github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-18 16:00:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Tensor] add ColoTensor TP1Dcol Embedding (#899)

Browse Source
This commit is contained in:
Ziyue Jiang
2022-04-28 17:45:06 +08:00
committed by GitHub
parent e46e423c00
commit 2c0d19d755
9 changed files with 173 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
colossalai/tensor/_ops/__init__.py
View File

@@ -2,3 +2,4 @@ from .linear import colo_linear
from .element_wise import *
from .layernorm import colo_layernorm
from .loss import colo_cross_entropy
from .embedding import colo_embedding

56
colossalai/tensor/_ops/embedding.py Normal file
View File

@@ -0,0 +1,56 @@
import torch
from colossalai.tensor.op_wrapper import colo_op_impl
from colossalai.context import ParallelMode
from colossalai.nn.layer.parallel_1d._utils import split_forward_gather_backward, reduce_input, \
gather_forward_split_backward, reduce_grad
from colossalai.nn.layer.utils import divide
from colossalai.core import global_context as gpc
from packaging import version
from colossalai.tensor import ComputePattern, TensorSpec, ComputePattern, ParallelAction, ColoTensor, ShardPattern
def colo_embedding_1Dcol(input_tensor: ColoTensor, weight: ColoTensor, args, kwargs) -> ColoTensor:
# embedding_1Dcol split the weight(lookup table)
# Gather splitted lookup table
parallel_action = weight.shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DCol_Embedding)
if not input_tensor.is_gathered():
input_tensor.gather()
output_parallel = torch.nn.functional.embedding(input_tensor.torch_tensor(), weight.torch_tensor(),
*args, **kwargs)
output = ColoTensor.init_from_torch_tensor(output_parallel)
out_parallel_action_list = [ParallelAction(priority=1, parallel_mode=parallel_action.parallel_mode)]
output_spec = TensorSpec(out_parallel_action_list)
output.set_spec(output_spec, shard=False)
output.set_shard_pattern(ShardPattern.Col)
output.gather()
return output
@colo_op_impl(torch.nn.functional.embedding)
def colo_embedding(types, args, kwargs, pg):
"""Handles ``__torch_function__`` dispatch for ``torch.nn.functional.embedding``.
This method looks up an embedding table.
"""
input_tensor = args[0]
weight = args[1]
args = args[2:]
if not isinstance(input_tensor, ColoTensor):
input_tensor = ColoTensor.init_from_torch_tensor(input_tensor)
if not isinstance(weight, ColoTensor):
weight = ColoTensor.init_from_torch_tensor(weight)
# Handle differen parallel actions.
if not weight.has_spec(): # No Model Parallel Applied
input_tensor = input_tensor.torch_tensor()
weight = weight.torch_tensor()
output = torch.nn.functional.embedding(input_tensor, weight, *args, **kwargs)
return ColoTensor.init_from_torch_tensor(output)
elif weight.shard_spec.num_action == 1: # Single Model Parallel Applied
compute_patterns = weight.shard_spec.compute_patterns
if ComputePattern.TP1DCol_Embedding in compute_patterns:
return colo_embedding_1Dcol(input_tensor, weight, args, kwargs)
else:
raise NotImplementedError
else:
raise NotImplementedError

2
colossalai/tensor/_ops/layernorm.py
View File

@@ -27,7 +27,7 @@ def colo_layernorm(types, args=(), kwargs=None, pg=None):
eps = kwargs['eps']
if isinstance(input_tensor, ColoTensor):
if input_tensor.is_activation() and not input_tensor.is_gathered():
if not input_tensor.is_gathered():
input_tensor.gather()
input_tensor = input_tensor.torch_tensor()
if isinstance(weight, ColoTensor):

