[V1] AsyncLLM data parallel (#13923)

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

vllm/v1/engine/__init__.py

@@ -128,12 +128,18 @@ class EngineCoreOutputs(
     #NOTE(Nick): We could consider ways to make this more compact,
     # e.g. columnwise layout
 
+    engine_index: int = 0
+
     # [num_reqs]
     outputs: list[EngineCoreOutput] = []
     scheduler_stats: Optional[SchedulerStats] = None
     timestamp: float = 0.0
 
     utility_output: Optional[UtilityOutput] = None
+    finished_requests: Optional[set[str]] = None
+
+    # In DP case, used to signal that the engine is paused.
+    engine_paused: bool = False
 
     def __post_init__(self):
         if self.timestamp == 0.0:
@@ -147,4 +153,5 @@ class EngineCoreRequestType(enum.Enum):
     """
     ADD = b'\x00'
     ABORT = b'\x01'
-    UTILITY = b'\x02'
+    START_DP = b'\x02'
+    UTILITY = b'\x03'

vllm/v1/engine/async_llm.py

@@ -66,11 +66,17 @@ class AsyncLLM(EngineClient):
 
         self.log_requests = log_requests
         self.log_stats = log_stats
-        self.stat_loggers: list[StatLoggerBase] = []
+
+        # Set up stat loggers; independent set for each DP rank.
+        self.stat_loggers: list[list[StatLoggerBase]] = []
         if self.log_stats:
-            if logger.isEnabledFor(logging.INFO):
-                self.stat_loggers.append(LoggingStatLogger())
-            self.stat_loggers.append(PrometheusStatLogger(vllm_config))
+            for i in range(vllm_config.parallel_config.data_parallel_size):
+                loggers: list[StatLoggerBase] = []
+                if logger.isEnabledFor(logging.INFO):
+                    loggers.append(LoggingStatLogger(engine_index=i))
+                loggers.append(
+                    PrometheusStatLogger(vllm_config, engine_index=i))
+                self.stat_loggers.append(loggers)
 
         # Tokenizer (+ ensure liveness if running in another process).
         self.tokenizer = init_tokenizer_from_configs(
@@ -329,6 +335,7 @@ class AsyncLLM(EngineClient):
                 # TODO(rob): make into a coroutine and launch it in
                 # background thread once Prometheus overhead is non-trivial.
                 self._record_stats(
+                    engine_index=outputs.engine_index,
                     scheduler_stats=outputs.scheduler_stats,
                     iteration_stats=iteration_stats,
                 )
@@ -350,12 +357,13 @@ class AsyncLLM(EngineClient):
         self,
         scheduler_stats: Optional[SchedulerStats],
         iteration_stats: Optional[IterationStats],
+        engine_index: int = 0,
     ):
         if not self.log_stats:
             return
 
         assert scheduler_stats is not None
-        for stat_logger in self.stat_loggers:
+        for stat_logger in self.stat_loggers[engine_index]:
             stat_logger.record(scheduler_stats=scheduler_stats,
                                iteration_stats=iteration_stats)
 
@@ -393,8 +401,9 @@ class AsyncLLM(EngineClient):
         scheduler_outputs=None,
         model_output=None,
     ) -> None:
-        for stat_logger in self.stat_loggers:
-            stat_logger.log()
+        for loggers in self.stat_loggers:
+            for stat_logger in loggers:
+                stat_logger.log()
 
     async def check_health(self) -> None:
         logger.debug("Called check_health.")

vllm/v1/engine/core.py

@@ -1,12 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
-
+import os
 import queue
 import signal
+import sys
 import threading
 import time
 from concurrent.futures import Future
 from inspect import isclass, signature
-from multiprocessing.connection import Connection
+from logging import DEBUG
 from typing import Any, Optional
 
 import msgspec
@@ -14,7 +15,9 @@ import psutil
 import zmq
 import zmq.asyncio
 
-from vllm.config import VllmConfig
+from vllm.config import ParallelConfig, VllmConfig
+from vllm.distributed import stateless_destroy_torch_distributed_process_group
+from vllm.executor.multiproc_worker_utils import _add_prefix
 from vllm.logger import init_logger
 from vllm.lora.request import LoRARequest
 from vllm.transformers_utils.config import (
@@ -91,6 +94,8 @@ class EngineCore:
             cache_config=vllm_config.cache_config,
             lora_config=vllm_config.lora_config,
             speculative_config=vllm_config.speculative_config,
+            include_finished_set=vllm_config.parallel_config.data_parallel_size
+            > 1,
             log_stats=self.log_stats,
             structured_output_manager=self.structured_output_manager,
         )
@@ -283,10 +288,10 @@ class EngineCoreProc(EngineCore):
         self,
         input_path: str,
         output_path: str,
-        ready_pipe: Connection,
         vllm_config: VllmConfig,
         executor_class: type[Executor],
         log_stats: bool,
+        engine_index: int = 0,
     ):
         super().__init__(vllm_config, executor_class, log_stats)
 
