Support Multi Process Tokenizer Manager(#6555) (#8964)

Signed-off-by: ybyang <ybyang7@iflytek.com>
Signed-off-by: huanglong <huanglong@linux.alibaba.com>
Co-authored-by: Huang Long <121648372+LLLL114@users.noreply.github.com>
Co-authored-by: huanglong <huanglong@linux.alibaba.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>

python/sglang/srt/managers/scheduler.py

@@ -84,6 +84,8 @@ from sglang.srt.managers.io_struct import (
     InitWeightsUpdateGroupReqInput,
     LoadLoRAAdapterReqInput,
     LoadLoRAAdapterReqOutput,
+    MultiTokenizerRegisterReq,
+    MultiTokenizerWarpper,
     OpenSessionReqInput,
     OpenSessionReqOutput,
     ProfileReq,
@@ -257,7 +259,6 @@ class Scheduler(
         # Init inter-process communication
         context = zmq.Context(2)
         self.idle_sleeper = None
-
         if self.pp_rank == 0 and self.attn_tp_rank == 0:
             self.recv_from_tokenizer = get_zmq_socket(
                 context, zmq.PULL, port_args.scheduler_input_ipc_name, False
@@ -540,6 +541,7 @@ class Scheduler(
                 (ExpertDistributionReq, self.expert_distribution_handle),
                 (LoadLoRAAdapterReqInput, self.load_lora_adapter),
                 (UnloadLoRAAdapterReqInput, self.unload_lora_adapter),
+                (MultiTokenizerRegisterReq, self.register_multi_tokenizer),
             ]
         )
 
@@ -1101,6 +1103,17 @@ class Scheduler(
                     )
                     self.send_to_tokenizer.send_pyobj(abort_req)
                     continue
+
+            # If it is a MultiTokenizerWarpper, unwrap it and handle the inner request.
+            if isinstance(recv_req, MultiTokenizerWarpper):
+                worker_id = recv_req.worker_id
+                recv_req = recv_req.obj
+                output = self._request_dispatcher(recv_req)
+                if output is not None:
+                    output = MultiTokenizerWarpper(worker_id, output)
+                    self.send_to_tokenizer.send_pyobj(output)
+                continue
+
             output = self._request_dispatcher(recv_req)
             if output is not None:
                 if isinstance(output, RpcReqOutput):
@@ -2474,6 +2487,10 @@ class Scheduler(
         result = self.tp_worker.unload_lora_adapter(recv_req)
         return result
 
+    def register_multi_tokenizer(self, recv_req: MultiTokenizerRegisterReq):
+        self.send_to_detokenizer.send_pyobj(recv_req)
+        return recv_req
+
     def slow_down(self, recv_req: SlowDownReqInput):
         t = recv_req.forward_sleep_time
         if t is not None and t <= 0: