Who is this for?

This Learning Path is for Edge AI developers who need to inspect model artifacts after backend delegation, understand graph structure and delegate coverage, and use those insights to reason about performance and behavior.

What will you learn?

Upon completion of this Learning Path, you will be able to:

  • Set up Google Model Explorer with the combined ExecuTorch extension and separate Tensor Operator Set Architecture (TOSA) and VGF adapters.
  • Open ExecuTorch deployment graphs and inspect delegate regions, work outside delegates, graph fragmentation, and backend-specific changes
  • Trace model transformations across PTE, TOSA, and VGF artifacts
  • Use ETRecord and ETDump overlays to connect exported graph structure with runtime profiling data

Prerequisites

Before starting, you will need the following:

  • Python 3.10, 3.11, or 3.12
  • Basic familiarity with PyTorch, ExecuTorch, or model deployment workflows
  • Git Large File Storage (LFS)

Summary

AI-assisted

This summary was drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
You’ll use pre-generated .pte, .tosa, .vgf, .etrecord, and .etdp artifacts to understand how Google Model Explorer and Arm extensions expose model export, lowering, backend delegation, and runtime behavior. After installing the ExecuTorch extension and the TOSA and VGF adapters, you’ll use a Cortex-M PTE to learn graph navigation, compare portable and XNNPACK Cortex-A programs, and diagnose no, clean, and fragmented Ethos-U delegation. You’ll then inspect TOSA intermediate graphs, relate TOSA fragments to delegation boundaries, and compare a VGF-backed PTE with the standalone VGF graph used by workflows with the ML extensions for Vulkan. Finally, you’ll open ETRecord files and load their matching ETDump data to identify whether delegate calls or non-delegated operators dominate runtime.

Frequently asked questions

AI-assisted

These FAQs were drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
Which extensions should I enable to load the artifacts used here?
Enable the ExecuTorch extension to open .pte and .etrecord files and add .etdp profiling overlays. Enable the TOSA adapter for .tosa files and the VGF adapter for .vgf files. By running all three, you can switch between formats in one session.
How do I recognize delegate regions and fragmentation in a PTE?
Inspect the graph for a single contiguous delegate region, multiple smaller delegate regions, or nodes that remain outside the delegate. Use this view to see how much of the model is delegated and where non-delegated work occurs.
What result should I expect when opening the Cortex-M .pte example?
The deployment graph renders in the browser with expandable nodes and metadata. Seeing the graph confirms the installation and that the ExecuTorch extension is active.
How do I compare portable and XNNPACK PTEs for Cortex-A CPU paths?
Open both PTE files and examine operator nodes and backend-specific changes. Look for differences in how operators are implemented and where execution paths differ from portable kernels.
How do I connect runtime data to the exported graph?
Open the matching .etrecord file and its .etdp profiling data with the ExecuTorch extension enabled. Model Explorer overlays align these data with graph nodes so you can correlate structure with runtime behavior.
Next