[SPEC] feat: init adaptive spec params from config (#27493)

Co-authored-by: Qiaolin Yu <liin1211@outlook.com>

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:
@@ -340,18 +333,20 @@ 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
         )
-        from sglang.srt.server_args import auto_choose_speculative_params
 
         (
             server_args.speculative_num_steps,
             server_args.speculative_eagle_topk,
             server_args.speculative_num_draft_tokens,
-        ) = auto_choose_speculative_params(server_args)
+        ) = _auto_choose_speculative_params(server_args, model_arch)
 
     if (
         server_args.attention_backend == "trtllm_mha"
@@ -466,3 +461,63 @@ 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
+
+
+def _auto_choose_speculative_params(
+    server_args: "ServerArgs", model_arch: str
+) -> tuple:
+    """
+    Automatically choose the parameters for speculative decoding.
+
+    You can tune them on your own models and prompts with scripts/playground/bench_speculative.py
+    """
+    if server_args.speculative_algorithm == "STANDALONE":
+        return (3, 1, 4)
+    if model_arch in ["LlamaForCausalLM"]:
+        return (5, 4, 8)
+    elif model_arch in [
+        "DeepseekV32ForCausalLM",
+        "DeepseekV3ForCausalLM",
+        "DeepseekV2ForCausalLM",
+        "GptOssForCausalLM",
+        "Glm4MoeForCausalLM",
+        "Glm4MoeLiteForCausalLM",
+        "GlmMoeDsaForCausalLM",
+        "BailingMoeForCausalLM",
+        "BailingMoeV2ForCausalLM",
+        "BailingMoeV2_5ForCausalLM",
+        "MistralLarge3ForCausalLM",
+        "PixtralForConditionalGeneration",
+        "MiMoV2ForCausalLM",
+        "MiMoV2FlashForCausalLM",
+    ]:
+        return (3, 1, 4)
+    elif model_arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
+        return (5, 4, 8)
+    else:
+        return (3, 1, 4)

python/sglang/srt/server_args.py

@@ -8448,41 +8448,3 @@ class PortArgs:
                 ).to_tcp(),
                 instance_id=instance_id,
             )
-
-
-def auto_choose_speculative_params(self: ServerArgs):
-    """
-    Automatically choose the parameters for speculative decoding.
-
-    You can tune them on your own models and prompts with scripts/playground/bench_speculative.py
-    """
-    hf_config = self.get_model_config().hf_config
-    arch = hf_config.architectures[0]
-    if self.speculative_algorithm == "STANDALONE":
-        # The default value for standalone speculative decoding
-        return (3, 1, 4)
-    if arch in ["LlamaForCausalLM"]:
-        # The default value for llama
-        return (5, 4, 8)
-    elif arch in [
-        "DeepseekV32ForCausalLM",
-        "DeepseekV3ForCausalLM",
-        "DeepseekV2ForCausalLM",
-        "GptOssForCausalLM",
-        "Glm4MoeForCausalLM",
-        "Glm4MoeLiteForCausalLM",
-        "GlmMoeDsaForCausalLM",
-        "BailingMoeForCausalLM",
-        "BailingMoeV2ForCausalLM",
-        "BailingMoeV2_5ForCausalLM",
-        "MistralLarge3ForCausalLM",
-        "PixtralForConditionalGeneration",
-        "MiMoV2ForCausalLM",
-        "MiMoV2FlashForCausalLM",
-    ]:
-        return (3, 1, 4)
-    elif arch in ["Grok1ForCausalLM", "Grok1VForCausalLM"]:
-        return (5, 4, 8)
-    else:
-        # The default value for all other models
-        return (3, 1, 4)

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.