[Feature] Spec-Overlap supporting DP-ATTN; PD-Disaggregation; npugraph mode (#12443)

test/srt/ascend/test_ascend_deepseek_mtp.py

@@ -43,6 +43,9 @@ class TestAscendDeepSeekMTP(CustomTestCase):
             32768,
             "--tp-size",
             16,
+            "--dp-size",
+            2,
+            "--enable-dp-attention",
             "--speculative-algorithm",
             "NEXTN",
             "--speculative-num-steps",
@@ -55,6 +58,8 @@ class TestAscendDeepSeekMTP(CustomTestCase):
 
         cls.extra_envs = {
             "SGLANG_NPU_USE_MLAPO": "1",
+            "SGLANG_ENABLE_SPEC_V2": "1",
+            "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
         }
         os.environ.update(cls.extra_envs)
 
@@ -91,26 +96,6 @@ class TestAscendDeepSeekMTP(CustomTestCase):
                 finally:
                     kill_process_tree(process.pid)
 
-    def test_b_throughput(self):
-        for model in self.models:
-            with self.subTest(model=model):
-                print(f"##=== Testing throughput: {model} ===##")
-
-                output_throughput = run_bench_offline_throughput(
-                    model,
-                    [
-                        *self.common_args,
-                    ],
-                )
-
-                print(f"##=== {model} throughput: {output_throughput} ===##")
-
-                if is_in_ci():
-                    self.assertGreater(
-                        output_throughput,
-                        TEST_MODEL_MATRIX[model]["output_throughput"],
-                    )
-
 
 if __name__ == "__main__":
     unittest.main()

test/srt/run_suite.py

@@ -181,6 +181,7 @@ suites = {
         TestFile("test_flash_attention_4.py", 300),
         TestFile("test_gpt_oss_4gpu.py", 600),
         TestFile("test_llama31_fp4.py", 300),
+        TestFile("test_eagle_infer_beta_dp_attention.py", 200),
     ],
     "per-commit-4-gpu-gb200": [
         TestFile("test_cutedsl_moe.py", 300),

test/srt/test_eagle_infer_beta_dp_attention.py

@@ -0,0 +1,97 @@
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+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,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
+
+
+class TestEagleDPAttnServerBase(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        os.environ["SGLANG_ENABLE_SPEC_V2"] = "1"
+        cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp-size",
+            "4",
+            "--dp-size",
+            "4",
+            "--enable-dp-attention",
+            "--attention-backend",
+            "trtllm_mla",
+            "--moe-runner-backend",
+            "flashinfer_trtllm",
+            "--quantization",
+            "modelopt_fp4",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+        ]
+        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)
+        if "SGLANG_ENABLE_SPEC_V2" in os.environ:
+            del os.environ["SGLANG_ENABLE_SPEC_V2"]
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        requests.get(self.base_url + "/flush_cache")
+
+        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=}")
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
+        print(f"{avg_spec_accept_length=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3-fp4 mtp)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+            self.assertGreater(metrics["accuracy"], 0.94)
+            self.assertGreater(avg_spec_accept_length, 2.04)
+
+
+if __name__ == "__main__":
+    unittest.main()