[Core][Bugfix][Perf] Introduce MQLLMEngine to avoid asyncio OH (#8157)
Co-authored-by: Nick Hill <nickhill@us.ibm.com> Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com> Co-authored-by: Simon Mo <simon.mo@hey.com>
This commit is contained in:
co-authored by
Nick Hill
rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Robert Shaw
Simon Mo
parent
9d104b5beb
commit
7c7714d856
@@ -601,9 +601,12 @@ class AsyncLLMEngine:
|
||||
return self._errored_with is not None
|
||||
|
||||
@property
|
||||
def limit_concurrency(self) -> Optional[int]:
|
||||
"""Maximum number of concurrently running requests."""
|
||||
return None
|
||||
def dead_error(self) -> BaseException:
|
||||
return AsyncEngineDeadError(
|
||||
"Background loop is not running. If it was running, "
|
||||
"inspect the output to find the stacktrace of the "
|
||||
"error that caused the background loop to stop "
|
||||
"(AsyncEngineDeadError).")
|
||||
|
||||
def set_errored(self, exc: Exception) -> None:
|
||||
self._errored_with = exc
|
||||
|
||||
@@ -1289,6 +1289,7 @@ class LLMEngine:
|
||||
# torch.distributed ops which may otherwise timeout, and unblocks
|
||||
# the RPC thread in the workers so that they can process any other
|
||||
# queued control plane messages, such as add/remove lora adapters.
|
||||
logger.debug("Stopping remote worker execution loop.")
|
||||
self.model_executor.stop_remote_worker_execution_loop()
|
||||
|
||||
return ctx.request_outputs
|
||||
|
||||
@@ -0,0 +1,73 @@
|
||||
from dataclasses import dataclass
|
||||
from enum import Enum
|
||||
from typing import List, Mapping, Optional, Union
|
||||
|
||||
from vllm.inputs import PromptInputs
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.outputs import RequestOutput
|
||||
from vllm.prompt_adapter.request import PromptAdapterRequest
|
||||
from vllm.sampling_params import SamplingParams
|
||||
|
||||
VLLM_RPC_SUCCESS_STR = "SUCCESS"
|
||||
|
||||
IPC_INPUT_EXT = "_input_socket"
|
||||
IPC_OUTPUT_EXT = "_output_socket"
|
||||
IPC_HEALTH_EXT = "_health_socket"
|
||||
IPC_DATA_EXT = "_data_socket"
|
||||
|
||||
|
||||
class MQEngineDeadError(RuntimeError):
|
||||
pass
|
||||
|
||||
|
||||
@dataclass
|
||||
class RPCGenerateRequest:
|
||||
inputs: PromptInputs
|
||||
sampling_params: SamplingParams
|
||||
request_id: str
|
||||
lora_request: Optional[LoRARequest] = None
|
||||
trace_headers: Optional[Mapping[str, str]] = None
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None
|
||||
|
||||
|
||||
@dataclass
|
||||
class RPCError:
|
||||
request_id: Optional[str]
|
||||
is_engine_errored: bool
|
||||
exception: BaseException
|
||||
|
||||
|
||||
@dataclass
|
||||
class RPCAbortRequest:
|
||||
request_id: str
|
||||
|
||||
|
||||
class RPCHealthRequest:
|
||||
pass
|
||||
|
||||
|
||||
class RPCStartupRequest(Enum):
|
||||
IS_SERVER_READY = 1
|
||||
|
||||
|
||||
@dataclass
|
||||
class RPCStartupResponse:
|
||||
tracing_enabled: bool
|
||||
|
||||
|
||||
RPC_REQUEST_T = Union[RPCGenerateRequest, RPCAbortRequest, RPCHealthRequest,
|
||||
RPCStartupRequest]
|
||||
|
||||
REQUEST_OUTPUTS_T = Union[List[RequestOutput], RPCError]
|
||||
|
||||
|
||||
def ENGINE_DEAD_ERROR(
|
||||
error: Optional[BaseException] = None) -> MQEngineDeadError:
|
||||
if error is None:
|
||||
return MQEngineDeadError(
|
||||
"Engine loop is not running. Inspect the stacktrace to "
|
||||
"find the original error")
|
||||
|
||||
return MQEngineDeadError(
|
||||
"Engine loop is not running. Inspect the stacktrace to "
|
||||
f"find the original error: {repr(error)}.")
