From 89ac8d971be7d2f598598c1551a21c646bdf8b0d Mon Sep 17 00:00:00 2001 From: Andrey Talman Date: Fri, 9 Oct 2026 12:59:17 -0400 Subject: [PATCH] [CI][Test] Allow 10% ngram acceptance drop under preemption on CUDA in test_async_scheduling_accuracy (#60715) Signed-off-by: Andrey Talman Co-authored-by: Andrey Talman --- tests/v1/e2e/general/test_async_scheduling_accuracy.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/v1/e2e/general/test_async_scheduling_accuracy.py b/tests/v1/e2e/general/test_async_scheduling_accuracy.py index aa785b3270..3ffcaefdc3 100644 --- a/tests/v1/e2e/general/test_async_scheduling_accuracy.py +++ b/tests/v1/e2e/general/test_async_scheduling_accuracy.py @@ -271,9 +271,9 @@ def run_tests( context = f"config=[{config}], params={params}" if spec_config.get("max_model_len") is None: # Keep the original acceptance-quality floor per batch. - relative_drop = ( - 0.10 if current_platform.is_rocm() and test_preemption else 0.05 - ) + # Preemption recomputes requests and loses n-gram drafts, + # so allow a larger drop there. + relative_drop = 0.10 if test_preemption else 0.05 assert actual >= baseline * (1 - relative_drop), ( f"{context}: acceptance={actual}, baseline={baseline}" )