Skip to content

GitLab

"tests/test_zero/test_legacy/test_shard_param.py" did not exist on "7aef75ca42537277ddf1c32a7ebe3b1c11897cf4"
Commit da3f0934 authored by zhuwenwen's avatar zhuwenwen
Browse files

delete unused files

parent c4dd1fd4
Show whitespace changes
Inline Side-by-side
Showing with 0 additions and 1094 deletions
colossalai/context/process_group_initializer/initializer_1d.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.global_variables import tensor_parallel_env as env
from colossalai.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
@DIST_GROUP_INITIALIZER.register_module
class Initializer_1D(ProcessGroupInitializer):
'''A ProcessGroupInitializer for 1d tensor parallelism.
'''
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.num_group = self.world_size // self.tensor_parallel_size
def init_dist_group(self):
"""Initialize 1D tensor parallel groups, and assign local_ranks and groups to each gpu.
: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.PARALLEL_1D
env.parallel_input_1d = False
for i in range(self.num_group):
ranks = [i * self.tensor_parallel_size + j for j in range(self.tensor_parallel_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
colossalai/context/process_group_initializer/initializer_2d.py deleted 100644 → 0
View file @ c4dd1fd4
import math
import torch.distributed as dist
from colossalai.registry import DIST_GROUP_INITIALIZER
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
from colossalai.global_variables import tensor_parallel_env as env
def _check_summa_env_var(summa_dim):
# check environment variable for SUMMA
env_summa_dim = env.summa_dim
if env_summa_dim:
assert int(env_summa_dim) == summa_dim, \
'SUMMA_DIM has been set in the current environment and ' \
'does not match with the value passed to this initialized'
else:
env.summa_dim = summa_dim
class Initializer_2D_Row(ProcessGroupInitializer):
"""2d tensor parallel initialization among rows.
:param num_group: The number of all tensor groups
:param summa_dim: The dimension of SUMMA
:param args: Args used to initialize base class
:param kwargs: Kwargs used to initialize base class
:type num_group: int
:type summa_dim: int
"""
def __init__(self, num_group, summa_dim, *args, **kwargs):
super(Initializer_2D_Row, self).__init__(*args, **kwargs)
self.num_group = num_group
self.summa_dim = summa_dim
def init_dist_group(self):
"""Initialize 2D tensor row parallel groups, and assign local_ranks and groups to each gpu.
:return: 2D tensor row parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2D_ROW
for i in range(self.num_group):
for j in range(self.summa_dim):
ranks = [i * self.tensor_parallel_size + j * self.summa_dim + k
for k in range(self.summa_dim)]
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
class Initializer_2D_Col(ProcessGroupInitializer):
"""2d tensor parallel initialization among cols.
:param num_group: The number of all tensor groups
:param summa_dim: The dimension of SUMMA
:param args: Args used to initialize base class
:param kwargs: Kwargs used to initialize base class
:type num_group: int
:type summa_dim: int
"""
def __init__(self, num_group, summa_dim, *args, **kwargs):
super(Initializer_2D_Col, self).__init__(*args, **kwargs)
self.num_group = num_group
self.summa_dim = summa_dim
def init_dist_group(self):
"""Initialize 2D tensor row parallel groups, and assign local_ranks and groups to each gpu.
:return: 2D tensor col parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2D_COL
for i in range(self.num_group):
for j in range(self.summa_dim):
ranks = [i * self.tensor_parallel_size + j + k * self.summa_dim
for k in range(self.summa_dim)]
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_2D(ProcessGroupInitializer):
"""
Serve as the single entry point to 2D parallel initialization.
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.num_group = self.world_size // self.tensor_parallel_size
self.summa_dim = int(math.sqrt(self.tensor_parallel_size))
assert self.tensor_parallel_size == self.summa_dim ** 2, \
"2D summa dim should equal to tensor parallel size ^ 0.5"
_check_summa_env_var(self.summa_dim)
self.col_initializer = Initializer_2D_Col(self.num_group, self.summa_dim, *args, **kwargs)
self.row_initializer = Initializer_2D_Row(self.num_group, self.summa_dim, *args, **kwargs)
def init_dist_group(self):
"""Initialize 2D tensor row and col parallel groups, and assign local_ranks and groups to each gpu.
