github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-07 20:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[legacy] clean up legacy code (#4743)

Browse Source 
* [legacy] remove outdated codes of pipeline (#4692)

* [legacy] remove cli of benchmark and update optim (#4690)

* [legacy] remove cli of benchmark and update optim

* [doc] fix cli doc test

* [legacy] fix engine clip grad norm

* [legacy] remove outdated colo tensor (#4694)

* [legacy] remove outdated colo tensor

* [test] fix test import

* [legacy] move outdated zero to legacy (#4696)

* [legacy] clean up utils (#4700)

* [legacy] clean up utils

* [example] update examples

* [legacy] clean up amp

* [legacy] fix amp module

* [legacy] clean up gpc (#4742)

* [legacy] clean up context

* [legacy] clean core, constants and global vars

* [legacy] refactor initialize

* [example] fix examples ci

* [example] fix examples ci

* [legacy] fix tests

* [example] fix gpt example

* [example] fix examples ci

* [devops] fix ci installation

* [example] fix examples ci
This commit is contained in:
Hongxin Liu
2023-09-18 16:31:06 +08:00
committed by GitHub
parent 32e7f99416
commit b5f9e37c70
342 changed files with 2919 additions and 4182 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
colossalai/legacy/context/process_group_initializer/__init__.py Normal file
View File

@@ -0,0 +1,15 @@
from .initializer_1d import Initializer_1D
from .initializer_2d import Initializer_2D
from .initializer_2p5d import Initializer_2p5D
from .initializer_3d import Initializer_3D
from .initializer_data import Initializer_Data
from .initializer_model import Initializer_Model
from .initializer_pipeline import Initializer_Pipeline
from .initializer_sequence import Initializer_Sequence
from .initializer_tensor import Initializer_Tensor
from .process_group_initializer import ProcessGroupInitializer
__all__ = [
'Initializer_Tensor', 'Initializer_Sequence', 'Initializer_Pipeline', 'Initializer_Data', 'Initializer_2p5D',
'Initializer_2D', 'Initializer_3D', 'Initializer_1D', 'ProcessGroupInitializer', 'Initializer_Model'
]

57
colossalai/legacy/context/process_group_initializer/initializer_1d.py Normal file
View File

@@ -0,0 +1,57 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.legacy.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.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
1D tensor parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode

155
colossalai/legacy/context/process_group_initializer/initializer_2d.py Normal file
View File

@@ -0,0 +1,155 @@
import math
import torch.distributed as dist
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
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.
Args:
num_group (int): The number of all tensor groups.
summa_dim (int): The dimension of SUMMA.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2D tensor row parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_2D_Col(ProcessGroupInitializer):
"""2d tensor parallel initialization among cols.
Args:
num_group (int): The number of all tensor groups.
summa_dim (int): The dimension of SUMMA.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2D tensor col parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_2D(ProcessGroupInitializer):
"""
Serve as the single entry point to 2D parallel initialization.
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.
"""
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.
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
2D tensor parallelism's information in a list of tuples.
"""
parallel_setting = [self.row_initializer.init_dist_group(), self.col_initializer.init_dist_group()]
return parallel_setting

298
colossalai/legacy/context/process_group_initializer/initializer_2p5d.py Normal file
View File

@@ -0,0 +1,298 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import math
import torch.distributed as dist
from colossalai.context import Config
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.legacy.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):
"""2.5d tensor parallel initialization among rows.
Args:
tesseract_dim (int): The dimension of tesseract.
tesseract_dep (int): The dimension of depth.
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.
"""
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 2.5D tensor row parallel groups, and assign local_ranks and groups to each gpu.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2.5D tensor row parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_2p5D_Col(ProcessGroupInitializer):
"""2.5d tensor parallel initialization among cols.
Args:
tesseract_dim (int): The dimension of tesseract.
tesseract_dep (int): The dimension of depth.
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.
"""
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
def init_dist_group(self):
"""Initialize 2.5D tensor col parallel groups, and assign local_ranks and groups to each gpu.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2.5D tensor col parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_2p5D_Dep(ProcessGroupInitializer):
"""2.5D tensor parallel initialization among depths.
Args:
tesseract_dim (int): The dimension of tesseract.
tesseract_dep (int): The dimension of depth.
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.
"""
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
def init_dist_group(self):
"""Initialize 2.5D tensor depth parallel groups, and assign local_ranks and groups to each gpu.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2.5D tensor depth parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
# i row j col k dep
class Initializer_2p5D_XZ(ProcessGroupInitializer):
"""2.5d tensor parallel initialization among cols times dep.
Args:
tesseract_dim (int): The dimension of tesseract.
tesseract_dep (int): The dimension of depth.
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.
"""
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
def init_dist_group(self):
"""Initialize 2.5D tensor colXdepth parallel groups, and assign local_ranks and groups to each gpu.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
2.5D tensor colXdepth parallelism's information in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_2p5D(ProcessGroupInitializer):
"""
Serve as the single entry point to Tesseract parallel initialization.
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.
depth (int): The depth of 2.5d parallel.
"""
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 2.5D tensor row, col, depth, and colXdepth parallel groups, and assign local_ranks and groups to each gpu.
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
Whole 2.5D tensor parallelism's information in a list of tuples.
"""
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

329
colossalai/legacy/context/process_group_initializer/initializer_3d.py Normal file
View File

