series VMs for your Azure subscription.

  Important

  You might need to request a quota increase for your subscription before you can use this series of VMs. For more information, see NCv3-series.

The information in this article is based on the online-endpoints-triton.ipynb notebook contained in the azureml-examples repository. To run the commands locally without having to copy/paste files, clone the repo, and then change directories to the sdk/endpoints/online/triton/single-model/ directory in the repo:

git clone https://github.com/Azure/azureml-examples --depth 1
cd azureml-examples/sdk/python/endpoints/online/triton/single-model/

• An Azure subscription. If you don't have an Azure subscription, create a trial subscription before you begin. Try the trial subscription.

• An Azure Machine Learning workspace. If you don't have one, use the steps in Manage Azure Machine Learning workspaces in the portal, or with the Python SDK to create one.

APPLIES TO: Azure CLI ml extension v2 (current)

This section shows how you can deploy to a managed online endpoint using the Azure CLI with the Machine Learning extension (v2).

1. To avoid typing in a path for multiple commands, use the following command to set a BASE_PATH environment variable. This variable points to the directory where the model and associated YAML configuration files are located:

   BASE_PATH=endpoints/online/triton/single-model
   

2. Use the following command to set the name of the endpoint that will be created. In this example, a random name is created for the endpoint:

   export ENDPOINT_NAME=triton-single-endpt-`echo $RANDOM`
   
   

3. Create a YAML configuration file for your endpoint. The following example configures the name and authentication mode of the endpoint. The one used in the following commands is located at /cli/endpoints/online/triton/single-model/create-managed-endpoint.yml in the azureml-examples repo you cloned earlier:

   create-managed-endpoint.yaml

$schema: https://azuremlschemas.azureedge.net/latest/managedOnlineEndpoint.schema.json
name: my-endpoint
auth_mode: aml_token

1. Create a YAML configuration file for the deployment. The following example configures a deployment named blue to the endpoint defined in the previous step. The one used in the following commands is located at /cli/endpoints/online/triton/single-model/create-managed-deployment.yml in the azureml-examples repo you cloned earlier:

   Important

   For Triton no-code-deployment (NCD) to work, setting type to triton_model​ is required, type: triton_model​. For more information, see CLI (v2) model YAML schema.

   This deployment uses a Standard_NC6s_v3 VM. You might need to request a quota increase for your subscription before you can use this VM. For more information, see NCv3-series.

$schema: https://azuremlschemas.azureedge.net/latest/managedOnlineDeployment.schema.json
name: blue
endpoint_name: my-endpoint
model:
  name: sample-densenet-onnx-model
  version: 1
  path: ./models
  type: triton_model
instance_count: 1
instance_type: Standard_NC6s_v3

APPLIES TO: Python SDK azure-ai-ml v2 (current)

This section shows how you can define a Triton deployment to deploy to a managed online endpoint using the Azure Machine Learning Python SDK (v2).

1. To connect to a workspace, we need identifier parameters - a subscription, resource group and workspace name.

   subscription_id = "<SUBSCRIPTION_ID>"
   resource_group = "<RESOURCE_GROUP>"
   workspace_name = "<AML_WORKSPACE_NAME>"
   

2. Use the following command to set the name of the endpoint that will be created. In this example, a random name is created for the endpoint:

   import random
   
   endpoint_name = f"endpoint-{random.randint(0, 10000)}"
   

3. We use these details above in the MLClient from azure.ai.ml to get a handle to the required Azure Machine Learning workspace. Check the configuration notebook for more details on how to configure credentials and connect to a workspace.

   from azure.ai.ml import MLClient
   from azure.identity import DefaultAzureCredential
   
   ml_client = MLClient(
       DefaultAzureCredential(),
       subscription_id,
       resource_group,
       workspace_name,
   )
   

4. Create a ManagedOnlineEndpoint object to configure the endpoint. The following example configures the name and authentication mode of the endpoint.