:return: 2D tensor parallelism's information
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
parallel_setting = [self.row_initializer.init_dist_group(), self.col_initializer.init_dist_group()]
return parallel_setting
colossalai/context/process_group_initializer/initializer_2p5d.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import math
import torch.distributed as dist
from colossalai.context import Config
from colossalai.global_variables import tensor_parallel_env as env
from colossalai.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
def _check_tesseract_env_var(tesseract_dim: int,
tesseract_dep: int):
# check global variable for TESSERACT
env_tesseract_dim = env.tesseract_dim
env_tesseract_dep = env.tesseract_dep
if env_tesseract_dim and env_tesseract_dep:
assert int(env_tesseract_dim) == tesseract_dim, \
'TESSERACT_DIM has been set in the current environment and ' \
'does not match with the value passed to this initialized'
assert int(env_tesseract_dep) == tesseract_dep, \
'TESSERACT_DEP has been set in the current environment and ' \
'does not match with the value passed to this initialized'
else:
env.tesseract_dim = tesseract_dim
env.tesseract_dep = tesseract_dep
# i row j col k dep
class Initializer_2p5D_ROW(ProcessGroupInitializer):
"""2p5d tensor parallel initialization among rows.
:param tesseract_dim: The dimension of tesseract
:param tesseract_dep: The dimension of depth
:param args: Args used to initialize base class
:type tesseract_dim: int
:type tesseract_dep: int
"""
def __init__(self,
tesseract_dim: int,
tesseract_dep: int,
*args):
super(Initializer_2p5D_ROW, self).__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.tesseract_dep = tesseract_dep
self.tesseract_dim = tesseract_dim
assert self.tensor_parallel_size == self.tesseract_dim ** 2 * self.tesseract_dep, \
"Tensor parallel size should be depth * dim ** 2 in 2.5D parallel"
def init_dist_group(self):
"""Initialize 2p5D tensor row parallel groups, and assign local_ranks and groups to each gpu.
:return: 2p5D tensor row parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2P5D_ROW
for h in range(self.num_group):
for j in range(self.tesseract_dim):
for k in range(self.tesseract_dep):
ranks = [h * self.tensor_parallel_size + i + self.tesseract_dim * (
j + self.tesseract_dim * k) for i in range(self.tesseract_dim)]
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
class Initializer_2p5D_Col(ProcessGroupInitializer):
"""2p5d tensor parallel initialization among cols.
:param tesseract_dim: The dimension of tesseract
:param tesseract_dep: The dimension of depth
:param args: Args used to initialize base class
:type tesseract_dim: int
:type tesseract_dep: int
"""
def __init__(self,
tesseract_dim: int,
tesseract_dep: int,
*args):
super(Initializer_2p5D_Col, self).__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.tesseract_dep = tesseract_dep
self.tesseract_dim = tesseract_dim
assert self.tensor_parallel_size == self.tesseract_dim ** 2 * self.tesseract_dep, \
"Tensor parallel size should be depth * dim ** 2 in 2.5D parallel"
def init_dist_group(self):
"""Initialize 2p5D tensor col parallel groups, and assign local_ranks and groups to each gpu.
:return: 2p5D tensor col parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2P5D_COL
for h in range(self.num_group):
for i in range(self.tesseract_dim):
for k in range(self.tesseract_dep):
ranks = [h * self.tensor_parallel_size + i + self.tesseract_dim * (
j + self.tesseract_dim * k) for j in range(self.tesseract_dim)]
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
class Initializer_2p5D_Dep(ProcessGroupInitializer):
"""2p5D tensor parallel initialization among depths.
