[PerfFix] Avoid separate thread for MP executor shm spin (take 2) (#28319)

Signed-off-by: Nick Hill <nhill@redhat.com>

tests/v1/executor/test_executor.py

@@ -4,6 +4,7 @@
 import asyncio
 import os
 from collections.abc import Callable
+from concurrent.futures import Future
 from typing import Any
 
 import pytest
@@ -27,7 +28,7 @@ class CustomMultiprocExecutor(MultiprocExecutor):
         kwargs: dict | None = None,
         non_block: bool = False,
         unique_reply_rank: int | None = None,
-    ) -> list[Any]:
+    ) -> Any | list[Any] | Future[Any | list[Any]]:
         # Drop marker to show that this was run
         with open(".marker", "w"):
             ...

tests/v1/kv_connector/unit/test_output_aggregator.py

@@ -89,14 +89,12 @@ def test_aggregate_workers_output():
 def test_async_aggregate_workers_output():
     aggregator = KVOutputAggregator(expected_finished_count=2)
 
-    future1: Future[DummyModelRunnerOutput] = Future()
+    future: Future[list[DummyModelRunnerOutput]] = Future()
-    future2: Future[DummyModelRunnerOutput] = Future()
+    result_future = aggregator.async_aggregate(future)
-    result_future = aggregator.async_aggregate([future1, future2])
 
     output1 = DummyModelRunnerOutput()
     output2 = DummyModelRunnerOutput()
-    future1.set_result(output1)
+    future.set_result([output1, output2])
-    future2.set_result(output2)
 
     assert result_future.done()
     aggregated = result_future.result()
@@ -106,16 +104,14 @@ def test_async_aggregate_workers_output():
     assert aggregated.finished_recving is None
     assert not aggregated.invalid_block_ids
 
-    future1 = Future()
+    future = Future()
-    future2 = Future()
+    result_future = aggregator.async_aggregate(future)
-    result_future = aggregator.async_aggregate([future1, future2])
 
     output1 = DummyModelRunnerOutput(
         finished_sending={"req1"}, finished_recving={"req2"}
     )
     output2 = DummyModelRunnerOutput(invalid_block_ids={1})
-    future1.set_result(output1)
+    future.set_result([output1, output2])
-    future2.set_result(output2)
 
     assert result_future.done()
     aggregated = result_future.result()
@@ -125,14 +121,12 @@ def test_async_aggregate_workers_output():
     assert aggregated.finished_recving is None
     assert aggregated.invalid_block_ids == {1}
 
-    future1 = Future()
+    future = Future()
-    future2 = Future()
+    result_future = aggregator.async_aggregate(future)
-    result_future = aggregator.async_aggregate([future1, future2])
 
     output1 = DummyModelRunnerOutput(invalid_block_ids={2})
     output2 = DummyModelRunnerOutput(finished_sending={"req1"})
-    future1.set_result(output1)
+    future.set_result([output1, output2])
-    future2.set_result(output2)
 
     assert result_future.done()
     aggregated = result_future.result()
@@ -142,16 +136,14 @@ def test_async_aggregate_workers_output():
     assert aggregated.finished_recving is None
     assert aggregated.invalid_block_ids == {2}
 
-    future1 = Future()
+    future = Future()
-    future2 = Future()
+    result_future = aggregator.async_aggregate(future)
-    result_future = aggregator.async_aggregate([future1, future2])
 
     output1 = DummyModelRunnerOutput(invalid_block_ids={3, 4})
     output2 = DummyModelRunnerOutput(
         finished_recving={"req2"}, invalid_block_ids={4, 5}
     )
-    future1.set_result(output1)
+    future.set_result([output1, output2])
-    future2.set_result(output2)
 
     assert result_future.done()
     aggregated = result_future.result()

vllm/distributed/kv_transfer/kv_connector/utils.py

@@ -4,6 +4,7 @@
 KV cache helper for store.
 """
 
