diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py
index f00b214b1ef18..90a1ad2a658ab 100644
--- a/vllm/model_executor/models/qwen2_vl.py
+++ b/vllm/model_executor/models/qwen2_vl.py
@@ -1218,6 +1218,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         grid_thw = image_input["image_grid_thw"]
         assert grid_thw.ndim == 2
+        grid_thw_list = grid_thw.tolist()
 
         if image_input["type"] == "image_embeds":
             image_embeds = image_input["image_embeds"]
@@ -1227,15 +1228,17 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         # Split concatenated embeddings for each image item.
         merge_size = self.visual.spatial_merge_size
-        sizes = grid_thw.prod(-1) // merge_size // merge_size
+        sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
+                 (merge_size * merge_size)).tolist()
 
-        return image_embeds.split(sizes.tolist())
+        return image_embeds.split(sizes)
 
     def _process_video_input(
             self, video_input: Qwen2VLVideoInputs) -> tuple[torch.Tensor, ...]:
 
         grid_thw = video_input["video_grid_thw"]
         assert grid_thw.ndim == 2
+        grid_thw_list = grid_thw.tolist()
 
         if video_input["type"] == "video_embeds":
             video_embeds = video_input["video_embeds"]
@@ -1245,9 +1248,10 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
 
         # Split concatenated embeddings for each video item.
         merge_size = self.visual.spatial_merge_size
-        sizes = grid_thw.prod(-1) // merge_size // merge_size
+        sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
+                 (merge_size * merge_size)).tolist()
 
-        return video_embeds.split(sizes.tolist())
+        return video_embeds.split(sizes)
 
     def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
         modalities = {}