[Core] Eliminate parallel worker per-step task scheduling overhead (#4894)
This commit is contained in:
@@ -1,11 +1,12 @@
|
||||
import asyncio
|
||||
from abc import abstractmethod
|
||||
from typing import Any, Dict, List, Optional, Set, Tuple
|
||||
from typing import Any, Awaitable, Dict, List, Optional, Set, Tuple, Union
|
||||
|
||||
from vllm.executor.executor_base import ExecutorAsyncBase
|
||||
from vllm.executor.gpu_executor import GPUExecutor
|
||||
from vllm.logger import init_logger
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.sequence import SamplerOutput
|
||||
from vllm.sequence import ExecuteModelRequest, SamplerOutput
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -13,6 +14,16 @@ logger = init_logger(__name__)
|
||||
class DistributedGPUExecutor(GPUExecutor):
|
||||
"""Abstract superclass of multi-GPU executor implementations."""
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
# This is non-None when the execute model loop is running
|
||||
# in the parallel workers. It's a coroutine in the AsyncLLMEngine case.
|
||||
self.parallel_worker_tasks: Optional[Union[Any, Awaitable[Any]]] = None
|
||||
# Updated by implementations that require additional args to be passed
|
||||
# to the _run_workers execute_model call
|
||||
self.extra_execute_model_run_workers_kwargs: Dict[str, Any] = {}
|
||||
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
def determine_num_available_blocks(self) -> Tuple[int, int]:
|
||||
"""Determine the number of available KV blocks.
|
||||
|
||||
@@ -52,13 +63,28 @@ class DistributedGPUExecutor(GPUExecutor):
|
||||
num_gpu_blocks=num_gpu_blocks,
|
||||
num_cpu_blocks=num_cpu_blocks)
|
||||
|
||||
def execute_model(self, *args, **kwargs) -> List[SamplerOutput]:
|
||||
all_outputs = self._run_workers("execute_model",
|
||||
driver_args=args,
|
||||
driver_kwargs=kwargs)
|
||||
def execute_model(
|
||||
self,
|
||||
execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
|
||||
if self.parallel_worker_tasks is None:
|
||||
self.parallel_worker_tasks = self._run_workers(
|
||||
"start_worker_execution_loop",
|
||||
async_run_remote_workers_only=True,
|
||||
**self.extra_execute_model_run_workers_kwargs)
|
||||
|
||||
# Only the driver worker returns the sampling results.
|
||||
return all_outputs[0]
|
||||
return self._driver_execute_model(execute_model_req)
|
||||
|
||||
def stop_remote_worker_execution_loop(self) -> None:
|
||||
if self.parallel_worker_tasks is None:
|
||||
return
|
||||
|
||||
self._driver_execute_model()
|
||||
parallel_worker_tasks = self.parallel_worker_tasks
|
||||
self.parallel_worker_tasks = None
|
||||
# Ensure that workers exit model loop cleanly
|
||||
# (this will raise otherwise)
|
||||
self._wait_for_tasks_completion(parallel_worker_tasks)
|
||||
|
||||
def add_lora(self, lora_request: LoRARequest) -> bool:
|
||||
assert lora_request.lora_int_id > 0, "lora_id must be greater than 0."
|
||||
@@ -88,39 +114,84 @@ class DistributedGPUExecutor(GPUExecutor):
|
||||
pattern=pattern,
|
||||
max_size=max_size)
|
||||
|
||||
@abstractmethod
|
||||
def _driver_execute_model(
|
||||
self,
|
||||
execute_model_req: Optional[ExecuteModelRequest] = None
|
||||
) -> List[SamplerOutput]:
|
||||
"""Run execute_model in the driver worker.
|
||||
|
||||
Passing None will cause the driver to stop the model execution
|
||||
loop running in each of the remote workers.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def _run_workers(
|
||||
self,
|
||||
method: str,
|
||||
*args,
|
||||
driver_args: Optional[Tuple[Any, ...]] = None,
|
||||
driver_kwargs: Optional[Dict[str, Any]] = None,
|
||||
async_run_remote_workers_only: bool = False,
|
||||
max_concurrent_workers: Optional[int] = None,
|
||||
**kwargs,
|
||||
) -> Any:
|
||||
"""Runs the given method on all workers."""
|
||||
"""Runs the given method on all workers.
|
||||
|
||||
Args:
|
||||
async_run_remote_workers_only: If True the method will be run only
|
||||
in the remote workers, not the driver worker. It will also be
|
||||
run asynchronously and return a list of futures rather than
|
||||
blocking on the results.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None:
|
||||
"""Wait for futures returned from _run_workers() with
|
||||
async_run_remote_workers_only to complete."""
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class DistributedGPUExecutorAsync(DistributedGPUExecutor, ExecutorAsyncBase):
|
||||
|
||||
@abstractmethod
|
||||
async def _run_workers_async(
|
||||
self,
|
||||
method: str,
|
||||
*args,
|
||||
driver_args: Optional[Tuple[Any, ...]] = None,
|
||||
driver_kwargs: Optional[Dict[str, Any]] = None,
|
||||
**kwargs,
|
||||
) -> Any:
|
||||
"""Runs the given method on all workers."""
|
||||
raise NotImplementedError
|
||||
|
||||
async def execute_model_async(self, *args,
|
||||
**kwargs) -> List[SamplerOutput]:
|
||||
all_outputs = await self._run_workers_async("execute_model",
|
||||
driver_args=args,
|
||||
driver_kwargs=kwargs)
|
||||
async def execute_model_async(
|
||||
self,
|
||||
execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
|
||||
if self.parallel_worker_tasks is None:
|
||||
# Start model execution loop running in the parallel workers
|
||||
self.parallel_worker_tasks = asyncio.create_task(
|
||||
self._start_worker_execution_loop())
|
||||
|
||||
# Only the driver worker returns the sampling results.
|
||||
return all_outputs[0]
|
||||
return await self._driver_execute_model_async(execute_model_req)
|
||||
|
||||
async def stop_remote_worker_execution_loop_async(self) -> None:
|
||||
if self.parallel_worker_tasks is None:
|
||||
return
|
||||
|
||||
await self._driver_execute_model_async()
|
||||
parallel_worker_tasks = self.parallel_worker_tasks
|
||||
self.parallel_worker_tasks = None
|
||||
# Ensure that workers exit model loop cleanly
|
||||
# (this will raise otherwise)
|
||||
await parallel_worker_tasks
|
||||
|
||||
@abstractmethod
|
||||
async def _driver_execute_model_async(
|
||||
self,
|
||||
execute_model_req: Optional[ExecuteModelRequest] = None
|
||||
) -> List[SamplerOutput]:
|
||||
"""Execute the model asynchronously in the driver worker.
|
||||
|
||||
Passing None will cause the driver to stop the model execution
|
||||
loop running in each of the remote workers.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
async def _start_worker_execution_loop(self):
|
||||
"""Run execution loop on all workers. It guarantees all workers run
|
||||
the loop or None of them is running the loop. Loop can be stopped by
|
||||
`stop_remote_worker_execution_loop`.
|
||||
The API is idempotent (guarantee only 1 loop run at any moment)."""
|
||||
raise NotImplementedError
|
||||
|
||||
Reference in New Issue
Block a user