fix: cast inputs to float16 to match model dtype

app.py

@@ -40,7 +40,7 @@ def predict(image):
 
     # Segmentation
     seg_inputs = seg_processor(images=image_resized, task_inputs=["semantic"], return_tensors="pt")
-    seg_inputs = {k: v.to(device) for k, v in seg_inputs.items()}
+    seg_inputs = {k: v.to(device, dtype=torch.float16) if v.is_floating_point() else v.to(device) for k, v in seg_inputs.items()}
 
     seg_outputs = seg_model(**seg_inputs)
     seg_result = seg_processor.post_process_semantic_segmentation(
@@ -56,7 +56,7 @@ def predict(image):
 
     # Depth estimation
     depth_inputs = depth_processor(images=image_resized, return_tensors="pt")
-    depth_inputs = {k: v.to(device) for k, v in depth_inputs.items()}
+    depth_inputs = {k: v.to(device, dtype=torch.float16) if v.is_floating_point() else v.to(device) for k, v in depth_inputs.items()}
 
     depth_outputs = depth_model(**depth_inputs)
     depth_map = depth_outputs.predicted_depth.squeeze().cpu().numpy()