fix(security): block request-level GPU video backend selection withou… (#47259)

Signed-off-by: jperezde <jperezde@redhat.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

Author: jperezdealgaba

tests/multimodal/media/test_video.py

@@ -16,7 +16,11 @@
     video_to_pil_images_list,
 )
 from vllm.multimodal.media import ImageMediaIO, VideoMediaIO
-from vllm.multimodal.video import VIDEO_LOADER_REGISTRY, VideoLoader
+from vllm.multimodal.video import (
+    PYNVVIDEOCODEC_VIDEO_BACKEND,
+    VIDEO_LOADER_REGISTRY,
+    VideoLoader,
+)
 
 from ..utils import cosine_similarity, create_video_from_image, normalize_image
 
@@ -357,3 +361,95 @@ def test_load_base64_jpeg_raises_on_zero_num_frames():
 
     with pytest.raises(ValueError, match="num_frames must be greater than 0 or -1"):
         videoio.load_base64("video/jpeg", data)
+
+
+# ---------------------------------------------------------------------------
+# GPU video backend policy tests
+# ---------------------------------------------------------------------------
+
+
+class TestMergeKwargsGpuBackendPolicy:
+    """Verify that merge_kwargs blocks request-level GPU backend selection
+    when the static (engine-level) config did not configure that backend."""
+
+    def test_pynvvideocodec_requires_gpu(self):
+        assert VIDEO_LOADER_REGISTRY.backend_requires_gpu(PYNVVIDEOCODEC_VIDEO_BACKEND)
+
+    def test_strips_video_backend_pynv_when_not_static(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs=None,
+            runtime_kwargs={"video_backend": "pynvvideocodec"},
+        )
+        assert "video_backend" not in result
+
+    def test_strips_backend_pynv_when_not_static(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs={"num_frames": 16},
+            runtime_kwargs={"backend": "pynvvideocodec"},
+        )
+        assert result.get("backend") != "pynvvideocodec"
+
+    def test_preserves_video_backend_pynv_when_static(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs={"video_backend": "pynvvideocodec"},
+            runtime_kwargs={"video_backend": "pynvvideocodec", "num_frames": 8},
+        )
+        assert result["video_backend"] == "pynvvideocodec"
+        assert result["num_frames"] == 8
+
+    def test_preserves_backend_pynv_when_static(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs={"backend": "pynvvideocodec"},
+            runtime_kwargs={"backend": "pynvvideocodec"},
+        )
+        assert result["backend"] == "pynvvideocodec"
+
+    @pytest.mark.parametrize("backend", ["opencv", "pyav", "torchcodec"])
+    def test_software_video_backend_passes_through(self, backend: str):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs=None,
+            runtime_kwargs={"video_backend": backend},
+        )
+        assert result["video_backend"] == backend
+
+    @pytest.mark.parametrize("backend", ["opencv", "pyav"])
+    def test_software_codec_backend_passes_through(self, backend: str):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs=None,
+            runtime_kwargs={"backend": backend},
+        )
+        assert result["backend"] == backend
+
+    def test_strips_both_keys_independently(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs=None,
+            runtime_kwargs={
+                "video_backend": "pynvvideocodec",
+                "backend": "pynvvideocodec",
+                "num_frames": 4,
+            },
+        )
+        assert "video_backend" not in result
+        assert result.get("backend") != "pynvvideocodec"
+        assert result["num_frames"] == 4
+
+    def test_other_kwargs_preserved_when_gpu_backend_stripped(self):
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs={"fps": 2},
+            runtime_kwargs={
+                "video_backend": "pynvvideocodec",
+                "num_frames": 16,
+            },
+        )
+        assert "video_backend" not in result
+        assert result["num_frames"] == 16
+
+    def test_static_pynv_with_different_runtime_gpu_backend(self):
+        """If static sets pynv via video_backend but runtime tries to set it
+        via the codec-level 'backend' key (without a static match), strip it."""
+        result = VideoMediaIO.merge_kwargs(
+            default_kwargs={"video_backend": "pynvvideocodec"},
+            runtime_kwargs={"backend": "pynvvideocodec"},
+        )
+        assert result.get("backend") != "pynvvideocodec"
+        assert result["video_backend"] == "pynvvideocodec"

vllm/multimodal/media/video.py

@@ -10,11 +10,14 @@
 from PIL import Image
 
 from vllm import envs
+from vllm.logger import init_logger
 
 from ..video import VIDEO_LOADER_REGISTRY
 from .base import MediaIO
 from .image import ImageMediaIO
 
