[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>
2026-09-21 05:27:29 +00:00
parent 6654bcbfca
commit 8ec00ec2bc
1 changed files with 3 additions and 1 deletions
@@ -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.