Refactor / Unify event loop across PD-Disagg, Overlap, DP-Attn cases (#12839)

Co-authored-by: cctry <17473714+cctry@users.noreply.github.com>

test/srt/run_suite.py

@@ -158,7 +158,6 @@ suites = {
     ],
     "per-commit-4-gpu": [
         TestFile("models/test_qwen3_next_models.py", 291),
-        TestFile("test_disaggregation_dp_attention.py", 155),
         TestFile("test_gpt_oss_4gpu.py", 300),
         TestFile("test_local_attn.py", 411),
         TestFile("test_multi_instance_release_memory_occupation.py", 64),
@@ -174,6 +173,7 @@ suites = {
         TestFile("quant/test_w4a8_deepseek_v3.py", 520),
         TestFile("test_disaggregation_different_tp.py", 600),
         TestFile("test_disaggregation_pp.py", 140),
+        TestFile("test_disaggregation_dp_attention.py", 155),
     ],
     "per-commit-4-gpu-b200": [
         TestFile("test_deepseek_v3_fp4_4gpu.py", 3600),

test/srt/test_disaggregation_dp_attention.py

@@ -1,4 +1,3 @@
-import os
 import unittest
 from types import SimpleNamespace
 
@@ -9,17 +8,21 @@ from sglang.test.test_utils import (
     DEFAULT_MODEL_NAME_FOR_TEST_MLA,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     popen_launch_pd_server,
+    try_cached_model,
 )
 
 
 class TestDisaggregationDPAttention(TestDisaggregationBase):
+    PREFILL_DP_SIZE = 4
+    DECODE_DP_SIZE = 4
+
     @classmethod
     def setUpClass(cls):
         super().setUpClass()
         # Temporarily disable JIT DeepGEMM
         envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
 
-        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
+        cls.model = try_cached_model(DEFAULT_MODEL_NAME_FOR_TEST_MLA)
 
         # Non blocking start servers
         cls.start_prefill()
@@ -38,9 +41,9 @@ class TestDisaggregationDPAttention(TestDisaggregationBase):
             "--disaggregation-mode",
             "prefill",
             "--tp",
-            "2",
+            str(cls.PREFILL_DP_SIZE),
             "--dp",
-            "2",
+            str(cls.PREFILL_DP_SIZE),
             "--enable-dp-attention",
         ]
         prefill_args += cls.transfer_backend + cls.rdma_devices
@@ -58,12 +61,13 @@ class TestDisaggregationDPAttention(TestDisaggregationBase):
             "--disaggregation-mode",
             "decode",
             "--tp",
-            "2",
+            str(cls.DECODE_DP_SIZE),
             "--dp",
-            "2",
+            str(cls.DECODE_DP_SIZE),
             "--enable-dp-attention",
             "--base-gpu-id",
-            "2",
+            str(cls.PREFILL_DP_SIZE),
+            "--prefill-round-robin-balance",
         ]
         decode_args += cls.transfer_backend + cls.rdma_devices
         cls.process_decode = popen_launch_pd_server(
@@ -77,7 +81,7 @@ class TestDisaggregationDPAttention(TestDisaggregationBase):
         args = SimpleNamespace(
             num_shots=5,
             data_path=None,
-            num_questions=200,
+            num_questions=1400,
             max_new_tokens=512,
             parallel=128,
             host=f"http://{self.base_host}",