11 papers
GEB-Bench: Abstract Structures Told in Many Voices
Tong Zhang, Zhiyuan Shi, Yun Peng +1
Can a model look at a river delta and a lightning bolt and see that they share a structure? We introduce GEB-Bench, a benchmark whose unit is an abstract structural motif--self-ref…
StrataCL: Fabric-Native Communication Library for Production Supernodes
Tiancheng Hu, Jin Qin, Yuzheng Wang +14
StrataCL is a communication library for large‑scale AI workloads on supernode clusters that removes redundant data copies by enabling direct user‑buffer communication and uses NPU‑…
When Does Recurrence Become an Algorithm? Convergence Selection in Weight-Tied Looped Transformers
Tong Zhang, Junhao Hu, Yun Peng +1
When does a weight-tied looped transformer -- one block applied T times -- implement an actual algorithm? We answer with four findings from controlled populations on group word pro…
Tessera: Unlocking Heterogeneous GPUs through Kernel-Granularity Disaggregation
Tiancheng Hu, Jin Qin, Zheng Wang +10
Disaggregation maps parts of an AI workload to different types of GPUs, offering a path to utilize modern heterogeneous GPU clusters. However, existing solutions operate at a coars…
Lil: Less is Less When Applying Post-Training Sparse-Attention Algorithms in Long-Decode Stage
Junhao Hu, Fangze Li, Mingtao Xu +9
Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference e…
Hummingbird: SLO-Oriented GPU Preemption at Microsecond-scale
Tiancheng Hu, Chenxi Wang, Ting Cao +9
Existing GPU-sharing techniques, including spatial and temporal sharing, aim to improve utilization but face challenges in simultaneously ensuring SLO adherence and maximizing effi…