@@ -302,14 +307,20 @@ class EngineCoreProc(EngineCore):
                          args=(input_path, ),
                          daemon=True).start()
         threading.Thread(target=self.process_output_socket,
-                         args=(output_path, ),
+                         args=(output_path, engine_index),
                          daemon=True).start()
 
-        # Send Readiness signal to EngineClient.
-        ready_pipe.send({"status": "READY"})
+        self.global_unfinished_reqs = False
+
+        self.step_fn = (self.step if self.batch_queue is None else
+                        self.step_with_batch_queue)
 
     @staticmethod
-    def run_engine_core(*args, **kwargs):
+    def run_engine_core(*args,
+                        dp_rank: int = 0,
+                        local_dp_rank: int = 0,
+                        ready_pipe,
+                        **kwargs):
         """Launch EngineCore busy loop in background process."""
 
         # Signal handler used for graceful termination.
@@ -331,9 +342,21 @@ class EngineCoreProc(EngineCore):
         signal.signal(signal.SIGINT, signal_handler)
 
         parent_process = psutil.Process().parent()
-        engine_core = None
+        engine_core: Optional[EngineCoreProc] = None
         try:
-            engine_core = EngineCoreProc(*args, **kwargs)
+            parallel_config: ParallelConfig = kwargs[
+                "vllm_config"].parallel_config
+            if parallel_config.data_parallel_size > 1:
+                # Set data parallel rank for this engine process.
+                parallel_config.data_parallel_rank = dp_rank
+                parallel_config.data_parallel_rank_local = local_dp_rank
+                engine_core = DPEngineCoreProc(*args, **kwargs)
+            else:
+                engine_core = EngineCoreProc(*args, **kwargs)
+
+            # Send Readiness signal to EngineClient.
+            ready_pipe.send({"status": "READY"})
+
             engine_core.run_busy_loop()
 
         except SystemExit:
@@ -351,28 +374,44 @@ class EngineCoreProc(EngineCore):
     def run_busy_loop(self):
         """Core busy loop of the EngineCore."""
 
-        step_fn = (self.step
-                   if self.batch_queue is None else self.step_with_batch_queue)
-
         # Loop until process is sent a SIGINT or SIGTERM
         while True:
             # 1) Poll the input queue until there is work to do.
-            while not self.scheduler.has_requests():
-                logger.debug("EngineCore busy loop waiting.")
-                req = self.input_queue.get()
-                self._handle_client_request(*req)
+            self._process_input_queue()
+            # 2) Step the engine core and return the outputs.
+            self._process_engine_step()
 
-            # 2) Handle any new client requests.
-            while not self.input_queue.empty():
-                req = self.input_queue.get_nowait()
-                self._handle_client_request(*req)
+    def _process_input_queue(self):
+        """Exits when an engine step needs to be performed."""
 
-            # 3) Step the engine core.
-            outputs = step_fn()
+        waited = False
+        while not self.global_unfinished_reqs and not (
+                self.scheduler.has_requests()):
+            if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
+                logger.debug("EngineCore waiting for work.")
+                waited = True
+            req = self.input_queue.get()
+            self._handle_client_request(*req)
 
-            # 4) Put EngineCoreOutputs into the output queue.
-            if outputs is not None:
-                self.output_queue.put_nowait(outputs)
+        if waited:
+            logger.debug(
+                "EngineCore loop active - local unfinished: %s, finished: %s.",
+                self.scheduler.has_unfinished_requests(),
+                self.scheduler.has_finished_requests())
+
+        # Handle any more client requests.
+        while not self.input_queue.empty():
+            req = self.input_queue.get_nowait()
+            self._handle_client_request(*req)
+
+    def _process_engine_step(self):
+        """Called only when there are unfinished local requests."""
+
+        # Step the engine core.
+        outputs = self.step_fn()
+        # Put EngineCoreOutputs into the output queue.
+        if outputs is not None:
+            self.output_queue.put_nowait(outputs)
 
