8 papers
ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels
Bingyang Wu, Chao Jin, Zili Zhang +6
LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-deco…
UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing
Xinming Wei, Chao Jin, Tuo Dai +10
Large-scale expert parallelism (EP) is becoming pivotal for training and serving frontier MoE models, but it also amplifies device-level expert load imbalance into compute straggle…
ReLibra: Routing-Replay-Guided Load Balancing for MoE Training in Reinforcement Learning
Chao Jin, Xinming Wei, Yinmin Zhong +6
Load imbalance is a long-standing challenge in Mixture-of-Experts (MoE) training and is exacerbated in reinforcement learning (RL) for LLMs, where hot experts can shift frequently…
Heddle: A Distributed Orchestration System for Agentic RL Rollout
Zili Zhang, Yinmin Zhong, Chengxu Yang +5
Agentic Reinforcement Learning (RL) enables LLMs to solve complex tasks by alternating between a data-collection rollout phase and a policy training phase. During rollout, the agen…
TokenLake: A Unified Segment-level Prefix Cache Pool for Fine-grained Elastic Long-Context LLM Serving
Bingyang Wu, Zili Zhang, Yinmin Zhong +4
Prefix caching is crucial to accelerate multi-turn interactions and requests with shared prefixes. At the cluster level, existing prefix caching systems are tightly coupled with re…
StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation
Yinmin Zhong, Zili Zhang, Xiaoniu Song +11
Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs). RL for LLMs involves two stages: generation and training. The LLM first ge…