[RL] re-abort_request when model_update_lock is still locked (#13338)

python/sglang/srt/managers/tokenizer_manager.py

@@ -1190,7 +1190,14 @@ class TokenizerManager(TokenizerCommunicatorMixin):
     async def pause_generation(self):
         async with self.is_pause_cond:
             self.is_pause = True
-            self.abort_request(abort_all=True)
+            # we are using the model_update_lock to check if there is still on-going requests.
+            while True:
+                # TODO: maybe make it async instead of fire-and-forget
+                self.abort_request(abort_all=True)
+                is_locked = await self.model_update_lock.is_locked()
+                if not is_locked:
+                    break
+                await asyncio.sleep(1.0)
 
     async def continue_generation(self):
         async with self.is_pause_cond:

python/sglang/srt/utils/aio_rwlock.py

@@ -75,6 +75,10 @@ class RWLock:
             # Wake up anyone waiting (readers or writers)
             self._cond.notify_all()
 
+    async def is_locked(self):
+        async with self._lock:
+            return self._writer_active or self._readers > 0
+
 
 class _ReaderLock:
     def __init__(self, rwlock: RWLock):