:param tesseract_dim: The dimension of tesseract
:param tesseract_dep: The dimension of depth
:param args: Args used to initialize base class
:type tesseract_dim: int
:type tesseract_dep: int
"""
def __init__(self,
tesseract_dim: int,
tesseract_dep: int,
*args):
super(Initializer_2p5D_Dep, self).__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.tesseract_dep = tesseract_dep
self.tesseract_dim = tesseract_dim
assert self.tensor_parallel_size == self.tesseract_dim ** 2 * self.tesseract_dep, \
"Tensor parallel size should be depth * dim ** 2 in 2.5D parallel"
def init_dist_group(self):
"""Initialize 2p5D tensor depth parallel groups, and assign local_ranks and groups to each gpu.
:return: 2p5D tensor depth parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2P5D_DEP
for h in range(self.num_group):
for i in range(self.tesseract_dim):
for j in range(self.tesseract_dim):
ranks = [h * self.tensor_parallel_size + i + self.tesseract_dim * (
j + self.tesseract_dim * k) for k in range(self.tesseract_dep)]
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
# i row j col k dep
class Initializer_2p5D_XZ(ProcessGroupInitializer):
"""2p5d tensor parallel initialization among cols times dep.
:param tesseract_dim: The dimension of tesseract
:param tesseract_dep: The dimension of depth
:param args: Args used to initialize base class
:type tesseract_dim: int
:type tesseract_dep: int
"""
def __init__(self,
tesseract_dim: int,
tesseract_dep: int,
*args):
super(Initializer_2p5D_XZ, self).__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.tesseract_dep = tesseract_dep
self.tesseract_dim = tesseract_dim
assert self.tensor_parallel_size == self.tesseract_dim ** 2 * self.tesseract_dep, \
"Tensor parallel size should be depth * dim ** 2 in 2.5D parallel"
def init_dist_group(self):
"""Initialize 2p5D tensor colXdepth parallel groups, and assign local_ranks and groups to each gpu.
:return: 2p5D tensor colXdepth parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_2P5D_XZ
for h in range(self.num_group):
for i in range(self.tesseract_dim):
ranks = [h * self.tensor_parallel_size + i + self.tesseract_dim * (
j + self.tesseract_dim * k) for k in range(self.tesseract_dep) for j in
range(self.tesseract_dim)]
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_2p5D(ProcessGroupInitializer):
"""
Serve as the single entry point to Tesseract parallel initialization.
:param rank: The rank of current process
:param world_size: Size of whole communication world
:param config: Running configuration
:param data_parallel_size: Size of data parallel
:param pipeline_parallel_size: Size of pipeline parallel
:param tensor_parallel_size: Size of tensor parallel
:param depth: The depth of 2p5d parallel
:type rank: int
:type world_size: int
:type config: Config
:type data_parallel_size: int
:type pipeline_parallel_size: int
:type tensor_parallel_size: int
:type depth: int
"""
def __init__(self,
rank: int,
world_size: int,
config: Config,
data_parallel_size: int,
pipeline_parallel_size: int,
tensor_parallel_size: int,
depth: int
):
args = (rank, world_size, config, data_parallel_size, pipeline_parallel_size, tensor_parallel_size)
super().__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.tesseract_dim = int(math.sqrt(self.tensor_parallel_size / depth))
self.tesseract_dep = depth
assert self.tensor_parallel_size == self.tesseract_dim ** 2 * self.tesseract_dep, \
"2.5D tesseract dim should equal to (tensor parallel size / tesseract dep) ^ 0.5"
_check_tesseract_env_var(self.tesseract_dim, self.tesseract_dep)
self.col_initializer = Initializer_2p5D_Col(self.tesseract_dim, self.tesseract_dep, *args)
self.row_initializer = Initializer_2p5D_ROW(self.tesseract_dim, self.tesseract_dep, *args)
self.dep_initializer = Initializer_2p5D_Dep(self.tesseract_dim, self.tesseract_dep, *args)
self.xz_initializer = Initializer_2p5D_XZ(self.tesseract_dim, self.tesseract_dep, *args)
def init_dist_group(self):
"""Initialize 2p5D tensor row, col, depth, and colXdepth parallel groups, and assign local_ranks and groups to each gpu.
