DUMP_GRAPH_FORMAT

Description

Controls the type of the dump file to be generated.

Available values:

onnx: model description structure based on ONNX. You can open this file using visualizer software such as Netron. The generated file name is ge_onnx*.pbtxt.
ge_proto: text file stored in Protobuf format. The generated file name is ge_proto*.txt.
readable: highly readable text file in Dynamo FX graph style. The generated file name is ge_readable*.txt. For details about the file content, see Readable File Parsing.

How to use:

The configuration is a character string separated by vertical bars (|). The type is in lowercase. Combined configuration is supported, for example, ge_proto|onnx|readable indicates that all types of files are dumped. Separate configuration is also supported, for example, ge_proto indicates that only ge_proto*.txt files of the ge_proto type are dumped.

DUMP_GRAPH_FORMAT takes effect only when DUMP_GE_GRAPH is enabled. The default value is ge_proto|onnx.

Example

export DUMP_GRAPH_FORMAT="ge_proto|onnx"

Restrictions

If this environment variable is set to an invalid value, undefined behavior may occur.
If the operator dump data collection function is enabled, see ge.exec.enableDump in Graph Development.
The subgraph ge_proto_xxxx_Build.txt is dumped even if the value of DUMP_GRAPH_FORMAT does not contain ge_proto.

Readable File Parsing

The following is a complete ge_readable*.txt file:

