chore: 迁移服务器前同步本地修改

kilostar/worker_cluster/worker_cluster.py

@@ -15,13 +15,7 @@
 import time
 import asyncio
 from collections import OrderedDict
-from kilostar.utils.ray_compat import actor_class, _STANDALONE
-from kilostar.utils.ray_hook import ray_actor_hook
-
-if _STANDALONE:
-    from asyncio import Queue
-else:
-    from ray.util.queue import Queue
+from asyncio import Queue
 from kilostar.worker_individual.base_individual import BaseIndividual
 from kilostar.worker_individual.skill_individual import SkillIndividual
 from kilostar.worker_individual.ordinary_individual import OrdinaryIndividual
@@ -31,15 +25,14 @@ from kilostar.worker_individual.special_individual import SpecialIndividual
 from kilostar.utils.logger import get_logger
 
 
-@actor_class
 class WorkerCluster:
     """
-    工作集群 Actor：管理和调度所有的 worker_individual
-    设计理念：按需加载，内存 LRU 淘汰，避免 Actor 爆炸
+    工作集群：管理和调度所有的 worker_individual
+    设计理念：按需加载，内存 LRU 淘汰，避免实例爆炸
 
-    分布式模式下每种 node_type 对应一个独立实例，Ray 根据自定义资源
-    ``kilostar_node_cpu`` / ``kilostar_node_core`` / ``kilostar_node_gpu``
-    将 Actor 调度到声明了对应资源的节点上。
+    ``submit_task`` 是对外的稳定执行边界——未来若要把重型 worker（如 vLLM 本地
+    推理）拆到独立进程/机器上，只需在这个边界后面换一个 executor 实现，上层不感知。
+    ``node_type`` 字段暂时保留，供未来按算力亲和性路由使用。
     """
 
     def __init__(self, max_capacity: int = 200, num_runners: int = 10, node_type: str = "cpu"):
@@ -70,11 +63,8 @@ class WorkerCluster:
 
         from kilostar.core.global_state_machine.gsm_snapshot import fetch_snapshot
 
-        global_state_machine = ray_actor_hook(
-            "global_state_machine"
-        ).global_state_machine
-        # 走快照读，避开 GSM actor RPC：高频唤醒路径不再是单 actor 瓶颈
-        snapshot = await fetch_snapshot(gsm_actor=global_state_machine)
+        # 走快照读：高频唤醒路径直接读进程内缓存快照
+        snapshot = await fetch_snapshot()
         agent_config = snapshot.individuals.get(agent_id)
 
         if not agent_config:
@@ -104,7 +94,7 @@ class WorkerCluster:
                 if self.task_queue is None:
                     await asyncio.sleep(0.1)
                     continue
-                task = await self.task_queue.get() if _STANDALONE else await self.task_queue.get_async()
+                task = await self.task_queue.get()
                 task_id = task.get("task_id")
                 agent_id = task.get("agent_id")
                 task_event = task.get("task_event")
@@ -150,10 +140,7 @@ class WorkerCluster:
         self.results_futures[task_id] = future
 
         task = {"task_id": task_id, "agent_id": agent_id, "task_event": task_event}
-        if _STANDALONE:
-            await self.task_queue.put(task)
-        else:
-            await self.task_queue.put_async(task)
+        await self.task_queue.put(task)
         self.logger.debug(f"[WorkerCluster] 任务 {task_id} 已加入队列。")
 
         try:
@@ -168,5 +155,5 @@ class WorkerCluster:
             "active_worker_count": len(self._active_workers),
             "max_capacity": self.max_capacity,
             "cached_agent_ids": list(self._active_workers.keys()),
-            "queue_size": self.task_queue.qsize() if _STANDALONE else self.task_queue.size(),
+            "queue_size": self.task_queue.qsize() if self.task_queue else 0,
         }