:return: Whole 2p5D tensor parallelism's information
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
parallel_setting = [self.col_initializer.init_dist_group(), self.row_initializer.init_dist_group(),
self.dep_initializer.init_dist_group(), self.xz_initializer.init_dist_group()]
return parallel_setting
colossalai/context/process_group_initializer/initializer_3d.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import math
import torch.distributed as dist
from colossalai.global_variables import tensor_parallel_env as env
from colossalai.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
def _check_depth_env_var(depth):
# check global variable
env_depth = env.depth_3d
if env_depth:
assert int(env_depth) == depth, \
'DEPTH_3D has been set in the current environment and ' \
'does not match with the value passed to this initialized'
else:
env.depth_3d = depth
class Initializer_3D_Input(ProcessGroupInitializer):
"""3D tensor parallel initialization among input.
:param num_group: The number of all tensor groups
:param depth: Depth of 3D parallelism
:param args: Args used in base class
:type num_group: int
:type depth: int
"""
def __init__(self, num_group: int, depth: int, *args):
super().__init__(*args)
self.num_group = num_group
self.depth = depth
def init_dist_group(self):
"""Initialize 3D tensor parallel groups among input, and assign local_ranks and groups to each gpu.
:return: 3D tensor parallelism's information among input
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_3D_INPUT
env.input_group_3d = mode
for h in range(self.num_group):
for i in range(self.depth):
for k in range(self.depth):
ranks = [h * self.depth**3 + i + self.depth * (j + self.depth * k) for j in range(self.depth)]
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
class Initializer_3D_Weight(ProcessGroupInitializer):
"""3D tensor parallel initialization among weight.
:param num_group: The number of all tensor groups
:param depth: Depth of 3D parallelism
:param args: Args used in base class
:type num_group: int
:type depth: int
"""
def __init__(self, num_group: int, depth: int, *args):
super().__init__(*args)
self.num_group = num_group
self.depth = depth
def init_dist_group(self):
"""Initialize 3D tensor parallel groups among weight, and assign local_ranks and groups to each gpu.
:return: 3D tensor parallelism's information among weight
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_3D_WEIGHT
env.weight_group_3d = mode
for h in range(self.num_group):
for k in range(self.depth):
for j in range(self.depth):
ranks = [h * self.depth**3 + i + self.depth * (j + self.depth * k) for i in range(self.depth)]
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
class Initializer_3D_Output(ProcessGroupInitializer):
"""3D tensor parallel initialization among output.
:param num_group: The number of all tensor groups
:param depth: Depth of 3D parallelism
:param args: Args used in base class
:type num_group: int
:type depth: int
"""
def __init__(self, num_group: int, depth: int, *args):
super().__init__(*args)
self.num_group = num_group
self.depth = depth
def init_dist_group(self):
"""Initialize 3D tensor parallel groups among output, and assign local_ranks and groups to each gpu.
:return: 3D tensor parallelism's information among output
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_3D_OUTPUT
env.output_group_3d = mode
for h in range(self.num_group):
for i in range(self.depth):
for j in range(self.depth):
ranks = [h * self.depth**3 + i + self.depth * (j + self.depth * k) for k in range(self.depth)]
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_3D(ProcessGroupInitializer):
"""Serve as the single entry point to 3D parallel initialization.
:param args: Args used to initialize ProcessGroupInitializer
"""
def __init__(self, *args):
super().__init__(*args)
self.num_group = self.world_size // self.tensor_parallel_size
self.depth = round(math.pow(self.tensor_parallel_size, 1 / 3))
assert self.tensor_parallel_size == self.depth ** 3, \
f'3D depth ({self.depth}) if not cube root of tensor parallel size ({self.tensor_parallel_size})'
_check_depth_env_var(self.depth)
self.input_initializer = Initializer_3D_Input(self.num_group, self.depth, *args)
self.weight_initializer = Initializer_3D_Weight(self.num_group, self.depth, *args)
self.output_initializer = Initializer_3D_Output(self.num_group, self.depth, *args)
def init_dist_group(self):
"""Initialize 3D tensor parallel groups, and assign local_ranks and groups to each gpu.
