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 # This implementation is equipped with RNN layers to replace the MHA
  14
  15 import math, warnings
  16
  17 import torch, einops
  18
  19 from torch import nn
  20 from torch.nn import functional as F
  21
  22 import ffutils
  23
  24 # import memload
  25
  26 ######################################################################
  27
  28 # A BracketedSequence is a BxTx... tensor with a first and a nb time
  29 # steps to compute.
  30
  31 # Modules able to process it expect that they will have to process a
  32 # first bracket starting at t=0, followed by a succession of brackets
  33 # that move forward in time, do not overlap, and cover the axis T with
  34 # no holes.
  35 #
  36 # Although it is more general, for a classical prompt-conditioned
  37 # auto-regressive process it will be a first bracket starting at 0 and
  38 # of arbitrary length for the "prompt", followed by brackets of length
  39 # 1 for the successive tokens.
  40 #
  41 # Modules able to process brackets may implement a cache that is
  42 # resetted when init_cache is True
  43
  44
  45 class BracketedSequence:
  46     def __init__(self, x, first=None, nb=None, init_cache=None):
  47         self.x = x
  48         assert (first is None and nb is None and init_cache is None) or (
  49             first is not None and nb is not None and init_cache is not None
  50         )
  51
  52         self.first = 0 if first is None else first
  53         self.nb = x.size(1) if nb is None else nb
  54         self.init_cache = True if init_cache is None else init_cache
  55
  56     def slice(self):
  57         return self.x[:, self.first : self.first + self.nb]
  58
  59     def complete(self):
  60         return self.first == 0 and self.nb == self.x.size(1)
  61
  62
  63 ######################################################################
  64
  65
  66 class CacheWrapper(nn.Module):
  67     def __init__(self, *f):
  68         super().__init__()
  69         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  70
  71     def forward(self, bs):
  72         if bs.init_cache:
  73             y = self.f(bs.slice())
  74             self.cache_y = y.new(*((y.size(0), bs.x.size(1)) + y.size()[2:]))
  75             self.cache_y[:, bs.first : bs.first + bs.nb] = y
  76         else:
  77             assert tuple(bs.x.size()[:2]) == tuple(self.cache_y.size()[:2])
  78             assert bs.first + bs.nb <= self.cache_y.size(1)
  79             self.cache_y[:, bs.first : bs.first + bs.nb] = self.f(bs.slice())
  80
  81         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
  82
  83
  84 ##############################
  85
  86
  87 class WithResidual(nn.Module):
  88     def __init__(self, *f):
  89         super().__init__()
  90         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  91
  92     def forward(self, bs):
  93         return BracketedSequence(bs.x + self.f(bs).x, bs.first, bs.nb, bs.init_cache)
  94
  95
  96 ##############################
  97
  98
  99 class AddPositionalEncoding(nn.Module):
 100     def __init__(self, len_max):
 101         super().__init__()
 102         self.len_max = len_max
 103
 104     # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
 105
 106     def forward(self, bs):
 107         if bs.init_cache:
 108             t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
 109                 :, None
 110             ]
 111             j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
 112                 None, :
 113             ]
 114             k = j % 2
 115             self.pe = torch.sin(
 116                 t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
 117             )
 118             self.cache_y = bs.x.new(bs.x.size())
 119
 120         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 121             bs.slice() + self.pe[bs.first : bs.first + bs.nb]
 122         )
 123
 124         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 125
 126
 127 import pscan
 128
 129 # X is /.../xTxD   A is /.../xT   Y_init is /.../xD
 130
 131
 132 def pscan_dim(A, X, Y_init, dim=-2):
 133     s = X.size()
 134     a, T, b = s[:dim].numel(), s[dim], s[dim + 1 :].numel()
 135
 136     A = A.reshape(a, T, *s[dim + 1 : -1])
 137     X = X.reshape(a, T, *s[dim + 1 : -1], -1)
 138
 139     if Y_init is None:
 140         Y_init = X.new_zeros(a, *s[dim + 1 : -1], X.size(-1))
 141     else:
 142         Y_init = Y_init.reshape(a, *s[dim + 1 : -1], -1)
 143
 144     Y = pscan.pscan(A, X, Y_init).reshape(s)
 145
 146     return Y
 147
 148
 149 def pscan_rgrad(grad_Y, A, X, Y_init, dim=-2, eps=1e-2):
 150     with torch.no_grad():
 151         s_A, s_X = 0, 0
 152         for t in range(X.size(dim) - 1, 0, -1):
 153             delta = (grad_Y[t] - s_A) / A[t].grad
 154             s_A += A[t].grad * delta
 155             A[t].grad = delta
 156             delta = (grad_Y[t] - s_X) / X[t].grad
 157             s_X += X[t].grad * delta
 158             X[t].grad = delta
 159
 160
 161 def pscan_shape(A, X, Y_init):
 162     s = X.size()
 163     A = A.reshape(-1, s[-2])
 164     X = X.reshape(-1, s[-2], s[-1])
 165
