[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>
This commit is contained in:
Divakar Verma
2026-10-01 23:37:36 +00:00
committed by GitHub
co-authored by Claude Sonnet 5
parent e35082e7b6
commit e6674262f1
+5 -1
View File
@@ -11,6 +11,7 @@ import pytest
import torch
from tests.utils import RemoteOpenAIServer
from vllm.platforms import current_platform
from vllm.utils.import_utils import has_deep_ep
# Detect Blackwell / B200 (compute capability 10.x)
@@ -30,7 +31,10 @@ DP_SIZE = 2
# GSM8K eval configuration
NUM_QUESTIONS = 256 # Fast eval for CI; but must be large enough to hit dbo thresholds
NUM_SHOTS = 5 # Few-shot examples
MIN_ACCURACY = 0.62 # Expected 0.64 with 2% buffer (based on vLLM test data)
# Expected 0.64 with 2% buffer (based on vLLM test data)
# On ROCm, widened to 0.61 after observing 0.617 amd-ci nightly flake twice.
MIN_ACCURACY = 0.61 if current_platform.is_rocm() else 0.62
# Increase max_num_seqs to trigger DBO for decode batches
# With 64 seqs, decode batches should exceed the 32 token threshold