:return: 3D tensor parallelism's information
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
parallel_setting = [self.input_initializer.init_dist_group(), self.weight_initializer.init_dist_group(),
self.output_initializer.init_dist_group()]
return parallel_setting
colossalai/context/process_group_initializer/initializer_data.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from torch import distributed as dist
from colossalai.registry import DIST_GROUP_INITIALIZER
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Data(ProcessGroupInitializer):
"""A ProcessGroupInitializer for data parallelism.
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.num_data_parallel_group = self.world_size // self.data_parallel_size
def init_dist_group(self):
"""Initialize data parallel groups, and assign local_ranks and groups to each gpu.
:return: Data parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.DATA
for i in range(self.num_data_parallel_group):
ranks = [i + j * self.num_data_parallel_group for j in range(self.data_parallel_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
colossalai/context/process_group_initializer/initializer_model.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.context import Config
from colossalai.registry import DIST_GROUP_INITIALIZER
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Model(ProcessGroupInitializer):
"""A ProcessGroupInitializer for model parallelism (model parallel group contains pipeline and tensor parallel
groups).
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model_parallel_size = self.tensor_parallel_size * self.pipeline_parallel_size
self.num_group = self.world_size // self.model_parallel_size
def init_dist_group(self):
"""Initialize model parallel groups, and assign local_ranks and groups to each gpu.
: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.MODEL
for i in range(self.num_group):
ranks = [i * self.model_parallel_size + j for j in range(self.model_parallel_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
colossalai/context/process_group_initializer/initializer_moe.py deleted 100644 → 0
View file @ c4dd1fd4
import torch.distributed as dist
from colossalai.registry import DIST_GROUP_INITIALIZER
from colossalai.global_variables import moe_env
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Moemodel(ProcessGroupInitializer):
"""Model parallel initialization for MoE system.
:param moe_moel: Size of moe model parallel
:param moe_data: Size of moe data parallel
:param args: Args used in base class
:param kwargs: Kwargs used in base class
:type moe_model: int
:type moe_data: int
"""
def __init__(self, moe_model, moe_data, *args, **kwargs):
super().__init__(*args, **kwargs)
self.moe_model = moe_model
self.moe_data = moe_data
def init_dist_group(self):
"""Initialize model parallel groups in moe parallel environment,
and assign local_ranks and groups to each gpu.
:return: MoE model parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.MOE_MODEL
for i in range(self.moe_data):
ranks = [i * self.moe_model + j for j in range(self.moe_model)]
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_Moedata(ProcessGroupInitializer):
"""Data parallel initialization for MoE system.
:param moe_moel: Size of moe model parallel
:param moe_data: Size of moe data parallel
:param args: Args used in base class
:param kwargs: Kwargs used in base class
:type moe_model: int
:type moe_data: int
"""
def __init__(self, moe_model, moe_data, *args, **kwargs):
super().__init__(*args, **kwargs)
self.moe_model = moe_model
self.moe_data = moe_data
def init_dist_group(self):
"""Initialize data parallel groups in moe parallel environment,
and assign local_ranks and groups to each gpu.
:return: MoE data parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.MOE_DATA
for i in range(self.moe_model):
ranks = [i + j * self.moe_model for j in range(self.moe_data)]
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_Moe(ProcessGroupInitializer):
"""Serves as the single entry point to MoE parallel initialization.
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.moe_model = moe_env.model_parallel_size
self.moe_data = moe_env.data_parallel_size
self.model_initializer = Initializer_Moemodel(
self.moe_model, self.moe_data, *args, **kwargs)
self.data_initializer = Initializer_Moedata(
self.moe_model, self.moe_data, *args, **kwargs)
def init_dist_group(self):
"""Initializes MoE parallel communication groups.
:return: MoE parallelism's information
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
parallel_setting = [self.model_initializer.init_dist_group(),
self.data_initializer.init_dist_group()]
return parallel_setting
colossalai/context/process_group_initializer/initializer_pipeline.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from torch import distributed as dist
from colossalai.registry import DIST_GROUP_INITIALIZER
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Pipeline(ProcessGroupInitializer):
"""A ProcessGroupInitializer for pipeline parallelism.
