SageMaker Python SDK v3 — Claude Code Skill

A Claude Code skill that guarantees correct SageMaker Python SDK v3 code every time.

A Claude Code skill that guarantees correct SageMaker Python SDK v3 code every time. Covers classical ML (XGBoost, SKLearn, PyTorch), LLM inference (DJL/vLLM), LLM fine-tuning (QLoRA/LoRA/Trainium), HyperPod inference, Model Monitor, and SageMaker Pipelines.

Verified against SDK v3.8.0 source code using CodeGraph analysis.

---

What This Skill Does

• Generates correct ModelTrainer-based launchers (never v2 estimators)
• Writes training scripts with proper paths, argparse types, and metric logging
• Produces inference hooks (model_fn, input_fn, predict_fn, output_fn)
• Scaffolds full training projects with README and run commands
• Provides SageMaker Pipelines patterns for multi-step workflows
• Deploys LLMs with DJL LMI / vLLM using the Core API pattern
• Deploys models on HyperPod clusters (JumpStart, custom, autoscaling)
• Sets up Model Monitor (Data Quality, Model Quality, Bias, Explainability)
• Fine-tunes LLMs with QLoRA, LoRA, and AWS training recipes (GPU + Trainium)

Guardrails it enforces on every output:

• ModelTrainer only — never v2 estimators
• output_data_config=OutputDataConfig(s3_output_path=...) — never output_path=
• entry_script= — never entry_point=
• logs=True (boolean) — never logs='All'
• Model saved to /opt/ml/model/ always
• LLM deployments use Core API — NOT ModelBuilder with DJL/vLLM
• All imports use v3 paths — never sagemaker.model, sagemaker.workflow, sagemaker.processing

---

Installation

Global (all projects)

ln -s /path/to/sagemaker-python-sdk-v3-skill ~/.claude/skills/sagemaker-sdk-v3

Project-local

mkdir -p .claude/skills
ln -s /path/to/sagemaker-python-sdk-v3-skill .claude/skills/sagemaker-sdk-v3

---

Usage

Invoke with /sagemaker-sdk-v3 in Claude Code, optionally with arguments:

/sagemaker-sdk-v3 xgboost train
/sagemaker-sdk-v3 sklearn deploy
/sagemaker-sdk-v3 pytorch scaffold
/sagemaker-sdk-v3 pipeline
/sagemaker-sdk-v3 llm deploy
/sagemaker-sdk-v3 llm train
/sagemaker-sdk-v3 hyperpod
/sagemaker-sdk-v3 monitor

Or just describe what you need:

• "Write a SageMaker training job launcher for XGBoost"
• "Deploy Llama 3.1 8B on SageMaker with vLLM"
• "Build a SageMaker Pipeline with preprocessing and training steps"
• "Fine-tune Llama with QLoRA on SageMaker"
• "Set up Model Monitor for my endpoint"

---

Requirements

• Python <= 3.13 (SDK v3 is not compatible with Python 3.14+)
• AWS credentials configured (aws configure or IAM role)

pip install sagemaker boto3

For HyperPod:

pip install sagemaker-hyperpod

---

File Structure

sagemaker-python-sdk-v3-skill/
├── SKILL.md                              # Main skill: guardrails, routing, architecture
├── README.md                             # This file
├── references/
│   ├── api-reference.md                  # SDK v3 imports, class signatures, gotchas
│   ├── training-patterns.md              # Framework-specific launcher patterns
│   ├── inference-patterns.md             # model_fn/input_fn/predict_fn patterns
│   ├── pipeline-patterns.md              # SageMaker Pipelines patterns
│   ├── llm-inference-patterns.md         # LLM deployment: DJL/vLLM, Core API, CUDA compat
│   ├── llm-training-patterns.md          # LLM fine-tuning: QLoRA, LoRA, Trainium, recipes
│   ├── hyperpod-inference-patterns.md    # HyperPod: JumpStart, custom models, autoscaling
│   └── model-monitor-patterns.md         # All 4 monitor types, baselines, schedules, bugs
└── templates/
    ├── launcher.py                       # Canonical SDK v3 launcher
    ├── train_xgboost.py                  # XGBoost training script
    ├── train_sklearn.py                  # SKLearn training script
    ├── train_pytorch.py                  # PyTorch training script
    └── inference.py                      # Multi-framework inference script