     def _handle_client_request(self, request_type: EngineCoreRequestType,
                                request: Any) -> None:
@@ -382,6 +421,10 @@ class EngineCoreProc(EngineCore):
             self.add_request(request)
         elif request_type == EngineCoreRequestType.ABORT:
             self.abort_requests(request)
+        elif request_type == EngineCoreRequestType.START_DP:
+            if not self.global_unfinished_reqs:
+                logger.debug("EngineCore starting idle loop.")
+                self.global_unfinished_reqs = True
         elif request_type == EngineCoreRequestType.UTILITY:
             call_id, method_name, args = request
             output = UtilityOutput(call_id)
@@ -432,7 +475,7 @@ class EngineCoreProc(EngineCore):
                 # Push to input queue for core busy loop.
                 self.input_queue.put_nowait((request_type, request))
 
-    def process_output_socket(self, output_path: str):
+    def process_output_socket(self, output_path: str, engine_index: int):
         """Output socket IO thread."""
 
         # Msgpack serialization encoding.
@@ -443,5 +486,114 @@ class EngineCoreProc(EngineCore):
         with zmq_socket_ctx(output_path, zmq.constants.PUSH) as socket:
             while True:
                 outputs = self.output_queue.get()
+                outputs.engine_index = engine_index
                 encoder.encode_into(outputs, buffer)
-                socket.send_multipart((buffer, ), copy=False)
+                socket.send(buffer, copy=False)
+
+
+ENGINE_PAUSED_OUTPUTS = EngineCoreOutputs(engine_paused=True)
+
+
+class DPEngineCoreProc(EngineCoreProc):
+    """ZMQ-wrapper for running EngineCore in background process
+    in a data parallel context."""
+
+    def __init__(
+        self,
+        input_path: str,
+        output_path: str,
+        vllm_config: VllmConfig,
+        executor_class: type[Executor],
+        log_stats: bool,
+    ):
+        # Add process-specific prefix to stdout and stderr before
+        # we initialize the engine.
+        from multiprocessing import current_process
+        process_name = current_process().name
+        pid = os.getpid()
+        _add_prefix(sys.stdout, process_name, pid)
+        _add_prefix(sys.stderr, process_name, pid)
+
+        dp_size = vllm_config.parallel_config.data_parallel_size
+        dp_rank = vllm_config.parallel_config.data_parallel_rank
+        local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local
+
+        assert dp_size > 1
+        assert 0 <= local_dp_rank <= dp_rank < dp_size
+
+        from vllm.platforms import current_platform
+        if current_platform.is_cuda_alike():
+            from vllm.platforms.cuda import device_id_to_physical_device_id
+            tp_size = vllm_config.parallel_config.tensor_parallel_size
+            os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(
+                str(device_id_to_physical_device_id(i))
+                for i in range(local_dp_rank * tp_size, (local_dp_rank + 1) *
+                               tp_size))
+
+        self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
+
+        # Initialize the engine after setting up environment.
+        super().__init__(input_path, output_path, vllm_config, executor_class,
+                         log_stats, dp_rank)
+
+        # Counts forward-passes of the model so that we can synchronize
+        # finished with DP peers every N steps.
+        self.counter = 0
+
+    def shutdown(self):
+        super().shutdown()
+        if dp_group := getattr(self, "dp_group", None):
+            stateless_destroy_torch_distributed_process_group(dp_group)
+
+    def run_busy_loop(self):
+        """Core busy loop of the EngineCore for data parallel case."""
+
+        # Loop until process is sent a SIGINT or SIGTERM
+        while True:
+            # 1) Poll the input queue until there is work to do.
+            self._process_input_queue()
+
+            local_unfinished_reqs = self.scheduler.has_unfinished_requests()
+
+            if local_unfinished_reqs:
+                # 2) Step the engine core.
+                self._process_engine_step()
+
+                # Check if we have now finished all requests.
+                local_unfinished_reqs = (
+                    self.scheduler.has_unfinished_requests())
+            else:
+                if self.scheduler.has_finished_requests():
+                    # There are no unfinished requests, but there are some
+                    # finished requests remaining to be removed from the
+                    # batch state. This engine step won't perform a forward
+                    # pass but will flush the finished requests to ensure
+                    # up-to-date state is returned in the engine outputs.
+                    self._process_engine_step()
+
+                if not self.global_unfinished_reqs:
+                    # All engines are idle.
+                    continue
+
+                # There must be unfinished requests in DP peers, run a
+                # dummy forward pass.
+                self.execute_dummy_batch()
+
+            # 3) All-reduce operation to determine global unfinished reqs.
+            self.global_unfinished_reqs = self._has_global_unfinished_reqs(
+                local_unfinished_reqs)
+
+            if not self.global_unfinished_reqs:
+                # Notify client that we are pausing the loop.
+                self.output_queue.put_nowait(ENGINE_PAUSED_OUTPUTS)
+
+    def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:
+
+        # Optimization - only perform finish-sync all-reduce every 16 steps.
+        self.counter += 1
+        if self.counter != 16:
+            return True
+        self.counter = 0
+
+        return ParallelConfig.has_unfinished_dp(self.dp_group,
+                                                local_unfinished)

