[BugFix] Improve internal DP load balancing (#21617)

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

vllm/entrypoints/openai/api_server.py

@@ -199,6 +199,8 @@ async def build_async_engine_client_from_engine_args(
 
         from vllm.v1.engine.async_llm import AsyncLLM
         async_llm: Optional[AsyncLLM] = None
+        client_count = client_config.pop(
+            "client_count") if client_config else 1
         client_index = client_config.pop(
             "client_index") if client_config else 0
         try:
@@ -208,6 +210,7 @@ async def build_async_engine_client_from_engine_args(
                 enable_log_requests=engine_args.enable_log_requests,
                 disable_log_stats=engine_args.disable_log_stats,
                 client_addresses=client_config,
+                client_count=client_count,
                 client_index=client_index)
 
             # Don't keep the dummy data in memory

vllm/v1/engine/async_llm.py

@@ -57,6 +57,7 @@ class AsyncLLM(EngineClient):
         start_engine_loop: bool = True,
         stat_loggers: Optional[list[StatLoggerFactory]] = None,
         client_addresses: Optional[dict[str, str]] = None,
+        client_count: int = 1,
         client_index: int = 0,
     ) -> None:
         """
@@ -120,6 +121,7 @@ class AsyncLLM(EngineClient):
             executor_class=executor_class,
             log_stats=self.log_stats,
             client_addresses=client_addresses,
+            client_count=client_count,
             client_index=client_index,
         )
 
@@ -156,6 +158,7 @@ class AsyncLLM(EngineClient):
             enable_log_requests: bool = False,
             disable_log_stats: bool = False,
             client_addresses: Optional[dict[str, str]] = None,
+            client_count: int = 1,
             client_index: int = 0,
             disable_log_requests: bool = True,  # Deprecated, will be removed
     ) -> "AsyncLLM":
@@ -176,6 +179,7 @@ class AsyncLLM(EngineClient):
             log_stats=not disable_log_stats,
             usage_context=usage_context,
             client_addresses=client_addresses,
+            client_count=client_count,
             client_index=client_index,
         )
 

vllm/v1/engine/coordinator.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import copy
 import multiprocessing
 import time
 import weakref
@@ -65,18 +66,14 @@ class DPCoordinator:
 
         # Assume coordinator is colocated with front-end procs when not in
         # either external or hybrid DP LB mode.
+        local_only = not (external_lb or hybrid_lb)
         front_publish_address = get_engine_client_zmq_addr(
-            local_only=not external_lb and not hybrid_lb, host=host)
+            local_only=local_only, host=host)
 
         local_only_eng = dp_size == parallel_config.data_parallel_size_local
         back_publish_address = get_engine_client_zmq_addr(local_only_eng, host)
         back_output_address = get_engine_client_zmq_addr(local_only_eng, host)
 
-        # When in external LB mode, load stats aren't published, only changes
-        # to request wave / running state, so we don't need to rate-limit the
-        # updates to the front-end proc(s).
-        min_stats_update_interval_ms = 0 if external_lb else 100
-
         context = get_mp_context()
         self.proc: multiprocessing.Process = context.Process(
             target=DPCoordinatorProc.run_coordinator,
@@ -86,7 +83,6 @@ class DPCoordinator:
                 "front_publish_address": front_publish_address,
                 "back_output_address": back_output_address,
                 "back_publish_address": back_publish_address,
-                "min_stats_update_interval_ms": min_stats_update_interval_ms,
             },
             daemon=True)
         self.proc.start()
@@ -125,10 +121,6 @@ class DPCoordinatorProc:
 
         self.stats_update_interval_ms = min_stats_update_interval_ms
 
-        self.current_wave = 0
-        self.engines_running = False
-        self.stats_changed = False
-
     @staticmethod
     def run_coordinator(
         engine_count: int,
@@ -155,6 +147,16 @@ class DPCoordinatorProc:
 
         decoder = MsgpackDecoder(EngineCoreOutputs)
 
