[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: