Zili Zhang

Latest

• ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels
• UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing
• ReLibra: Routing-Replay-Guided Load Balancing for MoE Training in Reinforcement Learning
• FastServe: Iteration-Level Preemptive Scheduling for Large Language Model Inference
• Heddle: A Distributed Orchestration System for Agentic RL Rollout
• Epiphron: Resource-Efficient Distributed Key-Value Storage
• TokenLake: A Unified Segment-level Prefix Cache Pool for Fine-grained Elastic Long-Context LLM Serving
• Optimizing RLHF Training for Large Language Models with Stage Fusion
• StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation
• dLoRA: Dynamically Orchestrating Requests and Adapters for LoRA LLM Serving
• Transparent GPU Sharing in Container Clouds for Deep Learning Workloads