📌 Part 1: Deploying a Hugging Face Model to SageMaker Endpoint (with Screenshot Guides)

1. Create an IAM Execution Role

Purpose: This role allows SageMaker to access your AWS resources (like ECR, S3, and other services).

Steps:

• Go to the AWS Console → IAM → Roles → Create role.
• Trusted Entity: Choose SageMaker.
• Permissions policies: Attach the following managed policies:
  • AmazonSageMakerFullAccess
  • AmazonS3FullAccess
• Role Name: Name it something like SageMakerExecutionRole.
• After creation, click on the role and copy the Role ARN — you’ll need this when deploying the model.

📷 Screenshot Tip:
✅ Screenshot the Create role page → Select trusted entity as SageMaker
✅ Screenshot the Attach permissions policies step showing both attached policies
✅ Screenshot the Role ARN after creation

2. Launch a SageMaker Notebook Instance

Purpose: This notebook instance will be used to run Python code to deploy your model.

Steps:

• Go to SageMaker → Notebook instances → Create notebook instance.
• Name: huggingface-llm-instance.
• Instance type: ml.t3.medium (or larger if needed).
• IAM Role: Choose the SageMakerExecutionRole you created.
• Git repositories: Skip this.
• Click Create notebook instance.

📷 Screenshot Tip:
✅ Screenshot the Create notebook instance screen showing the name and instance type
✅ Screenshot the IAM role selection

3. Add requirements.txt to Your Notebook Instance

Purpose: To ensure consistent environments, install all necessary Python dependencies at once.

Example requirements.txt:

transformers==4.53.2
torch==2.6.0
langchain==0.3.26
langchain-community==0.0.37
sagemaker==2.219.0
boto3==1.34.112

Upload Instructions:

• Open your notebook instance via JupyterLab.
• Drag and drop the requirements.txt file into your notebook environment.

📷 Screenshot Tip:
✅ Screenshot the JupyterLab interface with the uploaded requirements.txt file visible in the left sidebar

4. Install Python Packages

In a new Jupyter notebook cell, run:

# Install all necessary packages listed in requirements.txt
!pip install -r requirements.txt

✅ Note: The ! prefix is necessary to run shell commands within JupyterLab.

📷 Screenshot Tip:
✅ Screenshot the notebook cell running !pip install -r requirements.txt and the successful installation output

5. Deploy the Hugging Face Model to SageMaker Endpoint

We’ll use sagemaker SDK to deploy the MBZUAI LaMini T5 738M model.

from sagemaker.huggingface import HuggingFaceModel
import sagemaker

# Specify your execution role ARN
role = "arn:aws:iam::YOUR_ACCOUNT_ID:role/SageMakerExecutionRole"

# Initialize the SageMaker session
sess = sagemaker.Session()

# Specify the model checkpoint from Hugging Face
hub_model_id = "MBZUAI/LaMini-T5-738M"

# Configure environment variables for the model container
hub = {
    'HF_MODEL_ID': hub_model_id,
    'HF_TASK': 'text2text-generation',
}

# Create a HuggingFaceModel object
huggingface_model = HuggingFaceModel(
    role=role,
    transformers_version="4.37.0",
    pytorch_version="2.1.0",
    py_version="py310",
    env=hub
)

# Deploy the model to an endpoint
predictor = huggingface_model.deploy(
    initial_instance_count=1,
    instance_type="ml.g4dn.xlarge",  # GPU instance
    endpoint_name="lamini-t5-gpu-endpoint"
)

📷 Screenshot Tip:
✅ Screenshot the notebook cell after the endpoint is successfully deployed showing the predictor output
✅ Screenshot the SageMaker Console → Endpoints showing the new endpoint lamini-t5-gpu-endpoint

6. Finding Your Model Endpoint ARN

• Go to SageMaker Console → Endpoints.
• Find lamini-t5-gpu-endpoint.
• Click on it → Copy the Endpoint ARN (you will need this for Lambda permissions later).

✅ Example ARN Format:

arn:aws:sagemaker:us-east-2:YOUR_ACCOUNT_ID:endpoint/lamini-t5-gpu-endpoint

📷 Screenshot Tip:
✅ Screenshot the Endpoint configuration page showing the ARN field

✅ Summary Checklist:

• Created IAM execution role.
• Launched SageMaker notebook instance.
• Added and installed requirements.txt.
• Deployed Hugging Face model to endpoint.
• Saved endpoint name and ARN for future use.