 166     if Y_init is None:
 167         Y_init = X.new_zeros(X.size(0), s[-1])
 168     else:
 169         Y_init = Y_init.reshape(-1, s[-1])
 170
 171     Y = pscan.pscan(A, X, Y_init).reshape(s)
 172
 173     return Y
 174
 175
 176 def nsum_shape(X, Y_init):
 177     s = X.size()
 178     X = X.reshape(-1, s[-2], s[-1])  # ntd
 179
 180     Y = 0 if Y_init is None else Y_init.reshape(-1, s[-1])
 181     result = []
 182
 183     for k in range(X.size(1)):
 184         Y = Y + X[:, k]
 185         Y = Y / Y.norm(dim=-1, keepdim=True).clamp(min=1)
 186         result.append(Y)
 187
 188     return torch.cat(result, dim=1).reshape(s)
 189
 190
 191 ##############################
 192
 193
 194 class DumbRec(nn.Module):
 195     def __init__(
 196         self,
 197         dim_model,
 198         dim_qk,
 199         dim_v,
 200         nb_heads,
 201         nb_lines,
 202         attention_dropout=0.0,
 203         len_max=1e5,
 204         logger=print,
 205         **kwargs,
 206     ):
 207         super().__init__()
 208
 209         def randw(*d):
 210             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 211
 212         self.nb_lines = nb_lines
 213         self.attention_dropout = attention_dropout
 214
 215         self.k_star = randw(nb_lines, dim_qk)
 216
 217         self.w_qw = randw(nb_heads, dim_qk, dim_model)
 218         self.w_qr = randw(nb_heads, dim_qk, dim_model)
 219         # self.w_k = randw(nb_heads, dim_qk, dim_model)
 220         self.w_v = randw(nb_heads, dim_v, dim_model)
 221         self.w_o = randw(dim_v * nb_heads, dim_model)
 222
 223     def reset_inner_loss(self):
 224         self.acc_attention = 0
 225         self.acc_nb = 0
 226
 227     def get_inner_loss(self):
 228         warnings.warn("l2 regularization", RuntimeWarning)
 229         return (self.acc_attention / self.acc_nb).pow(2).sum()
 230         # return torch.tensor([0], device=self.w_qw.device)
 231
 232     def forward(self, bs):
 233         x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
 234
 235         if bs.init_cache:
 236             self.rec_v = x_q.new_zeros(
 237                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
 238             )
 239             # self.rec_k = x_q.new_zeros(
 240             # x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
 241             # )
 242             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 243
 244         ######################################################################
 245         # Prepare the keys
 246
 247         k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
 248
 249         warnings.warn("rotating key barrel", RuntimeWarning)
 250         k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
 251         t_barrel = torch.arange(t0, t1, device=k_star.device)
 252         t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
 253         l_barrel = (
 254             torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
 255         ) % k_star.size(0)
 256         k_star = k_star[l_barrel, t_barrel]
 257
 258         ######################################################################
 259         # Compute the recurrent state
 260
 261         qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
 262
 263         v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
 264         # k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
 265
 266         aw = torch.einsum(
 267             "nhtd,ltd->nhlt",
 268             qw,
 269             k_star,
 270         ) / math.sqrt(self.w_qw.size(1))
 271
 272         aw = aw.softmax(dim=2)  # nhlt
 273
 274         if self.train:
 275             self.acc_attention += aw.sum(dim=(0, 1, 3))
 276             self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
 277
 278         aw = F.dropout(aw, self.attention_dropout, self.training)
 279
 280         A = 1 - aw.sum(dim=1)  # nlt
 281
 282         V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
 283         # K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
 284
 285         if t0 == 0:
 286             V0 = None
 287             # K0 = None
 288         else:
 289             V0 = self.rec_v[:, :, t0 - 1]
 290             # K0 = self.rec_k[:, :, t0 - 1]
 291
 292         self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
 293         # self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
 294
 295         ######################################################################
 296         # compute the readout
 297
 298         qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
 299
 300         ar = torch.einsum(
 301             "nhtd,ld->nhlt",
 302             qr,
 303             # self.rec_k[:, :, t0:t1],
 304             self.k_star,
 305         ) / math.sqrt(self.w_qr.size(1))
 306
 307         ar = ar.softmax(dim=2)  # nhlt
 308
 309         ar = F.dropout(ar, self.attention_dropout, self.training)
 310
 311         y = torch.einsum(
 312             "nhlt,nltd->nthd",
 313             ar,
 314             self.rec_v[:, :, t0:t1],
 315         ).flatten(2)
 316
 317         self.cache_y[:, t0:t1] = y @ self.w_o
 318
 319         return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
