[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>
Andreas Karatzas committed
Sep 21, 2026 at 00:27 UTC
8ec00ec2bc2d294b742f7814041d3a3043e0a90e
1 file changed
+3
-1
tests/models/language/generation/test_common.py
+3
-1
@@ -158,7 +158,9 @@ def test_models(
158
hf_model.model.device
159
)
160
if prompt_embeds is not None:
161
- embed = hf_model.model.get_input_embeddings()(token_ids)
161
+ # Retained prompt values must not keep reference weights alive.
162
+ with torch.no_grad():
163
+ embed = hf_model.model.get_input_embeddings()(token_ids)
164
165
# MiniCPM models apply scale_emb to embeddings internally.
166
# vLLM expects pre-scaled embeddings when using inputs_embeds.