diff --git a/tests/models/language/generation/test_common.py b/tests/models/language/generation/test_common.py
index f70d5a0c68..450e8badce 100644
--- a/tests/models/language/generation/test_common.py
+++ b/tests/models/language/generation/test_common.py
@@ -158,7 +158,9 @@ def test_models(
                 hf_model.model.device
             )
             if prompt_embeds is not None:
-                embed = hf_model.model.get_input_embeddings()(token_ids)
+                # Retained prompt values must not keep reference weights alive.
+                with torch.no_grad():
+                    embed = hf_model.model.get_input_embeddings()(token_ids)
 
                 # MiniCPM models apply scale_emb to embeddings internally.
                 # vLLM expects pre-scaled embeddings when using inputs_embeds.