Sync from v0.13

This commit is contained in:
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions

View File

@@ -0,0 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from .abstract import Executor
from .uniproc_executor import UniProcExecutor
__all__ = ["Executor", "UniProcExecutor"]

View File

@@ -0,0 +1,352 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
from abc import ABC, abstractmethod
from collections.abc import Callable
from concurrent.futures import Future
from functools import cached_property
from typing import TYPE_CHECKING, Literal, TypeVar, overload
from vllm.config import VllmConfig
from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator
from vllm.distributed.kv_transfer.kv_connector.v1.base import (
KVConnectorHandshakeMetadata,
)
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.tasks import SupportedTask
from vllm.utils.import_utils import resolve_obj_by_qualname
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.engine import ReconfigureDistributedRequest
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
from vllm.v1.outputs import DraftTokenIds, ModelRunnerOutput
from vllm.v1.worker.worker_base import WorkerBase
if TYPE_CHECKING:
from vllm.distributed.kv_transfer.kv_connector.base import KVConnectorBase
logger = init_logger(__name__)
_R = TypeVar("_R")
FailureCallback = Callable[[], None]
class Executor(ABC):
"""Abstract base class for vLLM executors."
An executor is responsible for executing the model on one device,
or it can be a distributed executor that can execute the model on multiple devices.
"""
uses_ray: bool = False # whether the executor uses Ray for orchestration.
supports_pp: bool = False # whether the executor supports PP
@staticmethod
def get_class(vllm_config: VllmConfig) -> type["Executor"]:
executor_class: type[Executor]
parallel_config = vllm_config.parallel_config
distributed_executor_backend = parallel_config.distributed_executor_backend
# distributed_executor_backend must be set in VllmConfig.__post_init__
if isinstance(distributed_executor_backend, type):
if not issubclass(distributed_executor_backend, Executor):
raise TypeError(
"distributed_executor_backend must be a subclass of "
f"Executor. Got {distributed_executor_backend}."
)
executor_class = distributed_executor_backend
elif distributed_executor_backend == "ray":
from vllm.v1.executor.ray_executor import RayDistributedExecutor
executor_class = RayDistributedExecutor
elif distributed_executor_backend == "mp":
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
executor_class = MultiprocExecutor
elif distributed_executor_backend == "uni":
from vllm.v1.executor.uniproc_executor import UniProcExecutor
executor_class = UniProcExecutor
elif distributed_executor_backend == "external_launcher":
# TODO: make v1 scheduling deterministic
# to support external launcher
executor_class = ExecutorWithExternalLauncher
elif isinstance(distributed_executor_backend, str):
executor_class = resolve_obj_by_qualname(distributed_executor_backend)
if not issubclass(executor_class, Executor):
raise TypeError(
"distributed_executor_backend must be a subclass of "
f"Executor. Got {executor_class}."
)
else:
raise ValueError(
f"Unknown distributed executor backend: {distributed_executor_backend}"
)
return executor_class
def __init__(
self,
vllm_config: VllmConfig,
) -> None:
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.lora_config = vllm_config.lora_config
self.load_config = vllm_config.load_config
self.parallel_config = vllm_config.parallel_config
self.scheduler_config = vllm_config.scheduler_config
self.device_config = vllm_config.device_config
self.speculative_config = vllm_config.speculative_config
self.observability_config = vllm_config.observability_config
self._init_executor()
self.is_sleeping = False
self.sleeping_tags: set[str] = set()
self.kv_output_aggregator: KVOutputAggregator | None = None
@abstractmethod
def _init_executor(self) -> None:
raise NotImplementedError
def initialize_from_config(self, kv_cache_configs: list[KVCacheConfig]) -> None:
"""
Initialize the KV caches and begin the model execution loop of the
underlying workers.
"""
self.collective_rpc("initialize_from_config", args=(kv_cache_configs,))
self.collective_rpc("compile_or_warm_up_model")
def register_failure_callback(self, callback: FailureCallback): # noqa: B027
"""
Register a function to be called if the executor enters a permanent
failed state.
"""
pass
def determine_available_memory(self) -> list[int]: # in bytes
return self.collective_rpc("determine_available_memory")
def get_kv_cache_specs(self) -> list[dict[str, KVCacheSpec]]:
return self.collective_rpc("get_kv_cache_spec")
@overload
def collective_rpc(
self,
method: str | Callable[[WorkerBase], _R],
timeout: float | None = None,
args: tuple = (),
kwargs: dict | None = None,
non_block: Literal[False] = False,
) -> list[_R]:
"""
Execute an RPC call on all workers.
Args:
method: Name of the worker method to execute, or a callable that
is serialized and sent to all workers to execute.
If the method is a callable, it should accept an additional
`self` argument, in addition to the arguments passed in `args`
and `kwargs`. The `self` argument will be the worker object.
timeout: Maximum time in seconds to wait for execution. Raises a
[`TimeoutError`][] on timeout. `None` means wait indefinitely.
args: Positional arguments to pass to the worker method.
kwargs: Keyword arguments to pass to the worker method.
non_block: If `True`, returns a list of Futures instead of waiting
for the results.
Returns:
A list containing the results from each worker.
Note:
It is recommended to use this API to only pass control messages,
and set up data-plane communication to pass data.
"""
pass
@overload
def collective_rpc(
self,
method: str | Callable[[WorkerBase], _R],
timeout: float | None = None,
args: tuple = (),
kwargs: dict | None = None,
non_block: Literal[True] = True,
) -> Future[list[_R]]:
pass
@abstractmethod
def collective_rpc(
self, method, timeout=None, args=(), kwargs=None, non_block: bool = False
):
raise NotImplementedError
def get_kv_connector_handshake_metadata(
self,
) -> list[dict[int, KVConnectorHandshakeMetadata]]:
return self.collective_rpc("get_kv_connector_handshake_metadata")
@overload
def execute_model(
self, scheduler_output: SchedulerOutput, non_block: Literal[False] = False
) -> ModelRunnerOutput | None:
pass
@overload
def execute_model(
self, scheduler_output: SchedulerOutput, non_block: Literal[True] = True
) -> Future[ModelRunnerOutput | None]:
pass
def execute_model(
self, scheduler_output: SchedulerOutput, non_block: bool = False
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
output = self.collective_rpc( # type: ignore[call-overload]
"execute_model", args=(scheduler_output,), non_block=non_block
)
return output[0]
@overload
def sample_tokens(
self, grammar_output: GrammarOutput | None, non_block: Literal[False] = False
) -> ModelRunnerOutput:
pass
@overload
def sample_tokens(
self, grammar_output: GrammarOutput | None, non_block: Literal[True] = True
) -> Future[ModelRunnerOutput]:
pass
def sample_tokens(
self, grammar_output: GrammarOutput | None, non_block: bool = False
) -> ModelRunnerOutput | Future[ModelRunnerOutput]:
output = self.collective_rpc( # type: ignore[call-overload]
"sample_tokens", args=(grammar_output,), non_block=non_block
)
return output[0]
def execute_dummy_batch(self) -> None:
self.collective_rpc("execute_dummy_batch")
def take_draft_token_ids(self) -> DraftTokenIds | None:
output: list[DraftTokenIds] = self.collective_rpc("take_draft_token_ids")
return output[0]
@property
def max_concurrent_batches(self) -> int:
return 1
def profile(self, is_start: bool = True):
self.collective_rpc("profile", args=(is_start,))
def save_sharded_state(
self,
path: str,
pattern: str | None = None,
max_size: int | None = None,
) -> None:
self.collective_rpc(
"save_sharded_state",
kwargs=dict(path=path, pattern=pattern, max_size=max_size),
)
@abstractmethod
def check_health(self) -> None:
"""Checks if the executor is healthy. If not, it should raise an
exception."""
raise NotImplementedError
def shutdown(self) -> None:
"""Shutdown the executor."""
self.collective_rpc("shutdown")
def init_kv_output_aggregator(self, connector: "KVConnectorBase") -> None:
"""Init KVOutputAggregator"""
self.kv_output_aggregator = KVOutputAggregator.from_connector(
connector, self.parallel_config.world_size
)
@cached_property # Avoid unnecessary RPC calls
def supported_tasks(self) -> tuple[SupportedTask, ...]:
output: list[tuple[SupportedTask, ...]]
output = self.collective_rpc("get_supported_tasks")
return output[0]
def add_lora(self, lora_request: LoRARequest) -> bool:
assert lora_request.lora_int_id > 0, "lora_id must be greater than 0."
return all(self.collective_rpc("add_lora", args=(lora_request,)))
def remove_lora(self, lora_id: int) -> bool:
assert lora_id > 0, "lora_id must be greater than 0."
return all(self.collective_rpc("remove_lora", args=(lora_id,)))
def pin_lora(self, lora_id: int) -> bool:
assert lora_id > 0, "lora_id must be greater than 0."
return all(self.collective_rpc("pin_lora", args=(lora_id,)))
def list_loras(self) -> set[int]:
sets: list[set[int]] = self.collective_rpc("list_loras")
for s in sets:
assert s == sets[0], "All workers should have the same LORAs."
return sets[0]
def reset_mm_cache(self) -> None:
"""Reset the multi-modal cache in each worker."""
self.collective_rpc("reset_mm_cache")
def sleep(self, level: int = 1):
if self.is_sleeping:
logger.warning("Executor is already sleeping.")
return
time_before_sleep = time.perf_counter()
self.collective_rpc("sleep", kwargs=dict(level=level))
time_after_sleep = time.perf_counter()
self.sleeping_tags = {"weights", "kv_cache"}
self.is_sleeping = True
logger.info(
"It took %.6f seconds to fall asleep.", time_after_sleep - time_before_sleep
)
def wake_up(self, tags: list[str] | None = None):
if not self.is_sleeping:
logger.warning("Executor is not sleeping.")
return
if tags:
for tag in tags:
if tag not in self.sleeping_tags:
logger.warning(
"Tag %s is not in sleeping tags %s", tag, self.sleeping_tags
)
return
time_before_wakeup = time.perf_counter()
self.collective_rpc("wake_up", kwargs=dict(tags=tags))
time_after_wakeup = time.perf_counter()
logger.info(
"It took %.6f seconds to wake up tags %s.",
time_after_wakeup - time_before_wakeup,
tags if tags is not None else self.sleeping_tags,
)
if tags:
for tag in tags:
self.sleeping_tags.remove(tag)
else:
self.sleeping_tags.clear()
if not self.sleeping_tags:
self.is_sleeping = False
def reinitialize_distributed(
self, reconfig_request: ReconfigureDistributedRequest
) -> None:
raise NotImplementedError
from vllm.v1.executor.uniproc_executor import ( # noqa: E402
ExecutorWithExternalLauncher as _ExecutorWithExternalLauncher,
)
from vllm.v1.executor.uniproc_executor import ( # noqa: E402
UniProcExecutor as _UniProcExecutor,
)
# For backwards compatibility.
UniProcExecutor = _UniProcExecutor
ExecutorWithExternalLauncher = _ExecutorWithExternalLauncher

