SageMaker Python SDK v3.22.0

Highlights

• Added Instance Preferences for multi-instance-type training and processing, including resolved selected-instance fields.
• Added Feature Store UpdateRecord and Standard_V2 storage support.
• Added pre-trainer hyperparameter discovery with list_hyperparameters().
• Added SageMaker Hub validation for raw base-model names.
• Completed PipelineSession support for SFT, DPO, RLAIF, and RLVR trainers.
• Improved private-Hub model and alias resolution for ModelTrainer.
• Added and refreshed image URI configurations for supported inference and training frameworks.

Package versions

• sagemaker-core==2.22.0
• sagemaker-train==1.22.0
• sagemaker-serve==1.22.0
• sagemaker-mlops==1.22.0
• sagemaker==3.22.0

Compatibility

• sagemaker-core now requires boto3>=1.43.90,<2.0.0, ensuring botocore includes the public InstancePreferences API model.
• Internal dependency lower bounds move to the versions released together in 3.22.0.

New features

• feat(train): add list_hyperparameters() for pre-trainer hyperparameter discovery (#6149)
• feat(train): validate raw base model names in SageMaker Hub (#6227)
• feat(core,train): add InstancePreferences for multi-instance-type training and processing (#6249)
• feat(feature-store): add UpdateRecord API and Standard_V2 storage type (#6247)

Bug fixes

• fix(train): add PipelineSession support to SFT, DPO, RLAIF, and RLVR trainers (#6213)
• fix(feature-store): register HubContent Dataset from DatasetBuilder CSV paths (#6212)
• fix(local): detect Docker Compose v2+ when its version has no v prefix (#6231)
• fix(train): resolve private Hub models and aliased references for ModelTrainer (#6201)
• fix(core): resolve default training role from SageMaker configuration (#6228)
• fix(train): validate evaluator models against the live supported-model list (#6217)
• fix(train): complete PipelineSession support for SFT, DPO, RLAIF, and RLVR trainers (#6235)
• fix(train): preserve training_plan_arn during serverful compute reconstruction (#6258)

Other changes

• change(core): add image URI configs for DLC serving frameworks and Amazon Linux 2023 PyTorch (#6220)
• change(core): add image URI configs for vLLM and SGLang (#6218)
• ci(core): add botocore-sync workflows (#6226)
• change(core): add TensorFlow inference 2.20 and training 2.21 image URI configs (#6230)
• change(core): add Ray/llama-cpp CPU images and device-selectable DLC serving configs (#6229)
• change(core): refresh generated image URI configs (55c2a9bd)
• change(serve): emit the JumpStart model ID in ModelBuilder telemetry (#6234)

Tests and documentation

• fix(ci,train): stop integration tests from rerunning the shallow suite (#6216)
• docs(train): add guidance for maintaining shallow integration tests (#6219)
• fix(train): refresh MTRL attached-job integration fixtures (#6259)
• fix(train): add training_plan_arn to serverful test fixtures (#6270)