vllm/v1/engine/core_client.py

@@ -8,10 +8,11 @@ import threading
 import uuid
 import weakref
 from abc import ABC, abstractmethod
+from collections.abc import Awaitable, Sequence
 from concurrent.futures import Future
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from threading import Thread
-from typing import Any, Optional, Union
+from typing import Any, Callable, Optional, Union
 
 import zmq
 import zmq.asyncio
@@ -60,6 +61,9 @@ class EngineCoreClient(ABC):
                 "is not currently supported.")
 
         if multiprocess_mode and asyncio_mode:
+            if vllm_config.parallel_config.data_parallel_size > 1:
+                return DPAsyncMPClient(vllm_config, executor_class, log_stats)
+
             return AsyncMPClient(vllm_config, executor_class, log_stats)
 
         if multiprocess_mode and not asyncio_mode:
@@ -207,28 +211,74 @@ class InprocClient(EngineCoreClient):
         return self.engine_core.pin_lora(lora_id)
 
 
+class CoreEngine:
+    """One per data parallel rank."""
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        executor_class: type[Executor],
+        log_stats: bool,
+        ctx: Union[zmq.Context, zmq.asyncio.Context],
+        output_path: str,
+        index: int = 0,
+        local_dp_rank: int = 0,
+    ):
+        # Paths and sockets for IPC.
+        input_path = get_open_zmq_ipc_path()
+        self.input_socket = make_zmq_socket(ctx, input_path,
+                                            zmq.constants.PUSH)
+        try:
+            # Start EngineCore in background process.
+            self.proc_handle = BackgroundProcHandle(
+                input_path=input_path,
+                output_path=output_path,
+                process_name=f"EngineCore_{index}",
+                target_fn=EngineCoreProc.run_engine_core,
+                process_kwargs={
+                    "vllm_config": vllm_config,
+                    "dp_rank": index,
+                    "local_dp_rank": local_dp_rank,
+                    "executor_class": executor_class,
+                    "log_stats": log_stats,
+                })
+
+            self.num_reqs_in_flight = 0
+        finally:
+            if not hasattr(self, "num_reqs_in_flight"):
+                # Ensure socket is closed if process fails to start.
+                self.close()
+
+    def send_multipart(self, msg_parts: Sequence):
+        return self.input_socket.send_multipart(msg_parts, copy=False)
+
+    def close(self):
+        if proc_handle := getattr(self, "proc_handle", None):
+            proc_handle.shutdown()
+        if socket := getattr(self, "input_socket", None):
+            socket.close(linger=0)
+
+
 @dataclass
 class BackgroundResources:
     """Used as a finalizer for clean shutdown, avoiding
     circular reference back to the client object."""
 
-    ctx: zmq.Context
+    ctx: Union[zmq.Context]
+    core_engines: list[CoreEngine] = field(default_factory=list)
     output_socket: Optional[Union[zmq.Socket, zmq.asyncio.Socket]] = None
-    input_socket: Optional[Union[zmq.Socket, zmq.asyncio.Socket]] = None
-    proc_handle: Optional[BackgroundProcHandle] = None
     shutdown_path: Optional[str] = None
 
     def __call__(self):
         """Clean up background resources."""
 
-        if self.proc_handle is not None:
-            self.proc_handle.shutdown()
+        for core_engine in self.core_engines:
+            core_engine.close()
+
         # ZMQ context termination can hang if the sockets
         # aren't explicitly closed first.
         if self.output_socket is not None:
             self.output_socket.close(linger=0)
-        if self.input_socket is not None:
-            self.input_socket.close(linger=0)
         if self.shutdown_path is not None:
             # We must ensure that the sync output socket is
             # closed cleanly in its own thread.
@@ -284,7 +334,7 @@ class MPClient(EngineCoreClient):
         self.decoder = MsgpackDecoder(EngineCoreOutputs)
 
         # ZMQ setup.
-        sync_ctx = zmq.Context()
+        sync_ctx = zmq.Context(io_threads=2)
         self.ctx = zmq.asyncio.Context(sync_ctx) if asyncio_mode else sync_ctx
 
         # This will ensure resources created so far are closed
@@ -293,28 +343,38 @@ class MPClient(EngineCoreClient):
         self.resources = BackgroundResources(ctx=sync_ctx)
         self._finalizer = weakref.finalize(self, self.resources)
 
-        # Paths for IPC.
+        # Paths and sockets for IPC.
         self.output_path = get_open_zmq_ipc_path()
-        input_path = get_open_zmq_ipc_path()
 
-        # Start EngineCore in background process.
-        self.resources.proc_handle = BackgroundProcHandle(
-            input_path=input_path,
-            output_path=self.output_path,
-            process_name="EngineCore",
-            target_fn=EngineCoreProc.run_engine_core,
-            process_kwargs={
-                "vllm_config": vllm_config,
-                "executor_class": executor_class,
-                "log_stats": log_stats,
-            })
+        new_core_engine = lambda index, local_dp_rank=None: CoreEngine(
+            vllm_config, executor_class, log_stats, self.ctx, self.output_path,
+            index, local_dp_rank)
+
+        # Start engine core process(es).
+        self._init_core_engines(vllm_config, new_core_engine,
+                                self.resources.core_engines)
+
+        # Wait for engine core process(es) to start.
+        for engine in self.resources.core_engines:
+            engine.proc_handle.wait_for_startup()
 
-        # Create input socket.
-        self.resources.input_socket = make_zmq_socket(self.ctx, input_path,
-                                                      zmq.constants.PUSH)
-        self.input_socket = self.resources.input_socket
         self.utility_results: dict[int, AnyFuture] = {}
 
+    def _init_core_engines(
+        self,
+        vllm_config: VllmConfig,
+        new_core_engine: Callable[[int, Optional[int]], CoreEngine],
+        core_engines: list[CoreEngine],
+    ) -> None:
+
+        # Default case - single core engine.
+        dp_rank = vllm_config.parallel_config.data_parallel_rank
+        local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local
+        core_engine = new_core_engine(
+            dp_rank, local_dp_rank if local_dp_rank is not None else dp_rank)
+        core_engines.append(core_engine)
+        self.core_engine = core_engine
+
     def shutdown(self):
         self._finalizer()
 
@@ -370,7 +430,7 @@ class SyncMPClient(MPClient):
                         # shutdown signal, exit thread.
                         break
 
-                    (frame, ) = out_socket.recv_multipart(copy=False)
+                    frame = out_socket.recv(copy=False)
                     outputs = decoder.decode(frame.buffer)
                     if outputs.utility_output:
                         _process_utility_output(outputs.utility_output,
@@ -391,18 +451,15 @@ class SyncMPClient(MPClient):
     def get_output(self) -> EngineCoreOutputs:
         return self.outputs_queue.get()
 
-    def _send_input(self, request_type: EngineCoreRequestType,
-                    request: Any) -> None:
-
+    def _send_input(self, request_type: EngineCoreRequestType, request: Any):
         # (RequestType, SerializedRequest)
         msg = (request_type.value, self.encoder.encode(request))
-        self.input_socket.send_multipart(msg, copy=False)
+        self.core_engine.send_multipart(msg)
 
-    def _call_utility(self, method: str, *args) -> Any:
+    def call_utility(self, method: str, *args) -> Any:
         call_id = uuid.uuid1().int >> 64
         future: Future[Any] = Future()
         self.utility_results[call_id] = future
-
         self._send_input(EngineCoreRequestType.UTILITY,
                          (call_id, method, args))
 
@@ -419,34 +476,34 @@ class SyncMPClient(MPClient):
             self._send_input(EngineCoreRequestType.ABORT, request_ids)
 
     def profile(self, is_start: bool = True) -> None:
-        self._call_utility("profile", is_start)
+        self.call_utility("profile", is_start)
 
     def reset_prefix_cache(self) -> None:
-        self._call_utility("reset_prefix_cache")
+        self.call_utility("reset_prefix_cache")
 
     def add_lora(self, lora_request: LoRARequest) -> bool:
-        return self._call_utility("add_lora", lora_request)
+        return self.call_utility("add_lora", lora_request)
 
     def remove_lora(self, lora_id: int) -> bool:
-        return self._call_utility("remove_lora", lora_id)
+        return self.call_utility("remove_lora", lora_id)
 
     def list_loras(self) -> set[int]:
-        return self._call_utility("list_loras")
+        return self.call_utility("list_loras")
 
     def pin_lora(self, lora_id: int) -> bool:
-        return self._call_utility("pin_lora", lora_id)
+        return self.call_utility("pin_lora", lora_id)
 
     def sleep(self, level: int = 1) -> None:
-        self._call_utility("sleep", level)
+        self.call_utility("sleep", level)
 
     def wake_up(self) -> None:
-        self._call_utility("wake_up")
+        self.call_utility("wake_up")
 
     def is_sleeping(self) -> bool:
-        return self._call_utility("is_sleeping")
+        return self.call_utility("is_sleeping")
 
     def execute_dummy_batch(self) -> None:
-        self._call_utility("execute_dummy_batch")
+        self.call_utility("execute_dummy_batch")
 
 
 class AsyncMPClient(MPClient):
@@ -464,13 +521,21 @@ class AsyncMPClient(MPClient):
         self.outputs_queue: Optional[asyncio.Queue[EngineCoreOutputs]] = None
         self.queue_task: Optional[asyncio.Task] = None
 
-    async def _start_output_queue_task(self):
+        self.outputs_handler: Optional[Callable[
+            [AsyncMPClient, EngineCoreOutputs], Awaitable[None]]] = None
+
+    def _ensure_output_queue_task(self):
+        if self.outputs_queue is not None:
+            return
+
         # Perform IO in separate task to parallelize as much as possible.
         # Avoid task having direct reference back to the client.
         self.outputs_queue = asyncio.Queue()
         decoder = self.decoder
         utility_results = self.utility_results
         outputs_queue = self.outputs_queue
+        output_handler = self.outputs_handler
+        _self_ref = weakref.ref(self) if output_handler else None
         output_path = self.output_path
         output_socket = make_zmq_socket(self.ctx, output_path,
                                         zmq.constants.PULL)
@@ -483,34 +548,52 @@ class AsyncMPClient(MPClient):
                 if outputs.utility_output:
                     _process_utility_output(outputs.utility_output,
                                             utility_results)
-                else:
+                    continue
+
+                if output_handler is not None:
+                    assert _self_ref is not None
+                    _self = _self_ref()
+                    if not _self:
+                        # Client has been garbage collected, abort.
+                        return
+                    await output_handler(_self, outputs)
+
+                if outputs.outputs or outputs.scheduler_stats:
                     outputs_queue.put_nowait(outputs)
 
         self.queue_task = asyncio.create_task(process_outputs_socket(),
                                               name="EngineCoreOutputQueueTask")
 
     async def get_output_async(self) -> EngineCoreOutputs:
-        if self.outputs_queue is None:
-            await self._start_output_queue_task()
-            assert self.outputs_queue is not None
+        self._ensure_output_queue_task()
+        assert self.outputs_queue is not None
         return await self.outputs_queue.get()
 
     async def _send_input(self, request_type: EngineCoreRequestType,
                           request: Any) -> None:
+        await self.core_engine.send_multipart(
+            (request_type.value, self.encoder.encode(request)))
 
-        msg = (request_type.value, self.encoder.encode(request))
-        await self.input_socket.send_multipart(msg, copy=False)
+        self._ensure_output_queue_task()
 
-        if self.outputs_queue is None:
-            await self._start_output_queue_task()
+    async def call_utility_async(self, method: str, *args) -> Any:
+        return await self._call_utility_async(method,
+                                              *args,
+                                              engine=self.core_engine)
 
-    async def _call_utility_async(self, method: str, *args) -> Any:
+    async def _call_utility_async(
+        self,
+        method: str,
+        *args,
+        engine: CoreEngine,
+    ) -> Any:
         call_id = uuid.uuid1().int >> 64
         future = asyncio.get_running_loop().create_future()
         self.utility_results[call_id] = future
-        await self._send_input(EngineCoreRequestType.UTILITY,
-                               (call_id, method, args))
-
+        message = (EngineCoreRequestType.UTILITY.value,
+                   self.encoder.encode((call_id, method, args)))
+        await engine.send_multipart(message)
+        self._ensure_output_queue_task()
         return await future
 
     async def add_request_async(self, request: EngineCoreRequest) -> None:
@@ -524,31 +607,146 @@ class AsyncMPClient(MPClient):
             await self._send_input(EngineCoreRequestType.ABORT, request_ids)
 
     async def profile_async(self, is_start: bool = True) -> None:
-        await self._call_utility_async("profile", is_start)
+        await self.call_utility_async("profile", is_start)
 
     async def reset_prefix_cache_async(self) -> None:
-        await self._call_utility_async("reset_prefix_cache")
+        await self.call_utility_async("reset_prefix_cache")
 
     async def sleep_async(self, level: int = 1) -> None:
-        await self._call_utility_async("sleep", level)
+        await self.call_utility_async("sleep", level)
 
     async def wake_up_async(self) -> None:
-        await self._call_utility_async("wake_up")
+        await self.call_utility_async("wake_up")
 
     async def is_sleeping_async(self) -> bool:
-        return await self._call_utility_async("is_sleeping")
+        return await self.call_utility_async("is_sleeping")
 
     async def execute_dummy_batch_async(self) -> None:
-        await self._call_utility_async("execute_dummy_batch")
+        await self.call_utility_async("execute_dummy_batch")
 
     async def add_lora_async(self, lora_request: LoRARequest) -> bool:
-        return await self._call_utility_async("add_lora", lora_request)
+        return await self.call_utility_async("add_lora", lora_request)
 
     async def remove_lora_async(self, lora_id: int) -> bool:
-        return await self._call_utility_async("remove_lora", lora_id)
+        return await self.call_utility_async("remove_lora", lora_id)
 
     async def list_loras_async(self) -> set[int]:
-        return await self._call_utility_async("list_loras")
+        return await self.call_utility_async("list_loras")
 
     async def pin_lora_async(self, lora_id: int) -> bool:
-        return await self._call_utility_async("pin_lora", lora_id)
+        return await self.call_utility_async("pin_lora", lora_id)
+
+
+class DPAsyncMPClient(AsyncMPClient):
+    """Asyncio-compatible client for multi-proc, multi-engine (data parallel)
+    EngineCore."""
+
+    def __init__(self, vllm_config: VllmConfig, executor_class: type[Executor],
+                 log_stats: bool):
+        super().__init__(vllm_config, executor_class, log_stats)
+
+        assert len(self.core_engines) > 1
+
+        # Control message used for triggering dp idle mode loop.
+        self.start_dp_msg = (EngineCoreRequestType.START_DP.value,
+                             self.encoder.encode(None))
+
+        self.num_engines_running = 0
+        self.reqs_in_flight: dict[str, CoreEngine] = {}
+
+        self.outputs_handler = DPAsyncMPClient.process_engine_outputs  # type: ignore[assignment]
+
+    def _init_core_engines(
+        self,
+        vllm_config: VllmConfig,
+        new_core_engine: Callable[[int, Optional[int]], CoreEngine],
+        core_engines: list[CoreEngine],
+    ) -> None:
+
+        # Launch a core engine for each data parallel rank.
+        dp_size = vllm_config.parallel_config.data_parallel_size
+        for i in range(dp_size):
+            # Multi-node not yet supported so local_dp_rank == dp_rank.
+            core_engines.append(new_core_engine(i, i))
+
+        self.core_engines = core_engines
+
+    async def call_utility_async(self, method: str, *args) -> Any:
+        # Only the result from the first engine is returned.
+        return (await asyncio.gather(*[
+            self._call_utility_async(method, *args, engine=engine)
+            for engine in self.core_engines
+        ]))[0]
+
+    async def add_request_async(self, request: EngineCoreRequest) -> None:
+        # NOTE: text prompt is not needed in the core engine as it has been
+        # tokenized.
+        request.prompt = None
+
+        msg = (EngineCoreRequestType.ADD.value, self.encoder.encode(request))
+
+        chosen_engine = self.get_core_engine_for_request()
+        self.reqs_in_flight[request.request_id] = chosen_engine
+        chosen_engine.num_reqs_in_flight += 1
+        if self.num_engines_running >= len(self.core_engines):
+            await chosen_engine.send_multipart(msg)
+        else:
+            # Send request to chosen engine and dp start loop
+            # control message to all other engines.
+            self.num_engines_running += len(self.core_engines)
+            await asyncio.gather(*[
+                engine.send_multipart(msg if engine is
+                                      chosen_engine else self.start_dp_msg)
+                for engine in self.core_engines
+            ])
+
+        self._ensure_output_queue_task()
+
+    def get_core_engine_for_request(self) -> CoreEngine:
+        return min(self.core_engines, key=lambda e: e.num_reqs_in_flight)
+
+    @staticmethod
+    async def process_engine_outputs(self: "DPAsyncMPClient",
+                                     outputs: EngineCoreOutputs):
+        if self.reqs_in_flight:
+            for req_id in outputs.finished_requests or ():
+                if engine := self.reqs_in_flight.pop(req_id, None):
+                    engine.num_reqs_in_flight -= 1
+
+        if outputs.engine_paused:
+            assert self.num_engines_running >= 1
+            self.num_engines_running -= 1
+            if not self.num_engines_running and self.reqs_in_flight:
+                # If there are requests in flight here, they must have
+                # been sent after the engines paused. We must make
+                # sure to start the other engines:
+                self.num_engines_running = len(self.core_engines)
+                coros = [
+                    engine.send_multipart(self.start_dp_msg)
+                    for engine in self.core_engines
+                    if not engine.num_reqs_in_flight
+                ]
+                if coros:
+                    await asyncio.gather(*coros)
+
+    async def abort_requests_async(self, request_ids: list[str]) -> None:
+        if not request_ids:
+            return
+
+        if len(request_ids) == 1:
+            # Fast-path common case.
+            if engine := self.reqs_in_flight.get(request_ids[0]):
+                await self._abort_requests(request_ids, engine)
+            return
+
+        by_engine: dict[CoreEngine, list[str]] = {}
+        for req_id in request_ids:
+            if engine := self.reqs_in_flight.get(req_id):
+                by_engine.setdefault(engine, []).append(req_id)
+        for engine, req_ids in by_engine.items():
+            await self._abort_requests(req_ids, engine)
+
+    async def _abort_requests(self, request_ids: list[str],
+                              engine: CoreEngine) -> None:
+        await engine.send_multipart((EngineCoreRequestType.ABORT.value,
+                                     self.encoder.encode(request_ids)))

