Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow: Step-by-Step Guide

Minseok_Song · Jul 3, 2024

Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow

Spoiler (Highlight to read)

This blog series will have several versions, each covering different aspects and techniques. Check out the following resource:
- End-to-end (E2E) sample on Phi-3CookBook

This blog series will have several versions, each covering different aspects and techniques. Check out the following resource: - End-to-end (E2E) sample on Phi-3CookBook

Introduction

Phi-3 is a family of small language models (SLMs) developed by Microsoft that delivers exceptional performance and cost-effectiveness. In this tutorial, you will learn how to fine-tune the Phi-3 model and integrate it with Prompt Flow. By leveraging Azure Machine Learning, and Prompt flow you will establish a workflow for deploying and utilizing custom AI models. This tutorial is divided into three series:

Series 1: Set up Azure resources and Prepare the Environment
- In Series 1, You will learn how to configure the necessary Azure resources and set up your environment for model fine-tuning.
Series 2: Fine-Tune and Deploy the Phi-3 model
- In Series 2, you will fine-tune the Phi-3 model. After fine-tuning, you will deploy the model, making it accessible for integration with Prompt Flow.
Series 3: Integrate the custom phi-3 model with Prompt Flow
- In Series 3, You will Integrate the fine-tuned Phi-3 model with Prompt flow.

Here is an overview of this tutorial.

Note

For more detailed information and to explore additional resources about Phi-3, please visit the Phi-3CookBook.

Prerequisites

Request GPU Quotas in Azure Subscription

In this tutorial, you will learn how to fine-tune and deploy a Phi-3 model, using GPUs. For fine-tuning, you will use the Standard_NC6s_v3 GPU, which requires a quota request. For deployment, you will use the Standard_E4s_v3 CPU, which does not require a quota request.

Note

Only Pay-As-You-Go subscriptions (the standard subscription type) are eligible for GPU allocation; benefit subscriptions are not currently supported.

For those using benefit subscriptions (such as Visual Studio Enterprise Subscription) or those looking to quickly test the fine-tuning and deployment process, this tutorial also provides guidance for fine-tuning with a minimal dataset using a CPU. However, it is important to note that fine-tuning results are significantly better when using a GPU with larger datasets.

In this exercise, you will:

Request GPU Quotas in your Azure Subscription

Request GPU Quotas in Azure Subscription

Type subscriptions in the search bar at the top of the portal page and select Subscriptions from the options that appear.
Select the subscription you'd like to use.
Perform the following tasks:
- Select Usage + quotas from the left side tab.
- Select the Region you'd like to use.
- Select the Quota name to use. For example, select Standard NCSv3 Family vCPUs, which includes the Standard_NC6s_v3 GPU.
- Select the New Quota Request.
- Select the Enter a new limit.
- Inside the New Quota Request page, enter the New limit you'd like to use.
- Inside the New Quota Request page, select Submit to request the GPU quota.

Create Azure Machine Learning Workspace

In this exercise, you will:

Create an Azure Machine Learning Workspace.

Create an Azure Machine Learning Workspace

Type azure machine learning in the search bar at the top of the portal page and select Azure Machine Learning from the options that appear.
Select + Create from the navigation menu.
Select New workspace from the navigation menu.
Perform the following tasks:
- Select your Azure Subscription.
- Select the Resource group to use (create a new one if needed).
- Enter Workspace Name. It must be a unique value.
- Select the Region you'd like to use.
- Select the Storage account to use (create a new one if needed).
- Select the Key vault to use (create a new one if needed).
- Select the Application insights to use (create a new one if needed).
- Select the Container registry to None.
Tip

When you create or use a Storage account in Azure Machine Learning, a container named "azureml" is automatically created within the Storage account. This container is used for storing model artifacts, training outputs, and other data generated during the machine learning process. In this tutorial, you will use the "azureml" container to manage and store all the necessary files and outputs related to our machine learning workflows.

Click to expand...
Select Review + Create.
Select Create.

Add role assignment

To fine-tune and deploy your models, you must first ceate a User Assigned Managed Identity (UAI) and assign it the appropriate permissions. This UAI will be used for authentication during deployment, so it is critical to grant it access to the storage accounts, container registry, and resource group.

In this exercise, you will:

Create User Assigned Managed Identity(UAI).
Add Contributor role assignment to Managed Identity.
Add Storage Blob Data Reader role assignment to Managed Identity.
Add AcrPull role assignment to Managed Identity.

Create User Assigned Managed Identity(UAI)

Type managed identities in the search bar at the top of the portal page and select Managed Identities from the options that appear.
Select + Create.
Perform the following tasks to navigate to Add role assignment page:
- Select your Azure Subscription.
- Select the Resource group to use (create a new one if needed).
- Select the Region you'd like to use.
- Enter the Name. It must be a unique value.
Select Review + create.
Select + Create.

Add Contributor role assignment to Managed Identity

Navigate to the Managed Identity resource that you created.
Select Azure role assignments from the left side tab.
Select +Add role assignment from the navigation menu.
Inside Add role assignment page, Perform the following tasks:
- Select the Scope to Resource group.
- Select your Azure Subscription.
- Select the Resource group to use.
- Select the Role to Contributor.
Select Save.

