Driven by the rapid expansion of the digital economy and the implementation of the national “Eastern Data and Western Computing” strategy, the societal demand for computing power has become increasingly urgent. This presentation introduces full-stack performance acceleration solutions based on the Heterogeneous computing platform, spanning from operator library design to template library development, and from shared memory libraries to the design of communication-computation fused operators. It further explores compilation optimizations from MLIR to AscendNPU IR, as well as Triton-based linear representation and Cube-Vector (CV) fusion schemes. By sharing application cases focused on operator performance optimization and compilation-driven ecosystem enablement, this report provides effective parallel acceleration solutions for the complex, unified training-inference environments of LLM.

Biography

Lu Lu, has completed his PhD at the age of 28 years from Xi’an jiaotong University. He is professor and doctoral supervisor at the School of Computer Science and Engineering, South China University of Technology, and a dual-appointed professor at Shenzhen Pengcheng National Laboratory, mainly engages in scientific research in the fields of software system and architecture design, software reliability assurance, high-performance computing, and heterogeneous parallel acceleration. He has published more than 80 papers and applied for authorized 20 patents and software copyrights.