融合规则列表【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如下融合规则关闭后可能会有功能使用上的风险更多融合规则请参见《图融合和UB融合规则参考》。Ascend 950PR/Ascend 950DT不支持UB融合。图融合Ascend 950PR/Ascend 950DT支持的图融合规则AABiasaddConvFusionAddRmsNormFusionGraphPassADepthwiseFusionPassConv2DQuantProcessFusionPassDepthwiseDfFusionPassDepthwiseDwMulFusionPassEinsumPassFixPipeAbilityProcessPassInplaceAddRmsNormFusionPassSoftmaxGradExtFusionTfTagNoConstFoldingFusionPassWeightQuantBatchMatmulV2TransposeFusionPassZConfusionSoftmaxGradFusionPassConstToAttrStridedSliceV2FusionDepthwiseToConv2dFusionPassDWConv2DQuantProcessFusionPassGlobalavgpoolpass图融合非Ascend 950PR/Ascend 950DT支持的图融合规则AABiasaddConvFusionAddNFusionPassAddRmsNormFusionGraphPassADeformableConv2dPassADepthwiseFusionPassALSTMFusionPassApplyAddOutputPassAReduceMeanFusionPassAReduceSumFusionPassAvgPool3DFusionPassAvgPool3DGradFusionPassAvgPoolGradFusionPassAvgPoolQuantProcessFusionPassBatchMatmulV2QuantProcessFusionPassBatchNormBnInferFusionPassBatchNormGradBnInferGradFusionBatchNormGradInfGradFusionBatchMatMulFusionPassCastRemoveFusionPassConstToAttrPassConstToAttrReduceSumFusionConv2DQuantProcessFusionPassCommonLSTMFusionPassCommonSubexpressionEliminationPassGE公共表达式消除PassConstToAttrGatherV2FusionConstToAttrResizeNearestNeighborGradFusionConstToAttrStridedSliceV2FusionConv2DTDQuantProcessFusionPassCOPYPassDeConvQuantProcessFusionPassDepthwiseDfFusionPassDepthwiseDwMulFusionPassDepthwiseFusionPassDepthwiseToConv2dFusionPassDreluFusionPassDWConv2DQuantProcessFusionPassDynamicGRUV2GradFusionPassDynamicRNNFusionPassDynamicRNNGradAFusionPassDynamicRNNGradAlignFusionPassDynamicRNNGradDAlignFusionPassDynamicRNNGradDFusionPassDynamicRNNGradFusionPassDynamicRNNInsertTransposePassDynamicRNNSeqFusionPassEinsumPassFCQuantProcessFusionPassFixPipeAbilityProcessPassFlattenV2PassFusedBatchNormBertFusionPassFusedBatchnormFusionPassFusedBatchNormGradFusionPassGlobalavgpoolpassGroupConv2DQuantProcessFusionPassHostBNFusionPassInplaceAddRmsNormFusionPassMapIndexFusionPassMatMul2MatMulV2FusionPassMatmulV2QuantProcessFusionPassMaxPoolWithArgmaxFusionPassPackFusionPassPassThroughFusionPassPassThroughSecondFusionPassPermuteFusionPassPoolingQuantProcessFusionPassPReluGradFusionPassPriorBoxPassProposalFusionPassRNNFusionPassSingleBatchNormFusionSpatialTransformerDPassSPPPassSoftmaxGradExtFusionTbeAntiquantMaxpoolingFusionPassTbeConvDequantS16FusionPassTfMergeConv2DBackpropInputFusionPassTfTagNoConstFoldingFusionPassTransposedUpdateFusionPassTransdataCastFusionPassUnpackFusionPassV100NotRequantFusionPassV200NotRequantFusionPassWeightQuantBatchMatmulV2TransposeFusionPassYoloPassYoloV2DetectionOutputPassYoloV3DetectionOutputV2PassZConcatExt2FusionPassZConcatDFusionPassZConfusionSoftmaxGradFusionPassZSplitVDFusionPassZSplitVFusionPassUB融合ArgMaxWithFusionPassclip_by_norm_nodivsquaresumDeformableOffsetsFusionPassDepthwiseInsertTransDataFusionPassHostShapeOptimizationPassAI CPU断流水passNormalizeFusionPassTbeConvRequantFusionPassTbeConvCommonRules0FusionPassTbeConvCommonRules2FusionPassTfMergeSubFusionPass【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