feat(SpecEagleV2): add standalone_worker_v2 (#12625)

Co-authored-by: liugaoji.lgj <liugaoji.lgj@alibaba-inc.com>

test/registered/spec/test_standalone_speculative_decoding.py

@@ -1,3 +1,4 @@
+import os
 import unittest
 from types import SimpleNamespace
 
@@ -41,6 +42,25 @@ DEFAULT_SERVER_ARGS = [
     0.7,
 ]
 
+# Default server arguments for V2 tests
+DEFAULT_SERVER_ARGS_V2 = [
+    "--trust-remote-code",
+    "--cuda-graph-max-bs",
+    "8",
+    "--speculative-algorithm",
+    "STANDALONE",
+    "--speculative-draft-model-path",
+    DEFAULT_DRAFT_MODEL_STANDALONE,
+    "--speculative-num-steps",
+    "4",
+    "--speculative-eagle-topk",
+    "1",
+    "--speculative-num-draft-tokens",
+    "5",
+    "--mem-fraction-static",
+    0.7,
+]
+
 
 class TestStandaloneSpeculativeDecodingBase(CustomTestCase):
 
@@ -100,6 +120,67 @@ class TestStandaloneSpeculativeDecodingBase(CustomTestCase):
         self.assertGreater(avg_spec_accept_length, self.spec_decode_threshold)
 
 
+class TestStandaloneV2SpeculativeDecodingBase(CustomTestCase):
+
+    model = DEFAULT_TARGET_MODEL_STANDALONE
+    draft_model = DEFAULT_DRAFT_MODEL_STANDALONE
+    base_url = DEFAULT_URL_FOR_TEST
+    accuracy_threshold = 0.7  # derived tests need to override this
+    spec_decode_threshold = 3.6  # derived spec decoding tests need to override this
+
+    @classmethod
+    def get_server_args(cls):
+        """Return the arguments for the server launch. Override in subclasses."""
+        return DEFAULT_SERVER_ARGS_V2 + ["--attention-backend", "fa3"]
+
+    @classmethod
+    def setUpClass(cls):
+        # disable deep gemm precompile to make launch server faster
+        # please don't do this if you want to make your inference workload faster
+        envs.SGLANG_JIT_DEEPGEMM_PRECOMPILE.set(False)
+        envs.SGLANG_ENABLE_JIT_DEEPGEMM.set(False)
+        envs.SGLANG_ENABLE_SPEC_V2.set(True)  # Enable Speculative Decoding V2
+        model = cls.model
+        cls.process = popen_launch_server(
+            model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=cls.get_server_args(),
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+        if "SGLANG_ENABLE_SPEC_V2" in os.environ:
+            envs.SGLANG_ENABLE_SPEC_V2.set(False)
+
+    def test_gsm8k(self):
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=4,
+            num_questions=100,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+            data_path=GSM_DATASET_PATH,
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        # Use the appropriate metric key based on the test class
+        metric_key = "accuracy"
+        self.assertGreater(metrics[metric_key], self.accuracy_threshold)
+
+        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=}")
+        self.assertGreater(avg_spec_accept_length, self.spec_decode_threshold)
+
+
 class TestStandaloneSpeculativeDecodingTriton(TestStandaloneSpeculativeDecodingBase):
 
     @classmethod
@@ -115,5 +196,22 @@ class TestStandaloneSpeculativeDecodingFlashinfer(
         return DEFAULT_SERVER_ARGS + ["--attention-backend", "flashinfer"]
 
 
+class TestStandaloneV2SpeculativeDecodingTriton(
+    TestStandaloneV2SpeculativeDecodingBase
+):
+
+    @classmethod
+    def get_server_args(cls):
+        return DEFAULT_SERVER_ARGS_V2 + ["--attention-backend", "triton"]
+
+
+class TestStandaloneV2SpeculativeDecodingFlashinfer(
+    TestStandaloneV2SpeculativeDecodingBase
+):
+    @classmethod
+    def get_server_args(cls):
+        return DEFAULT_SERVER_ARGS_V2 + ["--attention-backend", "flashinfer"]
+
+
 if __name__ == "__main__":
     unittest.main()