[Bugfix][CI] Widen DBO+DP+EP GSM8K accuracy margin on ROCm (#59700)
Signed-off-by: Divakar Verma <divakar.verma@amd.com> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Divakar Verma committed
Oct 1, 2026 at 19:37 UTC
e6674262f19a0da389dad1c37b5bbd937db80778
1 file changed
+5
-1
tests/v1/distributed/test_dbo.py
+5
-1
@@ -11,6 +11,7 @@ import pytest
11
import torch
12
13
from tests.utils import RemoteOpenAIServer
14
+from vllm.platforms import current_platform
15
from vllm.utils.import_utils import has_deep_ep
16
17
# Detect Blackwell / B200 (compute capability 10.x)
@@ -30,7 +31,10 @@ DP_SIZE = 2
31
# GSM8K eval configuration
32
NUM_QUESTIONS = 256 # Fast eval for CI; but must be large enough to hit dbo thresholds
33
NUM_SHOTS = 5 # Few-shot examples
33
-MIN_ACCURACY = 0.62 # Expected 0.64 with 2% buffer (based on vLLM test data)
34
+
35
+# Expected 0.64 with 2% buffer (based on vLLM test data)
36
+# On ROCm, widened to 0.61 after observing 0.617 amd-ci nightly flake twice.
37
+MIN_ACCURACY = 0.61 if current_platform.is_rocm() else 0.62
38
39
# Increase max_num_seqs to trigger DBO for decode batches
40
# With 64 seqs, decode batches should exceed the 32 token threshold