[ROCm][CI] Add TRITON_ATTN score absolute tolerance floor (#46891)
Signed-off-by: pei.zhang <pei.zhang@amd.com> Co-authored-by: Claude <noreply@anthropic.com>
peizhang56 committed
Jun 26, 2026 at 23:35 UTC
091d13976c1c246714bb2112dd2e208561dda6a3
1 file changed
+8
-3
tests/entrypoints/pooling/scoring/test_cross_encoder_online_vision.py
+8
-3
@@ -38,12 +38,17 @@ BACKEND_TOL: dict[str, float] = {
38
"FLEX_ATTENTION": 0.045, # gfx950:~3.25%, gfx942:~1.10%
39
}
40
41
-# ROCm 7.2/gfx950 shows small absolute drift on the low text-vs-text
42
-# probability even though larger scores remain well inside the relative
43
-# tolerance. Keep the relative tolerances tight and add only a small floor.
41
+# Some ROCm attention backends show small absolute drift on the low
42
+# text-vs-text probability even though larger scores remain well inside the
43
+# relative tolerance. The absolute drift is uniform across score magnitudes
44
+# (~0.005-0.010), so it only exceeds the relative tolerance for the small
45
+# ~0.10 text-vs-text value. Keep the relative tolerances tight and add only a
46
+# small absolute floor for the affected backends.
47
+# TRITON_ATTN: gfx942/ROCm 7.2 drifts ~0.008 abs on text-vs-text (~7.9% rel).
48
BACKEND_ABS_TOL: dict[str, float] = {
49
"default": 0.0,
50
"ROCM_AITER_FA": 0.005,
51
+ "TRITON_ATTN": 0.009,
52
"FLEX_ATTENTION": 0.006,
53
}
54