Add Storage Blob Data Reader role assignment to Managed Identity

Type azure storage accounts in the search bar at the top of the portal page and select Storage accounts from the options that appear.
Select the storage account that associated with the Azure Machine Learning workspace. For example, finetunephistorage.
Perform the following tasks to navigate to Add role assignment page:
- Navigate to the Azure Storage account that you created.
- Select Access Control (IAM) from the left side tab.
- Select + Add from the navigation menu.
- Select Add role assignment from the navigation menu.
Inside Add role assignment page, Perform the following tasks:
- Inside the Role page, type Storage Blob Data Reader in the search bar and select Storage Blob Data Reader from the options that appear.
- Inside the Role page, select Next.
- Inside the Members page, select Assign access to Managed identity.
- Inside the Members page, select + Select members.
- Inside Select managed identities page, select your Azure Subscription.
- Inside Select managed identities page, select the Managed identity to Manage Identity.
- Inside Select managed identities page, select the Manage Identity that you created. For example, finetunephi-managedidentity.
- Inside Select managed identities page, select Select.
- Select Review + assign.

Add AcrPull role assignment to Managed Identity

Type container registries in the search bar at the top of the portal page and select Container registries from the options that appear.
Select the container registry that associated with the Azure Machine Learning workspace. For example, finetunephicontainerregistries
Perform the following tasks to navigate to Add role assignment page:
- Select Access Control (IAM) from the left side tab.
- Select + Add from the navigation menu.
- Select Add role assignment from the navigation menu.
Inside Add role assignment page, Perform the following tasks:
- Inside the Role page, Type AcrPull in the search bar and select AcrPull from the options that appear.
- Inside the Role page, select Next.
- Inside the Members page, select Assign access to Managed identity.
- Inside the Members page, select + Select members.
- Inside Select managed identities page, select your Azure Subscription.
- Inside Select managed identities page, select the Managed identity to Manage Identity.
- Inside Select managed identities page, select the Manage Identity that you created. For example, finetunephi-managedidentity.
- Inside Select managed identities page, select Select.
- Select Review + assign.

Set up the project and install the libraries

Now, you will create a folder to work in and set up a virtual environment to develop a program that interacts with users and uses stored chat history from Azure Cosmos DB to inform its responses.

In this exercise, you will

Create a folder to work inside it.
Create a virtual environment.
Install the required packages.

Create a folder to work inside it

Open a terminal window and type the following command to create a folder named finetune-phi in the default path.
Code:
```
mkdir finetune-phi
```
Type the following command inside your terminal to navigate to the finetune-phi folder you created.
Code:
```
cd finetune-phi
```

Create a virtual environment

Type the following command inside your terminal to create a virtual environment named .venv.
Code:
```
python -m venv .venv
```
Type the following command inside your terminal to activate the virtual environment.
Code:
```
.venv\Scripts\activate.bat
```

Note

If it worked, you should see (.venv) before the command prompt.

Install the required packages

Type the following commands inside your terminal to install the required packages.

Code:

pip install datasets==2.19.1
pip install transformers==4.41.1
pip install azure-ai-ml==1.16.0
pip install torch==2.3.1
pip install trl==0.9.4
pip install promptflow==1.12.0

Set up project files in Visual Studio Code

In this exercise, you will create the essential files for our project. These files include scripts for downloading the dataset, setting up the Azure Machine Learning environment, fine-tuning the Phi-3 model, and deploying the fine-tuned model. You will also create a conda.yml file to set up the fine-tuning environment.

In this exercise, you will:

Create a download_dataset.py file to download the dataset.
Create a setup_ml.py file to set up the Azure Machine Learning environment.
Create a fine_tune.py file in the finetuning_dir folder to fine-tune the Phi-3 model using the dataset.
Create a conda.yml file to setup fine-tuning environment.
Create a deploy_model.py file to deploy the fine-tuned model.
Create a integrate_with_promptflow.py file, to integrate the fine-tuned model and execute the model using Prompt Flow.
Create a flow.dag.yml file, to set up the workflow structure for Promptflow.
Create a config.py file to enter Azure information.

Note

Complete folder structure:

Code:

└── YourUserName
.    └── finetune-phi
.        ├── finetuning_dir
.        │      └── fine_tune.py
.        ├── conda.yml
.        ├── config.py
.        ├── deploy_model.py
.        ├── download_dataset.py
.        ├── flow.dag.yml
.        ├── integrate_with_promptflow.py
.        └── setup_ml.py

Create Project Files

Open Visual Studio Code.
Select File from the menu bar.
Select Open Folder.
Select the finetune-phi folder that you created, which is located at C:\Users\yourUserName\finetune-phi.
In the left pane of Visual Studio Code, right-click and select New File to create a new file named download_dataset.py.
In the left pane of Visual Studio Code, right-click and select New File to create a new file named setup_ml.py.
In the left pane of Visual Studio Code, right-click and select New File to create a new file named deploy_model.py.
In the left pane of Visual Studio Code, right-click and select New Folder to create a new forder named finetuning_dir.
In the finetuning_dir folder, create a new file named fine_tune.py.

Create and Configure conda.yml file

In the left pane of Visual Studio Code, right-click and select New File to create a new file named conda.yml.

Add the following code to the conda.yml file to set up the fine-tuning environment for the Phi-3 model.

