[Security] Cap GLMGA video sampling to prevent request-driven resource exhaustion (#54935)

Signed-off-by: jperezde <jperezde@redhat.com>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>

tests/multimodal/test_video.py

@@ -22,6 +22,7 @@ from vllm.multimodal.video import (
     DynamicVideoBackend,
     Glm5NextVideoBackend,
     GLM46VVideoBackend,
+    GLMGAVideoBackend,
     Molmo2VideoBackend,
     Qwen2VLVideoBackend,
     Qwen3VLVideoBackend,
@@ -1473,6 +1474,59 @@ def test_glm46v_duration_estimation_from_fps():
     assert all(0 <= idx < 90 for idx in indices)
 
 
+class TestGLMGASamplingCaps:
+    """Regression tests for GHSA-58v5-2m8f-94pr: request-controlled fps
+    and max_frames must not create intermediate allocations larger than
+    the actual frame count."""
+
+    @staticmethod
+    def _source(
+        total_frames: int, fps: float = 30.0, duration: float = 0
+    ) -> VideoSourceMetadata:
+        if duration == 0 and fps > 0 and total_frames > 1:
+            duration = round((total_frames - 1) / fps) + 1
+        return VideoSourceMetadata(total_frames, fps, duration)
+
+    def test_extreme_values_bounded_by_total_frames(self):
+        source = self._source(total_frames=2, fps=2.0, duration=1.0)
+        target = VideoTargetMetadata(num_frames=-1, fps=500_000, max_duration=-1)
+        indices = GLMGAVideoBackend.compute_frames_index_to_sample(
+            source,
+            target,
+            max_frames=500_000,
+        )
+        assert len(indices) <= 2
+
+    def test_class_cap_overrides_kwargs_max_frames(self):
+        source = self._source(total_frames=10_000, fps=30.0)
+        target = VideoTargetMetadata(num_frames=-1, fps=30, max_duration=-1)
+        indices = GLMGAVideoBackend.compute_frames_index_to_sample(
+            source,
+            target,
+            max_frames=100_000,
+        )
+        assert len(indices) <= GLMGAVideoBackend._MAX_FRAMES
+
+    def test_class_cap_overrides_target_fps(self):
+        source = self._source(total_frames=10_000, fps=30.0)
+        target = VideoTargetMetadata(num_frames=-1, fps=500_000, max_duration=-1)
+        indices = GLMGAVideoBackend.compute_frames_index_to_sample(
+            source,
+            target,
+        )
+        assert len(indices) <= GLMGAVideoBackend._MAX_FRAMES
+
+    def test_normal_operation_unchanged(self):
+        source = self._source(total_frames=1000, fps=30.0, duration=33.0)
+        target = VideoTargetMetadata(num_frames=-1, fps=2, max_duration=-1)
+        indices = GLMGAVideoBackend.compute_frames_index_to_sample(
+            source,
+            target,
+        )
+        assert 0 < len(indices) <= GLMGAVideoBackend._MAX_FRAMES
+        assert all(0 <= idx < 1000 for idx in indices)
+
+
 def test_glm5next_backend_selected_for_processor():
     """Glm5NextVideoProcessor maps to the glm5next loader so only the
     sampled frames are decoded instead of the whole container. Both the

vllm/multimodal/video.py

@@ -781,6 +781,9 @@ class Glm5NextVideoBackend(VideoBackend):
     video_processor="GlmgaVideoProcessor",
 )
 class GLMGAVideoBackend(VideoBackend):
+    _MAX_FRAMES: ClassVar[int] = 640
+    _MAX_FPS: ClassVar[int] = 30
+
     @classmethod
     def _prepare_source(cls, source: VideoSourceMetadata) -> VideoSourceMetadata:
         # Estimate duration from frame count and fps when the container
@@ -806,14 +809,14 @@ class GLMGAVideoBackend(VideoBackend):
         total_frames_num = source.total_frames_num
         duration = source.duration
         original_fps = source.original_fps
-        target_fps = target.fps
+        target_fps = min(target.fps, cls._MAX_FPS)
         max_frame_idx = source.total_frames_num - 1
-        max_frames = kwargs.get("max_frames", 640)
+        max_frames = min(kwargs.get("max_frames", cls._MAX_FRAMES), cls._MAX_FRAMES)
 
         duration = duration or round(max_frame_idx / original_fps) + 1
 
         extract_t = int(duration * target_fps)
-        extract_t = min(extract_t, max_frames)
+        extract_t = min(extract_t, max_frames, total_frames_num)
 
         duration_per_frame = 1 / original_fps