diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py
index 28c62e59bd1..9b8c42713f8 100644
--- a/vllm/model_executor/models/qwen3_vl.py
+++ b/vllm/model_executor/models/qwen3_vl.py
@@ -2269,6 +2269,8 @@ class Qwen3VLForConditionalGeneration(
         input_embeds for the LLM.
         """
 
+        device = video_embeddings.device
+
         # Generate video replacement token IDs using get_video_repl
         # This tokenizes each frame separator independently, then uses pre-tokenized
         # special tokens to ensure consistent tokenization regardless of
@@ -2283,10 +2285,8 @@ class Qwen3VLForConditionalGeneration(
             select_token_id=self.is_multimodal_pruning_enabled,
         )
 
-        repl_token_ids = torch.tensor(video_repl.full)
-        embed_token_id = _cached_tensor(
-            self.config.video_token_id, repl_token_ids.device
-        )
+        repl_token_ids = torch.tensor(video_repl.full, device=device)
+        embed_token_id = _cached_tensor(self.config.video_token_id, device=device)
         is_video_embed = torch.isin(repl_token_ids, embed_token_id)
 
         # Get text embeddings for indicator tokens (has only `visual_dim``).