@@ -0,0 +1,329 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import math
import torch.distributed as dist
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.legacy.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.
Args:
num_group (int): The number of all tensor groups.
depth (int): Depth of 3D parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
3D tensor parallelism's information among input in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_3D_Weight(ProcessGroupInitializer):
"""3D tensor parallel initialization among weight.
Args:
num_group (int): The number of all tensor groups.
depth (int): Depth of 3D parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
3D tensor parallelism's information among weight in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_3D_Output(ProcessGroupInitializer):
"""3D tensor parallel initialization among output.
Args:
num_group (int): The number of all tensor groups.
depth (int): Depth of 3D parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
3D tensor parallelism's information among output in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_3D_InputxWeight(ProcessGroupInitializer):
"""3D tensor parallel initialization among input.
Args:
num_group (int): The number of all tensor groups.
depth (int): Depth of 3D parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
3D tensor parallelism's information among input in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_3D_INPUT_X_WEIGHT
env.input_x_weight_group_3d = mode
for h in range(self.num_group):
for k in range(self.depth):
ranks = [
h * self.depth**3 + i + self.depth * (j + self.depth * k)
for j in range(self.depth)
for i in range(self.depth)
]
group = dist.new_group(ranks)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
class Initializer_3D_OutputxWeight(ProcessGroupInitializer):
"""3D tensor parallel initialization among input.
Args:
num_group (int): The number of all tensor groups.
depth (int): Depth of 3D parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
3D tensor parallelism's information among input in a tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_group = None
group_world_size = None
mode = ParallelMode.PARALLEL_3D_OUTPUT_X_WEIGHT
env.output_x_weight_group_3d = mode
for h in range(self.num_group):
for j in range(self.depth):
ranks = [
h * self.depth**3 + i + self.depth * (j + self.depth * k)
for k in range(self.depth)
for i in range(self.depth)
]
group = dist.new_group(ranks)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_3D(ProcessGroupInitializer):
"""Serve as the single entry point to 3D parallel initialization.
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.
"""
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)
self.input_x_weight_initializer = Initializer_3D_InputxWeight(self.num_group, self.depth, *args)
self.output_x_weight_initializer = Initializer_3D_OutputxWeight(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.
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
Whole 3D tensor parallelism's information in a list of tuples.
"""
parallel_setting = [
self.input_initializer.init_dist_group(),
self.weight_initializer.init_dist_group(),
self.output_initializer.init_dist_group(),
self.input_x_weight_initializer.init_dist_group(),
self.output_x_weight_initializer.init_dist_group()
]
return parallel_setting

55
colossalai/legacy/context/process_group_initializer/initializer_data.py Normal file
View File

@@ -0,0 +1,55 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from torch import distributed as dist
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Data(ProcessGroupInitializer):
"""A ProcessGroupInitializer for data parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
A Data parallelism's information tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode

57
colossalai/legacy/context/process_group_initializer/initializer_model.py Normal file
View File

@@ -0,0 +1,57 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Model(ProcessGroupInitializer):
"""A ProcessGroupInitializer for model parallelism (model parallel group contains pipeline and tensor parallel
groups).
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
A Model parallelism's information tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode

56
colossalai/legacy/context/process_group_initializer/initializer_pipeline.py Normal file
View File

@@ -0,0 +1,56 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from torch import distributed as dist
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Pipeline(ProcessGroupInitializer):
"""A ProcessGroupInitializer for pipeline parallelism.
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
"""
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.
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
A Pipeline parallelism's information in list of tuples.
"""
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)
group_cpu = dist.new_group(pipe_ranks, backend='gloo') if dist.get_backend() != 'gloo' else pipe_group
if self.rank in pipe_ranks:
local_rank = pipe_ranks.index(self.rank)
group_world_size = pipe_group_size
process_group = pipe_group
cpu_group = group_cpu
ranks_in_group = pipe_ranks
dist_settings.append(
tuple((local_rank, group_world_size, process_group, cpu_group, ranks_in_group,
ParallelMode.PIPELINE)))
return dist_settings

101
colossalai/legacy/context/process_group_initializer/initializer_sequence.py Normal file
View File

@@ -0,0 +1,101 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .initializer_tensor import Initializer_Tensor
from .process_group_initializer import ProcessGroupInitializer
@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
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
"""
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.
Returns:
Tuple: A tuple (local_rank, group_world_size, process_group, ranks_in_group, mode).
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Sequence(ProcessGroupInitializer):
"""A ProcessGroupInitializer for sequence parallelism.
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.
"""
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 partial query, key and value embedding to compute self attention values. The second is for
all-reduce to synchronize the model parameters.
Returns:
List[Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode)]:
A Sequence parallelism's information in list of tuples.
"""
parallel_setting = []
local_rank, group_world_size, process_group, cpu_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, cpu_group, ranks_in_group, mode))
parallel_setting.append(self._sequence_dp_initializer.init_dist_group())
return parallel_setting

55
colossalai/legacy/context/process_group_initializer/initializer_tensor.py Normal file
View File

@@ -0,0 +1,55 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch.distributed as dist
from colossalai.legacy.registry import DIST_GROUP_INITIALIZER
from ..parallel_mode import ParallelMode
from .process_group_initializer import ProcessGroupInitializer
@DIST_GROUP_INITIALIZER.register_module
class Initializer_Tensor(ProcessGroupInitializer):
"""A ProcessGroupInitializer for tensor parallelism.
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.
"""
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.
Returns:
Tuple (local_rank, group_world_size, process_group, ranks_in_group, mode):
A Tensor parallelism's information tuple.
"""
local_rank = None
ranks_in_group = None
process_group = None
cpu_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)
group_cpu = dist.new_group(ranks, backend='gloo') if dist.get_backend() != 'gloo' else group
if self.rank in ranks:
local_rank = ranks.index(self.rank)
group_world_size = len(ranks)
process_group = group
cpu_group = group_cpu
ranks_in_group = ranks
return local_rank, group_world_size, process_group, cpu_group, ranks_in_group, mode

33
colossalai/legacy/context/process_group_initializer/process_group_initializer.py Normal file
View File

@@ -0,0 +1,33 @@
#!/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.
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.
"""
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