Code:

name: phi-3-training-env
channels:
  - defaults
  - conda-forge
dependencies:
  - python=3.10
  - pip
  - numpy<2.0
  - pip:
      - torch~=2.0
      - torchvision~=0.18
      - trl==0.8.6
      - transformers~=4.41
      - datasets~=2.19
      - azureml-core~=1.30
      - azure-storage-blob==12.19
      - azure-ai-ml~=1.16
      - azure-identity~=1.16
      - accelerate~=0.30
      - mlflow==2.14.1
      - azureml-mlflow==1.56.0

Create and Configure config.py file

In the left pane of Visual Studio Code, right-click and select New File to create a new file named config.py.

Add the following code to the config.py file to include your Azure information.

Code:

# Azure settings
AZURE_SUBSCRIPTION_ID = "your_subscription_id"
AZURE_RESOURCE_GROUP_NAME = "your_resource_group_name" # "TestGroup"

# Azure Machine Learning settings
AZURE_ML_WORKSPACE_NAME = "your_workspace_name" # "finetunephi-workspace"

# Azure Managed Identity settings
AZURE_MANAGED_IDENTITY_CLIENT_ID = "your_azure_managed_identity_client_id"
AZURE_MANAGED_IDENTITY_NAME = "your_azure_managed_identity_name" # "finetunephi-mangedidentity"
AZURE_MANAGED_IDENTITY_RESOURCE_ID = f"/subscriptions/{AZURE_SUBSCRIPTION_ID}/resourceGroups/{AZURE_RESOURCE_GROUP_NAME}/providers/Microsoft.ManagedIdentity/userAssignedIdentities/{AZURE_MANAGED_IDENTITY_NAME}"

# Dataset file paths
TRAIN_DATA_PATH = "data/train_data.jsonl"
TEST_DATA_PATH = "data/test_data.jsonl"

# Fine-tuned model settings
AZURE_MODEL_NAME = "your_fine_tuned_model_name" # "finetune-phi-model"
AZURE_ENDPOINT_NAME = "your_fine_tuned_model_endpoint_name" # "finetune-phi-endpoint"
AZURE_DEPLOYMENT_NAME = "your_fine_tuned_model_deployment_name" # "finetune-phi-deployment"

AZURE_ML_API_KEY = "your_fine_tuned_model_api_key"
AZURE_ML_ENDPOINT = "your_fine_tuned_model_endpoint_uri" # "https://{your-endpoint-name}.{your-region}.inference.ml.azure.com/score"

Add Azure Environment Variables

Perform the following tasks to add the Azure Subscription ID:
- Type subscriptions in the search bar at the top of the portal page and select Subscriptions from the options that appear.
- Select the Azure Subscription you are currently using.
- Copy and paste your Subscription ID into the config.py file.
Perform the following tasks to add the Azure Workspace Name:
- Navigate to the Azure Machine Learning resource that you created.
- Copy and paste your account name into the config.py file.
Perform the following tasks to add the Azure Resource Group Name:
- Navigate to the Azure Machine Learning resource that you created.
- Copy and paste your Azure Resource Group Name into the config.py file.
Perform the following tasks to add the Azure Managed Identity name
- Navigate to the Managed Identities resource that you created.
- Copy and paste your Azure Managed Identity name into the config.py file.

Prepare Dataset for Fine-Tuning

In this exercise, you will run the download_data.py file to download the wikitext datasets to your local environment. You will then use this datasets to fine-tune the Phi-3 model in Azure Machine Learning.

In this exercise, you will:

Add code to the download_dataset.py file to download the datasets.
Run the download_dataset.py file to download datasets to your local environment.

Download your dataset using download_data.py

Open the download_data.py file in Visual Studio Code.

Add the following code into download_data.py.

Code:

import json
import os
from datasets import load_dataset
from config import (
    TRAIN_DATA_PATH,
    TEST_DATA_PATH)

def load_and_split_dataset(dataset_name, config_name, split_ratio):
    """
    Load and split a dataset.
    """
    # Load the dataset with the specified name and configuration
    dataset = load_dataset(dataset_name, config_name, split=split_ratio)
    print(f"Original dataset size: {len(dataset)}")
    
    # Split the dataset into train and test sets (80% train, 20% test)
    split_dataset = dataset.train_test_split(test_size=0.2)
    print(f"Train dataset size: {len(split_dataset['train'])}")
    print(f"Test dataset size: {len(split_dataset['test'])}")
    
    return split_dataset

def save_dataset_to_jsonl(dataset, filepath):
    """
    Save a dataset to a JSONL file.
    """
    # Create the directory if it does not exist
    os.makedirs(os.path.dirname(filepath), exist_ok=True)
    
    # Open the file in write mode
    with open(filepath, 'w', encoding='utf-8') as f:
        # Iterate over each record in the dataset
        for record in dataset:
            # Dump the record as a JSON object and write it to the file
            json.dump(record, f)
            # Write a newline character to separate records
            f.write('\n')
    
    print(f"Dataset saved to {filepath}")

def main():
    """
    Main function to load, split, and save the dataset.
    """
    # Load and split the dataset with a specific configuration and split ratio
    dataset = load_and_split_dataset("wikitext", 'wikitext-2-v1', 'train[:1%]')
    
    # Extract the train and test datasets from the split
    train_dataset = dataset['train']
    test_dataset = dataset['test']

    # Save the train dataset to a JSONL file
    save_dataset_to_jsonl(train_dataset, TRAIN_DATA_PATH)
    
    # Save the test dataset to a separate JSONL file
    save_dataset_to_jsonl(test_dataset, TEST_DATA_PATH)

if __name__ == "__main__":
    main()

Tip

Guidance for fine-tuning with a minimal dataset using a CPU

If you want to use a CPU for fine-tuning, this approach is ideal for those with benefit subscriptions (such as Visual Studio Enterprise Subscription) or to quickly test the fine-tuning and deployment process.

