Update image_processing_gemma3_tiled.py

image_processing_gemma3_tiled.py

@@ -172,6 +172,7 @@ class Gemma3TiledImageProcessor(BaseImageProcessor):
     """
     
     model_input_names = ["pixel_values", "tile_grid_shape", "num_crops"]
+    _auto_class = "AutoImageProcessor"  # Required for auto_map in preprocessor_config.json
     
     def __init__(
         self,
@@ -291,15 +292,20 @@ class Gemma3TiledImageProcessor(BaseImageProcessor):
         # num_crops is 0 for each image since we use tiling, not pan-and-scan
         num_crops = [0] * len(all_pixel_values)
         
+        # Concatenate all tiles into a single array for vLLM compatibility
+        # vLLM's flat_from_sizes expects a single tensor, not a list
+        if len(all_pixel_values) > 0:
+            concatenated_pixels = np.concatenate(all_pixel_values, axis=0)
+        else:
+            concatenated_pixels = np.array([])
+        
         data = {
-            "pixel_values": all_pixel_values,
+            "pixel_values": concatenated_pixels,
             "tile_grid_shape": all_grid_shapes,
             "num_crops": num_crops,
         }
         
-        # Don't convert to tensors here - pixel_values have inhomogeneous shapes
-        # (different images have different tile counts). Let the model handle it.
-        return BatchFeature(data=data, tensor_type=None)
+        return BatchFeature(data=data, tensor_type=return_tensors)
 
 
 __all__ = ["Gemma3TiledImageProcessor", "calculate_tile_grid", "tile_image"]