Quickstart

Quickstart#

Get started with SageMaker Python SDK V3 in minutes. This guide walks you through the essential steps to train and deploy your first model.

Prerequisites#

  • Python 3.9+ installed

  • AWS account with appropriate permissions

  • AWS credentials configured

Installation#

Install SageMaker Python SDK V3:

pip install sagemaker

Basic Setup#

Import the SDK and create a session:

from sagemaker.core.helper.session_helper import Session, get_execution_role

# Create a SageMaker session
session = Session()
role = get_execution_role()

print(f"Using role: {role}")
print(f"Default bucket: {session.default_bucket()}")

Training Your First Model#

Train a custom PyTorch model using the unified ModelTrainer:

from sagemaker.train import ModelTrainer
from sagemaker.train.configs import SourceCode

# Create ModelTrainer with custom code
model_trainer = ModelTrainer(
    training_image="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:1.13.1-cpu-py39",
    source_code=SourceCode(
        source_dir="./training_code",
        entry_script="train.py",
        requirements="requirements.txt"
    ),
    role=role
)

# Start training (this will create a training job)
training_job = model_trainer.train()
print(f"Training completed: {training_job.name}")

Deploying Your Model#

Deploy the trained model using the V3 workflow: build() → deploy() → invoke():

from sagemaker.serve.model_builder import ModelBuilder
from sagemaker.serve.builder.schema_builder import SchemaBuilder
from sagemaker.serve.utils.types import ModelServer

# Create schema for model input/output
sample_input = [[0.1, 0.2, 0.3, 0.4]]
sample_output = [[0.8, 0.2]]
schema_builder = SchemaBuilder(sample_input, sample_output)

# Create ModelBuilder from training job
model_builder = ModelBuilder(
    model=training_job, # Pass the trained ModelTrainer to use its trained model artifacts
    schema_builder=schema_builder,
    model_server=ModelServer.TORCHSERVE,
    role=role
)

# Build the model
model = model_builder.build(model_name="my-pytorch-model")

# Deploy to endpoint
endpoint = model_builder.deploy(
    endpoint_name="my-endpoint",
    instance_type="ml.m5.large",
    initial_instance_count=1
)

print(f"Endpoint deployed: {endpoint.endpoint_name}")

Making Predictions#

Use your deployed model to make predictions:

import json

# Sample tensor data for prediction
test_data = [[0.5, 0.3, 0.2, 0.1]]

# Make a prediction
result = endpoint.invoke(
    body=json.dumps(test_data),
    content_type="application/json"
)

# Parse the result
prediction = json.loads(result.body.read().decode('utf-8'))
print(f"Prediction: {prediction}")

Cleanup#

Don’t forget to clean up resources to avoid charges:

# Delete the endpoint
endpoint.delete()

print("Endpoint deleted")

Foundation Model Fine-Tuning#

Try V3’s new foundation model fine-tuning capabilities:

from sagemaker.train import SFTTrainer
from sagemaker.train.common import TrainingType

# Fine-tune a foundation model
sft_trainer = SFTTrainer(
    model="huggingface-textgeneration-gpt2",
    training_type=TrainingType.LORA,
    training_dataset="s3://your-bucket/training-data.jsonl",
    role=role
)

# Start fine-tuning
fine_tuning_job = sft_trainer.train()
print(f"Fine-tuning job: {fine_tuning_job.name}")

Next Steps#

Now that you’ve completed the quickstart:

  1. Explore Training: Learn more about Model Training capabilities

  2. Try Inference: Discover advanced Deploy Models for Inference features

  3. Model Customization: Experiment with Customizing Models

  4. Build Pipelines: Create workflows with Implement MLOps

  5. Use SageMaker Core: Access low-level resources with SageMaker Core

Contents