Support Qwen3 MoE context parallel (#18233)

Co-authored-by: Shunkang <182541032+Shunkangz@users.noreply.github.co>
Co-authored-by: Jiying Dong <87510204+dongjiyingdjy@users.noreply.github.com>

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

@@ -0,0 +1,77 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    kill_process_tree,
+    popen_launch_server,
+)
+
+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
+
+
+class TestQwen330B(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = QWEN3_30B_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--tp-size",
+                "4",
+                "--moe-dp-size",
+                "2",
+                "--ep-size",
+                "2",
+                "--attn-cp-size",
+                "2",
+                "--enable-prefill-context-parallel",
+                "--cuda-graph-max-bs",
+                "32",
+                "--max-running-requests",
+                "32",
+                "--trust-remote-code",
+                "--disable-piecewise-cuda-graph",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true, "num_threads": 64}',
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            model=self.model,
+            eval_name="gsm8k",
+            num_shots=5,
+            num_examples=200,
+            max_tokens=16000,
+            num_threads=128,
+            repeat=1,
+            temperature=0.6,
+            top_p=0.95,
+            top_k=20,
+            base_url=self.base_url,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreaterEqual(metrics["score"], GSM8K_BASELINE_ACCURACY)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/8-gpu-models/test_qwen3_235b.py

@@ -18,9 +18,10 @@ QWEN3_235B_EAGLE3_MODEL_PATH = (
 class TestQwen3235BFP8(unittest.TestCase):
     """Test class for Qwen3-235B-FP8 performance and accuracy.
 
-    Two variants:
+    Three variants:
     - basic: TP=8
     - eagle3: TP=8 + EP=2 + EAGLE3 speculative decoding
+    - TP8+CP2+EP2: TP=8 + CP=2 + EP=2 context parallel
 
     Each variant runs BOTH:
     - Performance test (using NightlyBenchmarkRunner)
@@ -69,6 +70,42 @@ class TestQwen3235BFP8(unittest.TestCase):
             ),
         )
 
+    def test_qwen3_235b_fp8_cp(self):
+        """Run performance and accuracy for Qwen3-235B-FP8 with context parallelism."""
+
+        BASE_ARGS = [
+            "--trust-remote-code",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true, "num_threads": 64}',
+        ]
+
+        DP_ARGS = [
+            "--tp=8",
+            "--moe-dp-size=2",
+            "--attn-cp-size=2",
+            "--ep-size=4",
+            "--enable-prefill-context-parallel",
+        ]
+
+        MTP_ARGS = [
+            "--cuda-graph-max-bs=32",
+            "--max-running-requests=32",
+        ]
+        variants = [
+            ModelLaunchSettings(
+                QWEN3_235B_FP8_MODEL_PATH,
+                tp_size=8,
+                extra_args=BASE_ARGS + DP_ARGS + MTP_ARGS,
+                variant="TP8+CP2+EP2",
+            ),
+        ]
+
+        run_combined_tests(
+            models=variants,
+            test_name="Qwen3-235B-FP8 Context Parallel",
+            accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.88),
+        )
+
 
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/managers/test_prefill_adder.py

@@ -1,6 +1,6 @@
 import unittest
 from types import SimpleNamespace
-from unittest.mock import MagicMock, patch
+from unittest.mock import MagicMock
 
 from sglang.srt.managers.schedule_batch import Req
 from sglang.srt.managers.schedule_policy import AddReqResult, PrefillAdder
@@ -8,6 +8,7 @@ from sglang.srt.mem_cache.base_prefix_cache import (
     DecLockRefResult,
     IncLockRefResult,
 )
+from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
@@ -17,14 +18,9 @@ register_amd_ci(est_time=2, suite="stage-b-test-small-1-gpu-amd")
 
 class TestPrefillAdder(CustomTestCase):
     def setUp(self):
+        set_global_server_args_for_scheduler(ServerArgs(model_path="dummy"))
         self.mock_tree_cache = self.create_tree_cache()
         self.mock_token_allocator = self.create_token_allocator()
-        patcher = patch(
-            "sglang.srt.managers.schedule_policy.is_nsa_prefill_cp_in_seq_split",
-            return_value=False,
-        )
-        self.mock_is_nsa = patcher.start()
-        self.addCleanup(patcher.stop)
 
     def create_tree_cache(
         self,