   from azure.ai.ml.entities import ManagedOnlineEndpoint
   
   endpoint = ManagedOnlineEndpoint(name=endpoint_name, auth_mode="key")
   

5. Create a ManagedOnlineDeployment object to configure the deployment. The following example configures a deployment named blue to the endpoint defined in the previous step and defines a local model inline.

   from azure.ai.ml.entities import ManagedOnlineDeployment, Model
   
   model_name = "densenet-onnx-model"
   model_version = 1
   
   deployment = ManagedOnlineDeployment(
       name="blue",
       endpoint_name=endpoint_name,
       model=Model(
           name=model_name, 
           version=model_version,
           path="./models",
           type="triton_model"
       ),
       instance_type="Standard_NC6s_v3",
       instance_count=1,
   )
   

This section shows how you can define a Triton deployment on a managed online endpoint using Azure Machine Learning studio.

1. Register your model in Triton format using the following YAML and CLI command. The YAML uses a densenet-onnx model from https://github.com/Azure/azureml-examples/tree/main/cli/endpoints/online/triton/single-model

   create-triton-model.yaml

   name: densenet-onnx-model
   version: 1
   path: ./models
   type: triton_model​
   description: Registering my Triton format model.
   

   az ml model create -f create-triton-model.yaml
   

   The following screenshot shows how your registered model will look on the Models page of Azure Machine Learning studio.

   

2. From studio, select your workspace and then use either the endpoints or models page to create the endpoint deployment:

   • Endpoints page
   • Models page

   1. From the Endpoints page, select Create.

      

   2. Provide a name and authentication type for the endpoint, and then select Next.

   3. When selecting a model, select the Triton model registered previously. Select Next to continue.

   4. When you select a model registered in Triton format, in the Environment step of the wizard, you don't need scoring script and environment.

      

APPLIES TO: Azure CLI ml extension v2 (current)

1. To create a new endpoint using the YAML configuration, use the following command:

   az ml online-endpoint create -n $ENDPOINT_NAME -f $BASE_PATH/create-managed-endpoint.yaml
   
   

2. To create the deployment using the YAML configuration, use the following command:

   az ml online-deployment create --name blue --endpoint $ENDPOINT_NAME -f $BASE_PATH/create-managed-deployment.yaml --all-traffic
   
   

APPLIES TO: Python SDK azure-ai-ml v2 (current)

1. To create a new endpoint using the ManagedOnlineEndpoint object, use the following command:

   endpoint = ml_client.online_endpoints.begin_create_or_update(endpoint)
   

2. To create the deployment using the ManagedOnlineDeployment object, use the following command:

   ml_client.online_deployments.begin_create_or_update(deployment)
   

3. Once the deployment completes, its traffic value will be set to 0%. Update the traffic to 100%.

   endpoint.traffic = {"blue": 100}
   ml_client.online_endpoints.begin_create_or_update(endpoint)
   

1. Complete the wizard to deploy to the endpoint.

   

2. Once the deployment completes, its traffic value will be set to 0%. Update the traffic to 100% from the Endpoint page by clicking Update Traffic on the second menu row.

APPLIES TO: Azure CLI ml extension v2 (current)

Once your deployment completes, use the following command to make a scoring request to the deployed endpoint.

1. To get the endpoint scoring uri, use the following command:

   scoring_uri=$(az ml online-endpoint show -n $ENDPOINT_NAME --query scoring_uri -o tsv)
   scoring_uri=${scoring_uri%/*}
   
   

2. To get an authentication key, use the following command:

   auth_token=$(az ml online-endpoint get-credentials -n $ENDPOINT_NAME --query accessToken -o tsv)
   
   

3. To score data with the endpoint, use the following command. It submits the image of a peacock (https://aka.ms/peacock-pic) to the endpoint:

   python $BASE_PATH/triton_densenet_scoring.py --base_url=$scoring_uri --token=$auth_token --image_path $BASE_PATH/data/peacock.jpg
   
