fix(gpt-oss): avoid duplicate MoE reduction with DP attention (#37199)

test/registered/models_e2e/test_gpt_oss_4gpu_mxfp4.py

@@ -3,8 +3,8 @@ import unittest
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.gpt_oss_common import BaseTestGptOss
 
-register_cuda_ci(est_time=220, stage="base-c", runner_config="4-gpu-h100")
-register_cuda_ci(est_time=220, stage="base-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=300, stage="base-c", runner_config="4-gpu-h100")
+register_cuda_ci(est_time=300, stage="base-c", runner_config="4-gpu-b200")
 
 
 class TestGptOss4GpuMxfp4(BaseTestGptOss):
@@ -23,6 +23,28 @@ class TestGptOss4GpuMxfp4(BaseTestGptOss):
             ],
         )
 
+    def test_mxfp4_120b_dpa(self):
+        self.run_test(
+            model_variant="120b",
+            quantization="mxfp4",
+            expected_score_of_reasoning_effort={
+                "low": 0.50,
+            },
+            other_args=[
+                "--tp",
+                "4",
+                "--dp",
+                "4",
+                "--ep",
+                "4",
+                "--enable-dp-attention",
+                "--moe-a2a-backend",
+                "none",
+                "--cuda-graph-max-bs-decode",
+                "200",
+            ],
+        )
+
 
 if __name__ == "__main__":
     unittest.main()