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.data_group_size = self.world_size // self.data_parallel_size
self.pipeline_stage_size = self.data_group_size // self.pipeline_parallel_size
def init_dist_group(self):
"""Initialize pipeline parallel groups, and assign local_ranks and groups to each gpu.
:return: Pipeline parallelism's information
:rtype: list of Tuples (local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
dist_settings = list()
for i in range(self.data_parallel_size):
for j in range(self.pipeline_stage_size):
pipe_ranks = list(
range(i * self.data_group_size + j,
(i + 1) * self.data_group_size,
self.pipeline_stage_size))
pipe_group_size = len(pipe_ranks)
pipe_group = dist.new_group(pipe_ranks)
if self.rank in pipe_ranks:
local_rank = pipe_ranks.index(self.rank)
group_world_size = pipe_group_size
process_group = pipe_group
ranks_in_group = pipe_ranks
dist_settings.append(
tuple((local_rank, group_world_size,
process_group, ranks_in_group,
ParallelMode.PIPELINE)))
return dist_settings
colossalai/context/process_group_initializer/initializer_sequence.py deleted 100644 → 0
View file @ c4dd1fd4
#!/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
colossalai/context/process_group_initializer/initializer_tensor.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.registry import DIST_GROUP_INITIALIZER
from .process_group_initializer import ProcessGroupInitializer
from ..parallel_mode import ParallelMode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Tensor(ProcessGroupInitializer):
"""A ProcessGroupInitializer for tensor parallelism.
:param args: Args used to initialize ProcessGroupInitializer
:param kwargs: Kwargs used to initialize ProcessGroupInitializer
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.num_tensor_parallel_group = self.world_size // self.tensor_parallel_size
def init_dist_group(self):
"""Initialize tensor parallel groups, and assign local_ranks and groups to each gpu.
:return: Tensor parallelism's information
:rtype: Tuple(local_rank, group_world_size, process_group, ranks_in_group, mode)
"""
local_rank = None
ranks_in_group = None
process_group = None
group_world_size = None
mode = ParallelMode.TENSOR
for i in range(self.num_tensor_parallel_group):
ranks = [i * self.tensor_parallel_size + j for j in range(self.tensor_parallel_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
colossalai/context/process_group_initializer/process_group_initializer.py deleted 100644 → 0
View file @ c4dd1fd4
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from abc import ABC, abstractmethod
from colossalai.context import Config
class ProcessGroupInitializer(ABC):
"""An object, knowing the parallelism configuration, that initializes parallel groups.
:param rank: The rank of current process
:param world_size: Size of whole communication world
:param config: Running configuration
:param data_parallel_size: Size of data parallel
:param pipeline_parallel_size: Size of pipeline parallel
:param tensor_parallel_size: Size of tensor parallel
:type rank: int
:type world_size: int
:type config: Config
:type data_parallel_size: int
:type pipeline_parallel_size: int
:type tensor_parallel_size: int
"""
def __init__(self,
rank: int,
world_size: int,
config: Config,
data_parallel_size: int,
pipeline_parallel_size: int,
tensor_parallel_size: int
):
self.rank = rank
self.world_size = world_size
self.data_parallel_size = data_parallel_size
self.config = config
self.pipeline_parallel_size = pipeline_parallel_size
self.tensor_parallel_size = tensor_parallel_size
super().__init__()
@abstractmethod
def init_dist_group(self):
pass
colossalai/context/random/__init__.py deleted 100644 → 0
View file @ c4dd1fd4
from ._helper import (seed, set_mode, with_seed, add_seed,
get_seeds, get_states, get_current_mode,
set_seed_states, sync_states, moe_set_seed)
__all__ = [
'seed', 'set_mode', 'with_seed', 'add_seed', 'get_seeds',
'get_states', 'get_current_mode', 'set_seed_states', 'sync_states',
'moe_set_seed'
]
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment