Clean up metrics code (#1972)

python/sglang/srt/managers/tokenizer_manager.py

@@ -391,8 +391,12 @@ class TokenizerManager:
 
             async with self.model_update_lock:
                 # wait for the previous generation requests to finish
-                while len(self.rid_to_state) > 0:
-                    await asyncio.sleep(0.001)
+                for i in range(3):
+                    while len(self.rid_to_state) > 0:
+                        await asyncio.sleep(0.001)
+                    # FIXME: We add some sleep here to avoid some race conditions.
+                    # We can use a read-write lock as a better fix.
+                    await asyncio.sleep(0.01)
                 self.send_to_scheduler.send_pyobj(obj)
                 self.model_update_result = asyncio.Future()