Overview

Overview#

Note

SageMaker Python SDK V3 contains breaking changes from V2. See the V3.0.0 release notes for release notes or visit the V2 documentation.

SageMaker Python SDK v3.0 introduces a modern, modular API for training, fine-tuning, deploying, and managing models on Amazon SageMaker. This release replaces legacy interfaces such as Estimator, Model, and Predictor with unified classes like ModelTrainer and ModelBuilder, reducing boilerplate and simplifying workflows. V3 also introduces comprehensive fine-tuning support with new trainer classes for foundation models.

What’s New in V3#

Model Customization (V3 Exclusive)

Revolutionary foundation model fine-tuning with specialized trainers:

  • SFTTrainer - Supervised fine-tuning for task-specific adaptation
  • DPOTrainer - Direct preference optimization without RL complexity
  • RLAIFTrainer - Reinforcement learning from AI feedback
  • RLVRTrainer - Reinforcement learning from verifiable rewards

Advanced techniques like LoRA, preference optimization, and RLHF that simply don't exist in V2.

Modular Architecture

Separate PyPI packages for specialized capabilities:

  • sagemaker-core - Low-level SageMaker resource management
  • sagemaker-train - Unified training with ModelTrainer
  • sagemaker-serve - Simplified inference with ModelBuilder
  • sagemaker-mlops - ML operations and pipeline management

Single classes replace multiple framework-specific implementations:

  • ModelTrainer replaces PyTorchEstimator, TensorFlowEstimator, SKLearnEstimator, etc.
  • ModelBuilder replaces PyTorchModel, TensorFlowModel, SKLearnModel, etc.

Capabilities#

Training with ModelTrainer#

Unified training interface replacing framework-specific estimators with intelligent defaults and streamlined workflows:

from sagemaker.train import ModelTrainer
from sagemaker.train.configs import InputData

trainer = ModelTrainer(
    training_image="your-training-image",
    role="your-sagemaker-role"
)

train_data = InputData(
    channel_name="training",
    data_source="s3://your-bucket/train-data"
)

training_job = trainer.train(input_data_config=[train_data])

Learn more about Training

Inference with ModelBuilder#

Simplified model deployment and inference with the ModelBuilder

from sagemaker.serve import ModelBuilder

model_builder = ModelBuilder(
    model="your-model",
    model_path="s3://your-bucket/model-artifacts"
)

model = model_builder.build(model_name="my-model")
endpoint = model_builder.deploy(
    endpoint_name="my-endpoint",
    instance_type="ml.m5.xlarge",
    initial_instance_count=1
)
result = endpoint.invoke(
    body={"inputs": "your-input-data"},
    content_type="application/json"
)

Learn more about Inference

ML Operations#

Comprehensive MLOps capabilities for building, deploying, and managing machine learning workflows at scale:

from sagemaker.mlops import Pipeline, TrainingStep, ModelStep

pipeline = Pipeline(name="production-ml-pipeline")

training_step = TrainingStep(
    name="train-model",
    training_config=TrainingConfig(
        algorithm_specification={
            "training_image": "your-training-image"
        }
    )
)

pipeline.add_step(training_step)

Learn more about ML Operations

SageMaker Core#

Low-level, object-oriented access to Amazon SageMaker resources with intelligent defaults and type safety:

from sagemaker.core.resources import TrainingJob

training_job = TrainingJob.create(
    training_job_name="my-training-job",
    role_arn="arn:aws:iam::123456789012:role/SageMakerRole",
    input_data_config=[{
        "channel_name": "training",
        "data_source": "s3://my-bucket/train"
    }]
)

Learn more about SageMaker Core

Getting Started#

Installation#

Install SageMaker Python SDK V3 to get started

Migration from V2#

Key changes when migrating from V2:

  • Replace Estimator classes with ModelTrainer

  • Replace Model classes with ModelBuilder

  • Use structured config objects instead of parameter dictionaries

Next Steps#

Get Started: Follow the Quickstart guide for a hands-on introduction

Contents