朝着无缝模型编译集成迈进 | KC23

Toward Seamless Model Compilation Integration - Hongyu Zhu, ByteDance

朝着无缝模型编译集成迈进 | Toward Seamless Model Compilation Integration - Hongyu Zhu, ByteDance

模型编译在AI加速中变得越来越重要。然而，在IT公司中采用模型编译用于生产模型并不是一件简单的事情。主要的负担包括来自不同领域、框架或格式的模型，从现有库过渡以及采用新的ASIC。为了解决这些问题，ByteIR被开发出来以提高模型编译的生产力。ByteIR建立在OpenXLA和LLVM/MLIR编译器基础设施之上。它包括前端、编译器和运行时组件，每个组件解决不同的问题。例如，前端专注于模型覆盖和无缝框架集成；编译器专注于性能、新ASIC的采用和向后兼容库的过渡；运行时提供统一且无缝的生产集成。这三个组件可以一起工作，也可以选择独立工作以满足不同的业务需求。

Model compilation is becoming more and more important in AI acceleration. However, it is not trivial to adopt model compilation for production models in IT companies. Burdens mainly include models coming from domains, frameworks, or formats, transition from existing libraries, and adoption of new ASICs. To address these issues, ByteIR is developed to improve model compilation productivity. ByteIR is built on top of OpenXLA and LLVM/MLIR compiler infrastructure. It includes frontends, compiler, and runtime components, each of which resolve different issues. For example, the frontends focus on model coverage and seamless framework integration; compiler focuses on performance, new ASIC adoption, and backward library transition; runtime provides unified and seamless integration for production. These three components can either work together, or selectively work independently to meet different business needs.

CNCF概况（幻灯片）

扫描二维码联系我们！

CNCF (Cloud Native Computing Foundation)成立于2015年12月，隶属于Linux  Foundation，是非营利性组织。 

CNCF（云原生计算基金会）致力于培育和维护一个厂商中立的开源生态系统，来推广云原生技术。我们通过将最前沿的模式民主化，让这些创新为大众所用。请关注CNCF微信公众号。