fix: fix video input for qwen3-vl (#11442)

python/sglang/srt/layers/rotary_embedding.py

@@ -1142,6 +1142,13 @@ class MRotaryEmbedding(RotaryEmbedding):
         second_per_grid_ts: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if (
+            model_type.startswith("qwen3_vl") or model_type.startswith("qwen3_vl_moe")
+        ) and video_grid_thw is not None:
+            video_grid_thw = torch.repeat_interleave(
+                video_grid_thw, video_grid_thw[:, 0], dim=0
+            )
+            video_grid_thw[:, 0] = 1
         mrope_position_deltas = []
         if input_ids is not None and (
             image_grid_thw is not None or video_grid_thw is not None