[SPEC] feat: init adaptive spec params from config

Author: alphabetc1

python/sglang/srt/arg_groups/speculative_hook.py

@@ -108,6 +108,11 @@ def handle_speculative_decoding(server_args: "ServerArgs") -> None:
             f"speculative_algorithm == EAGLE, got {server_args.speculative_algorithm}."
         )
 
+    if server_args.speculative_adaptive:
+        _maybe_disable_adaptive(server_args)
+        if server_args.speculative_adaptive:
+            _init_adaptive_speculative_params(server_args)
+
     if server_args.speculative_algorithm == "DFLASH":
         _handle_dflash(server_args)
     elif server_args.speculative_algorithm == "FROZEN_KV_MTP":
@@ -117,18 +122,6 @@ def handle_speculative_decoding(server_args: "ServerArgs") -> None:
     elif server_args.speculative_algorithm == "NGRAM":
         _handle_ngram(server_args)
 
-    if server_args.speculative_adaptive:
-        _maybe_disable_adaptive(server_args)
-        if server_args.speculative_adaptive:
-            from sglang.srt.speculative.adaptive_spec_params import (
-                validate_adaptive_initial_steps,
-            )
-
-            validate_adaptive_initial_steps(
-                server_args.speculative_num_steps,
-                server_args.speculative_adaptive_config,
-            )
-
 
 def _handle_dflash(server_args: "ServerArgs") -> None:
     if server_args.enable_dp_attention:
@@ -356,7 +349,10 @@ def _handle_eagle_family(server_args: "ServerArgs") -> None:
                     "DeepSeek MTP does not require setting speculative_draft_model_path."
                 )
 
-    if server_args.speculative_num_steps is None:
+    if (
+        not server_args.speculative_adaptive
+        and server_args.speculative_num_steps is None
+    ):
         assert (
             server_args.speculative_eagle_topk is None
             and server_args.speculative_num_draft_tokens is None
@@ -472,3 +468,31 @@ def _maybe_disable_adaptive(server_args: "ServerArgs") -> None:
             "Falling back to static speculative params."
         )
         server_args.speculative_adaptive = False
+
+
+def _init_adaptive_speculative_params(server_args: "ServerArgs") -> None:
+    from sglang.srt.speculative.adaptive_spec_params import (
+        resolve_candidate_steps_from_config,
+    )
+
+    candidate_steps = resolve_candidate_steps_from_config(
+        cfg_path=server_args.speculative_adaptive_config,
+    )
+
+    if server_args.speculative_eagle_topk is None:
+        server_args.speculative_eagle_topk = 1
+
+    if server_args.speculative_num_steps is None:
+        server_args.speculative_num_steps = candidate_steps[len(candidate_steps) // 2]
+
+    if server_args.speculative_num_steps not in candidate_steps:
+        raise ValueError(
+            f"--speculative-num-steps={server_args.speculative_num_steps} "
+            f"is not in the adaptive config candidate_steps {candidate_steps}. "
+            "Pass one of those values."
+        )
+
+    server_args.speculative_num_draft_tokens = (
+        server_args.speculative_num_steps + 1
+    )
+

python/sglang/srt/speculative/adaptive_spec_params.py

@@ -49,7 +49,10 @@ def adaptive_unsupported_reason(server_args: ServerArgs) -> str | None:
             f"speculative_algorithm={server_args.speculative_algorithm} "
             "(only EAGLE/EAGLE3 are supported)"
         )
-    if server_args.speculative_eagle_topk != 1:
+    if (
+        server_args.speculative_eagle_topk is not None
+        and server_args.speculative_eagle_topk != 1
+    ):
         return (
             f"speculative_eagle_topk={server_args.speculative_eagle_topk} "
             "(only topk=1 is supported)"
@@ -126,19 +129,6 @@ def resolve_candidate_steps_from_config(
     return sorted(all_steps)
 
