Supporting YOLO models: aten::_reinterpret_tensor not implemented

[oluwatimilehin]

Hi team, thanks for the fantastic work on PyTorchSIM.

Just flagging here that I've just attempted to run inference on a YOLO model on PyTorchSIM, and it failed with the error:

NotImplementedError: Could not run 'aten::_reinterpret_tensor' with arguments from the 'npu' backend. This could be because the operator doesn't exist for this backend, or was omitted during the selective/custom build process (if using custom build). 

This operator is registered here:

PyTorchSim/PyTorchSimFrontend/extension_device.cpp

Line 295 in eaf6da3

TORCH_LIBRARY_FRAGMENT(aten, m) {

but it's not picked up during inference.

Here's my config:

def run_yolo(batch, config):
    device = PyTorchSimRunner.setup_device().custom_device()
    
    model = torch.hub.load("ultralytics/yolov5", "yolov5s")
    url = "https://ultralytics.com/images/zidane.jpg"

    response = requests.get(url)
    img = Image.open(BytesIO(response.content)).convert("RGB")

    transform = transforms.Compose([
        transforms.Resize((640, 640)),
        transforms.ToTensor(),
    ])

    x = transform(img).unsqueeze(0)  
   
    # Move model and input tensors to the custom device
    x = x.to(device)   
    model.to(device)

    # Compile and run the model with PyTorchSim
    compiled_model = torch.compile(dynamic=False)(model)
    y = compiled_model(x)

[oluwatimilehin]

I fixed this error by modifying the declaration in extension_device.cpp to:

TORCH_LIBRARY_FRAGMENT(aten, m) {
m.def(
  "_reinterpret_tensor(Tensor self, int[] size, int[] stride, int offset_increment=0) -> Tensor",
  {at::Tag::pt2_compliant_tag});
}

TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
  m.impl("_reinterpret_tensor", _reinterpret_tensor);
}

TORCH_LIBRARY_IMPL(aten, AutogradPrivateUse1, m) {
  m.impl("_reinterpret_tensor", _reinterpret_tensor);
}

I'm encountering other errors when scheduling a YOLOv5s inference task, but I'll close this issue and upstream my changes if/when I can run it successfully.