 320
 321
 322 ##############################
 323
 324
 325 class KVRec(nn.Module):
 326     def __init__(
 327         self,
 328         dim_model,
 329         dim_qk,
 330         dim_v,
 331         nb_heads,
 332         nb_lines,
 333         attention_dropout=0.0,
 334         len_max=1e5,
 335         logger=print,
 336         **kwargs,
 337     ):
 338         super().__init__()
 339
 340         def randw(*d):
 341             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 342
 343         self.nb_lines = nb_lines
 344         self.attention_dropout = attention_dropout
 345
 346         self.k_star = randw(nb_lines, dim_qk)
 347
 348         self.w_qw = randw(nb_heads, dim_qk, dim_model)
 349         self.w_qr = randw(nb_heads, dim_qk, dim_model)
 350         self.w_k = randw(nb_heads, dim_qk, dim_model)
 351         self.w_v = randw(nb_heads, dim_v, dim_model)
 352         self.w_o = randw(dim_v * nb_heads, dim_model)
 353
 354     def reset_inner_loss(self):
 355         self.acc_attention = 0
 356         self.acc_nb = 0
 357
 358     def get_inner_loss(self):
 359         warnings.warn("l2 regularization", RuntimeWarning)
 360         return (self.acc_attention / self.acc_nb).pow(2).sum()
 361         # return torch.tensor([0], device=self.w_qw.device)
 362         # warnings.warn("side regularization", RuntimeWarning)
 363         # return (
 364         # (0.5 / self.nb_lines - self.acc_attention / self.acc_nb).clamp(min=0).sum()
 365         # )
 366         # return torch.tensor([0], device=self.w_qw.device)
 367
 368     def forward(self, bs):
 369         x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
 370
 371         if bs.init_cache:
 372             self.rec_v = x_q.new_zeros(
 373                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
 374             )
 375             self.rec_k = x_q.new_zeros(
 376                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
 377             )
 378             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 379
 380         ######################################################################
 381         # Prepare the keys
 382
 383         k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
 384
 385         warnings.warn("rotating key barrel", RuntimeWarning)
 386         k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
 387         t_barrel = torch.arange(t0, t1, device=k_star.device)
 388         t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
 389         l_barrel = (
 390             torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
 391         ) % k_star.size(0)
 392         k_star = k_star[l_barrel, t_barrel]
 393
 394         ######################################################################
 395         # Compute the recurrent state
 396
 397         qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
 398
 399         v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
 400         k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
 401
 402         aw = torch.einsum(
 403             "nhtd,ltd->nhlt",
 404             qw,
 405             k_star,
 406         ) / math.sqrt(self.w_qw.size(1))
 407
 408         aw = aw.softmax(dim=2)  # nhlt
 409
 410         if self.train:
 411             # We want all the memory lines to be used similarly
 412             self.acc_attention += aw.sum(dim=(0, 1, 3))  # Sum accross NxHx_xT
 413             self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
 414
 415         aw = F.dropout(aw, self.attention_dropout, self.training)
 416
 417         A = 1 - aw.sum(dim=1)  # nlt
 418
 419         V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
 420         K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
 421
 422         if t0 == 0:
 423             V0 = None
 424             K0 = None
 425         else:
 426             V0 = self.rec_v[:, :, t0 - 1]
 427             K0 = self.rec_k[:, :, t0 - 1]
 428
 429         self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
 430         self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
 431
 432         ######################################################################
 433         # compute the readout
 434
 435         qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
 436
 437         ar = torch.einsum(
 438             "nhtd,nltd->nhlt",
 439             qr,
 440             self.rec_k[:, :, t0:t1],
 441         ) / math.sqrt(self.w_qr.size(1))
 442
 443         ar = ar.softmax(dim=2)  # nhlt
 444
 445         ar = F.dropout(ar, self.attention_dropout, self.training)
 446
 447         y = torch.einsum(
 448             "nhlt,nltd->nthd",
 449             ar,
 450             self.rec_v[:, :, t0:t1],
 451         ).flatten(2)
 452
 453         self.cache_y[:, t0:t1] = y @ self.w_o
 454
 455         return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
 456
 457
 458 ##############################
 459
 460
 461 # Returns a tensor with an additional index at rank win_dim, that move
 462 # along the same dimension as dim, on a domain {0...win_size-1}, and
 463 # dim is restricted on a domain reduced by win_size-1 values.