Replace dataset = load_and_split_dataset("wikitext", 'wikitext-2-v1', 'train[:1%]') with dataset = load_and_split_dataset("wikitext", 'wikitext-2-v1', 'train[:10]')

Type the following command inside your terminal to run the script and download the dataset to your local environment.
Code:
```
python download_data.py
```
Verify that the datasets were saved successfully to your local finetune-phi/data directory.

Note

Note on dataset size and fine-tuning time

In this tutorial, you use only 1% of the dataset (split='train[:1%]'). This significantly reduces the amount of data, speeding up both the upload and fine-tuning processes. You can adjust the percentage to find the right balance between training time and model performance. Using a smaller subset of the dataset reduces the time required for fine-tuning, making the process more manageable for a tutorial.

Fine-Tune the Phi-3 Model

In this exercise, you will fine-tune the Phi-3 model using the provided dataset. First, you will define the fine-tuning process in the fine_tune.py file. Then, you will configure the Azure Machine Learning environment and initiate the fine-tuning process by running the setup_ml.py file. This script ensures that the fine-tuning occurs within the Azure Machine Learning environment.

By running setup_ml.py, you will run the fine-tuning process in the Azure Machine Learning environment.

In this exercise, you will:

Set up Azure CLI to authenticate environment
Add code to the fine_tune.py file to fine-tune the model.
Add code to and run the setup_ml.py file to initiate the fine-tuning process in Azure Machine Learning.
Run the setup_ml.py file to fine-tune the Phi-3 model using Azure Machine Learning.

Set up Azure CLI

You need to set up Azure CLI to authenticate your environment. Azure CLI allows you to manage Azure resources directly from the command line and provides the credentials necessary for Azure Machine Learning to access these resources. To get started install Azure CLI

Open a terminal window and type the following command to log in to your Azure account.
Code:
```
az login
```
Select your Azure account to use.
Select your Azure subscription to use.

Code:
Tip

Having trouble signing in to Azure? Try using a device code

Open a terminal window and type the following command to log in to your Azure account.

Code:

az login --use-device-code

Visit the website displayed in the terminal window and enter the provided code on that site.

Inside the website, select Next.

Inside the website, select the account to use in this tutorial

Inside the website, select continue to complete login.

After successful login, go back to your terminal and select your Azure subscription to use.

Add code to the fine_tune.py file

Navigate to the finetuning_dir folder and Open the fine_tune.py file in Visual Studio Code.

Add the following code into fine_tune.py.

Code:

import argparse
import sys
import logging
import os
from datasets import Dataset
import torch
import mlflow
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer

# To avoid the INVALID_PARAMETER_VALUE error in MLflow, disable MLflow integration
os.environ["DISABLE_MLFLOW_INTEGRATION"] = "True"

logger = logging.getLogger(__name__)

# Logging setup
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    handlers=[logging.StreamHandler(sys.stdout)],
    level=logging.WARNING
)

# Model and tokenizer setup
pretrained_model_name = "microsoft/Phi-3-mini-4k-instruct"
model_kwargs = dict(
    use_cache=False,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map=None,
    attn_implementation="eager"
)

def initialize_model_and_tokenizer(pretrained_model_name, model_kwargs):
    """
    Initialize the model and tokenizer with the given pretrained model name and arguments.
    """
    model = AutoModelForCausalLM.from_pretrained(pretrained_model_name, **model_kwargs)
    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name)
    tokenizer.model_max_length = 2048
    tokenizer.pad_token = tokenizer.unk_token
    tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
    tokenizer.padding_side = 'right'
    return model, tokenizer

def preprocess_function(examples, tokenizer):
    """
    Preprocess function for tokenizing the dataset.
    """
    tokens = tokenizer(examples['text'], padding="max_length", truncation=True, max_length=512)
    tokens['labels'] = tokens['input_ids'].copy()
    return tokens

def load_and_preprocess_data(train_filepath, test_filepath, tokenizer):
    """
    Load and preprocess the dataset.
    """
    train_dataset = Dataset.from_json(train_filepath)
    test_dataset = Dataset.from_json(test_filepath)
    train_dataset = train_dataset.map(lambda examples: preprocess_function(examples, tokenizer), batched=True)
    test_dataset = test_dataset.map(lambda examples: preprocess_function(examples, tokenizer), batched=True)
    return train_dataset, test_dataset

def main(train_file, eval_file, model_output_dir):
    """
    Main function to fine-tune the model.
    """
    with mlflow.start_run():
        model, tokenizer = initialize_model_and_tokenizer(pretrained_model_name, model_kwargs)
        train_dataset, test_dataset = load_and_preprocess_data(train_file, eval_file, tokenizer)

