ColossalAI/colossalai/legacy/nn/loss/__init__.py

42 lines
1.5 KiB
Python

from torch import nn
from torch.nn.modules.loss import *
from torch.nn.modules.loss import _Loss
from colossalai.global_variables import tensor_parallel_env as env
from colossalai.legacy.nn.layer.utils import get_tensor_parallel_mode
from .loss_1d import VocabParallelCrossEntropyLoss1D
from .loss_2d import CrossEntropyLoss2D, VocabParallelCrossEntropyLoss2D
from .loss_2p5d import CrossEntropyLoss2p5D, VocabParallelCrossEntropyLoss2p5D
from .loss_3d import CrossEntropyLoss3D, VocabParallelCrossEntropyLoss3D
_parallel_cross_entropy = {
'2d': CrossEntropyLoss2D,
'2.5d': CrossEntropyLoss2p5D,
'3d': CrossEntropyLoss3D,
}
_vocab_parallel_cross_entropy = {
'1d': VocabParallelCrossEntropyLoss1D,
'2d': VocabParallelCrossEntropyLoss2D,
'2.5d': VocabParallelCrossEntropyLoss2p5D,
'3d': VocabParallelCrossEntropyLoss3D,
}
class CrossEntropyLoss(_Loss):
def __init__(self, reduction: bool = True, *args, **kwargs):
super().__init__()
tensor_parallel = get_tensor_parallel_mode()
if tensor_parallel is not None and env.vocab_parallel:
self.loss = _vocab_parallel_cross_entropy[tensor_parallel](reduction=reduction, *args, **kwargs)
elif tensor_parallel is None or tensor_parallel == '1d':
reduction = 'mean' if reduction else 'none'
self.loss = nn.CrossEntropyLoss(reduction=reduction, *args, **kwargs)
else:
self.loss = _parallel_cross_entropy[tensor_parallel](reduction=reduction, *args, **kwargs)
def forward(self, *args):
return self.loss(*args)