View File

@@ -0,0 +1,890 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import multiprocessing
import os
import pickle
import queue
import signal
import threading
import time
import traceback
import weakref
from collections import deque
from collections.abc import Callable, Sequence
from concurrent.futures import Future, InvalidStateError
from contextlib import suppress
from dataclasses import dataclass
from enum import Enum, auto
from functools import cached_property, partial
from multiprocessing.connection import Connection
from multiprocessing.process import BaseProcess
from multiprocessing.synchronize import Lock as LockType
from threading import Thread
from typing import Any, cast
import cloudpickle
import torch
import vllm.envs as envs
from vllm.config import VllmConfig
from vllm.distributed import destroy_distributed_environment, destroy_model_parallel
from vllm.distributed.device_communicators.shm_broadcast import Handle, MessageQueue
from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator
from vllm.distributed.parallel_state import (
get_dcp_group,
get_dp_group,
get_ep_group,
get_inner_dp_world_group,
get_pcp_group,
get_pp_group,
get_tp_group,
)
from vllm.envs import enable_envs_cache
from vllm.logger import init_logger
from vllm.utils.network_utils import (
get_distributed_init_method,
get_loopback_ip,
get_open_port,
)
from vllm.utils.system_utils import (
_maybe_force_spawn,
decorate_logs,
get_mp_context,
set_process_title,
)
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.executor.abstract import Executor, FailureCallback
from vllm.v1.outputs import AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput
from vllm.v1.worker.worker_base import WorkerWrapperBase
logger = init_logger(__name__)
class FutureWrapper(Future):
def __init__(
self,
futures_queue: deque[tuple["FutureWrapper", Callable]],
aggregate: Callable = lambda x: x,
):
self.futures_queue = futures_queue
self.aggregate = aggregate
super().__init__()
def result(self, timeout=None):
if timeout is not None:
raise RuntimeError("timeout not implemented")
# Drain any futures ahead of us in the queue.
while not self.done():
future, get_response = self.futures_queue.pop()
future.wait_for_response(get_response)
return super().result()
def wait_for_response(self, get_response: Callable):
try:
response = self.aggregate(get_response())
with suppress(InvalidStateError):
self.set_result(response)
except Exception as e:
with suppress(InvalidStateError):
self.set_exception(e)
class MultiprocExecutor(Executor):
supports_pp: bool = True
def __init__(self, vllm_config: VllmConfig, monitor_workers: bool = True):
self.monitor_workers = monitor_workers
super().__init__(vllm_config)
def _init_executor(self) -> None:
# Call self.shutdown at exit to clean up
# and ensure workers will be terminated.
self._finalizer = weakref.finalize(self, self.shutdown)
self.is_failed = False
self.shutdown_event = threading.Event()
self.failure_callback: FailureCallback | None = None
self.world_size = self.parallel_config.world_size
assert self.world_size % self.parallel_config.nnodes_within_dp == 0, (
f"global world_size ({self.parallel_config.world_size}) must be "
f"divisible by nnodes_within_dp "
f"({self.parallel_config.nnodes_within_dp}). "
)
self.local_world_size = self.parallel_config.local_world_size
tp_size = self.parallel_config.tensor_parallel_size
pp_size = self.parallel_config.pipeline_parallel_size
pcp_size = self.parallel_config.prefill_context_parallel_size
assert self.world_size == tp_size * pp_size * pcp_size, (
f"world_size ({self.world_size}) must be equal to the "
f"tensor_parallel_size ({tp_size}) x pipeline"
f"_parallel_size ({pp_size}) x prefill_context"
f"_parallel_size ({pcp_size}). "
)
# Set multiprocessing envs
set_multiprocessing_worker_envs()
# use the loopback address get_loopback_ip() for communication.
distributed_init_method = get_distributed_init_method(
get_loopback_ip(), get_open_port()
)
self.rpc_broadcast_mq: MessageQueue | None = None
scheduler_output_handle: Handle | None = None
# Initialize worker and set up message queues for SchedulerOutputs
# and ModelRunnerOutputs
if self.parallel_config.node_rank_within_dp == 0:
# For leader node within each dp rank,
# each dp will have its own leader multiproc executor.
max_chunk_bytes = envs.VLLM_MQ_MAX_CHUNK_BYTES_MB * 1024 * 1024
self.rpc_broadcast_mq = MessageQueue(
self.world_size,
self.local_world_size,
max_chunk_bytes=max_chunk_bytes,
connect_ip=self.parallel_config.master_addr,
)
scheduler_output_handle = self.rpc_broadcast_mq.export_handle()
# Create workers
context = get_mp_context()
shared_worker_lock = context.Lock()
unready_workers: list[UnreadyWorkerProcHandle] = []
success = False
try:
global_start_rank = (
self.local_world_size * self.parallel_config.node_rank_within_dp
)
for local_rank in range(self.local_world_size):
global_rank = global_start_rank + local_rank
unready_workers.append(
WorkerProc.make_worker_process(
vllm_config=self.vllm_config,
local_rank=local_rank,
rank=global_rank,
distributed_init_method=distributed_init_method,
input_shm_handle=scheduler_output_handle,
shared_worker_lock=shared_worker_lock,
)
)
# Workers must be created before wait_for_ready to avoid
# deadlock, since worker.init_device() does a device sync.
# Wait for all local workers to be ready.
self.workers = WorkerProc.wait_for_ready(unready_workers)
# Start background thread to monitor worker health if not in headless mode.
if self.monitor_workers:
self.start_worker_monitor()
self.response_mqs = []
# Only leader node have remote response mqs
if self.parallel_config.node_rank_within_dp == 0:
for rank in range(self.world_size):
if rank < self.local_world_size:
local_message_queue = self.workers[rank].worker_response_mq
assert local_message_queue is not None
self.response_mqs.append(local_message_queue)
else:
remote_message_queue = self.workers[0].peer_worker_response_mqs[
rank
]
assert remote_message_queue is not None
self.response_mqs.append(remote_message_queue)
# Ensure message queues are ready. Will deadlock if re-ordered
# Must be kept consistent with the WorkerProc.
# Wait for all input mqs to be ready.
if self.rpc_broadcast_mq is not None:
self.rpc_broadcast_mq.wait_until_ready()
# Wait for all remote response mqs to be ready.
for response_mq in self.response_mqs:
response_mq.wait_until_ready()
success = True
finally:
if not success:
# Clean up the worker procs if there was a failure.
# Close death_writers first to signal workers to exit
for uw in unready_workers:
if uw.death_writer is not None:
uw.death_writer.close()
self._ensure_worker_termination([uw.proc for uw in unready_workers])
self.futures_queue = deque[tuple[FutureWrapper, Callable]]()
self.output_rank = self._get_output_rank()
def start_worker_monitor(self, inline=False) -> None:
workers = self.workers
self_ref = weakref.ref(self)
# Monitors worker process liveness. If any die unexpectedly,
# logs an error, shuts down the executor and invokes the failure
# callback to inform the engine.
def monitor_workers():
sentinels = [h.proc.sentinel for h in workers]
died = multiprocessing.connection.wait(sentinels)
_self = self_ref()
if not _self or getattr(_self, "shutting_down", False):
return
_self.is_failed = True
proc_name = next(h.proc.name for h in workers if h.proc.sentinel == died[0])
logger.error(
"Worker proc %s died unexpectedly, shutting down executor.", proc_name
)
_self.shutdown()
callback = _self.failure_callback
if callback is not None:
_self.failure_callback = None
callback()
if not inline:
Thread(
target=monitor_workers, daemon=True, name="MultiprocWorkerMonitor"
).start()
return
monitor_workers()
def register_failure_callback(self, callback: FailureCallback):
if self.is_failed:
callback()
else:
self.failure_callback = callback
def execute_model( # type: ignore[override]
self, scheduler_output: SchedulerOutput, non_block: bool = False
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
return self.collective_rpc(
"execute_model",
args=(scheduler_output,),
unique_reply_rank=self.output_rank,
non_block=non_block,
timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS,
kv_output_aggregator=self.kv_output_aggregator,
)
def sample_tokens( # type: ignore[override]
self, grammar_output: GrammarOutput | None, non_block: bool = False
) -> ModelRunnerOutput | Future[ModelRunnerOutput]:
return self.collective_rpc(
"sample_tokens",
args=(grammar_output,),
unique_reply_rank=self.output_rank,
non_block=non_block,
timeout=envs.VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS,
kv_output_aggregator=self.kv_output_aggregator,
)
def execute_dummy_batch(self) -> None:
self.collective_rpc("execute_dummy_batch", unique_reply_rank=self.output_rank)
def take_draft_token_ids(self) -> DraftTokenIds | None:
# OPTIMIZATION: Get output only from a single worker (output_rank)
return self.collective_rpc(
"take_draft_token_ids", unique_reply_rank=self.output_rank
)
