From 8c175f403acd2749086c298cd1c6f29ed9b4e2ba Mon Sep 17 00:00:00 2001 From: stefankoncarevic Date: Mon, 12 Oct 2026 15:20:35 +0200 Subject: [PATCH] [CI][ROCm] Compare greedy output only in the AITER attn+quant test (#61185) Signed-off-by: Stefan Koncarevic --- tests/compile/correctness_e2e/test_attn_quant.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/tests/compile/correctness_e2e/test_attn_quant.py b/tests/compile/correctness_e2e/test_attn_quant.py index e13a8f264a..1d8f73da4b 100644 --- a/tests/compile/correctness_e2e/test_attn_quant.py +++ b/tests/compile/correctness_e2e/test_attn_quant.py @@ -77,6 +77,10 @@ def test_rocm_aiter_static_attention_output_correctness(caplog_mp_spawn): env1=env, env2=env, method="generate", + # Fusing output quant drops a bf16 round-trip, so the two settings + # agree on the argmax but not on the exact distribution. Greedy + # decoding survives that, sampling at temperature 1.0 does not. + include_seeded_sampling=False, ) matches = [ int(count) for count in re.findall(r"'attn_quant_fusion': (\d+)", logs.text)