mygpt.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 # This is an implementation from scratch of a "GPT", that is a model
   9 # composed of several causal self-attention blocks. It is equipped
  10 # with a caching mechanism for keys and values to avoid a O(N^3) cost
  11 # for auto-regression.
  12
  13 import math
  14
  15 import torch
  16
  17 from torch import nn
  18 from torch.nn import functional as F
  19
  20 ######################################################################
  21
  22 # A BracketedSequence is a BxTx... tensor with a first and a nb time
  23 # steps to compute.
  24
  25 # Modules able to process it expect that they will have to process a
  26 # first bracket starting at t=0, followed by a succession of brackets
  27 # that move forward in time, do not overlap, and cover the axis T with
  28 # no holes.
  29 #
  30 # Although it is more general, for a classical prompt-conditioned
  31 # auto-regressive process it will be a first bracket starting at 0 and
  32 # of arbitrary length for the "prompt", followed by brackets of length
  33 # 1 for the successive tokens.
  34 #
  35 # Modules able to process brackets may implement a cache that is
  36 # resetted when the input bracket starts at t=0
  37
  38
  39 class BracketedSequence:
  40     def __init__(self, x, first=None, nb=None):
  41         self.x = x
  42         self.first = 0 if first is None else first
  43         self.nb = x.size(1) if nb is None else nb
  44
  45     def slice(self):
  46         return self.x[:, self.first : self.first + self.nb]
  47
  48     def complete(self):
  49         return self.first == 0 and self.nb == x.size(1)
  50
  51
  52 ######################################################################
  53
  54
  55 class CacheWrapper(nn.Module):
  56     def __init__(self, *f):
  57         super().__init__()
  58         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  59
  60     def forward(self, bs):
  61         if bs.first == 0:
  62             y = self.f(bs.slice())
  63             self.cache_y = y.new(*((y.size(0), bs.x.size(1)) + y.size()[2:]))
  64             self.cache_y[:, bs.first : bs.first + bs.nb] = y
  65         else:
  66             self.cache_y[:, bs.first : bs.first + bs.nb] = self.f(bs.slice())
  67
  68         return BracketedSequence(self.cache_y, bs.first, bs.nb)
  69
  70
  71 ##############################
  72
  73
  74 class WithResidual(nn.Module):
  75     def __init__(self, *f):
  76         super().__init__()
  77         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  78
  79     def forward(self, bs):
  80         return BracketedSequence(bs.x + self.f(bs).x, bs.first, bs.nb)
  81
  82
  83 ##############################
  84
  85
  86 class AddPositionalEncoding(nn.Module):
  87     def __init__(self, len_max):
  88         super().__init__()
  89         self.len_max = len_max
  90
  91     # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
  92
  93     def forward(self, bs):
  94         if bs.first == 0:
  95             t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
  96                 :, None
  97             ]
  98             j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
  99                 None, :
 100             ]
 101             k = j % 2
 102             self.pe = torch.sin(
 103                 t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
 104             )
 105             self.cache_y = bs.x.new(bs.x.size())
 106
 107         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 108             bs.slice() + self.pe[bs.first : bs.first + bs.nb]
 109         )
 110
 111         return BracketedSequence(self.cache_y, bs.first, bs.nb)
 112
 113
 114 ##############################
 115
 116
 117 class QKVAttention(nn.Module):
 118     def __init__(
 119         self,
 120         dim_in,
 121         dim_qk,
 122         dim_v,
 123         nb_heads=1,
 124         causal=False,
 125         attention_dropout=0.0,
 126     ):
 127         super().__init__()
 128
 129         def randw(*d):
 130             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 131
 132         self.causal = causal
 133         self.attention_dropout = attention_dropout
 134         self.record_attention = False
 135
 136         self.w_q = randw(nb_heads, dim_qk, dim_in)
 137         self.w_k = randw(nb_heads, dim_qk, dim_in)
 138         self.w_v = randw(nb_heads, dim_v, dim_in)
 139         self.w_o = randw(dim_v * nb_heads, dim_in)
 140
 141     def forward(self, bs_q):
 142         x_q = bs_q.x
 143
 144         assert (
 145             self.causal or bs_q.complete()
 146         ), "Partial evaluation is only possible for causal models"
 147
 148         if bs_q.first == 0:
 149             self.cache_k = x_q.new_zeros(
 150                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 151             )
 152             self.cache_v = x_q.new_zeros(
 153                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 154             )
 155             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 156
 157         q = torch.einsum(
 158             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
 159         )
 160
 161         self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 162             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
 163         )
 164         self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 165             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
 166         )
 167
 168         a = torch.einsum(
 169             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
 170         ) / math.sqrt(self.w_q.size(1))
 171
 172         if self.record_attention:
 173             self.a = a
 174
 175         if self.causal:
 176             if bs_q.first == 0:
 177                 self.cache_attzero = (
 178                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 179                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 180                 )
 181             a = a.masked_fill(
 182                 self.cache_attzero[
 183                     :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb
 184                 ],
 185                 float("-inf"),
 186             )
 187
 188         a = a.softmax(dim=3)
 189         a = F.dropout(a, self.attention_dropout, self.training)
 190
 191         y = torch.einsum(
 192             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs_q.first + bs_q.nb]
 193         ).flatten(2)
 194
 195         self.cache_y[:, bs_q.first : bs_q.first + bs_q.nb] = y @ self.w_o
 196
 197         return BracketedSequence(self.cache_y, bs_q.first, bs_q.nb)
 198
 199
 200 ##############################
 201
 202
 203 class MyGPT(nn.Module):
 204     def __init__(
 205         self,
 206         vocabulary_size,
 207         dim_model,
 208         dim_keys,
 209         dim_hidden,
 210         nb_heads,
 211         nb_blocks,
 212         causal=False,
 213         dropout=0.0,
 214         len_max=1e5,
 215     ):
 216         super().__init__()
 217
 218         assert dim_model % nb_heads == 0
 219
 220         self.embedding = nn.Sequential(
 221             CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
 222             AddPositionalEncoding(len_max),
 223         )
 224
 225         trunk_blocks = []
 226
 227         for b in range(nb_blocks):
 228             trunk_blocks += [
 229                 WithResidual(
 230                     CacheWrapper(nn.LayerNorm((dim_model,))),
 231                     QKVAttention(
 232                         dim_in=dim_model,
 233                         dim_qk=dim_keys,
 234                         dim_v=dim_model // nb_heads,
 235                         nb_heads=nb_heads,
 236                         causal=causal,
 237                         attention_dropout=dropout,
 238                     ),
 239                 ),
 240                 WithResidual(
 241                     CacheWrapper(
 242                         nn.LayerNorm((dim_model,)),
 243                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 244                         nn.ReLU(),
 245                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 246                         nn.Dropout(dropout),
 247                     ),
 248                 ),
 249             ]
 250
 251         self.trunk = nn.Sequential(*trunk_blocks)
 252
 253         self.readout = CacheWrapper(
 254             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 255         )
 256
 257         with torch.no_grad():
 258             for m in self.modules():
 259                 if isinstance(m, nn.Embedding):
 260                     m.weight.normal_(mean=0, std=2e-2)
 261                 elif isinstance(m, nn.LayerNorm):
 262                     m.bias.zero_()
 263                     m.weight.fill_(1.0)
 264
 265     def forward(self, bs):
 266         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
 267         bs = self.embedding(bs)
 268         bs = self.trunk(bs)
 269         bs = self.readout(bs)
 270         return bs
 271
 272     # ar_mask is a tensor with 0s and 1s, of same shape as input, with
 273     # 1s where tokens should be generated. The others are kept
 274     # unchanged.
 275
 276     def masked_inplace_autoregression(
 277         self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
 278     ):
 279         to_generate = (ar_mask.sum(0) > 0).nonzero()
 280         if to_generate.min() > 0:
 281             self(
 282                 BracketedSequence(input, 0, to_generate.min())
 283             )  # Needed to initialize the model's cache
 284         for s in range(to_generate.min(), to_generate.max() + 1):
 285             output = self(BracketedSequence(input, s, 1)).x
 286             logits = output[:, s]
 287             if forbidden_tokens is not None:
 288                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 289             if deterministic_synthesis:
 290                 t_next = logits.argmax(1)
 291             else:
 292                 dist = torch.distributions.categorical.Categorical(logits=logits)
 293                 t_next = dist.sample()
 294             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 295
 296     def record_attention(self, v=True):
 297         for m in self.modules():
 298             if isinstance(m, QKVAttention):
 299                 m.record_attention = v
 300
 301     def retrieve_attention(self):
 302         a = []
 303         for m in self.modules():
 304             if isinstance(m, QKVAttention):
 305                 a.append(m.a)
 306         return a
 307
 308
 309 ######################################################################
 310
 311 if __name__ == "__main__":
 312     print("Basic check.")
 313
 314     vocabulary_size = 3
 315     x = torch.randint(vocabulary_size, (1, 5))
 316
 317     model = MyGPT(
 318         vocabulary_size=vocabulary_size,
 319         dim_model=4,
 320         dim_keys=2,
 321         dim_hidden=2,
 322         nb_heads=2,
 323         nb_blocks=2,
 324         dropout=0.1,
 325         causal=True,
 326     )
 327
 328     model.eval()
 329     y1 = model(BracketedSequence(x)).x
 330     y2 = torch.randn_like(y1)
 331     for s in range(x.size(1)):
 332         z = model(BracketedSequence(x, s, 1))
 333         y2[:, s] = z.slice()
 334
 335     print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
 336
 337 ######################################################################