CANN/ge跨功能分析检查模板

Feature Cross Analysis【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/gePurpose: Ensure new requirements are covered under GEs core execution scenarios, avoiding omissions that lead to missing functionality or abnormal behavior in some scenarios. Analyze the relationship between this requirement and the following scenarios item by item, filling in Applicable/Not Applicable. Applicable scenarios need to provide corresponding solutions in the design section; not applicable scenarios need to explain reasons.ScenarioApplicabilityAnalysis DescriptionStatic Shape{Applicable/Not Applicable}{Describe the impact and design points of the requirement in this scenario}Dynamic Shape{Applicable/Not Applicable}{Describe the impact and design points of the requirement in this scenario}Dynamic Shape Static Subgraph (Graph Split){Applicable/Not Applicable}{Describe the impact of the requirement in mixed graph split scenarios, whether it affects split boundaries, subgraph behavior, etc.}Offline Scenario (atc Compilation){Applicable/Not Applicable}{Describe whether the requirement affects offline compilation flow, OM artifact structure, or serialization/deserialization}Online Scenario (Framework Adaptation){Applicable/Not Applicable}{Describe whether the requirement affects online compilation execution flow, interaction with Adapters, etc.}Analysis GuidanceStatic Shape: Does the requirement affect known shape compilation optimization (memory reuse, operator online compilation, sink scheduling)? Does it involveruntime/v1/path? Pay special attention toDavinciModelloading/initialization/execution flow.Dynamic Shape: Does the requirement involve operators or execution paths with uncertain shape? Does it involve runtime tiling, dynamic memory allocation? Does it involveruntime/v2/path? Pay special attention toLoweringGlobalData,ModelConverter,LoweringFileConstantNodeand other lowering phase components.Dynamic Shape Static Subgraph (Critical Cross Scenario): This is the most easily overlooked scenario. Static subgraphs in dynamic shape models go through v1 path viaDavinciModelKernel. Analysis points:Data crossing v2→v1 boundary: If the requirement adds data in v2sLoweringGlobalData, does this data need to be passed to v1sDavinciModel? Specifically, doesCreateDavinciModelOnInitRootinstatic_compiled_graph_converter.ccneed to pass the new data as input toDavinciModelCreate/DavinciModelCreateV2kernel?DavinciModel interface changes: If the requirement modifiesDavinciModelinterfaces (such as adding Set methods), doDavinciModelCreate/DavinciModelCreateV2inDavinciModelKernelneed to call the new interfaces? Doesdavinci_model_kernel.ccneed new KernelContext input indices?Check method: grepDavinciModelCreateandDavinciModelCreateV2to confirm all paths creating DavinciModel can access the new data.Offline Scenario: Does the requirement affect atc compilation flow, OM file format, model serialization/deserialization? Do offline compilation artifacts need new fields?Online Scenario: Does the requirement affect GEs interaction with frontend framework adaptation layers (TorchAir/TF Adapter)? Does it affectsession-level compilation and execution flow?【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