// the find
aws/sagemaker-python-sdk
A library for training and deploying machine learning models on Amazon SageMaker
The official AWS Python SDK for SageMaker — wraps the SageMaker API to let you launch training jobs, deploy endpoints, run batch transforms, and manage MLOps pipelines without writing boto3 calls by hand. V3, released recently, breaks the old Estimator/Model/Predictor pattern in favor of unified ModelTrainer and ModelBuilder classes. This is for teams running ML workloads on AWS who want something above raw boto3 but below a full MLOps platform.
V3's modular split into sagemaker-core, sagemaker-train, sagemaker-serve, and sagemaker-mlops lets you install only what you need instead of dragging in the entire kitchen sink. The new SFTTrainer/DPOTrainer/RLAIFTrainer classes for LLM fine-tuning are genuinely useful — they handle the distributed training config that would otherwise take days to wire up manually. The auto-generated sagemaker-core resource classes mirror the AWS API shape exactly, which means fewer surprises when something works in the console but not in the SDK. Active CI and a real CHANGELOG make it reasonably safe to track upgrades.
V3 drops Estimator, Model, and Predictor with no backwards compatibility — any existing codebase needs a full rewrite, not a migration. The repo ships random Jupyter notebooks (.ipynb files) at the root level alongside production source, which is sloppy and suggests docs and examples are an afterthought bolted on. Local mode testing still requires Docker and a functioning AWS session context, making pure unit testing awkward. Stars are low (2.2k) relative to the actual usage footprint — most people use this via SageMaker Studio and never star the repo — so the community signal here is misleading and you won't find much Stack Overflow coverage of V3-specific issues yet.