作者位次 #5/21OSDI 2026Oral· Operating Systems & Distributed Systems
MPK: A Compiler and Runtime for Mega-Kernelizing Tensor Programs
Xinhao Cheng · Zhihao Zhang · Yu Zhou · Jianan Ji · Jinchen Jiang · Zepeng Zhao · Ziruo Xiao · Zihao Ye · Yingyi Huang · Ruihang Lai · Hongyi Jin · Bohan Hou · Mengdi Wu · Yixin Dong · Anthony Yip · Zihao Ye · Songting Wang · Wenqin Yang · Xupeng Miao · Tianqi Chen · Zhihao Jia
摘要
We introduce Mirage Persistent Kernel (MPK), the first compiler and runtime system that automatically transforms multi-GPU model inference into a single high-performance mega-kernel. MPK introduces an SM-level graph representation that captures data dependencies at the granularity of individual s...