Documentation

Dump Graph Details

Before model conversion, you can set the following environment variables to dump the model graphs generated during build. You can view the dump graphs to observe the model changes.

export DUMP_GE_GRAPH=1 # Determines the size of a dump graph.
export DUMP_GRAPH_LEVEL=1 # Determines the number of dump graphs.

The following files are generated in the current path where the atc command is executed. For details about the environment variables, see 2.

  • ge_onnx*.pbtxt: model description structure based on ONNX. You can open this file using visualizer software such as Netron.
  • ge_proto*.txt: text file stored in Protobuf format. You can convert it into a JSON file to facilitate fault locating. This file appears in pair with the ge_onnx*.pbtxt file, but has more attributes of the string type than the ge_onnx*.pbtxt file, making the ge_proto*.txt file more comprehensive. You can open either of them.

    Compared with the ge_onnx*.pbtxt file, the ge_proto*.txt file has a smaller size. Therefore, setting DUMP_GE_GRAPH to 2 or 3 has the same effect on the ge_proto*.txt file, that is, dumping information without data such as weights.

  • ge_readable*.txt: (optional) highly readable text file in Dynamo FX graph style. This type of file is generated only when the DUMP_GRAPH_FORMAT environment variable is set and the value contains readable.

Each of the preceding files corresponds to a step in the model build process, for example, the build begins with the execution of the ge_onnx_xxxx_PreRunBegin.pbtxt file and ends with the execution of the ge_onnx_xxxx_PreRunAfterBuild.pbtxt file. Each file contains all operators involved in the corresponding step. For details about the subgraph functions in each phase of the dump graph, see Table 1. (The dump subgraphs may vary with models, but the workflow is basically the same.)

Table 1 Description of dump graph details

Subgraph

Phase

Trustlisted or Not

ge_proto_xxxx_GraphPreRunBegin.txt

Graph before build.

-

ge_proto_xxxx_AfterFlowGraphPartition.txt

Graph after flow partitioning (flow partitioning is a partitioning mode used in the data flow).

-

ge_proto_xxxx_AfterParallelPartitioner.txt

Graph after pipeline parallel partitioning (pipeline here refers to PP in the backend inference scenario).

-

ge_proto_xxxx_PreRunBegin.txt

Original graph structure.

Trustlisted

ge_proto_xxxx_RunCustomPassBeforeInfershape.txt

Custom pass exit graph before InferShape.

Trustlisted

ge_proto_xxxx_PreRunAfterInitPreparation.txt

Graph structure obtained after all initialization in the graph preparation phase.

-

ge_proto_xxxx_PreRunAfterHandleSummaryOp.txt

Graph structure obtained after Summary node processing.

-

ge_proto_xxxx_PrepareAfterCheckAndUpdateInput.txt

Graph structure obtained after the graph input is checked and updated.

-

ge_proto_xxxx_PrepareAfterGraphEquivalentTransformation.txt

Graph structure obtained after the for loop is equivalently transformed into while loop.

-

ge_proto_xxxx_PrepareAfterProcessOutput.txt

Graph structure obtained after graph data processing.

-

ge_proto_xxxx_PrepareAfterOptimizeAfterGraphNormalization.txt

Optimization exit graph after graph normalization.

-

ge_proto_xxxx_PrepareAfterInsertAipp.txt

Graph structure obtained after AIPP processing.

-

ge_proto_xxxx_PrepareAfterProcessAippNodesDataFormat.txt

Exit graph after AIPP node format update.

-

ge_proto_xxxx_PreRunAfterNormalizeGraph.txt

Exit graph after graph normalization.

Trustlisted

ge_proto_xxxx_PreRunAfterOptimizeGraphInit.txt

Exit graph after graph optimization and initialization.

-

ge_proto_xxxx_OptimizeGraph_TagNoConstFoldingAfter.txt

Used for quantization scenarios. The FE adds a tag for an operator to indicate no constant folding. When the GE checks that the tag exists, it does not perform constant folding.

-

ge_proto_xxxx_HcclAfterOptimizeGraphPrepare.txt

Graph obtained after optimization in the HCCL graph preparation phase.

-

ge_proto_xxxx_PreRunAfterOptimizeGraphPrepare.txt

Graph structure obtained after the original graphs in the operator information library are prepared (by calling the OptimizeGraphPrepare API).

-

ge_proto_xxxx_PrepareAfterProcessBeforeInfershape.txt

Graph structure obtained after "dead edges"of the conditional operator are eliminated.

