[VLM] Optimize GLM4.5-V-style video processing to only decode necessary frames (#24161)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

tests/multimodal/test_utils.py

@@ -204,6 +204,32 @@ async def test_fetch_video_http(video_url: str, num_frames: int):
     assert metadata_sync == metadata_async
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("video_url", TEST_VIDEO_URLS)
+@pytest.mark.parametrize("max_duration", [1, 60, 1800])
+@pytest.mark.parametrize("requested_fps", [2, 24])
+async def test_fetch_video_http_with_dynamic_loader(
+        video_url: str, max_duration: int, requested_fps: int,
+        monkeypatch: pytest.MonkeyPatch):
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_VIDEO_LOADER_BACKEND", "opencv_dynamic")
+        connector = MediaConnector(
+            media_io_kwargs={
+                "video": {
+                    "max_duration": max_duration,
+                    "requested_fps": requested_fps,
+                }
+            })
+
+        video_sync, metadata_sync = connector.fetch_video(video_url)
+        video_async, metadata_async = await connector.fetch_video_async(
+            video_url)
+
+        assert np.array_equal(video_sync, video_async)
+        assert metadata_sync == metadata_async
+        assert metadata_sync["video_backend"] == "opencv_dynamic"
+
+
 # Used for `test_argsort_mm_positions`.
 class TestCase(NamedTuple):
     mm_positions: "MultiModalPlaceholderDict"