File size: 10,200 Bytes
714cf46 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 | from typing import Optional
import torch
from torch import Tensor, nn
from . import vb_const as const
from .vb_layers_attention import AttentionPairBias
from .vb_layers_attentionv2 import AttentionPairBias as AttentionPairBiasV2
from .vb_layers_dropout import get_dropout_mask
from .vb_layers_transition import Transition
from .vb_tri_attn_attention import (
TriangleAttentionEndingNode,
TriangleAttentionStartingNode,
)
from .vb_layers_triangular_mult import (
TriangleMultiplicationIncoming,
TriangleMultiplicationOutgoing,
)
class PairformerLayer(nn.Module):
"""Pairformer module."""
def __init__(
self,
token_s: int,
token_z: int,
num_heads: int = 16,
dropout: float = 0.25,
pairwise_head_width: int = 32,
pairwise_num_heads: int = 4,
post_layer_norm: bool = False,
v2: bool = False,
) -> None:
super().__init__()
self.token_z = token_z
self.dropout = dropout
self.num_heads = num_heads
self.post_layer_norm = post_layer_norm
self.pre_norm_s = nn.LayerNorm(token_s)
if v2:
self.attention = AttentionPairBiasV2(token_s, token_z, num_heads)
else:
self.attention = AttentionPairBias(token_s, token_z, num_heads)
self.tri_mul_out = TriangleMultiplicationOutgoing(token_z)
self.tri_mul_in = TriangleMultiplicationIncoming(token_z)
self.tri_att_start = TriangleAttentionStartingNode(
token_z, pairwise_head_width, pairwise_num_heads, inf=1e9
)
self.tri_att_end = TriangleAttentionEndingNode(
token_z, pairwise_head_width, pairwise_num_heads, inf=1e9
)
self.transition_s = Transition(token_s, token_s * 4)
self.transition_z = Transition(token_z, token_z * 4)
self.s_post_norm = (
nn.LayerNorm(token_s) if self.post_layer_norm else nn.Identity()
)
def forward(
self,
s: Tensor,
z: Tensor,
mask: Tensor,
pair_mask: Tensor,
chunk_size_tri_attn: Optional[int] = None,
use_kernels: bool = False,
use_cuequiv_mul: bool = False,
use_cuequiv_attn: bool = False,
) -> tuple[Tensor, Tensor]:
# Compute pairwise stack
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_mul_out(
z, mask=pair_mask, use_kernels=use_cuequiv_mul or use_kernels
)
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_mul_in(
z, mask=pair_mask, use_kernels=use_cuequiv_mul or use_kernels
)
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_att_start(
z,
mask=pair_mask,
chunk_size=chunk_size_tri_attn,
use_kernels=use_cuequiv_attn or use_kernels,
)
dropout = get_dropout_mask(self.dropout, z, self.training, columnwise=True)
z = z + dropout * self.tri_att_end(
z,
mask=pair_mask,
chunk_size=chunk_size_tri_attn,
use_kernels=use_cuequiv_attn or use_kernels,
)
z = z + self.transition_z(z)
# Compute sequence stack
with torch.autocast("cuda", enabled=False):
s_normed = self.pre_norm_s(s.float())
s = s.float() + self.attention(
s=s_normed, z=z.float(), mask=mask.float(), k_in=s_normed
)
s = s + self.transition_s(s)
s = self.s_post_norm(s)
return s, z
class PairformerModule(nn.Module):
"""Pairformer module."""
def __init__(
self,
token_s: int,
token_z: int,
num_blocks: int,
num_heads: int = 16,
dropout: float = 0.25,
pairwise_head_width: int = 32,
pairwise_num_heads: int = 4,
post_layer_norm: bool = False,
activation_checkpointing: bool = False,
v2: bool = False,
**kwargs,
) -> None:
super().__init__()
self.token_z = token_z
self.num_blocks = num_blocks
self.dropout = dropout
self.num_heads = num_heads
self.post_layer_norm = post_layer_norm
self.activation_checkpointing = activation_checkpointing
self.layers = nn.ModuleList()
for _ in range(num_blocks):
self.layers.append(
PairformerLayer(
token_s,
token_z,
num_heads,
dropout,
pairwise_head_width,
pairwise_num_heads,
post_layer_norm,
v2,
),
)
def forward(
self,
s: Tensor,
z: Tensor,
mask: Tensor,
pair_mask: Tensor,
use_kernels: bool = False,
) -> tuple[Tensor, Tensor]:
"""Perform the forward pass.