-

ge_proto_xxxx_AfterFirstInferformat.txt

Graph structure obtained after format inference on the entire graph.

-

ge_proto_xxxx_AfterInfershape.txt

Graph structure obtained after shape inference on the entire graph, with constant folding.

Trustlisted

ge_proto_xxxx_PrepareAfterInferFormatAndShape.txt

Graph structure obtained after InferFormat and InferShape processing. This graph has undergone the second InferFormat processing, compared with the preceding graph.

-

ge_proto_xxxx_RunCustomPass_AfterInferShape.txt

Custom pass exit graph after InferShape.

Trustlisted

ge_proto_xxxx_AfterSecondInferformat.txt

Graph obtained after the second format inference.

-

ge_proto_xxxx_PrepareAfterCtrlFlowPreProcess.txt

Graph structure obtained after the conditional operator is preprocessed.

-

ge_proto_xxxx_PrepareAfterGetDynamicOutputShape.txt

Graph structure obtained after graph output processing in the presence of dynamic batch size profiles.

-

ge_proto_xxxx_PrepareAfterProcessAippStage2.txt

Graph structure obtained after graph input processing in AIPP mode.

-

ge_proto_xxxx_PrepareAfterPrepareOptimize.txt

Graph structure obtained after optimization in the graph preparation phase.

-

ge_proto_xxxx_PreRunAfterPrepare.txt

Graph structure obtained after graph preparation, the same as the preceding graph.

Trustlisted

ge_proto_xxxx_OptimizeQuantGraph_FeGraphFusionAfter.txt

Graph structure obtained after quantization in the graph optimization phase ends.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeGraphFusionAfter.txt

Graph structure obtained after graph fusion ends.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeTopoSortingAfter.txt

Graph structure obtained after graph fusion and topology sorting, for checking whether a ring is formed.

-

ge_proto_xxxx_OptimizeOriginalGraph_DSAFeOpJudgeAfter.txt

Graph with DataType and Format selected after dynamic shape analysis.

-

ge_proto_xxxx_HcclBeforeOptimizeOriginalGraph.txt

Graph before optimization of the original HCCL graph.

-

ge_proto_xxxx_HcclAfaterOptimizeOriginalGraph.txt

Graph after optimization of the original HCCL graph.

-

ge_proto_xxxx_RunCustomPassAfterBuiltinFusionPass.txt

Custom pass exit graph after internal pass.

Trustlisted

ge_proto_xxxx_PreRunAfterOptimizeOriginalGraph.txt

Graph structure obtained after the original graphs in the operator information library are optimized (by calling the OptimizeOriginalGraph API).

Trustlisted

ge_proto_xxxx_PrepareAfterUpdateInputOutputByUserOptions.txt

Graph structure obtained after the graph input and output are processed based on the user's command-line options.

-

ge_proto_xxxx_PrepareAfterUpdateVariableFormats.txt

Graph structure obtained after the variable formats are processed.

-

ge_proto_xxxx_PreRunAfterPrepareRunningFormatRefiner.txt

Same as the preceding graph.

-

ge_proto_xxxx_BeforeOptimizeOriginalGraphJudgeInsert.txt

Entry graph of the operator judging process.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeOpDtypeJudgeAfter.txt

Graph after precision mode selection.

-

ge_proto_xxxx_PreRunAfterRefineRunningPrecision.txt

Graph after precision selection.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeOpFormatJudgeAfter.txt

Graph with complete op_judge after format selection.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeDistHeavyFormatAfter.txt

Graph structure obtained after diffusion of heavy operators.

-

ge_proto_xxxx_OptimizeOriginalGraph_FeInsertTransNodeAfter.txt

Graph structure obtained after the transform operator is inserted.

-

ge_proto_xxxx_PreRunAfterRefineRunningFormat.txt

Graph structure obtained after each operator information library is optimized (by calling the OptimizeOriginalGraphJudgeInsert API).

-

ge_proto_xxxx_PreRunAfterSubexpressionMigration.txt

Graph structure after the common subexpression is extracted in the scenario of dynamic dimension size profiles.

-

ge_proto_xxxx_before_SameTransdataBreadthFusionPass.txt

SameTransdataBreadthFusionPass entry graph.

-

ge_proto_xxxx_after_SameTransdataBreadthFusionPass.txt

SameTransdataBreadthFusionPass exit graph.

-

ge_proto_xxxx_OptimizeStage1_1.txt

Graph structure obtained after graph optimization stage 1_1.

