[fix] fix qwen image_embeds input (#21049)

Signed-off-by: h-avsha <avshalom.manevich@hcompany.ai>

vllm/model_executor/models/qwen2_5_vl.py

@@ -974,7 +974,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
         grid_thw_list = grid_thw.tolist()
 
         if image_input["type"] == "image_embeds":
-            image_embeds = image_input["image_embeds"]
+            image_embeds = image_input["image_embeds"].type(self.visual.dtype)
         else:
             pixel_values = image_input["pixel_values"]
             image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list)
@@ -994,7 +994,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
         grid_thw_list = grid_thw.tolist()
 
         if video_input["type"] == "video_embeds":
-            video_embeds = video_input["video_embeds"]
+            video_embeds = video_input["video_embeds"].type(self.visual.dtype)
         else:
             pixel_values_videos = video_input["pixel_values_videos"]
             video_embeds = self.visual(pixel_values_videos,