[platform] allow the out-of-tree device to set executor class

vllm/engine/async_llm_engine.py

@@ -29,13 +29,14 @@
     get_guided_decoding_logits_processor)
 from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.outputs import PoolingRequestOutput, RequestOutput
+from vllm.platforms import current_platform
 from vllm.pooling_params import PoolingParams
 from vllm.prompt_adapter.request import PromptAdapterRequest
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import ExecuteModelRequest
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 from vllm.usage.usage_lib import UsageContext
-from vllm.utils import deprecate_kwargs, weak_bind
+from vllm.utils import deprecate_kwargs, resolve_obj_by_qualname, weak_bind
 
 logger = init_logger(__name__)
 ENGINE_ITERATION_TIMEOUT_S = envs.VLLM_ENGINE_ITERATION_TIMEOUT_S
@@ -630,6 +631,11 @@ def _get_executor_cls(
                     "distributed_executor_backend must be a subclass of "
                     f"ExecutorAsyncBase. Got {distributed_executor_backend}.")
             executor_class = distributed_executor_backend
+        elif current_platform.is_out_of_tree():
+            executor_cls = current_platform.get_executor_cls(
+                distributed_executor_backend=distributed_executor_backend,
+                is_async=True)
+            executor_class = resolve_obj_by_qualname(executor_cls)
         elif engine_config.device_config.device_type == "neuron":
             from vllm.executor.neuron_executor import NeuronExecutorAsync
             executor_class = NeuronExecutorAsync

vllm/engine/llm_engine.py

@@ -43,6 +43,7 @@
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
 from vllm.outputs import (PoolingRequestOutput, RequestOutput,
                           RequestOutputFactory)
+from vllm.platforms import current_platform
 from vllm.pooling_params import PoolingParams
 from vllm.prompt_adapter.request import PromptAdapterRequest
 from vllm.sampling_params import RequestOutputKind, SamplingParams
@@ -58,7 +59,8 @@
     BaseTokenizerGroup, init_tokenizer_from_configs)
 from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
                                   usage_message)
-from vllm.utils import Counter, Device, deprecate_kwargs, weak_bind
+from vllm.utils import (Counter, Device, deprecate_kwargs,
+                        resolve_obj_by_qualname, weak_bind)
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger(__name__)
@@ -445,6 +447,10 @@ def _get_executor_cls(cls,
             if distributed_executor_backend.uses_ray:  # type: ignore
                 initialize_ray_cluster(engine_config.parallel_config)
             executor_class = distributed_executor_backend
+        elif current_platform.is_out_of_tree():
+            executor_cls = current_platform.get_executor_cls(
+                distributed_executor_backend)
+            executor_class = resolve_obj_by_qualname(executor_cls)
         elif engine_config.device_config.device_type == "neuron":
             from vllm.executor.neuron_executor import NeuronExecutor
             executor_class = NeuronExecutor

vllm/platforms/interface.py

@@ -283,6 +283,13 @@ def get_punica_wrapper(cls) -> str:
         """
         raise NotImplementedError
 
+    @classmethod
+    def get_executor_cls(cls,
+                         distributed_executor_backend: Optional[str] = None,
+                         is_async: Optional[bool] = None) -> str:
+        """Get the executor class of the out-of-tree device."""
+        return ""
+
 
 class UnspecifiedPlatform(Platform):
     _enum = PlatformEnum.UNSPECIFIED