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