Allow engine recovery to be executed in parallel with generation and wait for aliveness

miles/ray/rollout/rollout_manager.py

@@ -186,9 +186,10 @@ async def onload_kv(self):
 
     # -------------------------- engine management -----------------------------
 
-    def get_updatable_engines_and_lock(self):
+    async def get_updatable_engines_and_lock(self):
         """Return engines eligible for weight updates."""
         srv = self._get_updatable_server()
+        await srv.wait_all_engines_alive()
         engines = [e.actor_handle for e in srv.engines] if srv else []
         gpu_counts = srv.engine_gpu_counts if srv else []
         gpu_offsets = srv.engine_gpu_offsets if srv else []

miles/ray/rollout/rollout_server.py

@@ -209,3 +209,10 @@ async def onload(self, tags: list[str] | None = None):
 
     async def check_weights(self, action: str):
         return await asyncio.gather(*[g.check_weights(action=action) for g in self.server_groups])
+
+    async def wait_all_engines_alive(self):
+        while True:
+            if all(e.is_alive for g in self.server_groups for e in g.all_engines):
+                return
+            await asyncio.sleep(2)
+            logger.info("wait_all_engines_alive looping...")

miles/ray/rollout/server_engine.py

@@ -42,6 +42,10 @@ def actor_handle(self) -> ray.actor.ActorHandle:
     def is_allocated(self) -> bool:
         return isinstance(self._state, _StateAllocatedBase)
 
+    @property
+    def is_alive(self) -> bool:
+        return isinstance(self._state, _StateAllocatedAlive)
+
     # TODO: unify w/ trainer `change_state`
     def _change_state(
         self,