[V1] [Feature] Collective RPC (#15444)

Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>

vllm/engine/llm_engine.py

@@ -7,8 +7,8 @@ from collections import deque
 from contextlib import contextmanager
 from dataclasses import dataclass
 from functools import partial
-from typing import (TYPE_CHECKING, Callable, ClassVar, Deque, Dict, Iterable,
-                    List, Mapping, NamedTuple, Optional)
+from typing import (TYPE_CHECKING, Any, Callable, ClassVar, Deque, Dict,
+                    Iterable, List, Mapping, NamedTuple, Optional)
 from typing import Sequence as GenericSequence
 from typing import Set, Type, Union, cast, overload
 
@@ -67,6 +67,7 @@ _LOCAL_LOGGING_INTERVAL_SEC = 5
 
 _G = TypeVar("_G", bound=BaseTokenizerGroup, default=BaseTokenizerGroup)
 _O = TypeVar("_O", RequestOutput, PoolingRequestOutput)
+_R = TypeVar("_R", default=Any)
 
 
 @dataclass
@@ -2123,6 +2124,14 @@ class LLMEngine:
 
         return sampling_params
 
+    def collective_rpc(self,
+                       method: Union[str, Callable[..., _R]],
+                       timeout: Optional[float] = None,
+                       args: tuple = (),
+                       kwargs: Optional[dict[str, Any]] = None) -> list[_R]:
+        return self.model_executor.collective_rpc(method, timeout, args,
+                                                  kwargs)
+
 
 if envs.is_set("VLLM_USE_V1") and envs.VLLM_USE_V1:
     from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine