| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Worker-side orchestration of Mooncake control, memory, and data planes. |
| 4 | |
| 5 | Consumer Workers expose rank-local reservations and publish received tensors. |
| 6 | Producer Workers reserve every destination shard, bind computed sources, run |
| 7 | batched Mooncake writes, and report asynchronous completion to the Scheduler. |
| 8 | """ |
| 9 | |
| 10 | from __future__ import annotations |
| 11 | |
| 12 | import math |
| 13 | import threading |
| 14 | import time |
| 15 | from collections.abc import Callable |
| 16 | from concurrent.futures import Future, ThreadPoolExecutor |
| 17 | from functools import partial |
| 18 | from typing import TYPE_CHECKING, Any, TypeVar, cast |
| 19 | |
| 20 | import torch |
| 21 | |
| 22 | from vllm.distributed.ec_transfer.ec_connector.mooncake.config import ( |
| 23 | _RESERVATION_TTL_SECONDS, |
| 24 | MooncakeECConfig, |
| 25 | ) |
| 26 | from vllm.distributed.ec_transfer.ec_connector.mooncake.control import ( |
| 27 | ConsumerControlServer, |
| 28 | ControlClient, |
| 29 | make_cancel_request, |
| 30 | ) |
| 31 | from vllm.distributed.ec_transfer.ec_connector.mooncake.memory import ( |
| 32 | ConsumerMemoryPool, |
| 33 | ProducerMemoryPool, |
| 34 | ) |
| 35 | from vllm.distributed.ec_transfer.ec_connector.mooncake.metadata import ( |
| 36 | ECMooncakeConnectorMetadata, |
| 37 | ECMooncakePushSpec, |
| 38 | ECMooncakeWorkerMetadata, |
| 39 | ) |
| 40 | from vllm.distributed.ec_transfer.ec_connector.mooncake.producer import ( |
| 41 | ProducerPushManager, |
| 42 | ProducerPushRecord, |
| 43 | ) |
| 44 | from vllm.distributed.ec_transfer.ec_connector.mooncake.reservation import ( |
| 45 | ConsumerReservationManager, |
| 46 | ConsumerReservationState, |
| 47 | ) |
| 48 | from vllm.distributed.ec_transfer.ec_connector.mooncake.transfer import ( |
| 49 | MooncakeTransfer, |
| 50 | ensure_mooncake_available, |
| 51 | ) |
| 52 | from vllm.distributed.ec_transfer.ec_connector.mooncake_store_embedding.data import ( |
| 53 | build_embedding_namespace, |
| 54 | ) |
| 55 | from vllm.logger import init_logger |
| 56 | from vllm.utils.network_utils import get_ip |
| 57 | |
| 58 | logger = init_logger(__name__) |
| 59 | |
| 60 | _T = TypeVar("_T") |
| 61 | |
| 62 | if TYPE_CHECKING: |
| 63 | from vllm.config import VllmConfig |
| 64 | |
| 65 | from ..mooncake_store_embedding.backend import ( |
| 66 | MooncakeEmbeddingStoreBackend, |
| 67 | ) |
| 68 | |
| 69 | _RESERVATION_REFRESH_SECONDS = _RESERVATION_TTL_SECONDS / 2 |
| 70 | _MAX_CANCELLED_TRANSFER_IDS = 1 << 16 |
| 71 | _CANCEL_ATTEMPTS = 2 |
| 72 | _TRANSFER_WORKERS = 4 |
| 73 | _CONTROL_WORKERS = 8 |
| 74 | |
| 75 | |
| 76 | class _FanoutError(RuntimeError): |
| 77 | """Retain shard outcomes after every started task settles.""" |
| 78 | |
| 79 | def __init__(self, error: BaseException, results: list[Any | None]) -> None: |
| 80 | self.results = results |
| 81 | super().__init__(str(error)) |
| 82 | |
| 83 | |
| 84 | class ECMooncakeWorker: |
| 85 | """Orchestrate consumer reservations and producer push batches. |
| 86 | |
| 87 | Consumers may use TP, PP, and DP. Only the first PP stage receives encoder |
| 88 | outputs, and each TP rank exposes a consecutive control port and receives |
| 89 | the same source concurrently. Producers remain unsharded and unreplicated. |
| 90 | With DP, the caller must route both halves of a request to the same replica |
| 91 | and pass that replica's control address to the producer. |
| 92 | """ |
| 93 | |
| 94 | def __init__(self, vllm_config: VllmConfig) -> None: |
| 95 | ensure_mooncake_available() |
| 96 | config = MooncakeECConfig.from_vllm_config(vllm_config) |
| 97 | self._store_config = ( |
| 98 | ( |
| 99 | build_embedding_namespace(vllm_config), |
| 100 | config.store_max_pending_items, |
| 101 | config.store_max_pending_bytes, |
| 102 | config.store_read_buffer_bytes, |
| 103 | ) |
| 104 | if config.cross_encoder_cache |
| 105 | else None |
| 106 | ) |
| 107 | self._output_store: MooncakeEmbeddingStoreBackend | None = None |
| 108 | self.is_producer = config.is_producer |
| 109 | self.is_consumer = config.is_consumer |
| 110 | self._buffer_device = config.buffer_device |
| 111 | self._control_host = config.control_host |
| 112 | self._control_port = config.control_port |
| 113 | self._transfer = MooncakeTransfer(get_ip(), config.protocol) |
| 114 | self._consumer_memory = ConsumerMemoryPool( |
| 115 | config.pool_size, |
| 116 | self._transfer, |
| 117 | ) |
| 118 | self._reservations = ConsumerReservationManager( |
| 119 | self._consumer_memory, |
| 120 | _RESERVATION_TTL_SECONDS, |
| 121 | _MAX_CANCELLED_TRANSFER_IDS, |
| 122 | ) |
| 123 | self._consumer_rank_resolved = False |
| 124 | self._is_receiving_rank = True |
| 125 | self._tp_rank = 0 |
| 126 | self._tp_size = 1 |
| 127 | self._control_server: ConsumerControlServer | None = None |
| 128 | # Worker producer |
| 129 | self._producer_memory = ProducerMemoryPool( |
| 130 | config.pool_size, |
| 131 | self._transfer, |
| 132 | ) |
| 133 | self._control_client = ControlClient(config.control_timeout_ms) |
| 134 | self._shutdown_drain_timeout_s = config.control_timeout_ms / 1000 |
| 135 | self._io_executor = ThreadPoolExecutor( |
| 136 | max_workers=_TRANSFER_WORKERS, |
| 137 | thread_name_prefix="ec-mooncake-transfer", |
| 138 | ) |
| 139 | self._control_executor = ThreadPoolExecutor( |
| 140 | max_workers=_CONTROL_WORKERS, |
| 141 | thread_name_prefix="ec-mooncake-control", |
| 142 | ) |
| 143 | self._shard_pool: ThreadPoolExecutor | None = None |
| 144 | self._shard_pool_lock = threading.Lock() |
| 145 | self._push_ready = threading.Event() |
| 146 | self._producer_pushes = ProducerPushManager(self._push_ready.set) |
| 147 | self._dispatch_stop = threading.Event() |
| 148 | self._dispatcher: threading.Thread | None = None |
| 149 | self._failed_saves: set[str] = set() |
| 150 | self._collecting_sources = False |
| 151 | self._completed_loads: set[str] = set() |
| 152 | self._failed_loads: set[str] = set() |
| 153 | self._shutdown = False |
| 154 | if self.is_producer: |
| 155 | self._dispatcher = threading.Thread( |
| 156 | target=self._dispatch_pushes, name="ec-mooncake-ready", daemon=True |
| 157 | ) |
| 158 | self._dispatcher.start() |
| 159 | |
| 160 | def _dispatch_pushes(self) -> None: |
| 161 | pending_event = False |
| 162 | while not self._dispatch_stop.is_set(): |
| 163 | self._push_ready.wait(timeout=0.001 if pending_event else None) |
| 164 | self._push_ready.clear() |
| 165 | if self._dispatch_stop.is_set(): |
| 166 | break |
| 167 | if self._collecting_sources: |
| 168 | pending_event = False |
| 169 | continue |
| 170 | pending_event = self._producer_pushes.submit_batches( |
| 171 | self._io_executor, self._push_batch |
| 172 | ) |
| 173 | |
| 174 | def _resolve_consumer_rank(self) -> None: |
| 175 | """Place this worker in the consumer receive topology.""" |
| 176 | if self._consumer_rank_resolved: |
| 177 | return |
| 178 | self._consumer_rank_resolved = True |
| 179 | try: |
| 180 | from vllm.distributed.parallel_state import get_pp_group, get_tp_group |
| 181 | |
| 182 | tp_group = get_tp_group() |
| 183 | self._tp_rank = tp_group.rank_in_group |
| 184 | self._tp_size = tp_group.world_size |
| 185 | self._is_receiving_rank = get_pp_group().is_first_rank |
| 186 | except AssertionError: |
| 187 | # Groups are only absent outside a distributed run, where this |
| 188 | # worker is the whole consumer. |
| 189 | self._tp_rank = 0 |
| 190 | self._tp_size = 1 |
| 191 | self._is_receiving_rank = True |
| 192 | |
| 193 | def start_services(self) -> None: |
| 194 | if self._store_config is not None and self._output_store is None: |
| 195 | from ..mooncake_store_embedding.backend import ( |
| 196 | MooncakeEmbeddingStoreBackend, |
| 197 | ) |
| 198 | from ..mooncake_store_embedding.store_client import ( |
| 199 | create_mooncake_embedding_store_client, |
| 200 | ) |
| 201 | |
| 202 | namespace, max_items, max_bytes, read_bytes = self._store_config |
| 203 | self._output_store = MooncakeEmbeddingStoreBackend( |
| 204 | create_mooncake_embedding_store_client(read_buffer_bytes=read_bytes), |
| 205 | namespace, |
| 206 | max_pending_items=max_items, |
| 207 | max_pending_bytes=max_bytes, |
| 208 | ) |
| 209 | if not self.is_consumer or self._control_server is not None: |
| 210 | return |
| 211 | self._resolve_consumer_rank() |
| 212 | if not self._is_receiving_rank: |
| 213 | # Later pipeline stages hold no encoder outputs, so they need |
| 214 | # neither a receive pool nor a control channel. |
| 215 | return |
| 216 | self._consumer_memory.prepare(torch.device(self._buffer_device)) |
| 217 | consumer_pool = self._consumer_memory.tensor |
| 218 | if consumer_pool is None: |
| 219 | raise RuntimeError( |
| 220 | "Mooncake push mode requires a registered consumer buffer pool." |
| 221 | ) |
| 222 | base_port = self._control_port |
| 223 | self._control_server = ConsumerControlServer( |
| 224 | # The reservation channel hands out registered-memory addresses |
| 225 | # and takes cancellations, so it listens where the Consumer |
| 226 | # advertises itself (`ec_ip`), not on every interface. |
| 227 | self._control_host, |
| 228 | base_port + self._tp_rank, |
| 229 | self._reserve_push_destination, |
| 230 | self._push_status, |
| 231 | self._reservations.complete, |
| 232 | self._reservations.cancel, |
| 233 | self._reservations.expire, |
| 234 | peer_ports=[base_port + rank for rank in range(self._tp_size)], |
| 235 | device=consumer_pool.device, |
| 236 | drain_ready=self._reservations.drain_ready, |
| 237 | ) |
| 238 | try: |
| 239 | self._control_server.start() |
| 240 | except Exception: |
| 241 | self._control_server.close() |
| 242 | self._control_server = None |
| 243 | raise |
| 244 | |
| 245 | def _reserve_push_destination(self, payload: dict[str, Any]) -> dict[str, Any]: |
| 246 | transfer_id = str(payload["transfer_id"]) |
| 247 | mm_hash = str(payload["mm_hash"]) |
| 248 | nbytes = int(payload["nbytes"]) |
| 249 | shape = tuple(int(value) for value in payload["shape"]) |
| 250 | dtype_name = str(payload["dtype"]) |
| 251 | dtype = getattr(torch, dtype_name, None) |
| 252 | if not isinstance(dtype, torch.dtype): |
| 253 | raise ValueError(f"Unsupported torch dtype string: {dtype_name!r}") |
| 254 | expected_nbytes = math.prod(shape) * dtype.itemsize |
| 255 | if expected_nbytes != nbytes: |
| 256 | raise ValueError("shape and dtype do not match nbytes") |
| 257 | |
| 258 | self._reservations.expire() |
| 259 | reservation, should_write = self._reservations.reserve( |
| 260 | transfer_id, mm_hash, nbytes, shape, dtype_name, dtype |
| 261 | ) |
| 262 | if reservation is None: |
| 263 | raise RuntimeError("EC consumer buffer pool is full") |
| 264 | if reservation.state in { |
| 265 | ConsumerReservationState.CANCEL_PENDING, |
| 266 | ConsumerReservationState.CANCELLED, |
| 267 | }: |
| 268 | return { |
| 269 | "reservation_id": "", |
| 270 | "dst_session": "", |
| 271 | "dst_ptr": 0, |
| 272 | "nbytes": nbytes, |
| 273 | "write": False, |
| 274 | "ready": False, |
| 275 | "cancelled": True, |
| 276 | } |
| 277 | assert reservation.allocation is not None |
| 278 | |
| 279 | return { |
| 280 | "reservation_id": reservation.reservation_id, |
| 281 | "dst_session": self._transfer.local_session(), |
| 282 | "dst_ptr": reservation.allocation.tensor.data_ptr(), |
| 283 | "nbytes": reservation.allocation.tensor.nbytes, |
| 284 | "write": should_write, |
| 285 | "ready": reservation.state is ConsumerReservationState.READY, |
| 286 | "cached": reservation.lease is not None, |
| 287 | } |
| 288 | |
| 289 | def _push_status(self, transfer_id: str) -> dict[str, Any] | None: |
| 290 | reservation = self._reservations.status(transfer_id) |
| 291 | if reservation is None: |
| 292 | return None |
| 293 | assert reservation.allocation is not None |
| 294 | return { |
| 295 | "mm_hash": reservation.mm_hash, |
| 296 | "ready": reservation.state is ConsumerReservationState.READY, |
| 297 | "reservation_id": reservation.reservation_id, |
| 298 | "nbytes": reservation.allocation.tensor.nbytes, |
| 299 | "shape": list(reservation.shape), |
| 300 | "dtype": reservation.dtype, |
| 301 | } |
| 302 | |
| 303 | def _shard_executor(self) -> ThreadPoolExecutor: |
| 304 | """Use a separate pool so nested shard fan-out cannot deadlock.""" |
| 305 | with self._shard_pool_lock: |
| 306 | if self._shard_pool is None: |
| 307 | self._shard_pool = ThreadPoolExecutor( |
| 308 | max_workers=32, thread_name_prefix="ec-mooncake-shard" |
| 309 | ) |
| 310 | return self._shard_pool |
| 311 | |
| 312 | def _reserve_one(self, addr: str, spec: ECMooncakePushSpec) -> dict[str, Any]: |
| 313 | result = self._control_client.request( |
| 314 | addr, |
| 315 | { |
| 316 | "op": "reserve", |
| 317 | "transfer_id": spec.transfer_id, |
| 318 | "mm_hash": spec.mm_hash, |
| 319 | "nbytes": spec.nbytes, |
| 320 | "shape": list(spec.shape), |
| 321 | "dtype": spec.dtype, |
| 322 | }, |
| 323 | ) |
| 324 | if not isinstance(result, dict): |
| 325 | raise RuntimeError("Invalid EC reservation response") |
| 326 | result["_received_at"] = time.monotonic() |
| 327 | result["addr"] = addr |
| 328 | return result |
| 329 | |
| 330 | def _run_fanout( |
| 331 | self, |
| 332 | tasks: list[Callable[[], _T]], |
| 333 | on_submit: Callable[[int, Future[_T]], None] | None = None, |
| 334 | ) -> list[_T]: |
| 335 | """Run every started shard task and retain partial results on failure. |
| 336 | |
| 337 | Waiting for all submitted tasks is what makes source-memory release and |
| 338 | partial-reservation cleanup safe after one shard fails. |
| 339 | """ |
| 340 | if not tasks: |
| 341 | return [] |
| 342 | futures: list[tuple[int, Future[_T]]] = [] |
| 343 | results: list[_T | None] = [None] * len(tasks) |
| 344 | error: BaseException | None = None |
| 345 | for index, task in enumerate(tasks[1:], 1): |
| 346 | try: |
| 347 | future = self._shard_executor().submit(task) |
| 348 | except Exception as exc: |
| 349 | error = exc |
| 350 | break |
| 351 | futures.append((index, future)) |
| 352 | if on_submit is not None: |
| 353 | on_submit(index, future) |
| 354 | if error is None: |
| 355 | try: |
| 356 | results[0] = tasks[0]() |
| 357 | except Exception as exc: |
| 358 | error = exc |
| 359 | for index, future in futures: |
| 360 | try: |
| 361 | results[index] = future.result() |
| 362 | except Exception as exc: |
| 363 | if error is None: |
| 364 | error = exc |
| 365 | if error is not None: |
| 366 | raise _FanoutError(error, results) |
| 367 | return cast(list[_T], results) |
| 368 | |