        # Fine-tuning settings
        finetuning_settings = {
            "bf16": True,
            "do_eval": True,
            "output_dir": model_output_dir,
            "eval_strategy": "epoch",
            "learning_rate": 1e-4,
            "logging_steps": 20,
            "lr_scheduler_type": "linear",
            "num_train_epochs": 3,
            "overwrite_output_dir": True,
            "per_device_eval_batch_size": 4,
            "per_device_train_batch_size": 4,
            "remove_unused_columns": True,
            "save_steps": 500,
            "seed": 0,
            "gradient_checkpointing": True,
            "gradient_accumulation_steps": 1,
            "warmup_ratio": 0.2,
        }

        training_args = TrainingArguments(
            **finetuning_settings
        )
        trainer = SFTTrainer(
            model=model,
            args=training_args,
            train_dataset=train_dataset,
            eval_dataset=test_dataset,
            max_seq_length=2048,
            dataset_text_field="text",
            tokenizer=tokenizer,
            packing=True
        )

        train_result = trainer.train()

        metrics = train_result.metrics

        trainer.log_metrics("train", metrics)

        mlflow.transformers.log_model(
            transformers_model={"model": trainer.model, "tokenizer": tokenizer},
            artifact_path=model_output_dir,  # This is a relative path to save model files within MLflow run
        )

        # Evaluation
        tokenizer.padding_side = 'left'
        metrics = trainer.evaluate()
        metrics["eval_samples"] = len(test_dataset)
        trainer.log_metrics("eval", metrics)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--train-file", type=str, required=True, help="Path to the training data")
    parser.add_argument("--eval-file", type=str, required=True, help="Path to the evaluation data")
    parser.add_argument("--model_output_dir", type=str, required=True, help="Directory to save the fine-tuned model")
    args = parser.parse_args()
    main(args.train_file, args.eval_file, args.model_output_dir)

Save and close the fine_tune.py file.

Add code to the setup_ml.py file

Open the setup_ml.py file in Visual Studio Code.

Add the following code into setup_ml.py.

Code:

import logging
from azure.ai.ml import MLClient, command, Input
from azure.ai.ml.entities import Environment, AmlCompute
from azure.identity import AzureCliCredential
from config import (
    AZURE_SUBSCRIPTION_ID,
    AZURE_RESOURCE_GROUP_NAME,
    AZURE_ML_WORKSPACE_NAME,
    TRAIN_DATA_PATH,
    TEST_DATA_PATH
)

# Constants

# Uncomment the following lines to use a CPU instance for training
# COMPUTE_INSTANCE_TYPE = "Standard_E16s_v3" # cpu
# COMPUTE_NAME = "cpu-e16s-v3"
# DOCKER_IMAGE_NAME = "mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest"
# CONDA_FILE = "conda_cpu.yml"

# Uncomment the following lines to use a GPU instance for training
COMPUTE_INSTANCE_TYPE = "Standard_NC6s_v3"
COMPUTE_NAME = "gpu-nc6s-v3"
DOCKER_IMAGE_NAME = "mcr.microsoft.com/azureml/curated/acft-hf-nlp-gpu:59"
CONDA_FILE = "conda.yml"

LOCATION = "eastus2" # Replace with the location of your compute cluster
FINETUNING_DIR = "./finetuning_dir" # Path to the fine-tuning script
TRAINING_ENV_NAME = "phi-3-training-environment" # Name of the training environment
MODEL_OUTPUT_DIR = "./model_output" # Path to the model output directory in azure ml

# Logging setup to track the process
logger = logging.getLogger(__name__)
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    level=logging.WARNING
)

def get_ml_client():
    """
    Initialize the ML Client using Azure CLI credentials.
    """
    credential = AzureCliCredential()
    return MLClient(credential, AZURE_SUBSCRIPTION_ID, AZURE_RESOURCE_GROUP_NAME, AZURE_ML_WORKSPACE_NAME)

def create_or_get_environment(ml_client):
    """
    Create or update the training environment in Azure ML.
    """
    env = Environment(
        image=DOCKER_IMAGE_NAME,  # Docker image for the environment
        conda_file=CONDA_FILE,  # Conda environment file
        name=TRAINING_ENV_NAME,  # Name of the environment
    )
    return ml_client.environments.create_or_update(env)

def create_or_get_compute_cluster(ml_client, compute_name, COMPUTE_INSTANCE_TYPE, location):
    """
    Create or update the compute cluster in Azure ML.
    """
    try:
        compute_cluster = ml_client.compute.get(compute_name)
        logger.info(f"Compute cluster '{compute_name}' already exists. Reusing it for the current run.")
    except Exception:
        logger.info(f"Compute cluster '{compute_name}' does not exist. Creating a new one with size {COMPUTE_INSTANCE_TYPE}.")
        compute_cluster = AmlCompute(
            name=compute_name,
            size=COMPUTE_INSTANCE_TYPE,
            location=location,
            tier="Dedicated",  # Tier of the compute cluster
            min_instances=0,  # Minimum number of instances
            max_instances=1  # Maximum number of instances
        )
        ml_client.compute.begin_create_or_update(compute_cluster).wait()  # Wait for the cluster to be created
    return compute_cluster

def create_fine_tuning_job(env, compute_name):
    """
    Set up the fine-tuning job in Azure ML.
    """
    return command(
        code=FINETUNING_DIR,  # Path to fine_tune.py
        command=(
            "python fine_tune.py "
            "--train-file ${{inputs.train_file}} "
            "--eval-file ${{inputs.eval_file}} "
            "--model_output_dir ${{inputs.model_output}}"
        ),
        environment=env,  # Training environment
        compute=compute_name,  # Compute cluster to use
        inputs={
            "train_file": Input(type="uri_file", path=TRAIN_DATA_PATH),  # Path to the training data file
            "eval_file": Input(type="uri_file", path=TEST_DATA_PATH),  # Path to the evaluation data file
            "model_output": MODEL_OUTPUT_DIR
        }
    )

def main():
    """
    Main function to set up and run the fine-tuning job in Azure ML.
    """
    # Initialize ML Client
    ml_client = get_ml_client()

