github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-02 09:38:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

[misc] update pre-commit and run all files (#4752)

Browse Source 
* [misc] update pre-commit

* [misc] run pre-commit

* [misc] remove useless configuration files

* [misc] ignore cuda for clang-format
This commit is contained in:
Hongxin Liu
2023-09-19 14:20:26 +08:00
committed by GitHub
parent 3c6b831c26
commit 079bf3cb26
1268 changed files with 50037 additions and 38444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
colossalai/legacy/nn/layer/parallel_2d/__init__.py
View File

@@ -10,6 +10,13 @@ from .layers import (
)
__all__ = [
'split_batch_2d', 'reduce_by_batch_2d', 'Linear2D', 'LayerNorm2D', 'Classifier2D', 'PatchEmbedding2D',
'Embedding2D', 'VocabParallelEmbedding2D', 'VocabParallelClassifier2D'
"split_batch_2d",
"reduce_by_batch_2d",
"Linear2D",
"LayerNorm2D",
"Classifier2D",
"PatchEmbedding2D",
"Embedding2D",
"VocabParallelEmbedding2D",
"VocabParallelClassifier2D",
]

303
colossalai/legacy/nn/layer/parallel_2d/_operation.py
View File

@@ -5,10 +5,9 @@ import torch.distributed as dist
from torch import Tensor
from torch.cuda.amp import custom_bwd, custom_fwd
from colossalai.legacy.communication.collective import all_gather, all_reduce, reduce, reduce_scatter
from colossalai.legacy.communication.collective import all_gather, all_reduce, reduce_scatter
from colossalai.legacy.context.parallel_mode import ParallelMode
from colossalai.legacy.core import global_context as gpc
from colossalai.legacy.global_variables import tensor_parallel_env as env
from colossalai.utils import get_current_device
@@ -49,17 +48,30 @@ def matmul_2d(
col_rank = gpc.get_local_rank(row_parallel_mode)
data_parallel_rank = 0 if not gpc.is_initialized(ParallelMode.DATA) else gpc.get_local_rank(ParallelMode.DATA)
pipeline_parallel_rank = 0 if not gpc.is_initialized(ParallelMode.PIPELINE) else gpc.get_local_rank(
ParallelMode.PIPELINE)
pipeline_parallel_size = 1 if not gpc.is_initialized(ParallelMode.PIPELINE) else gpc.get_world_size(
ParallelMode.PIPELINE)
pipeline_parallel_rank = (
0 if not gpc.is_initialized(ParallelMode.PIPELINE) else gpc.get_local_rank(ParallelMode.PIPELINE)
)
pipeline_parallel_size = (
1 if not gpc.is_initialized(ParallelMode.PIPELINE) else gpc.get_world_size(ParallelMode.PIPELINE)
)
tensor_parallel_size = summa_dim**2
return Matmul_AB_2D(a, b, summa_dim, out_shape, row_rank, col_rank, row_parallel_mode, col_parallel_mode,
data_parallel_rank, pipeline_parallel_rank, pipeline_parallel_size, tensor_parallel_size)
return Matmul_AB_2D(
a,
b,
summa_dim,
out_shape,
row_rank,
col_rank,
row_parallel_mode,
col_parallel_mode,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size,
)
class _Classifier2D(torch.autograd.Function):
@staticmethod
@custom_fwd(cast_inputs=torch.float16)
def forward(
@@ -132,10 +144,21 @@ class _Classifier2D(torch.autograd.Function):
return A_grad, B_grad, bias_grad, None, None, None, None, None, None, None, None, None, None
def classifier_2d(A: Tensor, B: Tensor, bias: Optional[Tensor], summa_dim: int, out_shape: Tuple[int, ...],
row_rank: int, col_rank: int, row_parallel_mode: ParallelMode, col_parallel_mode: ParallelMode,
data_parallel_rank: int, pipeline_parallel_rank: int, pipeline_parallel_size: int,
tensor_parallel_size: int) -> Tensor:
def classifier_2d(
A: Tensor,
B: Tensor,
bias: Optional[Tensor],
summa_dim: int,
out_shape: Tuple[int, ...],
row_rank: int,
col_rank: int,
row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode,
data_parallel_rank: int,
pipeline_parallel_rank: int,
pipeline_parallel_size: int,
tensor_parallel_size: int,
) -> Tensor:
r"""2D parallel classifier.