+        # For tracking request wave progression.
+        current_wave = 0
+        engines_running = False
+
+        # For tracking request counts for internal load-balancing.
+        stats_changed = False
+        last_stats_step = -1
+        last_stats_wave = -1
+        last_step_counts: Optional[list[list[int]]] = None
+
         with make_zmq_socket(
                 path=front_publish_address,  # IPC
                 ctx=self.ctx,
@@ -191,21 +193,33 @@ class DPCoordinatorProc:
             while True:
                 elapsed = int(time.time() * 1000) - last_publish_time
                 # Send at stats_update_interval_ms interval if the stats have
-                # changed, or otherwise every 4 seconds.
+                # changed, or otherwise every 5 seconds.
                 wait_for = (self.stats_update_interval_ms
-                            if self.stats_changed else 4000)
-                events = poller.poll(timeout=max(0, wait_for - elapsed))
+                            if stats_changed else 5000)
+
+                # Wait at least 50ms to ensure we've received all stats for
+                # the current step.
+                min_timeout = 50 if last_step_counts is None else 0
+
+                events = poller.poll(timeout=max(min_timeout, wait_for -
+                                                 elapsed))
                 if not events:
                     # Poller timeout - publish current stats to front-ends.
-                    engine_req_counts_list = self._get_engine_counts()
-                    to_publish = (engine_req_counts_list, self.current_wave,
-                                  self.engines_running)
+                    if last_step_counts is not None:
+                        engine_req_counts_list = last_step_counts
+                        last_step_counts = None
+                    else:
+                        engine_req_counts_list = self._get_engine_counts()
+                        stats_changed = False
+
+                    to_publish = (engine_req_counts_list, current_wave,
+                                  engines_running)
                     publish_front.send(msgspec.msgpack.encode(to_publish))
                     last_publish_time = int(time.time() * 1000)
-                    self.stats_changed = False
                     continue
 
                 events = dict(events)
+                wave_state_changed = False
 
                 if publish_front in events:
                     buffer = publish_front.recv()
@@ -232,7 +246,7 @@ class DPCoordinatorProc:
                             # current_wave
                             # we note that 0 is the wave number for the new
                             # engine
-                            self.engines_running = False
+                            engines_running = False
                             logger.info(
                                 "DPCoordinator scaled up from %s to %s "
                                 "engines", current_count, new_engine_count)
@@ -248,15 +262,15 @@ class DPCoordinatorProc:
                     # engines are paused, so that we can wake the other
                     # engines.
                     engine_to_exclude, wave = decoded
-                    if not self.engines_running:
-                        if wave < self.current_wave:
+                    if not engines_running:
+                        if wave < current_wave:
                             # If the wave number is stale, ensure the message
                             # is handled by all the engines.
                             engine_to_exclude = None
 
-                        self.engines_running = True
-                        self.stats_changed = True
-                        self._send_start_wave(publish_back, self.current_wave,
+                        engines_running = True
+                        wave_state_changed = True
+                        self._send_start_wave(publish_back, current_wave,
                                               engine_to_exclude)
 
                 if output_back in events:
@@ -274,36 +288,56 @@ class DPCoordinatorProc:
                         # 1. Updated request load stats - update our local
                         # state with these.
                         stats = self.engines[eng_index].request_counts
+                        stats_step = scheduler_stats.step_counter
+                        stats_wave = scheduler_stats.current_wave
+                        if (stats_wave > last_stats_wave
+                                or stats_wave == last_stats_wave
+                                and stats_step > last_stats_step):
+                            if stats_changed:
+                                last_step_counts = self._get_engine_counts(
+                                    do_copy=True)
+                            last_stats_step = stats_step
+                            last_stats_wave = stats_wave
+                        elif stats_wave != last_stats_wave or (
+                                stats_step != last_stats_step):
+                            logger.warning(
+                                "Received stats for out-of-order "
+                                "step (%d, %d) from engine %d (expected "
+                                "> (%d, %d))", stats_wave, stats_step,
+                                eng_index, last_stats_wave, last_stats_step)
                         stats[0] = scheduler_stats.num_waiting_reqs
                         stats[1] = scheduler_stats.num_running_reqs
-                        self.stats_changed = True
+                        stats_changed = True
 