| 369 | def _cancel_reservations( |
| 370 | self, |
| 371 | spec: ECMooncakePushSpec, |
| 372 | reservations: list[dict[str, Any]], |
| 373 | *, |
| 374 | refresh: bool = False, |
| 375 | record: ProducerPushRecord | None = None, |
| 376 | ) -> None: |
| 377 | reservations = [ |
| 378 | shard for shard in reservations if not shard.get("cancelled", False) |
| 379 | ] |
| 380 | if not reservations: |
| 381 | return |
| 382 | |
| 383 | def cancel(shard: dict[str, Any]) -> dict[str, Any]: |
| 384 | addr = shard.get("addr") |
| 385 | if not isinstance(addr, str) or not addr: |
| 386 | raise RuntimeError( |
| 387 | "EC reservation is missing a confirmed shard address" |
| 388 | ) |
| 389 | result = self._control_client.request( |
| 390 | addr, |
| 391 | make_cancel_request( |
| 392 | spec.transfer_id, |
| 393 | str(shard.get("reservation_id", "")), |
| 394 | abandon=True, |
| 395 | refresh=refresh, |
| 396 | ), |
| 397 | ) |
| 398 | if not isinstance(result, dict) or not result.get("cancelled"): |
| 399 | raise RuntimeError( |
| 400 | f"Could not cancel EC reservation for mm_hash={spec.mm_hash}" |
| 401 | ) |
| 402 | return shard |
| 403 | |
| 404 | def track(_index: int, future: Future[dict[str, Any]]) -> None: |
| 405 | if record is not None: |
| 406 | self._producer_pushes.track_shard_futures([record], [future]) |
| 407 | |
| 408 | self._run_fanout([partial(cancel, shard) for shard in reservations], track) |
| 409 | |
| 410 | def _retry_cancel_reservations( |
| 411 | self, |
| 412 | spec: ECMooncakePushSpec, |
| 413 | reservations: list[dict[str, Any]], |
| 414 | *, |
| 415 | record: ProducerPushRecord | None = None, |
| 416 | ) -> None: |
| 417 | pending = [shard for shard in reservations if not shard.get("cancelled", False)] |
| 418 | error: _FanoutError | None = None |
| 419 | for _ in range(_CANCEL_ATTEMPTS): |
| 420 | try: |
| 421 | self._cancel_reservations(spec, pending, record=record) |
| 422 | except _FanoutError as exc: |
| 423 | error = exc |
| 424 | pending = [ |
| 425 | shard |
| 426 | for index, shard in enumerate(pending) |
| 427 | if exc.results[index] is None |
| 428 | ] |
| 429 | continue |
| 430 | return |
| 431 | assert error is not None |
| 432 | raise error |
| 433 | |
| 434 | def _reserve_remote(self, spec: ECMooncakePushSpec) -> list[dict[str, Any]]: |
| 435 | """Reserve a destination on every shard of the consumer.""" |
| 436 | result = self._reserve_remote_many([spec])[0] |
| 437 | if isinstance(result, Exception): |
| 438 | raise result |
| 439 | return result |
| 440 | |
| 441 | def _refresh_remote_reservations( |
| 442 | self, |
| 443 | spec: ECMooncakePushSpec, |
| 444 | reservations: list[dict[str, Any]], |
| 445 | record: ProducerPushRecord | None = None, |
| 446 | ) -> list[dict[str, Any]]: |
| 447 | stale = [ |
| 448 | shard |
| 449 | for shard in reservations |
| 450 | if not shard.get("ready", False) |
| 451 | and not shard.get("cached", False) |
| 452 | and not shard.get("cancelled", False) |
| 453 | ] |
| 454 | try: |
| 455 | self._cancel_reservations(spec, stale, refresh=True, record=record) |
| 456 | except _FanoutError as exc: |
| 457 | pending = [ |
| 458 | shard for index, shard in enumerate(stale) if exc.results[index] is None |
| 459 | ] |
| 460 | try: |
| 461 | self._retry_cancel_reservations(spec, pending, record=record) |
| 462 | except _FanoutError as cleanup_error: |
| 463 | raise exc from cleanup_error |
| 464 | raise |
| 465 | return self._reserve_remote(spec) |
| 466 | |
| 467 | @staticmethod |
| 468 | def _validate_push_source(push: ProducerPushRecord) -> None: |
| 469 | tensor = push.source_tensor |
| 470 | assert tensor is not None |
| 471 | spec = push.spec |
| 472 | if tuple(tensor.shape) != tuple(spec.shape): |
| 473 | raise ValueError(f"EC source shape mismatch for mm_hash={spec.mm_hash}") |
| 474 | if str(tensor.dtype).split(".")[-1] != spec.dtype: |
| 475 | raise ValueError(f"EC source dtype mismatch for mm_hash={spec.mm_hash}") |
| 476 | if not tensor.is_contiguous(): |
| 477 | raise ValueError(f"EC source must be contiguous for mm_hash={spec.mm_hash}") |
| 478 | if tensor.nbytes != spec.nbytes: |
| 479 | raise ValueError(f"EC source size mismatch for mm_hash={spec.mm_hash}") |
| 480 | |
| 481 | def start_save_caches( |
| 482 | self, |
| 483 | metadata: ECMooncakeConnectorMetadata, |
| 484 | encoder_cache: dict[str, torch.Tensor] | None = None, |
| 485 | **kwargs: Any, |
| 486 | ) -> None: |
| 487 | self._collecting_sources = True |
| 488 | new: dict[str, list[ProducerPushRecord]] = {} |
| 489 | for spec in metadata.pushes: |
| 490 | try: |
| 491 | record, created = self._producer_pushes.reserve(spec, Future) |
| 492 | if created: |
| 493 | new.setdefault(spec.consumer_zmq, []).append(record) |
| 494 | except ValueError: |
| 495 | logger.warning("Rejected conflicting EC push", exc_info=True) |
| 496 | self._failed_saves.add(spec.request_id) |
| 497 | for records in new.values(): |
| 498 | future = self._control_executor.submit(self._reserve_batch, records) |
| 499 | future.add_done_callback(partial(self._reservation_batch_done, records)) |
| 500 | if not isinstance(encoder_cache, dict): |
| 501 | return |
| 502 | for mm_hash in dict.fromkeys(spec.mm_hash for spec in metadata.pushes): |
| 503 | tensor = encoder_cache.get(mm_hash) |
| 504 | if tensor is not None: |
| 505 | self._bind_push_source(tensor, mm_hash) |
| 506 | if self._output_store is not None: |
| 507 | loaded = self._output_store.resolve_inputs( |
| 508 | metadata.store_candidates, encoder_cache, self._buffer_device |
| 509 | ) |
| 510 | # Loaded entries precede the runner's new-output snapshot, so they |
| 511 | # need explicit P2P source binding here. |
| 512 | for mm_hash in loaded: |
| 513 | self._bind_push_source(encoder_cache[mm_hash], mm_hash) |
| 514 | |
| 515 | def _reserve_batch(self, records: list[ProducerPushRecord]) -> None: |
| 516 | """Resolve independent item futures with one reserve RPC per shard.""" |
| 517 | if len(records) == 1: |
| 518 | record = records[0] |
| 519 | try: |
| 520 | record.reservation_future.set_result(self._reserve_remote(record.spec)) |
| 521 | except Exception as exc: |
| 522 | record.reservation_future.set_exception(exc) |
| 523 | return |
| 524 | outcomes = self._reserve_remote_many([record.spec for record in records]) |
| 525 | self._producer_pushes.finish_reservations(list(zip(records, outcomes))) |
| 526 | |
| 527 | def _reserve_remote_many( |
| 528 | self, specs: list[ECMooncakePushSpec] |
| 529 | ) -> list[list[dict[str, Any]] | Exception]: |
| 530 | shards = None |
| 531 | for _ in range(_CANCEL_ATTEMPTS): |
| 532 | shards = self._control_client.discover_shards(specs[0].consumer_zmq) |
| 533 | if shards is not None: |
| 534 | break |
| 535 | if shards is None: |
| 536 | raise RuntimeError( |
| 537 | f"Could not discover every EC consumer shard at {specs[0].consumer_zmq}" |
| 538 | ) |
| 539 | |
| 540 | def reserve(addr: str) -> list[dict[str, Any]]: |
| 541 | if len(specs) == 1: |
| 542 | return [{"ok": True, "result": self._reserve_one(addr, specs[0])}] |
| 543 | response = self._control_client.request( |
| 544 | addr, |
| 545 | { |
| 546 | "op": "reserve_batch", |
| 547 | "items": [ |
| 548 | { |
| 549 | "transfer_id": spec.transfer_id, |
| 550 | "mm_hash": spec.mm_hash, |
| 551 | "nbytes": spec.nbytes, |
| 552 | "shape": list(spec.shape), |
| 553 | "dtype": spec.dtype, |
| 554 | } |
| 555 | for spec in specs |
| 556 | ], |
| 557 | }, |
| 558 | ) |
| 559 | items = response["items"] |
| 560 | if len(items) != len(specs): |
| 561 | raise RuntimeError("Malformed EC reservation batch response") |
| 562 | for item in items: |
| 563 | if item.get("ok"): |
| 564 | item["result"]["_received_at"] = time.monotonic() |
| 565 | return items |
| 566 | |
| 567 | fanout_error = None |
| 568 | results: list[Any] |
| 569 | try: |
| 570 | results = self._run_fanout([partial(reserve, addr) for addr in shards]) |
| 571 | except _FanoutError as exc: |
| 572 | results = exc.results |
| 573 | fanout_error = exc |
| 574 | outcomes: list[list[dict[str, Any]] | Exception] = [] |
| 575 | for index, spec in enumerate(specs): |
| 576 | reservations: list[dict[str, Any]] = [] |
| 577 | error = None |
| 578 | for addr, items in zip(shards, results): |
| 579 | item = items[index] if items is not None else {} |
| 580 | if item.get("ok"): |
| 581 | reservations.append({**item["result"], "addr": addr}) |
| 582 | else: |
| 583 | error = fanout_error or RuntimeError(item["error"]) |
| 584 | reservations.append({"addr": addr, "reservation_id": ""}) |
| 585 | if error is None: |
| 586 | outcomes.append(reservations) |
| 587 | else: |
| 588 | failure = _FanoutError(error, list(reservations)) |
| 589 | try: |
| 590 | self._retry_cancel_reservations(spec, reservations) |
| 591 | except Exception as cleanup_error: |
| 592 | failure.__cause__ = cleanup_error |
| 593 | logger.exception("Failed to release partial EC reservation batch") |
| 594 | outcomes.append(failure) |
| 595 | return outcomes |
| 596 | |
| 597 | @staticmethod |
| 598 | def _reservation_batch_done( |
| 599 | records: list[ProducerPushRecord], future: Future[None] |
| 600 | ) -> None: |
| 601 | error = future.exception() |
| 602 | if error is not None: |
| 603 | for record in records: |
| 604 | if not record.reservation_future.done(): |
| 605 | record.reservation_future.set_exception(error) |
| 606 | |
| 607 | def start_load_caches( |
| 608 | self, |
| 609 | metadata: ECMooncakeConnectorMetadata, |
| 610 | encoder_cache: dict[str, torch.Tensor], |
| 611 | **kwargs: Any, |
| 612 | ) -> None: |
| 613 | self._resolve_consumer_rank() |
| 614 | if not self._is_receiving_rank: |
| 615 | # Later pipeline stages never gather multimodal embeddings. |
| 616 | return |
| 617 | self._transfer.ensure_ready() |
