[Serve LLM] Fix enable_log_requests=False not being forwarded to vLLM AsyncLLM

python/ray/llm/_internal/serve/engines/vllm/vllm_engine.py

@@ -475,6 +475,7 @@ def _start_async_llm_engine(
         engine_client = AsyncLLM(
             vllm_config=vllm_engine_config,
             executor_class=executor_class,
+            log_requests=vllm_engine_args.enable_log_requests,
             log_stats=not vllm_engine_args.disable_log_stats,
             stat_loggers=custom_stat_loggers,
         )

python/ray/llm/_internal/serve/engines/vllm/vllm_models.py

@@ -147,15 +147,7 @@ def get_initialization_kwargs(self) -> dict:
             # For GPU mode, use "ray" backend (default)
             engine_kwargs["distributed_executor_backend"] = "ray"
 
-        # TODO (Nikhil): Remove this once vLLM fully deprecates disable_log_requests.
-        if "disable_log_requests" in engine_kwargs:
-            logger.warning(
-                "disable_log_requests is set in engine_kwargs, but vLLM does not support it. Converting to enable_log_requests."
-            )
-            engine_kwargs["enable_log_requests"] = not engine_kwargs.pop(
-                "disable_log_requests"
-            )
-        elif "enable_log_requests" not in engine_kwargs:
+        if "enable_log_requests" not in engine_kwargs:
             engine_kwargs["enable_log_requests"] = False
 
         return engine_kwargs

python/ray/serve/_private/logging_utils.py

@@ -337,6 +337,8 @@ def configure_component_logger(
         stream_handler.setFormatter(serve_formatter)
         stream_handler.addFilter(log_to_stderr_filter)
         stream_handler.addFilter(ServeContextFilter())
+        if logging_config.enable_access_log is False:
+            stream_handler.addFilter(log_access_log_filter)
         logger.addHandler(stream_handler)
 
     # Skip setting up file handler and stdout/stderr redirect if `stream_handler_only`

python/ray/serve/tests/test_logging.py

@@ -944,6 +944,45 @@ def __call__(self, req: starlette.requests.Request):
             with pytest.raises(AssertionError):
                 check_log_file(resp["logs_path"], [".*model_not_show.*"])
 
+    def test_access_log_suppressed_in_stderr(self, serve_and_ray_shutdown):
+        """Test that enable_access_log=False suppresses access logs in stderr.
+
+        Regression test: the log_access_log_filter was previously only applied
+        to the file handler (memory_handler), not the stream handler (stderr).
+        Since RAY_SERVE_LOG_TO_STDERR defaults to True, access logs appeared in
+        stderr even when enable_access_log=False.
+        """
+        logger = logging.getLogger("ray.serve")
+
+        @serve.deployment(
+            logging_config={"enable_access_log": False},
+        )
+        class Model:
+            def __call__(self, req: starlette.requests.Request):
+                logger.info("user_log_should_appear")
+                return "ok"
+
+        serve.run(Model.bind())
+        url = get_application_url(use_localhost=True)
+
+        f = io.StringIO()
+        with redirect_stderr(f):
+            for _ in range(5):
+                resp = httpx.get(url)
+                assert resp.status_code == 200
+
+            # Give logs time to flush.
+            time.sleep(2)
+
+            stderr_output = f.getvalue()
+
+            # Normal user logs should still appear in stderr.
+            assert "user_log_should_appear" in stderr_output
+
+            # Access logs (replica-side "CALL" and proxy-side "GET /") should
+            # NOT appear in stderr when enable_access_log=False.
+            assert "CALL __call__ OK" not in stderr_output
+
     @pytest.mark.parametrize("encoding_type", ["TEXT", "JSON"])
     def test_additional_log_standard_attrs(self, serve_and_ray_shutdown, encoding_type):
         """Test additional log standard attrs"""