Tiny add PD disaggregation + DP attention test (#11167)

test/srt/run_suite.py

@@ -147,6 +147,7 @@ suites = {
         TestFile("hicache/test_hicache_storage_mooncake_backend.py", 800),
         TestFile("lora/test_lora_llama4.py", 600),
         TestFile("test_disaggregation.py", 499),
+        TestFile("test_disaggregation_dp_attention.py", 155),
         TestFile("test_disaggregation_different_tp.py", 155),
         TestFile("test_disaggregation_pp.py", 60),
         TestFile("test_full_deepseek_v3.py", 333),

test/srt/test_disaggregation_dp_attention.py

@@ -0,0 +1,107 @@
+import os
+import time
+import unittest
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_disaggregation_utils import TestDisaggregationBase
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST_MLA,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_pd_server,
+)
+
+
+class TestDisaggregationDPAttention(TestDisaggregationBase):
+    @classmethod
+    def setUpClass(cls):
+        # Temporarily disable JIT DeepGEMM
+        cls.original_jit_deepgemm = os.environ.get("SGL_ENABLE_JIT_DEEPGEMM")
+        os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = "false"
+
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
+        parsed_url = urlparse(DEFAULT_URL_FOR_TEST)
+        cls.base_host = parsed_url.hostname
+        base_port = str(parsed_url.port)
+        cls.lb_port = base_port
+        cls.prefill_port = f"{int(base_port) + 100}"
+        cls.decode_port = f"{int(base_port) + 200}"
+        cls.prefill_url = f"http://{cls.base_host}:{cls.prefill_port}"
+        cls.decode_url = f"http://{cls.base_host}:{cls.decode_port}"
+        cls.lb_url = f"http://{cls.base_host}:{cls.lb_port}"
+        print(f"{cls.base_host=} {cls.lb_port=} {cls.prefill_port=} {cls.decode_port=}")
+
+        # 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",
+            "--dp",
+            "2",
+            "--enable-dp-attention",
+            "--disaggregation-ib-device",
+            "mlx5_roce0,mlx5_roce1",
+        ]
+        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",
+            "--base-gpu-id",
+            "2",
+            "--disaggregation-ib-device",
+            "mlx5_roce2,mlx5_roce3",
+        ]
+        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"], 0.60)
+
+
+if __name__ == "__main__":
+    unittest.main()