[BugFix] Fix GC bug for LLM class (#2882)

tests/test_regression.py

@@ -4,6 +4,10 @@ It should include tests that are reported by users and making sure they
 will never happen again.
 
 """
+import gc
+
+import torch
+
 from vllm import LLM, SamplingParams
 
 
@@ -35,6 +39,20 @@ def test_max_tokens_none():
     assert len(prompts) == len(outputs)
 
 
+def test_gc():
+    llm = LLM("facebook/opt-125m", enforce_eager=True)
+    del llm
+
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    # The memory allocated for model and KV cache should be released.
+    # The memory allocated for PyTorch and others should be less than 50MB.
+    # Usually, it's around 10MB.
+    allocated = torch.cuda.memory_allocated()
+    assert allocated < 50 * 1024 * 1024
+
+
 if __name__ == "__main__":
     import pytest
     pytest.main([__file__])