diff --git a/llama_index/llms/vllm.py b/llama_index/llms/vllm.py
index f7dce9a5f45f0..5233bd9d05ebc 100644
--- a/llama_index/llms/vllm.py
+++ b/llama_index/llms/vllm.py
@@ -186,6 +186,17 @@ def __init__(
             output_parser=output_parser,
         )
 
+    def __del__(self) -> None:
+        import torch
+        from vllm.model_executor.parallel_utils.parallel_state import (
+            destroy_model_parallel,
+        )
+
+        destroy_model_parallel()
+        del self._client
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()
+
     @classmethod
     def class_name(cls) -> str:
         return "Vllm"