fix bounding box detection not accounting for batch dimension

modeling_moondream3.py

@@ -1147,12 +1147,15 @@ class Moondream3ForConditionalGeneration(Moondream3PreTrainedModel, GenerationMi
                         torch.stack([w, h],dim=-1).to(size_logits.dtype)
                     )
                 ).unsqueeze(1)
+                x_center = x_center.squeeze(1)
+                y_center = y_center.squeeze(1)
                 bbox = [
-                    x_center.item() - w.item() / 2,
-                    y_center.item() - h.item() / 2,
-                    x_center.item() + w.item() / 2,
-                    y_center.item() + h.item() / 2,
+                    x_center - w / 2,
+                    y_center - h / 2,
+                    x_center + w / 2,
+                    y_center + h / 2,
                 ]
+                bbox = torch.stack(bbox, dim=1)
                 bbox = bbox * (batch_mask).unsqueeze(1)
                 pos_ids += 1
                 cache_pos = cache_pos + 1