+logger = init_logger(__name__)
+
 
 class VideoMediaIO(MediaIO[tuple[npt.NDArray, dict[str, Any]]]):
     """Configuration values can be user-provided either by --media-io-kwargs or
@@ -28,6 +31,24 @@ def merge_kwargs(
         default_kwargs: dict[str, Any] | None,
         runtime_kwargs: dict[str, Any] | None,
     ) -> dict[str, Any]:
+        if runtime_kwargs:
+            # Block request-level selection of GPU video backends that
+            # were not configured (and VRAM-reserved) at startup.
+            for key in ("video_backend", "backend"):
+                requested = runtime_kwargs.get(key)
+                if requested and VIDEO_LOADER_REGISTRY.backend_requires_gpu(requested):
+                    static_val = (default_kwargs or {}).get(key)
+                    if static_val != requested:
+                        logger.warning_once(
+                            "Stripping request-level %s=%r: GPU video "
+                            "backend not configured at startup.",
+                            key,
+                            requested,
+                        )
+                        runtime_kwargs = {
+                            k: v for k, v in runtime_kwargs.items() if k != key
+                        }
+
         merged = super().merge_kwargs(default_kwargs, runtime_kwargs)
         # fps and num_frames interact with each other, so if either is
         # overridden at request time, wipe the other from defaults to

vllm/multimodal/video.py

@@ -39,6 +39,7 @@ class VideoLoaderRegistry(ExtensionManager):
     def __init__(self) -> None:
         super().__init__()
         self.processor2backend: dict[str, str] = {}
+        self._requires_gpu: dict[str, bool] = {}
 
     @staticmethod
     def _normalize_registered_video_processors(
@@ -62,11 +63,13 @@ def register(
         name: str,
         *,
         video_processor: str | tuple[str, ...] | None = None,
+        requires_gpu: bool = False,
     ):
         processors = self._normalize_registered_video_processors(video_processor)
 
         def wrap(cls_to_register):
             self.name2class[name] = cls_to_register
+            self._requires_gpu[name] = requires_gpu
             for processor_name in processors:
                 self.processor2backend[processor_name] = name
             return cls_to_register
@@ -82,6 +85,9 @@ def get_backend_for_video_processor(
 
         return self.processor2backend.get(video_processor)
 
+    def backend_requires_gpu(self, name: str) -> bool:
+        return self._requires_gpu.get(name, False)
+
 
 def get_video_loader_backend_for_processor(
     video_processor: str | None,
@@ -909,7 +915,7 @@ def load_bytes(
         )
 
 
-@VIDEO_LOADER_REGISTRY.register(PYNVVIDEOCODEC_VIDEO_BACKEND)
+@VIDEO_LOADER_REGISTRY.register(PYNVVIDEOCODEC_VIDEO_BACKEND, requires_gpu=True)
 class PyNvVideoCodecVideoBackend(VideoBackend):
     """Hardware-accelerated video backend using PyNvVideoCodec.
 

vllm/v1/worker/gpu_worker.py

@@ -55,7 +55,7 @@
     PYNVVIDEOCODEC_CUDA_CONTEXT_BYTES,
     PYNVVIDEOCODEC_DECODER_GPU_MEMORY_BYTES,
     PYNVVIDEOCODEC_MAX_RETAINED_DECODERS,
-    PYNVVIDEOCODEC_VIDEO_BACKEND,
+    VIDEO_LOADER_REGISTRY,
 )
 from vllm.platforms import current_platform
 from vllm.profiler.wrapper import CudaProfilerWrapper, TorchProfilerWrapper
@@ -584,15 +584,15 @@ def determine_available_memory(self) -> int:
         )
 
     @staticmethod
-    def _uses_pynvvideocodec_video_backend(mm_config) -> bool:
+    def _uses_gpu_video_backend(mm_config) -> bool:
         video_kwargs = mm_config.media_io_kwargs.get("video", {})
         video_loader_backend = (
             video_kwargs.get("video_backend") or envs.VLLM_VIDEO_LOADER_BACKEND
         )
         codec_backend = video_kwargs.get("backend")
-        return (
-            video_loader_backend == PYNVVIDEOCODEC_VIDEO_BACKEND
-            or codec_backend == PYNVVIDEOCODEC_VIDEO_BACKEND
+        return VIDEO_LOADER_REGISTRY.backend_requires_gpu(video_loader_backend) or (
+            codec_backend is not None
+            and VIDEO_LOADER_REGISTRY.backend_requires_gpu(codec_backend)
         )
 
     def _reserve_mm_ipc_gpu_memory(self, available_kv_cache_memory_bytes: int) -> int:
@@ -623,7 +623,7 @@ def _reserve_mm_ipc_gpu_memory(self, available_kv_cache_memory_bytes: int) -> in
         )
         decoder_reserved_bytes = (
             num_api_servers * per_server_decoder_bytes
-            if self._uses_pynvvideocodec_video_backend(mm_config)
+            if self._uses_gpu_video_backend(mm_config)
             else 0
         )
         reserved_bytes = raw_frame_reserved_bytes + decoder_reserved_bytes