def collective_rpc( # type: ignore[override]
self,
method: str | Callable,
timeout: float | None = None,
args: tuple = (),
kwargs: dict | None = None,
non_block: bool = False,
unique_reply_rank: int | None = None,
kv_output_aggregator: KVOutputAggregator | None = None,
) -> Any:
"""Returns single result if unique_reply_rank and/or kv_output_aggregator
is provided, otherwise list."""
assert self.rpc_broadcast_mq is not None, (
"collective_rpc should not be called on follower node"
)
if self.is_failed:
raise RuntimeError("Executor failed.")
deadline = None if timeout is None else time.monotonic() + timeout
kwargs = kwargs or {}
if kv_output_aggregator is not None:
output_rank = None
aggregate: Callable[[Any], Any] = partial(
kv_output_aggregator.aggregate, output_rank=unique_reply_rank or 0
)
else:
output_rank = unique_reply_rank
aggregate = lambda x: x
if isinstance(method, str):
send_method = method
else:
send_method = cloudpickle.dumps(method, protocol=pickle.HIGHEST_PROTOCOL)
self.rpc_broadcast_mq.enqueue((send_method, args, kwargs, output_rank))
response_mqs: Sequence[MessageQueue] = self.response_mqs
if output_rank is not None:
response_mqs = (response_mqs[output_rank],)
shutdown_event = self.shutdown_event
def get_response():
responses = []
for mq in response_mqs:
dequeue_timeout = (
None if deadline is None else (deadline - time.monotonic())
)
try:
status, result = mq.dequeue(
timeout=dequeue_timeout, cancel=shutdown_event
)
except TimeoutError as e:
raise TimeoutError(f"RPC call to {method} timed out.") from e
if status != WorkerProc.ResponseStatus.SUCCESS:
raise RuntimeError(
f"Worker failed with error '{result}', please check the"
" stack trace above for the root cause"
)
responses.append(result)
return responses[0] if output_rank is not None else responses
if non_block:
future = FutureWrapper(self.futures_queue, aggregate=aggregate)
self.futures_queue.appendleft((future, get_response))
return future
# First drain any pending futures in the queue.
while self.futures_queue:
future, get_fut_response = self.futures_queue.pop()
future.wait_for_response(get_fut_response)
return aggregate(get_response())
@staticmethod
def _ensure_worker_termination(worker_procs: list[BaseProcess]):
"""Ensure that all worker processes are terminated. Assumes workers have
received termination requests. Waits for processing, then sends
termination and kill signals if needed."""
def wait_for_termination(procs, timeout):
if not time:
# If we are in late stage shutdown, the interpreter may replace
# `time` with `None`.
return all(not proc.is_alive() for proc in procs)
start_time = time.time()
while time.time() - start_time < timeout:
if all(not proc.is_alive() for proc in procs):
return True
time.sleep(0.1)
return False
# Send SIGTERM if still running
active_procs = [proc for proc in worker_procs if proc.is_alive()]
for p in active_procs:
p.terminate()
if not wait_for_termination(active_procs, 4):
# Send SIGKILL if still running
active_procs = [p for p in active_procs if p.is_alive()]
for p in active_procs:
p.kill()
def shutdown(self):
"""Properly shut down the executor and its workers"""
if not getattr(self, "shutting_down", False):
self.shutting_down = True
# Make sure all the worker processes are terminated first.
if workers := getattr(self, "workers", None):
for w in workers:
# Close death_writer to signal child processes to exit
if w.death_writer is not None:
w.death_writer.close()
w.death_writer = None
w.worker_response_mq = None
self._ensure_worker_termination([w.proc for w in workers])
self.shutdown_event.set()
self.rpc_broadcast_mq = None
def check_health(self) -> None:
self.collective_rpc("check_health", timeout=10)
return
@cached_property
def max_concurrent_batches(self) -> int:
if self.scheduler_config.async_scheduling:
return 2
return self.parallel_config.pipeline_parallel_size
def _get_output_rank(self) -> int:
# Only returns ModelRunnerOutput from TP rank=0 and PP rank=-1
# (the first TP worker of the last PP stage).
# Example:
# Assuming TP=8, PP=4, then the world_size=32
# 0-7, PP rank 0
# 8-15, PP rank 1
# 16-23, PP rank 2
# 24-31, PP rank 3
# so world_size - tp_size = 32 - 8 = 24 should be PP rank = -1 (i.e. 3)
return (
self.world_size
- self.parallel_config.tensor_parallel_size
* self.parallel_config.prefill_context_parallel_size
)
@dataclass
class UnreadyWorkerProcHandle:
"""WorkerProcess handle before READY."""
proc: BaseProcess
rank: int
ready_pipe: Connection
death_writer: Connection | None = None
@dataclass
class WorkerProcHandle:
proc: BaseProcess
rank: int
# The worker process writes to this MQ in single-node mode
worker_response_mq: MessageQueue | None
# This is only non empty on driver node,
# the peer worker process i writes to MQ
# `peer_worker_response_mqs[i]`
peer_worker_response_mqs: list[MessageQueue | None]
death_writer: Connection | None = None
@classmethod
def from_unready_handle(
cls,
unready_handle: UnreadyWorkerProcHandle,
worker_response_mq: MessageQueue | None,
peer_worker_response_mqs: list[MessageQueue | None],
) -> "WorkerProcHandle":
return cls(
proc=unready_handle.proc,
rank=unready_handle.rank,
worker_response_mq=worker_response_mq,
peer_worker_response_mqs=peer_worker_response_mqs,
death_writer=unready_handle.death_writer,
)
class WorkerProc:
"""Wrapper that runs one Worker in a separate process."""
READY_STR = "READY"
rpc_broadcast_mq: MessageQueue | None
worker_response_mq: MessageQueue | None
def _init_message_queues(
self, input_shm_handle: Handle, vllm_config: VllmConfig
) -> None:
if vllm_config.parallel_config.nnodes_within_dp == 1:
# Initialize MessageQueue for receiving SchedulerOutput
self.rpc_broadcast_mq = MessageQueue.create_from_handle(
input_shm_handle, self.worker.rank
)
# Initializes a message queue for sending the model output
self.worker_response_mq = MessageQueue(1, 1)
self.peer_response_handles = []
else:
# Initialize remote MessageQueue for receiving SchedulerOutput across nodes
self.rpc_broadcast_mq = get_inner_dp_world_group().create_mq_broadcaster(
external_writer_handle=input_shm_handle,
# Since there is external_writer_handle from executor proc,
# where the ready signal from actual writer is sent out of the
# create_mq_broadcaster method and after this setup, we make it
# non blocking. The handshake will be triggered when
# worker.rpc_broadcast_mq.wait_until_ready() is called
blocking=False,
)
# Initializes remote message queue for sending the model output to the
# driver worker, exposing peer_response_handles for driver worker
# that include handles for all ranks
self.worker_response_mq, self.peer_response_handles = (
get_inner_dp_world_group().create_single_reader_mq_broadcasters(
reader_rank_in_group=0
)
)
def __init__(
self,
vllm_config: VllmConfig,
local_rank: int,
rank: int,
distributed_init_method: str,
input_shm_handle: Handle,
shared_worker_lock: LockType,
):
self.rank = rank
wrapper = WorkerWrapperBase(
vllm_config=vllm_config, rpc_rank=local_rank, global_rank=rank
)
# TODO: move `init_worker` to executor level as a collective rpc call
all_kwargs: list[dict] = [
{} for _ in range(vllm_config.parallel_config.world_size)
]
is_driver_worker = rank % vllm_config.parallel_config.tensor_parallel_size == 0
all_kwargs[local_rank] = {
"vllm_config": vllm_config,
"local_rank": local_rank,
"rank": rank,
"distributed_init_method": distributed_init_method,
"is_driver_worker": is_driver_worker,
"shared_worker_lock": shared_worker_lock,
}
wrapper.init_worker(all_kwargs)
self.worker = wrapper
scheduler_config = vllm_config.scheduler_config
self.use_async_scheduling = scheduler_config.async_scheduling
if self.use_async_scheduling:
self.async_output_queue: queue.Queue = queue.Queue()
self.async_output_copy_thread = Thread(
target=self.async_output_busy_loop,
daemon=True,
name="WorkerAsyncOutputCopy",
)
self.async_output_copy_thread.start()
# Initialize device
self.worker.init_device()
# Set process title and log prefix
self.setup_proc_title_and_log_prefix(
enable_ep=vllm_config.parallel_config.enable_expert_parallel
)
# Load model
self._init_message_queues(input_shm_handle, vllm_config)
self.worker.load_model()
# Enable environment variable cache (e.g. assume no more
# environment variable overrides after this point)
enable_envs_cache()
@staticmethod
def make_worker_process(
vllm_config: VllmConfig,
local_rank: int,
rank: int,
distributed_init_method: str,
input_shm_handle, # Receive SchedulerOutput
shared_worker_lock: LockType,
) -> UnreadyWorkerProcHandle:
context = get_mp_context()
# (reader, writer)
reader, writer = context.Pipe(duplex=False)
# Create death pipe to detect parent process exit
death_reader, death_writer = context.Pipe(duplex=False)
process_kwargs = {
"vllm_config": vllm_config,
"local_rank": local_rank,
"rank": rank,
"distributed_init_method": distributed_init_method,
"input_shm_handle": input_shm_handle,
"ready_pipe": (reader, writer),
"death_pipe": death_reader,
"shared_worker_lock": shared_worker_lock,
}
# Run EngineCore busy loop in background process.
proc = context.Process(
target=WorkerProc.worker_main,
kwargs=process_kwargs,
name=f"VllmWorker-{rank}",
daemon=True,
)
proc.start()
writer.close()
# Keep death_writer open in parent - when parent exits,
# death_reader in child will get EOFError
return UnreadyWorkerProcHandle(proc, rank, reader, death_writer)
@staticmethod
def wait_for_response_handle_ready(
handles: dict[str, Any], proc_handle: UnreadyWorkerProcHandle
) -> WorkerProcHandle:
response_handle = handles["handle"]
worker_response_mq: MessageQueue | None = None
if len(response_handle.local_reader_ranks) > 0:
worker_response_mq = MessageQueue.create_from_handle(response_handle, 0)
peer_response_handles = handles["peer_response_handles"]
peer_worker_response_mqs = [
MessageQueue.create_from_handle(handle, -1)
if handle.remote_subscribe_addr is not None
else None
for handle in peer_response_handles
]
return WorkerProcHandle.from_unready_handle(
proc_handle,
worker_response_mq,
peer_worker_response_mqs=peer_worker_response_mqs,
)
@staticmethod
def wait_for_ready(
unready_proc_handles: list[UnreadyWorkerProcHandle],
) -> list[WorkerProcHandle]:
e = Exception(
"WorkerProc initialization failed due to "
"an exception in a background process. "
"See stack trace for root cause."
)
pipes = {handle.ready_pipe: handle for handle in unready_proc_handles}
ready_proc_handles: list[WorkerProcHandle | None] = [None] * len(
unready_proc_handles
)
while pipes:
ready = multiprocessing.connection.wait(pipes.keys())
for pipe in ready:
assert isinstance(pipe, Connection)
try:
# Wait until the WorkerProc is ready.
unready_proc_handle = pipes.pop(pipe)
response: dict[str, Any] = pipe.recv()
if response["status"] != "READY":
raise e
idx = unready_proc_handle.rank % len(ready_proc_handles)
ready_proc_handles[idx] = WorkerProc.wait_for_response_handle_ready(
response, unready_proc_handle
)
except EOFError:
e.__suppress_context__ = True
raise e from None
finally:
# Close connection.
pipe.close()
return cast(list[WorkerProcHandle], ready_proc_handles)
def shutdown(self):
self.worker.shutdown()
self.rpc_broadcast_mq = None
self.worker_response_mq = None
destroy_model_parallel()
destroy_distributed_environment()
@staticmethod
def worker_main(*args, **kwargs):
"""Worker initialization and execution loops.
This runs a background process"""
# Signal handler used for graceful termination.
# SystemExit exception is only raised once to allow this and worker
# processes to terminate without error
shutdown_requested = False
def signal_handler(signum, frame):
nonlocal shutdown_requested
if not shutdown_requested:
shutdown_requested = True
raise SystemExit()
# Either SIGTERM or SIGINT will terminate the worker
signal.signal(signal.SIGTERM, signal_handler)
signal.signal(signal.SIGINT, signal_handler)
worker = None
# tuple[Connection, Connection]
reader, ready_writer = kwargs.pop("ready_pipe")
death_pipe = kwargs.pop("death_pipe", None)
shutdown_event = threading.Event()
# Start death monitoring thread if death_pipe is provided
if death_pipe is not None:
def monitor_parent_death():
try:
# This will block until parent process exits (pipe closes)
death_pipe.recv()
except EOFError:
# Parent process has exited, terminate this worker
logger.info_once("Parent process exited, terminating worker")
# Send signal to self to trigger clean shutdown
shutdown_event.set()
except Exception as e:
logger.warning("Death monitoring error: %s", e)
death_monitor = Thread(
target=monitor_parent_death, daemon=True, name="WorkerDeathMonitor"
)
death_monitor.start()
try:
reader.close()
worker = WorkerProc(*args, **kwargs)
assert worker.worker_response_mq is not None
# Send READY once we know everything is loaded
ready_writer.send(
{
"status": WorkerProc.READY_STR,
"handle": worker.worker_response_mq.export_handle(),
"peer_response_handles": worker.peer_response_handles,
}
)
# Ensure message queues are ready. Will deadlock if re-ordered.
# Must be kept consistent with the Executor
if worker.rpc_broadcast_mq is not None:
worker.rpc_broadcast_mq.wait_until_ready()
worker.worker_response_mq.wait_until_ready()
ready_writer.close()
ready_writer = None
worker.worker_busy_loop(cancel=shutdown_event)
except Exception:
# NOTE: if an Exception arises in busy_loop, we send
# a FAILURE message over the MQ RPC to notify the Executor,
# which triggers system shutdown.
# TODO(rob): handle case where the MQ itself breaks.
if ready_writer is not None:
logger.exception("WorkerProc failed to start.")
elif shutdown_event.is_set():
logger.info("WorkerProc shutting down.")
else:
logger.exception("WorkerProc failed.")
# The parent sends a SIGTERM to all worker processes if
# any worker dies. Set this value so we don't re-throw
# SystemExit() to avoid zmq exceptions in __del__.
shutdown_requested = True
finally:
if ready_writer is not None:
ready_writer.close()
if death_pipe is not None:
death_pipe.close()
# Clean up once worker exits busy loop
if worker is not None:
worker.shutdown()
class ResponseStatus(Enum):
SUCCESS = auto()
FAILURE = auto()
def enqueue_output(self, output: Any):
"""Prepares output from the worker and enqueues it to the
worker_response_mq. If the output is an Exception, it is
converted to a FAILURE response.
"""
if isinstance(output, AsyncModelRunnerOutput):
output = output.get_output()
if isinstance(output, Exception):
result = (WorkerProc.ResponseStatus.FAILURE, str(output))
else:
result = (WorkerProc.ResponseStatus.SUCCESS, output)
if (response_mq := self.worker_response_mq) is not None:
response_mq.enqueue(result)
def handle_output(self, output: Any):
"""Handles output from the worker. If async scheduling is enabled,
it is passed to the async_output_busy_loop thread. Otherwise, it is
enqueued directly to the worker_response_mq.
"""
if self.use_async_scheduling:
self.async_output_queue.put(output)
else:
self.enqueue_output(output)
def async_output_busy_loop(self):
"""Entrypoint for the thread which handles outputs asynchronously."""
while True:
output = self.async_output_queue.get()
self.enqueue_output(output)
def worker_busy_loop(self, cancel: threading.Event | None = None):
"""Main busy loop for Multiprocessing Workers"""
assert self.rpc_broadcast_mq is not None
while True:
method, args, kwargs, output_rank = self.rpc_broadcast_mq.dequeue(
cancel=cancel, indefinite=True
)
try:
if isinstance(method, str):
func = getattr(self.worker, method)
elif isinstance(method, bytes):
func = partial(cloudpickle.loads(method), self.worker)
output = func(*args, **kwargs)
except Exception as e:
# Notes have been introduced in python 3.11
if hasattr(e, "add_note"):
e.add_note(traceback.format_exc())
logger.exception("WorkerProc hit an exception.")
# exception might not be serializable, so we convert it to
# string, only for logging purpose.
if output_rank is None or self.rank == output_rank:
self.handle_output(e)
continue
if output_rank is None or self.rank == output_rank:
self.handle_output(output)
@staticmethod
def setup_proc_title_and_log_prefix(enable_ep: bool) -> None:
dp_size = get_dp_group().world_size
dp_rank = get_dp_group().rank_in_group
pp_size = get_pp_group().world_size
pp_rank = get_pp_group().rank_in_group
pcp_size = get_pcp_group().world_size
pcp_rank = get_pcp_group().rank_in_group
tp_size = get_tp_group().world_size
tp_rank = get_tp_group().rank_in_group
dcp_size = get_dcp_group().world_size
dcp_rank = get_dcp_group().rank_in_group
process_name = "Worker"
if dp_size > 1:
process_name += f"_DP{dp_rank}"
if pp_size > 1:
process_name += f"_PP{pp_rank}"
if pcp_size > 1:
process_name += f"_PCP{pcp_rank}"
if tp_size > 1:
process_name += f"_TP{tp_rank}"
if dcp_size > 1:
process_name += f"_DCP{dcp_rank}"
if enable_ep:
ep_rank = get_ep_group().rank_in_group
process_name += f"_EP{ep_rank}"
set_process_title(name=process_name)
decorate_logs(process_name)
def set_multiprocessing_worker_envs():
"""Set up environment variables that should be used when there are workers
in a multiprocessing environment. This should be called by the parent
process before worker processes are created"""
_maybe_force_spawn()
# Configure thread parallelism if OMP_NUM_THREADS isn't set
#
# Helps to avoid CPU contention. The default of spawning a thread per
# core combined with multiprocessing for each GPU can have a negative
# impact on performance. The contention is amplified when running in a
# container where CPU limits can cause throttling.
default_omp_num_threads = 1
if (
"OMP_NUM_THREADS" not in os.environ
and (current_parallelism := torch.get_num_threads()) > default_omp_num_threads
):
logger.warning(
"Reducing Torch parallelism from %d threads to %d to avoid "
"unnecessary CPU contention. Set OMP_NUM_THREADS in the "
"external environment to tune this value as needed.",
current_parallelism,
default_omp_num_threads,
)
os.environ["OMP_NUM_THREADS"] = str(default_omp_num_threads)
torch.set_num_threads(default_omp_num_threads)

View File

@@ -0,0 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.v1.executor.ray_executor import (
RayDistributedExecutor as _RayDistributedExecutor,
)
# For backwards compatibility.
RayDistributedExecutor = _RayDistributedExecutor

View File

@@ -0,0 +1,626 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
from collections import defaultdict
from collections.abc import Callable
from concurrent.futures import Future
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any
import cloudpickle
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.ray.ray_env import get_env_vars_to_copy
from vllm.utils.network_utils import (
get_distributed_init_method,
get_ip,
get_open_port,
)
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
from vllm.v1.executor.abstract import Executor
from vllm.v1.executor.ray_utils import (
FutureWrapper,
RayWorkerWrapper,
initialize_ray_cluster,
ray,
)
from vllm.v1.outputs import ModelRunnerOutput
if ray is not None:
from ray.actor import ActorHandle
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
else:
ActorHandle = None
if TYPE_CHECKING:
from ray.util.placement_group import PlacementGroup
logger = init_logger(__name__)
COMPLETED_NONE_FUTURE: Future[ModelRunnerOutput | None] = Future()
COMPLETED_NONE_FUTURE.set_result(None)
@dataclass
class RayWorkerMetaData:
"""
Metadata for a Ray worker.
The order of ray worker creation can be random,
and we need to reset the rank after creating all workers.
"""
worker: ActorHandle
created_rank: int
adjusted_rank: int = -1
ip: str = ""
class RayDistributedExecutor(Executor):
"""Ray-based distributed executor"""
# These env vars are worker-specific, therefore are NOT copied
# from the driver to the workers
WORKER_SPECIFIC_ENV_VARS = {
"VLLM_HOST_IP",
"VLLM_HOST_PORT",
"LOCAL_RANK",
"CUDA_VISIBLE_DEVICES",
}
# These non-vLLM env vars are copied from the driver to workers
ADDITIONAL_ENV_VARS = {"HF_TOKEN", "HUGGING_FACE_HUB_TOKEN"}
uses_ray: bool = True
supports_pp: bool = True
def _init_executor(self) -> None:
self.forward_dag: ray.dag.CompiledDAG | None = None
# For TPU or XPU, avoid compiling NVIDIA's NCCL
if current_platform.is_tpu() or current_platform.is_xpu():
os.environ["VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE"] = "shm"
assert self.uses_ray
initialize_ray_cluster(self.parallel_config)
placement_group = self.parallel_config.placement_group
# Disable Ray usage stats collection.
ray_usage = os.environ.get("RAY_USAGE_STATS_ENABLED", "0")
if ray_usage != "1":
os.environ["RAY_USAGE_STATS_ENABLED"] = "0"
# Create the parallel GPU workers.
self._init_workers_ray(placement_group)
# KV connector setup
self.has_connector = self.vllm_config.kv_transfer_config is not None
self.uses_sampler = self.vllm_config.model_config.runner_type != "pooling" and (
self.vllm_config.ec_transfer_config is None
or not self.vllm_config.ec_transfer_config.is_ec_producer
)
self.scheduler_output: SchedulerOutput | None = None
@property
def max_concurrent_batches(self) -> int:
"""Ray distributed executor supports pipeline parallelism,
meaning that it allows PP size batches to be executed concurrently.
"""
if self.scheduler_config.async_scheduling:
return 2
return self.parallel_config.pipeline_parallel_size
def shutdown(self) -> None:
if logger:
# Somehow logger can be None here.
logger.info(
"Shutting down Ray distributed executor. If you see error log "
"from logging.cc regarding SIGTERM received, please ignore "
"because this is the expected termination process in Ray."
)
if hasattr(self, "forward_dag") and self.forward_dag is not None:
self.forward_dag.teardown()
import ray
for worker in self.workers:
ray.kill(worker)
self.forward_dag = None
def _configure_ray_workers_use_nsight(self, ray_remote_kwargs) -> dict[str, Any]:
# If nsight profiling is enabled, we need to set the profiling
# configuration for the ray workers as runtime env.
runtime_env = ray_remote_kwargs.setdefault("runtime_env", {})
runtime_env.update(
{
"nsight": {
"t": "cuda,cudnn,cublas",
"o": "'worker_process_%p'",
"cuda-graph-trace": "node",
}
}
)
return ray_remote_kwargs
# child class could overwrite this to return actual env vars.
def _get_env_vars_to_be_updated(self):
return self._env_vars_for_all_workers
def _init_workers_ray(self, placement_group: "PlacementGroup", **ray_remote_kwargs):
num_gpus = envs.VLLM_RAY_PER_WORKER_GPUS
# The driver dummy worker does not actually use any resources.
# It holds the resource for the driver worker.
self.driver_dummy_worker: RayWorkerWrapper | None = None
# The remaining workers are the actual ray actors.
self.workers: list[RayWorkerWrapper] = []
# Used in ray compiled DAG: indexed first by PP rank,
# and then TP rank. In other words, the inner list is
# the TP group of workers for a PP rank.
self.pp_tp_workers: list[list[RayWorkerWrapper]] = []
if self.parallel_config.ray_workers_use_nsight:
ray_remote_kwargs = self._configure_ray_workers_use_nsight(
ray_remote_kwargs
)
# Create the workers.
bundle_indices: list[int]
if envs.VLLM_RAY_BUNDLE_INDICES:
# Use the bundle indices specified by the user.
bundle_indices = list(map(int, envs.VLLM_RAY_BUNDLE_INDICES.split(",")))
assert len(bundle_indices) == self.parallel_config.world_size, (
"VLLM_RAY_BUNDLE_INDICES must have the same size"
f" as the world size, but got {bundle_indices=} "
f"and {self.parallel_config.world_size=}"
)
assert len(set(bundle_indices)) == len(bundle_indices), (
"VLLM_RAY_BUNDLE_INDICES cannot have duplicate values,"
f" but got {bundle_indices=}"
)
else:
# use the first N bundles that have GPU resources.
bundle_indices = []
for bundle_id, bundle in enumerate(placement_group.bundle_specs):
if bundle.get(current_platform.ray_device_key, 0):
bundle_indices.append(bundle_id)
bundle_indices = bundle_indices[: self.parallel_config.world_size]
worker_metadata: list[RayWorkerMetaData] = []
driver_ip = get_ip()
for rank, bundle_id in enumerate(bundle_indices):
scheduling_strategy = PlacementGroupSchedulingStrategy(
placement_group=placement_group,
placement_group_capture_child_tasks=True,
placement_group_bundle_index=bundle_id,
)
if current_platform.ray_device_key == "GPU":
# NV+AMD GPUs, and Intel XPUs
worker = ray.remote(
num_cpus=0,
num_gpus=num_gpus,
scheduling_strategy=scheduling_strategy,
**ray_remote_kwargs,
)(RayWorkerWrapper).remote( # type: ignore[attr-defined]
vllm_config=self.vllm_config, rpc_rank=rank
)
else:
worker = ray.remote(
num_cpus=0,
num_gpus=0,
resources={current_platform.ray_device_key: num_gpus},
scheduling_strategy=scheduling_strategy,
**ray_remote_kwargs,
)(RayWorkerWrapper).remote( # type: ignore[attr-defined]
vllm_config=self.vllm_config, rpc_rank=rank
)
worker_metadata.append(RayWorkerMetaData(worker=worker, created_rank=rank))
worker_ips = ray.get(
[
each.worker.get_node_ip.remote() # type: ignore[attr-defined]
for each in worker_metadata
]
)
for each, ip in zip(worker_metadata, worker_ips):
each.ip = ip
logger.debug("workers: %s", worker_metadata)
logger.debug("driver_dummy_worker: %s", self.driver_dummy_worker)
ip_counts: dict[str, int] = {}
for ip in worker_ips:
ip_counts[ip] = ip_counts.get(ip, 0) + 1
def sort_by_driver_then_worker_ip(item: RayWorkerMetaData):
"""
Sort the workers based on 3 properties:
1. If the worker is on the same node as the driver (vllm engine),
it should be placed first.
2. Then, if the worker is on a node with fewer workers, it should
be placed first.
3. Finally, if the work is on a node with smaller IP address, it
should be placed first.
"""
ip = item.ip
return 0 if ip == driver_ip else 1, ip_counts[ip], ip
# After sorting, the workers on the same node will be
# close to each other, and the workers on the driver
# node will be placed first.
sorted_worker_metadata = sorted(
worker_metadata, key=sort_by_driver_then_worker_ip
)
for i, item in enumerate(sorted_worker_metadata):
item.adjusted_rank = i
self.workers = [item.worker for item in sorted_worker_metadata]
rerank_mapping = {
item.created_rank: item.adjusted_rank for item in sorted_worker_metadata
}
self.collective_rpc("adjust_rank", args=(rerank_mapping,))
# Get the set of GPU IDs used on each node.
worker_node_and_gpu_ids = []
for worker in [self.driver_dummy_worker] + self.workers:
if worker is None:
# driver_dummy_worker can be None when using ray spmd worker.
continue
worker_node_and_gpu_ids.append(
ray.get(worker.get_node_and_gpu_ids.remote())
) # type: ignore[attr-defined]
node_workers = defaultdict(list) # node id -> list of worker ranks
node_gpus = defaultdict(list) # node id -> list of gpu ids
for i, (node_id, gpu_ids) in enumerate(worker_node_and_gpu_ids):
node_workers[node_id].append(i)
# `gpu_ids` can be a list of strings or integers.
# convert them to integers for consistency.
# NOTE: gpu_ids can be larger than 9 (e.g. 16 GPUs),
# string sorting is not sufficient.
# see https://github.com/vllm-project/vllm/issues/5590
gpu_ids = [int(x) for x in gpu_ids]
node_gpus[node_id].extend(gpu_ids)
for node_id, gpu_ids in node_gpus.items():
node_gpus[node_id] = sorted(gpu_ids)
all_ips = set(worker_ips + [driver_ip])
n_ips = len(all_ips)
n_nodes = len(node_workers)
if n_nodes != n_ips:
raise RuntimeError(
f"Every node should have a unique IP address. Got {n_nodes}"
f" nodes with node ids {list(node_workers.keys())} and "
f"{n_ips} unique IP addresses {all_ips}. Please check your"
" network configuration. If you set `VLLM_HOST_IP`"
" environment variable, make sure it is unique for"
" each node."
)
# Set environment variables for the driver and workers.
all_args_to_update_environment_variables = [
{
current_platform.device_control_env_var: ",".join(
map(str, node_gpus[node_id])
),
}
for (node_id, _) in worker_node_and_gpu_ids
]
# Environment variables to copy from driver to workers
env_vars_to_copy = get_env_vars_to_copy(
exclude_vars=self.WORKER_SPECIFIC_ENV_VARS,
additional_vars=set(current_platform.additional_env_vars).union(
self.ADDITIONAL_ENV_VARS
),
destination="workers",
)
# Copy existing env vars to each worker's args
for args in all_args_to_update_environment_variables:
# TODO: refactor platform-specific env vars
for name in env_vars_to_copy:
if name in os.environ:
args[name] = os.environ[name]
self._env_vars_for_all_workers = all_args_to_update_environment_variables
self.collective_rpc(
"update_environment_variables", args=(self._get_env_vars_to_be_updated(),)
)
if len(node_gpus) == 1:
# in single node case, we don't need to get the IP address.
# the loopback address is sufficient
# NOTE: a node may have several IP addresses, one for each
# network interface. `get_ip()` might return any of them,
# while they might not work for communication inside the node
# if the network setup is complicated. Using the loopback address
# solves this issue, as it always works for communication inside
# the node.
driver_ip = "127.0.0.1"
distributed_init_method = get_distributed_init_method(
driver_ip, get_open_port()
)
# Initialize the actual workers inside worker wrapper.
all_kwargs = []
for rank, (node_id, _) in enumerate(worker_node_and_gpu_ids):
local_rank = node_workers[node_id].index(rank)
kwargs = dict(
vllm_config=self.vllm_config,
local_rank=local_rank,
rank=rank,
distributed_init_method=distributed_init_method,
is_driver_worker=(not self.parallel_config)
or (rank % self.parallel_config.tensor_parallel_size == 0),
)
all_kwargs.append(kwargs)
self.collective_rpc("init_worker", args=(all_kwargs,))
self.collective_rpc("init_device")
self.collective_rpc("load_model")
for pp_rank in range(self.parallel_config.pipeline_parallel_size):
self.pp_tp_workers.append([])
for tp_rank in range(self.parallel_config.tensor_parallel_size):
# PP=2, TP=4
# pp_tp_workers = [[0, 1, 2, 3], [4, 5, 6, 7]]
rank = (pp_rank * self.parallel_config.tensor_parallel_size) + tp_rank
assert len(self.pp_tp_workers[pp_rank]) == tp_rank
assert pp_rank < len(self.pp_tp_workers)
self.pp_tp_workers[pp_rank].append(self.workers[rank])
def reinitialize_distributed(
self, reconfig_request: ReconfigureDistributedRequest
) -> None:
self.collective_rpc("reinitialize_distributed", args=(reconfig_request,))
if (
reconfig_request.new_data_parallel_rank
== ReconfigureRankType.SHUTDOWN_CURRENT_RANK
):
self.shutdown()
def execute_model( # type: ignore[override]
self,
scheduler_output: SchedulerOutput,
non_block: bool = False,
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
if self.scheduler_output is not None:
raise RuntimeError(
"State error: sample_tokens() must be called "
"after execute_model() returns None."
)
if not self.uses_sampler or not scheduler_output.total_num_scheduled_tokens:
# Model will not execute, call model runner immediately.
return self._execute_dag(scheduler_output, None, non_block)
# Model will execute, defer to sample_tokens() call.
self.scheduler_output = scheduler_output
return COMPLETED_NONE_FUTURE if non_block else None
def sample_tokens( # type: ignore[override]
self,
grammar_output: "GrammarOutput | None",
non_block: bool = False,
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
"""Execute the model on the Ray workers.
The scheduler output to use should have been provided in
a prior call to execute_model().
Args:
grammar_output: The structured outputs grammar bitmask, if applicable.
non_block: If True, the method will return a Future.
Returns:
The model runner output.
"""
scheduler_output = self.scheduler_output
if scheduler_output is None:
return COMPLETED_NONE_FUTURE if non_block else None
self.scheduler_output = None
return self._execute_dag(scheduler_output, grammar_output, non_block)
def _execute_dag(
self,
scheduler_output: SchedulerOutput,
grammar_output: "GrammarOutput | None",
non_block: bool = False,
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
# Build the compiled DAG for the first time.
if self.forward_dag is None: # type: ignore
self.forward_dag = self._compiled_ray_dag(enable_asyncio=False)
refs = self.forward_dag.execute((scheduler_output, grammar_output)) # type: ignore
if not self.has_connector:
# Get output only from a single worker (output_rank)
# When PP is not used, we block here until the result is available.
if not non_block:
return refs[0].get()
# When PP is used, we return a FutureWrapper immediately so that
# the scheduler can yield to the next batch.
return FutureWrapper(refs[0])
# Get output from all workers when connector is present
assert self.kv_output_aggregator is not None
if not non_block:
# Block and get results from all workers
return self.kv_output_aggregator.aggregate(ray.get(refs))
# Return a future that will aggregate outputs from all workers
return FutureWrapper(refs, self.kv_output_aggregator)
def collective_rpc( # type: ignore[override]
self,
method: str | Callable,
timeout: float | None = None,
args: tuple = (),
kwargs: dict[str, Any] | None = None,
non_block: bool = False,
) -> list[Any] | Future[list[Any]]:
"""Runs the given method on all workers."""
sent_method = method if isinstance(method, str) else cloudpickle.dumps(method)
del method
if kwargs is None:
kwargs = {}
ray_worker_outputs = [
worker.execute_method.remote( # type: ignore[attr-defined]
sent_method, *args, **kwargs
)
for worker in self.workers
]
# Get the results of the ray workers.
if non_block:
return FutureWrapper(ray_worker_outputs)
return ray.get(ray_worker_outputs, timeout=timeout)
def _check_ray_cgraph_installation(self):
import importlib.metadata
from packaging import version
required_version = version.parse("2.43.0")
current_version = version.parse(importlib.metadata.version("ray"))
if current_version < required_version:
raise ValueError(
f"Ray version {required_version} is "
f"required, but found {current_version}"
)
import importlib.util
cgraph_spec = importlib.util.find_spec("ray.experimental.compiled_dag_ref")
if cgraph_spec is None:
raise ValueError(
"Ray Compiled Graph is not installed. "
"Run `pip install ray[cgraph]` to install it."
)
cupy_spec = importlib.util.find_spec("cupy")
if cupy_spec is None and envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE == "nccl":
raise ValueError(
"cupy is not installed but required since "
"VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE is set to 'nccl'. "
"Run `pip install ray[cgraph]` and check cupy installation."
)
def _compiled_ray_dag(self, enable_asyncio: bool):
assert self.parallel_config.use_ray
self._check_ray_cgraph_installation()
# Enlarge the default value of "RAY_CGRAPH_get_timeout" to 300 seconds
# (it is 10 seconds by default). This is a Ray environment variable to
# control the timeout of getting result from a compiled graph execution,
# i.e., the distributed execution that includes model forward runs and
# intermediate tensor communications, in the case of vllm.
# Note: we should set this env var before importing
# ray.dag, otherwise it will not take effect.
os.environ.setdefault("RAY_CGRAPH_get_timeout", "300") # noqa: SIM112
from ray.dag import InputNode, MultiOutputNode
logger.info(
"RAY_CGRAPH_get_timeout is set to %s",
os.environ["RAY_CGRAPH_get_timeout"], # noqa: SIM112
)
logger.info(
"VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE = %s",
envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE,
)
logger.info(
"VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM = %s",
envs.VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM,
)
channel_type = envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE
if channel_type not in ("auto", "nccl", "shm"):
raise ValueError(
"Invalid value for VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: "
f"{channel_type}. Valid values are: 'auto', 'nccl', or 'shm'."
)
with InputNode() as input_data:
# Example DAG: PP=2, TP=4
#
# SchedulerOutput -> 0 -> (SchedulerOutput, IntermediateTensors) -> 4 -> ModelRunnerOutput # noqa: E501
# SchedulerOutput -> 1 -> (SchedulerOutput, IntermediateTensors) -> 5 -> ModelRunnerOutput # noqa: E501
# SchedulerOutput -> 2 -> (SchedulerOutput, IntermediateTensors) -> 6 -> ModelRunnerOutput # noqa: E501
# SchedulerOutput -> 3 -> (SchedulerOutput, IntermediateTensors) -> 7 -> ModelRunnerOutput # noqa: E501
# All workers in the first TP group will take in the
# ExecuteModelRequest as input.
outputs = [input_data for _ in self.pp_tp_workers[0]]
for pp_rank, tp_group in enumerate(self.pp_tp_workers):
# Each PP worker takes in the output of the previous PP worker,
# and the TP group executes in SPMD fashion.
outputs = [
worker.execute_model_ray.bind(outputs[i]) # type: ignore[attr-defined]
for i, worker in enumerate(tp_group)
]
last_pp_rank = len(self.pp_tp_workers) - 1
if (
pp_rank < last_pp_rank
and envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE != "shm"
):
# Specify how intermediate tensors should be passed
# between pp stages, no need to specify for the last
# pp stage or when using shared memory (the default).
transport = envs.VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE
outputs = [
output.with_tensor_transport(transport=transport)
for output in outputs
]
forward_dag = MultiOutputNode(outputs)
if envs.VLLM_USE_RAY_WRAPPED_PP_COMM:
from ray.experimental.channel.accelerator_context import (
register_accelerator_context,
)
from vllm.distributed.device_communicators.ray_communicator import (
RayPPCommunicator,
)
register_accelerator_context(
torch_module_name="cuda", communicator_cls=RayPPCommunicator
)
logger.info(
"Using RayPPCommunicator "
"(which wraps vLLM _PP GroupCoordinator) "
"for Ray Compiled Graph communication."
)
else:
logger.info(
"Using Ray's NCCL communicator for Ray Compiled Graph communication."
)
return forward_dag.experimental_compile(
enable_asyncio=enable_asyncio,
_overlap_gpu_communication=envs.VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM,
)
def __del__(self):
self.shutdown()
def check_health(self) -> None:
# Assume that the Ray workers are healthy.
# TODO: check the health of the Ray workers
return