                     if (wave := outputs.wave_complete) is not None:
                         # 2. Notification from rank 0 engine that we've
                         # moved into the global paused state
                         # (engines_running==False).
-                        if self.current_wave <= wave:
+                        if current_wave <= wave:
                             new_wave = wave + 1
                             logger.debug("Moving DP wave from %d to %d.",
-                                         self.current_wave, new_wave)
-                            self.current_wave = new_wave
-                            self.engines_running = False
-                            self.stats_changed = True
+                                         current_wave, new_wave)
+                            current_wave = new_wave
+                            engines_running = False
+                            wave_state_changed = True
                     elif (wave := outputs.start_wave) is not None and (
-                            wave > self.current_wave or
-                        (wave == self.current_wave
-                         and not self.engines_running)):
+                            wave > current_wave or
+                        (wave == current_wave and not engines_running)):
                         # 3. The engine received request for a non-current wave
                         # so we must ensure that other engines progress to the
                         # next wave (race condition handling).
                         logger.debug(
                             "Starting wave %d after notification of "
                             "stale wave request from engine.", wave)
-                        self.current_wave = wave
-                        self.engines_running = True
-                        self.stats_changed = True
+                        current_wave = wave
+                        engines_running = True
+                        wave_state_changed = True
                         self._send_start_wave(publish_back, wave, eng_index)
 
+                if wave_state_changed:
+                    message = (None, current_wave, engines_running)
+                    publish_front.send(msgspec.msgpack.encode(message))
+
     @staticmethod
     def _send_start_wave(socket: zmq.Socket, wave: int,
                          exclude_engine_index: Optional[int]):
@@ -316,6 +350,8 @@ class DPCoordinatorProc:
         socket.send_multipart(
             (EngineCoreRequestType.START_DP_WAVE.value, wave_encoded))
 
-    def _get_engine_counts(self) -> list[list[int]]:
+    def _get_engine_counts(self, do_copy=False) -> list[list[int]]:
         """Return list of [waiting, running] count lists for each engine."""
+        if do_copy:
+            return [copy.copy(e.request_counts) for e in self.engines]
         return [e.request_counts for e in self.engines]

vllm/v1/engine/core.py

@@ -928,7 +928,7 @@ class DPEngineCoreProc(EngineCoreProc):
     ):
         # Counts forward-passes of the model so that we can synchronize
         # finished with DP peers every N steps.
-        self.counter = 0
+        self.step_counter = 0
         self.current_wave = 0
         self.last_counts = (0, 0)
 
@@ -999,7 +999,9 @@ class DPEngineCoreProc(EngineCoreProc):
         counts = self.scheduler.get_request_counts()
         if counts != self.last_counts:
             self.last_counts = counts
-            stats = SchedulerStats(*counts)
+            stats = SchedulerStats(*counts,
+                                   step_counter=self.step_counter,
+                                   current_wave=self.current_wave)
             self.output_queue.put_nowait(
                 (-1, EngineCoreOutputs(scheduler_stats=stats)))
 
@@ -1041,15 +1043,16 @@ class DPEngineCoreProc(EngineCoreProc):
                     self.output_queue.put_nowait(
                         (client_index,
                          EngineCoreOutputs(wave_complete=self.current_wave)))
+                # Increment wave count and reset step counter.
                 self.current_wave += 1
+                self.step_counter = 0
 
     def _has_global_unfinished_reqs(self, local_unfinished: bool) -> bool:
 
         # Optimization - only perform finish-sync all-reduce every 32 steps.
-        self.counter += 1
-        if self.counter != 32:
+        self.step_counter += 1
+        if self.step_counter % 32 != 0:
             return True
-        self.counter = 0
 
         return ParallelConfig.has_unfinished_dp(self.dp_group,
                                                 local_unfinished)

vllm/v1/engine/core_client.py

@@ -86,11 +86,12 @@ class EngineCoreClient(ABC):
         executor_class: type[Executor],
         log_stats: bool,
         client_addresses: Optional[dict[str, str]] = None,
+        client_count: int = 1,
         client_index: int = 0,
     ) -> "MPClient":
         parallel_config = vllm_config.parallel_config
         client_args = (vllm_config, executor_class, log_stats,
-                       client_addresses, client_index)
+                       client_addresses, client_count, client_index)
         if parallel_config.data_parallel_size > 1:
             if parallel_config.data_parallel_external_lb:
                 # External load balancer - client per DP rank.
@@ -727,6 +728,7 @@ class AsyncMPClient(MPClient):
                  executor_class: type[Executor],
                  log_stats: bool,
                  client_addresses: Optional[dict[str, str]] = None,
+                 client_count: int = 1,
                  client_index: int = 0):
         super().__init__(
             asyncio_mode=True,
@@ -929,11 +931,12 @@ class DPAsyncMPClient(AsyncMPClient):
                  executor_class: type[Executor],
                  log_stats: bool,
                  client_addresses: Optional[dict[str, str]] = None,
+                 client_count: int = 1,
                  client_index: int = 0):
         self.current_wave = 0
 
         super().__init__(vllm_config, executor_class, log_stats,
-                         client_addresses, client_index)
+                         client_addresses, client_count, client_index)
 
