v2.0.0
Release Notes - v2.0.0
Announcing major updates to Amazon SageMaker HyperPod recipes (v2.0.0)
Key Features
Expanded Model Support
- Llama: 3, 3.1, 3.2, 3.3 (1B to 70B)
- DeepSeek: R1 Distilled Llama (8B, 70B), R1 Distilled Qwen (1.5B - 32B)
- GPT-OSS: 20B, 120B
- Qwen: 2.5 (0.5B - 72B), 3 (0.6B - 32B)
Expanded training techniques
| Technique | Available Methods |
|---|---|
| Supervised Fine-Tuning (SFT) | LoRA, Full Fine-Tuning |
| Direct Preference Optimization (DPO) | LoRA, Full Fine-Tuning |
| Reinforcement Learning from AI Feedback (RLAIF) | LoRA, Full Fine-Tuning |
| Reinforcement Learning with Verifiable Rewards (RLVR) | LoRA, Full Fine-Tuning |
Support for new training frameworks and techniques
- LLMFT: Advanced fine-tuning support for SFT and DPO with LoRA
- VERL: Reinforcement learning support using GRPO algorithm for RLVR, RLAIF
- Checkpointless training: Memory-efficient training for large models
- Elastic training: Dynamic resource scaling capabilities
Infrastructure support
- NVIDIA H100, A100, and A10G accelerators
- Built-in logging support (TensorBoard, MLflow)
- Choice of training infrastructure across SageMaker training jobs, SageMaker HyperPod with Amazon EKS, and Slurm
Documentation
- Refer to the
README.mdfor detailed usage instructions and examples - Refer to
recipes_collectionfor updated recipe collection - Refer to
launcher_scriptsfor launcher script examples
Contributing
We welcome contributions to enhance the capabilities of sagemaker-hyperpod-recipes. Please refer to our contributing guidelines for more information.
Thank you for choosing sagemaker-hyperpod-recipes for your model training!