[BUG] Phi 4 Mini GPU does not run on the GPU

[nmetulev]

Describe the bug
When I select Phi 4 Mini GPU under Custom models, the model runs on the CPU. Phi 3 Mini DirectML and Mistral 7B DirectML both work fine on the GPU

To Reproduce
Steps to reproduce the behavior:

1. Go to Generate Text sample
2. Select Phi 4 Mini GPU
3. Run sample
4. See memory, cpu, and gpu usage in Task manager

Expected behavior
The model should run on the GPU

Please complete the following information:

• App Version [e.g. 0.4.1 from the Microsoft Store]: 0.5.4 from the store

Additional context
My GPU is RTX 3070 with 8GB of VRAM

[luhan2017]

It works after I added the dml config in genai_config.json

"provider_options": [{
"dml": {}
}]

[haoliuu]

Looks like the provider_options for Phi 4 Mini is empty compared with the Phi 3 Mini DirectML and Mistral 7B DirectML.

One possible solution is to set up a local map between HardwareAccelerator and provider for GenAI, which would allow us to override the provider when the model's provider_options is null. However, it seems better to ensure the model includes the provider information it supports.

[weiyuanyue]

Root Cause: Incorrect configuration in the HuggingFace model repository.

Evidence 1: Configuration Removed in HuggingFace

Commit: 4f2996c5 (Sept 27, 2025)
Message: "Clear provider options for GPU"
Change: genai_config.json size reduced from 1663 to 1568 bytes (-95 bytes)

Evidence 2: Detection Logic

UserAddedModelUtil.cs iterates through provider_options array to detect execution providers:

var allProviderOptions = GetAllProviderOptions(config);
foreach (var provider in allProviderOptions) {
    if (provider.HasProvider("dml")) {
        return HardwareAccelerator.DML;  // GPU
    }
}
return HardwareAccelerator.CPU;  // Default if no provider found

When provider_options in genai_config.json is empty or missing, we allow the runtime to fall back to using the HardwareAccelerator metadata from the .modelgroup.json configuration file. This locally defined metadata could serve as a fallback source for determining the appropriate execution provider for the model runtime.