         # List of [waiting, running] pair per engine.
         # Used only by DPLBAsyncMPClient subclass.
@@ -1029,7 +1032,11 @@ class DPAsyncMPClient(AsyncMPClient):
                     counts, wave, running = msgspec.msgpack.decode(buf)
                     self.current_wave = wave
                     self.engines_running = running
-                    self.lb_engines = counts[count_slice]
+                    if counts is not None:
+                        sliced_counts = counts[count_slice]
+                        self.lb_engines = sliced_counts
+                        logger.debug("Received counts: %s (%s)", sliced_counts,
+                                     count_slice)
 
         resources.stats_update_task = asyncio.create_task(
             run_engine_stats_update_task())
@@ -1065,40 +1072,45 @@ class DPLBAsyncMPClient(DPAsyncMPClient):
                  executor_class: type[Executor],
                  log_stats: bool,
                  client_addresses: Optional[dict[str, str]] = None,
+                 client_count: int = 1,
                  client_index: int = 0):
 
+        self.client_count = client_count
+
         # To route aborts to the correct engine.
         self.reqs_in_flight: dict[str, EngineIdentity] = {}
 
         super().__init__(vllm_config, executor_class, log_stats,
-                         client_addresses, client_index)
+                         client_addresses, client_count, client_index)
 
         assert len(self.core_engines) > 1
 
+        self.eng_start_index = (len(self.core_engines) *
+                                self.client_index) // client_count
+
     def get_core_engine_for_request(
             self, request: EngineCoreRequest) -> EngineIdentity:
         # Engines are in rank order.
+        current_counts = self.lb_engines
         if (eng_index := request.data_parallel_rank) is None:
-            if not self.lb_engines:
+            if not current_counts:
                 return self.core_engine
             # TODO use P2C alg for larger DP sizes
-            num_engines = len(self.lb_engines)
-            min_counts = [sys.maxsize, sys.maxsize]
+            num_engines = len(current_counts)
+            min_score = sys.maxsize
             eng_index = 0
             for i in range(num_engines):
                 # Start from client_index to help with balancing when engines
                 # are empty.
-                idx = (self.client_index + i) % num_engines
-                counts = self.lb_engines[idx]
-                if counts < min_counts:
-                    min_counts = counts
+                idx = (self.eng_start_index + i) % num_engines
+                waiting, running = current_counts[idx]
+                score = waiting * 4 + running
+                if score < min_score:
+                    min_score = score
                     eng_index = idx
-            # Adjust local counts for better balancing between stats updates
-            # from the coordinator (which happen every 100ms).
-            if min_counts[0]:
-                min_counts[0] += 1
-            else:
-                min_counts[1] += 1
+            # Increment local waiting count for better balancing between stats
+            # updates from the coordinator (which happen every 100ms).
+            current_counts[eng_index][0] += self.client_count
 
         chosen_engine = self.core_engines[eng_index]
         # Record which engine is chosen for this request, to handle aborts.

vllm/v1/metrics/stats.py

@@ -33,6 +33,10 @@ class SchedulerStats:
     num_running_reqs: int = 0
     num_waiting_reqs: int = 0
 
+    # These are used for internal DP load-balancing.
+    step_counter: int = 0
+    current_wave: int = 0
+
     kv_cache_usage: float = 0.0
 
     prefix_cache_stats: PrefixCacheStats = field(

vllm/v1/utils.py

@@ -154,6 +154,7 @@ class APIServerProcessManager:
             client_config = {
                 "input_address": in_addr,
                 "output_address": out_addr,
+                "client_count": num_servers,
                 "client_index": i
             }
             if stats_update_address is not None: