File size: 8,868 Bytes
522e1ad |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 |
# -*- coding: utf-8 -*-
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
from michelangelo.models.modules.checkpoint import checkpoint
def init_linear(l, stddev):
nn.init.normal_(l.weight, std=stddev)
if l.bias is not None:
nn.init.constant_(l.bias, 0.0)
class MultiheadAttention(nn.Module):
def __init__(
self,
*,
device: torch.device,
dtype: torch.dtype,
n_ctx: int,
width: int,
heads: int,
init_scale: float,
qkv_bias: bool,
flash: bool = False
):
super().__init__()
self.n_ctx = n_ctx
self.width = width
self.heads = heads
self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias, device=device, dtype=dtype)
self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
self.attention = QKVMultiheadAttention(device=device, dtype=dtype, heads=heads, n_ctx=n_ctx, flash=flash)
init_linear(self.c_qkv, init_scale)
init_linear(self.c_proj, init_scale)
def forward(self, x):
x = self.c_qkv(x)
x = checkpoint(self.attention, (x,), (), True)
x = self.c_proj(x)
return x
class QKVMultiheadAttention(nn.Module):
def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_ctx: int, flash: bool = False):
super().__init__()
self.device = device
self.dtype = dtype
self.heads = heads
self.n_ctx = n_ctx
self.flash = flash
def forward(self, qkv):
bs, n_ctx, width = qkv.shape
attn_ch = width // self.heads // 3
scale = 1 / math.sqrt(math.sqrt(attn_ch))
qkv = qkv.view(bs, n_ctx, self.heads, -1)
q, k, v = torch.split(qkv, attn_ch, dim=-1)
if self.flash:
out = F.scaled_dot_product_attention(q, k, v)
else:
weight = torch.einsum(
"bthc,bshc->bhts", q * scale, k * scale
) # More stable with f16 than dividing afterwards
wdtype = weight.dtype
weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
return out
class ResidualAttentionBlock(nn.Module):
def __init__(
self,
*,
device: torch.device,
dtype: torch.dtype,
n_ctx: int,
width: int,
heads: int,
init_scale: float = 1.0,
qkv_bias: bool = True,
flash: bool = False,
use_checkpoint: bool = False
):
super().__init__()
self.use_checkpoint = use_checkpoint
self.attn = MultiheadAttention(
device=device,
dtype=dtype,
n_ctx=n_ctx,
width=width,
heads=heads,
init_scale=init_scale,
qkv_bias=qkv_bias,
flash=flash
)
self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
self.ln_2 = nn.LayerNorm(width, device=device, dtype=dtype)
def _forward(self, x: torch.Tensor):
x = x + self.attn(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
def forward(self, x: torch.Tensor):
return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
class MultiheadCrossAttention(nn.Module):
def __init__(
self,
*,
device: torch.device,
dtype: torch.dtype,
width: int,
heads: int,
init_scale: float,
qkv_bias: bool = True,
flash: bool = False,
n_data: Optional[int] = None,
data_width: Optional[int] = None,
):
super().__init__()
self.n_data = n_data
self.width = width
self.heads = heads
self.data_width = width if data_width is None else data_width
self.c_q = nn.Linear(width, width, bias=qkv_bias, device=device, dtype=dtype)
self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias, device=device, dtype=dtype)
self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
self.attention = QKVMultiheadCrossAttention(
device=device, dtype=dtype, heads=heads, n_data=n_data, flash=flash
)
init_linear(self.c_q, init_scale)
init_linear(self.c_kv, init_scale)
init_linear(self.c_proj, init_scale)
def forward(self, x, data):
x = self.c_q(x)
data = self.c_kv(data)
x = checkpoint(self.attention, (x, data), (), True)
x = self.c_proj(x)
return x
class QKVMultiheadCrossAttention(nn.Module):
def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int,
flash: bool = False, n_data: Optional[int] = None):
super().__init__()
self.device = device
self.dtype = dtype
self.heads = heads
self.n_data = n_data
self.flash = flash
def forward(self, q, kv):
_, n_ctx, _ = q.shape
bs, n_data, width = kv.shape
attn_ch = width // self.heads // 2
scale = 1 / math.sqrt(math.sqrt(attn_ch))
q = q.view(bs, n_ctx, self.heads, -1)
kv = kv.view(bs, n_data, self.heads, -1)
k, v = torch.split(kv, attn_ch, dim=-1)
if self.flash:
out = F.scaled_dot_product_attention(q, k, v)
else:
weight = torch.einsum(
"bthc,bshc->bhts", q * scale, k * scale
) # More stable with f16 than dividing afterwards
wdtype = weight.dtype
weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
return out
class ResidualCrossAttentionBlock(nn.Module):
def __init__(
self,
*,
device: Optional[torch.device],
dtype: Optional[torch.dtype],
n_data: Optional[int] = None,
width: int,
heads: int,
data_width: Optional[int] = None,
init_scale: float = 0.25,
qkv_bias: bool = True,
flash: bool = False
):
super().__init__()
if data_width is None:
data_width = width
self.attn = MultiheadCrossAttention(
device=device,
dtype=dtype,
n_data=n_data,
width=width,
heads=heads,
data_width=data_width,
init_scale=init_scale,
qkv_bias=qkv_bias,
flash=flash,
)
self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
self.ln_2 = nn.LayerNorm(data_width, device=device, dtype=dtype)
self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
self.ln_3 = nn.LayerNorm(width, device=device, dtype=dtype)
def forward(self, x: torch.Tensor, data: torch.Tensor):
x = x + self.attn(self.ln_1(x), self.ln_2(data))
x = x + self.mlp(self.ln_3(x))
return x
class MLP(nn.Module):
def __init__(self, *,
device: Optional[torch.device],
dtype: Optional[torch.dtype],
width: int,
init_scale: float):
super().__init__()
self.width = width
self.c_fc = nn.Linear(width, width * 4, device=device, dtype=dtype)
self.c_proj = nn.Linear(width * 4, width, device=device, dtype=dtype)
self.gelu = nn.GELU()
init_linear(self.c_fc, init_scale)
init_linear(self.c_proj, init_scale)
def forward(self, x):
return self.c_proj(self.gelu(self.c_fc(x)))
class Transformer(nn.Module):
def __init__(
self,
*,
device: Optional[torch.device],
dtype: Optional[torch.dtype],
n_ctx: int,
width: int,
layers: int,
heads: int,
init_scale: float = 0.25,
qkv_bias: bool = True,
flash: bool = False,
use_checkpoint: bool = False
):
super().__init__()
self.n_ctx = n_ctx
self.width = width
self.layers = layers
self.resblocks = nn.ModuleList(
[
ResidualAttentionBlock(
device=device,
dtype=dtype,
n_ctx=n_ctx,
width=width,
heads=heads,
init_scale=init_scale,
qkv_bias=qkv_bias,
flash=flash,
use_checkpoint=use_checkpoint
)
for _ in range(layers)
]
)
def forward(self, x: torch.Tensor):
for block in self.resblocks:
x = block(x)
return x
|