address comments

Signed-off-by: Ming Yang <minos.future@gmail.com>

Author: minosfuture

vllm/v1/engine/core.py

@@ -607,9 +607,12 @@ def __init__(
 
             self._init_data_parallel(vllm_config)
 
-            super().__init__(
-                vllm_config, executor_class, log_stats, executor_fail_callback
-            )
+            from vllm.config import set_current_vllm_config
+
+            with set_current_vllm_config(vllm_config):
+                super().__init__(
+                    vllm_config, executor_class, log_stats, executor_fail_callback
+                )
 
             # Background Threads and Queues for IO. These enable us to
             # overlap ZMQ socket IO with GPU since they release the GIL,

vllm/v1/worker/gpu_worker.py

@@ -269,9 +269,10 @@ def init_device(self):
     # to hijack tensor allocation.
     def load_model(self) -> None:
         eep_scale_up = os.environ.get("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH") == "1"
-        with self._maybe_get_memory_pool_context(
-            tag="weights"
-        ) and set_current_vllm_config(self.vllm_config):
+        with (
+            self._maybe_get_memory_pool_context(tag="weights"),
+            set_current_vllm_config(self.vllm_config),
+        ):
             self.model_runner.load_model(eep_scale_up=eep_scale_up)
 
     def update_config(self, overrides: dict[str, Any]) -> None: