1# Debug Backend Delegate 2 3We provide a list of util functions to give users insights on what happened to the graph modules during the `to_backend()` stage. 4 5## Get delegation summary 6The `get_delegation_info()` method provides a summary of what happened to the model after the `to_backend()` call: 7 8```python 9from executorch.devtools.backend_debug import get_delegation_info 10from tabulate import tabulate 11 12# ... After call to to_backend(), but before to_executorch() 13graph_module = edge_manager.exported_program().graph_module 14delegation_info = get_delegation_info(graph_module) 15print(delegation_info.get_summary()) 16df = delegation_info.get_operator_delegation_dataframe() 17print(tabulate(df, headers="keys", tablefmt="fancy_grid")) 18``` 19 20Example printout: 21``` 22Total delegated subgraphs: 86 23Number of delegated nodes: 473 24Number of non-delegated nodes: 430 25``` 26 27 28| | op_type | occurrences_in_delegated_graphs | occurrences_in_non_delegated_graphs | 29|----|---------------------------------|------- |-----| 30| 0 | aten__softmax_default | 12 | 0 | 31| 1 | aten_add_tensor | 37 | 0 | 32| 2 | aten_addmm_default | 48 | 0 | 33| 3 | aten_arange_start_step | 0 | 25 | 34| | ... | | | 35| 23 | aten_view_copy_default | 170 | 48 | 36| | ... | | | 37| 26 | Total | 473 | 430 | 38 39From the table, the operator `aten_view_copy_default` appears 170 times in delegate graphs and 48 times in non-delegated graphs. Users can use information like this to debug. 40 41## Visualize delegated graph 42To see a more detailed view, use the `format_delegated_graph()` method to get a str of printout of the whole graph or use `print_delegated_graph()` to print directly: 43 44```python 45from executorch.exir.backend.utils import format_delegated_graph 46graph_module = edge_manager.exported_program().graph_module 47print(format_delegated_graph(graph_module)) # or call print_delegated_graph(graph_module) 48``` 49It will print the whole model as well as the subgraph consumed by the backend. The generic debug function provided by fx like `print_tabular()` or `print_readable()` will only show `call_delegate` but hide the the subgraph consumes by the backend, while this function exposes the contents inside the subgraph. 50 51In the example printout below, observe that `embedding` and `add` operators are delegated to `XNNPACK` while the `sub` operator is not. 52 53``` 54%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 = {}) 55 %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 = {}) 56 %lowered_module_0 : [num_users=1] = get_attr[target=lowered_module_0] 57 backend_id: XnnpackBackend 58 lowered graph(): 59 %aten_embedding_default : [num_users=1] = placeholder[target=aten_embedding_default] 60 %aten_embedding_default_1 : [num_users=1] = placeholder[target=aten_embedding_default_1] 61 %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 = {}) 62 return (aten_add_tensor,) 63 %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 = {}) 64 %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 = {}) 65``` 66