|
||||
@@ -0,0 +1,452 @@
|
||||
import asyncio
|
||||
import copy
|
||||
import pickle
|
||||
from contextlib import contextmanager, suppress
|
||||
from typing import (Any, AsyncGenerator, Dict, Iterator, Mapping, Optional,
|
||||
Union)
|
||||
|
||||
import cloudpickle
|
||||
import zmq
|
||||
import zmq.asyncio
|
||||
from zmq import Frame # type: ignore[attr-defined]
|
||||
from zmq.asyncio import Socket
|
||||
|
||||
from vllm.config import DecodingConfig, EngineConfig, ModelConfig
|
||||
from vllm.engine.arg_utils import AsyncEngineArgs
|
||||
# yapf conflicts with isort for this block
|
||||
# yapf: disable
|
||||
from vllm.engine.multiprocessing import (ENGINE_DEAD_ERROR, IPC_DATA_EXT,
|
||||
IPC_HEALTH_EXT, IPC_INPUT_EXT,
|
||||
IPC_OUTPUT_EXT, RPC_REQUEST_T,
|
||||
VLLM_RPC_SUCCESS_STR, RPCAbortRequest,
|
||||
RPCError, RPCGenerateRequest,
|
||||
RPCHealthRequest, RPCStartupRequest,
|
||||
RPCStartupResponse)
|
||||
# yapf: enable
|
||||
from vllm.envs import VLLM_RPC_TIMEOUT
|
||||
from vllm.inputs import PromptInputs
|
||||
from vllm.logger import init_logger
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.outputs import EmbeddingRequestOutput, RequestOutput
|
||||
from vllm.prompt_adapter.request import PromptAdapterRequest
|
||||
from vllm.sampling_params import SamplingParams
|
||||
from vllm.transformers_utils.tokenizer_group import init_tokenizer_from_configs
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class MQClientClosedError(Exception):
|
||||
"""Exception class raised when the client is used post-close.
|
||||
|
||||
The client can be closed, which closes the ZMQ context. This normally
|
||||
happens on server shutdown. In some cases, methods like abort and
|
||||
do_log_stats will still be called and then try to open a socket, which
|
||||
causes a ZMQError and creates a huge stack trace.
|
||||
So, we throw this error such that we can suppress it.
|
||||
"""
|
||||
|
||||
|
||||
class MQLLMEngineClient:
|
||||
"""A client wrapper for MQLLMEngine that conforms to the
|
||||
EngineClient protocol.
|
||||
|
||||
MQLLMEngine and MQLLMEngineClient are intended to run in separate
|
||||
processes communicating via zeromq ipc sockets.
|
||||
|
||||
The entrypoint to MQLLMEngineClient is through the generate()
|
||||
method. On generate() MQLLMEngine does three things:
|
||||
- Creates an asyncio output queue
|
||||
- Sends a RPCGenerateRequest to the MQLLMEngine via zmq
|
||||
- Pulls RequestOutputs from its queue and yields them
|
||||
|
||||
MQLLMEngine runs two background loops:
|
||||
- output_loop: the output loop pulls List[RequestOutput]
|
||||
from the MQLLMEngine via zmq (each list is the output
|
||||
of one engine_step in the LLMEngine). It then parses
|
||||
the list and pushes individual request_outputs into
|
||||
the corresponding output_queue such that they can be
|
||||
consumed by the .generate() method.
|
||||
- health_loop: the health loop queries the health socket
|
||||
every N seconds, confirming the engine is healthy
|
||||
"""
|
||||
|
||||
def __init__(self, ipc_path: str, engine_config: EngineConfig):
|
||||
self.context = zmq.asyncio.Context()
|
||||
self._errored_with: Optional[BaseException] = None
|
||||
|
||||
# Get the configs.
|
||||
self.model_config = engine_config.model_config
|
||||
self.decoding_config = engine_config.decoding_config
|
||||
|
||||
# Create the tokenizer group.
|
||||
self.tokenizer = init_tokenizer_from_configs(
|
||||
model_config=self.model_config,
|
||||
scheduler_config=engine_config.scheduler_config,
|
||||
parallel_config=engine_config.parallel_config,
|
||||
enable_lora=bool(engine_config.lora_config),
|
||||
)
|
||||
|
||||
# Send RPCGenerateRequest to the MQLLMEngine.
|
||||
self.input_socket: Socket = self.context.socket(zmq.constants.PUSH)
|
||||
self.input_socket.connect(f"{ipc_path}{IPC_INPUT_EXT}")
|
||||
|
||||
# Receive streams of RequestOutput from the MQLLMEngine.
|
||||
self.output_socket: Socket = self.context.socket(zmq.constants.PULL)
|
||||
self.output_socket.connect(f"{ipc_path}{IPC_OUTPUT_EXT}")
|
||||
|
||||
# IPC path for ack of check_health requests.
|
||||
self.health_socket: Socket = self.context.socket(zmq.constants.PULL)
|
||||
self.health_socket.connect(f"{ipc_path}{IPC_HEALTH_EXT}")
|
||||
|
||||
# IPC path for the data socket.
|
||||
self.data_ipc_path = f"{ipc_path}{IPC_DATA_EXT}"
|
||||
|
||||
# Stream for each individual request.
|
||||
self.output_queues: Dict[str, asyncio.Queue] = {}
|
||||
self.output_loop = asyncio.create_task(self.run_output_handler_loop())