    # Create Environment
    env = create_or_get_environment(ml_client)
    
    # Create or get existing compute cluster
    create_or_get_compute_cluster(ml_client, COMPUTE_NAME, COMPUTE_INSTANCE_TYPE, LOCATION)

    # Create and Submit Fine-Tuning Job
    job = create_fine_tuning_job(env, COMPUTE_NAME)
    returned_job = ml_client.jobs.create_or_update(job)  # Submit the job
    ml_client.jobs.stream(returned_job.name)  # Stream the job logs
    
    # Capture the job name
    job_name = returned_job.name
    print(f"Job name: {job_name}")

if __name__ == "__main__":
    main()

Replace COMPUTE_INSTANCE_TYPE, COMPUTE_NAME, and LOCATION with your specific details.

Code:

# Uncomment the following lines to use a GPU instance for training
COMPUTE_INSTANCE_TYPE = "Standard_NC6s_v3"
COMPUTE_NAME = "gpu-nc6s-v3"
...
LOCATION = "eastus2" # Replace with the location of your compute cluster

Code:
Tip

Guidance for fine-tuning with a minimal dataset using a CPU

If you want to use a CPU for fine-tuning, this approach is ideal for those with benefit subscriptions (such as Visual Studio Enterprise Subscription) or to quickly test the fine-tuning and deployment process.

Open the conda.yml file.

Delete torchvision~=0.18 as it is only compatible with GPU environments.

Open the setup_ml file.

Replace COMPUTE_INSTANCE_TYPE, COMPUTE_NAME, and DOCKER_IMAGE_NAME with the following. If you do not have access to Standard_E16s_v3, you can use an equivalent CPU instance or request a new quota.

Replace LOCATION with your specific details.

Code:

# Uncomment the following lines to use a CPU instance for training COMPUTE_INSTANCE_TYPE = "Standard_E16s_v3" # cpu COMPUTE_NAME = "cpu-e16s-v3" DOCKER_IMAGE_NAME = "mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest" CONDA_FILE = "conda.yml" LOCATION = "eastus2" # Replace with the location of your compute cluster

Type the following command to run the setup_ml.py script and start the fine-tuning process in Azure Machine Learning.
Code:
```
python setup_ml.py
```
In this exercise, you successfully fine-tuned the Phi-3 model using Azure Machine Learning. By running the setup_ml.py script, you have set up the Azure Machine Learning environment and initiated the fine-tuning process defined in fine_tune.py file. Please note that the fine-tuning process can take a considerable amount of time. After running the python setup_ml.py command, you need to wait for the process to complete. You can monitor the status of the fine-tuning job by following the link provided in the terminal to the Azure Machine Learning portal. In the next series, you will deploy the fine-tuned model and integrate it with Prompt Flow.

Deploy the fine-tuned model

To integrate the fine-tuned Phi-3 model with Prompt Flow, you need to deploy the model to make it accessible for real-time inference. This process involves registering the model, creating an online endpoint, and deploying the model.

In this exercise, you will:

Set the model name, endpoint name, and deployment name for deployment.
Register the fine-tuned model in the Azure Machine Learning workspace.
Create an online endpoint.
Deploy the registered fine-tuned Phi-3 model.

Set the model name, endpoint name, and deployment name for deployment

Open config.py file.
Replace AZURE_MODEL_NAME = "your_fine_tuned_model_name" with the desired name for your model.
Replace AZURE_ENDPOINT_NAME = "your_fine_tuned_model_endpoint_name" with the desired name for your endpoint.
Replace AZURE_DEPLOYMENT_NAME = "your_fine_tuned_model_deployment_name" with the desired name for your deployment.

Deploy the fine-tuned model

Running the deploy_model.py file automates the entire deployment process. It registers the model, creates an endpoint, and executes the deployment based on the settings specified in the config.py file, which includes the model name, endpoint name, and deployment name.

Open the deploy_model.py file in Visual Studio Code.

Add the following code into deploy_model.py.

Code:

import logging
from azure.identity import AzureCliCredential
from azure.ai.ml import MLClient
from azure.ai.ml.entities import Model, ProbeSettings, ManagedOnlineEndpoint, ManagedOnlineDeployment, IdentityConfiguration, ManagedIdentityConfiguration, OnlineRequestSettings
from azure.ai.ml.constants import AssetTypes

# Configuration imports
from config import (
    AZURE_SUBSCRIPTION_ID,
    AZURE_RESOURCE_GROUP_NAME,
    AZURE_WORKSPACE_NAME,
    AZURE_MANAGED_IDENTITY_RESOURCE_ID,
    AZURE_MANAGED_IDENTITY_CLIENT_ID,
    AZURE_MODEL_NAME,
    AZURE_ENDPOINT_NAME,
    AZURE_DEPLOYMENT_NAME
)

# Constants
JOB_NAME = "your-job-name"
COMPUTE_INSTANCE_TYPE = "Standard_E4s_v3"

deployment_env_vars = {
    "SUBSCRIPTION_ID": AZURE_SUBSCRIPTION_ID,
    "RESOURCE_GROUP_NAME": AZURE_RESOURCE_GROUP_NAME,
    "UAI_CLIENT_ID": AZURE_MANAGED_IDENTITY_CLIENT_ID,
}

