Foundry Local SDK: CUDA is being ignored on Windows

My RTX 5070 is basically a paperweight right now because foundry-local-sdk (v1.2.3) refuses to acknowledge it exists. Despite every configuration attempt, the SDK is forcing CPUExecutionProvider and crashing when I try to push it to the GPU.

The setup is Windows 11, Python 3.14 (venv), and I'm using the foundry-local-sdk-winml package. Every time I initialize the FoundryLocalManager with CUDA settings, I get hit with this:

System.InvalidOperationException: Cannot load model 'Phi-3.5-mini-instruct-cuda-gpu:2': 
it requires the 'CUDAExecutionProvider' execution provider, which is not available. 
Available EPs: [CPUExecutionProvider].

I've already tried the usual suspects for a deep dive into the environment:
1. Installed both onnxruntime-genai-cuda and onnxruntime-genai-directml.
2. Explicitly set the config to use the GPU:

config = Configuration(
    app_name="my-app",
    additional_settings={
        "execution_provider": "CUDAExecutionProvider",
        "device": "gpu"
    }
)

The model Phi-3.5-mini-instruct-cuda-gpu:2 is sitting right there in the cache, but the SDK just won't see the provider. I have a sneaking suspicion that foundry-local-sdk-winml is bundled with a runtime that's hardcoded for CPU/WinML, effectively making the Configuration object a piece of fiction.

Since I'm a data engineer who values my sanity (and security/compliance), I'm trying to figure out if there's a way to force the CUDA provider or if I'm just fighting a losing battle with the winml variant.

Is there a non-winml version of the package that actually plays nice with CUDA? Or is the only real-world fix to just grab the model path and run it through onnxruntime-genai directly, bypassing the SDK entirely?