|
||||
|
||||
# Loop to check health of the LLMEngine periodically.
|
||||
# Started after the MQLLMEngine is ready.
|
||||
self.health_loop: Optional[asyncio.Task] = None
|
||||
|
||||
@staticmethod
|
||||
def is_unsupported_config(engine_args: AsyncEngineArgs):
|
||||
if engine_args.pipeline_parallel_size > 1:
|
||||
return True
|
||||
|
||||
is_embedding = ModelConfig(
|
||||
model=engine_args.model,
|
||||
revision=engine_args.revision,
|
||||
tokenizer=engine_args.model,
|
||||
tokenizer_mode="auto",
|
||||
trust_remote_code=engine_args.trust_remote_code,
|
||||
quantization=engine_args.quantization,
|
||||
seed=0,
|
||||
dtype="auto").embedding_mode
|
||||
|
||||
return is_embedding
|
||||
|
||||
@contextmanager
|
||||
def get_data_socket(self) -> Iterator[Socket]:
|
||||
socket = self.context.socket(zmq.constants.DEALER)
|
||||
try:
|
||||
socket.connect(self.data_ipc_path)
|
||||
yield socket
|
||||
finally:
|
||||
socket.close(linger=0)
|
||||
|
||||
async def run_check_health_loop(self, timeout: int):
|
||||
"""Background loop that continually probes the RPCServer for health.
|
||||
|
||||
The loop sends CHECK_HEALTH requests to the INPUT_SOCKET, which
|
||||
the MQLLMEngine server is blocking on.
|
||||
|
||||
The Server replies on the HEALTH_SOCKET (rather than on the
|
||||
OUTPUT_SOCKET such that the messages are not intermingled with
|
||||
output streaming).
|
||||
"""
|
||||
|
||||
try:
|
||||
while True:
|
||||
if await self.health_socket.poll(timeout=timeout) == 0:
|
||||
# Wakeup every N seconds and do a health probe.
|
||||
await self._send_one_way_rpc_request(
|
||||
RPCHealthRequest(), self.input_socket)
|
||||
|
||||
# Wait for ack from the health socket.
|
||||
await self._await_ack(error_message="Health check failed.",
|
||||
socket=self.health_socket)
|
||||
else:
|
||||
# Server sent a health status message unprompted.
|
||||
await self._check_success(
|
||||
error_message="Health check failed.",
|
||||
socket=self.health_socket)
|
||||
|
||||
logger.debug("Health probe successful.")
|
||||
|
||||
except asyncio.CancelledError:
|
||||
logger.debug("Shutting down MQLLMEngineClient check health loop.")
|
||||
|
||||
except Exception as e:
|
||||
self._set_errored(e)
|
||||
|
||||
async def run_output_handler_loop(self):
|
||||
"""Get RequestOutputs from Engine and stream to Request Queues"""
|
||||
|
||||
try:
|
||||
while True:
|
||||
# Poll, checking for ENGINE_DEAD
|
||||
while await self.output_socket.poll(timeout=VLLM_RPC_TIMEOUT
|
||||
) == 0:
|
||||
logger.debug("Waiting for output from MQLLMEngine.")
|
||||
|
||||
# If errored, alert all running requests.
|
||||
if self.errored:
|
||||
for queue_j in tuple(self.output_queues.values()):
|
||||
queue_j.put_nowait(
|
||||
ENGINE_DEAD_ERROR(self._errored_with))
|
||||
return
|
||||
|
||||
message: Frame = await self.output_socket.recv(copy=False)
|
||||
request_outputs = pickle.loads(message.buffer)
|
||||
|
||||
is_error = isinstance(request_outputs,
|
||||
(BaseException, RPCError))
|
||||
if is_error:
|
||||
if isinstance(request_outputs, RPCError):
|
||||
rpc_error: RPCError = request_outputs
|
||||
request_id = rpc_error.request_id
|
||||
exception = rpc_error.exception
|
||||
is_engine_errored = rpc_error.is_engine_errored
|
||||
else:
|
||||
# MPLLMEngine should always return an RPCError to
|
||||
# the output_socket when an issue arises.
|
||||
# If we are here, we are in a bad state and
|
||||
# should shut down the server.
|
||||
error: BaseException = request_outputs
|
||||
logger.error(
|
||||
"Received Exception %s rather than RPCError from "
|
||||
"MPLLMEngine. This should never happen.", error)
|
||||
request_id = None
|
||||
exception = error
|
||||
is_engine_errored = True
|
||||
|
||||
# Set to error state only on engine critical error
|
||||
# (and record only the first one)
|
||||
if is_engine_errored and not self._errored_with:
|
||||
self._errored_with = exception
|
||||
|
||||
if request_id is None:
|
||||
for queue_i in tuple(self.output_queues.values()):
|
||||
queue_i.put_nowait(exception)
|
||||
else:
|
||||
queue = self.output_queues.get(request_id)
|
||||
if queue is not None:
|
||||
queue.put_nowait(exception)
|
||||
else:
|
||||
# Put each output into the appropriate steam.
|
||||
for request_output in request_outputs:
|
||||
queue = self.output_queues.get(
|
||||
request_output.request_id)
|
||||
if queue is not None:
|
||||
queue.put_nowait(request_output)
|
||||
|
||||
except asyncio.CancelledError:
|
||||
logger.debug("Shutting down MQLLMEngineClient output handler.")
|
||||
|
||||
async def setup(self):
|
||||
"""Setup the client before it starts sending server requests."""
|
||||
|
||||
with self.get_data_socket() as socket:
|
||||
# Wait until server is ready.
|
||||
response = await self._wait_for_server_rpc(socket)
|
||||
|
||||
self.tracing_flag = response.tracing_enabled
|
||||
|
||||
# Start health_loop.
|
||||
self.health_loop = asyncio.create_task(
|
||||
self.run_check_health_loop(timeout=VLLM_RPC_TIMEOUT))
|
||||
|
||||
def close(self):
|
||||
"""Destroy the ZeroMQ Context."""
|
||||
# Close all sockets and terminate the context.
|
||||
self.context.destroy(linger=0)
|
||||
|
||||
# Cancel background tasks.
|
||||
if self.health_loop is not None:
|
||||
self.health_loop.cancel()
|
||||
self.output_loop.cancel()
|
||||
|
||||
def _set_errored(self, e: BaseException):
|
||||
logger.exception(repr(e))
|
||||
if self._errored_with is None:
|
||||
self._errored_with = e
|
||||
|
||||
@staticmethod
|
||||
async def _send_get_data_rpc_request(request: RPCStartupRequest,
|
||||
expected_type: Any,
|
||||
error_message: str,
|
||||
socket: Socket) -> Any:
|
||||
"""Send an RPC request that is expecting data back."""
|
||||
|
||||
# Ping RPCServer with a request.
|
||||
await socket.send_multipart((pickle.dumps(request), ), copy=False)
|
||||
|
||||
# Make sure the server responds in time.
|
||||
if await socket.poll(timeout=VLLM_RPC_TIMEOUT) == 0:
|
||||
raise TimeoutError("RPCServer didn't reply within "
|
||||
f"{VLLM_RPC_TIMEOUT} ms")
|
||||
|
||||
# Await the data from the Server.
|
||||
frame = await socket.recv(copy=False)
|
||||
data = pickle.loads(frame.buffer)
|
||||
|
||||
if isinstance(data, BaseException):
|
||||
raise data
|
||||
elif not isinstance(data, expected_type):
|
||||
raise ValueError(error_message)
|
||||
|
||||
return data
|
||||
|
||||
@staticmethod
|
||||
async def _send_one_way_rpc_request(request: RPC_REQUEST_T,
|
||||
socket: Socket):
|
||||
"""Send one-way RPC request to trigger an action."""
|
||||
|
||||
if socket.closed:
|
||||
raise MQClientClosedError()
|
||||
|
||||
await socket.send_multipart((pickle.dumps(request), ))
|
||||
|
||||
async def _await_ack(self, error_message: str, socket: Socket):
|
||||
"""Await acknowledgement that a request succeeded."""
|
||||
|
||||
if socket.closed:
|
||||
raise MQClientClosedError()
|
||||
|
||||
if await socket.poll(timeout=VLLM_RPC_TIMEOUT) == 0:
|
||||
raise TimeoutError("MQLLMEngine didn't reply within "
|
||||
f"{VLLM_RPC_TIMEOUT}ms")
|
||||
|
||||
await self._check_success(error_message, socket)
|
||||
|
||||
@staticmethod
|
||||
async def _check_success(error_message: str, socket: Socket):
|
||||
"""Confirm that socket has a VLLM_RPC_SUCCESS_STR message"""
|
||||
|
||||
if socket.closed:
|
||||
raise MQClientClosedError()
|
||||
|
||||
frame = await socket.recv(copy=False)
|
||||
response = pickle.loads(frame.buffer)
|
||||
|
||||
# Raise error if unsuccessful
|
||||
if isinstance(response, BaseException):
|
||||
raise response
|
||||
elif (not isinstance(response, str)
|
||||
or response != VLLM_RPC_SUCCESS_STR):
|
||||
raise ValueError(error_message)
|
||||
|
||||
async def get_tokenizer(self, lora_request: LoRARequest):
|
||||
return await self.tokenizer.get_lora_tokenizer_async(lora_request)
|
||||
|
||||
async def get_decoding_config(self) -> DecodingConfig:
|
||||
return self.decoding_config
|
||||
|
||||
async def get_model_config(self) -> ModelConfig:
|
||||
return self.model_config
|
||||
|
||||
async def is_tracing_enabled(self) -> bool:
|
||||
return self.tracing_flag
|
||||
|
||||
async def _wait_for_server_rpc(self, socket: Socket) -> RPCStartupResponse:
|
||||
"""Wait for the RPCServer to start up."""
