[PD] Fix multi-tokenizer disaggregation metrics labels (#30412)

Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/managers/multi_tokenizer_mixin.py

@@ -36,7 +36,7 @@ import setproctitle
 import zmq
 import zmq.asyncio
 
-from sglang.srt.disaggregation.utils import DisaggregationMode, TransferBackend
+from sglang.srt.disaggregation.utils import TransferBackend
 from sglang.srt.managers.disagg_service import start_disagg_service
 from sglang.srt.managers.io_struct import (
     BaseBatchReq,
@@ -634,24 +634,16 @@ class TokenizerWorker(TokenizerManager):
         import torch
 
         torch.set_num_threads(1)
-        # prevent init prefill bootstrapserver again
-        disaggregation_mode = server_args.disaggregation_mode
-        server_args.override(
-            "tokenizer_worker.suppress_bootstrap", disaggregation_mode="null"
+        super().__init__(
+            server_args,
+            port_args,
+            start_pd_bootstrap_service=False,
         )
-        super().__init__(server_args, port_args)
 
         self.worker_id = os.getpid()
         self.tokenizer_ipc_name = port_args.tokenizer_ipc_name
 
-        # For PD disaggregtion
-        self.server_args.override(
-            "tokenizer_worker.restore_disaggregation_mode",
-            disaggregation_mode=disaggregation_mode,
-        )
-        self.disaggregation_mode = DisaggregationMode(
-            self.server_args.disaggregation_mode
-        )
+        # For PD disaggregation
         self.disaggregation_transfer_backend = TransferBackend(
             self.server_args.disaggregation_transfer_backend
         )

python/sglang/srt/managers/tokenizer_manager.py

@@ -279,6 +279,8 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         self,
         server_args: ServerArgs,
         port_args: PortArgs,
+        *,
+        start_pd_bootstrap_service: bool = True,
     ):
         # Parse args
         self.server_args = server_args
@@ -318,7 +320,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         self.init_lora()
 
         # Init PD disaggregation and encoder disaggregation
-        self.init_disaggregation()
+        self.init_disaggregation(start_pd_bootstrap_service=start_pd_bootstrap_service)
 
         # Init metric collector and watchdog
         self.init_metric_collector_watchdog()
@@ -522,13 +524,17 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             for lora_ref in self.server_args.lora_paths:
                 self.lora_ref_cache[lora_ref.lora_name] = lora_ref
 
-    def init_disaggregation(self):
+    def init_disaggregation(self, *, start_pd_bootstrap_service: bool = True):
         # PD Disaggregation
         self.disaggregation_mode = DisaggregationMode(
             self.server_args.disaggregation_mode
         )
         # Keep a reference so the bootstrap server is not garbage-collected.
-        self.bootstrap_server = start_disagg_service(self.server_args)
+        self.bootstrap_server = (
+            start_disagg_service(self.server_args)
+            if start_pd_bootstrap_service
+            else None
+        )
         # Single-source counter for auto-assigning fake bootstrap_room.
         self.fake_bootstrap_room_counter = 0