From ecf9d83520eb217401b47d8a5451a27c5231b8c2 Mon Sep 17 00:00:00 2001 From: Oxana Korzh Date: Thu, 18 Jun 2026 22:06:56 -0600 Subject: [PATCH] [AMD][CI] Fix Language Models Test (Extended Generation) failures (#45509) Signed-off-by: Oxana Korzh Co-authored-by: Claude Co-authored-by: Cursor --- tests/models/language/generation/test_common.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/tests/models/language/generation/test_common.py b/tests/models/language/generation/test_common.py index 1b6c8ef558..50c87d7729 100644 --- a/tests/models/language/generation/test_common.py +++ b/tests/models/language/generation/test_common.py @@ -130,8 +130,12 @@ def test_models( monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1") if model == "TitanML/tiny-mixtral": # Untrained model: near-uniform logits make argmax sensitive to - # AITER's bfloat16 rounding error in plain rms_norm. + # AITER's bfloat16 rounding error. Route the plain rms_norm and the + # fused MoE (whose near-uniform router logits flip expert selection + # under ~1 ULP drift) through the native kernels for this model. + # See ROCm/aiter#3806 for the tracking issue and minimal repro. monkeypatch.setenv("VLLM_ROCM_USE_AITER_RMSNORM", "0") + monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "0") elif use_rocm_aiter and model not in AITER_MODEL_LIST: # Skip model that are not using AITER tests. # When more AITER kernels are added, this list will not be