# Logging setup
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    level=logging.DEBUG
)
logger = logging.getLogger(__name__)

def get_ml_client():
    """Initialize and return the ML Client."""
    credential = AzureCliCredential()
    return MLClient(credential, AZURE_SUBSCRIPTION_ID, AZURE_RESOURCE_GROUP_NAME, AZURE_WORKSPACE_NAME)

def register_model(ml_client, model_name, job_name):
    """Register a new model."""
    model_path = f"azureml://jobs/{job_name}/outputs/artifacts/paths/model_output"
    logger.info(f"Registering model {model_name} from job {job_name} at path {model_path}.")
    run_model = Model(
        path=model_path,
        name=model_name,
        description="Model created from run.",
        type=AssetTypes.MLFLOW_MODEL,
    )
    model = ml_client.models.create_or_update(run_model)
    logger.info(f"Registered model ID: {model.id}")
    return model

def delete_existing_endpoint(ml_client, endpoint_name):
    """Delete existing endpoint if it exists."""
    try:
        endpoint_result = ml_client.online_endpoints.get(name=endpoint_name)
        logger.info(f"Deleting existing endpoint {endpoint_name}.")
        ml_client.online_endpoints.begin_delete(name=endpoint_name).result()
        logger.info(f"Deleted existing endpoint {endpoint_name}.")
    except Exception as e:
        logger.info(f"No existing endpoint {endpoint_name} found to delete: {e}")

def create_or_update_endpoint(ml_client, endpoint_name, description=""):
    """Create or update an endpoint."""
    delete_existing_endpoint(ml_client, endpoint_name)
    logger.info(f"Creating new endpoint {endpoint_name}.")
    endpoint = ManagedOnlineEndpoint(
        name=endpoint_name,
        description=description,
        identity=IdentityConfiguration(
            type="user_assigned",
            user_assigned_identities=[ManagedIdentityConfiguration(resource_id=AZURE_MANAGED_IDENTITY_RESOURCE_ID)]
        )
    )
    endpoint_result = ml_client.online_endpoints.begin_create_or_update(endpoint).result()
    logger.info(f"Created new endpoint {endpoint_name}.")
    return endpoint_result

def create_or_update_deployment(ml_client, endpoint_name, deployment_name, model):
    """Create or update a deployment."""

    logger.info(f"Creating deployment {deployment_name} for endpoint {endpoint_name}.")
    deployment = ManagedOnlineDeployment(
        name=deployment_name,
        endpoint_name=endpoint_name,
        model=model.id,
        instance_type=COMPUTE_INSTANCE_TYPE,
        instance_count=1,
        environment_variables=deployment_env_vars,
        request_settings=OnlineRequestSettings(
            max_concurrent_requests_per_instance=3,
            request_timeout_ms=180000,
            max_queue_wait_ms=120000
        ),
        liveness_probe=ProbeSettings(
            failure_threshold=30,
            success_threshold=1,
            period=100,
            initial_delay=500,
        ),
        readiness_probe=ProbeSettings(
            failure_threshold=30,
            success_threshold=1,
            period=100,
            initial_delay=500,
        ),
    )
    deployment_result = ml_client.online_deployments.begin_create_or_update(deployment).result()
    logger.info(f"Created deployment {deployment.name} for endpoint {endpoint_name}.")
    return deployment_result

def set_traffic_to_deployment(ml_client, endpoint_name, deployment_name):
    """Set traffic to the specified deployment."""
    try:
        # Fetch the current endpoint details
        endpoint = ml_client.online_endpoints.get(name=endpoint_name)
        
        # Log the current traffic allocation for debugging
        logger.info(f"Current traffic allocation: {endpoint.traffic}")
        
        # Set the traffic allocation for the deployment
        endpoint.traffic = {deployment_name: 100}
        
        # Update the endpoint with the new traffic allocation
        endpoint_poller = ml_client.online_endpoints.begin_create_or_update(endpoint)
        updated_endpoint = endpoint_poller.result()
        
        # Log the updated traffic allocation for debugging
        logger.info(f"Updated traffic allocation: {updated_endpoint.traffic}")
        logger.info(f"Set traffic to deployment {deployment_name} at endpoint {endpoint_name}.")
        return updated_endpoint
    except Exception as e:
        # Log any errors that occur during the process
        logger.error(f"Failed to set traffic to deployment: {e}")
        raise


def main():
    ml_client = get_ml_client()

    registered_model = register_model(ml_client, AZURE_MODEL_NAME, JOB_NAME)
    logger.info(f"Registered model ID: {registered_model.id}")

    endpoint = create_or_update_endpoint(ml_client, AZURE_ENDPOINT_NAME, "Endpoint for finetuned Phi-3 model")
    logger.info(f"Endpoint {AZURE_ENDPOINT_NAME} is ready.")

    try:
        deployment = create_or_update_deployment(ml_client, AZURE_ENDPOINT_NAME, AZURE_DEPLOYMENT_NAME, registered_model)
        logger.info(f"Deployment {AZURE_DEPLOYMENT_NAME} is created for endpoint {AZURE_ENDPOINT_NAME}.")

        set_traffic_to_deployment(ml_client, AZURE_ENDPOINT_NAME, AZURE_DEPLOYMENT_NAME)
        logger.info(f"Traffic is set to deployment {AZURE_DEPLOYMENT_NAME} at endpoint {AZURE_ENDPOINT_NAME}.")
    except Exception as e:
        logger.error(f"Failed to create or update deployment: {e}")

if __name__ == "__main__":
    main()

