[BugFix] Preserve tokenizer worker fanout when skip_tokenizer_init is enabled (#30682)

Co-authored-by: hjzhang <zhanghjzzz@qq.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>

python/sglang/srt/managers/scheduler_components/ipc_channels.py

@@ -53,7 +53,8 @@ class SchedulerIpcChannels:
                 context, zmq.PUSH, port_args.tokenizer_ipc_name, False
             )
             if skip_tokenizer_init:
-                # Directly send to the TokenizerManager
+                # No decode work: send outputs straight to the tokenizer side
+                # (MultiTokenizerRouter fans out when tokenizer_worker_num > 1).
                 send_to_detokenizer_raw = get_zmq_socket(
                     context, zmq.PUSH, port_args.tokenizer_ipc_name, False
                 )

python/sglang/srt/managers/tokenizer_manager.py

@@ -601,6 +601,8 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 (FreezeGCReq, lambda x: None),
                 # For handling case when scheduler skips detokenizer and forwards back to the tokenizer manager, we ignore it.
                 (HealthCheckOutput, lambda x: None),
+                # Same skip-detokenizer forwarding case as above.
+                (ConfigureLoggingReq, lambda x: None),
                 (ActiveRanksOutput, self.update_active_ranks),
             ]
         )

python/sglang/srt/server_args.py

@@ -6089,16 +6089,12 @@ class ServerArgs:
             )
 
         if self.skip_tokenizer_init:
-            if self.tokenizer_worker_num != 1:
-                logger.warning(
-                    "skip_tokenizer_init=True disables tokenizer workers; forcing tokenizer_worker_num=1 "
-                    f"(requested {self.tokenizer_worker_num})."
-                )
-                self.tokenizer_worker_num = 1
+            # Tokenizer workers still serve HTTP / state / output work, so
+            # their fanout is preserved; detokenizer workers only decode.
             if self.detokenizer_worker_num != 1:
                 logger.warning(
-                    "skip_tokenizer_init=True disables detokenizer workers; forcing detokenizer_worker_num=1 "
-                    f"(requested {self.detokenizer_worker_num})."
+                    "skip_tokenizer_init=True leaves no decode work for detokenizer workers; "
+                    f"forcing detokenizer_worker_num=1 (requested {self.detokenizer_worker_num})."
                 )
                 self.detokenizer_worker_num = 1