| 618 | if self._buffer_device == "cuda" and not torch.accelerator.is_available(): |
| 619 | raise RuntimeError( |
| 620 | "ECMooncakeConnector requires CUDA for ec_buffer_device=cuda" |
| 621 | ) |
| 622 | self._reservations.retire_stale(encoder_cache, metadata.freed) |
| 623 | |
| 624 | for spec in metadata.loads: |
| 625 | if spec.mm_hash in encoder_cache: |
| 626 | if not spec.local: |
| 627 | # The spec's id is one shard's; cancel by transfer. |
| 628 | self._reservations.cancel(spec.transfer_id, "") |
| 629 | self._completed_loads.add(spec.mm_hash) |
| 630 | continue |
| 631 | if spec.local: |
| 632 | tensor = self._consumer_memory.take_resident( |
| 633 | spec.mm_hash, tuple(spec.shape), spec.dtype |
| 634 | ) |
| 635 | else: |
| 636 | try: |
| 637 | allocation = self._reservations.take(spec.transfer_id, spec.mm_hash) |
| 638 | tensor = allocation.tensor |
| 639 | except RuntimeError as e: |
| 640 | logger.warning("EC Mooncake pushed load failed: %s", e) |
| 641 | tensor = None |
| 642 | if tensor is None: |
| 643 | self._failed_loads.add(spec.mm_hash) |
| 644 | else: |
| 645 | encoder_cache[spec.mm_hash] = tensor |
| 646 | self._completed_loads.add(spec.mm_hash) |
| 647 | |
| 648 | def _push_batch(self, pushes: list[ProducerPushRecord]) -> None: |
| 649 | started_at = time.monotonic() |
| 650 | ready: list[tuple[ProducerPushRecord, dict[str, Any]]] = [] |
| 651 | written_pushes: dict[str, ProducerPushRecord] = {} |
| 652 | failure: Exception | None = None |
| 653 | try: |
| 654 | for push in pushes: |
| 655 | self._validate_push_source(push) |
| 656 | reservations = self._producer_pushes.resolve_reservations(push) |
| 657 | stale = [ |
| 658 | index |
| 659 | for index, shard in enumerate(reservations) |
| 660 | if not shard.get("ready", False) |
| 661 | and not shard.get("cancelled", False) |
| 662 | and time.monotonic() - float(shard.get("_received_at", started_at)) |
| 663 | >= _RESERVATION_REFRESH_SECONDS |
| 664 | ] |
| 665 | if stale: |
| 666 | reservations = self._refresh_remote_reservations( |
| 667 | push.spec, reservations, push |
| 668 | ) |
| 669 | self._producer_pushes.replace_reservations(push, reservations) |
| 670 | self._producer_pushes.begin_writing(push) |
| 671 | writable = [ |
| 672 | shard |
| 673 | for shard in reservations |
| 674 | if not shard.get("cached", False) |
| 675 | and not shard.get("cancelled", False) |
| 676 | and shard.get("write", True) |
| 677 | ] |
| 678 | source = push.source_tensor |
| 679 | assert source is not None |
| 680 | for shard in writable: |
| 681 | if int(shard["nbytes"]) != source.nbytes: |
| 682 | raise RuntimeError( |
| 683 | "Reserved EC size does not match tensor for " |
| 684 | f"mm_hash={push.spec.mm_hash}" |
| 685 | ) |
| 686 | ready.append((push, shard)) |
| 687 | written_pushes.setdefault(push.spec.transfer_id, push) |
| 688 | if ready: |
| 689 | # Stage each source once, then write it to every destination. |
| 690 | source_index = { |
| 691 | push.spec.transfer_id: index |
| 692 | for index, push in enumerate(written_pushes.values()) |
| 693 | } |
| 694 | tensors = [ |
| 695 | cast(torch.Tensor, push.source_tensor) |
| 696 | for push in written_pushes.values() |
| 697 | if push.source_tensor is not None |
| 698 | ] |
| 699 | lengths = [tensor.nbytes for tensor in tensors] |
| 700 | staged = self._producer_memory.stage(tensors) |
| 701 | registered_sources: list[int] = [] |
| 702 | if staged is not None: |
| 703 | sources = staged.tensors |
| 704 | else: |
| 705 | sources = tensors |
| 706 | registered_sources = self._transfer.acquire_sources(tensors) |
| 707 | addresses = [tensor.data_ptr() for tensor in sources] |
| 708 | try: |
| 709 | by_session: dict[str, list[tuple[int, int]]] = {} |
| 710 | session_records: dict[str, dict[str, ProducerPushRecord]] = {} |
| 711 | for push, shard in ready: |
| 712 | session = str(shard["dst_session"]) |
| 713 | by_session.setdefault(session, []).append( |
| 714 | (source_index[push.spec.transfer_id], int(shard["dst_ptr"])) |
| 715 | ) |
| 716 | session_records.setdefault(session, {})[ |
| 717 | push.spec.transfer_id |
| 718 | ] = push |
| 719 | |
| 720 | def write(session: str, items: list[tuple[int, int]]) -> None: |
| 721 | self._transfer.write( |
| 722 | session, |
| 723 | [addresses[index] for index, _ in items], |
| 724 | [dst for _, dst in items], |
| 725 | [lengths[index] for index, _ in items], |
| 726 | ) |
| 727 | |
| 728 | sessions = list(by_session.items()) |
| 729 | |
| 730 | # Write shards concurrently to avoid serial TP latency. |
| 731 | def track_write(index: int, future: Future[None]) -> None: |
| 732 | session = sessions[index][0] |
| 733 | self._producer_pushes.track_shard_futures( |
| 734 | list(session_records[session].values()), [future] |
| 735 | ) |
| 736 | |
| 737 | writes: list[Callable[[], None]] = [ |
| 738 | partial(write, *session) for session in sessions |
| 739 | ] |