Parameters
----------
s : Tensor
The sequence stack.
z : Tensor
The pairwise stack.
mask : Tensor
The mask.
pair_mask : Tensor
The pairwise mask.
use_kernels : bool
Whether to use kernels.
"""
if not self.training:
if z.shape[1] > const.chunk_size_threshold:
chunk_size_tri_attn = 128
else:
chunk_size_tri_attn = 512
else:
chunk_size_tri_attn = None
for layer in self.layers:
if self.activation_checkpointing and self.training:
s, z = torch.utils.checkpoint.checkpoint(
layer,
s,
z,
mask,
pair_mask,
chunk_size_tri_attn,
use_kernels,
)
else:
s, z = layer(s, z, mask, pair_mask, chunk_size_tri_attn, use_kernels)
return s, z
class PairformerNoSeqLayer(nn.Module):
"""Pairformer module without sequence track."""
def __init__(
self,
token_z: int,
dropout: float = 0.25,
pairwise_head_width: int = 32,
pairwise_num_heads: int = 4,
post_layer_norm: bool = False,
) -> None:
super().__init__()
self.token_z = token_z
self.dropout = dropout
self.post_layer_norm = post_layer_norm
self.tri_mul_out = TriangleMultiplicationOutgoing(token_z)
self.tri_mul_in = TriangleMultiplicationIncoming(token_z)
self.tri_att_start = TriangleAttentionStartingNode(
token_z, pairwise_head_width, pairwise_num_heads, inf=1e9
)
self.tri_att_end = TriangleAttentionEndingNode(
token_z, pairwise_head_width, pairwise_num_heads, inf=1e9
)
self.transition_z = Transition(token_z, token_z * 4)
def forward(
self,
z: Tensor,
pair_mask: Tensor,
chunk_size_tri_attn: Optional[int] = None,
use_kernels: bool = False,
use_cuequiv_mul: bool = False,
use_cuequiv_attn: bool = False,
) -> Tensor:
# Compute pairwise stack
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_mul_out(
z, mask=pair_mask, use_kernels=use_cuequiv_mul or use_kernels
)
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_mul_in(
z, mask=pair_mask, use_kernels=use_cuequiv_mul or use_kernels
)
dropout = get_dropout_mask(self.dropout, z, self.training)
z = z + dropout * self.tri_att_start(
z,
mask=pair_mask,
chunk_size=chunk_size_tri_attn,
use_kernels=use_cuequiv_attn or use_kernels,
)
dropout = get_dropout_mask(self.dropout, z, self.training, columnwise=True)
z = z + dropout * self.tri_att_end(
z,
mask=pair_mask,
chunk_size=chunk_size_tri_attn,
use_kernels=use_cuequiv_attn or use_kernels,
)
z = z + self.transition_z(z)
return z
class PairformerNoSeqModule(nn.Module):
"""Pairformer module without sequence track."""
def __init__(
self,
token_z: int,
num_blocks: int,
dropout: float = 0.25,
pairwise_head_width: int = 32,
pairwise_num_heads: int = 4,
post_layer_norm: bool = False,
activation_checkpointing: bool = False,
**kwargs,
) -> None:
super().__init__()
self.token_z = token_z
self.num_blocks = num_blocks
self.dropout = dropout
self.post_layer_norm = post_layer_norm
self.activation_checkpointing = activation_checkpointing
self.layers = nn.ModuleList()
for i in range(num_blocks):
self.layers.append(
PairformerNoSeqLayer(
token_z,
dropout,
pairwise_head_width,
pairwise_num_heads,
post_layer_norm,
),
)
def forward(
self,
z: Tensor,
pair_mask: Tensor,
use_kernels: bool = False,
) -> Tensor:
if not self.training:
if z.shape[1] > const.chunk_size_threshold:
chunk_size_tri_attn = 128
else:
chunk_size_tri_attn = 512
else:
chunk_size_tri_attn = None
for layer in self.layers:
if self.activation_checkpointing and self.training:
z = torch.utils.checkpoint.checkpoint(
layer,
z,
pair_mask,
chunk_size_tri_attn,
use_kernels,
)
else:
z = layer(
z,
pair_mask,
chunk_size_tri_attn,
use_kernels,
)
return z
|