nvidia--tensorrt/tools/Polygraphy/how-to/use_debug_subtools_effectively.md

Using debug Subtools Effectively

Table Of Contents

• Introduction
• Tips And Tricks
  • Using Interactive Mode To Get Started
  • What Counts As A "Failure"?

Introduction

The debug subtools work on the same general principles:

1. Iteratively perform some task that generates some output
2. Evaluate the generated output to determine if it should be considered good or bad
3. Sort any tracked artifacts into good and bad directories based on (2)
4. Make changes if required and then repeat the process

This guide provides some general information as well as tips and tricks on how to use debug subtools effectively.

Tips And Tricks

Using Interactive Mode To Get Started

Automatic mode (i.e. when a --check command is provided) in the debug subtools provides a totally hands-off approach. However, if you're just starting out, it can also be harder to use and reason about. The debug tools in Polygraphy also provide an interactive mode which will prompt and guide you through the process.

To use interactive mode, you can simply omit the --check option. For example, with debug reduce:

polygraphy debug reduce <model.onnx> -o <reduced.onnx>

What Counts As A "Failure"?

In automatic mode (i.e. when a --check command is provided), debug subtools will assume that a non-zero exit status indicates a failure. There are a few cases where this assumption may not be desirable:

• You may want to ignore certain types of failures so that you can get a minimal model that fails in a very specific way.

• Your --check command may not exit with a non-zero status at all! Instead, you may want to trigger a failure based on the output of such a command instead of its exit status.

For these cases, you can use the --fail-regex option, which will trigger a failure only when some part of the output from the --check command (on either stdout or stderr) matches one of the specified regular expressions.