v3.20.0

fix(serve): pre-deploy JumpStart benchmark data + public HuggingFace …

v3.19.0

SageMaker Python SDK 3.19.0

New Features

• feat(train): Add SequenceLength support for SFT, DPO, RLVR, RLAIF trainers (#5965)
• feat(train): Job notifications, log streaming, and metrics visualization for training jobs (#6051)
• feat(serve): Model Reuse — reuse previously built models across deployments (#6051)
• feat(core): Nova follow-ups — IAM role resolution and training utils (#6051)

Bug Fixes

• fix(serve): propagate additional_model_data_sources for JumpStart models (#6151)
• fix(serve): forward tolerance flags from get_jumpstart_configs (#6137)
• fix(serve): create local model_path dir before using it as download dir (#6147)
• fix(serve): remove IC data-source collapse hack from recommendation deploy (#6101)

v2.257.6

v2.257.6 (2026-08-10)

Enhancements

• Add Zimmer deployment and lineage pipeline step types (v2 backport) (#6153)
• Update SDK to use latest LMI v27 image for sdk v2.x (#5977)

Bug Fixes

• Forward tolerance flags from get_jumpstart_configs (#6136)

Documentation Changes

• Show V2 deprecation note on every documentation page (#6172)

v3.18.0

SageMaker Python SDK v3.18.0

Modular release of the SageMaker Python SDK v3 (pip install sagemaker):
sagemaker-core 2.18.0 · sagemaker-train 1.18.0 · sagemaker-serve 1.18.0 ·
sagemaker-mlops 1.18.0 · sagemaker (meta) 3.18.0

sagemaker-core (2.18.0)

Bug Fixes

• Use the caller-provided session for the sagemaker control-plane client (#6107)
• Remove dev-only endpoint override and fix SageMakerClient singleton pinning (#6120)
• Drop tags before rebuilding TransformJob in transform() (#6110)
• Telemetry INFO logging now prints only once per process (#6083)

sagemaker-train (1.18.0)

Bug Fixes

• Assign SDK-managed channels to instance groups on heterogeneous clusters (#6116)
• Fix agent_run_time inference from the attached input trainer (#6115)

sagemaker-serve (1.18.0)

New Features

• Make BenchmarkResult sweep-aware for concurrency search runs (#6098)

Bug Fixes

• Repack source_code for image_uri / ModelTrainer builds (#6112)

sagemaker-mlops (1.18.0)

New Features

• Add lineage registration to DatasetBuilder (feature-store) (#6014)

Cross-cutting (build / test infrastructure)

• Bump torch in sagemaker-train test data (#5432, #5328)
• Bump black 24.3.0 → 26.3.1 in /requirements/tox (#6119)
• Remove codecov after_n_builds so single-submodule PRs report coverage (#6056)
• Fix undefined STAGE in S3_TRAINING_DATA path (#6134)

Full changelog:
v3.17.0...v3.18.0

v3.17.0

SageMaker Python SDK v3.17.0

Modular release of the SageMaker Python SDK v3 (pip install sagemaker):
sagemaker-core 2.17.0 · sagemaker-train 1.17.0 · sagemaker-serve 1.17.0 ·
sagemaker-mlops 1.17.0 · sagemaker (meta) 3.17.0

sagemaker-core (2.17.0)

New Features

• Update SDK to use the latest LMI v27 image for SDK v3.x (#5976)

Bug Fixes

• Fix source_dir in FrameworkProcessor (#6047)

sagemaker-train (1.17.0)

Bug Fixes

• Correct Networking field names in ModelTrainer intelligent defaults (#6064)
• Resolve MTRL eval base-model ARN against the configured hub (#6040)
• Fall back to the public hub when the private hub lacks the base model (#6092)
• Fix datamixing recipe path (#6073)

Tests

• Doc update and added SFT integ test (#6018)
• Fix role issue in MTRL integ tests (#6070)
• Fix GPU integ test failure due to outdated MPG (#6097)

sagemaker-serve (1.17.0)

New Features

• Support fine-tuned models in the deployment-config API (#6041)

Bug Fixes

• Fix private hub (#6036)
• Support aliased hub content names in private hub deploys (#6039)
• Dedicated INFERENCE_RECOMMENDER telemetry feature + type workload param (#6028)
• Fix EULA check to rely on the HostingEulaUri field (#6077)

Tests

• Move two serve tests to gpu-integ-tests (#6096)

sagemaker-mlops (1.17.0)

New Features

• Wire BatchWriteRecord and ListRecords into ingest_dataframe (#6026)

Tests

• Isolate feature-processor pipeline names to fix flaky integ tests (#6095)

Cross-cutting (test infrastructure)

• Absorb iam:SimulatePrincipalPolicy throttling across suites (#6081)
• Let exhausted IAM throttling fail instead of skipping (#6094)

Full changelog:
v3.16.0...v3.17.0

v3.16.0

v3.16.0 (2026-07-15)

New Features

• feat: actionable guidance for removed v2 interfaces (#6004)
• feat(serve): add SageMaker GenAI inference benchmarking and recommendation (#5874)
• feat(feature-store): add BatchWriteRecord and ListRecords to FeatureGroup (#5983)

Bug Fixes

• fix(iam): scope repo-level ECR actions to prevent false deny in preflight validation (#6024)
• fix: filter full recipe template from serverless train() (#6021)
• Fix sm-train unit tests + use single logger in base trainer (#6030)

