Brief
Google AI Edge Model Explorer adds visual graph debugging for Arm’s custom edge compilers
Arm used the new Model Explorer features to validate VGF exports, compare NPU and CPU graphs, and map profiling traces onto execution graphs.
By Felo News Desk · Published
On Oct. 20, 2026 Google’s AI Edge team announced that Model Explorer now offers visual graph debugging for custom edge compilers, a capability demonstrated by Arm’s integration of the tool.
What happened
Arm engineers built open‑source adapters that let Model Explorer ingest Arm‑specific intermediate representations such as TOSA, VGF and PTE. The visual workflow lets developers inspect hierarchical graphs, view node attributes and overlay cycle‑level profiling data directly on the execution graph. In a case study, a PyTorch‑trained Neural Super Sampling model exported as a VGF file showed unintended TRANSPOSE operators after a backend change, which were identified and fixed using Model Explorer’s node‑detail view.
What the reports add
The blog post notes that the split‑pane view lets engineers compare a CPU‑only graph with an NPU‑delegated graph for a Ultralytics YOLO model, revealing structural differences such as layer grouping on the Ethos‑U85 NPU. It also highlights that the tool can map SPIR‑V tensor connections for Vulkan ML extensions, aiding developers working on graphics‑accelerated inference.
What was said
Alan Ma, a staff software engineer at Google, wrote that Model Explorer’s “hierarchical graph visualization and plugin architecture” enable Arm to “make edge model transformations easier to inspect.” The post describes how the visual debugger reduces reliance on raw flatbuffers or text logs, especially for large language models and complex neural networks.
How it came about
Running machine‑learning models on edge devices requires multiple compilation steps—quantization, graph lowering, operator fusion and hardware delegation. Arm faced challenges when changes in its ExecuTorch backend introduced hidden operators that could affect performance or correctness. By extending Model Explorer, Arm created a workflow that lets engineers answer three key questions: whether graph lowering preserved structural integrity, whether the model was fully delegated to the intended accelerator, and which nodes consumed the most compute cycles.
Key facts
- Google’s AI Edge Model Explorer now supports visual debugging of custom compiler IRs such as TOSA, VGF and PTE. (developers.googleblog.com)
- Arm used the tool to detect unintended TRANSPOSE operators in a VGF export of a Neural Super Sampling model. (developers.googleblog.com)
- The split‑pane view lets developers compare CPU and NPU graphs side‑by‑side for models like Ultralytics YOLO. (developers.googleblog.com)
- Model Explorer can overlay cycle‑level profiling traces onto the execution graph. (developers.googleblog.com)
Sources
- [1] developers.googleblog.com — originally reported as “Bringing Visual Graph Debugging to Custom Edge Compilers with Model Explorer- Google Developers Blog”