10
colossalai/tensor/_ops/linear.py
View File

@@ -9,8 +9,8 @@ from packaging import version
from colossalai.tensor import ComputePattern, TensorSpec, ComputePattern, ParallelAction, ColoTensor, ShardPattern
def colo_linear_1Drow(input_tensor: ColoTensor, weight: ColoTensor, bias: ColoTensor) -> ColoTensor:
parallel_action = weight.shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DRow)
def colo_linear_1Drow(input_tensor: ColoTensor, weight: ColoTensor, bias:ColoTensor) -> ColoTensor:
parallel_action = weight.shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DRow_Linear)
# Input:S[1] x Weight:S[0] = Output:P
# All-Reduce(Output) + bias = res
# Input:S[1]
@@ -47,7 +47,7 @@ def colo_linear_1Dcol(input_tensor: ColoTensor, weight: ColoTensor, bias: ColoTe
# Input:B x Weight:S[1] + Bias:S[1] = Output:S[1]
# All-Gather(Output)
# Input:B
parallel_action = weight.shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DCol)
parallel_action = weight.shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DCol_Linear)
if input_tensor.is_gathered():
# Not splited yet.
assert input_tensor.shape[-1] == weight.size(-1), \
@@ -108,9 +108,9 @@ def colo_linear(types, args, kwargs, pg):
return ColoTensor.init_from_torch_tensor(torch.nn.functional.linear(input_tensor, weight, bias))
elif weight.shard_spec.num_action == 1: # Single Model Parallel Applied
compute_patterns = weight.shard_spec.compute_patterns
if ComputePattern.TP1DRow in compute_patterns:
if ComputePattern.TP1DRow_Linear in compute_patterns:
return colo_linear_1Drow(input_tensor, weight, bias)
elif ComputePattern.TP1DCol in compute_patterns:
elif ComputePattern.TP1DCol_Linear in compute_patterns:
return colo_linear_1Dcol(input_tensor, weight, bias)
else:
raise NotImplementedError

21
colossalai/tensor/colo_tensor.py
View File

@@ -142,14 +142,19 @@ class ColoTensor(object):
if self._shard_pattern is not ShardPattern.NA: # reshard
self.gather()
# Model Parameters
if ComputePattern.TP1DRow in self._shard_spec.compute_patterns:
parallel_action = self._shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DRow)
self._shard_1d(parallel_action=parallel_action, dim=-1)
self._shard_pattern = ShardPattern.Col # We bind our ComputePattern on weight, which has to be transposed when linear().
elif ComputePattern.TP1DCol in self._shard_spec.compute_patterns:
parallel_action = self._shard_spec.get_action_by_compute_pattern(ComputePattern.TP1DCol)
self._shard_1d(parallel_action=parallel_action, dim=0)
self._shard_pattern = ShardPattern.Row
if self._shard_spec.num_action == 1:
parallel_action = self._shard_spec.get_action_by_compute_pattern(
self._shard_spec.compute_patterns[0])
if parallel_action.compute_pattern in [ComputePattern.TP1DRow_Linear, \
ComputePattern.TP1DCol_Embedding]:
self._shard_1d(parallel_action=parallel_action, dim=-1)
self._shard_pattern = ShardPattern.Col # We bind our ComputePattern on weight, which has to be transposed when linear().
elif parallel_action.compute_pattern in [ComputePattern.TP1DCol_Linear, \
ComputePattern.TP1DRow_Embedding]:
self._shard_1d(parallel_action=parallel_action, dim=0)
self._shard_pattern = ShardPattern.Row
else:
raise NotImplementedError
def gather(self):
assert self.is_activation(), 'Currently we only support gather Activation ColoTensor.'

16
colossalai/tensor/spec.py
View File

@@ -4,10 +4,12 @@ from colossalai.context.parallel_mode import ParallelMode
class ComputePattern(Enum):
TP1DRow = 1
TP1DCol = 2
ZeRO = 3
DP = 4
TP1DRow_Linear = 1
TP1DCol_Linear = 2
TP1DRow_Embedding = 3
TP1DCol_Embedding = 4
ZeRO = 5
DP = 6
class ShardPattern(Enum):
@@ -43,14 +45,14 @@ class TensorSpec(object):
# using ZeRO with DP-degree = 4 and 1DRowTP with TP-degree = 2.
# parallel_action_list = [
# ParallelAction(10, ComputePattern.ZeRO, gpc.get_group(ParallelMode.DATA)),
# ParallelAction(1, ComputePattern.TP1DRow, gpc.get_group(ParallelMode.PARALLEL_1D))
# ParallelAction(1, ComputePattern.TP1DRow_Linear, gpc.get_group(ParallelMode.PARALLEL_1D))
# ]
# When the ColoTensor is initialized,
# we first splitting tensor according to ParallelAction of ZeRO,
# then splitting tensor according to ParallelAction of TP1DRow.
# then splitting tensor according to ParallelAction of TP1DRow_Linear.
# During Linear computation
# Before Linear Op, we gather the tensors according to ZeRO.
# We perform Linear Op according to compute pattern of TP1DRow.
# We perform Linear Op according to compute pattern of TP1DRow_Linear.
# After Linear Op, we split the tensors according to ZeRO.
def __init__(self, parallel_action_list: List[ParallelAction] = [], shard_pattern: ShardPattern = ShardPattern.NA):