PD deployment support without router (#7412)

Author: juncaipeng

examples/splitwise/start_v0_tp1.sh

@@ -2010,13 +2010,13 @@ def __init__(
             and self.router_config
             and self.router_config.router
         ):
-            # For RL scenario: version.yaml will be required for models in future releases.
+            # For RL scenario, version.yaml is required for models
             # Temporarily enforce use router to be enabled.
             self.model_config.read_model_version()
 
         self.read_from_config()
         self.postprocess()
-        self.init_cache_info()
+        self.init_pd_info()
         if test_mode:
             return
         self.check()
@@ -2371,18 +2371,17 @@ def print(self):
                 logger.info("{:<20}:{:<6}{}".format(k, "", v))
         logger.info("=============================================================")
 
-    def init_cache_info(self):
+    def init_pd_info(self):
         """
-        initialize cache info
+        initialize info for pd deployment
         """
-        # TODO: group the splitiwse params
         # There are two methods for splitwise deployment:
         # 1. v0 splitwise_scheduler or dp_scheduler
-        # 2. v1 local_scheduler + router
+        # 2. v1 local_scheduler + router (optional)
         self.splitwise_version = None
         if self.scheduler_config.name in ("splitwise", "dp"):
             self.splitwise_version = "v0"
-        elif self.scheduler_config.name == "local" and self.router_config and self.router_config.router:
+        elif self.scheduler_config.name == "local":
             self.splitwise_version = "v1"
 
         # the information for registering this server to router or splitwise_scheduler

fastdeploy/config.py

@@ -600,10 +600,15 @@ def __post_init__(self):
                 raise NotImplementedError("Only ENABLE_V1_KVCACHE_SCHEDULER=1 support max_logprobs=-1")
 
         if self.splitwise_role != "mixed":
-            if self.scheduler_name == "local" and self.router is None:
+            if self.scheduler_name == "splitwise":
                 raise ValueError(
-                    f"When using {self.splitwise_role} role and the {self.scheduler_name} "
-                    f"scheduler, please provide --router argument."
+                    "Setting scheduler_name as splitwise is not supported in pd deployment, "
+                    "please use router as scheduler."
+                )
+            if self.scheduler_name == "local" and self.router is None:
+                console_logger.warning(
+                    f"Running {self.splitwise_role} role with {self.scheduler_name} "
+                    f"scheduler without --router. Router registration and request routing will be disabled."
                 )
 
         if not (

fastdeploy/engine/args_utils.py

@@ -109,7 +109,7 @@ def start(self, ipc_signal_suffix, local_data_parallel_id):
         if envs.FD_ENABLE_RETURN_TEXT:
             self.engine.create_data_processor()
         if self.cfg.scheduler_config.name == "dp":
-            self.cfg.init_cache_info()
+            self.cfg.init_pd_info()
             self.engine.scheduler.start(local_data_parallel_id)
 
         if ipc_signal_suffix is not None:
@@ -122,7 +122,7 @@ def start(self, ipc_signal_suffix, local_data_parallel_id):
         self.llm_logger.info(f"start expert service {local_data_parallel_id}")
 
         if self.cfg.scheduler_config.name == "splitwise":
-            self.cfg.init_cache_info()
+            self.cfg.init_pd_info()
             role = self.cfg.scheduler_config.splitwise_role
             host_ip = self.cfg.host_ip
             self.engine.scheduler.start(role, host_ip, self.cfg.register_info)