Merge remote-tracking branch 'nm/lwilkinson/fix-flashmla-full-cudagraph' into wide_ep_working_branch

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

vllm/utils/__init__.py

@@ -2827,6 +2827,9 @@ def make_zmq_socket(
     if linger is not None:
         socket.setsockopt(zmq.LINGER, linger)
 
+    if socket_type == zmq.XPUB:
+        socket.setsockopt(zmq.XPUB_VERBOSE, True)
+
     # Determine if the path is a TCP socket with an IPv6 address.
     # Enable IPv6 on the zmq socket if so.
     scheme, host, _ = split_zmq_path(path)

vllm/v1/attention/backends/mla/flashmla.py

@@ -67,6 +67,20 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
         self.cg_buf_tile_scheduler_metadata = None
         self.cg_buf_num_splits = None
 
+        device_properties = torch.cuda.get_device_properties(self.device)
+        num_sms = device_properties.multi_processor_count
+
+        if self.compilation_config.full_cuda_graph:
+            self.cg_buf_tile_scheduler_metadata = torch.empty(
+                (num_sms, 8),  # TileSchedulerMetaDataSize == 8
+                device=self.device,
+                dtype=torch.int32,
+            )
+            self.cg_buf_num_splits = torch.empty(
+                (vllm_config.scheduler_config.max_num_seqs + 1),
+                device=self.device,
+                dtype=torch.int32)
+
     def _build_decode(self, block_table_tensor: torch.Tensor,
                       seq_lens: torch.Tensor) -> FlashMLADecodeMetadata:
         tile_scheduler_metadata, num_splits = \
@@ -77,28 +91,24 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
         )
 
         if self.compilation_config.full_cuda_graph:
-            # First time around (CUDAGraph capture), allocate the static buffer
-            if self.cg_buf_tile_scheduler_metadata is None:
-                self.cg_buf_tile_scheduler_metadata = tile_scheduler_metadata
-                self.cg_buf_num_splits = num_splits
-            else:
-                assert self.cg_buf_num_splits is not None
+            assert self.cg_buf_tile_scheduler_metadata is not None
+            assert self.cg_buf_num_splits is not None
 
-                # Metadata per-SM, fixed size (#SMs, TileMetadataSize)
-                assert (self.cg_buf_tile_scheduler_metadata.size() ==
-                        tile_scheduler_metadata.size())
-                self.cg_buf_tile_scheduler_metadata.\
-                    copy_(tile_scheduler_metadata)
-                tile_scheduler_metadata = self.cg_buf_tile_scheduler_metadata
+            # Metadata per-SM, fixed size (#SMs, TileMetadataSize)
+            assert (self.cg_buf_tile_scheduler_metadata.size() ==
+                    tile_scheduler_metadata.size())
+            self.cg_buf_tile_scheduler_metadata.\
+                copy_(tile_scheduler_metadata)
+            tile_scheduler_metadata = self.cg_buf_tile_scheduler_metadata
 
-                # Num splits is per-batch, varying size (batch_size,)
-                n = num_splits.size(0)
-                # make sure static buffer is large enough
-                assert n <= self.cg_buf_num_splits.size(0)
-                num_splits_view = self.cg_buf_num_splits[:n]
-                num_splits_view.copy_(num_splits)
-                self.cg_buf_num_splits[n:].fill_(0)  # fill the rest with 0s
-                num_splits = num_splits_view
+            # Num splits is per-batch, varying size (batch_size,)
+            n = num_splits.size(0)
+            # make sure static buffer is large enough
+            assert n <= self.cg_buf_num_splits.size(0)
+            num_splits_view = self.cg_buf_num_splits[:n]
+            num_splits_view.copy_(num_splits)
+            self.cg_buf_num_splits[n:].fill_(0)  # fill the rest with 0s
+            num_splits = num_splits_view
 
         return FlashMLADecodeMetadata(
             block_table=block_table_tensor,

vllm/v1/engine/coordinator.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import copy
 import multiprocessing
 import time
 import weakref
@@ -66,18 +67,14 @@ 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,
@@ -87,7 +84,6 @@ 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()
@@ -126,10 +122,6 @@ 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,
@@ -156,6 +148,13 @@ class DPCoordinatorProc:
 
         decoder = MsgpackDecoder(EngineCoreOutputs)
 
+        current_wave = 0
+        engines_running = False
+
+        stats_changed = False
+        last_stats_step = -1
+        last_step_counts: Optional[list[list[int]]] = None
+
         with make_zmq_socket(
                 path=front_publish_address,  # IPC
                 ctx=self.ctx,
@@ -173,6 +172,18 @@ class DPCoordinatorProc:
                 bind=True,
         ) as publish_back:
 
+            # Wait until all engines subscribe.
+            for _ in self.engines:
+                if publish_back.recv() != b'\x01':
+                    logger.error(
+                        "DP Coordinator received unexpected message while"
+                        "waiting for engines to subscribe")
+                    return
+            # Send ready message to engines.
+            publish_back.send(b"READY")
+
+            logger.info("All engine subscriptions received by DP coordinator")
+
             poller = zmq.Poller()
             poller.register(publish_front, zmq.POLLIN)
             poller.register(output_back, zmq.POLLIN)
@@ -180,21 +191,33 @@ 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()
@@ -221,7 +244,7 @@ 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)
@@ -237,15 +260,15 @@ 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:
@@ -263,36 +286,47 @@ 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
+                        if stats_changed and stats_step != last_stats_step:
+                            last_step_counts = self._get_engine_counts(
+                                do_copy=True)
+                        elif stats_step < last_stats_step:
+                            logger.warning("Received stats for out-of-order "
+                                           "step from engine {eng_index}")
                         stats[0] = scheduler_stats.num_waiting_reqs
                         stats[1] = scheduler_stats.num_running_reqs
-                        self.stats_changed = True
+                        last_stats_step = stats_step
+                        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]):
@@ -305,6 +339,8 @@ 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

