diff --git a/python/sglang/srt/layers/rotary_embedding.py b/python/sglang/srt/layers/rotary_embedding.py
index 55c3121ba..a18f0bff9 100644
--- a/python/sglang/srt/layers/rotary_embedding.py
+++ b/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
diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index c034c37b9..c521b1112 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -25,7 +25,6 @@ import signal
 import sys
 import threading
 import time
-import uuid
 from collections import deque
 from contextlib import nullcontext
 from datetime import datetime
@@ -360,7 +359,8 @@ class TokenizerManager(TokenizerCommunicatorMixin):
                 (
                     FreezeGCReq,
                     lambda x: None,
-                ),  # For handling case when scheduler skips detokenizer and forwards back to the tokenizer manager, we ignore it.
+                ),
+                # For handling case when scheduler skips detokenizer and forwards back to the tokenizer manager, we ignore it.
                 (HealthCheckOutput, lambda x: None),
             ]
         )
@@ -587,9 +587,9 @@ class TokenizerManager(TokenizerCommunicatorMixin):
             )
 
         if self.mm_processor and obj.contains_mm_input():
-            if not isinstance(obj.image_data, list):
+            if not isinstance(obj.image_data, list) and obj.image_data:
                 obj.image_data = [obj.image_data]
-            if not isinstance(obj.audio_data, list):
+            if not isinstance(obj.audio_data, list) and obj.audio_data:
                 obj.audio_data = [obj.audio_data]
             mm_inputs: Dict = await self.mm_processor.process_mm_data_async(
                 image_data=obj.image_data,
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 3977ad01e..d59378a72 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -196,7 +196,6 @@ MAMBA_CACHE_SIZE_MAX_RUNNING_REQUESTS_RATIO = 3
 
 logger = logging.getLogger(__name__)
 
-
 if _is_npu:
     import torch_npu
 
@@ -636,6 +635,22 @@ 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", None
+                )
+            ) 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.")
 
diff --git a/test/srt/test_vision_openai_server_a.py b/test/srt/test_vision_openai_server_a.py
index e8e0d62e9..b6861b99c 100644
--- a/test/srt/test_vision_openai_server_a.py
+++ b/test/srt/test_vision_openai_server_a.py
@@ -50,6 +50,27 @@ class TestQwen2VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
         cls.base_url += "/v1"
 
 
+class TestQwen3VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-VL-30B-A3B-Instruct"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            api_key=cls.api_key,
+            other_args=[
+                "--mem-fraction-static",
+                "0.80",
+                "--cuda-graph-max-bs",
+                "4",
+            ],
+        )
+        cls.base_url += "/v1"
+
+
 class TestQwen2_5_VLServer(ImageOpenAITestMixin, VideoOpenAITestMixin):
     @classmethod
     def setUpClass(cls):
diff --git a/test/srt/test_vision_openai_server_common.py b/test/srt/test_vision_openai_server_common.py
index 792636060..6af8f099c 100644
--- a/test/srt/test_vision_openai_server_common.py
+++ b/test/srt/test_vision_openai_server_common.py
@@ -494,7 +494,7 @@ class VideoOpenAITestMixin(TestOpenAIOmniServerBase):
             **(self.get_vision_request_kwargs()),
         )
 
-        video_response = response.choices[0].message.content
+        video_response = response.choices[0].message.content.lower()
 
         print("-" * 30)
         print(f"Video response:\n{video_response}")
@@ -502,9 +502,10 @@ class VideoOpenAITestMixin(TestOpenAIOmniServerBase):
 
         # Add assertions to validate the video response
         assert (
-            "iPod" in video_response
+            "ipod" in video_response
             or "device" in video_response
             or "microphone" in video_response
+            or "phone" in video_response
         ), f"video_response: {video_response}, should contain 'iPod' or 'device'"
         assert (
             "man" in video_response