   

   The response from the script is similar to the following text:

   Is server ready - True
   Is model ready - True
   /azureml-examples/cli/endpoints/online/triton/single-model/densenet_labels.txt
   84 : PEACOCK
   

APPLIES TO: Python SDK azure-ai-ml v2 (current)

1. To get the endpoint scoring uri, use the following command:

   endpoint = ml_client.online_endpoints.get(endpoint_name)
   scoring_uri = endpoint.scoring_uri
   

2. To get an authentication key, use the following command: keys = ml_client.online_endpoints.list_keys(endpoint_name) auth_key = keys.primary_key

3. The following scoring code uses the Triton Inference Server Client to submit the image of a peacock to the endpoint. This script is available in the companion notebook to this example - Deploy a model to online endpoints using Triton.

   # Test the blue deployment with some sample data
   import requests
   import gevent.ssl
   import numpy as np
   import tritonclient.http as tritonhttpclient
   from pathlib import Path
   import prepost
   
   img_uri = "http://aka.ms/peacock-pic"
   
   # We remove the scheme from the url
   url = scoring_uri[8:]
   
   # Initialize client handler
   triton_client = tritonhttpclient.InferenceServerClient(
       url=url,
       ssl=True,
       ssl_context_factory=gevent.ssl._create_default_https_context,
   )
   
   # Create headers
   headers = {}
   headers["Authorization"] = f"Bearer {auth_key}"
   
   # Check status of triton server
   health_ctx = triton_client.is_server_ready(headers=headers)
   print("Is server ready - {}".format(health_ctx))
   
   # Check status of model
   model_name = "model_1"
   status_ctx = triton_client.is_model_ready(model_name, "1", headers)
   print("Is model ready - {}".format(status_ctx))
   
   if Path(img_uri).exists():
       img_content = open(img_uri, "rb").read()
   else:
       agent = f"Python Requests/{requests.__version__} (https://github.com/Azure/azureml-examples)"
       img_content = requests.get(img_uri, headers={"User-Agent": agent}).content
   
   img_data = prepost.preprocess(img_content)
   
   # Populate inputs and outputs
   input = tritonhttpclient.InferInput("data_0", img_data.shape, "FP32")
   input.set_data_from_numpy(img_data)
   inputs = [input]
   output = tritonhttpclient.InferRequestedOutput("fc6_1")
   outputs = [output]
   
   result = triton_client.infer(model_name, inputs, outputs=outputs, headers=headers)
   max_label = np.argmax(result.as_numpy("fc6_1"))
   label_name = prepost.postprocess(max_label)
   print(label_name)
   

4. The response from the script is similar to the following text:

   Is server ready - True
   Is model ready - True
   /azureml-examples/sdk/endpoints/online/triton/single-model/densenet_labels.txt
   84 : PEACOCK
   

Triton Inference Server requires using Triton Client for inference, and it supports tensor-typed input. Azure Machine Learning studio doesn't currently support this. Instead, use CLI or SDK to invoke endpoints with Triton.

APPLIES TO: Azure CLI ml extension v2 (current)

1. Once you're done with the endpoint, use the following command to delete it:

   az ml online-endpoint delete -n $ENDPOINT_NAME --yes
   
   

2. Use the following command to archive your model:

   az ml model archive --name $MODEL_NAME --version $MODEL_VERSION
   

APPLIES TO: Python SDK azure-ai-ml v2 (current)

1. Delete the endpoint. Deleting the endpoint also deletes any child deployments, however it will not archive associated Environments or Models.

   ml_client.online_endpoints.begin_delete(name=endpoint_name)
   

2. Archive the model with the following code.

   ml_client.models.archive(name=model_name, version=model_version)
   

1. From the endpoint's page, click Delete in the second row below the endpoint's name.

2. From the model's page, click Delete in the first row below the model's name.