File size: 11,321 Bytes
a5d67fc | 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 | from typing import Optional, Type
import torch
import torch.nn as nn
import torch.nn.functional as F
def make_attn_mask(query_pad: torch.Tensor, key_pad: torch.Tensor) -> torch.Tensor:
"""
Build an additive attention mask of shape (B, Q, K) from
query/key padding masks.
Args:
query_pad: (B, Q) bool or 0/1 tensor. 1/True = padded query position.
key_pad: (B, K) bool or 0/1 tensor. 1/True = padded key position.
Returns:
attn_mask: (B, Q, K) float tensor, where masked positions are -inf
and valid positions are 0.0 (for use with SDPA).
"""
# Ensure boolean
q_pad = query_pad.bool() # (B, Q)
k_pad = key_pad.bool() # (B, K)
# A position (q, k) is invalid if *either* the query or key is padded
# Shape: (B, Q, K)
pad = q_pad.unsqueeze(-1) | k_pad.unsqueeze(-2)
# Build float mask with -inf on padded positions, 0 elsewhere
attn_mask = torch.zeros_like(pad, dtype=torch.float32)
attn_mask.masked_fill_(pad, float("-inf"))
return attn_mask
class MLP(nn.Module):
def __init__(
self,
in_dim,
out_dim,
hidden_dim=256,
num_hidden_layers=1,
dropout=0,
norm=False,
activation=nn.GELU(approximate="tanh"),
output_activation=nn.Identity(),
norm_layer=nn.LayerNorm,
):
super().__init__()
layers = []
layers.append(nn.Linear(in_dim, hidden_dim))
# layers.append(norm_layer(hidden_dim) if norm else nn.Identity())
layers.append(activation)
for _ in range(num_hidden_layers - 1):
layers.append(nn.Dropout(dropout))
layers.append(norm_layer(hidden_dim) if norm else nn.Identity())
layers.append(nn.Linear(hidden_dim, hidden_dim))
layers.append(activation)
layers.append(nn.Dropout(dropout))
layers.append(norm_layer(hidden_dim) if norm else nn.Identity())
layers.append(nn.Linear(hidden_dim, out_dim))
layers.append(output_activation)
self.layers = nn.Sequential(*layers)
# self.init_weights()
def forward(self, x):
return self.layers(x)
class SwiGLU(nn.Module):
def __init__(self, in_dim, out_dim, hidden_dim=384, dropout=0):
super().__init__()
hidden_dim = round(hidden_dim * 2 / 3)
self.fc1 = nn.Linear(in_dim, hidden_dim)
self.fc2 = nn.Linear(in_dim, hidden_dim)
self.fc3 = nn.Linear(hidden_dim, out_dim)
self.activation = nn.SiLU()
self.dropout = nn.Dropout(dropout)
def forward(self, x):
x = self.fc1(x) * self.activation(self.fc2(x))
return self.dropout(self.fc3(x))
class Attention(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = False,
qk_norm: bool = False,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
norm_layer: Type[nn.Module] = nn.LayerNorm,
) -> None:
super().__init__()
assert dim % num_heads == 0, "dim should be divisible by num_heads"
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim**-0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x: torch.Tensor, attn_mask: torch.Tensor | None = None) -> torch.Tensor:
if x.ndim == 3:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv.unbind(0) # (B, num_heads, N, head_dim)
q, k = self.q_norm(q), self.k_norm(k)
x = F.scaled_dot_product_attention(
q,
k,
v,
dropout_p=self.attn_drop.p if self.training else 0.0,
attn_mask=attn_mask,
)
x = x.transpose(1, 2).reshape(B, N, C)
elif x.ndim == 4:
B, M, N, C = x.shape
qkv = self.qkv(x).reshape(B, M, N, 3, self.num_heads, self.head_dim).permute(3, 0, 4, 1, 2, 5)
q, k, v = qkv.unbind(0) # (B, num_heads, M, N, head_dim)
q, k = self.q_norm(q), self.k_norm(k)
# print('q', q.shape, 'k', k.shape, 'v', v.shape, 'attn_mask', attn_mask.shape if attn_mask is not None else "None")
x = F.scaled_dot_product_attention(
q,
k,
v,
dropout_p=self.attn_drop.p if self.training else 0.0,
attn_mask=attn_mask.unsqueeze(1) if attn_mask is not None else None,
)
x = x.permute(0, 2, 3, 1, 4).reshape(B, M, N, C)
else:
raise ValueError(f"Unsupported input dimension: {x.ndim}")
x = self.proj(x)
x = self.proj_drop(x)
return x
class CrossAttention(nn.Module):
def __init__(
self,
q_dim: int, # dim of x
kv_dim: Optional[int] = None, # dim of m (defaults to q_dim)
num_heads: int = 8,
qkv_bias: bool = False,
qk_norm: bool = False,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
norm_layer: Type[nn.Module] = nn.LayerNorm,
) -> None:
super().__init__()
