Fix EVS crash when using `video_embeds` inputs in Qwen2.5-VL (#29232)

Signed-off-by: zitian.zhao <zitian.zhao@tencentmusic.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/model_executor/models/qwen2_5_vl.py

@@ -230,6 +230,9 @@ class Qwen2_5_VLVideoEmbeddingInputs(TensorSchema):
         - hidden_size must match the hidden size of language model backbone.
         - video_grid_thw shape: (num_videos, 3) in (grid_t, grid_h, grid_w)
           format
+        - second_per_grid_ts: The video time interval (in seconds) for each
+          grid along the temporal dimension in the 3D position IDs. Returned
+          when `videos` is not `None`.
     """
 
     type: Literal["video_embeds"]
@@ -244,6 +247,11 @@ class Qwen2_5_VLVideoEmbeddingInputs(TensorSchema):
         TensorShape("nv", 3),
     ]
 
+    second_per_grid_ts: Annotated[
+        torch.Tensor | None,
+        TensorShape("nv"),
+    ] = None
+
 
 Qwen2_5_VLVideoInputs: TypeAlias = (
     Qwen2_5_VLVideoPixelInputs | Qwen2_5_VLVideoEmbeddingInputs
@@ -1311,6 +1319,7 @@ class Qwen2_5_VLForConditionalGeneration(
                 type="video_embeds",
                 video_embeds=video_embeds,
                 video_grid_thw=video_grid_thw,
+                second_per_grid_ts=second_per_grid_ts,
             )
 
     def _process_image_input(
@@ -1422,7 +1431,13 @@ class Qwen2_5_VLForConditionalGeneration(
 
         # Cast to long to match the original code
         # https://github.com/huggingface/transformers/blob/41980ce93e775f6c88500c51c8db7946fc6a2add/src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py#L491 # noqa
-        second_per_grid_ts = video_input["second_per_grid_ts"].long()
+        second_per_grid_ts = video_input.get("second_per_grid_ts")
+        if second_per_grid_ts is None:
+            raise ValueError(
+                "second_per_grid_ts is required when video_pruning_rate > 0 "
+                "is enabled for video inputs, including the video_embeds path."
+            )
+        second_per_grid_ts = second_per_grid_ts.long()
         tokens_per_second = self.config.vision_config.tokens_per_second
 
         video_embeds_out = []