|
||||
|
||||
return await self._send_get_data_rpc_request(
|
||||
request=RPCStartupRequest.IS_SERVER_READY,
|
||||
expected_type=RPCStartupResponse,
|
||||
error_message="Unable to start RPC Server",
|
||||
socket=socket)
|
||||
|
||||
async def abort(self, request_id: str):
|
||||
"""Send an ABORT_REQUEST signal to the RPC Server"""
|
||||
|
||||
with suppress(MQClientClosedError):
|
||||
await self._send_one_way_rpc_request(
|
||||
request=RPCAbortRequest(request_id), socket=self.input_socket)
|
||||
|
||||
async def do_log_stats(self):
|
||||
"""Ignore do_log_stats (handled on MQLLMEngine polling)"""
|
||||
pass
|
||||
|
||||
async def check_health(self):
|
||||
"""
|
||||
The check health loop probes the health status of the
|
||||
Engine's health every N seconds and sets _errored_with
|
||||
if the engine is unhealthy.
|
||||
"""
|
||||
if self._errored_with is not None:
|
||||
raise self._errored_with
|
||||
|
||||
@property
|
||||
def is_running(self) -> bool:
|
||||
return not self.errored
|
||||
|
||||
@property
|
||||
def is_stopped(self) -> bool:
|
||||
return self.errored
|
||||
|
||||
@property
|
||||
def errored(self) -> bool:
|
||||
return self._errored_with is not None
|
||||
|
||||
async def generate(
|
||||
self,
|
||||
inputs: PromptInputs,
|
||||
sampling_params: SamplingParams,
|
||||
request_id: str,
|
||||
lora_request: Optional[LoRARequest] = None,
|
||||
trace_headers: Optional[Mapping[str, str]] = None,
|
||||
prompt_adapter_request: Optional[PromptAdapterRequest] = None
|
||||
) -> AsyncGenerator[RequestOutput, None]:
|
||||
"""Send an RPCGenerateRequest to the RPCServer and stream responses."""
|
||||
|
||||
# If already dead, error out.
|
||||
if self._errored_with is not None:
|
||||
raise ENGINE_DEAD_ERROR(self._errored_with)
|
||||
|
||||
# 1) Create output queue for this requests.
|
||||
queue: asyncio.Queue[Union[RequestOutput,
|
||||
BaseException]] = asyncio.Queue()
|
||||
self.output_queues[request_id] = queue
|
||||
|
||||
try:
|
||||
# 2) Detach logits processors so that they can be pickled
|
||||
# separately (may require cloudpickle which is slower)
|
||||
if sampling_params.logits_processors:
|
||||
# Defensive shallow copy
|
||||
sampling_params = copy.copy(sampling_params)
|
||||
logits_processors = sampling_params.logits_processors
|
||||
sampling_params.logits_processors = None
|
||||
lp_bytes = cloudpickle.dumps(logits_processors)
|
||||
else:
|
||||
lp_bytes = None
|
||||
|
||||
request_bytes = pickle.dumps(
|
||||
RPCGenerateRequest(
|
||||
inputs=inputs,
|
||||
sampling_params=sampling_params,
|
||||
request_id=request_id,
|
||||
lora_request=lora_request,
|
||||
trace_headers=trace_headers,
|
||||
prompt_adapter_request=prompt_adapter_request))
|
||||
|
||||
# 3) Send the RPCGenerateRequest to the MQLLMEngine.
|
||||
parts = (request_bytes,
|
||||
lp_bytes) if lp_bytes else (request_bytes, )
|
||||
await self.input_socket.send_multipart(parts, copy=False)
|
||||
|
||||
# 4) Stream the RequestOutputs from the output queue. Note
|
||||
# that the output_loop pushes RequestOutput objects to this
|
||||
# queue after pulling them from the zmq socket.
|
||||
finished = False
|
||||
try:
|
||||
while not finished:
|
||||
request_output = await queue.get()
|
||||
|
||||
if isinstance(request_output, BaseException):
|
||||
raise request_output
|
||||
|
||||
finished = request_output.finished
|
||||
yield request_output
|
||||
finally:
|
||||
# Request was canceled by the client.
|
||||
if not finished and not self.errored:
|
||||
await self.abort(request_id)
|
||||
finally:
|
||||
self.output_queues.pop(request_id)
|
||||
|
||||
async def encode(self, *args,
|
||||
**kwargs) -> AsyncGenerator[EmbeddingRequestOutput, None]:
|
||||
raise NotImplementedError(
|
||||
"Embeddings not supported with multiprocessing backend")
|
||||
@@ -0,0 +1,321 @@
|
||||
import pickle
|
||||
import signal
|
||||
from contextlib import contextmanager
|
||||
from typing import Iterator, List, Optional, Union
|
||||
|
||||
import cloudpickle
|
||||
import zmq
|
||||
|
||||
from vllm import AsyncEngineArgs, LLMEngine
|
||||
from vllm.config import (DecodingConfig, LoRAConfig, ModelConfig,
|
||||
ParallelConfig, SchedulerConfig)
|
||||
# yapf conflicts with isort for this block
|
||||
# yapf: disable
|
||||
from vllm.engine.multiprocessing import (ENGINE_DEAD_ERROR, IPC_DATA_EXT,
|
||||
IPC_HEALTH_EXT, IPC_INPUT_EXT,
|
||||
IPC_OUTPUT_EXT, REQUEST_OUTPUTS_T,
|
||||
VLLM_RPC_SUCCESS_STR, RPCAbortRequest,
|
||||
RPCError, RPCGenerateRequest,
|
||||
RPCHealthRequest, RPCStartupRequest,
|
||||
RPCStartupResponse)
|
||||
# yapf: enable
|
||||
from vllm.logger import init_logger
|
||||
from vllm.outputs import RequestOutput
|
||||
from vllm.usage.usage_lib import UsageContext
|
||||
|
||||
CONFIG_TYPE = Union[ModelConfig, DecodingConfig, ParallelConfig,
|
||||
SchedulerConfig, LoRAConfig]
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
POLLING_TIMEOUT_MS = 10000
|
||||
HEALTHY_RESPONSE = (pickle.dumps(VLLM_RPC_SUCCESS_STR), )
|
||||
|
||||
|
||||
class MQLLMEngine:
|
||||
"""A multiprocessing wrapper for :class:`LLMEngine`.
|
||||
|
||||
This class is used to wrap the :class:`LLMEngine` class to enable use
|
||||
in concurrnet manner. It runs a background loop and uses zeromq to
|
||||
receive new requests and stream outputs incrementally via ipc.
|
||||
|
||||
The :class:`LLMEngine.generate` is kicked off when a new
|
||||
RPCGenerateRequest is received by the input_socket.
|
||||
|
||||
The self.engine_loop checks the input_socket for new requests,
|
||||
adds them to the LLMEngine if there are any, calls the internal
|
||||
:class:`LLMEngine.step()`, and sends the RequestOutputs back over
|
||||
the output_socket.
|
||||
|
||||
If use_async_sockets is set, the logic associated with reading new
|
||||
requests from the socket and sending data to the socket is passed
|
||||
as a callback to the llm_engine, which calls the logic asynchronously
|
||||
such that the IPC can be overlapped with the GPU.
|
||||
|
||||
Args:
|
||||
ipc_path: Base path for zeromq interprocess messaging
|
||||
use_async_sockets: Whether to make send/recv async with GPU
|
||||
log_requests: Whether to log the requests.
|
||||
*args: Arguments for :class:`LLMEngine`.
|
||||
**kwargs: Arguments for :class:`LLMEngine`.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
ipc_path: str,
|
||||
use_async_sockets: bool,
|
||||
*args,
|
||||
log_requests: bool = True,
|
||||
**kwargs) -> None:
|
||||
self.engine = LLMEngine(*args, **kwargs)
|
||||
self.log_requests = log_requests
|
||||
|
||||
self.use_async_sockets = use_async_sockets
|
||||
if self.use_async_sockets:
|
||||
self.engine.process_request_outputs_callback = \
|
||||
self._async_socket_engine_callback
|
||||
|
||||
self.ctx = zmq.Context() # type: ignore[attr-defined]
|
||||
|
||||
# Receive input from the client.
|
||||
self.input_socket = self.ctx.socket(zmq.constants.PULL)
|
||||
self.input_socket.bind(f"{ipc_path}{IPC_INPUT_EXT}")
|
||||
|
||||
# Send output stream back to client.
|
||||
self.output_socket = self.ctx.socket(zmq.constants.PUSH)
|
||||
self.output_socket.bind(f"{ipc_path}{IPC_OUTPUT_EXT}")
|
||||
|
||||
# Send health status back to client.
|
||||
self.health_socket = self.ctx.socket(zmq.constants.PUSH)
|
||||
self.health_socket.bind(f"{ipc_path}{IPC_HEALTH_EXT}")
|
||||
|
||||
# IPC path for the data socket.
|
||||
self.data_ipc_path = f"{ipc_path}{IPC_DATA_EXT}"
|
||||
|
||||
# Error state.
|
||||
self._errored_with: Optional[BaseException] = None
|
||||
|
||||
@property
|
||||
def dead_error(self) -> BaseException:
|
||||
if self._errored_with is not None:
|
||||
return ENGINE_DEAD_ERROR(self._errored_with)
|
||||
else:
|
||||
return ENGINE_DEAD_ERROR()
|
||||
|
||||
@classmethod
|
||||
def from_engine_args(cls, engine_args: AsyncEngineArgs,
|
||||
usage_context: UsageContext, ipc_path: str):
|
||||
"""Creates an MQLLMEngine from the engine arguments."""