-

ge_proto_xxxx_OptimizeStage1_2.txt

Graph structure obtained after graph optimization stage 1_2.

-

ge_proto_xxxx_PreRunAfterOptimize1.txt

Graph structure obtained after optimization stage 1 for all graphs.

-

ge_proto_xxxx_PreRunAfterOptimizeAfterStage1.txt

Graph structure obtained after each operator information library is optimized (by calling the OptimizeAfterStage1 API).

Trustlisted

ge_proto_xxxx_PreRunAfterInferShape2.txt

Graph structure obtained after the second InferShape processing.

-

ge_proto_xxxx_BeforeStagePartition.txt

Graph before stage partitioning.

-

ge_proto_xxxx_AfterStagePartition.txt

Graph after stage partitioning.

-

ge_proto_xxxx_AfterEnginePlacer.txt

Graph after engine selection.

-

ge_proto_xxxx_Before_DSP.txt

Graph before partitioning static and dynamic models.

-

ge_proto_xxxx_After_DSP.txt

Graph after partitioning static and dynamic models.

-

ge_proto_xxxx_AfterDynamicShapePartition.txt

Graph structure obtained after graph partitioning with a dynamic shape.

-

ge_proto_xxxx_MergedComputeGraphAfterCompositeEnginePartition.txt

Structure of the merged graph obtained after opposite subgraph partitioning and subgraph optimization.

-

ge_proto_xxxx_partition0_rank0_inputNodeGraph_AtomicEnginePartitioning.txt

Structure of the input node subgraph obtained after graph partitioning based on the atomic engine rules.

-

ge_proto_xxxx_partition0_rank1_new_sub_graph1_AtomicEnginePartitioning.txt

Structure of subgraph 1 obtained after graph partitioning based on the atomic engine rules.

-

ge_proto_xxxx_partition0_rank2_new_sub_graph110_AtomicEnginePartitioning.txt

Structure of subgraph 110 obtained after graph partitioning based on the atomic engine rules.

-

ge_proto_xxxx_OptimizeSubgraphPreProc.txt

Exit graph after subgraph optimization preprocessing.

-

ge_proto_xxxx_DNN_VM_RTS_OptimizeSubGraphBefore.txt

-

-

ge_proto_xxxx_DNN_VM_RTS_OptimizeSubGraphAfter.txt

-

-

ge_proto_xxxx_AIcoreEngine_OptimizeSubGraphBefore.txt

Entry graph before AI Core subgraph optimization.

-

ge_proto_xxxx_OptimizeSubGraphBefore.txt

Subgraph structure obtained before optimization. Subgraphs have the same name but different sequence numbers, depending on the number of subgraphs.

-

ge_proto_xxxx_OptimizeSubGraphAfter.txt

Subgraph structure obtained after optimization. Subgraphs have the same name but different sequence numbers, depending on the number of subgraphs.

-

ge_proto_xxxx_partition0_rank1_new_sub_graph1_lxfusion_input.txt

SGAT input graph in the ATC and AOE baseline scenarios.

-

ge_proto_xxxx_partition0_rank1_new_sub_graph1_after_rebuild.txt

UB fusion graph of the AOE SGAT internal process.

-

ge_proto_xxxx_AIcoreEngine_OptimizeSubGraphAfter.txt

Exit graph after AI Core subgraph optimization.

-

ge_proto_xxxx_OptimizeSubgraphPostProc.txt

Exit graph after subgraph optimization postprocessing.

-

ge_proto_xxxx_mergedComputeGraph.txt

Structure of the merged graph, the same as the preceding graph.

-

ge_proto_xxxx_MergedComputeGraphAfterAtomicEnginePartition.txt

Structure of the merged graph obtained after opposite subgraph partitioning and subgraph optimization.

-

ge_proto_xxxx_PreRunAfterOptimizeSubgraph.txt

Subgraph structure obtained after optimization.

Trustlisted

ge_proto_xxxx_OptimizeWholeGraphaicpu_tf_optimizer.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphaicpu_ascend_optimizer.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphdvpp_graph_optimizer.txt

Exit graph after entire graph optimization and DVPP optimization.

-

ge_proto_xxxx_OptimizeWholeGraphhccl_alltoallvc_fusion_optimizer.txt

Graph after fusion optimization in the original HCCL graph optimization phase.

-

ge_proto_xxxx_OptimizeWholeGraphAIcoreEngine.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphDSAEngine.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphhccl_graph_optimizer.txt

Graph optimized in the original HCCL graph optimization phase.