 464
 465
 466 def moving_window(x, dim, win_dim, win_size):
 467     size, stride = x.size(), x.stride()
 468     size = size[:dim] + (size[dim] - win_size + 1,) + size[dim + 1 :]
 469     size = size[:win_dim] + (win_size,) + size[win_dim:]
 470     stride = stride[:win_dim] + (stride[dim],) + stride[win_dim:]
 471
 472     return x.as_strided(size=size, stride=stride)
 473
 474
 475 ##############################
 476
 477
 478 class Caterpillar(nn.Module):
 479     def __init__(
 480         self,
 481         dim_model,
 482         dim_qk,
 483         dim_v,
 484         nb_heads,
 485         caterpillar_length,
 486         caterpillar_height,
 487         attention_dropout=0.0,
 488         len_max=1e5,
 489         logger=print,
 490         **kwargs,
 491     ):
 492         super().__init__()
 493
 494         warnings.warn("Caterpillar", RuntimeWarning)
 495
 496         def randw(*d, amplitude=None):
 497             if amplitude is None:
 498                 amplitude = 1 / math.sqrt(d[-1])
 499             return nn.Parameter(amplitude * torch.randn(*d))
 500
 501         self.caterpillar_length = caterpillar_length
 502         self.caterpillar_height = caterpillar_height
 503         self.attention_dropout = attention_dropout
 504
 505         ######################################################################
 506         # sup_args
 507
 508         x = kwargs.get("gate_dropout")
 509         if x is None:
 510             self.proba_gate_dropout = 0.0
 511         else:
 512             self.proba_gate_dropout = float(x)
 513
 514         logger(f"self.proba_gate_dropout {self.proba_gate_dropout}")
 515
 516         x = kwargs.get("default_bg")
 517         if x is None:
 518             default_bg = -math.log(caterpillar_height - 1)
 519         else:
 520             default_bg = float(x)
 521
 522         logger(f"default_bg {default_bg}")
 523
 524         ######################################################################
 525
 526         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
 527         self.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), default_bg))
 528
 529         self.w_K = randw(nb_heads, dim_qk, dim_model)
 530         self.w_V = randw(nb_heads, dim_v, dim_model)
 531         self.w_Q = randw(nb_heads, dim_qk, dim_model)
 532         self.w_O = randw(dim_v * nb_heads, dim_model)
 533
 534         self.init_K_rec = randw(
 535             caterpillar_height,
 536             caterpillar_length,
 537             dim_qk,
 538         )
 539         self.init_V_rec = randw(
 540             caterpillar_height,
 541             caterpillar_length,
 542             dim_v,
 543         )
 544
 545     def reset_inner_loss(self):
 546         self.acc_attention = 0
 547         self.acc_nb = 0
 548
 549     def get_inner_loss(self):
 550         # warnings.warn("l2 regularization", RuntimeWarning)
 551         # return (self.acc_attention / self.acc_nb).pow(2).sum()
 552         return torch.tensor([0], device=self.w_Q.device)
 553
 554     def forward(self, bs):
 555         # Dimensions to make the source a bit clearer, that's needed
 556
 557         X, t0, t1 = bs.slice(), bs.first, bs.first + bs.nb
 558
 559         N = bs.x.size(0)
 560         T = bs.x.size(1)
 561         H = self.w_V.size(0)
 562         DV = self.w_V.size(1)
 563         DK = self.w_K.size(1)
 564         DM = self.w_O.size(1)
 565         R = self.caterpillar_height
 566         L = self.caterpillar_length
 567
 568         assert (
 569             t0 >= L and (t1 - t0) % L == 0
 570         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 571
 572         # We cache values to deal efficiently with auto-regression
 573
 574         if bs.init_cache:
 575             self.rec_V = X.new_zeros(N, R, T, DV)
 576             self.rec_K = X.new_zeros(N, R, T, DK)
 577             # We start the recurrent sequences with optimizable
 578             # initial values. No idea if it helps.