View File

@@ -0,0 +1,465 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
import time
from collections import defaultdict
from concurrent.futures import Future
from typing import TYPE_CHECKING, Union
import vllm.platforms
from vllm.config import ParallelConfig
from vllm.distributed import get_pp_group
from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.utils.network_utils import get_ip
from vllm.v1.outputs import AsyncModelRunnerOutput
from vllm.v1.worker.worker_base import WorkerWrapperBase
if TYPE_CHECKING:
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.outputs import ModelRunnerOutput
logger = init_logger(__name__)
PG_WAIT_TIMEOUT = 1800
try:
import ray
from ray.util import placement_group_table
from ray.util.placement_group import PlacementGroup
try:
from ray._private.state import available_resources_per_node
except ImportError:
# Ray 2.9.x doesn't expose `available_resources_per_node`
from ray._private.state import state as _state
available_resources_per_node = _state._available_resources_per_node
class RayWorkerWrapper(WorkerWrapperBase):
"""Ray wrapper for vllm.worker.Worker, allowing Worker to be
lazily initialized after Ray sets CUDA_VISIBLE_DEVICES."""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
# Since the compiled DAG runs a main execution
# in a different thread that calls cuda.set_device.
# The flag indicates is set_device is called on
# that thread.
self.compiled_dag_cuda_device_set = False
def get_node_ip(self) -> str:
return get_ip()
def get_node_and_gpu_ids(self) -> tuple[str, list[int]]:
node_id = ray.get_runtime_context().get_node_id()
device_key = vllm.platforms.current_platform.ray_device_key
if not device_key:
raise RuntimeError(
"current platform %s does not support ray.",
vllm.platforms.current_platform.device_name,
)
gpu_ids = ray.get_runtime_context().get_accelerator_ids()[device_key]
return node_id, gpu_ids
def setup_device_if_necessary(self):
# TODO(swang): This is needed right now because Ray CG executes
# on a background thread, so we need to reset torch's current
# device.
# We can remove this API after it is fixed in compiled graph.
assert self.worker is not None, "Worker is not initialized"
if not self.compiled_dag_cuda_device_set:
if current_platform.is_tpu():
# Not needed
pass
else:
assert self.worker.device is not None
current_platform.set_device(self.worker.device)
self.compiled_dag_cuda_device_set = True
def execute_model_ray(
self,
execute_model_input: tuple["SchedulerOutput", "GrammarOutput"]
| tuple["SchedulerOutput", "GrammarOutput", "IntermediateTensors"],
) -> Union[
"ModelRunnerOutput",
tuple["SchedulerOutput", "GrammarOutput", "IntermediateTensors"],
]:
# This method is used by Ray Compiled Graph to execute the model,
# and it needs a special logic of self.setup_device_if_necessary()
self.setup_device_if_necessary()
assert self.worker is not None, "Worker is not initialized"
if len(execute_model_input) == 3:
scheduler_output, grammar_output, intermediate_tensors = (
execute_model_input
)
else:
scheduler_output, grammar_output = execute_model_input
intermediate_tensors = None
assert self.worker.model_runner is not None
output = self.worker.model_runner.execute_model(
scheduler_output, intermediate_tensors
)
if isinstance(output, IntermediateTensors):
output = scheduler_output, grammar_output, output
elif not get_pp_group().is_last_rank:
# Case where there are no scheduled requests
# but may still be finished requests.
assert not output or not output.req_ids
output = scheduler_output, grammar_output, None
elif output is None:
output = self.worker.model_runner.sample_tokens(grammar_output)
# Ensure outputs crossing Ray compiled DAG are serializable.
# AsyncModelRunnerOutput holds CUDA events and cannot be
# pickled.
if isinstance(output, AsyncModelRunnerOutput):
output = output.get_output()
return output
def override_env_vars(self, vars: dict[str, str]):
os.environ.update(vars)
ray_import_err = None
except ImportError as e:
ray = None # type: ignore
# only capture string to avoid variable references in the traceback that can
# prevent garbage collection in some cases
ray_import_err = str(e)
RayWorkerWrapper = None # type: ignore
class FutureWrapper(Future):
"""A wrapper around Ray output reference to meet the interface
of .execute_model(): The top level (core busy loop) expects .result() api
to block and return a single output.
If aggregator is provided, the outputs from all workers are aggregated upon
the result() call. If not only the first worker's output is returned.
"""
def __init__(self, ref_or_refs, aggregator: KVOutputAggregator | None = None):
super().__init__()
self.ref_or_refs = ref_or_refs
self.aggregator = aggregator
def result(self, timeout=None):
outputs = ray.get(self.ref_or_refs, timeout=timeout)
if self.aggregator is None:
return outputs
return self.aggregator.aggregate(outputs, output_rank=0)
def ray_is_available() -> bool:
"""Returns True if Ray is available."""
return ray is not None
def assert_ray_available():
"""Raise an exception if Ray is not available."""
if ray is None:
raise ValueError(
f"Failed to import Ray: {ray_import_err}."
"Please install Ray with `pip install ray`."
)
def _verify_bundles(
placement_group: "PlacementGroup", parallel_config: ParallelConfig, device_str: str
):
"""Verify a given placement group has bundles located in the right place.
There are 2 rules.
- Warn if all tensor parallel workers cannot fit in a single node.
- Fail if driver node is not included in a placement group.
"""
assert ray.is_initialized(), (
"Ray is not initialized although distributed-executor-backend is ray."
)
pg_data = placement_group_table(placement_group)
# bundle_idx -> node_id
bundle_to_node_ids = pg_data["bundles_to_node_id"]
# bundle_idx -> bundle (e.g., {"GPU": 1})
bundles = pg_data["bundles"]
# node_id -> List of bundle (e.g., {"GPU": 1})
node_id_to_bundle: dict[str, list[dict[str, float]]] = defaultdict(list)
for bundle_idx, node_id in bundle_to_node_ids.items():
node_id_to_bundle[node_id].append(bundles[bundle_idx])
driver_node_id = ray.get_runtime_context().get_node_id()
if driver_node_id not in node_id_to_bundle:
raise RuntimeError(
f"driver node id {driver_node_id} is not included in a placement "
f"group {placement_group.id}. Node id -> bundles "
f"{node_id_to_bundle}. "
"You don't have enough GPUs available in a current node. Check "
"`ray status` and `ray list nodes` to see if you have available "
"GPUs in a node `{driver_node_id}` before starting an vLLM engine."
)
for node_id, bundles in node_id_to_bundle.items():
if len(bundles) < parallel_config.tensor_parallel_size:
logger.warning(
"tensor_parallel_size=%d "
"is bigger than a reserved number of %ss (%d "
"%ss) in a node %s. Tensor parallel workers can be "
"spread out to 2+ nodes which can degrade the performance "
"unless you have fast interconnect across nodes, like "
"Infiniband. To resolve this issue, make sure you have more "
"than %d GPUs available at each node.",
parallel_config.tensor_parallel_size,
device_str,
len(bundles),
device_str,
node_id,
parallel_config.tensor_parallel_size,
)
def _wait_until_pg_ready(current_placement_group: "PlacementGroup"):
"""Wait until a placement group is ready.
It prints the informative log messages if the placement group is
not created within time.
"""
# Wait until PG is ready - this will block until all
# requested resources are available, and will time out
# if they cannot be provisioned.
placement_group_specs = current_placement_group.bundle_specs
s = time.time()
pg_ready_ref = current_placement_group.ready()
wait_interval = 10
while time.time() - s < PG_WAIT_TIMEOUT:
ready, _ = ray.wait([pg_ready_ref], timeout=wait_interval)
if len(ready) > 0:
break
# Exponential backoff for warning print.
wait_interval *= 2
logger.info(
"Waiting for creating a placement group of specs for "
"%d seconds. specs=%s. Check `ray status` and "
"`ray list nodes` to see if you have enough resources,"
" and make sure the IP addresses used by ray cluster"
" are the same as VLLM_HOST_IP environment variable"
" specified in each node if you are running on a multi-node.",
int(time.time() - s),
placement_group_specs,
)
try:
ray.get(pg_ready_ref, timeout=0)
except ray.exceptions.GetTimeoutError:
# Provide more helpful error message when GPU count is exceeded
total_gpu_required = sum(spec.get("GPU", 0) for spec in placement_group_specs)
# If more than one GPU is required for the placement group, provide a
# more specific error message.
# We use >1 here because multi-GPU (tensor parallel) jobs are more
# likely to fail due to insufficient cluster resources, and users may
# need to adjust tensor_parallel_size to fit available GPUs.
if total_gpu_required > 1:
raise ValueError(
f"Cannot provide a placement group requiring "
f"{total_gpu_required} GPUs "
f"(placement_group_specs={placement_group_specs}) within "
f"{PG_WAIT_TIMEOUT} seconds.\n"
f"Tensor parallel size may exceed available GPUs in your "
f"cluster. Check resources with `ray status` and "
f"`ray list nodes`.\n"
f"If running on K8s with limited GPUs, consider reducing "
f"--tensor-parallel-size to match available GPU resources."
) from None
else:
raise ValueError(
"Cannot provide a placement group of "
f"{placement_group_specs=} within "
f"{PG_WAIT_TIMEOUT} seconds. See "
"`ray status` and `ray list nodes` to make sure the cluster "
"has enough resources."
) from None
def _wait_until_pg_removed(current_placement_group: "PlacementGroup"):
ray.util.remove_placement_group(current_placement_group)
s = time.time()
wait_interval = 10
while time.time() - s < PG_WAIT_TIMEOUT:
pg = ray.util.get_current_placement_group()
if pg is None:
break
# Exponential backoff for warning print.
wait_interval *= 2
logger.info(
"Waiting for removing a placement group of specs for %d seconds.",
int(time.time() - s),
)
time.sleep(wait_interval)
def initialize_ray_cluster(
parallel_config: ParallelConfig,
ray_address: str | None = None,
):
"""Initialize the distributed cluster with Ray.
it will connect to the Ray cluster and create a placement group
for the workers, which includes the specification of the resources
for each distributed worker.
Args:
parallel_config: The configurations for parallel execution.
ray_address: The address of the Ray cluster. If None, uses
the default Ray cluster address.
"""
assert_ray_available()
from vllm.platforms import current_platform
# Prevalidate GPU requirements before Ray processing
if current_platform.is_cuda() and parallel_config.world_size > 1:
from vllm.utils.torch_utils import cuda_device_count_stateless
available_gpus = cuda_device_count_stateless()
if parallel_config.world_size > available_gpus:
logger.warning(
"Tensor parallel size (%d) exceeds available GPUs (%d). "
"This may result in Ray placement group allocation failures. "
"Consider reducing tensor_parallel_size to %d or less, "
"or ensure your Ray cluster has %d GPUs available.",
parallel_config.world_size,
available_gpus,
available_gpus,
parallel_config.world_size,
)
if ray.is_initialized():
logger.info("Ray is already initialized. Skipping Ray initialization.")
elif current_platform.is_rocm() or current_platform.is_xpu():
# Try to connect existing ray instance and create a new one if not found
try:
ray.init("auto")
except ConnectionError:
logger.warning(
"No existing RAY instance detected. "
"A new instance will be launched with current node resources."
)
ray.init(
address=ray_address,
num_gpus=parallel_config.world_size,
runtime_env=parallel_config.ray_runtime_env,
)
else:
ray.init(address=ray_address, runtime_env=parallel_config.ray_runtime_env)
device_str = current_platform.ray_device_key
if not device_str:
raise ValueError(
f"current platform {current_platform.device_name} does not support ray."
)
# Create or get the placement group for worker processes
if parallel_config.placement_group:
current_placement_group = parallel_config.placement_group
else:
current_placement_group = ray.util.get_current_placement_group()
if current_placement_group:
logger.info("Using the existing placement group")
# We are in a placement group
bundles = current_placement_group.bundle_specs
# Verify that we can use the placement group.
device_bundles = 0
for bundle in bundles:
bundle_devices = bundle.get(device_str, 0)
if bundle_devices > 1:
raise ValueError(
f"Placement group bundle cannot have more than 1 {device_str}."
)
if bundle_devices:
device_bundles += 1
if parallel_config.world_size > device_bundles:
raise ValueError(
f"The number of required {device_str}s exceeds the total "
f"number of available {device_str}s in the placement group. "
f"Required number of devices: {parallel_config.world_size}. "
f"Total number of devices: {device_bundles}."
)
else:
logger.info("No current placement group found. Creating a new placement group.")
num_devices_in_cluster = ray.cluster_resources().get(device_str, 0)
# Log a warning message and delay resource allocation failure response.
# Avoid immediate rejection to allow user-initiated placement group
# created and wait cluster to be ready
if parallel_config.world_size > num_devices_in_cluster:
logger.warning(
"The number of required %ss exceeds the total "
"number of available %ss in the placement group.",
device_str,
device_str,
)
# Create a new placement group
placement_group_specs: list[dict[str, float]] = [
{device_str: 1.0} for _ in range(parallel_config.world_size)
]
# vLLM engine is also a worker to execute model with an accelerator,
# so it requires to have the device in a current node. Check if
# the current node has at least one device.
current_ip = get_ip()
current_node_id = ray.get_runtime_context().get_node_id()
current_node_resource = available_resources_per_node()[current_node_id]
if current_node_resource.get(device_str, 0) < 1:
raise ValueError(
f"Current node has no {device_str} available. "
f"{current_node_resource=}. vLLM engine cannot start without "
f"{device_str}. Make sure you have at least 1 {device_str} "
f"available in a node {current_node_id=} {current_ip=}."
)
# This way, at least bundle is required to be created in a current
# node.
placement_group_specs[0][f"node:{current_ip}"] = 0.001
# By default, Ray packs resources as much as possible.
current_placement_group = ray.util.placement_group(
placement_group_specs, strategy="PACK"
)
_wait_until_pg_ready(current_placement_group)
assert current_placement_group is not None
_verify_bundles(current_placement_group, parallel_config, device_str)
# Set the placement group in the parallel config
parallel_config.placement_group = current_placement_group
def get_num_tpu_nodes() -> int:
from ray._private.accelerators import TPUAcceleratorManager
cluster_resources = ray.cluster_resources()
total_tpus = int(cluster_resources["TPU"])
tpus_per_node = TPUAcceleratorManager.get_current_node_num_accelerators()
assert total_tpus % tpus_per_node == 0
return total_tpus // tpus_per_node
def get_num_nodes_in_placement_group() -> int:
pg_table = ray.util.placement_group_table()
current_pg = ray.util.get_current_placement_group()
num_nodes = 0
if current_pg:
nodes_in_pg = set()
for pg_key, pg in pg_table.items():
if pg_key == current_pg.id.hex():
for _, node in pg["bundles_to_node_id"].items():
nodes_in_pg.add(node)
num_nodes = len(nodes_in_pg)
return num_nodes

