03.DeployPhi3VisionOnAzure.md

Lab 3 - Deploy Phi-3-vision on Azure Machine Learning Service

We use NPU to complete the production deployment of local code, and then we want to introduce the ability to introduce PHI-3-VISION through it to achieve pictures to generate code.

In this introduction, we can quickly build a Model As Service Phi-3 Vision service in Azure Machine Learning Service.

Note： Phi-3 Vision requires computing power to generate content at a faster speed. We need cloud computing power to help us achieve this.

1. Create Azure Machine Learning Service

We need to create an Azure Machine Learning Service in the Azure Portal. If you want to learn how, please visit this link https://learn.microsoft.com/azure/machine-learning/quickstart-create-resources?view=azureml-api-2

2. Choose Phi-3 Vision in Azure Machine Learning Service

3. Deploy Phi-3-Vision in Azure

4. Test Endpoint in Postman

Note

1. The parameters to be transmitted must include Authorization, azureml-model-deployment, and Content-Type. You need to check the deployment information to obtain it.

2. To transmit parameters, Phi-3-Vision needs to transmit an image link. Please refer to the GPT-4-Vision method to transmit parameters, such as

{
  "input_data":{
    "input_string":[
      {
        "role":"user",
        "content":[ 
          {
            "type": "text",
            "text": "You are a Python coding assistant.Please create Python code for image "
          },
          {
              "type": "image_url",
              "image_url": {
                "url": "https://ajaytech.co/wp-content/uploads/2019/09/index.png"
              }
          }
        ]
      }
    ],
    "parameters":{
          "temperature": 0.6,
          "top_p": 0.9,
          "do_sample": false,
          "max_new_tokens": 2048
    }
  }
}

3. Call /score using the Post method

Congratulations ！You have completed the fast PHI-3-VISION deployment and tried how to use pictures to generate code. Next, we can build applications in combination with NPUs and clouds