@@ -445,8 +445,11 @@ class EngineCoreProc(EngineCore):
             self.has_coordinator = addresses.coordinator_output is not None
             self.frontend_stats_publish_address = (
                 addresses.frontend_stats_publish_address)
+            logger.debug("Has DP Coordinator: %s, stats publish address: %s",
+                         self.has_coordinator,
+                         self.frontend_stats_publish_address)
             # Only publish request queue stats to coordinator for "internal"
-            # LB mode.
+            # and "hybrid" LB modes .
             self.publish_dp_lb_stats = (
                 self.has_coordinator
                 and not vllm_config.parallel_config.data_parallel_external_lb)
@@ -456,25 +459,38 @@ class EngineCoreProc(EngineCore):
             super().__init__(vllm_config, executor_class, log_stats,
                              executor_fail_callback)
 
+            # Background Threads and Queues for IO. These enable us to
+            # overlap ZMQ socket IO with GPU since they release the GIL,
+            # and to overlap some serialization/deserialization with the
+            # model forward pass.
+            # Threads handle Socket <-> Queues and core_busy_loop uses Queue.
+            ready_event = threading.Event()
+            input_thread = threading.Thread(target=self.process_input_sockets,
+                                            args=(addresses.inputs,
+                                                  addresses.coordinator_input,
+                                                  identity, ready_event),
+                                            daemon=True)
+            input_thread.start()
+
+            self.output_thread = threading.Thread(
+                target=self.process_output_sockets,
+                args=(addresses.outputs, addresses.coordinator_output,
+                      self.engine_index),
+                daemon=True)
+            self.output_thread.start()
+
+            # Don't complete handshake until DP coordinator ready message is
+            # received.
+            while not ready_event.wait(10):
+                if not input_thread.is_alive():
+                    raise RuntimeError(
+                        "Input socket thread died during startup")
+                assert addresses.coordinator_input is not None
+                logger.info("Waiting for READY message from DP Coordinator")
+
         self.step_fn = (self.step if self.batch_queue is None else
                         self.step_with_batch_queue)
 
-        # Background Threads and Queues for IO. These enable us to
-        # overlap ZMQ socket IO with GPU since they release the GIL,
-        # and to overlap some serialization/deserialization with the
-        # model forward pass.
-        # Threads handle Socket <-> Queues and core_busy_loop uses Queue.
-        threading.Thread(target=self.process_input_sockets,
-                         args=(addresses.inputs, addresses.coordinator_input,
-                               identity),
-                         daemon=True).start()
-        self.output_thread = threading.Thread(
-            target=self.process_output_sockets,
-            args=(addresses.outputs, addresses.coordinator_output,
-                  self.engine_index),
-            daemon=True)
-        self.output_thread.start()
-
     @contextmanager
     def _perform_handshakes(
         self,
@@ -489,10 +505,10 @@ class EngineCoreProc(EngineCore):
 
         For DP=1 or offline mode, this is with the colocated front-end process.
 
-        For DP>1 with internal loadbalancing this is with the shared front-end
+        For DP>1 with internal load-balancing this is with the shared front-end
         process which may reside on a different node.
 
-        For DP>1 with external or hybrid loadbalancing, two handshakes are
+        For DP>1 with external or hybrid load-balancing, two handshakes are
         performed:
             - With the rank 0 front-end process which retrieves the
               DP Coordinator ZMQ addresses and DP process group address.
@@ -751,7 +767,7 @@ class EngineCoreProc(EngineCore):
 
     def process_input_sockets(self, input_addresses: list[str],
                               coord_input_address: Optional[str],
-                              identity: bytes):
+                              identity: bytes, ready_event: threading.Event):
         """Input socket IO thread."""
 
         # Msgpack serialization decoding.
@@ -788,9 +804,14 @@ class EngineCoreProc(EngineCore):
                 # back to us.
                 input_socket.send(b'')
                 poller.register(input_socket, zmq.POLLIN)
+
             if coord_socket is not None:
+                # Wait for ready message from coordinator.
+                assert coord_socket.recv() == b"READY"
                 poller.register(coord_socket, zmq.POLLIN)
 
+            ready_event.set()
+            del ready_event
             while True:
                 for input_socket, _ in poller.poll():
                     # (RequestType, RequestData)
@@ -887,7 +908,7 @@ 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)
 
@@ -967,7 +988,7 @@ 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)
             self.output_queue.put_nowait(
                 (-1, EngineCoreOutputs(scheduler_stats=stats)))
 
@@ -1014,10 +1035,10 @@ class DPEngineCoreProc(EngineCoreProc):
     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:
             return True
-        self.counter = 0
+        self.step_counter = 0
 
         return ParallelConfig.has_unfinished_dp(self.dp_group,
                                                 local_unfinished)

vllm/v1/engine/core_client.py

@@ -986,7 +986,12 @@ 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
+                        #TODO TBD whether to keep this debug log
+                        logger.debug("Received counts: %s (%s)",
+                                     sliced_counts, count_slice)
 
         resources.stats_update_task = asyncio.create_task(
             run_engine_stats_update_task())
@@ -1035,27 +1040,26 @@ class DPLBAsyncMPClient(DPAsyncMPClient):
     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
+                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] += 1
 
         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 +33,8 @@ class SchedulerStats:
     num_running_reqs: int = 0
     num_waiting_reqs: int = 0
 
+    step_counter: int = 0
+
     kv_cache_usage: float = 0.0
 
     prefix_cache_stats: PrefixCacheStats = field(