File size: 6,238 Bytes
2e3ebcb eefe43c 2e3ebcb eefe43c 2e3ebcb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 |
# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mlp.py
# Commit id: c3b219665292c61a51153d0ded4473c494296382
# Copyright (c) 2023, Tri Dao.
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributed import ProcessGroup
try:
from flash_attn.ops.activations import swiglu
except ImportError:
swiglu = None
try:
from flash_attn.ops.fused_dense import ColumnParallelLinear, RowParallelLinear
except ImportError:
ColumnParallelLinear, RowParallelLinear = None, None
try:
from flash_attn.ops.fused_dense import FusedMLP, ParallelFusedMLP
except ImportError:
FusedMLP, ParallelFusedMLP = None, None
class Mlp(nn.Module):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
activation=F.gelu,
bias1=True,
bias2=True,
return_residual=False,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
out_features = out_features if out_features is not None else in_features
hidden_features = hidden_features if hidden_features is not None else in_features * 4
self.return_residual = return_residual
self.fc1 = nn.Linear(in_features, hidden_features, bias=bias1, **factory_kwargs)
self.activation = activation
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
def forward(self, x):
y = self.fc1(x, task='clustering')
y = self.activation(y)
y = self.fc2(y, task='sts')
return y if not self.return_residual else (y, x)
class ParallelMLP(nn.Module):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
activation=F.gelu,
process_group: ProcessGroup = None,
sequence_parallel=True,
bias1=True,
bias2=True,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
assert ColumnParallelLinear is not None, "Need to install fused_dense"
assert RowParallelLinear is not None, "Need to install fused_dense"
out_features = out_features if out_features is not None else in_features
hidden_features = hidden_features if hidden_features is not None else in_features * 4
self.fc1 = ColumnParallelLinear(
in_features,
hidden_features,
process_group,
bias=bias1,
sequence_parallel=sequence_parallel,
**factory_kwargs,
)
self.activation = activation
self.fc2 = RowParallelLinear(
hidden_features,
out_features,
process_group,
bias=bias2,
sequence_parallel=sequence_parallel,
**factory_kwargs,
)
def forward(self, x):
y = self.fc1(x)
y = self.activation(y)
y = self.fc2(y)
return y
class GatedMlp(nn.Module):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
activation=F.sigmoid,
bias1=True,
bias2=True,
multiple_of=128,
return_residual=False,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
out_features = out_features if out_features is not None else in_features
hidden_features = (
hidden_features if hidden_features is not None else int(8 * in_features / 3)
)
hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
self.return_residual = return_residual
self.fc1 = nn.Linear(in_features, 2 * hidden_features, bias=bias1, **factory_kwargs)
self.activation = activation
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
def forward(self, x):
y = self.fc1(x)
if self.activation == F.sigmoid: # Special case for GLU
y = F.glu(y, dim=-1)
elif self.activation == F.silu and swiglu is not None: # Special case for SwiGLU
y, gate = y.chunk(2, dim=-1)
y = swiglu(gate, y)
else:
y, gate = y.chunk(2, dim=-1)
y = y * self.activation(gate)
y = self.fc2(y)
return y if not self.return_residual else (y, x)
class ParallelGatedMlp(nn.Module):
"""Parallel GatedMlp"""
def __init__(
self,
in_features,
process_group,
hidden_features=None,
out_features=None,
activation=F.sigmoid,
bias1=True,
bias2=True,
multiple_of=128,
sequence_parallel=True,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
out_features = out_features if out_features is not None else in_features
hidden_features = (
hidden_features if hidden_features is not None else int(8 * in_features / 3)
)
hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
if ColumnParallelLinear is None or RowParallelLinear is None:
raise ImportError("fused_dense is not installed")
self.fc1 = ColumnParallelLinear(
in_features,
2 * hidden_features,
process_group,
bias=bias1,
sequence_parallel=sequence_parallel,
**factory_kwargs,
)
self.activation = activation
self.fc2 = RowParallelLinear(
hidden_features,
out_features,
process_group,
bias=bias2,
sequence_parallel=sequence_parallel,
**factory_kwargs,
)
def forward(self, x):
y = self.fc1(x)
if self.activation == F.sigmoid: # Special case for GLU
y = F.glu(y, dim=-1)
else:
y, gate = y.chunk(2, dim=-1)
y = y * self.activation(gate)
y = self.fc2(y)
return y |