TileLang : A Composable Tiled Programming Model for AI Systems

Lei Wang §, Yu Cheng §, Yining Shi §, Zhengju Tang, Zhiwen Mo, Wenhao Xie, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang · 2025

Read TileLang : A Composable Tiled Programming Model for AI Systems by Lei Wang §, Yu Cheng §, Yining Shi §, Zhengju Tang, Zhiwen Mo, Wenhao Xie, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-17577.txt.

Description

Lei Wang and co-authors present TileLang, a programming model that makes writing fast AI kernels less painful. Their 2025 paper notes that modern AI runs on small optimized routines that shuttle tiles of data between memory levels. Getting peak speed from those routines demands fiddly, hardware-specific tuning that most programmers would rather avoid.

TileLang separates what a kernel computes from how it is scheduled. Users describe the data flow, then adjust thread binding, layout, and pipelining through compact annotations. The compiler handles the rest of the hardware detail. Across common accelerators and standard kernels, the system reached state-of-the-art speed in tests. The authors argue the unified block-and-thread view gives both power and flexibility. It targets the people who build the engines under today's AI tools.

Catalog record

CREATORLei Wang §, Yu Cheng §, Yining Shi §, Zhengju Tang, Zhiwen Mo, Wenhao Xie, Lingxiao Ma, Yuqing Xia, Jilong Xue, Fan Yang, Zhi Yang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-17577

Browse the full library · About the Tower of Babel