Flexible RL Dataflows
Enables building diverse reinforcement learning algorithms by constructing dataflows in a few lines of code using a hybrid programming model.
Seamless LLM Integration
Provides modular APIs for integration with existing LLM frameworks such as PyTorch FSDP, Megatron-LM, vLLM, SGLang, and HuggingFace models.
Scalable GPU Parallelism
Supports flexible device mapping and parallelism across different GPU sets and cluster sizes to optimize resource utilization.
High Throughput Training and Inference
Leverages state-of-the-art LLM training and inference tools to achieve efficient generation and training throughput.
Memory and Communication Optimization
Uses 3D-HybridEngine for actor model resharding to reduce memory redundancy and communication overhead during training-generation transitions.