[Multimodal] Improve max video embedding length estimation in V1 (#24312)

Signed-off-by: Roger Wang <hey@rogerw.me>
Co-authored-by: Roger Wang <hey@rogerw.me>

vllm/model_executor/models/llava_onevision.py

@@ -216,12 +216,9 @@ class LlavaOnevisionProcessingInfo(LlavaNextProcessingInfo):
         seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> int:
-        max_images = mm_counts.get("image", 0)
         max_videos = mm_counts.get("video", 0)
 
-        max_image_tokens = self.get_max_image_tokens() * max_images
-        max_total_frames = self._get_max_video_frames(seq_len -
-                                                      max_image_tokens)
+        max_total_frames = self._get_max_video_frames(seq_len)
         max_frames_per_video = min(max_total_frames // max(max_videos, 1),
                                    _MAX_FRAMES_PER_VIDEO)
 

vllm/model_executor/models/qwen2_vl.py

@@ -915,12 +915,9 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
         seq_len: int,
         mm_counts: Mapping[str, int],
     ) -> int:
-        max_images = mm_counts.get("image", 0)
         max_videos = mm_counts.get("video", 0)
 
-        max_image_tokens = self.get_max_image_tokens() * max_images
-        max_total_frames = self._get_max_video_frames(seq_len -
-                                                      max_image_tokens)
+        max_total_frames = self._get_max_video_frames(seq_len)
         max_frames_per_video = min(max_total_frames // max(max_videos, 1),
                                    _MAX_FRAMES_PER_VIDEO)