graph("MakeTransformerSubGraph"):
  %input_0 : [#users=1] = Node[type=Data] (attrs = {index: 0})
  %input_1 : [#users=1] = Node[type=Data] (attrs = {index: 1})
  %input_2 : [#users=1] = Node[type=Data] (attrs = {index: 2})
  %Const_0 : [#users=1] = Node[type=Const] (attrs = {value: [-1 7168]})
  %Reshape_1 : [#users=1] = Node[type=Reshape] (inputs = (x=%input_0, shape=%Const_0), attrs = {axis: 0, num_axes: -1})
  %Cast_2 : [#users=1] = Node[type=Cast] (inputs = (x=%Reshape_1), attrs = {dst_type: 0})
  %Cast_3 : [#users=1] = Node[type=Cast] (inputs = (x=%input_1), attrs = {dst_type: 0})
  %Const_4 : [#users=1] = Node[type=Const] (attrs = {value: [1 0]})
  %Transpose_5 : [#users=1] = Node[type=Transpose] (inputs = (x=%Cast_3, perm=%Const_4))
  %MatMul_6 : [#users=1] = Node[type=MatMul] (inputs = (x1=%Cast_2, x2=%Transpose_5), attrs = {transpose_x1: false, transpose_x2: false})
  %Sigmoid_7 : [#users=1] = Node[type=Sigmoid] (inputs = (x=%MatMul_6))
  %Const_8 : [#users=1] = Node[type=Const] (attrs = {value: [-1 256]})
  %Reshape_9 : [#users=1] = Node[type=Reshape] (inputs = (x=%Sigmoid_7, shape=%Const_8), attrs = {axis: 0, num_axes: -1})
  %Unsqueeze_10 : [#users=1] = Node[type=Unsqueeze] (inputs = (x=%input_2), attrs = {axes: {0}})
  %Cast_11 : [#users=1] = Node[type=Cast] (inputs = (x=%Unsqueeze_10), attrs = {dst_type: 0})
  %Add_12 : [#users=1] = Node[type=Add] (inputs = (x1=%Reshape_9, x2=%Cast_11))
  %Const_13 : [#users=1] = Node[type=Const] (attrs = {value: [2]})
  %TopKV2_14 : [#users=2] = Node[type=TopKV2] (inputs = (x=%Add_12, k=%Const_13), attrs = {sorted: true, dim: -1, largest: true, indices_dtype: 3})
  %ret : [users=1] = get_element[node=%TopKV2_14](0)
  %ret_1 : [users=0] = get_element[node=%TopKV2_14](1)
  %Const_15 : [#users=1] = Node[type=Const] (attrs = {value: [-1]})
  %ReduceSum_16 : [#users=1] = Node[type=ReduceSum] (inputs = (x=%ret, axes=%Const_15), attrs = {keep_dims: false, noop_with_empty_axes: true})
  %Const_17 : [#users=1] = Node[type=Const] (attrs = {value: [4]})
  %TopKV2_18 : [#users=2] = Node[type=TopKV2] (inputs = (x=%ReduceSum_16, k=%Const_17), attrs = {sorted: false, dim: -1, largest: true, indices_dtype: 3})
  %ret_2 : [users=0] = get_element[node=%TopKV2_18](0)
  %ret_3 : [users=1] = get_element[node=%TopKV2_18](1)
  %Cast_19 : [#users=1] = Node[type=Cast] (inputs = (x=%ret_3), attrs = {dst_type: 9})
  %ZerosLike_20 : [#users=1] = Node[type=ZerosLike] (inputs = (x=%ReduceSum_16))
  %Shape_21 : [#users=1] = Node[type=Shape] (inputs = (x=%Cast_19), attrs = {dtype: 3})
  %Const_22 : [#users=1] = Node[type=Const] (attrs = {value: [1.000000]})
  %Cast_23 : [#users=1] = Node[type=Cast] (inputs = (x=%Const_22), attrs = {dst_type: 0})
  %Fill_24 : [#users=1] = Node[type=Fill] (inputs = (dims=%Shape_21, value=%Cast_23))
  %ScatterElements_25 : [#users=1] = Node[type=ScatterElements] (inputs = (data=%ZerosLike_20, indices=%Cast_19, updates=%Fill_24), attrs = {axis: 0, reduction: "none"})
  %Unsqueeze_26 : [#users=1] = Node[type=Unsqueeze] (inputs = (x=%ScatterElements_25), attrs = {axes: {-1}})
  %Const_27 : [#users=1] = Node[type=Const] (attrs = {value: [256 256]})
  %BroadcastTo_28 : [#users=1] = Node[type=BroadcastTo] (inputs = (x=%Unsqueeze_26, shape=%Const_27))
  %Identity_29 : [#users=1] = Node[type=Identity] (inputs = (x=%BroadcastTo_28))
  %Const_30 : [#users=1] = Node[type=Const] (attrs = {value: [256 256]})
  %Reshape_31 : [#users=1] = Node[type=Reshape] (inputs = (x=%Identity_29, shape=%Const_30), attrs = {axis: 0, num_axes: -1})
  %Cast_32 : [#users=1] = Node[type=Cast] (inputs = (x=%Reshape_31), attrs = {dst_type: 12})
  %LogicalNot_33 : [#users=1] = Node[type=LogicalNot] (inputs = (x=%Cast_32))
  %Const_34 : [#users=1] = Node[type=Const] (attrs = {value: [0.000000]})
  %MaskedFill_35 : [#users=1] = Node[type=MaskedFill] (inputs = (x=%Add_12, mask=%LogicalNot_33, value=%Const_34))
  %Const_36 : [#users=1] = Node[type=Const] (attrs = {value: [4]})
  %TopKV2_37 : [#users=2] = Node[type=TopKV2] (inputs = (x=%MaskedFill_35, k=%Const_36), attrs = {sorted: false, dim: -1, largest: true, indices_dtype: 3})
  %ret_4 : [users=0] = get_element[node=%TopKV2_37](0)
  %ret_5 : [users=1] = get_element[node=%TopKV2_37](1)
  %Cast_38 : [#users=1] = Node[type=Cast] (inputs = (x=%ret_5), attrs = {dst_type: 9})
  %GatherElements_39 : [#users=1] = Node[type=GatherElements] (inputs = (x=%Sigmoid_7, index=%Cast_38), attrs = {dim: 1})
  %Const_40 : [#users=1] = Node[type=Const] (attrs = {value: [0.000001]})
  %RealDiv_41 : [#users=1] = Node[type=RealDiv] (inputs = (x1=%GatherElements_39, x2=%Const_40))
  %Const_42 : [#users=1] = Node[type=Const] (attrs = {value: [2.500000]})
  %Mul_43 : [#users=1] = Node[type=Mul] (inputs = (x1=%RealDiv_41, x2=%Const_42))
  %Cast_44 : [#users=1] = Node[type=Cast] (inputs = (x=%Mul_43), attrs = {dst_type: 0})

  return (output_0=%Cast_38, output_1=%Cast_44)

The following is an example of a file that contains subgraphs:

graph("TransformerBlockSubgraph"):
  %input_0 : [#users=1] = Node[type=Data] (attrs = {index: 0})
  %pred_1 : [#users=1] = Node[type=Data] (attrs = {index: 1})
  %If_0 : [#users=1] = Node[type=If] (inputs = (cond=%pred_1, input_0=%input_0), attrs = {then_branch: %If_then, else_branch: %If_else})
  %Const_1 : [#users=1] = Node[type=Const] (attrs = {value: [0]})
  %Const_2 : [#users=1] = Node[type=Const] (attrs = {value: [4]})
  %Const_3 : [#users=1] = Node[type=Const] (attrs = {value: [1]})
  %For_6 : [#users=2] = Node[type=For] (inputs = (start=%Const_1, limit=%Const_2, delta=%Const_3, input_1=%If_0), attrs = {body: %For_body})
  %ret : [users=1] = get_element[node=%For_6](0)
  %ret_1 : [users=1] = get_element[node=%For_6](1)

