Mark DBO test as flaky on b200 for Distributed B200 test (#29913)

Signed-off-by: dougbtv <dosmith@redhat.com>

tests/v1/distributed/test_dbo.py

@@ -9,10 +9,22 @@ correctly with the DeepSeek-V2-Lite model using GSM8K evaluation.
 """
 
 import pytest
+import torch
 
 from tests.evals.gsm8k.gsm8k_eval import evaluate_gsm8k
 from tests.utils import RemoteOpenAIServer
 
+# Detect Blackwell / B200 (compute capability 10.x)
+try:
+    if torch.cuda.is_available():
+        cap = torch.cuda.get_device_capability(0)
+        IS_BLACKWELL = cap[0] >= 10
+    else:
+        IS_BLACKWELL = False
+except Exception:
+    # Be conservative: if we can't detect, don't xfail by default
+    IS_BLACKWELL = False
+
 MODEL_NAME = "deepseek-ai/DeepSeek-V2-Lite-Chat"
 DP_SIZE = 2
 
@@ -33,6 +45,13 @@ DEEPEP_BACKENDS = [
 
 
 @pytest.mark.parametrize("all2all_backend", DEEPEP_BACKENDS)
+@pytest.mark.xfail(
+    IS_BLACKWELL,
+    reason=(
+        "Temporary: DBO accuracy unstable on Blackwell "
+        "(doesn't meet expectation of MIN_ACCURACY = 0.62)"
+    ),
+)
 def test_dbo_dp_ep_gsm8k(all2all_backend: str, num_gpus_available):
     """
     Test DBO with DP+EP using GSM8K evaluation.