Args:
@@ -157,9 +180,21 @@ def classifier_2d(A: Tensor, B: Tensor, bias: Optional[Tensor], summa_dim: int,
The parallel_mode should be concluded in ``ParallelMode``. More details about ``ParallelMode`` could be found
in `parallel_mode <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/context/parallel_mode.py>`_
"""
return _Classifier2D.apply(A, B, bias, summa_dim, out_shape, row_rank, col_rank, row_parallel_mode,
col_parallel_mode, data_parallel_rank, pipeline_parallel_rank, pipeline_parallel_size,
tensor_parallel_size)
return _Classifier2D.apply(
A,
B,
bias,
summa_dim,
out_shape,
row_rank,
col_rank,
row_parallel_mode,
col_parallel_mode,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size,
)
class Matmul_AB_2D(torch.autograd.Function):
@@ -205,8 +240,7 @@ class Matmul_AB_2D(torch.autograd.Function):
# B: [h / q, s / q]
# C: [b / q, s, s / q] -> [(b * s) / q, s / q]
assert A.shape[-1] == B.shape[-2], \
'Invalid shapes: A={}, B={} for AB.'.format(A.shape, B.shape)
assert A.shape[-1] == B.shape[-2], "Invalid shapes: A={}, B={} for AB.".format(A.shape, B.shape)
if ctx:
ctx.save_for_backward(A, B)
@@ -226,10 +260,16 @@ class Matmul_AB_2D(torch.autograd.Function):
row_group = gpc.get_group(row_parallel_mode)
col_group = gpc.get_group(col_parallel_mode)
src_a = summa_dim * row_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_b = col_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_a = (
summa_dim * row_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
src_b = (
col_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
opa = [None] * 2
opb = [None] * 2
@@ -278,14 +318,34 @@ class Matmul_AB_2D(torch.autograd.Function):
def backward(ctx: Any, output_grad: Tensor) -> Tuple[Tensor, ...]:
A, B = ctx.saved_tensors
with torch.no_grad():
A_grad = Matmul_ABT_2D.apply(output_grad, B, ctx.summa_dim, ctx.A_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
B_grad = Matmul_ATB_2D.apply(A, output_grad, ctx.summa_dim, ctx.B_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
A_grad = Matmul_ABT_2D.apply(
output_grad,
B,
ctx.summa_dim,
ctx.A_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
B_grad = Matmul_ATB_2D.apply(
A,
output_grad,
ctx.summa_dim,
ctx.B_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
return A_grad, B_grad, None, None, None, None, None, None, None, None, None, None
@@ -329,9 +389,7 @@ class Matmul_ABT_2D(torch.autograd.Function):
pipeline_parallel_size: int,
tensor_parallel_size: int,
) -> Tensor:
assert A.shape[-1] == B.shape[-1], \
'Invalid shapes: A={}, B={} for ABT.'.format(A.shape, B.shape)
assert A.shape[-1] == B.shape[-1], "Invalid shapes: A={}, B={} for ABT.".format(A.shape, B.shape)
if ctx:
ctx.save_for_backward(A, B)
@@ -351,10 +409,16 @@ class Matmul_ABT_2D(torch.autograd.Function):
row_group = gpc.get_group(row_parallel_mode)
col_group = gpc.get_group(col_parallel_mode)
src_b = col_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_c = summa_dim * row_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_b = (
col_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
src_c = (
summa_dim * row_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
opb = [None] * 2
opr = [None] * 2
@@ -412,14 +476,34 @@ class Matmul_ABT_2D(torch.autograd.Function):
A, B = ctx.saved_tensors
with torch.no_grad():
A_grad = Matmul_AB_2D.apply(output_grad, B, ctx.summa_dim, ctx.A_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
B_grad = Matmul_ATB_2D.apply(output_grad, A, ctx.summa_dim, ctx.B_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
A_grad = Matmul_AB_2D.apply(
output_grad,
B,
ctx.summa_dim,
ctx.A_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
B_grad = Matmul_ATB_2D.apply(
output_grad,
A,
ctx.summa_dim,
ctx.B_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
return A_grad, B_grad, None, None, None, None, None, None, None, None, None, None
@@ -462,9 +546,7 @@ class Matmul_ATB_2D(torch.autograd.Function):
pipeline_parallel_size: int,
tensor_parallel_size: int,
) -> Tensor:
assert A.shape[-2] == B.shape[-2], \
'Invalid shapes: A={}, B={} for ATB.'.format(A.shape, B.shape)
assert A.shape[-2] == B.shape[-2], "Invalid shapes: A={}, B={} for ATB.".format(A.shape, B.shape)
if ctx:
ctx.save_for_backward(A, B)
@@ -484,10 +566,16 @@ class Matmul_ATB_2D(torch.autograd.Function):
row_group = gpc.get_group(row_parallel_mode)
col_group = gpc.get_group(col_parallel_mode)