+import contextlib
 from collections.abc import Sequence
 from concurrent.futures import CancelledError, Future
 from typing import TYPE_CHECKING, Literal
@@ -221,38 +222,38 @@ class KVOutputAggregator:
 
     def async_aggregate(
         self,
-        output_futures: Sequence[Future[ModelRunnerOutput | None]],
+        output_future: Future[Sequence[ModelRunnerOutput | None]],
         output_rank: int = 0,
     ) -> Future[ModelRunnerOutput | None]:
-        """Takes a list of futures and returns a single future which resolves
+        """Takes a future that resolves to a list of outputs and returns a future
-        to the respective list of outputs."""
+        which resolves to a single aggregated output."""
         result_future: Future[ModelRunnerOutput | None] = Future()
 
-        outputs: list[ModelRunnerOutput | None] = [None] * len(output_futures)
+        def callback(fut):
-        remaining = len(output_futures)
+            if result_future.done():
+                return
+            try:
+                result_future.set_result(self.aggregate(fut.result(), output_rank))
+            except CancelledError:
+                result_future.cancel()
+            except Exception as e:
+                result_future.set_exception(e)
 
-        def make_callback(idx):
+        output_future.add_done_callback(callback)
-            def callback(fut):
-                if result_future.done():
-                    return
 
-                try:
+        from vllm.v1.executor.multiproc_executor import FutureWrapper
-                    outputs[idx] = fut.result()
-                except CancelledError:
-                    result_future.cancel()
-                except Exception as e:
-                    result_future.set_exception(e)
 
-                # this check assumes io_thread_pool uses a single thread
+        if isinstance(output_future, FutureWrapper):
-                nonlocal remaining
+            # Due to the threadless implementation of multiproc FutureWrapper,
-                remaining -= 1
+            # we must block on the delegate future's result() method.
-                if not remaining:
+            delegate_result = result_future.result
-                    result_future.set_result(self.aggregate(outputs, output_rank))
 
-            return callback
+            def result(timeout=None):
+                with contextlib.suppress(Exception):
+                    output_future.result(timeout=timeout)
+                return delegate_result()
 
-        for i, output_future in enumerate(output_futures):
+            result_future.result = result  # type: ignore[method-assign]
-            output_future.add_done_callback(make_callback(i))
 
         return result_future
 

vllm/v1/executor/abstract.py

@@ -171,7 +171,7 @@ class Executor(ABC):
         args: tuple = (),
         kwargs: dict | None = None,
         non_block: Literal[True] = True,
-    ) -> list[Future[_R]]:
+    ) -> Future[list[_R]]:
         pass
 
     @abstractmethod
@@ -219,7 +219,7 @@ class Executor(ABC):
 
     def sample_tokens(
         self, grammar_output: GrammarOutput | None, non_block: bool = False
-    ) -> ModelRunnerOutput | Future[ModelRunnerOutput]:
+    ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
         output = self.collective_rpc(  # type: ignore[call-overload]
             "sample_tokens", args=(grammar_output,), non_block=non_block
         )

vllm/v1/executor/multiproc_executor.py

@@ -9,8 +9,10 @@ import threading
 import time
 import traceback
 import weakref
+from collections import deque
 from collections.abc import Callable
-from concurrent.futures import Future, ThreadPoolExecutor
+from concurrent.futures import Future, InvalidStateError
+from contextlib import suppress
 from dataclasses import dataclass
 from enum import Enum, auto
 from functools import cached_property, partial
@@ -54,6 +56,30 @@ from vllm.v1.worker.worker_base import WorkerWrapperBase
 logger = init_logger(__name__)
 
 
+class FutureWrapper(Future):
+    def __init__(self, futures_queue: deque[tuple["FutureWrapper", Callable]]):
+        self.futures_queue = futures_queue
+        super().__init__()
+
+    def result(self, timeout=None):
+        if timeout is not None:
+            raise RuntimeError("timeout not implemented")
+        # Drain any futures ahead of us in the queue.
+        while not self.done():
+            future, get_response = self.futures_queue.pop()
+            future.wait_for_response(get_response)
+        return super().result()
+
+    def wait_for_response(self, get_response: Callable):
+        try:
+            response = get_response()
+            with suppress(InvalidStateError):
+                self.set_result(response)
+        except Exception as e:
+            with suppress(InvalidStateError):
+                self.set_exception(e)
+
+
 class MultiprocExecutor(Executor):
     supports_pp: bool = True
 
