Release 1.3.24

New Features

Training Job Email Notifications (SMTJ, SMHP)

• Added support for job notifications via email for training on the SMTJ and SMHP platforms via the NotificationManager class.
• Provided CloudFormation notification templates to allow users to deploy a stack of resources for receiving status updates.

SMHP Cluster Instance Group Scaling

• Added the scale_cluster() and get_instance_groups() functions to the SMHPRuntimeManager to allow users to scale their SMHP cluster restricted instance groups (RIGs) from the SDK.

RFT Lambda Deploy and Validation via RuntimeManager

• Added deploy_lambda() to RuntimeManager to package a local python file into zip, create/update a Lambda function, and store the resulting ARN on rft_lambda_arn.
• Added validate_lambda() method to invoke a deployed lambda with sample data from S3 to validate correctness before training.

Enhancements

• Simplified ECS infrastructure for RFT Multi-Turn and improved reliability
  • Removed Docker/ECR image push requirements
  • Switched ECS tasks to a public Amazon Linux 2023 base image
  • Improved CF stack state and virtual environment validation
  • Note: User might need to update their IAM policies to remove ECR permissions and add new SSM/EC2/CloudFormation permissions for RFT multiturn.
• Updated troubleshooting documentation to provide guidance on Bedrock deployment with permission restraints.
• Improved CloudWatchLogMonitor usability by allowing job start time resolution from platform APIs when not explicitly provided.

Bug Fixes

• Fixed evaluate() passing the incorrect training method to the RecipeBuilder to instead pass TrainingMethod.EVALUATION.
• Added an optional “recipe” flag to the “validation_config” to bypass recipe validation when needed.

Release 1.3.18

Release 1.3.18

Bug Fixes

• Fixed incorrect input data upload format for RFT SageMaker Training Jobs

Release 1.3.17

Release 1.3.17

Bug Fixes

• Fixed incorrect input data upload format for Nova 1.0 SageMaker Training Jobs

Release 1.3.16

Release 1.3.16

Bug Fixes

• Removed requirements.txt to address installation issue for Mac OS

Release 1.3.14

Release 1.3.14

The package has been renamed to Nova Forge SDK.
All internal imports are now referenced as amzn_nova_forge.

Upgrade Instructions

To use the latest version of Nova Forge SDK, run:

# Remove the old SDK
pip uninstall amzn-nova-customization-sdk

# Install the new SDK
pip install amzn-nova-forge

New Features

Bedrock Fine‑tuning

• Added support for Supervised Fine‑Tuning (SFT) and Reinforcement Fine‑Tuning (RFT) with Low‑Rank Adaptation (LoRA) on Bedrock.
• Introduced a new runtime manager, platform support, and extensive infrastructure to integrate Bedrock as a training platform.
• Implemented job creation, status tracking, and cleanup for Bedrock jobs.

Limitations & Constraints

Limitation	Details
Evaluation & batch inference	Not supported on Bedrock
Supported methods	Only SFT LoRA and RFT LoRA
Validation datasets	Not supported for Nova Lite 2 models
Monitoring	MLFlow monitoring is not available for Bedrock jobs

Serverless SageMaker Training Jobs

• Enabled serverless SageMaker Training Jobs for SFT and Direct Preference Optimization (DPO) with Full‑Rank and LoRA.

Plot Training Metrics

• Enabled plotting of:
  • Training loss curve for SFT and Continuous Pre‑Training (CPT) jobs.
  • Reward score curve for RFT jobs.

Enhancements

• Data Prep operations refactored
  These changes are backward compatible via deprecation warnings:

  • transform()/validate() → method parameter renamed to training_method; the new parameter selects the operation type (default: SCHEMA).
  • column_mappings moved from loader constructor to transform() kwarg.
  • save_data() renamed to save().
  • split_data() renamed to split().
  • New guide: docs/data_prep.md.
  • Updated README and QuickStart notebook.

• SageMaker SDK V3 Upgrade

• Security documentation
  SECURITY.md contents moved to README.md.

• Added Image URI override validation support.

Bug Fixes

• Added missing IAM permissions for the RFT multiturn guidance.
• Updated replicas override to use the customer‑provided instance_count instead of the recipe template value.

Release 1.1.2

Release 1.1.2

New Features

Reinforcement Fine-Tuning (RFT) Multiturn

• Added RFT Multiturn data transformer and validator.
• Added support for session restoration using dump() and load().
• Enabled custom starter kit path via starter_kit_path to use custom environments.
• Deprecated start_training_environment() and start_evaluation_environment() in favor of start_environment().
• Added a feature to detect duplicate environments running on same stack and infrastructure.

Reinforcement Fine-Tuning (RFT) Singleturn

• Introduced RFT Lambda verification via the validation_config parameter in NovaModelCustomizer.

Job Caching

• Added enable_job_caching to NovaModelCustomizer to save job results to disk and reuse them on subsequent calls with matching parameters.

Enhancements

Validation for SageMaker Inference (SMI) Configs

• Added validation for context length and concurrency settings based on model and instance type when deploying to SMI endpoints.