 579             self.rec_V[:, :, t0 - L : t0, :] = self.init_V_rec[None, :, :, :]
 580             self.rec_K[:, :, t0 - L : t0, :] = self.init_K_rec[None, :, :, :]
 581
 582             self.cache_Y = X.new_zeros(N, T, DM)
 583
 584         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
 585         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
 586
 587         ######################################################################
 588         # Compute the recurrent state
 589
 590         # This is the Gating sequence that modulates the storing of
 591         # the new key and value in the R pairs of the current
 592         # stack. There are R independent gating values, which means
 593         # that the current K/V may be stored in multiple pairs of the
 594         # recurrent state, or not at all.
 595
 596         G = (
 597             torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
 598         ).sigmoid()
 599
 600         # warnings.warn("softmax gating", RuntimeWarning)
 601
 602         # G = (
 603         # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
 604         # ).softmax(dim=2)
 605
 606         ######################################################################
 607         # The "flashbacks"
 608
 609         if self.training and self.proba_gate_dropout > 0.0:
 610             # This is a better implementation of "flashbacks".
 611
 612             # G is NxHxExT where e is the caterpillar's row.
 613
 614             warnings.warn("gate dropout", RuntimeWarning)
 615
 616             kill = (
 617                 torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout
 618             ).float()
 619
 620             alpha = G / (1 - self.proba_gate_dropout)
 621
 622             G = alpha * (1 - kill)
 623
 624         def recurrence(G, V, K):
 625             # Clip the gating to avoid values greater than 1 when several
 626             # heads hit the same row
 627
 628             G = G / G.sum(1, keepdim=True).clamp(min=1)
 629
 630             # We prepare the arguments for the parallel scan
 631
 632             A = 1 - G.sum(1)
 633
 634             gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
 635             gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
 636
 637             # We start from cached values, which matters in inference
 638
 639             init_rec_V = self.rec_V[:, :, t0 - L : t0]
 640             init_rec_K = self.rec_K[:, :, t0 - L : t0]
 641
 642             # Associative scan
 643
 644             # Here there is a trick: Since the stack at position t is
 645             # computed by updating that at position t-L, the parallel
 646             # scan operates with a period of L. To do so we split the
 647             # sequence indexing in two axes, the second of size L, and
 648             # run the parallel scan using the first as the sequence index.
