init

vllm/executor/gpu_executor.py

@@ -0,0 +1,150 @@
+from typing import Any, Dict, List, Optional, Set, Tuple
+
+from vllm.executor.executor_base import ExecutorAsyncBase, ExecutorBase
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.sequence import ExecuteModelRequest, SamplerOutput
+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__)
+
+
+class GPUExecutor(ExecutorBase):
+
+    def _init_executor(self) -> None:
+        """Initialize the worker and load the model.
+
+        If speculative decoding is enabled, we instead create the speculative
+        worker.
+        """
+        if self.speculative_config is None:
+            self._init_non_spec_worker()
+        else:
+            self._init_spec_worker()
+
+    def _get_worker_kwargs(
+            self,
+            local_rank: int = 0,
+            rank: int = 0,
+            distributed_init_method: Optional[str] = None) -> Dict[str, Any]:
+        """Return worker init args for a given rank."""
+        if distributed_init_method is None:
+            distributed_init_method = get_distributed_init_method(
+                get_ip(), get_open_port())
+        return dict(
+            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=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,
+        )
+
+    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()
+
+    def _init_spec_worker(self):
+        """Initialize a SpecDecodeWorker, using a draft model for proposals.
+        """
+        assert self.speculative_config is not None
+
+        from vllm.spec_decode.spec_decode_worker import SpecDecodeWorker
+
+        target_worker = self._create_worker()
+
+        draft_worker_kwargs = self._get_worker_kwargs()
+        # Override draft-model specific worker args.
+        draft_worker_kwargs.update(
+            model_config=self.speculative_config.draft_model_config,
+            parallel_config=self.speculative_config.draft_parallel_config,
+            # TODO allow draft-model specific load config.
+            #load_config=self.load_config,
+        )
+
+        spec_decode_worker = SpecDecodeWorker.create_worker(
+            scorer_worker=target_worker,
+            draft_worker_kwargs=draft_worker_kwargs,
+        )
+
+        assert self.parallel_config.world_size == 1, (
+            "GPUExecutor only supports single GPU.")
+
+        self.driver_worker = spec_decode_worker
+
+        # Load model handled in spec decode worker.
+        self.driver_worker.init_device()
+
+    def determine_num_available_blocks(self) -> Tuple[int, int]:
+        """Determine the number of available KV blocks by invoking the
+        underlying worker.
+        """
+        return self.driver_worker.determine_num_available_blocks()
+
+    def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None:
+        """Initialize the KV cache by invoking the underlying worker.
+        """
+        # NOTE: This is logged in the executor because there can be >1 worker
+        # with other executors. We could log in the engine level, but work
+        # remains to abstract away the device for non-GPU configurations.
+        logger.info("# GPU blocks: %d, # CPU blocks: %d", num_gpu_blocks,
+                    num_cpu_blocks)
+
+        self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks)
+
+    def execute_model(
+            self,
+            execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
+        output = self.driver_worker.execute_model(execute_model_req)
+        return output
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        assert lora_request.lora_int_id > 0, "lora_id must be greater than 0."
+        return self.driver_worker.add_lora(lora_request)
+
+    def remove_lora(self, lora_id: int) -> bool:
+        assert lora_id > 0, "lora_id must be greater than 0."
+        return self.driver_worker.remove_lora(lora_id)
+
+    def list_loras(self) -> Set[int]:
+        return self.driver_worker.list_loras()
+
+    def check_health(self) -> None:
+        # GPUExecutor will always be healthy as long as
+        # it's running.
+        return
+
+
+class GPUExecutorAsync(GPUExecutor, ExecutorAsyncBase):
+
+    async def execute_model_async(
+        self,
+        execute_model_req: ExecuteModelRequest,
+    ) -> List[SamplerOutput]:
+        output = await make_async(self.driver_worker.execute_model
+                                  )(execute_model_req=execute_model_req, )
+        return output