[Bugfix] Add init_cached_hf_modules to RayWorkerWrapper (#4286)

vllm/executor/ray_gpu_executor.py

@@ -100,6 +100,7 @@ class RayGPUExecutor(ExecutorBase):
             )(RayWorkerWrapper).remote(
                 worker_module_name="vllm.worker.worker",
                 worker_class_name="Worker",
+                trust_remote_code=self.model_config.trust_remote_code,
             )
 
             worker_ip = ray.get(worker.get_node_ip.remote())
@@ -110,6 +111,7 @@ class RayGPUExecutor(ExecutorBase):
                 self.driver_worker = RayWorkerWrapper(
                     worker_module_name="vllm.worker.worker",
                     worker_class_name="Worker",
+                    trust_remote_code=self.model_config.trust_remote_code,
                 )
             else:
                 # Else, added to the list of workers.

vllm/worker/worker_base.py

@@ -103,10 +103,15 @@ class WorkerWrapperBase:
 
     def __init__(self,
                  worker_module_name=None,
-                 worker_class_name=None) -> None:
+                 worker_class_name=None,
+                 trust_remote_code: bool = False) -> None:
         self.worker_module_name = worker_module_name
         self.worker_class_name = worker_class_name
         self.worker = None
+        if trust_remote_code:
+            # note: lazy import to avoid importing torch before initializing
+            from vllm.utils import init_cached_hf_modules
+            init_cached_hf_modules()
 
     @staticmethod
     def update_environment_variables(envs: Dict[str, str]) -> None: