From 356c11d5d90b7c62c073ef9cae1b35fea1a4e4ed Mon Sep 17 00:00:00 2001
From: longxin9715 <59550463+longxin9715@users.noreply.github.com>
Date: Tue, 28 Jul 2026 09:24:40 +0800
Subject: [PATCH] [Fix] --mm-process-config crash when video config contains
 (#30260)

---
 .../npu/modules/qwen_vl_processor.py          |  2 +-
 .../multimodal/processors/base_processor.py   | 10 +++-
 .../srt/multimodal/processors/qwen_vl.py      | 43 ++++++++++---
 .../unit/managers/test_mm_process_config.py   | 60 +++++++++++++++++++
 4 files changed, 105 insertions(+), 10 deletions(-)

diff --git a/python/sglang/srt/hardware_backend/npu/modules/qwen_vl_processor.py b/python/sglang/srt/hardware_backend/npu/modules/qwen_vl_processor.py
index 5dadafac6..dfbbbf927 100644
--- a/python/sglang/srt/hardware_backend/npu/modules/qwen_vl_processor.py
+++ b/python/sglang/srt/hardware_backend/npu/modules/qwen_vl_processor.py
@@ -29,7 +29,7 @@ def transform_patches_to_flatten(
     patch_size: int,
     merge_size: int,
 ) -> torch.Tensor:
-    patches = patches.view(
+    patches = patches.reshape(
         batch_size * grid_t,
         temporal_patch_size * channel,
         grid_h // merge_size,
diff --git a/python/sglang/srt/multimodal/processors/base_processor.py b/python/sglang/srt/multimodal/processors/base_processor.py
index 65b0f4151..dafa4889c 100644
--- a/python/sglang/srt/multimodal/processors/base_processor.py
+++ b/python/sglang/srt/multimodal/processors/base_processor.py
@@ -489,6 +489,7 @@ class BaseMultimodalProcessor(ABC):
         videos=None,
         audios=None,
         processor=None,
+        processor_video_config: Optional[Dict[str, Any]] = None,
         **kwargs,
     ) -> dict:
         """
@@ -502,8 +503,13 @@ class BaseMultimodalProcessor(ABC):
                 kwargs.setdefault("images_kwargs", {}).update(self.image_config)
         if videos:
             kwargs["videos"] = videos
-            if self.video_config:
-                kwargs.setdefault("videos_kwargs", {}).update(self.video_config)
+            video_config = (
+                self.video_config
+                if processor_video_config is None
+                else processor_video_config
+            )
+            if video_config:
+                kwargs.setdefault("videos_kwargs", {}).update(video_config)
         if audios:
             if processor.__class__.__name__ in {
                 "Gemma3nProcessor",
diff --git a/python/sglang/srt/multimodal/processors/qwen_vl.py b/python/sglang/srt/multimodal/processors/qwen_vl.py
index c68f31529..a53e93730 100644
--- a/python/sglang/srt/multimodal/processors/qwen_vl.py
+++ b/python/sglang/srt/multimodal/processors/qwen_vl.py
@@ -59,6 +59,30 @@ FPS = 2.0
 FPS_MIN_FRAMES = 4
 FPS_MAX_FRAMES = 768
 
+QWEN_VIDEO_PREPROCESS_CONFIG_KEYS = frozenset(
+    {
+        "fps",
+        "nframes",
+        "min_frames",
+        "max_frames",
+        "min_pixels",
+        "max_pixels",
+        "total_pixels",
+        "resized_height",
+        "resized_width",
+    }
+)
+
+
+def _get_processor_video_config(video_config, video_metadata):
+    if video_metadata and all(metadata is not None for metadata in video_metadata):
+        return {
+            key: value
+            for key, value in video_config.items()
+            if key not in QWEN_VIDEO_PREPROCESS_CONFIG_KEYS
+        }
+    return None
+
 
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
@@ -719,6 +743,13 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
 
         preprocess_time = time.perf_counter()
 
