Allow onnx DML backend

[FNsi]

An accident caused last pull request closed....it's my fault.
This is only the 'control V' from the closed one #3045

[FNsi]

See detail #3045

[joeyballentine]

Thanks

[FNsi]

Thanks

😁

Should I say you're welcome?
But I'm only a User.

So ,
Thank you
Haha

[Kim2091]

@FNsi I'm on the latest nightly build which includes this. DML is not showing up as an option with all default dependencies installed.

Manually installing onnxruntime-directml==1.17.1 in chaiNNer's python environment does not allow it to show in the settings menu either. TensorRT, CUDA, and CPU are the only options.

Uninstalling other ONNX related libraries and only installing directml allows it to appear, but of course it won't function like that.

[FNsi]

@Kim2091
Quick fix in my mind, add DML in default provider search.

In source:
backend/src/packages/chaiNNer_onnx/settings.py

Direct fix in installed location:

/chiNNer/resources/src/packages/chaiNNer_onnx/settings.py

def get_providers():
    providers = cast(List[str], ort.get_available_providers())

    default = providers[0]
    if "DmlExecutionProvider" in providers:
        default = "DmlExecutionProvider"
    elif "CUDAExecutionProvider" in providers:
        default = "CUDAExecutionProvider"
    elif "CPUExecutionProvider" in providers:
        default = "CPUExecutionProvider"

    return providers, default


def get_provider_label(identifier: str) -> str:
    label = identifier.replace("ExecutionProvider", "")
    if label.lower() == "tensorrt":
        label = "TensorRT"
    return label


execution_providers, default_provider = get_providers()

That should display the DML and cuda at same time. test needed since I don't have NV.

[Kim2091]

@FNsi Thank you for the suggestion. I'm unable to compile chaiNNer right now, but I did try the fix in the final install, but to no avail.

[Kim2091]

This may be an issue with ORT itself. I printed the available providers with the default chaiNNer ONNX packages + the ONNX DML package:

print(providers)
['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']

[FNsi]

print(providers)
['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']

Also I found out people in other repo compliant about that 😅😅, I think you are right.