Support PD disaggregation with different TP/DP size for Qwen3-Next (#16056)

Co-authored-by: xjx392321 <xjx392321@alibaba-inc.com>
2026-01-25 15:34:02 +08:00
parent 30ece5e1d6
commit 9bd92ba0f6
6 changed files with 258 additions and 13 deletions
@@ -153,5 +153,79 @@ class TestDisaggregationHybridAttentionMambaExtraBuffer(PDDisaggregationServerBa
        self.assertGreater(metrics["accuracy"], 0.93)


+class TestDisaggregationHybridAttentionMambaDPDecode(PDDisaggregationServerBase):
+    """Test with prefill tp=2 and decode tp=2/dp=2 with dp-attention enabled."""
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct"
+
+        # Non blocking start servers
+        cls.start_prefill()
+        cls.start_decode()
+
+        # Block until both
+        cls.wait_server_ready(cls.prefill_url + "/health")
+        cls.wait_server_ready(cls.decode_url + "/health")
+
+        cls.launch_lb()
+
+    @classmethod
+    def start_prefill(cls):
+        prefill_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "prefill",
+            "--tp",
+            "2",
+        ]
+        prefill_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_prefill = popen_launch_pd_server(
+            cls.model,
+            cls.prefill_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=prefill_args,
+        )
+
+    @classmethod
+    def start_decode(cls):
+        decode_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "decode",
+            "--tp",
+            "2",
+            "--dp",
+            "2",
+            "--enable-dp-attention",
+            "--enable-dp-lm-head",
+            "--base-gpu-id",
+            "2",
+        ]
+        decode_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host=f"http://{self.base_host}",
+            port=int(self.lb_port),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"Evaluation metrics: {metrics}")
+
+        self.assertGreater(metrics["accuracy"], 0.93)
+
+
 if __name__ == "__main__":
    unittest.main()