CANN/GE自定义融合Pass开发指南

Custom Fusion Pass【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/geThis directory provides development documentation and samples for GE custom fusion passes.If this is your first development, its recommended to read in the following order:Fusion Pattern Pass Mechanism: First understand pattern, matching, filtering, replacement and boundary rules.Python Fusion Pass Development Guide: Supports runtime integration andpatternexpression syntax.C Fusion Pass Development Guide: Suitable for product delivery after compiling into.so.Sample DirectoryDirectoryDescriptionpattern_base_passRecommended to reference first. Develop pattern-based fusion rules throughPatternFusionPassorDecomposePassgraph_base_passSamples that directly modify graph through graph interfaces, suitable for scenarios requiring complete manual control of graph modification【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考