 649
 650             A = A.unflatten(2, (-1, L))
 651             gated_V = gated_V.unflatten(2, (-1, L))
 652             gated_K = gated_K.unflatten(2, (-1, L))
 653
 654             next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
 655             next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
 656
 657             next_V = next_V.flatten(2, 3)
 658             next_K = next_K.flatten(2, 3)
 659
 660             return next_V, next_K
 661
 662         #################################################################
 663
 664         next_V, next_K = recurrence(G, V, K)
 665
 666         self.rec_V[:, :, t0:t1] = next_V
 667         self.rec_K[:, :, t0:t1] = next_K
 668
 669         ######################################################################
 670         # compute the readout
 671
 672         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 673
 674         # We build tensors NxHxTxFxL where N is the sample index, H
 675         # the head, T the time, F the row in the caterpillar, and L
 676         # the column in the caterpillar
 677
 678         windowed_V = moving_window(
 679             self.rec_V[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
 680         )
 681
 682         windowed_K = moving_window(
 683             self.rec_K[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
 684         )
 685
 686         # We have an attention score for each of the RxL values
 687
 688         ar = torch.einsum(
 689             "nhtd,nftld->nhtfl",
 690             Q,
 691             windowed_K,
 692         ) / math.sqrt(DK)
 693
 694         # softmax can operate only on one dimension, hence the
 695         # flattening
 696
 697         ar = ar.flatten(3).softmax(dim=3).view(ar.size())
 698
 699         ar = F.dropout(ar, self.attention_dropout, self.training)
 700
 701         # Compute the output for each head, flatten to concatenate
 702
 703         Y = torch.einsum(
 704             "nhtfl,nftld->nthd",
 705             ar,
 706             windowed_V,
 707         ).flatten(2)
 708
 709         # Compute the final output
 710
 711         self.cache_Y[:, t0:t1] = Y @ self.w_O
 712
 713         return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
 714
 715
 716 ##############################
 717
 718
 719 class QKVAttention(nn.Module):
 720     def __init__(
 721         self,
 722         dim_model,
 723         dim_qk,
 724         dim_v,
 725         nb_heads=1,
 726         causal=False,
 727         attention_dropout=0.0,
 728         logger=print,
 729         **kwargs,
 730     ):
 731         super().__init__()
 732
 733         def randw(*d):
 734             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 735
 736         self.causal = causal
 737         self.attention_dropout = attention_dropout
 738         self.record_attention = False
 739
 740         self.w_q = randw(nb_heads, dim_qk, dim_model)
 741         self.w_k = randw(nb_heads, dim_qk, dim_model)
 742         self.w_v = randw(nb_heads, dim_v, dim_model)
 743         self.w_o = randw(dim_v * nb_heads, dim_model)
 744
 745     def forward(self, bs):
 746         x_q = bs.x
 747
 748         assert (
 749             self.causal or bs.complete()
 750         ), "Partial evaluation is only possible for causal models"
 751
 752         if bs.init_cache:
 753             self.cache_k = x_q.new_zeros(
 754                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 755             )
 756             self.cache_v = x_q.new_zeros(
 757                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 758             )
 759             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 760
 761         q = torch.einsum("ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_q)
 762
 763         self.cache_k[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 764             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_k
 765         )
 766         self.cache_v[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 767             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_v
 768         )
 769
 770         a = torch.einsum(
 771             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs.first + bs.nb]
 772         ) / math.sqrt(self.w_q.size(1))
 773
 774         if self.causal:
 775             if bs.init_cache:
 776                 self.cache_attzero = (
 777                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 778                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 779                 )
 780             a = a.masked_fill(
 781                 self.cache_attzero[
 782                     :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb
 783                 ],
 784                 float("-inf"),
 785             )
 786
 787         a = a.softmax(dim=3)
 788
 789         if self.record_attention:
 790             self.a = a
 791
 792         a = F.dropout(a, self.attention_dropout, self.training)
 793
 794         y = torch.einsum(
 795             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs.first + bs.nb]
 796         ).flatten(2)
 797
 798         self.cache_y[:, bs.first : bs.first + bs.nb] = y @ self.w_o
 799
 800         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 801
 802
 803 ##############################
 804
 805
 806 class MyGPT(nn.Module):
 807     def __init__(
 808         self,
 809         vocabulary_size,
 810         dim_model,
 811         dim_keys,
 812         dim_hidden,
 813         nb_heads,
 814         nb_blocks,
 815         nb_lines=None,
 816         caterpillar_height=None,
 817         causal=False,
 818         dropout=0.0,
 819         len_max=1e5,
 820         attention_layer="kvrec",
 821         logger=print,
 822         **kwargs,
 823     ):
 824         super().__init__()
 825
 826         assert attention_layer in {
 827             "mha",
 828             "dumbrec",
 829             "kvrec",
 830             "caterpillar",
 831         }, f"Unknown attention operator {attention_layer}."
