From 8ec00ec2bc2d294b742f7814041d3a3043e0a90e Mon Sep 17 00:00:00 2001
From: Andreas Karatzas <akaratza@amd.com>
Date: Mon, 21 Sep 2026 00:27:29 -0500
Subject: [PATCH] [Tests] Release HF embedding weights after prompt extraction
 (#57868)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
---
 tests/models/language/generation/test_common.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

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.