[CI][ROCm] Compare greedy output only in the AITER attn+quant test (#61185)
Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
This commit is contained in:
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user