| 740 | self._run_fanout(writes, track_write) |
| 741 | finally: |
| 742 | if staged is not None: |
| 743 | self._producer_memory.release(staged) |
| 744 | self._transfer.release_sources(registered_sources) |
| 745 | |
| 746 | self._producer_pushes.begin_notifying(pushes) |
| 747 | self._notify_completions(ready) |
| 748 | self._producer_pushes.complete(pushes) |
| 749 | except Exception as exc: |
| 750 | # Report asynchronously; raising here would fail EngineCore. |
| 751 | failure = exc |
| 752 | logger.exception( |
| 753 | "EC Mooncake push batch failed for mm_hashes=%s", |
| 754 | [push.spec.mm_hash for push in pushes], |
| 755 | ) |
| 756 | self._producer_pushes.settle_all(pushes) |
| 757 | self._abandon_pushes(pushes) |
| 758 | finally: |
| 759 | if failure is not None: |
| 760 | self._producer_pushes.fail(pushes, failure) |
| 761 | |
| 762 | def _notify_completions( |
| 763 | self, notifications: list[tuple[ProducerPushRecord, dict[str, Any]]] |
| 764 | ) -> None: |
| 765 | """Tell the consumer, in one message per destination, what landed.""" |
| 766 | if not notifications: |
| 767 | return |
| 768 | by_destination: dict[str, list[tuple[ProducerPushRecord, dict[str, Any]]]] = {} |
| 769 | for push, reservation in notifications: |
| 770 | by_destination.setdefault( |
| 771 | str(reservation.get("addr", push.spec.consumer_zmq)), [] |
| 772 | ).append((push, reservation)) |
| 773 | destinations = list(by_destination.items()) |
| 774 | |
| 775 | def notify( |
| 776 | consumer_zmq: str, |
| 777 | items: list[tuple[ProducerPushRecord, dict[str, Any]]], |
| 778 | ) -> None: |
| 779 | result = self._control_client.request( |
| 780 | consumer_zmq, |
| 781 | { |
| 782 | "op": "complete_batch", |
| 783 | "items": [ |
| 784 | { |
| 785 | "transfer_id": push.spec.transfer_id, |
| 786 | "reservation_id": reservation["reservation_id"], |
| 787 | } |
| 788 | for push, reservation in items |
| 789 | ], |
| 790 | }, |
| 791 | ) |
| 792 | completions = result.get("items", []) if isinstance(result, dict) else [] |
| 793 | if len(completions) != len(items): |
| 794 | raise RuntimeError("Malformed EC completion response") |
| 795 | for (push, _), completion in zip(items, completions): |
| 796 | if not completion.get("completed"): |
| 797 | raise RuntimeError( |
| 798 | f"Unknown EC reservation for mm_hash={push.spec.mm_hash}" |
| 799 | ) |
| 800 | |
| 801 | def track(index: int, future: Future[None]) -> None: |
| 802 | records = { |
| 803 | push.spec.transfer_id: push for push, _ in destinations[index][1] |
| 804 | } |
| 805 | self._producer_pushes.track_shard_futures(list(records.values()), [future]) |
| 806 | |
| 807 | self._run_fanout( |
| 808 | [ |
| 809 | partial(notify, destination, items) |
| 810 | for destination, items in destinations |
| 811 | ], |
| 812 | track, |
| 813 | ) |
| 814 | |
| 815 | @staticmethod |
| 816 | def _known_reservations(record: ProducerPushRecord) -> list[dict[str, Any]]: |
| 817 | if record.reservations: |
| 818 | return list(record.reservations) |
| 819 | try: |
| 820 | return list(record.reservation_future.result()) |
| 821 | except _FanoutError as exc: |
| 822 | return [result for result in exc.results if isinstance(result, dict)] |
| 823 | except Exception: |
| 824 | return [] |
| 825 | |
| 826 | def _abandon_pushes(self, pushes: list[ProducerPushRecord]) -> None: |
| 827 | """Release the consumer-side reservations of a batch that failed.""" |
| 828 | for push in pushes: |
| 829 | shards = self._known_reservations(push) |
| 830 | if not shards: |
| 831 | # No confirmed topology means there is no safe address to |
| 832 | # cancel. The original push failure remains observable via |
| 833 | # ProducerPushManager.fail below. |
| 834 | continue |
| 835 | try: |
| 836 | self._retry_cancel_reservations(push.spec, shards, record=push) |
| 837 | except _FanoutError: |
| 838 | logger.exception( |
| 839 | "Failed to abandon EC reservations for transfer_id=%s", |
| 840 | push.spec.transfer_id, |
| 841 | ) |
| 842 | |
| 843 | def _flush_pending_pushes(self) -> None: |
| 844 | self._producer_pushes.submit_batches( |
| 845 | self._io_executor, |
| 846 | self._push_batch, |
| 847 | ) |
| 848 | |
| 849 | def _bind_push_source(self, tensor: torch.Tensor, mm_hash: str) -> None: |
| 850 | ready_event = None |
| 851 | if tensor.device.type == "cuda": |
| 852 | ready_event = torch.Event() |
| 853 | ready_event.record(torch.accelerator.current_stream(tensor.device)) |
| 854 | self._producer_pushes.bind_source(mm_hash, tensor, ready_event) |
| 855 | |
| 856 | def _cancel_orphaned_reservation(self, record: ProducerPushRecord) -> None: |
| 857 | resolution_error: Exception | None = None |
| 858 | try: |
| 859 | reservations = self._producer_pushes.resolve_reservations(record) |
| 860 | except Exception as exc: |
| 861 | resolution_error = exc |
| 862 | reservations = self._known_reservations(record) |
