AIFoundryModel strongly-typed model catalog fails on local, GPU-enhanced models.

[karolz-ms]

Is there an existing issue for this?

• I have searched the existing issues

Describe the bug

Seems like the strongly-typed model catalog does not handle CPU/GPU/NPU model variations when using Foundry Local.

Expected Behavior

No response

Steps To Reproduce

(the following steps need to be run on a machine with a GPU supported by Foundry Local, e.g. nVidia RTX series)

1. Create an Aspire starter app and add reference to Aspire.Hosting.Azure.AIFoundry nuget package to application host.
2. Install Foundry Local Aspire.Hosting.Azure.AIFoundry
3. Download phi-4-mini model, e.g. using the CLI: foundry model download phi-4-mini
4. Add the following to the app host code

var localFoundry = builder.AddAzureAIFoundry("foundry")
    .RunAsFoundryLocal();

_ = localFoundry.AddDeployment("chat", "phi-4-mini", "1", "Microsoft"); // this works
_ = localFoundry.AddDeployment("chat2", AIFoundryModel.Microsoft.Phi4MiniInstruct); // this does not

5. Run the app

Expected

Both "model deployments" should work

Actual

chat2 sub-resource fails to start with the error:

Failed to start Phi-4-mini-instruct. Error: Model 'Phi-4-mini-instruct' was not found in the catalogue

This is because neither the model ID, nor the model alias match what the strongly typed model catalog contains:

> foundry cache ls
Models cached on device:
   Alias                         Model ID
💾 phi-4-mini                    Phi-4-mini-instruct-cuda-gpu

Exceptions (if any)

No response

.NET Version info

Aspire version 9.5.0-preview.1.25468.4
Dotnet 10.0.0-rc.1.25451.107

Anything else?

No response

[MihuBot]

I'm a bot. Here is a possible related and/or duplicate issue (I may be wrong):

• AIFoundryModel.Microsoft.Phi35MiniInstruct fails to load #11288

[karolz-ms]

#11288 is almost a duplicate of this except that the fix (#11577) does not really enable me to use phi-4-mini model, which is quite useful with Foundry Local (significantly smaller than phi-4, which the Aspire strongly-typed catalog does support)

[karolz-ms]

It may be that we just lack support for phi-4-mini in the catalog.

[davidfowl]

cc @eerhardt @sebastienros

[sebastienros]

We won't be able to have something working reliably for foundry local until we generate a custom enum. This model is an example where the logical names on hosted and local are different. I already filed an issue ang got feedback from the team, they states they would provide a way to list all the model including the ones your local GPU doesn't support.