DUMP_GRAPH_FORMAT
功能描述
控制需要生成的dump文件类型。
支持的取值包括:
- onnx:基于ONNX的模型描述结构,可以使用Netron等可视化软件打开,生成文件名为ge_onnx*.pbtxt。
- ge_proto:protobuf格式存储的文本文件,生成文件名为ge_proto*.txt。
- readable:类似Dynamo fx图风格的高可读性文本文件,生成文件名为ge_readable*.txt。文件内容解析请参见readable文件解析。
使用方式:
配置按照“|”分隔的字符串,类型全部为小写,支持组合配置,例如“ge_proto|onnx|readable”,表示dump全量类型的文件;也支持单独配置,例如“ge_proto”,表示只dump ge_proto类型的文件ge_proto*.txt。
DUMP_GRAPH_FORMAT环境变量只有在DUMP_GE_GRAPH开启时才生效,默认为“ge_proto|onnx”。
配置示例
export DUMP_GRAPH_FORMAT="ge_proto|onnx"
使用约束
readable文件解析
- 如下为一个完整的ge_readable*.txt文件:
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) - 如下为一个包含子图的文件:
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)
字段解释如下:
- 图名称:graph("<图名称>")。
- 节点实例:%<节点实例名称> : [#users=<出度>] = Node[type=<节点类型>](inputs = (<输入名称1>=%<输入实例1>, ...), attrs = {<属性名称1>: <属性值1>, ...})。
- <节点实例名称>:节点实例名称。
- #users=<出度> :节点输出个数。
- Node[type=<节点类型>]:节点对应的算子类型,例如MatMul节点显示为Node[type=MatMul]。
- inputs = (<输入名称1>=%<输入实例1>, ...):节点输入以“参数名=实例名”形式展示;参数名解析异常时,回退为 _input_N 序列。当节点无输入时该字段缺省。
- attrs = {<属性名称1>: <属性值1>, ...}:节点属性集合,包含子图属性项与普通属性项,当属性为空时该字段缺省。
- 多输出节点的输出引用:%ret/ret_#cnt : [#users=<消费者个数>] = get_element[node=%<节点实例名称>](<输出index>)。
- %ret/ret_#cnt:多输出节点的每一路输出统一使用ret命名ret, ret_1, ret_2, ...。
- [#users=<消费者个数>]:该输出的消费者数量。
- get_element[node=%<节点实例名称>](<输出index>):以get_element表示“从多输出节点提取第 <输出index> 路输出”。
- 图输出:return (<输出列表>)
- 表示图的输出,对应NetOutput节点的输入。
- 单输出:return (%<输出实例>)。
- 多输出:return (output_0=%<输出实例0>, output_1=%<输出实例1>, ...)。输出使用output_#cnt形式编号(output_0, output_1, ...)。
- 子图表示:包含子图的节点按以下规则表示:
- 子图声明:在父节点的attrs中声明:attrs = {<子图属性名称>: %<子图实例名>, ... };子图属性名称解析异常时,回退为_graph_N序列。
- 输入对应关系:父节点的输入input_#cnt(或 args_#cnt)对应子图中index属性值为cnt的Data节点。例如:父节点的input_0 →子图的Data(attrs = {index: 0})。
- 输出对应关系:
- 单输出:子图返回值直接作为父节点输出。
- 多输出:子图的输出output_#cnt对应父节点的第cnt路输出,父节点通过get_element[node=%<父节点>](cnt)提取对应输出。
- 子图展示位置:子图内容在父图主体输出结束后单独展示;每个子图以 graph("<子图实例名>"):开始。
支持的型号
Atlas 350 加速卡
父主题: 图编译