|
||||
|
||||
engine_config = engine_args.create_engine_config()
|
||||
|
||||
executor_class = LLMEngine._get_executor_cls(engine_config)
|
||||
|
||||
return cls(
|
||||
ipc_path=ipc_path,
|
||||
use_async_sockets=engine_config.model_config.use_async_output_proc,
|
||||
**engine_config.to_dict(),
|
||||
executor_class=executor_class,
|
||||
log_requests=not engine_args.disable_log_requests,
|
||||
log_stats=not engine_args.disable_log_stats,
|
||||
usage_context=usage_context)
|
||||
|
||||
def start(self):
|
||||
try:
|
||||
try:
|
||||
logger.debug("Starting Startup Loop.")
|
||||
self.run_startup_loop()
|
||||
logger.debug("Starting Engine Loop.")
|
||||
self.run_engine_loop()
|
||||
except Exception as e:
|
||||
logger.exception(repr(e))
|
||||
except KeyboardInterrupt:
|
||||
logger.debug("Shutting down MQLLMEngine.")
|
||||
finally:
|
||||
logger.debug("MQLLMEngine is shut down.")
|
||||
self.cleanup()
|
||||
|
||||
def cleanup(self):
|
||||
"""Cleanup zeromq state on shutdown."""
|
||||
# Closes all sockets and destroys context.
|
||||
self.ctx.destroy(linger=0)
|
||||
del self.engine
|
||||
|
||||
@contextmanager
|
||||
def make_data_socket(
|
||||
self) -> Iterator[zmq.Socket]: # type: ignore[name-defined]
|
||||
socket = self.ctx.socket(zmq.constants.ROUTER)
|
||||
try:
|
||||
socket.bind(self.data_ipc_path)
|
||||
yield socket
|
||||
finally:
|
||||
socket.close(linger=0)
|
||||
|
||||
def run_startup_loop(self) -> None:
|
||||
"""Startup loop for sending data from Engine -> Client."""
|
||||
|
||||
with self.make_data_socket() as socket:
|
||||
response: Union[RPCStartupResponse, BaseException]
|
||||
try:
|
||||
identity, message = socket.recv_multipart(copy=False)
|
||||
request: RPCStartupRequest = pickle.loads(message.buffer)
|
||||
|
||||
# Handle the query from the Client.
|
||||
if request == RPCStartupRequest.IS_SERVER_READY:
|
||||
tracing_enabled = self.engine.is_tracing_enabled()
|
||||
response = RPCStartupResponse(
|
||||
tracing_enabled=tracing_enabled)
|
||||
|
||||
except Exception as e:
|
||||
response = e
|
||||
|
||||
socket.send_multipart((identity, pickle.dumps(response)),
|
||||
copy=False)
|
||||
|
||||
def run_engine_loop(self):
|
||||
"""Core busy loop of the LLMEngine."""
|
||||
|
||||
while True:
|
||||
if not self.engine.has_unfinished_requests():
|
||||
# Poll until there is work to do.
|
||||
while self.input_socket.poll(timeout=POLLING_TIMEOUT_MS) == 0:
|
||||
self.engine.do_log_stats()
|
||||
logger.debug("Waiting for new requests in engine loop.")
|
||||
|
||||
# Handle any input from the client.
|
||||
self.handle_new_input()
|
||||
|
||||
# Engine step.
|
||||
request_outputs = self.engine_step()
|
||||
|
||||
# Send request outputs (if async, done in engine_step callback).
|
||||
if not self.use_async_sockets:
|
||||
self._send_outputs(request_outputs)
|
||||
|
||||
def engine_step(self) -> List[RequestOutput]:
|
||||
"""Engine step wrapper with error handling."""
|
||||
|
||||
try:
|
||||
return self.engine.step()
|
||||
except SystemExit:
|
||||
raise
|
||||
except BaseException as e:
|
||||
self._set_errored(e)
|
||||
rpc_err = RPCError(request_id=None,
|
||||
is_engine_errored=True,
|
||||
exception=e)
|
||||
self._send_outputs(rpc_err)
|
||||
raise e
|
||||
|
||||
def handle_new_input(self):
|
||||
"""Handle new input from the socket"""
|
||||
try:
|
||||
while self.input_socket.poll(timeout=0) != 0:
|
||||
frames = self.input_socket.recv_multipart(copy=False)
|
||||
request = pickle.loads(frames[0].buffer)
|
||||
|
||||
if isinstance(request, RPCGenerateRequest):
|
||||
if len(frames) > 1:
|
||||
# Use cloudpickle for logits processors
|
||||
lprocs = cloudpickle.loads(frames[1].buffer)
|
||||
request.sampling_params.logits_processors = lprocs
|
||||
self._handle_generate_request(request)
|
||||
elif isinstance(request, RPCAbortRequest):
|
||||
self._handle_abort_request(request)
|
||||
elif isinstance(request, RPCHealthRequest):
|
||||
self._handle_health_request()
|
||||
else:
|
||||
raise ValueError("Unknown RPCRequest Type: {request}")
|
||||
|
||||
except Exception as e:
|
||||
self._set_errored(e)
|
||||
self._send_unhealthy(e)
|
||||
raise e
|
||||
|
||||
def _handle_generate_request(self, request: RPCGenerateRequest):
|
||||
"""Handle RPCGenerateRequest by adding it to the LLMEngine."""
