调试后端委托

我们提供了一个 util 函数列表，让用户能够深入了解图形模块在阶段中发生了什么。to_backend()

获取委托摘要

该方法提供了调用后模型所发生情况的摘要：get_delegation_info()to_backend()

from executorch.devtools.backend_debug import get_delegation_info
from tabulate import tabulate

# ... After call to to_backend(), but before to_executorch()
graph_module = edge_manager.exported_program().graph_module
delegation_info = get_delegation_info(graph_module)
print(delegation_info.get_summary())
df = delegation_info.get_operator_delegation_dataframe()
print(tabulate(df, headers="keys", tablefmt="fancy_grid"))

打印输出示例：

Total  delegated  subgraphs:  86
Number  of  delegated  nodes:  473
Number  of  non-delegated  nodes:  430

	op_type	occurrences_in_delegated_graphs	occurrences_in_non_delegated_graphs
0	aten__softmax_default	12	0
1	aten_add_tensor	37	0
2	aten_addmm_default	48	0
3	aten_arange_start_step	0	25
	…
23	aten_view_copy_default	170	48
	…
26	总	473	430

从表中可以看出，该运算符在委托图中出现 170 次，在非委托图中出现 48 次。用户可以使用此类信息进行调试。aten_view_copy_default

可视化委托图

要查看更详细的视图，请使用该方法获取整个图形的 str 打印输出，或者直接用于打印：format_delegated_graph()print_delegated_graph()

from executorch.exir.backend.utils import format_delegated_graph
graph_module = edge_manager.exported_program().graph_module
print(format_delegated_graph(graph_module)) # or call print_delegated_graph(graph_module)

它将打印整个模型以及后端消耗的子图。fx like or 提供的通用调试函数只会显示但隐藏后端使用的子图，而此函数则公开子图内部的内容。print_tabular()print_readable()call_delegate

在下面的示例打印输出中，请注意 and 运算符被委托给，而运算符未被委托。embeddingaddXNNPACKsub

%aten_unsqueeze_copy_default_22 : [num_users=1] = call_function[target=executorch.exir.dialects.edge._ops.aten.unsqueeze_copy.default](args = (%aten_arange_start_step_23, -2), kwargs = {})
  %aten_unsqueeze_copy_default_23 : [num_users=1] = call_function[target=executorch.exir.dialects.edge._ops.aten.unsqueeze_copy.default](args = (%aten_arange_start_step_24, -1), kwargs = {})
  %lowered_module_0 : [num_users=1] = get_attr[target=lowered_module_0]
    backend_id: XnnpackBackend
    lowered graph():
      %aten_embedding_default : [num_users=1] = placeholder[target=aten_embedding_default]
      %aten_embedding_default_1 : [num_users=1] = placeholder[target=aten_embedding_default_1]
      %aten_add_tensor : [num_users=1] = call_function[target=executorch.exir.dialects.edge._ops.aten.add.Tensor](args = (%aten_embedding_default, %aten_embedding_default_1), kwargs = {})
      return (aten_add_tensor,)
  %executorch_call_delegate : [num_users=1] = call_function[target=torch.ops.higher_order.executorch_call_delegate](args = (%lowered_module_0, %aten_embedding_default, %aten_embedding_default_1), kwargs = {})
  %aten_sub_tensor : [num_users=1] = call_function[target=executorch.exir.dialects.edge._ops.aten.sub.Tensor](args = (%aten_unsqueeze_copy_default, %aten_unsqueeze_copy_default_1), kwargs = {})