From 757dd3b2c4eb8a5efaa28503f0e91e8cb19d03ce Mon Sep 17 00:00:00 2001 From: tzhouam Date: Sat, 3 Oct 2026 22:51:54 +0800 Subject: [PATCH] Leave Code2Wav workspace headroom in two-L4 serving fixture Signed-off-by: tzhouam --- tests/e2e/online_serving/test_qwen2_5_omni_expansion.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/e2e/online_serving/test_qwen2_5_omni_expansion.py b/tests/e2e/online_serving/test_qwen2_5_omni_expansion.py index de31172b6..b5ecd3435 100644 --- a/tests/e2e/online_serving/test_qwen2_5_omni_expansion.py +++ b/tests/e2e/online_serving/test_qwen2_5_omni_expansion.py @@ -24,9 +24,9 @@ models = ["Qwen/Qwen2.5-Omni-7B"] stage_configs = [ modify_stage_config( get_deploy_config_path("ci/qwen2_5_omni.yaml"), - # This suite requests two cards. The general CI overlay puts Code2Wav - # on a third card; share the Talker's card within the existing budgets. - updates={"stages": {2: {"devices": "1"}}}, + # This suite requests two cards. Share the Talker's card with + # Code2Wav and leave room for its whole-sequence DiT workspace. + updates={"stages": {1: {"gpu_memory_utilization": 0.3}, 2: {"devices": "1"}}}, ) ]