Perform the following tasks to get the JOB_NAME:
- Navigate to Azure Machine Learning resource that you created.
- Select Studio web URL to open the Azure Machine Learning workspace.
- Select Jobs from the left side tab.
- Select the experiment for fine-tuning. For example, finetunephi.
- Select the job that you created.
- Copy and paste your job Name into the JOB_NAME = "your-job-name" in deploy_model.py file.
Replace COMPUTE_INSTANCE_TYPE with your specific details.
Type the following command to run the deploy_model.py script and start the deployment process in Azure Machine Learning.
Code:
```
python deploy_model.py
```

Warning

To avoid additional charges to your account, make sure to delete the created endpoint in the Azure Machine Learning workspace.

Check deployment status in Azure Machine Learning Workspace

Navigate to Azure Machine Learning resource that you created.
Select Studio web URL to open the Azure Machine Learning workspace.
Select Endpoints from the left side tab.
Select endpoint that you created.
On this page, you can manage the endpoints created during the deployment process.

Integrate the custom Phi-3 model with Prompt Flow

After successfully deploying your fine-tuned model, you can now integrate it with Prompt Flow to use your model in real-time applications, enabling a variety of interactive tasks with your custom Phi-3 model.

In this exercise, you will:

Set api key and endpoint uri of the fine-tuned Phi-3 model.
Add code to the flow.dag.yml file.
Add code to the integrate_with_promptflow.py file.
Test your custom Phi-3 model on Prompt flow.

Set api key and endpoint uri of the fine-tuned Phi-3 model

Navigate to the Azure Machine learning workspace that you created.
Select Endpoints from the left side tab.
Select endpoint that you created.
Select Consume from the navigation menu.
Copy and paste your REST endpoint into the config.py file, replacing AZURE_ML_ENDPOINT = "your_fine_tuned_model_endpoint_uri" with your REST endpoint.
Copy and paste your Primary key into the config.py file, replacing AZURE_ML_API_KEY = "your_fine_tuned_model_api_key" with your Primary key.

Add code to the flow.dag.yml file

Open the flow.dag.yml file in Visual Studio Code.

Add the following code into flow.dag.yml.

Code:

inputs:
  input_data:
    type: string
    default: "Who founded Microsoft?"

outputs:
  answer:
    type: string
    reference: ${integrate_with_promptflow.output}

nodes:
- name: integrate_with_promptflow
  type: python
  source:
    type: code
    path: integrate_with_promptflow.py
  inputs:
    input_data: ${inputs.input_data}

Add code to the integrate_with_promptflow.py file

Open the integrate_with_promptflow.py file in Visual Studio Code.

Add the following code into integrate_with_promptflow.py.

Code:

import logging
import requests
from promptflow.core import tool
import asyncio
import platform
from config import (
    AZURE_ML_ENDPOINT,
    AZURE_ML_API_KEY
)

# Logging setup
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    level=logging.DEBUG
)
logger = logging.getLogger(__name__)

def query_azml_endpoint(input_data: list, endpoint_url: str, api_key: str) -> str:
    """
    Send a request to the Azure ML endpoint with the given input data.
    """
    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {api_key}"
    }
    data = {
        "input_data": [input_data],
        "params": {
            "temperature": 0.7,
            "max_new_tokens": 128,
            "do_sample": True,
            "return_full_text": True
        }
    }
    try:
        response = requests.post(endpoint_url, json=data, headers=headers)
        response.raise_for_status()
        result = response.json()[0]
        logger.info("Successfully received response from Azure ML Endpoint.")
        return result
    except requests.exceptions.RequestException as e:
        logger.error(f"Error querying Azure ML Endpoint: {e}")
        raise

def setup_asyncio_policy():
    """
    Setup asyncio event loop policy for Windows.
    """
    if platform.system() == 'Windows':
        asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())
        logger.info("Set Windows asyncio event loop policy.")

@tool
def my_python_tool(input_data: str) -> str:
    """
    Tool function to process input data and query the Azure ML endpoint.
    """
    setup_asyncio_policy()
    return query_azml_endpoint(input_data, AZURE_ML_ENDPOINT, AZURE_ML_API_KEY)

Type the following command to run the deploy_model.py script and start the deployment process in Azure Machine Learning.
Code:
```
pf flow serve --source ./ --port 8080 --host localhost
```
Here's an example of the results: Now you can chat with your custom Phi-3 model. It is recommended to ask questions based on the data used for fine-tuning.

Congratulations!

You've completed this tutorial

Congratulations! You have successfully completed the tutorial on fine-tuning and integrating custom Phi-3 models with Prompt Flow. This tutorial introduced the simplest method of fine-tuning, avoiding additional techniques such as LoRA or QLoRA, and using MLflow to streamline the fine-tuning and deployment process. Advanced techniques and detailed explanations will be covered in the next series.

Clean Up Azure Resources

Cleanup your Azure resources to avoid additional charges to your account. Go to the Azure portal and delete the following resources:

The Azure Machine Learning resource.
The Azure Machine Learning model endpoint.

Source Code for the Tutorial

You can find the complete source code for this tutorial in the following repository:

skytin1004/Fine-Tune-and-Integrate-Custom-Phi-3-Models-with-Prompt-Flow

Reference

Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow: Step-by-Step Guide

Minseok_Song