+        processor_kwargs = {}
+        processor_video_config = _get_processor_video_config(
+            self.video_config, video_metadata
+        )
+        if processor_video_config is not None:
+            processor_kwargs["processor_video_config"] = processor_video_config
+
         # NOTE: for qwen3-vl, video_meta need to be passed in, since do_sample_frames is already done in preprocess_video
         if self.hf_config.model_type in (
             "qwen3_vl",
@@ -727,16 +758,14 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
             "qwen3_5_moe",
             "intern_s2_preview",
         ):
-            mm_items, input_ids, ret = await self.process_and_combine_mm_data_async(
-                base_output,
-                self.mm_tokens,
+            processor_kwargs.update(
                 video_metadata=video_metadata,
                 do_sample_frames=False,
             )
-        else:
-            mm_items, input_ids, ret = await self.process_and_combine_mm_data_async(
-                base_output, self.mm_tokens
-            )
+
+        mm_items, input_ids, ret = await self.process_and_combine_mm_data_async(
+            base_output, self.mm_tokens, **processor_kwargs
+        )
 
         audio_feature_lengths = None
 
diff --git a/test/registered/unit/managers/test_mm_process_config.py b/test/registered/unit/managers/test_mm_process_config.py
index 3d4495df8..63970ee7d 100644
--- a/test/registered/unit/managers/test_mm_process_config.py
+++ b/test/registered/unit/managers/test_mm_process_config.py
@@ -395,6 +395,37 @@ class TestProcessMmDataKwargs(unittest.TestCase):
             call_kwargs.kwargs.get("videos_kwargs"), {"fps": 3, "max_frames": 60}
         )
 
+    def test_preprocessed_video_config_is_filtered_before_single_call(self):
+        config = {
+            "video": {
+                "fps": 3,
+                "max_frames": 60,
+                "do_normalize": False,
+            }
+        }
+        proc, mock_proc, _ = self._make_base_processor(config)
+
+        proc.process_mm_data(
+            "test",
+            videos=["vid1"],
+            processor_video_config={"do_normalize": False},
+        )
+
+        self.assertEqual(mock_proc.__call__.call_count, 1)
+        self.assertEqual(
+            mock_proc.__call__.call_args.kwargs.get("videos_kwargs"),
+            {"do_normalize": False},
+        )
+
+    def test_processor_error_is_not_retried(self):
+        proc, mock_proc, _ = self._make_base_processor({"video": {"max_frames": 60}})
+        mock_proc.__call__.side_effect = ValueError("processor failure")
+
+        with self.assertRaisesRegex(ValueError, "processor failure"):
+            proc.process_mm_data("test", videos=["vid1"])
+
+        self.assertEqual(mock_proc.__call__.call_count, 1)
+
     def test_no_collision_with_overlapping_keys(self):
         """Core test: image and video both have max_pixels but stay separate."""
         config = {
@@ -527,6 +558,35 @@ class TestOverrideProcessorsConfigInjection(unittest.TestCase):
         self.assertTrue(audio_kw.get("truncation"))
 
 
+class TestQwenVideoConfigRouting(unittest.TestCase):
+    def test_preprocessed_video_drops_sglang_owned_config(self):
+        from sglang.srt.multimodal.processors.qwen_vl import (
+            _get_processor_video_config,
+        )
+
+        video_config = {
+            "fps": 3,
+            "nframes": 12,
+            "max_frames": 60,
+            "max_pixels": 500000,
+            "do_normalize": False,
+        }
+
+        processor_config = _get_processor_video_config(video_config, [{"fps": 30.0}])
+
+        self.assertEqual(processor_config, {"do_normalize": False})
+
+    def test_unprocessed_video_uses_original_config(self):
+        from sglang.srt.multimodal.processors.qwen_vl import (
+            _get_processor_video_config,
+        )
+
+        video_config = {"fps": 3, "max_frames": 60}
+
+        self.assertIsNone(_get_processor_video_config(video_config, None))
+        self.assertIsNone(_get_processor_video_config(video_config, [None]))
+
+
 class TestDoubleBosGuard(unittest.TestCase):
     """Regression test for the multimodal double-BOS bug.