kv_dim = kv_dim if kv_dim is not None else q_dim
assert q_dim % num_heads == 0, "q_dim must be divisible by num_heads"
self.num_heads = num_heads
self.head_dim = q_dim // num_heads
self.q = nn.Linear(q_dim, q_dim, bias=qkv_bias)
self.kv = nn.Linear(kv_dim, 2 * q_dim, bias=qkv_bias) # produce k and v in the SAME head dim as q
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(q_dim, q_dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
def forward(
self,
x: torch.Tensor, # (B, Nq, q_dim)
m: torch.Tensor, # (B, Nk, kv_dim)
attn_mask: Optional[torch.Tensor] = None, # broadcastable to (B, num_heads, Nq, Nk) or (Nq, Nk)
is_causal: bool = False,
) -> torch.Tensor:
if x.ndim == 3:
B, Nq, Cq = x.shape
_, Nk, _ = m.shape
q = self.q(x).reshape(B, Nq, self.num_heads, self.head_dim).permute(0, 2, 1, 3) # (B, H, Nq, Hd)
kv = self.kv(m).reshape(B, Nk, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
k, v = kv.unbind(0) # (B, H, Nk, Hd)
q, k = self.q_norm(q), self.k_norm(k)
x = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=attn_mask,
dropout_p=self.attn_drop.p if self.training else 0.0,
is_causal=is_causal,
) # (B, H, Nq, Hd)
x = x.transpose(1, 2).reshape(B, Nq, Cq) # back to (B, Nq, q_dim)
elif x.ndim == 4:
B, M, Nq, Cq = x.shape
_, Nk, _ = m.shape
q = self.q(x).reshape(B, M, Nq, self.num_heads, self.head_dim).permute(0, 3, 1, 2, 4) # (B, H, M, Nq, Hd)
kv = self.kv(m).reshape(B, Nk, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
k, v = kv.unbind(0) # (B, H, Nk, Hd)
q, k = self.q_norm(q), self.k_norm(k)
x = F.scaled_dot_product_attention(
q,
k.unsqueeze(2),
v.unsqueeze(2),
attn_mask=attn_mask.unsqueeze(1) if attn_mask is not None else None,
dropout_p=self.attn_drop.p if self.training else 0.0,
is_causal=is_causal,
) # (B, H, M, Nq, Hd)
x = x.permute(0, 2, 3, 1, 4).reshape(B, M, Nq, Cq)
else:
raise ValueError(f"Unsupported input dimension: {x.ndim}")
x = self.proj_drop(self.proj(x))
return x
class TransformerBlock(nn.Module):
"""
A standard Transformer block.
"""
def __init__(
self,
d_model,
num_heads,
mlp_ratio=4.0,
dropout=0.1,
norm_first=True,
norm_layer=nn.LayerNorm,
mlp_type="mlp",
):
super().__init__()
self.norm_first = norm_first
self.norm1 = norm_layer(d_model, elementwise_affine=True, eps=1e-6)
self.attn = Attention(d_model, num_heads, qkv_bias=True, attn_drop=dropout, proj_drop=dropout)
self.norm2 = norm_layer(d_model, elementwise_affine=True, eps=1e-6)
if mlp_type == "swiglu":
self.mlp = SwiGLU(d_model, d_model, hidden_dim=int(mlp_ratio * d_model), dropout=dropout)
elif mlp_type == "mlp":
self.mlp = MLP(
in_dim=d_model,
out_dim=d_model,
hidden_dim=int(mlp_ratio * d_model),
dropout=dropout,
)
else:
raise ValueError(f"Unsupported MLP type: {mlp_type}")
self.dropout = nn.Dropout(dropout)
def forward(self, x, attn_mask=None):
if self.norm_first:
x = x + self.attn(self.norm1(x), attn_mask)
x = x + self.dropout(self.mlp(self.norm2(x)))
else:
x = self.norm1(x + self.attn(x, attn_mask))
x = self.norm2(x + self.dropout(self.mlp(x)))
return x
class TransformerBlockCrossAttention(nn.Module):
def __init__(
self,
d_model,
num_heads,
d_cond=None,
mlp_ratio=4.0,
dropout=0.1,
norm_first=True,
norm_layer=nn.LayerNorm,
mlp_type="mlp",
):
super().__init__()
d_cond = d_cond if d_cond is not None else d_model
self.norm_first = norm_first
self.norm1 = norm_layer(d_model, elementwise_affine=True, eps=1e-6)
self.attn = CrossAttention(
d_model,
d_cond,
num_heads,
qkv_bias=True,
attn_drop=dropout,
proj_drop=dropout,
)
self.norm2 = norm_layer(d_model, elementwise_affine=True, eps=1e-6)
if mlp_type == "swiglu":
self.mlp = SwiGLU(d_model, d_model, hidden_dim=int(mlp_ratio * d_model), dropout=dropout)
elif mlp_type == "mlp":
self.mlp = MLP(
in_dim=d_model,
out_dim=d_model,
hidden_dim=int(mlp_ratio * d_model),
dropout=dropout,
)
else:
raise ValueError(f"Unsupported MLP type: {mlp_type}")
self.dropout = nn.Dropout(dropout)
def forward(self, x, m, attn_mask=None):
if self.norm_first:
x = x + self.attn(self.norm1(x), m, attn_mask)
x = x + self.dropout(self.mlp(self.norm2(x)))
else:
x = self.norm1(x + self.attn(x, m, attn_mask))
x = self.norm2(x + self.dropout(self.mlp(x)))
return x
|