File size: 9,782 Bytes
0bd8c04 |
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 |
from functools import lru_cache
from typing import Optional
import torch
import torch.nn as nn
from torch.nn import RMSNorm
from torch.nn import functional as F
def drop_path(
x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
'survival rate' as the argument.
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (
x.ndim - 1
) # work with diff dim tensors, not just 2D ConvNets
random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
if keep_prob > 0.0 and scale_by_keep:
random_tensor.div_(keep_prob)
return x * random_tensor
class DropPath(torch.nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
self.scale_by_keep = scale_by_keep
def forward(self, x):
return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
def extra_repr(self):
return f"drop_prob={round(self.drop_prob,3):0.3f}"
def find_multiple(n: int, k: int):
if n % k == 0:
return n
return n + k - (n % k)
@lru_cache(maxsize=16)
def get_causal_mask(seq_q, seq_k, device):
offset = seq_k - seq_q
i = torch.arange(seq_q, device=device).unsqueeze(1)
j = torch.arange(seq_k, device=device).unsqueeze(0)
causal_mask = (j > (offset + i)).bool()
causal_mask = causal_mask.unsqueeze(0).unsqueeze(0)
return causal_mask
class Attention(nn.Module):
def __init__(
self,
dim,
n_head,
attn_dropout_p,
resid_dropout_p,
# causal: bool = True,
):
super().__init__()
assert dim % n_head == 0
self.dim = dim
self.head_dim = dim // n_head
self.scale = self.head_dim**-0.5
self.n_head = n_head
total_kv_dim = (self.n_head * 3) * self.head_dim
self.wqkv = nn.Linear(dim, total_kv_dim, bias=False)
self.wo = nn.Linear(dim, dim, bias=False)
self.attn_dropout_p = attn_dropout_p
self.resid_dropout = nn.Dropout(resid_dropout_p)
# self.causal = causal
self.k_cache = None
self.v_cache = None
self.kv_cache_size = None
def enable_kv_cache(self, bsz, max_seq_len):
if self.kv_cache_size != (bsz, max_seq_len):
device = self.wo.weight.device
dtype = self.wo.weight.dtype
self.k_cache = torch.zeros(
(bsz, self.n_head, max_seq_len, self.head_dim),
device=device,
dtype=dtype,
)
self.v_cache = torch.zeros(
(bsz, self.n_head, max_seq_len, self.head_dim),
device=device,
dtype=dtype,
)
self.kv_cache_size = (bsz, max_seq_len)
def update_kv_cache(
self, start_pos, end_pos, keys: torch.Tensor, values: torch.Tensor
):
self.k_cache[:, :, start_pos:end_pos, :] = keys
self.v_cache[:, :, start_pos:end_pos, :] = values
return (
self.k_cache[:, :, :end_pos, :],
self.v_cache[:, :, :end_pos, :],
)
def naive_attention(self, xq, keys, values, mask):
xq = xq * self.scale
# q: [B, H, 1, D], k: [B, H, D, L] -> attn [B, H, 1, L]
attn = xq @ keys.transpose(-1, -2)
seq_q, seq_k = attn.shape[-2], attn.shape[-1]
if seq_q > 1:
# causal_mask = get_causal_mask(seq_q, seq_k, attn.device)
# attn.masked_fill_(mask, float("-inf"))
attn = attn + mask
attn = torch.softmax(attn, dim=-1)
if self.attn_dropout_p > 0 and self.training:
attn = F.dropout(attn, p=self.attn_dropout_p, training=self.training)
# [B, H, 1, L] @ [B, H, L, D] -> [B, H, 1, D]
return attn @ values
def forward(
self,
x: torch.Tensor,
mask: torch.Tensor,
freqs_cis: torch.Tensor = None,
start_pos: Optional[int] = None,
end_pos: Optional[int] = None,
):
bsz, seqlen, _ = x.shape
xq, xk, xv = self.wqkv(x).chunk(3, dim=-1)
xq = xq.view(bsz, seqlen, self.n_head, self.head_dim)
xk = xk.view(bsz, seqlen, self.n_head, self.head_dim)
xv = xv.view(bsz, seqlen, self.n_head, self.head_dim)
if freqs_cis is not None:
xq = apply_rotary_emb(xq, freqs_cis)
xk = apply_rotary_emb(xk, freqs_cis)
# is_causal = self.causal
if self.k_cache is not None and start_pos is not None:
xq, xk, xv = map(lambda x: x.transpose(1, 2), (xq, xk, xv))
keys, values = self.update_kv_cache(start_pos, end_pos, xk, xv)
output = self.naive_attention(xq, keys, values, mask)
output = output.transpose(1, 2).contiguous()
else:
xq, xk, xv = map(lambda x: x.transpose(1, 2), (xq, xk, xv))
output = F.scaled_dot_product_attention(xq, xk, xv, attn_mask=mask, is_causal=False)
output = output.transpose(1, 2).contiguous()
output = output.view(bsz, seqlen, self.dim)
output = self.resid_dropout(self.wo(output))
return output
class FeedForward(nn.Module):
def __init__(self, dim, dropout_p=0.1, mlp_ratio=4.0):
super().__init__()
hidden_dim = mlp_ratio * dim
hidden_dim = int(2 * hidden_dim / 3)
hidden_dim = find_multiple(hidden_dim, 256)
self.w1 = nn.Linear(dim, hidden_dim * 2, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
self.ffn_dropout = nn.Dropout(dropout_p)
def forward(self, x):
h1, h2 = self.w1(x).chunk(2, dim=-1)
return self.ffn_dropout(self.w2(F.silu(h1) * h2))
class TransformerBlock(nn.Module):
def __init__(
self,
dim,
n_head,
attn_dropout_p: float = 0.0,
resid_dropout_p: float = 0.0,
drop_path: float = 0.0,
# causal: bool = True,
):
super().__init__()
self.attention = Attention(
dim=dim,
n_head=n_head,
attn_dropout_p=attn_dropout_p,
resid_dropout_p=resid_dropout_p,
# causal=causal,
)
self.feed_forward = FeedForward(
dim=dim,
dropout_p=resid_dropout_p,
)
self.attention_norm = RMSNorm(dim, eps=1e-6)
self.ffn_norm = RMSNorm(dim, eps=1e-6)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
def forward(
self,
x: torch.Tensor,
mask: torch.Tensor,
freqs_cis: torch.Tensor,
):
h = x + self.drop_path(self.attention(self.attention_norm(x), mask, freqs_cis))
out = h + self.drop_path(self.feed_forward(self.ffn_norm(h)))
return out
def forward_onestep(
self,
x: torch.Tensor,
mask: torch.Tensor,
freqs_cis: torch.Tensor,
start_pos: int,
end_pos: int,
):
h = x + self.drop_path(
self.attention(self.attention_norm(x), mask, freqs_cis, start_pos, end_pos)
)
out = h + self.drop_path(self.feed_forward(self.ffn_norm(h)))
return out
def get_2d_pos(resolution, patch_size, num_scales=1):
max_pos = resolution // patch_size
coords_list = []
for i in range(num_scales):
scale = 2 ** (num_scales - i - 1)
P = max(resolution // scale // patch_size, 1)
edge = float(max_pos) / P
centers = (torch.arange(P, dtype=torch.float32) + 0.5) * edge
grid_y, grid_x = torch.meshgrid(centers, centers, indexing="ij")
coords = torch.stack([grid_x.reshape(-1), grid_y.reshape(-1)], dim=1)
coords_list.append(coords)
return torch.cat(coords_list, dim=0)
def precompute_freqs_cis_2d(
pos_2d, n_elem: int, base: float = 10000, cls_token_num=120
):
# split the dimension into half, one for x and one for y
half_dim = n_elem // 2
freqs = 1.0 / (
base ** (torch.arange(0, half_dim, 2)[: (half_dim // 2)].float() / half_dim)
)
t = pos_2d + 1.0
if cls_token_num > 0:
t = torch.cat(
[torch.zeros((cls_token_num, 2), device=freqs.device), t],
dim=0,
)
freqs = torch.outer(t.flatten(), freqs).view(*t.shape[:-1], -1)
return torch.stack([torch.cos(freqs), torch.sin(freqs)], dim=-1)
def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor):
# x: (bs, seq_len, n_head, head_dim)
# freqs_cis (seq_len, head_dim // 2, 2)
xshaped = x.float().reshape(
*x.shape[:-1], -1, 2
) # (bs, seq_len, n_head, head_dim//2, 2)
freqs_cis = freqs_cis.view(
1, xshaped.size(1), 1, xshaped.size(3), 2
) # (1, seq_len, 1, head_dim//2, 2)
x_out2 = torch.stack(
[
xshaped[..., 0] * freqs_cis[..., 0] - xshaped[..., 1] * freqs_cis[..., 1],
xshaped[..., 1] * freqs_cis[..., 0] + xshaped[..., 0] * freqs_cis[..., 1],
],
dim=-1,
)
x_out2 = x_out2.flatten(3)
return x_out2.type_as(x)
|