vllm/v1/engine/llm_engine.py

@@ -8,6 +8,7 @@ from typing_extensions import TypeVar
 
 import vllm.envs as envs
 from vllm.config import ParallelConfig, VllmConfig
+from vllm.distributed import stateless_destroy_torch_distributed_process_group
 from vllm.engine.arg_utils import EngineArgs
 from vllm.engine.metrics_types import StatLoggerBase
 from vllm.inputs import INPUT_REGISTRY, InputRegistry, PromptType
@@ -60,11 +61,13 @@ class LLMEngine:
         self.cache_config = vllm_config.cache_config
 
         # important: init dp group before init the engine_core
-        self.parallel_config = vllm_config.parallel_config
-        self.dp_enabled = self.parallel_config.data_parallel_size > 1  # noqa
+        # In the decoupled engine case this is handled in EngineCoreProc.
+        parallel_config = vllm_config.parallel_config
+        if not multiprocess_mode and parallel_config.data_parallel_size > 1:
+            self.dp_group = parallel_config.stateless_init_dp_group()
+        else:
+            self.dp_group = None
         self.should_execute_dummy_batch = False
-        if self.dp_enabled:
-            self.dp_group = self.parallel_config.stateless_init_dp_group()
 
         # Tokenizer (+ ensure liveness if running in another process).
         self.tokenizer = init_tokenizer_from_configs(
@@ -148,7 +151,7 @@ class LLMEngine:
 
     def has_unfinished_requests(self) -> bool:
         has_unfinished = self.output_processor.has_unfinished_requests()
-        if not self.dp_enabled:
+        if self.dp_group is None:
             return has_unfinished
         return self.has_unfinished_requests_dp(has_unfinished)
 
@@ -280,3 +283,7 @@ class LLMEngine:
     def pin_lora(self, lora_id: int) -> bool:
         """Prevent an adapter from being evicted."""
         return self.engine_core.pin_lora(lora_id)
+
+    def __del__(self):
+        if dp_group := getattr(self, "dp_group", None):
+            stateless_destroy_torch_distributed_process_group(dp_group)