ColossalAI/colossalai/context/process_group_initializer/initializer_sequence.py

102 lines
4.0 KiB
Python
Raw Normal View History

2021-10-28 16:21:23 +00:00
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
2022-01-20 05:44:51 +00:00
import torch.distributed as dist
2021-10-28 16:21:23 +00:00
from colossalai.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
2021-10-28 16:21:23 +00:00
from .initializer_tensor import Initializer_Tensor
from .process_group_initializer import ProcessGroupInitializer
2022-01-20 05:44:51 +00:00
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Sequence_DP(ProcessGroupInitializer):
2022-01-21 02:44:30 +00:00
"""A ProcessGroupInitializer for sequence parallelism all-reduce.
2022-01-20 05:44:51 +00:00
2022-01-21 02:44:30 +00:00
In Sequence Parallelism, each GPU holds the full copy of model weights,
2022-01-20 05:44:51 +00:00
thus, gradient all-reduce occurs across all processes in the same pipeline stage
2022-03-25 05:02:39 +00:00
Args:
rank (int): The rank of current process
world_size (int): Size of whole communication world
config (Config): Running configuration
data_parallel_size (int): Size of data parallel
pipeline_parallel_size (int): Size of pipeline parallel
tensor_parallel_size (int): Size of tensor parallel
2022-01-21 02:44:30 +00:00
"""
2022-01-20 05:44:51 +00:00
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):
2022-01-21 02:44:30 +00:00
"""Initialize Sequence Parallel process groups used for gradient all-reduce.
2022-03-25 05:02:39 +00:00
Returns:
Tuple: A tuple (local_rank, group_world_size, process_group, ranks_in_group, mode).
2022-01-21 02:44:30 +00:00
"""
2022-01-20 05:44:51 +00:00
local_rank = None
ranks_in_group = None
process_group = None
cpu_group = None
2022-01-20 05:44:51 +00:00
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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
2022-01-20 05:44:51 +00:00
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
2022-01-20 05:44:51 +00:00
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
2022-01-20 05:44:51 +00:00
2021-10-28 16:21:23 +00:00
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Sequence(ProcessGroupInitializer):
2022-01-21 02:44:30 +00:00
"""A ProcessGroupInitializer for sequence parallelism.
2021-10-28 16:21:23 +00:00
2022-03-25 05:02:39 +00:00
Args:
rank (int): The rank of current process.
world_size (int): Size of whole communication world.
config (Config): Running configuration.
data_parallel_size (int): Size of data parallel.
pipeline_parallel_size (int): Size of pipeline parallel.
tensor_parallel_size (int): Size of tensor parallel.
2022-01-21 02:44:30 +00:00
"""
def __init__(self, *args, **kwargs):
2021-10-28 16:21:23 +00:00
super().__init__(*args, **kwargs)
2022-01-20 05:44:51 +00:00
# reuse tensor parallel initializer code
self._sequence_initializer = Initializer_Tensor(*args, **kwargs)
self._sequence_dp_initializer = Initializer_Sequence_DP(*args, **kwargs)
2021-10-28 16:21:23 +00:00
def init_dist_group(self):
2022-01-21 02:44:30 +00:00
"""Initialize Sequence parallel process groups and assign local_ranks and groups to each gpu.
2022-01-20 05:44:51 +00:00
Sequence parallelism requires 2 process groups. The first is for model forward where several processes
2022-03-25 05:02:39 +00:00
exchange partial query, key and value embedding to compute self attention values. The second is for
2022-01-20 05:44:51 +00:00
all-reduce to synchronize the model parameters.
2021-10-28 16:21:23 +00:00
2022-03-25 05:02:39 +00:00
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
A Sequence parallelism's information in list of tuples.
2022-01-21 02:44:30 +00:00
"""
2022-01-20 05:44:51 +00:00
parallel_setting = []
local_rank, group_world_size, process_group, cpu_grop, ranks_in_group, mode = \
self._sequence_initializer.init_dist_group()
2021-10-28 16:21:23 +00:00
# change mode to sequence
mode = ParallelMode.SEQUENCE
2022-01-20 05:44:51 +00:00
parallel_setting.append((local_rank, group_world_size, process_group, cpu_grop, ranks_in_group, mode))
2022-01-20 05:44:51 +00:00
parallel_setting.append(self._sequence_dp_initializer.init_dist_group())
return parallel_setting