|
|
|
#!/usr/bin/env python
|
|
|
|
# -*- encoding: utf-8 -*-
|
|
|
|
import torch.distributed as dist
|
|
|
|
|
|
|
|
from colossalai.registry import DIST_GROUP_INITIALIZER
|
|
|
|
from .initializer_tensor import Initializer_Tensor
|
|
|
|
from .process_group_initializer import ProcessGroupInitializer
|
|
|
|
from ..parallel_mode import ParallelMode
|
|
|
|
|
|
|
|
|
|
|
|
@DIST_GROUP_INITIALIZER.register_module
|
|
|
|
class Initializer_Sequence_DP(ProcessGroupInitializer):
|
|
|
|
"""A ProcessGroupInitializer for sequence parallelism all-reduce.
|
|
|
|
|
|
|
|
In Sequence Parallelism, each GPU holds the full copy of model weights,
|
|
|
|
thus, gradient all-reduce occurs across all processes in the same pipeline stage
|
|
|
|
|
|
|
|
:param args: Args used to initialize ProcessGroupInitializer
|
|
|
|
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
|
|
|
|
"""
|
|
|
|
|
|
|
|
def __init__(self, *args, **kwargs):
|
|
|
|
super().__init__(*args, **kwargs)
|
|
|
|
self.dp_size = self.world_size // self.pipeline_parallel_size
|
|
|
|
self.num_group = self.pipeline_parallel_size
|
|
|
|
|
|
|
|
def init_dist_group(self):
|
|
|
|
"""Initialize Sequence Parallel process groups used for gradient all-reduce.
|
|
|
|
|
|
|
|
:return: (local_rank, group_world_size, process_group, ranks_in_group, mode)
|
|
|
|
:rtype: Tuple
|
|
|
|
"""
|
|
|
|
local_rank = None
|
|
|
|
ranks_in_group = None
|
|
|
|
process_group = None
|
|
|
|
group_world_size = None
|
|
|
|
mode = ParallelMode.SEQUENCE_DP
|
|
|
|
|
|
|
|
for i in range(self.num_group):
|
|
|
|
ranks = [i * self.dp_size + j for j in range(self.dp_size)]
|
|
|
|
group = dist.new_group(ranks)
|
|
|
|
|
|
|
|
if self.rank in ranks:
|
|
|
|
local_rank = ranks.index(self.rank)
|
|
|
|
group_world_size = len(ranks)
|
|
|
|
process_group = group
|
|
|
|
ranks_in_group = ranks
|
|
|
|
return local_rank, group_world_size, process_group, ranks_in_group, mode
|
|
|
|
|
|
|
|
|
|
|
|
@DIST_GROUP_INITIALIZER.register_module
|
|
|
|
class Initializer_Sequence(ProcessGroupInitializer):
|
|
|
|
"""A ProcessGroupInitializer for sequence parallelism.
|
|
|
|
|
|
|
|
:param args: Args used to initialize ProcessGroupInitializer
|
|
|
|
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
|
|
|
|
"""
|
|
|
|
def __init__(self,
|
|
|
|
*args, **kwargs):
|
|
|
|
super().__init__(*args, **kwargs)
|
|
|
|
# reuse tensor parallel initializer code
|
|
|
|
self._sequence_initializer = Initializer_Tensor(*args, **kwargs)
|
|
|
|
self._sequence_dp_initializer = Initializer_Sequence_DP(*args, **kwargs)
|
|
|
|
|
|
|
|
def init_dist_group(self):
|
|
|
|
"""Initialize Sequence parallel process groups and assign local_ranks and groups to each gpu.
|
|
|
|
|
|
|
|
Sequence parallelism requires 2 process groups. The first is for model forward where several processes
|
|
|
|
exchange paritial query, key and value embedding to compute self attention values. The second is for
|
|
|
|
all-reduce to synchronize the model parameters.
|
|
|
|
|
|
|
|
:return: Sequence parallelism's information
|
|
|
|
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
|
|
|
|
"""
|
|
|
|
|
|
|
|
parallel_setting = []
|
|
|
|
|
|
|
|
local_rank, group_world_size, process_group, ranks_in_group, mode = self._sequence_initializer.init_dist_group()
|
|
|
|
# change mode to sequence
|
|
|
|
mode = ParallelMode.SEQUENCE
|
|
|
|
|
|
|
|
parallel_setting.append((local_rank, group_world_size, process_group, ranks_in_group, mode))
|
|
|
|
parallel_setting.append(self._sequence_dp_initializer.init_dist_group())
|
|
|
|
return parallel_setting
|