Using the ExecuTorch SDK to Profile a Model

Author: Jack Khuu

The ExecuTorch SDK is a set of tools designed to provide users with the ability to profile, debug, and visualize ExecuTorch models.

This tutorial will show a full end-to-end flow of how to utilize the SDK. Specifically, it will:

1. Generate the artifacts consumed by the SDK (ETRecord, ETDump).

2. Create an Inspector class consuming these artifacts.

3. Utilize the Inspector class to analyze the model.

Prerequisites

To run this tutorial, you’ll first need to Set up your ExecuTorch environment.

Generate ETRecord (Optional)

The first step is to generate an ETRecord. ETRecord contains model graphs and metadata for linking runtime results (such as profiling) to the eager model. This is generated via executorch.sdk.generate_etrecord.

executorch.sdk.generate_etrecord takes in an output file path (str), the edge dialect model (EdgeProgramManager), the ExecuTorch dialect model (ExecutorchProgramManager), and an optional dictionary of additional models.

In this tutorial, an example model (shown below) is used to demonstrate.

import copy

import torch
import torch.nn as nn
import torch.nn.functional as F

from executorch.exir import (
    EdgeCompileConfig,
    EdgeProgramManager,
    ExecutorchProgramManager,
    to_edge,
)
from executorch.sdk import generate_etrecord
from torch.export import export, ExportedProgram


# Generate Model
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        # 1 input image channel, 6 output channels, 5x5 square convolution
        # kernel
        self.conv1 = nn.Conv2d(1, 6, 5)
        self.conv2 = nn.Conv2d(6, 16, 5)
        # an affine operation: y = Wx + b
        self.fc1 = nn.Linear(16 * 5 * 5, 120)  # 5*5 from image dimension
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        # Max pooling over a (2, 2) window
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        # If the size is a square, you can specify with a single number
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = torch.flatten(x, 1)  # flatten all dimensions except the batch dimension
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


model = Net()

aten_model: ExportedProgram = export(
    model,
    (torch.randn(1, 1, 32, 32),),
)

edge_program_manager: EdgeProgramManager = to_edge(
    aten_model, compile_config=EdgeCompileConfig(_check_ir_validity=True)
)
edge_program_manager_copy = copy.deepcopy(edge_program_manager)
et_program_manager: ExecutorchProgramManager = edge_program_manager.to_executorch()


# Generate ETRecord
etrecord_path = "etrecord.bin"
generate_etrecord(etrecord_path, edge_program_manager_copy, et_program_manager)

Warning

Users should do a deepcopy of the output of to_edge() and pass in the deepcopy to the generate_etrecord API. This is needed because the subsequent call, to_executorch(), does an in-place mutation and will lose debug data in the process.

Generate ETDump

Next step is to generate an ETDump. ETDump contains runtime results from executing a Bundled Program Model.

In this tutorial, a Bundled Program is created from the example model above.

import torch

from executorch.exir import to_edge
from executorch.sdk import BundledProgram

from executorch.sdk.bundled_program.config import MethodTestCase, MethodTestSuite
from executorch.sdk.bundled_program.serialize import (
    serialize_from_bundled_program_to_flatbuffer,
)
from torch.export import export

# Step 1: ExecuTorch Program Export
m_name = "forward"
method_graphs = {m_name: export(model, (torch.randn(1, 1, 32, 32),))}

# Step 2: Construct Method Test Suites
inputs = [[torch.randn(1, 1, 32, 32)] for _ in range(2)]

method_test_suites = [
    MethodTestSuite(
        method_name=m_name,
        test_cases=[
            MethodTestCase(inputs=inp, expected_outputs=getattr(model, m_name)(*inp))
            for inp in inputs
        ],
    )
]

# Step 3: Generate BundledProgram
executorch_program = to_edge(method_graphs).to_executorch()
bundled_program = BundledProgram(executorch_program, method_test_suites)

# Step 4: Serialize BundledProgram to flatbuffer.
serialized_bundled_program = serialize_from_bundled_program_to_flatbuffer(
    bundled_program
)
save_path = "bundled_program.bp"
with open(save_path, "wb") as f:
    f.write(serialized_bundled_program)

Use CMake (follow these instructions to set up cmake) to execute the Bundled Program to generate the ETDump:

cd executorch
./examples/sdk/build_sdk_example_runner.sh
cmake-out/examples/sdk/sdk_example_runner --bundled_program_path="bundled_program.bp"

Creating an Inspector

Final step is to create the Inspector by passing in the artifact paths. Inspector takes the runtime results from ETDump and correlates them to the operators of the Edge Dialect Graph.

Recall: An ETRecord is not required. If an ETRecord is not provided, the Inspector will show runtime results without operator correlation.

To visualize all runtime events, call Inspector’s print_data_tabular.

from executorch.sdk import Inspector

etdump_path = "etdump.etdp"
inspector = Inspector(etdump_path=etdump_path, etrecord=etrecord_path)
inspector.print_data_tabular()

False

Analyzing with an Inspector

Inspector provides 2 ways of accessing ingested information: EventBlocks and DataFrames. These mediums give users the ability to perform custom analysis about their model performance.

Below are examples usages, with both EventBlock and DataFrame approaches.

# Set Up
import pprint as pp

import pandas as pd

pd.set_option("display.max_colwidth", None)
pd.set_option("display.max_columns", None)

If a user wants the raw profiling results, they would do something similar to finding the raw runtime data of an addmm.out event.

for event_block in inspector.event_blocks:
    # Via EventBlocks
    for event in event_block.events:
        if event.name == "native_call_addmm.out":
            print(event.name, event.perf_data.raw)

    # Via Dataframe
    df = event_block.to_dataframe()
    df = df[df.event_name == "native_call_addmm.out"]
    print(df[["event_name", "raw"]])
    print()

If a user wants to trace an operator back to their model code, they would do something similar to finding the module hierarchy and stack trace of the slowest convolution.out call.

for event_block in inspector.event_blocks:
    # Via EventBlocks
    slowest = None
    for event in event_block.events:
        if event.name == "native_call_convolution.out":
            if slowest is None or event.perf_data.p50 > slowest.perf_data.p50:
                slowest = event
    if slowest is not None:
        print(slowest.name)
        print()
        pp.pprint(slowest.stack_traces)
        print()
        pp.pprint(slowest.module_hierarchy)

    # Via Dataframe
    df = event_block.to_dataframe()
    df = df[df.event_name == "native_call_convolution.out"]
    if len(df) > 0:
        slowest = df.loc[df["p50"].idxmax()]
        print(slowest.event_name)
        print()
        pp.pprint(slowest.stack_traces)
        print()
        pp.pprint(slowest.module_hierarchy)

If a user wants the total runtime of a module, they can use find_total_for_module.

print(inspector.find_total_for_module("L__self__"))
print(inspector.find_total_for_module("L__self___conv2"))

0.0
0.0

Note: find_total_for_module is a special first class method of Inspector

Conclusion

In this tutorial, we learned about the steps required to consume an ExecuTorch model with the ExecuTorch SDK. It also showed how to use the Inspector APIs to analyze the model run results.

Links Mentioned

• ExecuTorch SDK

• ETRecord

• ETDump

• Inspector