Fix: Align model and input data types to float16 for inference

## Description

This PR fixes a runtime error in the example code caused by a mismatch between the input tensor type and the model weights during inference.

## Changes

- Converted the model to float16:
model = model.to("cuda").half()

- Converted float32 input tensors from the processor to float16:
inputs = {k: v.half() if isinstance(v, torch.Tensor) and v.dtype == torch.float32 else v for k, v in inputs.items()}

## Testing

The code has been successfully tested and runs without error.

## Note

This contribution is part of an ongoing research initiative to systematically identify and correct faulty example code in Hugging Face Model Cards.
We would appreciate a timely review and integration of this patch to support code reliability and enhance reproducibility for downstream users.

README.md

@@ -62,7 +62,9 @@ input_points = [[[450, 600]]] # 2D localization of a window
 
 
 ```python
+model = model.to("cuda").half()
 inputs = processor(raw_image, input_points=input_points, return_tensors="pt").to("cuda")
+inputs = {k: v.half() if isinstance(v, torch.Tensor) and v.dtype == torch.float32 else v for k, v in inputs.items()}
 outputs = model(**inputs)
 masks = processor.image_processor.post_process_masks(outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu())
 scores = outputs.iou_scores