File size: 3,039 Bytes
bcdf9fa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Dict, Optional
import ray
from verl.single_controller.base.megatron.worker import DistGlobalInfo, DistRankInfo
from verl.single_controller.base.megatron.worker_group import MegatronWorkerGroup
from .base import RayClassWithInitArgs, RayResourcePool, RayWorkerGroup
# NOTE(sgm): for open-source megatron-core
class NVMegatronRayWorkerGroup(RayWorkerGroup, MegatronWorkerGroup):
"""
MegatronWorkerGroup will query each worker of its megatron rank info and store it inside the WorkerGroup
so that the dispatcher can use it to dispatch data.
"""
def __init__(self, resource_pool: RayResourcePool, ray_cls_with_init: RayClassWithInitArgs, **kwargs):
super().__init__(resource_pool=resource_pool, ray_cls_with_init=ray_cls_with_init, **kwargs)
self._megatron_rank_info: DistRankInfo = self.execute_all_sync(method_name="get_megatron_rank_info")
self._megatron_global_info: DistGlobalInfo = ray.get(self.execute_rank_zero_async(method_name="get_megatron_global_info"))
class MegatronRayWorkerGroup(RayWorkerGroup, MegatronWorkerGroup):
"""
MegatronWorkerGroup will query each worker of its megatron rank info and store it inside the WorkerGroup
so that the dispatcher can use it to dispatch data.
"""
def __init__(
self,
resource_pool: RayResourcePool,
ray_cls_with_init: RayClassWithInitArgs,
default_megatron_kwargs: Dict = None,
**kwargs,
):
super().__init__(
resource_pool=resource_pool,
ray_cls_with_init=ray_cls_with_init,
default_megatron_kwargs=default_megatron_kwargs,
**kwargs,
)
self.init_megatron(default_megatron_kwargs=default_megatron_kwargs)
self._megatron_rank_info: DistRankInfo = self.execute_all_sync(method_name="get_megatron_rank_info")
self._megatron_global_info: DistGlobalInfo = ray.get(self.execute_rank_zero_async(method_name="get_megatron_global_info"))
def init_megatron(self, default_megatron_kwargs: Optional[Dict] = None):
# after super, we will call init of each worker
if not self._is_init_with_detached_workers:
# only init_megatron if the WorkerGroup is created from scratch
self.execute_all_sync(method_name="init_megatron", default_megatron_kwargs=default_megatron_kwargs)
|