Enable the qwen3 test (#21195)

Co-authored-by: Shunkang <182541032+Shunkangz@users.noreply.github.co>

python/sglang/srt/models/qwen3_moe.py

@@ -33,6 +33,7 @@ from sglang.srt.distributed import (
     get_moe_tensor_parallel_world_size,
     get_pp_group,
     get_tensor_model_parallel_rank,
+    moe_expert_parallel_all_reduce,
     moe_tensor_model_parallel_all_reduce,
 )
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
@@ -321,6 +322,9 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
         topk_output = self.topk(hidden_states, router_logits)
         final_hidden_states = self.experts(hidden_states, topk_output)
 
+        if self.ep_size > 1 and not should_allreduce_fusion:
+            final_hidden_states = moe_expert_parallel_all_reduce(final_hidden_states)
+
         if (
             self.tp_size > 1
             and not should_allreduce_fusion

test/registered/4-gpu-models/test_qwen3_30b.py

@@ -11,11 +11,8 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(
-    est_time=300,
-    suite="stage-c-test-4-gpu-h100",
-    disabled="Temporarily disable the flaky test.",
-)
+register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-h100")
+
 QWEN3_30B_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"
 
 GSM8K_BASELINE_ACCURACY = 0.85