@@ -64,7 +90,6 @@ class MultiprocExecutor(Executor):
         self.is_failed = False
         self.shutdown_event = threading.Event()
         self.failure_callback: FailureCallback | None = None
-        self.io_thread_pool: ThreadPoolExecutor | None = None
 
         self.world_size = self.parallel_config.world_size
         tensor_parallel_size = self.parallel_config.tensor_parallel_size
@@ -132,12 +157,7 @@ class MultiprocExecutor(Executor):
                         uw.death_writer.close()
                 self._ensure_worker_termination([uw.proc for uw in unready_workers])
 
-        # Note: must use only 1 IO thread to keep dequeue sequence
+        self.futures_queue = deque[tuple[FutureWrapper, Callable]]()
-        # from the response queue.
-        # _async_aggregate_workers_output also assumes a single IO thread.
-        self.io_thread_pool = ThreadPoolExecutor(
-            max_workers=1, thread_name_prefix="mp_exec_io"
-        )
 
         self.output_rank = self._get_output_rank()
         self.has_connector = self.vllm_config.kv_transfer_config is not None
@@ -195,14 +215,13 @@ class MultiprocExecutor(Executor):
     ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
         if not self.has_connector:
             # get output only from a single worker (output_rank)
-            (output,) = self.collective_rpc(
+            return self.collective_rpc(
                 method,
                 args=args,
                 unique_reply_rank=self.output_rank,
                 non_block=non_block,
                 timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS,
             )
-            return output
 
         # get output from all workers
         outputs = self.collective_rpc(
@@ -223,12 +242,11 @@ class MultiprocExecutor(Executor):
 
     def take_draft_token_ids(self) -> DraftTokenIds | None:
         # OPTIMIZATION: Get output only from a single worker (output_rank)
-        outputs = self.collective_rpc(
+        return self.collective_rpc(
             "take_draft_token_ids", unique_reply_rank=self.output_rank
         )
-        return outputs[0]
 
-    def collective_rpc(
+    def collective_rpc(  # type: ignore[override]
         self,
         method: str | Callable,
         timeout: float | None = None,
@@ -236,7 +254,9 @@ class MultiprocExecutor(Executor):
         kwargs: dict | None = None,
         non_block: bool = False,
         unique_reply_rank: int | None = None,
-    ) -> list[Any]:
+    ) -> Any | list[Any] | Future[Any | list[Any]]:
+        """Returns single result if unique_reply_rank is provided, otherwise list."""
+
         if self.is_failed:
             raise RuntimeError("Executor failed.")
 
@@ -246,63 +266,52 @@ class MultiprocExecutor(Executor):
         # NOTE: If the args are heterogeneous, then we pack them into a list,
         # and unpack them in the method of every worker, because every worker
         # knows their own rank.
-        try:
-            if isinstance(method, str):
-                send_method = method
-            else:
-                send_method = cloudpickle.dumps(
-                    method, protocol=pickle.HIGHEST_PROTOCOL
-                )
-            self.rpc_broadcast_mq.enqueue(
-                (send_method, args, kwargs, unique_reply_rank)
-            )
 