Bug Fixes

• Fixed validation logic for save_steps for RFT multiturn to accept integer values.
• Removed pinned version constraint for numpy (numpy<=2.2.6) to resolve dependency conflicts.

Release 1.0.97

Bug Features

• Addresses bug reported in #31
• Updates error message with correct Github repo link

Release 1.0.96

New Features

SageMaker Model Deployment

• Introduced SageMaker as a deployment platform option alongside Bedrock On-Demand and Provisioned Throughput

• Introduced invoke_inference() method for real-time inference supporting both SageMaker and Bedrock platforms

Reinforcement Fine-Tuning (RFT) Multiturn

• Enabled Reinforcement Fine-Tuning (RFT) multiturn training and evaluation for Nova 2.0 models (Forge-subscribed feature only)

• Introduced CustomEnvironment and RFTMultiturnInfrastructure classes for setting up custom Reinforcement Learning (RL) environments on local, EC2, or ECS infrastructure for training and evaluation

Inference to SageMaker and Bedrock Endpoints

• Added support for streaming and non-streaming inference requests for SageMaker text models

• Added support for non-streaming inference requests for Bedrock text models

MLflow Monitoring

• Added get_presigned_url() function to MLflowMonitor for generating presigned URLs to access the MLflow tracking server UI

Enhancements

Region Support

• Added support for us-west-2

Model deployment interface (NovaModelCustomizer.deploy)

• Renamed pt_units parameter to unit_count for broader applicability across deployment platforms

• Replaced bedrock_execution_role_name parameter with execution_role_name for flexible role configuration across platforms

• Added sagemaker_instance_type parameter with default value ml.g5.4xlarge for SageMaker deployments

• Added sagemaker_environment_variablesparameter for SageMaker environment configuration

Bug Fixes

• Fix recipe overrides by creating deep copies to avoid shared references especially for overrides that have the same values.

• Fix validation logic for the "temperature" parameter to accept both int and float types.

Release 1.0.83

New Features

Training & Model Support

• Adds Direct preference optimization (DPO) support (LoRA and Full) for SMTJ and SMHP on Nova 1.0

Memory Management

• Refactors dataset loading to use iterators and lazy loading, allowing us to load large datasets with bounded host-memory utilization

Enhancements

Documentation

• Improves documentation clarity by reducing bulk in README.md
• Moves allowed instance types information from README.md to its own document
• Expands SECURITY.md with security best practices and code examples
• Adds CPT/DPO examples to JumpStart notebook
• Updates spec.md with latest SDK changes and additional AWS documentation links

Installation & Setup

• Improves guidance for Forge set-up
• Improves notebook markdown formatting

Logging

• Adds warning for checkpoint resolution failures with base model fallback in evaluate and batch inference

Bug Fixes

• Fixed DatasetValidator for Multi-modal data with Nova models

Release 1.0.72

Hey Nova builders! We have a bunch of features and Quality-of-Life enhancements for this release, and also improved some edge-case behaviors from the initial release.

New Features

MLflow Integration

• Track training experiments with Amazon SageMaker MLflow tracking servers
• Auto-discover DefaultMLFlowApp in your AWS account or specify a custom tracking URI
• Log metrics, hyperparameters, and model artifacts automatically during training

Continued Pre-Training (CPT) Support

• Pre-train Nova models on your own datasets

Data Mixing for SFT and CPT (Nova Forge customers only)

• Blend your custom training data with Nova's high-quality curated datasets

OpenAI Messages Format Conversion

• Transform datasets from OpenAI chat format to the Converse API format for use with Nova models

Multimodal SFT Dataset Validation

• Validate image content in SFT datasets (PNG, JPEG, GIF formats)
• Validate video content in SFT datasets (MOV, MKV, MP4 formats)
• Validate document content for Nova 2.0 (PDF format)

Enhancements

Dataset Validation

• Enhanced data-format validation for SFT, RFT, CPT, and Eval jobs, including message role ordering, content types, and tool specifications / uses.

IAM and Security

• We now have more granular validations of whether the IAM calling role has the required permissions for a job. Where possible, we also validate the SMTJ execution role.
  • IAM validation can be disabled via the validation_config parameter of NovaModelCustomizer
• New create_bedrock_execution_role() utility function generates scoped-down IAM roles for Bedrock model deployment with minimal required permissions
• New VPC configuration parameters (subnets, security_group_ids) in SMTJRuntimeManager allow training jobs to run within your VPC for network isolation
• New kms_key_id parameter encrypts training job output artifacts and inter-container traffic with your KMS key

Recipe Management

• Recipes are now automatically pulled from SageMaker JumpStart, ensuring you always use the latest supported configurations
• When starting evaluation, batch inference, or model deployment, we now automatically extract the model checkpoint from your most recent training job output unless explicitly specified.
  • You can also provide a TrainingResult object to extract the model checkpoint from it.

Documentation

• Added SECURITY.md for vulnerability reporting

Bug Fixes

• Fixed edge-case where evaluation jobs would fail if data_s3_path was set on the customizer but not needed for the evaluation task
• Improved validation error messages with specific field locations
• Improved README setup documentation and fixed some errors in the examples