WindowsML python

Browse code Download ZIP

Steps

Download test model

..\Download-Model.ps1

Prepare the python environment

  • Create a python environment with Python 3.10 - 3.13
  • Ensure that your Python installation is not from the Microsoft Store (you can install an unpackaged version from python.org or via winget). The sample depends on using the Windows App SDK dynamic dependency API, which is only valid for unpackaged apps.
  • Install the latest WinML python packages
.\Install-Requirements.ps1

The pinned Windows ML 2.3 package includes the compiled-model compatibility APIs used by this sample.

Install WindowsAppRuntime

Please install the WindowsAppRuntime that matches the version of the python package wasdk-Microsoft.Windows.ApplicationModel.DynamicDependency.Bootstrap

For experimental or preview WASDK. The version tags -xxxN are changed to .devN to fit Python's version requirements.

Run the example

python main.py

Compiled model compatibility

The sample uses the PREFER_NPU execution provider device policy and caches the compiled model as Model\SqueezeNet_ctx.onnx. Before reusing that file, it:

  1. Calls get_ep_devices and orders the returned devices for policy selection.
  2. For the PREFER_NPU policy, selects the preferred NPU when one is available and the CPU fallback devices.
  3. Reads each relevant execution provider's metadata with get_compatibility_info_from_model and validates it with get_model_compatibility_for_ep_devices.
  4. Separates the selected devices into single-EP groups because devices from different execution providers cannot be passed together.
  5. Requires the preferred execution provider group to have matching metadata, and reuses the cached model only when every selected group with metadata reports OrtCompiledModelCompatibility.EP_SUPPORTED_OPTIMAL. Missing metadata is tolerated only for fallback groups that do not compile the model.

If the cached model is missing, has no matching metadata, or is not optimal, the sample compiles the original model in a temporary directory with EP context embedded in the ONNX file. It verifies that no external sidecars were produced, and the temporary model must also report EP_SUPPORTED_OPTIMAL before it replaces the cache. If compilation or validation fails, temporary outputs are removed and inference uses the original SqueezeNet.onnx; an existing stale compiled model is never used for that run.