-

ge_proto_xxxx_OptimizeWholeGraphDNN_VM_RTS_GRAPH_OPTIMIZER_STORE.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphDNN_VM_RTS_FFTS_PLUS_GRAPH_OPTIMIZER_STORE.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphDNN_VM_HOST_CPU_OPTIMIZER.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphge_local_graph_optimizer.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_OptimizeWholeGraphffts_plus.txt

Graph information obtained after the original graph optimization API of each engine is called. OptimizeWholeGraph is followed by the engine name.

-

ge_proto_xxxx_PreRunAfterOptimizeWholeGraph.txt

Graph structure obtained after each operator information library is optimized (by calling the OptimizeWholeGraph API).

-

ge_proto_xxxx_PreRunAfterOptimize2.txt

Graph structure obtained after optimization stage 2 for all graphs.

-

ge_proto_xxxx_PreRunAfterOptimizeGraphBeforeBuild.txt

Entry graph for model build.

Trustlisted

ge_proto_xxxx_BeforeHandleMemConflict.txt

Graph before the memory conflict is handled.

-

ge_proto_xxxx_AfterHandleMemConflict.txt

Graph after the memory conflict is handled.

-

ge_proto_xxxx_BeforeHandleMemoryLayoutConflict.txt

Entry graph before the memory layout conflict is handled.

-

ge_proto_xxxx_PreRunAfterMemConflictProc.txt

Exit graph after the memory read/write conflict is handled.

-

ge_proto_xxxx_partition0_rank0_inputNodeGraph_SecondPartitioning.txt

Structure of the input node subgraph obtained after second partitioning.

-

ge_proto_xxxx_partition0_rank1_new_sub_graph1_SecondPartitioning.txt

Structure of subgraph 1 obtained after second partitioning.

-

ge_proto_xxxx_partition0_rank2_new_sub_graph110_SecondPartitioning.txt

Structure of subgraph 110 obtained after second partitioning.

-

ge_proto_xxxx_BeforeUnfoldSubgraphs.txt

Graph before dynamic-shape graph expansion.

-

ge_proto_xxxx_AfterUnfoldSubgraphs.txt

Graph after dynamic-shape graph expansion.

-

ge_proto_xxxx_RunCustomPass_BeforeAssignLogicStream*.txt

Graph before custom stream allocation pass processing. * indicates the specific pass name.

Trustlisted

ge_proto_xxxx_RunCustomPass_AfterAssignLogicStream*.txt

Graph after custom stream allocation pass processing. * indicates the specific pass name.

Trustlisted

ge_proto_xxxx_AfterAssignResource.txt

Graph after stream allocation, memory allocation, and GenTask.

-

ge_proto_xxxx_Build.txt

Exit graph for model build.

Trustlisted

ge_proto_xxxx_PreRunAfterBuild.txt

Same as the preceding graph.

-

ge_proto_xxxx_BeforeAttrsCompress.txt

Graph before offline model compression.

-

ge_proto_xxxx_AfterAttrsCompress.txt

Graph after offline model compression.

-

ge_proto_xxxx_BeforeAttrsDecompress.txt

Graph before offline model decompression.

-

ge_proto_xxxx_AfterAttrsDecompress.txt

Graph after offline model decompression.

-

ge_proto_xxxx_ComputeGraphBeforeLowering.txt

Computational graph before lowering.

Trustlisted

ge_proto_xxxx_Before_MultiStream_LoweringFirstLastEventSync.txt

Execution graph before multi-stream event insertion.

-

ge_proto_xxxx_ExeGraphBeforeOptimize.txt

Execution graph before graph optimization.

Trustlisted

ge_proto_xxxx_AfterZeroCopy.txt

Execution graph after zero-copy optimization.

-

ge_proto_xxxx_AfterCEM.txt

Execution graph after CEM optimization.

-

ge_proto_xxxx_AfterCopyFlowLaunch.txt

Execution graph after in-line copy optimization.

-

ge_proto_xxxx_TrustOutTensorAfter.txt

Execution graph after TrustOutTensor optimization.

-

ge_proto_xxxx_ AicpuFuseHostInputsAfter.txt

Graph after AicpuFuseHostInputs optimization.

-

ge_proto_xxxx_ExecuteGraphAfterSplit.txt

Final execution graph of the dynamic shape.

Trustlisted
Parent topic: References
This site uses cookies to improve your browsing experience. By continuing to browse this site, you agree to our use of cookies.View Details