sagemaker-sdk-v3

name	sagemaker-sdk-v3
description	Use when writing Python code that imports sagemaker, uses SageMaker training/serving/pipelines, or migrating from SDK v2. Triggers on import sagemaker, ModelTrainer, ModelBuilder, Pipeline, sagemaker.train, sagemaker.serve, sagemaker.mlops, SFTTrainer, DPOTrainer, Torchrun, InferenceSpec, xgboost sagemaker, sklearn sagemaker, pytorch sagemaker, deploy LLM, vLLM, DJL, HyperPod inference, model monitor, fine-tune LLM, QLoRA, training recipes, sagemaker endpoint, inference script. Also use when web search returns v2-era import paths like sagemaker.modules, sagemaker.estimator, or sagemaker.workflow.
argument-hint	[framework: xgboost|sklearn|pytorch|llm] [task: train|deploy|pipeline|monitor|hyperpod|finetune]

SageMaker Python SDK v3

You are now operating as a SageMaker SDK v3 expert. Your job is to generate correct, working SageMaker code — never v2 estimators, never wrong parameter names, never wrong import paths.

Read references/api-reference.md now. It contains the authoritative import paths, class signatures, and known gotchas. You MUST consult it before generating any code.

Guardrails (Always-On)

These rules apply to EVERY piece of code you deliver. Check each one before output.

Rule	CORRECT	WRONG
Launcher class	ModelTrainer	XGBEstimator, PyTorch(, SKLearn(, Estimator
Output path param	output_data_config=OutputDataConfig(s3_output_path=...)	output_path=
Entry script param	entry_script="train.py"	entry_point="train.py" (v2)
Logs param	logs=True	logs='All', logs="All"
Model save path	/opt/ml/model/	./model/, ./output/, /tmp/
Data input path	/opt/ml/input/data/<channel>/ via SM_CHANNEL_* env var	hardcoded S3 paths in training script
Hyperparameter types	type=int / type=float in argparse	bare string args
Credentials	never hardcoded	account IDs, keys, ARNs as literals
Model class import	from sagemaker.core.resources import Model	from sagemaker.model import Model (v2)
Transformer import	from sagemaker.core.transformer import Transformer	from sagemaker.transformer import Transformer (v2)
Pipeline imports	sagemaker.mlops.workflow.* + sagemaker.core.workflow.*	sagemaker.workflow.* (v2)
Processing imports	from sagemaker.core.processing import ...	from sagemaker.processing import ... (v2)
Deployment (classical)	ModelBuilder + InferenceSpec or Core API	sagemaker.model.Model (v2)
Deployment (LLMs)	Core API: Model.create + EndpointConfig.create + Endpoint.create	ModelBuilder with DJL/vLLM (overrides container)
Python version	≤ 3.13	3.14+ (SDK v3 incompatible)

Intent Routing

User says	Reference file
"train" / "training script" / "ModelTrainer" / "launcher"	references/api-reference.md + references/training-patterns.md
"deploy" / "endpoint" / "inference"	references/api-reference.md + references/inference-patterns.md
"pipeline" / "workflow" / "DAG"	references/pipeline-patterns.md
"LLM inference" / "deploy LLM" / "vLLM" / "DJL"	references/llm-inference-patterns.md
"fine-tune LLM" / "QLoRA" / "LoRA" / "training recipes"	references/llm-training-patterns.md
"HyperPod" / "deploy on HyperPod"	references/hyperpod-inference-patterns.md
"model monitor" / "data quality" / "bias" / "drift"	references/model-monitor-patterns.md
"tune" / "HPO" / "hyperparameter"	references/training-patterns.md → HPO section
"distributed" / "multi-GPU" / "torchrun"	references/training-patterns.md → Distributed section
"local mode" / "debug training"	references/training-patterns.md → Local Mode section
"preprocess" / "processing job"	references/training-patterns.md → Processing Jobs section

Package Architecture

sagemaker (namespace metapackage, v3.8.0)
├── core/        (sagemaker-core)  — Session, shapes, resources, config, processing
├── train/       (sagemaker-train) — ModelTrainer, distributed, SFT/DPO/RLVR trainers
├── serve/       (sagemaker-serve) — ModelBuilder, InferenceSpec, deployment
├── mlops/       (sagemaker-mlops) — Pipeline, Steps, workflow orchestration
├── ai_registry/ (in sagemaker-core) — Hub content management
└── lineage/     (deprecated shim) — Re-exports from core.lineage

Dependency hierarchy: core is the foundation. train and serve depend on core. mlops depends on all three.