  return (output_0=%ret, output_1=%ret_1)

graph("If_then"):
  %input_0 : [#users=1] = Node[type=Data] (attrs = {index: 0})
  %Const_0 : [#users=1] = Node[type=Const] (attrs = {value: [0.900000]})
  %Mul_1 : [#users=1] = Node[type=Mul] (inputs = (x1=%input_0, x2=%Const_0))
  return (%Mul_1)

graph("If_else"):
  %input_0 : [#users=1] = Node[type=Data] (attrs = {index: 0})
  %Identity_0 : [#users=1] = Node[type=Identity] (inputs = (x=%input_0))
  return (%Identity_0)

graph("For_body"):
  %iter : [#users=1] = Node[type=Data] (attrs = {index: 0})
  %hidden : [#users=1] = Node[type=Data] (attrs = {index: 1})
  %Const_0 : [#users=1] = Node[type=Const] (attrs = {value: [1]})
  %Add_0 : [#users=1] = Node[type=Add] (inputs = (x1=%iter, x2=%Const_0))
  %Const_1 : [#users=1] = Node[type=Const] (attrs = {value: [0.500000]})
  %Mul_1 : [#users=1] = Node[type=Mul] (inputs = (x1=%hidden, x2=%Const_1))
  %Add_2 : [#users=1] = Node[type=Add] (inputs = (x1=%hidden, x2=%Mul_1))

  return (output_0=%Add_0, output_1=%Add_2)

The following describes each field:

Graph name: graph("<Graph name>")
Node instance: %<Node instance name> : [#users=<Out-degree>] = Node[type=<Node type>](inputs = (<Input name 1>=%<Input instance 1>, ...), attrs = {<Attribute name 1>: <Attribute value 1>, ...})
- <Node instance name>: name of a node instance
- #users=<Out-degree>: number of node outputs
- Node[type=<Node type>]: operator type corresponding to the node. For example, the MatMul node is displayed as Node[type=MatMul].
- inputs = (<Input name 1>=%<Input instance 1>, ...): The node input is displayed in the format of Parameter name=Instance name. If the parameter name fails to be parsed, the input is rolled back to the _input_N sequence. The default value is used when the node has no input.
  Dynamic input: The dynamic input parameters are numbered in the format of Input name_#cnt (Input name_0, Input name_1, ...).
- attrs = {<Attribute name 1>: <Attribute value 1>, ...}: node attribute set, including subgraph attribute items and common attribute items. The default values are used when there are no attributes.
Output reference of a multi-output node: %ret/ret_#cnt : [#users=<Number of consumers>] = get_element[node=%<Node instance name>](<Output index>).
- %ret/ret_#cnt: Each output of a multi-output node is named in this format: ret, ret_1, ret_2, ...
- [#users=<Number of consumers>]: number of consumers of the output
- get_element[node=%<Node instance name>](<Output index>): Indicates that the <Output index>^th output is extracted from the multi-output node.
Graph output: return (<Output list>)
Graph output, corresponding to the input of the NetOutput node.
- Single output: return (%<Output instance>)
- Multi-output: return (output_0=%<Output instance 0>, output_1=%<Output instance 1>, ...) Outputs are numbered in the format of output_#cnt (output_0, output_1, ...).
Subgraph representation: Nodes that contain subgraphs are represented as follows:
- Subgraph declaration: Declare in the attrs of the parent node in this format: attrs = {<Subgraph attribute name>: %<Subgraph instance name>, ... }; If the subgraph attribute name fails to be parsed, rollback is performed to the _graph_N sequence.
- Input mapping: The input_#cnt (or args_#cnt) of the parent node corresponds to the Data node whose index value is cnt in the subgraph. For example, input_0 of the parent node corresponds to Data(attrs = {index: 0}) of the subgraph.
- Output mapping:
  - Single output: The return value of the subgraph is directly used as the output of the parent node.
  - Multiple outputs: The output_#cnt of the subgraph corresponds to the cnt^th output of the parent node. The parent node extracts the corresponding output through get_element[node=%<Parent node>](cnt).
- Subgraph display position: The subgraph content is displayed separately after the main output of the parent graph is complete. Each subgraph starts with graph("<Subgraph instance name>").

Applicability

Atlas training product

Atlas inference product

Atlas A2 training product / Atlas A2 inference product

Atlas A3 training product / Atlas A3 inference product

Atlas 350 Accelerator Card

Atlas 200I/500 A2 inference product

Parent topic: Graph Building