[DLLM] Basic dLLM scheduling strategy and implementation (#17484)

Signed-off-by: Zehuan Li <lizehuan.lzh@antgroup.com>

test/registered/dllm/test_dllm_batching.py

@@ -1,71 +0,0 @@
-from sglang.test.ci.ci_register import register_cuda_ci
-
-register_cuda_ci(est_time=500, suite="stage-b-test-large-1-gpu")
-
-import unittest
-from types import SimpleNamespace
-
-from sglang.srt.utils import kill_process_tree
-from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
-from sglang.test.test_utils import (
-    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-    DEFAULT_URL_FOR_TEST,
-    CustomTestCase,
-    popen_launch_server,
-)
-
-"""
-Test dLLM batching capability on CUDA GPUs.
-
-As current dLLM batching performance is suboptimal to BS=1, this test only verifies correctness.
-The test will be removed once dLLM batching performance improves.
-"""
-
-
-class TestBatching(CustomTestCase):
-    @classmethod
-    def setUpClass(cls):
-        cls.model = "inclusionAI/LLaDA2.0-mini"
-        cls.base_url = DEFAULT_URL_FOR_TEST
-
-        other_args = [
-            "--trust-remote-code",
-            "--mem-fraction-static",
-            "0.9",
-            "--max-running-requests",
-            "4",
-            "--attention-backend",
-            "flashinfer",
-            "--dllm-algorithm",
-            "LowConfidence",
-        ]
-
-        cls.process = popen_launch_server(
-            cls.model,
-            cls.base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=other_args,
-        )
-
-    @classmethod
-    def tearDownClass(cls):
-        kill_process_tree(cls.process.pid)
-
-    def test_gsm8k(self):
-        args = SimpleNamespace(
-            num_shots=5,
-            data_path=None,
-            num_questions=200,
-            max_new_tokens=512,
-            parallel=128,
-            host="http://127.0.0.1",
-            port=int(self.base_url.split(":")[-1]),
-        )
-        metrics = run_eval_few_shot_gsm8k(args)
-        print(f"{metrics=}")
-
-        self.assertGreater(metrics["accuracy"], 0.88)
-
-
-if __name__ == "__main__":
-    unittest.main()

test/registered/dllm/test_llada2_mini.py

@@ -28,14 +28,21 @@ class TestLLaDA2Mini(CustomTestCase):
 
         other_args = [
             "--trust-remote-code",
+            "--tp-size",
+            "1",
             "--mem-fraction-static",
             "0.9",
             "--max-running-requests",
-            "1",
+            "4",
             "--attention-backend",
             "flashinfer",
             "--dllm-algorithm",
-            "LowConfidence",  # TODO: Add dLLM configurations
+            "LowConfidence",
+            "--cuda-graph-bs",
+            "1",
+            "2",
+            "3",
+            "4",
         ]
 
         cls.process = popen_launch_server(
@@ -66,7 +73,7 @@ class TestLLaDA2Mini(CustomTestCase):
         if is_in_amd_ci():
             self.assertGreater(metrics["output_throughput"], 80)
         else:
-            self.assertGreater(metrics["output_throughput"], 150)
+            self.assertGreater(metrics["output_throughput"], 250)
 
     def test_bs_1_speed(self):
         args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)