[Core] RayWorkerVllm --> WorkerWrapper to reduce duplication (#4024)
[Core] replace narrow-usage RayWorkerVllm to general WorkerWrapper to reduce code duplication (#4024)
This commit is contained in:
@@ -1,55 +1,28 @@
|
||||
import pickle
|
||||
from typing import Callable, List, Optional, Tuple
|
||||
from typing import List, Optional, Tuple
|
||||
|
||||
from vllm.config import ParallelConfig
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils import get_ip, is_hip, set_cuda_visible_devices
|
||||
from vllm.worker.worker import Worker
|
||||
from vllm.utils import get_ip, is_hip
|
||||
from vllm.worker.worker_base import WorkerWrapperBase
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
try:
|
||||
import ray
|
||||
|
||||
class RayWorkerVllm:
|
||||
class RayWorkerWrapper(WorkerWrapperBase):
|
||||
"""Ray wrapper for vllm.worker.Worker, allowing Worker to be
|
||||
lazliy initialized after Ray sets CUDA_VISIBLE_DEVICES."""
|
||||
|
||||
def __init__(self, init_cached_hf_modules=False) -> None:
|
||||
if init_cached_hf_modules:
|
||||
from transformers.dynamic_module_utils import init_hf_modules
|
||||
init_hf_modules()
|
||||
self._worker: Optional[Worker] = None
|
||||
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 init_worker(self, worker_init_fn: Callable[[], Worker]):
|
||||
self._worker = worker_init_fn()
|
||||
|
||||
@property
|
||||
def worker(self) -> Worker:
|
||||
assert self._worker is not None
|
||||
return self._worker
|
||||
|
||||
def __getattr__(self, name):
|
||||
return getattr(self.worker, name)
|
||||
|
||||
def execute_method(self, method, *args, **kwargs):
|
||||
try:
|
||||
executor = getattr(self, method)
|
||||
return executor(*args, **kwargs)
|
||||
except Exception as e:
|
||||
# exceptions in ray worker may cause deadlock
|
||||
# see https://github.com/vllm-project/vllm/issues/3455
|
||||
# print the error and inform the user to solve the error
|
||||
msg = (f"Error executing method {method}. "
|
||||
"This might cause deadlock in distributed execution.")
|
||||
logger.exception(msg)
|
||||
raise e
|
||||
|
||||
def get_node_ip(self) -> str:
|
||||
return get_ip()
|
||||
|
||||
@@ -58,9 +31,6 @@ try:
|
||||
gpu_ids = ray.get_gpu_ids()
|
||||
return node_id, gpu_ids
|
||||
|
||||
def set_cuda_visible_devices(self, device_ids) -> None:
|
||||
set_cuda_visible_devices(device_ids)
|
||||
|
||||
def execute_model_compiled_dag_remote(self, ignored):
|
||||
"""Used only when compiled DAG is enabled."""
|
||||
import torch
|
||||
@@ -77,7 +47,7 @@ except ImportError as e:
|
||||
"For distributed inference, please install Ray with "
|
||||
"`pip install ray`.")
|
||||
ray = None # type: ignore
|
||||
RayWorkerVllm = None # type: ignore
|
||||
RayWorkerWrapper = None # type: ignore
|
||||
|
||||
|
||||
def initialize_ray_cluster(
|
||||
|
||||
Reference in New Issue
Block a user