[Core] Centralize GPU Worker construction (#4419)

vllm/executor/gpu_executor.py

@@ -1,4 +1,4 @@
-from typing import Dict, List, Set, Tuple
+from typing import Any, Dict, List, Optional, Set, Tuple
 
 from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase
 from vllm.logger import init_logger
@@ -6,6 +6,7 @@ from vllm.lora.request import LoRARequest
 from vllm.sequence import SamplerOutput, SequenceGroupMetadata
 from vllm.utils import (get_distributed_init_method, get_ip, get_open_port,
                         make_async)
+from vllm.worker.worker_base import WorkerWrapperBase
 
 logger = init_logger(__name__)
 
@@ -23,30 +24,47 @@ class GPUExecutor(ExecutorBase):
         else:
             self._init_spec_worker()
 
-    def _init_non_spec_worker(self):
+    def _get_worker_kwargs(
-        # Lazy import the Worker to avoid importing torch.cuda/xformers
+            self,
-        # before CUDA_VISIBLE_DEVICES is set in the Worker
+            local_rank: int = 0,
-        from vllm.worker.worker import Worker
+            rank: int = 0,
-
+            distributed_init_method: Optional[str] = None) -> Dict[str, Any]:
-        assert self.parallel_config.world_size == 1, (
+        """Return worker init args for a given rank."""
-            "GPUExecutor only supports single GPU.")
+        if distributed_init_method is None:
-
+            distributed_init_method = get_distributed_init_method(
-        distributed_init_method = get_distributed_init_method(
+                get_ip(), get_open_port())
-            get_ip(), get_open_port())
+        return dict(
-        self.driver_worker = Worker(
             model_config=self.model_config,
             parallel_config=self.parallel_config,
             scheduler_config=self.scheduler_config,
             device_config=self.device_config,
             cache_config=self.cache_config,
             load_config=self.load_config,
-            local_rank=0,
+            local_rank=local_rank,
-            rank=0,
+            rank=rank,
             distributed_init_method=distributed_init_method,
             lora_config=self.lora_config,
             vision_language_config=self.vision_language_config,
-            is_driver_worker=True,
+            is_driver_worker=rank == 0,
         )
+
+    def _create_worker(self,
+                       local_rank: int = 0,
+                       rank: int = 0,
+                       distributed_init_method: Optional[str] = None):
+        wrapper = WorkerWrapperBase(
+            worker_module_name="vllm.worker.worker",
+            worker_class_name="Worker",
+        )
+        wrapper.init_worker(**self._get_worker_kwargs(local_rank, rank,
+                                                      distributed_init_method))
+        return wrapper.worker
+
+    def _init_non_spec_worker(self):
+        assert self.parallel_config.world_size == 1, (
+            "GPUExecutor only supports single GPU.")
+
+        self.driver_worker = self._create_worker()
         self.driver_worker.init_device()
         self.driver_worker.load_model()
 
@@ -57,41 +75,18 @@ class GPUExecutor(ExecutorBase):
 
         from vllm.spec_decode.multi_step_worker import MultiStepWorker
         from vllm.spec_decode.spec_decode_worker import SpecDecodeWorker
-        from vllm.worker.worker import Worker
 
-        distributed_init_method = get_distributed_init_method(
+        target_worker = self._create_worker()
-            get_ip(), get_open_port())
 
-        target_worker = Worker(
+        draft_worker_kwargs = self._get_worker_kwargs()
-            model_config=self.model_config,
+        # Override draft-model specific worker args.
-            parallel_config=self.parallel_config,
+        draft_worker_kwargs.update(
-            scheduler_config=self.scheduler_config,
-            device_config=self.device_config,
-            cache_config=self.cache_config,
-            load_config=self.load_config,
-            local_rank=0,
-            rank=0,
-            distributed_init_method=distributed_init_method,
-            lora_config=self.lora_config,
-            vision_language_config=self.vision_language_config,
-            is_driver_worker=True,
-        )
-
-        draft_worker = MultiStepWorker(
             model_config=self.speculative_config.draft_model_config,
             parallel_config=self.speculative_config.draft_parallel_config,
-            scheduler_config=self.scheduler_config,
-            device_config=self.device_config,
-            cache_config=self.cache_config,
             # TODO allow draft-model specific load config.
-            load_config=self.load_config,
+            #load_config=self.load_config,
-            local_rank=0,
-            rank=0,
-            distributed_init_method=distributed_init_method,
-            lora_config=self.lora_config,
-            vision_language_config=self.vision_language_config,
-            is_driver_worker=True,
         )
+        draft_worker = MultiStepWorker(**draft_worker_kwargs)
 
         spec_decode_worker = SpecDecodeWorker.from_workers(
             proposer_worker=draft_worker, scorer_worker=target_worker)

vllm/executor/ray_gpu_executor.py

@@ -153,29 +153,14 @@ class RayGPUExecutor(DistributedGPUExecutor):
         distributed_init_method = get_distributed_init_method(
             driver_ip, get_open_port())
 
-        def collect_arg_helper_func(**kwargs):
-            # avoid writing `{"name": value}` manually
-            return kwargs
-
         # Initialize the actual workers inside worker wrapper.
-        init_worker_all_kwargs = []
+        init_worker_all_kwargs = [
-        for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids):
+            self._get_worker_kwargs(
-            local_rank = node_workers[node_id].index(rank)
+                local_rank=node_workers[node_id].index(rank),
-            init_worker_all_kwargs.append(
+                rank=rank,
-                collect_arg_helper_func(
+                distributed_init_method=distributed_init_method,
-                    model_config=self.model_config,
+            ) for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids)
-                    parallel_config=self.parallel_config,
+        ]
-                    scheduler_config=self.scheduler_config,
-                    device_config=self.device_config,
-                    cache_config=self.cache_config,
-                    load_config=self.load_config,
-                    local_rank=local_rank,
-                    rank=rank,
-                    distributed_init_method=distributed_init_method,
-                    lora_config=self.lora_config,
-                    vision_language_config=self.vision_language_config,
-                    is_driver_worker=rank == 0,
-                ))
         self._run_workers("init_worker", all_kwargs=init_worker_all_kwargs)
 
         self._run_workers("init_device")
@@ -201,8 +186,7 @@ class RayGPUExecutor(DistributedGPUExecutor):
             use_ray_compiled_dag=USE_RAY_COMPILED_DAG)
 
         # Only the driver worker returns the sampling results.
-        output = all_outputs[0]
+        return all_outputs[0]
-        return output
 
     def _run_workers(
         self,