A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 01:00 UTC
Mark DBO test as flaky on b200 for Distributed B200 test (#29913)
Signed-off-by: dougbtv <dosmith@redhat.com>
Doug Smith committed
Dec 4, 2025 at 10:38 UTC
5b4b42c0b6e190dacbf6dbfed3506c9e58bfc8de
1 file changed
+19
tests/v1/distributed/test_dbo.py
+19
@@ -9,10 +9,22 @@ correctly with the DeepSeek-V2-Lite model using GSM8K evaluation.
9
"""
10
11
import pytest
12
+import torch
13
14
from tests.evals.gsm8k.gsm8k_eval import evaluate_gsm8k
15
from tests.utils import RemoteOpenAIServer
16
17
+# Detect Blackwell / B200 (compute capability 10.x)
18
+try:
19
+ if torch.cuda.is_available():
20
+ cap = torch.cuda.get_device_capability(0)
21
+ IS_BLACKWELL = cap[0] >= 10
22
+ else:
23
+ IS_BLACKWELL = False
24
+except Exception:
25
+ # Be conservative: if we can't detect, don't xfail by default
26
+ IS_BLACKWELL = False
27
+
28
MODEL_NAME = "deepseek-ai/DeepSeek-V2-Lite-Chat"
29
DP_SIZE = 2
30
@@ -33,6 +45,13 @@ DEEPEP_BACKENDS = [
45
46
47
@pytest.mark.parametrize("all2all_backend", DEEPEP_BACKENDS)
48
+@pytest.mark.xfail(
49
+ IS_BLACKWELL,
50
+ reason=(
51
+ "Temporary: DBO accuracy unstable on Blackwell "
52
+ "(doesn't meet expectation of MIN_ACCURACY = 0.62)"
53
+ ),
54
+)
55
def test_dbo_dp_ep_gsm8k(all2all_backend: str, num_gpus_available):
56
"""
57
Test DBO with DP+EP using GSM8K evaluation.