-            workers = (
+        if isinstance(method, str):
-                (self.workers[unique_reply_rank],)
+            send_method = method
-                if unique_reply_rank is not None
+        else:
-                else self.workers
+            send_method = cloudpickle.dumps(method, protocol=pickle.HIGHEST_PROTOCOL)
-            )
+        self.rpc_broadcast_mq.enqueue((send_method, args, kwargs, unique_reply_rank))
+
+        workers = (
+            (self.workers[unique_reply_rank],)
+            if unique_reply_rank is not None
+            else self.workers
+        )
+
+        shutdown_event = self.shutdown_event
+
+        def get_response():
             responses = []
-
+            for w in workers:
-            def get_response(
+                dequeue_timeout = (
-                w: WorkerProcHandle,
+                    None if deadline is None else (deadline - time.monotonic())
-                dequeue_timeout: float | None = None,
-                cancel_event: threading.Event | None = None,
-            ):
-                status, result = w.worker_response_mq.dequeue(
-                    timeout=dequeue_timeout, cancel=cancel_event
                 )
-
+                try:
+                    status, result = w.worker_response_mq.dequeue(
+                        timeout=dequeue_timeout, cancel=shutdown_event
+                    )
+                except TimeoutError as e:
+                    raise TimeoutError(f"RPC call to {method} timed out.") from e
                 if status != WorkerProc.ResponseStatus.SUCCESS:
                     raise RuntimeError(
                         f"Worker failed with error '{result}', please check the"
                         " stack trace above for the root cause"
                     )
-                return result
-
-            for w in workers:
-                dequeue_timeout = (
-                    None if deadline is None else (deadline - time.monotonic())
-                )
-
-                if self.io_thread_pool is not None:
-                    # We must consume worker_response_mq from a single thread.
-                    result = self.io_thread_pool.submit(  # type: ignore
-                        get_response, w, dequeue_timeout, self.shutdown_event
-                    )
-                    if not non_block:
-                        result = result.result()
-                elif not non_block:
-                    result = get_response(w, dequeue_timeout, self.shutdown_event)
-                else:
-                    raise RuntimeError(
-                        "non_block can only be used when max_concurrent_batches > 1"
-                    )
                 responses.append(result)
+            return responses[0] if unique_reply_rank is not None else responses
 
-            return responses
+        if non_block:
-        except TimeoutError as e:
+            future = FutureWrapper(self.futures_queue)
-            raise TimeoutError(f"RPC call to {method} timed out.") from e
+            self.futures_queue.appendleft((future, get_response))
+            return future
+
+        # First drain any pending futures in the queue.
+        while self.futures_queue:
+            future, get_fut_response = self.futures_queue.pop()
+            future.wait_for_response(get_fut_response)
+
+        return get_response()
 
     @staticmethod
     def _ensure_worker_termination(worker_procs: list[BaseProcess]):
@@ -348,9 +357,6 @@ class MultiprocExecutor(Executor):
                 self._ensure_worker_termination([w.proc for w in workers])
 
             self.shutdown_event.set()
-            if self.io_thread_pool is not None:
-                self.io_thread_pool.shutdown(wait=False, cancel_futures=True)
-                del self.io_thread_pool
 
         self.rpc_broadcast_mq = None
 

vllm/v1/executor/ray_executor.py

@@ -435,26 +435,25 @@ class RayDistributedExecutor(Executor):
 
             # When PP is used, we return a FutureWrapper immediately so that
             # the scheduler can yield to the next batch.
-            return FutureWrapper(refs)
+            return FutureWrapper(refs[0])
 
         # Get output from all workers when connector is present
         assert self.kv_output_aggregator is not None
         if not non_block:
             # Block and get results from all workers
-            outputs = [ref.get() for ref in refs]
+            return self.kv_output_aggregator.aggregate(ray.get(refs))
-            return self.kv_output_aggregator.aggregate(outputs)
 
         # Return a future that will aggregate outputs from all workers
         return FutureWrapper(refs, self.kv_output_aggregator)
 
-    def collective_rpc(
+    def collective_rpc(  # type: ignore[override]
         self,
         method: str | Callable,
         timeout: float | None = None,
         args: tuple = (),
         kwargs: dict[str, Any] | None = None,
         non_block: bool = False,
-    ) -> list[Any]:
+    ) -> list[Any] | Future[list[Any]]:
         """Runs the given method on all workers."""
         sent_method = method if isinstance(method, str) else cloudpickle.dumps(method)
         del method
@@ -470,7 +469,7 @@ class RayDistributedExecutor(Executor):
 
         # Get the results of the ray workers.
         if non_block:
-            return [FutureWrapper((output,)) for output in ray_worker_outputs]
+            return FutureWrapper(ray_worker_outputs)
 
         return ray.get(ray_worker_outputs, timeout=timeout)
 

vllm/v1/executor/ray_utils.py

@@ -141,19 +141,16 @@ class FutureWrapper(Future):
     the result() call. If not only the first worker's output is returned.
     """
 