Tests

• test(mlops): Skip non-PEP440 version keys in sklearn_latest_version (#6022)

v2.257.5

v2.257.5 (2026-07-14)

Bug Fixes

• Read the Docs build failure (#6023)

3.15.1

v3.15.1 (2026-07-09)

New Features

• feat: Add granular telemetry signals decorator params and error classification (#5963)

Bug Fixes

• fix: always apply evaluator identity keywords and allow explicit domain_id (#5989)
• fix: ModelBuilder resolves private hub artifacts correctly (#5985)
• fix: refresh LLMAsJudgeEvaluator allowed evaluator models (#5987)
• fix: define LAMBDA_ARN_REGEX in finetune_utils to fix RLVRTrainer NameError (#5988)
• fix: RLVR validation bugfix (#6000)
• fix: drop claude-sonnet-4-20250514 from evaluator allowlist (#6009)
• fix(serve): Invoke pip without shell in xgboost install_package (#5981)
• fix: Correct DJL-LMI ISO/ADC accounts + add THF/ISO-E (djl-lmi, huggingface-llm-neuronx) (#5980)

Documentation

• docs: Add AGENTS.md and llms.txt for AI agent v3 guidance (#5982)
• docs: Add SDK-first guidance to AGENTS.md and llms.txt (#5997)
• docs: Add Version Lifecycle page under Getting Started (#5994)
• docs: serve robots.txt opting V2 docs out of AI training crawls (#6003)

Other

• Add Triton Server v26.05 image URI config (#5999)
• Add sklearn 1.4-2-py312 and xgboost 3.2-0 image URI configs (#6008)
• test: Fix/v3 tests (#5996)
• test: wip nova hyperpod integ tests (#5990)

v2.257.4

v2.257.4 (2026-07-13)

Enhancements

• Add v2 -> v3 runtime migration warnings (#5978)
• Update SDK to use latest LMIv26 image for sdk v2.x (#5955)

Bug Fixes

• Harden S3 download path handling (#5984)

Documentation Changes

• Add noindex and canonical tags to deprecated V2 docs (#6001)
• Add deprecation banner to V2 docs (#5991)
• Add v2 -> v3 version lifecycle table to README (#5979)

Other Changes

• Mark capacity-flaky GPU integ tests as slow_test (#5998)
• Disable HF Xet/hf_transfer in serve integ tests (#5992)
• Fix slow tests in v2 (#5944)
• Fix canaries-v2 (#5932)

3.15.0

Nova Forge SDK Reconciliation (v3.15.0)

Release: v3.15.0 (2026-06-22)
Packages: sagemaker 3.15.0 · sagemaker-core 2.15.0 · sagemaker-train 1.15.0 · sagemaker-serve 1.15.0 · sagemaker-mlops 1.15.0

This release converges the standalone Nova Forge SDK into the SageMaker Python SDK, making Nova model customization (training, evaluation, deployment) first-class under the sagemaker.* namespace.

---

New Features

Training — Recipes

• feat(train): add 3-level recipe override support with get_resolved_recipe() (#2034)
• feat: recipe override handling (#2056)
• feat: add Nova-specific recipe validations (#2052)
• feat(train): auto-resolve HyperPod recipe from Hub (#2050)

Training — Compute / Infra

• feat(train): add Serverless / SMTJ / HyperPod support to trainers and evaluators (#2045)
• feat: add infra validation (#2049)

Training — Methods & Data

• feat(train): enable Data Mixing for Nova models (#2047, #2054)
• feat: add is_multimodal utils function (multimodal data auto-detection) (#2033)
• feat: RLVRTrainer Lambda ARN support (#2025)
• feat: RLVR reward Lambda validation (#2036)
• feat: Iterative training for SMTJ serverful and Hyperpod
• feat: New CPTTrainer class

Evaluation

• feat: InspectAI evaluator (#2039)
• feat: add Nova as a target for LLM-as-a-Judge (LLMAJ) (#2059)
• feat: support serverful training job checkpoint resolution in InspectAI evaluator (#2066)

Deploy / Setup / Validation

• feat: add Nova SMI config bounds validation to ModelBuilder (#2040)
• feat: Bedrock deployment from S3 checkpoint

---

Recipes

• fix: reject unknown recipe overrides (serverless + serverful) and untrusted IAM roles (#2071)
• fix: apply recipe overrides to hyperparameters in SMTJ serverful path (#2070)
• fix: recipe override errors in evaluator (#2067)
• fix: hub-content IAM perms, recipe dataset paths, and log markup escaping (#2065)

IAM / Security

• IAM role validations and util methods to create IAM roles

HyperPod / Compute

• fix: add HyperPod validation in train and evaluate (#2051)
• fix: evaluation on HyperPod (#2061)
• fix: resolve HyperPod training image from EKS payload template (#2069)
• fix: skip model_package_group validation when HyperPod compute is provided in CPTTrainer (#2068)
• fix: use compute param in get fine-tuning utils (#2060)

Data / Serverless

• fix: set Converse as S3DataType for Nova models in SMTJ Serverful for SFT and DPO (#2064)
• fix(train): use Converse S3DataType for Nova SFT/DPO in serverless flow (#2079)

Evaluation / Misc

• fix: MLflow error causing OSS model eval to fail (#2075)
• fix: RLVR setup and reward Lambda handling (#2076)

---

Full Changelog: v3.14.0...v3.15.0