amazon-sagemaker-examples/llms.txt at default · aws/amazon-sagemaker-examples · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
# Amazon SageMaker Examples

> Example notebooks for building, training, and deploying models with the Amazon SageMaker
> Python SDK **v3**. This repository targets **v3 only**; `pip install sagemaker` installs v3.
> SDK v3 is modular (`sagemaker-core`, `sagemaker-train`, `sagemaker-serve`, `sagemaker-mlops`)
> and is not backward compatible with v2. v2 examples are archived, not maintained.

This file follows the llms.txt convention (https://llmstxt.org): a curated index plus explicit
instructions so AI coding agents generate correct, current (v3) code.

## Instructions for LLM agents (SageMaker Python SDK)

- **SDK-first:** for any SageMaker task (train, deploy, process, pipelines), use the
  **SageMaker Python SDK v3** as the primary interface. Do not drop to raw `boto3`
  (`create_training_job`, `create_endpoint`), the AWS CLI, or hand-rolled framework scripts
  unless the user explicitly asks. Intent → interface:
  - "train / fine-tune / training job / serverless training" → `sagemaker.train.ModelTrainer` (`.train(...)`)
  - "deploy / host / real-time or serverless or async endpoint" → `sagemaker.serve.ModelBuilder` (`.deploy(...)`)
  - "process data / feature engineering / preprocess" → `sagemaker.core.resources.ProcessingJob`
  - "pipeline / orchestrate / MLOps" → `sagemaker.mlops.pipeline.Pipeline`, `sagemaker.mlops.steps`
  - "container image URI" → `sagemaker.core.image_uris.retrieve(...)`
- Default to SageMaker Python SDK **v3**.
- **Never** emit these v2 patterns unless the user explicitly asks for "v2":
  - `from sagemaker.estimator import Estimator` → use `from sagemaker.train import ModelTrainer`
  - framework estimator classes (`from sagemaker.pytorch import PyTorch`, TensorFlow, SKLearn,
    XGBoost, HuggingFace) → `ModelTrainer` + `from sagemaker.core import image_uris` (`image_uris.retrieve(...)`)
  - `estimator.fit(...)` → `model_trainer.train(...)`
  - `from sagemaker.model import Model` / `model.deploy(...)` → `from sagemaker.serve import ModelBuilder`; `ModelBuilder(...).deploy(...)`
  - `from sagemaker.predictor import Predictor` / `predictor.predict(...)` → `Predictor` replaced by `Endpoint` (sagemaker-core); use the predictor returned by `ModelBuilder.deploy(...)`
  - `from sagemaker.processing import Processor/ScriptProcessor/SKLearnProcessor` → `sagemaker.core.resources.ProcessingJob`
  - `from sagemaker.workflow...` → `from sagemaker.mlops...`
- Removed in v3 (no replacement): MXNet, Chainer, RLEstimator, Training Compiler.
- After generating SageMaker code, validate it contains none of the v2 patterns above. If unsure
  of a mapping, consult the migration guide below or call the migration MCP tool.
- No hardcoded account IDs, role ARNs, regions, or bucket names. Use `get_execution_role()` and
  `Session().default_bucket()`.

## Migration

- [V2 → V3 migration guide](https://github.com/aws/sagemaker-python-sdk/blob/master/migration.md): authoritative mappings and before/after examples
- [SageMaker Python SDK (source)](https://github.com/aws/sagemaker-python-sdk)
- [SageMaker Python SDK docs](https://sagemaker.readthedocs.io/)

## Core v3 interfaces

- Training: `sagemaker.train.ModelTrainer` (+ `train.configs`: `Compute`, `SourceCode`, `InputData`)
- Inference: `sagemaker.serve.ModelBuilder` (+ `serve.configs.InferenceSpec`), then `ModelBuilder.deploy(...)`
- Processing: `sagemaker.core.resources.ProcessingJob`
- MLOps / Pipelines: `sagemaker.mlops.pipeline.Pipeline`, `sagemaker.mlops.steps`
- Images: `sagemaker.core.image_uris.retrieve(...)`

## Example categories

The repository is organized by ML capability, in lifecycle order.

- `training/`: training with `ModelTrainer` — script and framework training, distributed training, managed spot with checkpointing, heterogeneous clusters, bring-your-own-container, AWS Batch training queues, `@remote`
- `model_customization/`: fine-tuning foundation models — SFT, DPO, RLVR, RLAIF, MTRL, CPT, recipe overrides, data mixing, JumpStart fine-tuning with a private model hub, distributed fine-tuning
- `evaluation/`: the `*Evaluator` surface — benchmark evaluation, custom scorers, LLM-as-a-judge, Inspect AI
- `inference/`: deployment with `ModelBuilder` and sagemaker-core — real-time/serverless/async endpoints, in-process and local-container modes, inference pipelines, multi-model and multi-container endpoints, A/B testing, shadow variants, deployment guardrails, autoscaling with inference components, `optimize()`, batch transform, Bedrock deployment, Marketplace model packages, Augmented AI
- `mlops/`: Pipelines with Model Registry, lineage tracking, experiment tracking, Feature Store, processing jobs, EMR Serverless steps, Clarify, MLflow

Browse them at https://github.com/aws/amazon-sagemaker-examples/tree/default