-    def __init__(self, refs, aggregator: KVOutputAggregator | None = None):
+    def __init__(self, ref_or_refs, aggregator: KVOutputAggregator | None = None):
         super().__init__()
-        self.refs = refs
+        self.ref_or_refs = ref_or_refs
         self.aggregator = aggregator
 
     def result(self, timeout=None):
-        if timeout is not None:
+        outputs = ray.get(self.ref_or_refs, timeout=timeout)
-            raise NotImplementedError("timeout is not supported")
-
         if self.aggregator is None:
-            return self.refs[0].get()
+            return outputs
 
-        outputs = [ref.get() for ref in self.refs]
         return self.aggregator.aggregate(outputs, output_rank=0)
 
 

vllm/v1/executor/uniproc_executor.py

@@ -13,9 +13,10 @@ import torch.distributed as dist
 import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.utils.network_utils import get_distributed_init_method, get_ip, get_open_port
+from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
 from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
 from vllm.v1.executor.abstract import Executor
-from vllm.v1.outputs import AsyncModelRunnerOutput
+from vllm.v1.outputs import AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput
 from vllm.v1.serial_utils import run_method
 from vllm.v1.worker.worker_base import WorkerWrapperBase
 
@@ -58,32 +59,60 @@ class UniProcExecutor(Executor):
     def max_concurrent_batches(self) -> int:
         return 2 if self.scheduler_config.async_scheduling else 1
 
-    def collective_rpc(
+    def collective_rpc(  # type: ignore[override]
         self,
         method: str | Callable,
         timeout: float | None = None,
         args: tuple = (),
         kwargs: dict | None = None,
         non_block: bool = False,
-    ) -> list[Any]:
+        single_value: bool = False,
+    ) -> Any | list[Any] | Future[Any | list[Any]]:
         if kwargs is None:
             kwargs = {}
 
         if not non_block:
-            return [run_method(self.driver_worker, method, args, kwargs)]
+            result = run_method(self.driver_worker, method, args, kwargs)
+            return result if single_value else [result]
 
         try:
             result = run_method(self.driver_worker, method, args, kwargs)
             if isinstance(result, AsyncModelRunnerOutput):
                 if (async_thread := self.async_output_thread) is not None:
-                    return [async_thread.submit(result.get_output)]
+                    get_output = result.get_output
+                    if not single_value:
+                        get_output = lambda go=result.get_output: [go()]
+                    return async_thread.submit(get_output)
                 result = result.get_output()
             future = Future[Any]()
-            future.set_result(result)
+            future.set_result(result if single_value else [result])
         except Exception as e:
             future = Future[Any]()
             future.set_exception(e)
-        return [future]
+        return future
+
+    def execute_model(  # type: ignore[override]
+        self, scheduler_output: SchedulerOutput, non_block: bool = False
+    ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
+        return self.collective_rpc(
+            "execute_model",
+            args=(scheduler_output,),
+            non_block=non_block,
+            single_value=True,
+        )
+
+    def sample_tokens(  # type: ignore[override]
+        self, grammar_output: GrammarOutput | None, non_block: bool = False
+    ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
+        return self.collective_rpc(
+            "sample_tokens",
+            args=(grammar_output,),
+            non_block=non_block,
+            single_value=True,
+        )
+
+    def take_draft_token_ids(self) -> DraftTokenIds | None:
+        return self.collective_rpc("take_draft_token_ids", single_value=True)
 
     def check_health(self) -> None:
         # UniProcExecutor will always be healthy as long as

vllm/v1/worker/gpu_worker.py

@@ -524,7 +524,7 @@ class Worker(WorkerBase):
 
     @torch.inference_mode()
     def sample_tokens(
-        self, grammar_output: "GrammarOutput"
+        self, grammar_output: "GrammarOutput | None"
     ) -> ModelRunnerOutput | AsyncModelRunnerOutput:
         return self.model_runner.sample_tokens(grammar_output)