Triton-distributed: Programming Overlapping Kernels on Distributed AI Systems with the Triton Compiler

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Zheng, Size, Bao, Wenlei, Hou, Qi, Zheng, Xuegui, Fang, Jin, Huang, Chenhui, Li, Tianqi, Duanmu, Haojie, Chen, Renze, Xu, Ruifan, Guo, Yifan, Zheng, Ningxin, Jiang, Ziheng, Di, Xinyi, Wang, Dongyang, Ye, Jianxi, Lin, Haibin, Chang, Li-Wen, Lu, Liqiang, Liang, Yun, Zhai, Jidong, Liu, Xin
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916780182601728
author Zheng, Size
Bao, Wenlei
Hou, Qi
Zheng, Xuegui
Fang, Jin
Huang, Chenhui
Li, Tianqi
Duanmu, Haojie
Chen, Renze
Xu, Ruifan
Guo, Yifan
Zheng, Ningxin
Jiang, Ziheng
Di, Xinyi
Wang, Dongyang
Ye, Jianxi
Lin, Haibin
Chang, Li-Wen
Lu, Liqiang
Liang, Yun
Zhai, Jidong
Liu, Xin
author_facet Zheng, Size
Bao, Wenlei
Hou, Qi
Zheng, Xuegui
Fang, Jin
Huang, Chenhui
Li, Tianqi
Duanmu, Haojie
Chen, Renze
Xu, Ruifan
Guo, Yifan
Zheng, Ningxin
Jiang, Ziheng
Di, Xinyi
Wang, Dongyang
Ye, Jianxi
Lin, Haibin
Chang, Li-Wen
Lu, Liqiang
Liang, Yun
Zhai, Jidong
Liu, Xin
contents In this report, we propose Triton-distributed, an extension of existing Triton compiler, to overcome the programming challenges in distributed AI systems. Triton-distributed is the first compiler that supports native overlapping optimizations for distributed AI workloads, providing a good coverage of existing optimizations from different frameworks. First, we integrate communication primitives compliant with the OpenSHMEM standard into the compiler. This enables programmers to utilize these primitives with a higher-level Python programming model. Second, we illustrate how to achieve complex joint optimization of computation, memory access, and communication with the assistance of the compiler. In particular, we show how to use overlapping techniques to hide latency and present our compiler-based programming methods in both single-node and multi-node scenarios. Finally, we showcase the performance of the code generated by our compiler. In a test environment with up to 64 devices, our compiler can fully utilize heterogeneous communication and computation resources to provide effective overlapping and high performance. In many cases, the performance of the generated code can even outperform hand-optimized code. Moreover, the development difficulty and the time cost for development using our compiler are far less than those of low-level programming such as CUDA/C++, which clearly demonstrates significant productivity advantages.
format Preprint
id arxiv_https___arxiv_org_abs_2504_19442
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Triton-distributed: Programming Overlapping Kernels on Distributed AI Systems with the Triton Compiler
Zheng, Size
Bao, Wenlei
Hou, Qi
Zheng, Xuegui
Fang, Jin
Huang, Chenhui
Li, Tianqi
Duanmu, Haojie
Chen, Renze
Xu, Ruifan
Guo, Yifan
Zheng, Ningxin
Jiang, Ziheng
Di, Xinyi
Wang, Dongyang
Ye, Jianxi
Lin, Haibin
Chang, Li-Wen
Lu, Liqiang
Liang, Yun
Zhai, Jidong
Liu, Xin
Distributed, Parallel, and Cluster Computing
In this report, we propose Triton-distributed, an extension of existing Triton compiler, to overcome the programming challenges in distributed AI systems. Triton-distributed is the first compiler that supports native overlapping optimizations for distributed AI workloads, providing a good coverage of existing optimizations from different frameworks. First, we integrate communication primitives compliant with the OpenSHMEM standard into the compiler. This enables programmers to utilize these primitives with a higher-level Python programming model. Second, we illustrate how to achieve complex joint optimization of computation, memory access, and communication with the assistance of the compiler. In particular, we show how to use overlapping techniques to hide latency and present our compiler-based programming methods in both single-node and multi-node scenarios. Finally, we showcase the performance of the code generated by our compiler. In a test environment with up to 64 devices, our compiler can fully utilize heterogeneous communication and computation resources to provide effective overlapping and high performance. In many cases, the performance of the generated code can even outperform hand-optimized code. Moreover, the development difficulty and the time cost for development using our compiler are far less than those of low-level programming such as CUDA/C++, which clearly demonstrates significant productivity advantages.
title Triton-distributed: Programming Overlapping Kernels on Distributed AI Systems with the Triton Compiler
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2504.19442