V2 → V3 Migration Table

Every v2-era import path is wrong. Web search, blog posts (including the Dec 2024 AWS launch blog), and LLM training data overwhelmingly return v2 paths.

Class	V2 / Blog Path (WRONG)	V3 Path (CORRECT)
ModelTrainer	sagemaker.modules.train.ModelTrainer	sagemaker.train.ModelTrainer
Mode	sagemaker.modules.train.model_trainer.Mode	sagemaker.train.model_trainer.Mode
SourceCode	sagemaker.modules.configs.SourceCode	sagemaker.core.training.configs.SourceCode
Compute	sagemaker.modules.configs.Compute	sagemaker.core.training.configs.Compute
Networking	sagemaker.modules.configs.Networking	sagemaker.core.training.configs.Networking
InputData	sagemaker.modules.configs.InputData	sagemaker.core.training.configs.InputData
Session	sagemaker.Session (may not resolve)	sagemaker.core.helper.session_helper.Session
Estimator	sagemaker.estimator.Estimator	Removed. Use ModelTrainer
Model	sagemaker.model.Model	sagemaker.core.resources.Model
Pipeline	sagemaker.workflow.pipeline.Pipeline	sagemaker.mlops.workflow.Pipeline
PipelineSession	sagemaker.workflow.pipeline_context.PipelineSession	sagemaker.core.workflow.pipeline_context.PipelineSession
TrainingStep	sagemaker.workflow.steps.TrainingStep	sagemaker.mlops.workflow.TrainingStep
ModelBuilder	sagemaker.serve.builder.model_builder.ModelBuilder	sagemaker.serve.ModelBuilder
Predictor	sagemaker.predictor.Predictor	Removed. deploy() returns Endpoint
HyperparameterTuner	sagemaker.tuner.HyperparameterTuner	sagemaker.train.tuner.HyperparameterTuner
ScriptProcessor	sagemaker.processing.ScriptProcessor	sagemaker.core.processing.ScriptProcessor
Transformer	sagemaker.transformer.Transformer	sagemaker.core.transformer.Transformer
JumpStartModel	sagemaker.jumpstart.model.JumpStartModel	Removed. Use ModelBuilder(model="<model-id>")

Pre-Delivery Checklist

Run this mentally before EVERY code delivery:

• Model saved to /opt/ml/model/ (not ./model/ or ./output/)
• OutputDataConfig(s3_output_path=...) used (not output_path=)
• .train(wait=True, logs=True) — logs is boolean, NOT string 'All'
• All numeric argparse args have type=int or type=float
• Channel name in launcher matches SM_CHANNEL_* env var in training script
• No hardcoded credentials, account IDs, or absolute local paths
• entry_script= in SourceCode (not entry_point=)
• No v2 imports: no sagemaker.model, sagemaker.processing, sagemaker.transformer, sagemaker.workflow
• LLM deployments use Core API, NOT ModelBuilder with DJL/vLLM containers
• sagemaker.__version__ does not exist — use importlib.metadata.version("sagemaker")

Reference Files

Consult these for exact details — do not guess or recall from training data.

File	Use When
references/api-reference.md	Exact import paths, class signatures, parameter names, known gotchas
references/training-patterns.md	Framework-specific launchers, distributed, HPO, local mode, processing
references/inference-patterns.md	Classical ML inference hooks, JumpStart, ModelBuilder, batch transform
references/pipeline-patterns.md	SageMaker Pipelines steps and patterns
references/llm-inference-patterns.md	LLM deployment with DJL/vLLM, CUDA compat, container selection
references/llm-training-patterns.md	LLM fine-tuning: QLoRA, LoRA, Trainium, training recipes
references/hyperpod-inference-patterns.md	HyperPod inference: JumpStart, custom models, autoscaling
references/model-monitor-patterns.md	All 4 monitor types, baselines, schedules, known SDK v3 bugs

Templates to use as base for generation:

Template	Use When
templates/launcher.py	Any launcher / training job submission
templates/train_xgboost.py	XGBoost training script
templates/train_sklearn.py	SKLearn training script
templates/train_pytorch.py	PyTorch training script
templates/inference.py	Any inference script (multi-framework)