View File

@@ -0,0 +1,186 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
from collections.abc import Callable
from concurrent.futures import Future, ThreadPoolExecutor
from functools import cached_property
from multiprocessing import Lock
from typing import Any
import torch
import torch.distributed as dist
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.utils.network_utils import get_distributed_init_method, get_ip, get_open_port
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
from vllm.v1.executor.abstract import Executor
from vllm.v1.outputs import AsyncModelRunnerOutput, DraftTokenIds, ModelRunnerOutput
from vllm.v1.serial_utils import run_method
from vllm.v1.worker.worker_base import WorkerWrapperBase
logger = init_logger(__name__)
class UniProcExecutor(Executor):
def _init_executor(self) -> None:
"""Initialize the worker and load the model."""
self.driver_worker = WorkerWrapperBase(vllm_config=self.vllm_config, rpc_rank=0)
distributed_init_method, rank, local_rank = self._distributed_args()
kwargs = dict(
vllm_config=self.vllm_config,
local_rank=local_rank,
rank=rank,
distributed_init_method=distributed_init_method,
is_driver_worker=True,
shared_worker_lock=Lock(),
)
self.async_output_thread: ThreadPoolExecutor | None = None
if self.max_concurrent_batches > 1:
self.async_output_thread = ThreadPoolExecutor(
max_workers=1, thread_name_prefix="WorkerAsyncOutput"
)
self.driver_worker.init_worker(all_kwargs=[kwargs])
self.driver_worker.init_device()
self.driver_worker.load_model()
def _distributed_args(self) -> tuple[str, int, int]:
"""Return (distributed_init_method, rank, local_rank)."""
distributed_init_method = get_distributed_init_method(get_ip(), get_open_port())
# set local rank as the device index if specified
device_info = self.vllm_config.device_config.device.__str__().split(":")
local_rank = int(device_info[1]) if len(device_info) > 1 else 0
return distributed_init_method, 0, local_rank
@cached_property
def max_concurrent_batches(self) -> int:
return 2 if self.scheduler_config.async_scheduling else 1
def collective_rpc( # type: ignore[override]
self,
method: str | Callable,
timeout: float | None = None,
args: tuple = (),
kwargs: dict | None = None,
non_block: bool = False,
single_value: bool = False,
) -> Any:
if kwargs is None:
kwargs = {}
if not non_block:
result = run_method(self.driver_worker, method, args, kwargs)
return result if single_value else [result]
try:
result = run_method(self.driver_worker, method, args, kwargs)
if isinstance(result, AsyncModelRunnerOutput):
if (async_thread := self.async_output_thread) is not None:
if single_value:
return async_thread.submit(result.get_output)
def get_output_list() -> list[Any]:
return [result.get_output()]
return async_thread.submit(get_output_list)
result = result.get_output()
future = Future[Any]()
future.set_result(result if single_value else [result])
except Exception as e:
future = Future[Any]()
future.set_exception(e)
return future
def execute_model( # type: ignore[override]
self, scheduler_output: SchedulerOutput, non_block: bool = False
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
return self.collective_rpc(
"execute_model",
args=(scheduler_output,),
non_block=non_block,
single_value=True,
)
def sample_tokens( # type: ignore[override]
self, grammar_output: GrammarOutput | None, non_block: bool = False
) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
return self.collective_rpc(
"sample_tokens",
args=(grammar_output,),
non_block=non_block,
single_value=True,
)
def take_draft_token_ids(self) -> DraftTokenIds | None:
return self.collective_rpc("take_draft_token_ids", single_value=True)
def check_health(self) -> None:
# UniProcExecutor will always be healthy as long as
# it's running.
return
def reinitialize_distributed(
self, reconfig_request: ReconfigureDistributedRequest
) -> None:
self.driver_worker.reinitialize_distributed(reconfig_request)
if (
reconfig_request.new_data_parallel_rank
== ReconfigureRankType.SHUTDOWN_CURRENT_RANK
):
self.shutdown()
def shutdown(self) -> None:
if worker := self.driver_worker:
worker.shutdown()
class ExecutorWithExternalLauncher(UniProcExecutor):
"""An executor that uses external launchers to launch engines,
specially designed for torchrun-compatible launchers, for
offline inference with tensor parallelism.
see https://github.com/vllm-project/vllm/issues/11400 for
the motivation, and examples/offline_inference/torchrun_example.py
for the usage example.
The key idea: although it is tensor-parallel inference, we only
create one worker per executor, users will launch multiple
engines with torchrun-compatible launchers, and all these engines
work together to process the same prompts. When scheduling is
deterministic, all the engines will generate the same outputs,
and they don't need to synchronize the states with each other.
"""
def _init_executor(self) -> None:
"""Initialize the worker and load the model."""
assert not envs.VLLM_ENABLE_V1_MULTIPROCESSING, (
"To get deterministic execution, "
"please set VLLM_ENABLE_V1_MULTIPROCESSING=0"
)
super()._init_executor()
def _distributed_args(self) -> tuple[str, int, int]:
# engines are launched in torchrun-compatible launchers
# so we can use the env:// method.
# required env vars:
# - RANK
# - LOCAL_RANK
# - MASTER_ADDR
# - MASTER_PORT
distributed_init_method = "env://"
rank = int(os.environ["RANK"])
local_rank = int(os.environ["LOCAL_RANK"])
return distributed_init_method, rank, local_rank
def determine_available_memory(self) -> list[int]: # in bytes
# we need to get the min across all ranks.
memory = super().determine_available_memory()
from vllm.distributed.parallel_state import get_world_group
cpu_group = get_world_group().cpu_group
memory_tensor = torch.tensor([memory], device="cpu", dtype=torch.int64)
dist.all_reduce(memory_tensor, group=cpu_group, op=dist.ReduceOp.MIN)
return [memory_tensor.item()]