 832
 833         if attention_layer == "caterpillar":
 834             assert nb_lines % caterpillar_height == 0
 835             self.caterpillar_length = nb_lines // caterpillar_height
 836             self.caterpillar_height = caterpillar_height
 837         else:
 838             self.caterpillar_length = -1
 839             self.caterpillar_height = -1
 840
 841         assert dim_model % nb_heads == 0
 842
 843         self.embedding = nn.Sequential(
 844             CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
 845             AddPositionalEncoding(len_max),
 846         )
 847
 848         trunk_blocks = []
 849
 850         def attlayer():
 851             if attention_layer == "mha":
 852                 return QKVAttention(
 853                     dim_model=dim_model,
 854                     dim_qk=dim_keys,
 855                     dim_v=dim_model // nb_heads,
 856                     nb_heads=nb_heads,
 857                     causal=causal,
 858                     attention_dropout=dropout,
 859                     logger=logger,
 860                     **kwargs,
 861                 )
 862             elif attention_layer == "dumbrec":
 863                 return DumbRec(
 864                     dim_model=dim_model,
 865                     dim_qk=dim_keys,
 866                     dim_v=dim_model // nb_heads,
 867                     nb_heads=nb_heads,
 868                     nb_lines=nb_lines,
 869                     attention_dropout=dropout,
 870                     logger=logger,
 871                     **kwargs,
 872                 )
 873             elif attention_layer == "kvrec":
 874                 return KVRec(
 875                     dim_model=dim_model,
 876                     dim_qk=dim_keys,
 877                     dim_v=dim_model // nb_heads,
 878                     nb_heads=nb_heads,
 879                     nb_lines=nb_lines,
 880                     attention_dropout=dropout,
 881                     logger=logger,
 882                     **kwargs,
 883                 )
 884             elif attention_layer == "caterpillar":
 885                 return Caterpillar(
 886                     dim_model=dim_model,
 887                     dim_qk=dim_keys,
 888                     dim_v=dim_model // nb_heads,
 889                     nb_heads=nb_heads,
 890                     caterpillar_length=self.caterpillar_length,
 891                     caterpillar_height=self.caterpillar_height,
 892                     attention_dropout=dropout,
 893                     logger=logger,
 894                     **kwargs,
 895                 )
 896             else:
 897                 raise ValueError(f"Unknown attention type {attention_layer}.")
 898
 899         for b in range(nb_blocks):
 900             trunk_blocks += [
 901                 WithResidual(
 902                     CacheWrapper(nn.LayerNorm((dim_model,))),
 903                     attlayer(),
 904                 ),
 905                 WithResidual(
 906                     CacheWrapper(
 907                         nn.LayerNorm((dim_model,)),
 908                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 909                         nn.ReLU(),
 910                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 911                         nn.Dropout(dropout),
 912                     ),
 913                 ),
 914             ]
 915
 916         self.trunk = nn.Sequential(*trunk_blocks)
 917
 918         self.readout = CacheWrapper(
 919             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 920         )
 921
 922         with torch.no_grad():
 923             for m in self.modules():
 924                 if isinstance(m, nn.Embedding):
 925                     m.weight.normal_(mean=0, std=2e-2)
 926                 elif isinstance(m, nn.LayerNorm):
 927                     m.bias.zero_()
 928                     m.weight.fill_(1.0)
 929
 930         self.reset_inner_loss()
 931
 932     def forward(self, bs):
 933         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb, bs.init_cache)
 934
 935         # To make the code simpler in the Caterpillar layer, we pad
 936         # here. It's unclear if/how much it hurts computationaly by
 937         # increasing the sequence length for the other layers
 938
 939         if self.caterpillar_length > 0:
 940             original_nb = bs.nb
 941             if bs.nb % self.caterpillar_length > 0:
 942                 bs.nb += self.caterpillar_length - bs.nb % self.caterpillar_length
 943
 944             bs = BracketedSequence(
 945                 F.pad(bs.x, (self.caterpillar_length, self.caterpillar_length)),
 946                 bs.first + self.caterpillar_length,
 947                 bs.nb,
 948                 bs.init_cache,
 949             )
 950
 951         bs = self.embedding(bs)
 952         bs = self.trunk(bs)
 953         bs = self.readout(bs)
 954
 955         if self.caterpillar_length > 0:
 956             bs = BracketedSequence(
 957                 F.pad(bs.x, (0, 0, -self.caterpillar_length, -self.caterpillar_length)),
 958                 bs.first - self.caterpillar_length,
 959                 original_nb,
 960                 bs.init_cache,
 961             )
 962
 963         return bs
 964
 965     # ar_mask is a tensor with 0s and 1s, of same shape as input, with
 966     # 1s where tokens should be generated. The others are kept
 967     # unchanged.
