[AMD][CI] Fix Language Models Test (Extended Generation) failures (#45509)
Signed-off-by: Oxana Korzh <okorzh@amd.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Cursor <cursoragent@cursor.com>
Oxana Korzh committed
Jun 18, 2026 at 22:06 UTC
ecf9d83520eb217401b47d8a5451a27c5231b8c2
1 file changed
+5
-1
tests/models/language/generation/test_common.py
+5
-1
@@ -130,8 +130,12 @@ def test_models(
130
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
131
if model == "TitanML/tiny-mixtral":
132
# Untrained model: near-uniform logits make argmax sensitive to
133
- # AITER's bfloat16 rounding error in plain rms_norm.
133
+ # AITER's bfloat16 rounding error. Route the plain rms_norm and the
134
+ # fused MoE (whose near-uniform router logits flip expert selection
135
+ # under ~1 ULP drift) through the native kernels for this model.
136
+ # See ROCm/aiter#3806 for the tracking issue and minimal repro.
137
monkeypatch.setenv("VLLM_ROCM_USE_AITER_RMSNORM", "0")
138
+ monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "0")
139
elif use_rocm_aiter and model not in AITER_MODEL_LIST:
140
# Skip model that are not using AITER tests.
141
# When more AITER kernels are added, this list will not be