ES (Eager Style) Documentation【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/geOverviewES (Eager Style) is a functional interface module in GraphEngine for building computation graphs, providing convenient graph building functionality. The ES module supports multiple programming languages (C, C, Python), offering flexible and easy-to-use graph building methods.Core Characteristics of ES Series API:Auto-generation: API is not handwritten but auto-generated based on operator prototype definitions, reducing operator developer burden, especially for custom operators that also need to use ES API for graph building scenarios.Multi-language support: Native support for C, C, and extensible support for Python, meeting different development habits.Full-dimension compatibility: Through good API design and code generation mechanism, combined with IR semantic compatibility handling, achieves forward/backward API and ABI compatibility guarantee from the source.Three Key Components of ES:ES Basic Data Structures(GE Repository/GE Package): Provides core infrastructure likeEsGraphBuilder(graph builder),EsTensorHolder(tensor holder), which are the foundation for graph building.ES Code Generator(GE Repository/GE Package): Core toolgen_esb, responsible for reading operator prototype definitions and auto-generating graph building API code for each operator.Generated ES API(OPP Package): The final artifact built by generator, included in operator package, is the function interface that users directly call.Quick NavigationAPI Reference DocumentationPython API- Eager Style Graph Builder Python Interface DocumentationMain interfaces of GraphBuilder, TensorHolder and other basic classesC/C API- Eager Style Graph Builder C/C Interface DocumentationMain interfaces of EsGraphBuilder, EsTensorHolder and other basic classesDesign DocumentationOverall Architecture Design- ES module overall architecture design explanationOwnership Relationship Analysis- ES module Python layer and C layer resource management mechanism explanationTool Documentationgen_esb Code Generation Tool- ES code generation tool usage instructionsTool functionality explanationUsage method and parameter explanationgenerate_es_package.cmake- ES code generation, build, install one-stop cmake functionFunctionality explanationUsage method and parameter explanationRFC DocumentationRFC Directory- Design proposal documentsNew feature proposalsArchitecture improvement plansAPI change suggestionsDocument Structuredocs/es/ ├── README.md # This file, document navigation entry ├── api/ # API reference documentation ├── design/ # Design and technical analysis documentation ├── tools/ # Tool usage documentation └── rfc/ # RFC proposal documentsSampleMulti-language samples using ES for graph buildingMulti-language samples for custom ES API graph buildingDevelopment RoadmapWe first introduced ES functionality in 2025, aiming to provide convenient multi-language graph building capability, support full operator auto API generation, and achieve good forward/backward compatibility. In Q1 2026, we will focus on completing API integration and C graph building backward compatibility capability, allowing users to directly install ops package and use ES graph building capability, without manually generating code. Specific development roadmap is as follows:Core Architecture[December 2025]ES core architecture design completed and implemented, supporting full prototype code generation and multi-language (C, C, Python) graph building.API Integration[December 2025]math sub-package completed ES API integration.[February 2026]nn, cv, transformer completed ES API integration.[June 2026]hcom sub-package to complete ES API integration.Note: Before full prototype ES API is packaged into various operator sub-packages, if you need to use ES API, you can refer to generate_es_package.cmake in tool documentation for code generation and integration.Samples and Documentation[December 2025]Multi-language samples provided, covering common use cases.[December 2025]Detailed documentation provided, i.e., this directory.[January 2026]Collective communication samples delivered.Backward Compatibility[December 2025]Python API backward compatibility design completed and implemented.[January 2026]C API backward compatibility design completed; C API backward compatibility depends on completingHistorical Prototype Librarydesign plan.[March 2026]C API backward compatibility code development completed; achieve complete backward compatibility capability for C graph building scenarios.Custom Operator Project Integration[TBD]ascendc custom operator project integration of ES code generation capability, no plan yet.【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