 968
 969     def masked_inplace_autoregression(
 970         self,
 971         input_src,
 972         ar_mask_src,
 973         forbidden_tokens=None,
 974         deterministic_synthesis=False,
 975     ):
 976         input = input_src.to(self.readout.f.weight.device)
 977         ar_mask = ar_mask_src.to(self.readout.f.weight.device)
 978         to_generate = (ar_mask.sum(0) > 0).nonzero()
 979         if to_generate.min() > 0:
 980             self(
 981                 BracketedSequence(input, 0, to_generate.min(), True)
 982             )  # Needed to initialize the model's cache
 983         for s in range(to_generate.min(), to_generate.max() + 1):
 984             output = self(BracketedSequence(input, s, 1, s == 0)).x
 985             logits = output[:, s]
 986             if forbidden_tokens is not None:
 987                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 988             if deterministic_synthesis:
 989                 t_next = logits.argmax(1)
 990             else:
 991                 dist = torch.distributions.categorical.Categorical(logits=logits)
 992                 t_next = dist.sample()
 993             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 994
 995         input_src.copy_(input)
 996
 997     def reset_inner_loss(self):
 998         for m in self.modules():
 999             if m is not self and hasattr(m, "reset_inner_loss"):
1000                 m.reset_inner_loss()
1001
1002     def get_inner_loss(self):
1003         l = torch.tensor([0.0], device=self.readout.f.weight.device)
1004         for m in self.modules():
1005             if m is not self and hasattr(m, "get_inner_loss"):
1006                 l += m.get_inner_loss()
1007         return l
1008
1009     def record_attention(self, v=True):
1010         for m in self.modules():
1011             if isinstance(m, QKVAttention):
1012                 m.record_attention = v
1013
1014     def retrieve_attention(self):
1015         a = []
1016         for m in self.modules():
1017             if isinstance(m, QKVAttention):
1018                 a.append(m.a)
1019         return a
1020
1021
1022 ######################################################################
1023
1024 if __name__ == "__main__":
1025     print("Basic check.")
1026
1027     m = Caterpillar(
1028         dim_model=4,
1029         dim_qk=3,
1030         dim_v=7,
1031         nb_heads=1,
1032         caterpillar_length=7,
1033         caterpillar_height=3,
1034         attention_dropout=0.0,
1035     )
1036
1037     m.reset_inner_loss()
1038     x = torch.randn(1, 21 + 2 * 7, 4)
1039     y1 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
1040     y2 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
1041     y3a = m(BracketedSequence(x, first=7, nb=14, init_cache=True)).x[:, 7:21]
1042     y3b = m(BracketedSequence(x, first=21, nb=7, init_cache=False)).x[:, 21:28]
1043     print((y1 - y2).abs().max())
1044     print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
1045     exit(0)
1046
1047     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
1048
1049     vocabulary_size = 128
1050     x = torch.randint(vocabulary_size, (6, 1024))
1051
1052     model = MyGPT(
1053         vocabulary_size=vocabulary_size,
1054         dim_model=512,
1055         dim_keys=64,
1056         dim_hidden=2048,
1057         nb_heads=8,
1058         nb_lines=128,
1059         nb_blocks=12,
1060         dropout=0.1,
1061         causal=True,
1062     )
1063
1064     x = x.to(device)
1065     model.to(device)
1066
1067     import time, sys
1068
1069     # import torchvision.models as models
1070     # from torch.profiler import profile, record_function, ProfilerActivity
1071
1072     # with profile(activities=[ProfilerActivity.CPU,  ProfilerActivity.CUDA], profile_memory=True, record_shapes=True) as prof:
1073     # with record_function("model_inference"):
1074
1075     model.eval()
1076     for i in range(3):
1077         start_time = time.perf_counter()
1078         for k in range(10):
1079             model(BracketedSequence(x))
1080         duration = time.perf_counter() - start_time
1081         print(duration)
1082         sys.stdout.flush()
1083
1084     # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
1085     # print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
1086
1087     # print("##############################################################")
1088     # y2 = torch.randn_like(y1)
1089     # for s in range(x.size(1)):
1090     # z = model(BracketedSequence(x, s, 1))
1091     # y2[:, s : s + 1] = z.slice()
1092
1093     # print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
1094
1095 ######################################################################