src_a = summa_dim * row_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_c = col_rank + data_parallel_rank * pipeline_parallel_size * tensor_parallel_size + \
pipeline_parallel_rank * tensor_parallel_size
src_a = (
summa_dim * row_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
src_c = (
col_rank
+ data_parallel_rank * pipeline_parallel_size * tensor_parallel_size
+ pipeline_parallel_rank * tensor_parallel_size
)
opa = [None] * 2
opr = [None] * 2
@@ -545,19 +633,38 @@ class Matmul_ATB_2D(torch.autograd.Function):
A, B = ctx.saved_tensors
with torch.no_grad():
A_grad = Matmul_ABT_2D.apply(B, output_grad, ctx.summa_dim, ctx.A_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
B_grad = Matmul_AB_2D.apply(A, output_grad, ctx.summa_dim, ctx.B_shape, ctx.row_rank, ctx.col_rank,
ctx.row_parallel_mode, ctx.col_parallel_mode, ctx.data_parallel_rank,
ctx.pipeline_parallel_rank, ctx.pipeline_parallel_size,
ctx.tensor_parallel_size)
A_grad = Matmul_ABT_2D.apply(
B,
output_grad,
ctx.summa_dim,
ctx.A_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
B_grad = Matmul_AB_2D.apply(
A,
output_grad,
ctx.summa_dim,
ctx.B_shape,
ctx.row_rank,
ctx.col_rank,
ctx.row_parallel_mode,
ctx.col_parallel_mode,
ctx.data_parallel_rank,
ctx.pipeline_parallel_rank,
ctx.pipeline_parallel_size,
ctx.tensor_parallel_size,
)
return A_grad, B_grad, None, None, None, None, None, None, None, None, None, None
class _Add_Bias_2D(torch.autograd.Function):
@staticmethod
@custom_fwd(cast_inputs=torch.float16)
def forward(
@@ -608,10 +715,20 @@ class _Add_Bias_2D(torch.autograd.Function):
return output_grad, grad, None, None, None, None, None, None, None, None, None, None
def add_bias_2d(input_: Tensor, bias: Tensor, output_size_per_partition: int, row_rank: int, col_rank: int,
row_parallel_mode: ParallelMode, col_parallel_mode: ParallelMode, skip_bias_add: bool,
data_parallel_rank: int, pipeline_parallel_rank: int, pipeline_parallel_size: int,
tensor_parallel_size: int) -> Tensor:
def add_bias_2d(
input_: Tensor,
bias: Tensor,
output_size_per_partition: int,
row_rank: int,
col_rank: int,
row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode,
skip_bias_add: bool,
data_parallel_rank: int,
pipeline_parallel_rank: int,
pipeline_parallel_size: int,
tensor_parallel_size: int,
) -> Tensor:
r"""Matrix add bias: :math:`C = A + b`.
Args:
@@ -633,17 +750,34 @@ def add_bias_2d(input_: Tensor, bias: Tensor, output_size_per_partition: int, ro
The parallel_mode should be concluded in ``ParallelMode``. More details about ``ParallelMode`` could be found
in `parallel_mode <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/context/parallel_mode.py>`_
"""
return _Add_Bias_2D.apply(input_, bias, output_size_per_partition, row_rank, col_rank, row_parallel_mode,
col_parallel_mode, skip_bias_add, data_parallel_rank, pipeline_parallel_rank,
pipeline_parallel_size, tensor_parallel_size)
return _Add_Bias_2D.apply(
input_,
bias,
output_size_per_partition,
row_rank,
col_rank,
row_parallel_mode,
col_parallel_mode,
skip_bias_add,
data_parallel_rank,
pipeline_parallel_rank,
pipeline_parallel_size,
tensor_parallel_size,
)
class _Layernorm_2D(torch.autograd.Function):
@staticmethod
@custom_fwd(cast_inputs=torch.float32)
def forward(ctx: Any, input_: Tensor, E_x: Tensor, Var_x: Tensor, hidden_size: int, row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode) -> Tensor:
def forward(
ctx: Any,
input_: Tensor,
E_x: Tensor,
Var_x: Tensor,
hidden_size: int,
row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode,
) -> Tensor:
input_ = input_ - E_x
# in here, input = x - E[x], Var_x = 1 / sqrt(Var[x] + eps)
ctx.normalized_shape = hidden_size
@@ -657,7 +791,7 @@ class _Layernorm_2D(torch.autograd.Function):
@custom_bwd
def backward(ctx: Any, output_grad: Tensor) -> Tuple[Tensor, ...]:
row_parallel_mode = ctx.row_parallel_mode
col_parallel_mode = ctx.col_parallel_mode
ctx.col_parallel_mode
x, Var_x = ctx.saved_tensors
# in here, Var_x = 1 / sqrt(Var[x] + eps), x = (x - E[x]) * Var_x
output_grad_sum = torch.sum(output_grad, dim=-1, keepdim=True)
@@ -676,8 +810,14 @@ class _Layernorm_2D(torch.autograd.Function):
return input_grad, None, None, None, None, None
def layernorm_2d(input_: Tensor, E_x: Tensor, Var_x: Tensor, hidden_size: int, row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode) -> Tensor:
def layernorm_2d(
input_: Tensor,
E_x: Tensor,
Var_x: Tensor,
hidden_size: int,
row_parallel_mode: ParallelMode,
col_parallel_mode: ParallelMode,
) -> Tensor:
r"""Layernorm.
Args:
@@ -696,7 +836,6 @@ def layernorm_2d(input_: Tensor, E_x: Tensor, Var_x: Tensor, hidden_size: int, r
class _AllGatherTensor2D(torch.autograd.Function):
@staticmethod
@custom_fwd(cast_inputs=torch.float16)
def forward(ctx: Any, inputs: Tensor, dim: int, parallel_mode: ParallelMode) -> Tensor:
@@ -744,15 +883,14 @@ def split_batch_2d(input_: Tensor, dim: int = 0) -> Tensor:
if world_size <= 1:
return input_
assert dim_size % world_size == 0, \
f'The batch size ({dim_size}) is not a multiple of 2D size ({world_size}).'
assert dim_size % world_size == 0, f"The batch size ({dim_size}) is not a multiple of 2D size ({world_size})."
return torch.chunk(input_, gpc.get_world_size(ParallelMode.PARALLEL_2D_COL),
dim=dim)[gpc.get_local_rank(ParallelMode.PARALLEL_2D_COL)].contiguous()
return torch.chunk(input_, gpc.get_world_size(ParallelMode.PARALLEL_2D_COL), dim=dim)[
gpc.get_local_rank(ParallelMode.PARALLEL_2D_COL)
].contiguous()
class _ReduceTensor2D(torch.autograd.Function):
@staticmethod
def forward(ctx, input_, parallel_mode):
return all_reduce(input_, parallel_mode)
@@ -777,7 +915,6 @@ def reduce_tensor_2d(input_: Tensor, parallel_mode: ParallelMode) -> Tensor:
class _ReduceScatterTensor2D(torch.autograd.Function):
@staticmethod
def forward(ctx, input_, dim, parallel_mode):
ctx.dim = dim
@@ -803,14 +940,12 @@ def reduce_scatter_tensor_2d(tensor: Tensor, dim: int, parallel_mode: ParallelMo
"""
dim_size = tensor.size(dim)
world_size = gpc.get_world_size(parallel_mode)
assert dim_size % world_size == 0, \
f'The batch size ({dim_size}) is not a multiple of 2D size ({world_size}).'
assert dim_size % world_size == 0, f"The batch size ({dim_size}) is not a multiple of 2D size ({world_size})."
return _ReduceScatterTensor2D.apply(tensor, dim, parallel_mode)
class _ReduceByBatch2D(torch.autograd.Function):
@staticmethod
def symbolic(graph, input_, reduce_mean: bool = False):
output = all_reduce(input_, ParallelMode.PARALLEL_2D_COL)

16
colossalai/legacy/nn/layer/parallel_2d/_utils.py
View File

@@ -6,15 +6,17 @@ from colossalai.legacy.global_variables import tensor_parallel_env as env
def get_summa_dim_from_env() -> int:
try:
summa_dim = env.summa_dim
assert summa_dim > 0, 'SUMMA_DIM must be larger than zero'
assert summa_dim > 0, "SUMMA_DIM must be larger than zero"
return summa_dim
except KeyError as e:
raise EnvironmentError('SUMMA_DIM is not found in the current environment, '
'please make sure that you have used the correct process group initializer')
except KeyError:
raise EnvironmentError(
"SUMMA_DIM is not found in the current environment, "
"please make sure that you have used the correct process group initializer"
)
def assert_summa_initialization():
assert gpc.is_initialized(ParallelMode.PARALLEL_2D_COL) and \
gpc.is_initialized(ParallelMode.PARALLEL_2D_ROW), \
'Both TWO_DIMENSION_COL and TWO_DIMENSION_ROW must be initialized by the process group initializer'
assert gpc.is_initialized(ParallelMode.PARALLEL_2D_COL) and gpc.is_initialized(
ParallelMode.PARALLEL_2D_ROW
), "Both TWO_DIMENSION_COL and TWO_DIMENSION_ROW must be initialized by the process group initializer"

615
colossalai/legacy/nn/layer/parallel_2d/layers.py
View File

@@ -55,14 +55,16 @@ class Linear2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_.
