Spaces:
Sleeping
Sleeping
File size: 4,162 Bytes
e45d058 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 |
# Copyright (c) 2022, Tri Dao.
# Adapted from https://github.com/NVIDIA/apex/blob/master/apex/contrib/layer_norm/layer_norm.py
import torch
from torch.nn import init
from flash_attn.ops.layer_norm import (
DropoutAddLayerNormFn,
DropoutAddLayerNormParallelResidualFn,
DropoutAddLayerNormSubsetFn,
)
def rms_norm(x, weight, epsilon):
return DropoutAddLayerNormFn.apply(
x, None, weight, None, None, None, 0.0, epsilon, False, False, True
)
def dropout_add_rms_norm(
x0,
residual,
weight,
bias,
dropout_p,
epsilon,
rowscale=None,
layerscale=None,
prenorm=False,
residual_in_fp32=False,
return_dropout_mask=False,
):
"""residual_in_fp32 only has an effect if residual is None.
Otherwise residual dtype is residual.dtype.
"""
return DropoutAddLayerNormFn.apply(
x0,
residual,
weight,
bias,
rowscale,
layerscale,
dropout_p,
epsilon,
residual_in_fp32,
prenorm,
True,
return_dropout_mask,
)
def dropout_add_rms_norm_subset(
x0,
residual,
weight,
bias,
dropout_p,
epsilon,
layerscale=None,
x0_subset=None,
out_subset=None,
rowscale_const=1.0,
out_numrows=0,
prenorm=False,
residual_in_fp32=False,
return_dropout_mask=False,
):
"""residual_in_fp32 only has an effect if residual is None.
Otherwise residual dtype is residual.dtype.
"""
return DropoutAddLayerNormSubsetFn.apply(
x0,
residual,
weight,
bias,
layerscale,
x0_subset,
out_subset,
dropout_p,
epsilon,
rowscale_const,
out_numrows,
residual_in_fp32,
prenorm,
True,
return_dropout_mask,
)
def dropout_add_rms_norm_parallel_residual(
x0,
x1,
residual,
weight0,
bias0,
weight1,
bias1,
dropout_p,
epsilon,
prenorm=False,
residual_in_fp32=False,
return_dropout_mask=False,
):
"""residual_in_fp32 only has an effect if residual is None.
Otherwise residual dtype is residual.dtype.
"""
return DropoutAddLayerNormParallelResidualFn.apply(
x0,
x1,
residual,
weight0,
bias0,
weight1,
bias1,
dropout_p,
epsilon,
residual_in_fp32,
prenorm,
True,
return_dropout_mask,
)
class RMSNorm(torch.nn.Module):
def __init__(self, hidden_size, eps=1e-5, device=None, dtype=None):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.eps = eps
self.weight = torch.nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
self.register_parameter("bias", None)
self.reset_parameters()
def reset_parameters(self):
init.ones_(self.weight)
def forward(self, x):
return rms_norm(x, self.weight, self.eps)
class DropoutAddRMSNorm(torch.nn.Module):
def __init__(
self,
hidden_size,
prenorm=False,
p=0.0,
eps=1e-5,
residual_in_fp32=False,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.prenorm = prenorm
self.p = p
self.eps = eps
self.residual_in_fp32 = residual_in_fp32
self.weight = torch.nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
self.register_parameter("bias", None)
self.reset_parameters()
def reset_parameters(self):
init.ones_(self.weight)
def forward(self, x0, residual=None):
return dropout_add_rms_norm(
x0,
residual,
self.weight,
None,
self.p if self.training else 0.0,
self.eps,
prenorm=self.prenorm,
residual_in_fp32=self.residual_in_fp32,
)
|