All Docs

Extension Docs

Topics Overview Overview Node.js Python ASP.NET Core Debug Docker Compose Registries Deploy to Azure Choose a Dev Environment Customize Develop with Kubernetes Tips and Tricks Overview Jupyter Notebooks Data Science Tutorial Python Interactive Data Wrangler Quick Start Data Wrangler PyTorch Support Azure Machine Learning Manage Jupyter Kernels Jupyter Notebooks on the Web Data Science in Microsoft Fabric Foundry Toolkit Overview Foundry Toolkit Copilot Tools Create Agents Models Playground Agent Builder Agent Inspector Hosted Agents (Preview) Evaluation Tool Catalog Fine-Tuning (Automated Setup) Fine-Tuning (Project Template) Model Conversion Tracing Profiling (Windows ML) FAQ Reference File Structure Manual Model Conversion Manual Model Conversion on GPU Setup Environment Without Foundry Toolkit Update Project Template Project Migrating from Visualizer to Agent Inspector Overview Getting Started Resources View Deployment VS Code for the Web - Azure Containers Azure Kubernetes Service Kubernetes MongoDB Remote Debugging for Node.js Overview Tutorial Attach to Container Create Dev Container Advanced Containers devcontainer.json Dev Container CLI Tips and Tricks FAQ

Profiling an app using Windows Machine Learning

Profiling is a tool designed to help developers and AI engineers to diagnose CPU, GPU, and NPU resource usage of processes, profile ONNX models on different execution providers, and capture Windows ML events.

In this article, you learn how to start profiling and how to inspect the resource usage view and the events view.

Prerequisites

Install the latest version of Visual Studio Code.
Install the Foundry Toolkit VS Code extension. For more information, see install Foundry Toolkit.

Profile on app startup

In this mode, the profiling tool profiles the next app that is started and that is sending out Windows ML events. This option is ideal for testing a run-once app. In this case, you start profiling, then run the app, and the resource usage will appear.

[Screenshot that shows how to start by the next session]

The tool starts profiling a newly started app. This means that for profiling a Python notebook, if the kernel is already running, you need to restart the kernel to begin profiling for it. Just starting a new notebook does not automatically start profiling.

Important

To receive Windows ML events, the tool needs to be run in admin mode. If VS Code is not started in admin mode, a notification shows up and guides you to restart VS Code. You need to close all other VS Code instances to make the restart in admin mode work. [Screenshot that shows a notification to restart VS Code in admin mode]

Profile a running app

In this mode, the profiling tool starts profiling an already running app. You can select a process based on these criteria:

Process ID: like 12345
Process name: usually the name of app without .exe. The first match will be profiled.
Process path: like c:\Users\xxx\Inference.Service.Agent.exe. The first match will be profiled.

This option is ideal for profiling an app that is already running and you're unable to restart it for profiling purposes.

[Screenshot that shows how to start by process id or name]

Profile an ONNX model

In this mode, the profiling tool starts profiling an ONNX model file on a target execution provider (EP) for a given duration. You can see the resource usage while it's running.

This option is ideal for profiling an ONNX model on different EPs.

[Screenshot that shows how to start by model file]

After profiling, a notification shows up to guide you to open or save the report.

[Screenshot that shows the succeeded notification]

The report contains detailed profiling statistics and results for the ONNX model.

[Screenshot that shows the report data]

Benchmark time for each operation

If OP Profiling is enabled, op level data will be generated to allow you to inspect the model in more detail.

[Screenshot that shows the succeeded notification with OP profiling enabled]

The report contains detailed latencies for each op.

[Screenshot that shows the report data for each OP]

Enable op level data for Intel (OpenVINO) EP

To see op level data when running on the Intel (OpenVINO) EP, do the following:

Download Intel Unified Telemetry
Extract the content to a path like C:\Users\XXX\Downloads\ut-tool-ext-v0.2.0-beta1.1
In Foundry Toolkit settings, set Model Lab Intel Unified Telemetry Path to C:\Users\XXX\Downloads\ut-tool-ext-v0.2.0-beta1.1\bin

Intel Unified Telemetry requires administrator privileges to run. If VS Code is not running as administrator, profiling will prompt for elevation and open a new terminal window showing its progress. Don't close this window until the process has completed.

If VS Code is already running as administrator, you won't be prompted and no additional window will open.

Profile an ONNX GenAI model

In this mode, the profiling tool starts profiling an ONNX GenAI model on a target execution provider (EP) for a specified number of prompts. You can see the resource usage while it's running.

[Screenshot that shows how to start by genai model]

Note

You need to select the folder of the GenAI model, this is the folder that contains genai_config.json.

Resource Usage view

In the main window, the plot on the top shows usage of CPU, GPU, NPU, and memory. The usage is updated every second, and is kept for 10 minutes. You can use the tools on the top right to navigate the timeline by zooming in, zooming out, and panning.

[Screenshot that shows the resource usage view]

Note

This feature uses performance counters. To achieve higher accuracy, you could also try Windows Performance Recorder.

Windows ML Events view

In the main window, the plot on the bottom shows Windows ML events. Its timeline is synchronized with the Resource Usage view, so you can easily determine how resources are used when certain events occur.