4 papers
HBM Is Not All You Need: Efficient Disaggregated LLM Serving across Memory-heterogeneous Accelerators
Zhixiang Wei, Yun Wang, James Yen +2
LLM inference comprises a compute-bound prefill phase and a memory-bound decode phase, and recent systems disaggregate them onto separate hardware. Yet today's datacenter GPUs rely…
The Phantom of PCIe: Constraining Generative Artificial Intelligences for Practical Peripherals Trace Synthesizing
Zhibai Huang, Chen Chen, James Yen +9
Peripheral Component Interconnect Express (PCIe) is the de facto interconnect standard for high-speed peripherals and CPUs. The development of PCIe devices for emerging application…
BuddyMoE: Exploiting Expert Redundancy to Accelerate Memory-Constrained Mixture-of-Experts Inference
Yun Wang, Lingyun Yang, Senhao Yu +5
Mixture-of-Experts (MoE) architectures scale language models by activating only a subset of specialized expert networks for each input token, thereby reducing the number of floatin…
Equinox: Holistic Fair Scheduling in Serving Large Language Models
Zhixiang Wei, James Yen, Jingyi Chen +11
We address the limitations of current LLM serving with a dual-counter framework separating user and operator perspectives. The User Fairness Counter measures quality of service via…