|
||||
request_id = request.request_id
|
||||
|
||||
if self._errored_with is not None:
|
||||
rpc_err = RPCError(request_id=request_id,
|
||||
is_engine_errored=True,
|
||||
exception=ENGINE_DEAD_ERROR(self._errored_with))
|
||||
self._send_outputs(rpc_err)
|
||||
|
||||
try:
|
||||
self.engine.add_request(
|
||||
request_id=request_id,
|
||||
inputs=request.inputs,
|
||||
params=request.sampling_params,
|
||||
lora_request=request.lora_request,
|
||||
trace_headers=request.trace_headers,
|
||||
prompt_adapter_request=request.prompt_adapter_request)
|
||||
|
||||
if self.log_requests:
|
||||
logger.info("Added request %s.", request.request_id)
|
||||
|
||||
except Exception as e:
|
||||
# We do not set self._errored = True here, since the error
|
||||
# is due to an issue adding this request to the engine,
|
||||
# rather than an issue with the engine itself.
|
||||
is_errored = self._errored_with is not None
|
||||
rpc_err = RPCError(request_id=request_id,
|
||||
is_engine_errored=is_errored,
|
||||
exception=e)
|
||||
self._send_outputs(rpc_err)
|
||||
|
||||
# Remove request from the engine.
|
||||
self.engine.abort_request(request_id)
|
||||
|
||||
def _handle_abort_request(self, request: RPCAbortRequest):
|
||||
self.engine.abort_request(request.request_id)
|
||||
if self.log_requests:
|
||||
logger.info("Aborted request %s.", request.request_id)
|
||||
|
||||
def _handle_health_request(self):
|
||||
if self._errored_with is not None:
|
||||
self._send_unhealthy(self._errored_with)
|
||||
|
||||
# Raises error if unhealthy.
|
||||
self.engine.check_health()
|
||||
self._send_healthy()
|
||||
|
||||
def _send_outputs(self, outputs: REQUEST_OUTPUTS_T):
|
||||
"""Send List of RequestOutput to RPCClient."""
|
||||
if outputs:
|
||||
output_bytes = pickle.dumps(outputs)
|
||||
self.output_socket.send_multipart((output_bytes, ), copy=False)
|
||||
|
||||
def _send_healthy(self):
|
||||
"""Send HEALTHY message to RPCClient."""
|
||||
self.health_socket.send_multipart(HEALTHY_RESPONSE, copy=False)
|
||||
|
||||
def _send_unhealthy(self, error: BaseException):
|
||||
"""Send UNHEALTHY message to RPCClient."""
|
||||
error_bytes = pickle.dumps(error)
|
||||
self.health_socket.send_multipart((error_bytes, ), copy=False)
|
||||
|
||||
def _async_socket_engine_callback(self,
|
||||
request_outputs: REQUEST_OUTPUTS_T):
|
||||
"""Callback used by engine to make socket handling async with GPU."""
|
||||
self._send_outputs(request_outputs)
|
||||
self.handle_new_input()
|
||||
|
||||
def _set_errored(self, e: BaseException):
|
||||
"""Log and set errored status if this is the first issue."""
|
||||
if self._errored_with is None:
|
||||
self._errored_with = e
|
||||
|
||||
|
||||
def run_mp_engine(engine_args: AsyncEngineArgs, usage_context: UsageContext,
|
||||
ipc_path: str):
|
||||
|
||||
def signal_handler(*_) -> None:
|
||||
# Interrupt server on sigterm
|
||||
raise KeyboardInterrupt("MQLLMEngine terminated")
|
||||
|
||||
signal.signal(signal.SIGTERM, signal_handler)
|
||||
|
||||
engine = MQLLMEngine.from_engine_args(engine_args=engine_args,
|
||||
usage_context=usage_context,
|
||||
ipc_path=ipc_path)
|
||||
engine.start()
|
||||
@@ -14,8 +14,8 @@ from vllm.transformers_utils.tokenizer import AnyTokenizer
|
||||
|
||||
|
||||
@runtime_checkable
|
||||
class AsyncEngineClient(Protocol):
|
||||
"""Protocol class for Clients to AsyncLLMEngine"""
|
||||
class EngineClient(Protocol):
|
||||
"""Protocol class for Clients to Engine"""
|
||||
|
||||
@property
|
||||
def is_running(self) -> bool:
|
||||
@@ -30,8 +30,8 @@ class AsyncEngineClient(Protocol):
|
||||
...
|
||||
|
||||
@property
|
||||
def limit_concurrency(self) -> Optional[int]:
|
||||
"""Maximum number of concurrently running requests."""
|
||||
def dead_error(self) -> BaseException:
|
||||
...
|
||||
|
||||
def generate(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user