Ensure metrics are logged regardless of requests (#2347)

vllm/engine/async_llm_engine.py

@@ -506,3 +506,9 @@ class AsyncLLMEngine:
                      max_log_len=engine_args.max_log_len,
                      start_engine_loop=start_engine_loop)
         return engine
+
+    async def do_log_stats(self) -> None:
+        if self.engine_use_ray:
+            await self.engine.do_log_stats.remote()
+        else:
+            self.engine.do_log_stats()

vllm/engine/llm_engine.py

@@ -641,6 +641,9 @@ class LLMEngine:
 
         return self._process_model_outputs(output, scheduler_outputs)
 
+    def do_log_stats(self) -> None:
+        self._log_system_stats(False, 0)
+
     def _log_system_stats(
         self,
         prompt_run: bool,