| 863 | reservations = [ |
| 864 | shard for shard in reservations if not shard.get("cancelled", False) |
| 865 | ] |
| 866 | if not reservations: |
| 867 | if resolution_error is not None: |
| 868 | self._producer_pushes.fail([record], resolution_error) |
| 869 | else: |
| 870 | self._producer_pushes.finish_cancel(record) |
| 871 | return |
| 872 | try: |
| 873 | self._retry_cancel_reservations(record.spec, reservations, record=record) |
| 874 | except Exception as cleanup_error: |
| 875 | if resolution_error is not None: |
| 876 | combined_error = RuntimeError( |
| 877 | f"EC reservation resolution failed ({resolution_error}); " |
| 878 | f"cleanup also failed ({cleanup_error})" |
| 879 | ) |
| 880 | combined_error.__cause__ = resolution_error |
| 881 | cleanup_error = combined_error |
| 882 | self._producer_pushes.fail([record], cleanup_error) |
| 883 | return |
| 884 | if resolution_error is not None: |
| 885 | self._producer_pushes.fail([record], resolution_error) |
| 886 | else: |
| 887 | self._producer_pushes.finish_cancel(record) |
| 888 | |
| 889 | def get_finished( |
| 890 | self, finished_req_ids: set[str] |
| 891 | ) -> tuple[set[str] | None, set[str] | None]: |
| 892 | if not self.is_producer: |
| 893 | return None, None |
| 894 | |
| 895 | if self._output_store is not None: |
| 896 | self._output_store.reap() |
| 897 | for record in self._producer_pushes.cancel_requests(finished_req_ids): |
| 898 | self._producer_pushes.submit_cancel( |
| 899 | record, |
| 900 | self._io_executor, |
| 901 | self._cancel_orphaned_reservation, |
| 902 | ) |
| 903 | return None, None |
| 904 | |
| 905 | def save_caches( |
| 906 | self, encoder_cache: dict[str, torch.Tensor], mm_hash: str, **kwargs: Any |
| 907 | ) -> None: |
| 908 | if not self.is_producer: |
| 909 | return |
| 910 | tensor = encoder_cache[mm_hash] |
| 911 | self._bind_push_source(tensor, mm_hash) |
| 912 | if self._output_store is not None: |
| 913 | self._output_store.save_output(mm_hash, tensor) |
| 914 | |
| 915 | def build_connector_worker_meta(self) -> ECMooncakeWorkerMetadata | None: |
| 916 | if self.is_consumer and not self._is_receiving_rank: |
| 917 | # `loaded` is intersected across reporting ranks, so a stage that |
| 918 | # never loads must not report at all rather than report nothing. |
| 919 | return None |
| 920 | |
| 921 | self._collecting_sources = False |
| 922 | self._push_ready.set() |
| 923 | self._flush_pending_pushes() |
| 924 | failures = self._producer_pushes.poll() |
| 925 | for mm_hash, error in failures: |
| 926 | logger.error( |
| 927 | "EC Mooncake async save failed for mm_hash=%s: %s", |
| 928 | mm_hash, |
| 929 | error, |
| 930 | ) |
| 931 | reclaimed = self._consumer_memory.drain_reclaimed() |
| 932 | meta = ECMooncakeWorkerMetadata( |
| 933 | loaded=self._completed_loads, |
| 934 | failed_loads=self._failed_loads, |
| 935 | reclaimed=reclaimed, |
| 936 | pending_saves=self._producer_pushes.pending, |
| 937 | failed_saves=self._failed_saves, |
| 938 | ) |
| 939 | self._failed_saves = set() |
| 940 | self._completed_loads = set() |
| 941 | self._failed_loads = set() |
| 942 | return meta |
| 943 | |
| 944 | def close(self) -> None: |
| 945 | if self._shutdown: |
| 946 | return |
| 947 | self._shutdown = True |
| 948 | dispatcher = getattr(self, "_dispatcher", None) |
| 949 | if dispatcher is not None: |
| 950 | self._dispatch_stop.set() |
| 951 | self._push_ready.set() |
| 952 | dispatcher.join() |
| 953 | if self._control_server is not None: |
| 954 | self._reservations.begin_shutdown() |
| 955 | self._flush_pending_pushes() |
| 956 | for record in self._producer_pushes.cancel_requests(None): |
| 957 | self._producer_pushes.submit_cancel( |
| 958 | record, |
| 959 | self._io_executor, |
| 960 | self._cancel_orphaned_reservation, |
| 961 | ) |
| 962 | self._control_executor.shutdown(wait=True) |
| 963 | self._producer_pushes.submit_batches( |
| 964 | self._io_executor, self._push_batch, wait=True |
| 965 | ) |
| 966 | self._io_executor.shutdown(wait=True) |
| 967 | if self._shard_pool is not None: |
| 968 | self._shard_pool.shutdown(wait=True) |
| 969 | # Every producer-side thread that could hold a control socket is stopped. |
| 970 | self._control_client.close() |
| 971 | drained = True |
| 972 | if self._control_server is not None: |
| 973 | drained = self._reservations.wait_for_writers( |
| 974 | self._shutdown_drain_timeout_s |
| 975 | ) |
| 976 | self._control_server.close() |
| 977 | if drained: |
| 978 | self._consumer_memory.close() |
| 979 | else: |
| 980 | logger.error( |
| 981 | "Timed out waiting for Mooncake EC writers; keeping the consumer " |
| 982 | "receive pool registered" |
| 983 | ) |
| 984 | self._producer_memory.close() |
| 985 | self._transfer.close() |
| 986 | if self._output_store is not None: |
| 987 | self._output_store.shutdown() |
| 988 | self._output_store = None |