 
-def validate_adaptive_initial_steps(
-    initial_steps: int,
-    cfg_path: str | None = None,
-) -> None:
-    """Require the initial step to be a candidate of some BS slot."""
-    candidate_steps = resolve_candidate_steps_from_config(cfg_path)
-    if initial_steps not in candidate_steps:
-        raise ValueError(
-            f"--speculative-num-steps={initial_steps} is not in the adaptive "
-            f"config candidate_steps {candidate_steps}. Pass one of those values."
-        )
-
-
 class AdaptiveStepSlot:
     """Tracks acceptance rate via EMA and adapts num_steps accordingly.
 

test/registered/spec/eagle/test_adaptive_speculative.py

@@ -72,12 +72,6 @@ def setUpClass(cls):
                     "EAGLE",
                     "--speculative-draft-model-path",
                     cls.draft_model,
-                    "--speculative-num-steps",
-                    "1",
-                    "--speculative-eagle-topk",
-                    "1",
-                    "--speculative-num-draft-tokens",
-                    "2",
                     "--speculative-adaptive",
                     "--speculative-adaptive-config",
                     cls.adaptive_config_path,

test/registered/unit/server_args/test_server_args.py

@@ -522,6 +522,38 @@ def test_external_corpus_max_tokens_must_be_positive(self):
         self.assertIn("external-corpus-max-tokens", str(context.exception))
 
 
+class TestAdaptiveSpecArgs(CustomTestCase):
+    def test_adaptive_defaults_to_config_step_when_spec_params_omitted(self):
+        with tempfile.NamedTemporaryFile("w", suffix=".json") as f:
+            json.dump(
+                {
+                    "1": {"candidate_steps": [1, 3, 5]},
+                    "8": {"candidate_steps": [1]},
+                },
+                f,
+            )
+            f.flush()
+
+            args = ServerArgs(model_path="dummy")
+            args.speculative_algorithm = "EAGLE"
+            args.speculative_adaptive = True
+            args.speculative_adaptive_config = f.name
+            args.device = "cuda"
+            args.get_model_config = lambda: SimpleNamespace(
+                hf_config=SimpleNamespace(
+                    architectures=["LlamaForCausalLM"],
+                    get_text_config=lambda: SimpleNamespace(),
+                )
+            )
+
+            handle_speculative_decoding(args)
+
+        self.assertTrue(args.speculative_adaptive)
+        self.assertEqual(args.speculative_eagle_topk, 1)
+        self.assertEqual(args.speculative_num_steps, 3)
+        self.assertEqual(args.speculative_num_draft_tokens, 4)
+
+
 class TestDeepEPWaterfillArgs(CustomTestCase):
     def test_waterfill_enforces_shared_experts_fusion(self):
         server_args = ServerArgs(

test/registered/unit/spec/test_adaptive_spec_params.py

@@ -6,7 +6,6 @@
     AdaptiveSpeculativeParams,
     AdaptiveStepSlot,
     resolve_candidate_steps_from_config,
-    validate_adaptive_initial_steps,
 )
 from sglang.test.ci.ci_register import register_cpu_ci, register_xpu_ci
 
@@ -403,25 +402,5 @@ def test_unions_and_dedups_across_slots(self):
         self.assertEqual(steps, [1, 3, 5, 7])
 
 
-class TestValidateAdaptiveInitialSteps(unittest.TestCase):
-    def test_accepts_value_from_any_slot(self):
-        with tempfile.NamedTemporaryFile("w", suffix=".json") as f:
-            json.dump(
-                {
-                    "1": {"candidate_steps": [1, 5]},
-                    "8": {"candidate_steps": [1, 3, 7]},
-                },
-                f,
-            )
-            f.flush()
-            # Membership in any slot is enough: 5 lives in the smallest slot,
-            # 7 only in a larger slot -- both accepted.
-            validate_adaptive_initial_steps(5, cfg_path=f.name)
-            validate_adaptive_initial_steps(7, cfg_path=f.name)
-            # 9 is in no slot -> rejected.
-            with self.assertRaises(ValueError):
-                validate_adaptive_initial_steps(9, cfg_path=f.name)
-
-
 if __name__ == "__main__":
     unittest.main()