"""
def __init__(self,
in_features: int,
out_features: int,
bias: bool = True,
dtype: torch.dtype = None,
skip_bias_add: bool = False,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1)):
def __init__(
self,
in_features: int,
out_features: int,
bias: bool = True,
dtype: torch.dtype = None,
skip_bias_add: bool = False,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1),
):
super().__init__()
self.in_features = in_features
@@ -80,15 +82,16 @@ class Linear2D(ParallelLayer):
self.hidden_size_per_partition = divide(self.out_features, self.summa_dim)
# create weight, shape: [k/q, h/q]
factory_kwargs = {'device': get_current_device(), 'dtype': dtype}
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
self.weight = Parameter(
torch.empty(self.input_size_per_partition, self.hidden_size_per_partition, **factory_kwargs))
torch.empty(self.input_size_per_partition, self.hidden_size_per_partition, **factory_kwargs)
)
# create bias, shape: [h/q]
if bias:
self.bias = Parameter(torch.empty(divide(self.out_features, self.summa_dim**2), **factory_kwargs))
else:
self.register_parameter('bias', None)
self.register_parameter("bias", None)
# initialize parameters
with seed(ParallelMode.TENSOR):
@@ -108,8 +111,8 @@ class Linear2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
weight = state_dict.pop(weight_key, None)
@@ -126,34 +129,22 @@ class Linear2D(ParallelLayer):
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
# partition in column groups
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
local_state = OrderedDict({weight_key: self.weight})
if self.bias is not None:
local_state[bias_key] = self.bias
@@ -162,14 +153,8 @@ class Linear2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
# gather in row groups
@@ -177,14 +162,8 @@ class Linear2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
@@ -196,22 +175,53 @@ class Linear2D(ParallelLayer):
# output: [m/q, n/q, h/q]
out_shape = x.shape[:-1] + (self.hidden_size_per_partition,)
output = Matmul_AB_2D.apply(x, self.weight, self.summa_dim, out_shape, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, self.data_parallel_rank,
self.pipeline_parallel_rank, self.pipeline_parallel_size, self.tensor_parallel_size)
output = Matmul_AB_2D.apply(
x,
self.weight,
self.summa_dim,
out_shape,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
if self.bias is not None:
if self.skip_bias_add:
bias = add_bias_2d(None, self.bias, self.hidden_size_per_partition, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, True,
self.data_parallel_rank, self.pipeline_parallel_rank, self.pipeline_parallel_size,
self.tensor_parallel_size)
bias = add_bias_2d(
None,
self.bias,
self.hidden_size_per_partition,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
True,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
return output, bias
else:
output = add_bias_2d(output, self.bias, self.hidden_size_per_partition, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, False,
self.data_parallel_rank, self.pipeline_parallel_rank, self.pipeline_parallel_size,
self.tensor_parallel_size)
output = add_bias_2d(
output,
self.bias,
self.hidden_size_per_partition,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
False,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
return output
else:
return output
@@ -249,7 +259,7 @@ class LayerNorm2D(ParallelLayer):
self.partitioned_partition = divide(normalized_shape, self.summa_dim**2)
# create parameters
factory_kwargs = {'device': get_current_device(), 'dtype': dtype}
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
self.weight = Parameter(torch.ones(self.partitioned_partition, **factory_kwargs))
if bias:
@@ -266,8 +276,8 @@ class LayerNorm2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
weight = state_dict.pop(weight_key, None)
@@ -283,34 +293,22 @@ class LayerNorm2D(ParallelLayer):
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
# partition in column groups
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
local_state = OrderedDict({weight_key: self.weight})
if self.bias is not None:
local_state[bias_key] = self.bias
@@ -319,14 +317,8 @@ class LayerNorm2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
# gather in row groups
@@ -334,14 +326,8 @@ class LayerNorm2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
@@ -349,29 +335,51 @@ class LayerNorm2D(ParallelLayer):
def forward(self, x: Tensor) -> Tensor:
with torch.no_grad():
E_x = torch.sum(x, dim=-1, keepdim=True) # [b/q, s, 1]
E_x = torch.sum(x, dim=-1, keepdim=True) # [b/q, s, 1]
torch.distributed.all_reduce(E_x, group=gpc.get_group(ParallelMode.PARALLEL_2D_ROW))
E_x /= self.normalized_shape
# Var_x in the block below is the sum of input^2
Var_x = torch.sum(x * x, dim=-1, keepdim=True) # [b/q, s, 1]
Var_x = torch.sum(x * x, dim=-1, keepdim=True) # [b/q, s, 1]
torch.distributed.all_reduce(Var_x, group=gpc.get_group(ParallelMode.PARALLEL_2D_ROW))
Var_x /= self.normalized_shape
Var_x = Var_x - E_x * E_x # variance of x [b/q, s, 1]
Var_x = Var_x - E_x * E_x # variance of x [b/q, s, 1]
# this time 1/sqrt(Var_x + epsilon)
Var_x = 1.0 / torch.sqrt(Var_x + self.variance_epsilon)
output = layernorm_2d(x, E_x, Var_x, self.normalized_shape, ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL)
scale = add_bias_2d(None, self.weight, self.partitioned_partition, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, True, self.data_parallel_rank,
self.pipeline_parallel_rank, self.pipeline_parallel_size, self.tensor_parallel_size)
output = layernorm_2d(
x, E_x, Var_x, self.normalized_shape, ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL
)
scale = add_bias_2d(
None,
self.weight,
self.partitioned_partition,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
True,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
if self.bias is not None:
bias = add_bias_2d(None, self.bias, self.partitioned_partition, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, True,
self.data_parallel_rank, self.pipeline_parallel_rank, self.pipeline_parallel_size,
self.tensor_parallel_size)
bias = add_bias_2d(
None,
self.bias,
self.partitioned_partition,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
True,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
output = torch.addcmul(bias, scale, output)
else:
output = torch.mul(scale, output)
@@ -400,16 +408,18 @@ class PatchEmbedding2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_.
"""
def __init__(self,
img_size: int,
patch_size: int,
in_chans: int,
embed_size: int,
flatten: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1),
position_embed_initializer: Callable = init.zeros_()):
def __init__(
self,
img_size: int,
patch_size: int,
in_chans: int,
embed_size: int,
flatten: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1),
position_embed_initializer: Callable = init.zeros_(),
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
@@ -426,17 +436,22 @@ class PatchEmbedding2D(ParallelLayer):
with seed(ParallelMode.TENSOR):
self.weight = Parameter(
torch.empty((self.embed_size_per_partition, in_chans, *self.patch_size),
device=get_current_device(),
dtype=dtype))
torch.empty(
(self.embed_size_per_partition, in_chans, *self.patch_size),
device=get_current_device(),
dtype=dtype,
)
)
self.bias = Parameter(torch.empty(self.embed_size_per_partition, device=get_current_device(), dtype=dtype))
self.cls_token = Parameter(
torch.zeros((1, 1, self.embed_size_per_partition), device=get_current_device(), dtype=dtype))
torch.zeros((1, 1, self.embed_size_per_partition), device=get_current_device(), dtype=dtype)
)
self.pos_embed = Parameter(
torch.zeros((1, self.num_patches + 1, self.embed_size_per_partition),
device=get_current_device(),
dtype=dtype))
torch.zeros(
(1, self.num_patches + 1, self.embed_size_per_partition), device=get_current_device(), dtype=dtype
)
)
self.reset_parameters(weight_initializer, bias_initializer, position_embed_initializer)
self._set_tensor_parallel_attribute()
@@ -457,10 +472,10 @@ class PatchEmbedding2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
cls_token_key = prefix + 'cls_token'
pos_embed_key = prefix + 'pos_embed'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
cls_token_key = prefix + "cls_token"
pos_embed_key = prefix + "pos_embed"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
weight = state_dict.pop(weight_key, None)
@@ -484,67 +499,34 @@ class PatchEmbedding2D(ParallelLayer):
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: 0,
bias_key: 0,
cls_token_key: -1,
pos_embed_key: -1
},
partition_states={
weight_key: True,
bias_key: True,
cls_token_key: True,
pos_embed_key: True
},
dims={weight_key: 0, bias_key: 0, cls_token_key: -1, pos_embed_key: -1},
partition_states={weight_key: True, bias_key: True, cls_token_key: True, pos_embed_key: True},
)
# partition in column groups
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0,
cls_token_key: -1,
pos_embed_key: -1
},
partition_states={
weight_key: True,
bias_key: True,
cls_token_key: True,
pos_embed_key: True
},
dims={weight_key: 0, bias_key: 0, cls_token_key: -1, pos_embed_key: -1},
partition_states={weight_key: True, bias_key: True, cls_token_key: True, pos_embed_key: True},
)
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
cls_token_key = prefix + 'cls_token'
pos_embed_key = prefix + 'pos_embed'
local_state = OrderedDict({
weight_key: self.weight,
bias_key: self.bias,
cls_token_key: self.cls_token,
pos_embed_key: self.pos_embed
})
weight_key = prefix + "weight"
bias_key = prefix + "bias"
cls_token_key = prefix + "cls_token"
pos_embed_key = prefix + "pos_embed"
local_state = OrderedDict(
{weight_key: self.weight, bias_key: self.bias, cls_token_key: self.cls_token, pos_embed_key: self.pos_embed}
)
# gather in column groups
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0,
cls_token_key: -1,
pos_embed_key: -1
},
partition_states={
weight_key: True,
bias_key: True,
cls_token_key: True,
pos_embed_key: True
},
dims={weight_key: 0, bias_key: 0, cls_token_key: -1, pos_embed_key: -1},
partition_states={weight_key: True, bias_key: True, cls_token_key: True, pos_embed_key: True},
keep_vars=keep_vars,
)
# gather in row groups
@@ -552,18 +534,8 @@ class PatchEmbedding2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: 0,
bias_key: 0,
cls_token_key: -1,
pos_embed_key: -1
},
partition_states={
weight_key: True,
bias_key: True,
cls_token_key: True,
pos_embed_key: True
},
dims={weight_key: 0, bias_key: 0, cls_token_key: -1, pos_embed_key: -1},
partition_states={weight_key: True, bias_key: True, cls_token_key: True, pos_embed_key: True},
keep_vars=keep_vars,
)
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
@@ -573,15 +545,16 @@ class PatchEmbedding2D(ParallelLayer):
input_ = split_batch_2d(input_)
B, C, H, W = input_.shape
assert H == self.img_size[0] and W == self.img_size[1], \
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
assert (
H == self.img_size[0] and W == self.img_size[1]
), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
weight = all_gather_tensor_2d(self.weight, 0, ParallelMode.PARALLEL_2D_COL)
bias = all_gather_tensor_2d(self.bias, 0, ParallelMode.PARALLEL_2D_COL)
output = F.conv2d(input_, weight, bias, stride=self.patch_size)
if self.flatten:
output = output.flatten(2).transpose(1, 2) # BCHW -> BNC
output = output.flatten(2).transpose(1, 2) # BCHW -> BNC
cls_token = all_gather_tensor_2d(self.cls_token, -1, ParallelMode.PARALLEL_2D_COL)
pos_embed = all_gather_tensor_2d(self.pos_embed, -1, ParallelMode.PARALLEL_2D_COL)
@@ -623,14 +596,16 @@ class Embedding2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_
"""
def __init__(self,
num_embeddings: int,
embedding_dim: int,
padding_idx: int = None,
dtype: torch.dtype = None,
weight_initializer: Callable = init.normal_(),
*args,
**kwargs):
def __init__(
self,
num_embeddings: int,
embedding_dim: int,
padding_idx: int = None,
dtype: torch.dtype = None,
weight_initializer: Callable = init.normal_(),
*args,
**kwargs,
):
super().__init__()
assert_summa_initialization()
@@ -644,7 +619,8 @@ class Embedding2D(ParallelLayer):
self.embed_kwargs = kwargs
self.weight = Parameter(
torch.empty((num_embeddings, embed_dim_per_partition), device=get_current_device(), dtype=dtype))
torch.empty((num_embeddings, embed_dim_per_partition), device=get_current_device(), dtype=dtype)
)
self.reset_parameters(weight_initializer)
self._set_tensor_parallel_attributes()
@@ -665,7 +641,7 @@ class Embedding2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
weight_key = prefix + "weight"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
weight = state_dict.pop(weight_key, None)
@@ -691,7 +667,7 @@ class Embedding2D(ParallelLayer):
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
weight_key = prefix + "weight"
local_state = OrderedDict({weight_key: self.weight})
# gather in column groups
@@ -754,14 +730,16 @@ class VocabParallelEmbedding2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_.
"""
def __init__(self,
num_embeddings: int,
embedding_dim: int,
padding_idx: int = None,
dtype: torch.dtype = None,
weight_initializer: Callable = init.normal_(),
*args,
**kwargs):
def __init__(
self,
num_embeddings: int,
embedding_dim: int,
padding_idx: int = None,
dtype: torch.dtype = None,
weight_initializer: Callable = init.normal_(),
*args,
**kwargs,
):
super().__init__()
self.num_embeddings = num_embeddings
self.embed_dim = embedding_dim
@@ -778,9 +756,12 @@ class VocabParallelEmbedding2D(ParallelLayer):
self.vocab_end_index = self.vocab_start_index + self.num_embeddings_per_partition
self.weight = Parameter(
torch.empty((self.num_embeddings_per_partition, self.embed_dim_per_partition),
device=get_current_device(),
dtype=dtype))
torch.empty(
(self.num_embeddings_per_partition, self.embed_dim_per_partition),
device=get_current_device(),
dtype=dtype,
)
)
self.reset_parameters(weight_initializer)
self._set_tensor_parallel_attributes()
@@ -796,14 +777,17 @@ class VocabParallelEmbedding2D(ParallelLayer):
self._fill_padding_idx_with_zero()
def _fill_padding_idx_with_zero(self) -> None:
if self.padding_idx is not None and \
self.padding_idx >= self.vocab_start_index and self.padding_idx < self.vocab_end_index:
if (
self.padding_idx is not None
and self.padding_idx >= self.vocab_start_index
and self.padding_idx < self.vocab_end_index
):
with torch.no_grad():
self.weight[self.padding_idx - self.vocab_start_index].fill_(0)
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
weight_key = prefix + "weight"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
weight = state_dict.pop(weight_key, None)
@@ -829,7 +813,7 @@ class VocabParallelEmbedding2D(ParallelLayer):
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
weight_key = prefix + "weight"
local_state = OrderedDict({weight_key: self.weight})
# gather in column groups
@@ -857,10 +841,11 @@ class VocabParallelEmbedding2D(ParallelLayer):
masked_input = input_.clone() - self.vocab_start_index
masked_input[input_mask] = 0
output_parallel = F.embedding(masked_input, self.weight, self.padding_idx, *self.embed_args,
**self.embed_kwargs)
output_parallel = F.embedding(
masked_input, self.weight, self.padding_idx, *self.embed_args, **self.embed_kwargs
)
output_parallel[input_mask, :] = 0.
output_parallel[input_mask, :] = 0.0
output = reduce_scatter_tensor_2d(output_parallel, 0, ParallelMode.PARALLEL_2D_COL)
return output
@@ -884,14 +869,16 @@ class Classifier2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_.
"""
def __init__(self,
in_features: int,
num_classes: int,
weight: Parameter = None,
bias: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1)):
def __init__(
self,
in_features: int,
num_classes: int,
weight: Parameter = None,
bias: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1),
):
super().__init__()
self.in_features = in_features
self.num_classes = num_classes
@@ -908,7 +895,8 @@ class Classifier2D(ParallelLayer):
self.has_weight = False
else:
self.weight = Parameter(
torch.empty(self.num_classes, self.input_size_per_partition, device=get_current_device(), dtype=dtype))
torch.empty(self.num_classes, self.input_size_per_partition, device=get_current_device(), dtype=dtype)
)
self.has_weight = True
if bias:
self.bias = Parameter(torch.zeros(self.num_classes, device=get_current_device(), dtype=dtype))
@@ -938,8 +926,8 @@ class Classifier2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
if self.has_weight:
@@ -957,34 +945,22 @@ class Classifier2D(ParallelLayer):
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: False
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: False},
)
# partition in column groups
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: False
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: False},
)
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
local_state = OrderedDict()
if self.has_weight:
local_state[weight_key] = self.weight
@@ -995,14 +971,8 @@ class Classifier2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: False
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: False},
keep_vars=keep_vars,
)
# gather in row groups
@@ -1010,14 +980,8 @@ class Classifier2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: False
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: False},
keep_vars=keep_vars,
)
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
@@ -1026,9 +990,21 @@ class Classifier2D(ParallelLayer):
def forward(self, input_: Tensor) -> Tensor:
out_shape = input_.shape[:-1] + (self.num_classes,)
return classifier_2d(input_, self.weight, self.bias, self.summa_dim, out_shape, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, self.data_parallel_rank,
self.pipeline_parallel_rank, self.pipeline_parallel_size, self.tensor_parallel_size)
return classifier_2d(
input_,
self.weight,
self.bias,
self.summa_dim,
out_shape,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
@LAYERS.register_module
@@ -1050,14 +1026,16 @@ class VocabParallelClassifier2D(ParallelLayer):
`init <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/nn/init.py>`_.
"""
def __init__(self,
in_features: int,
num_classes: int,
weight: Parameter = None,
bias: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1)):
def __init__(
self,
in_features: int,
num_classes: int,
weight: Parameter = None,
bias: bool = True,
dtype: torch.dtype = None,
weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)),
bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1),
):
super().__init__()
self.in_features = in_features
@@ -1074,13 +1052,14 @@ class VocabParallelClassifier2D(ParallelLayer):
self.output_size_per_partition = divide(num_classes, self.summa_dim)
# create weight, shape: [k/q, h/q]
factory_kwargs = {'device': get_current_device(), 'dtype': dtype}
factory_kwargs = {"device": get_current_device(), "dtype": dtype}
if weight is not None:
self.weight = weight
self.has_weight = False
else:
self.weight = Parameter(
torch.empty(self.output_size_per_partition, self.input_size_per_partition, **factory_kwargs))
torch.empty(self.output_size_per_partition, self.input_size_per_partition, **factory_kwargs)
)
self.has_weight = True
# create bias, shape: [h/q]
if bias:
@@ -1109,8 +1088,8 @@ class VocabParallelClassifier2D(ParallelLayer):
def _load_from_global_state_dict(self, state_dict, prefix, *args, **kwargs):
local_state = OrderedDict()
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
# weight
if self.has_weight:
@@ -1128,34 +1107,22 @@ class VocabParallelClassifier2D(ParallelLayer):
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
# partition in column groups
local_state = partition_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
)
super()._load_from_global_state_dict(local_state, prefix, *args, **kwargs)
def _save_to_global_state_dict(self, destination, prefix, keep_vars):
weight_key = prefix + 'weight'
bias_key = prefix + 'bias'
weight_key = prefix + "weight"
bias_key = prefix + "bias"
local_state = OrderedDict()
if self.has_weight:
local_state[weight_key] = self.weight
@@ -1166,14 +1133,8 @@ class VocabParallelClassifier2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_COL,
dims={
weight_key: 0,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: 0, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
# gather in row groups
@@ -1181,14 +1142,8 @@ class VocabParallelClassifier2D(ParallelLayer):
local_state = gather_tensor_parallel_state_dict(
local_state,
ParallelMode.PARALLEL_2D_ROW,
dims={
weight_key: -1,
bias_key: 0
},
partition_states={
weight_key: True,
bias_key: True
},
dims={weight_key: -1, bias_key: 0},
partition_states={weight_key: True, bias_key: True},
keep_vars=keep_vars,
)
if gpc.get_local_rank(ParallelMode.TENSOR) == 0:
@@ -1200,14 +1155,34 @@ class VocabParallelClassifier2D(ParallelLayer):
# output: [m/q, n/q, h/q]
out_shape = x.shape[:-1] + (self.output_size_per_partition,)
output = Matmul_ABT_2D.apply(x, self.weight, self.summa_dim, out_shape, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL,
self.data_parallel_rank, self.pipeline_parallel_rank, self.pipeline_parallel_size,
self.tensor_parallel_size)
output = Matmul_ABT_2D.apply(
x,
self.weight,
self.summa_dim,
out_shape,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
if self.bias is not None:
output = add_bias_2d(output, self.bias, self.output_size_per_partition, self.row_rank, self.col_rank,
ParallelMode.PARALLEL_2D_ROW, ParallelMode.PARALLEL_2D_COL, False,
self.data_parallel_rank, self.pipeline_parallel_rank, self.pipeline_parallel_size,
self.tensor_parallel_size)
output = add_bias_2d(
output,
self.bias,
self.output_size_per_partition,
self.row_rank,
self.col_rank,
ParallelMode.PARALLEL_2D_ROW,
ParallelMode.PARALLEL_2D_COL,
False,
self.data_parallel_rank,
self.pipeline_parallel_rank,
self.pipeline_parallel_size,
self.tensor_parallel_size,
)
return output