fix: fix video input for qwen3-vl (#11361)

python/sglang/srt/layers/rotary_embedding.py

@@ -1126,6 +1126,11 @@ class MRotaryEmbedding(RotaryEmbedding):
         second_per_grid_ts: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if model_type.startswith("qwen3_vl") 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

python/sglang/srt/model_executor/model_runner.py

@@ -186,7 +186,6 @@ UNBALANCED_MODEL_LOADING_TIMEOUT_S = 300
 
 logger = logging.getLogger(__name__)
 
-
 if _is_npu:
     import torch_npu
 
@@ -625,6 +624,22 @@ class ModelRunner:
                     "Setting hicache_io_backend to vanilla I/O, which may lead to suboptimal performance with small page sizes."
                 )
 
+        if self.model_config.hf_config.model_type == "qwen3_vl_moe":
+            if (
+                quantization_config := getattr(
+                    self.model_config.hf_config, "quantization_config"
+                )
+            ) is not None:
+                text_config = self.model_config.hf_text_config
+                weight_block_size_n = quantization_config["weight_block_size"][0]
+                if (
+                    text_config.moe_intermediate_size
+                    // (self.tp_size // self.moe_ep_size)
+                ) % weight_block_size_n != 0:
+                    raise ValueError(
+                        f"For qwen3-vl-fp8 models, please make sure ({text_config.moe_intermediate_size=} // ({self.tp_size=} // {self.moe_ep_size=})) % {weight_block_size_n=} == 0"
+                    )
+
     def init_torch_distributed(self):
         logger.info("Init torch distributed begin.")