[SPEC] fix: use effective max draft tokens for adaptive spec initiali… (#26354)

Co-authored-by: maodoudou168 <maodoudou168@users.noreply.github.com>

python/sglang/srt/managers/tokenizer_manager.py

@@ -297,7 +297,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             # so we need to reserve the space for the draft tokens.
             self.num_reserved_tokens = max(
                 server_args.speculative_eagle_topk * server_args.speculative_num_steps,
-                server_args.speculative_num_draft_tokens,
+                server_args.max_speculative_num_draft_tokens,
             )
         else:
             self.num_reserved_tokens = 0

python/sglang/srt/managers/utils.py

@@ -240,7 +240,7 @@ def get_alloc_len_per_decode(server_args: Optional[ServerArgs] = None) -> int:
 
     spec_steps = server_args.speculative_num_steps or 1
     spec_topk = server_args.speculative_eagle_topk or 1
-    spec_tokens = server_args.speculative_num_draft_tokens
+    spec_tokens = server_args.max_speculative_num_draft_tokens
     page_size = server_args.page_size
 
     if page_size == 1 or spec_topk == 1:

python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py

@@ -285,9 +285,7 @@ class ModelRunnerKVCacheMixin:
         # Initialize req_to_token_pool
         if self.req_to_token_pool is None:
             # FIXME(lsyin): this is the temporary fix for the context length issue when using speculative decoding
-            max_spec_draft_tokens = (
-                self.server_args.effective_max_speculative_num_draft_tokens()
-            )
+            max_spec_draft_tokens = self.server_args.max_speculative_num_draft_tokens
             extra_max_context_len = 4
             if max_spec_draft_tokens is not None:
                 extra_max_context_len += max_spec_draft_tokens

python/sglang/srt/server_args.py

@@ -25,6 +25,7 @@ import logging
 import os
 import random
 import tempfile
+from functools import cached_property
 from typing import Any, Callable, Dict, List, Literal, Optional, Union
 
 from sglang.srt.arg_groups.argparse_actions import (
@@ -590,8 +591,6 @@ class ServerArgs:
     speculative_moe_runner_backend: Optional[str] = None
     speculative_moe_a2a_backend: Optional[str] = None
     speculative_draft_model_quantization: Optional[str] = None
-    speculative_adaptive: bool = False
-    speculative_adaptive_config: Optional[str] = None
     speculative_skip_dp_mlp_sync: bool = False
 
     # Speculative decoding (ngram)
@@ -605,6 +604,10 @@ class ServerArgs:
     speculative_ngram_external_corpus_max_tokens: int = 10000000
     enable_multi_layer_eagle: bool = False
 
+    # Adaptive speculative decoding
+    speculative_adaptive: bool = False
+    speculative_adaptive_config: Optional[str] = None
+
     # Expert parallelism
     ep_size: int = 1
     moe_a2a_backend: Literal[
@@ -7050,8 +7053,9 @@ class ServerArgs:
     def enable_mamba_extra_buffer(self) -> bool:
         return self.mamba_scheduler_strategy == "extra_buffer"
 
-    def effective_max_speculative_num_draft_tokens(self) -> Optional[int]:
-        """Return the maximum draft-token count runtime speculative decoding may use."""
+    @cached_property
+    def max_speculative_num_draft_tokens(self) -> Optional[int]:
+        """Return the maximum draft-token count speculative decoding may use."""
         if self.speculative_num_draft_tokens is None:
             return None
         if not self.speculative_adaptive: