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):
 480             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 481
 482         self.caterpillar_length = caterpillar_length
 483         self.caterpillar_height = caterpillar_height
 484         self.attention_dropout = attention_dropout
 485
 486         self.proba_gate_dropout = 0.0
 487
 488         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
 489         self.b_G = nn.Parameter(
 490             torch.full(
 491                 (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
 492             )
 493         )
 494
 495         self.w_K = randw(nb_heads, dim_qk, dim_model)
 496         self.w_V = randw(nb_heads, dim_v, dim_model)
 497         self.w_Q = randw(nb_heads, dim_qk, dim_model)
 498         self.w_O = randw(dim_v * nb_heads, dim_model)
 499
 500         self.init_K_rec = randw(caterpillar_height, caterpillar_length, dim_qk)
 501         self.init_V_rec = randw(caterpillar_height, caterpillar_length, dim_v)
 502
 503     def reset_inner_loss(self):
 504         self.acc_attention = 0
 505         self.acc_nb = 0
 506
 507     def get_inner_loss(self):
 508         # warnings.warn("l2 regularization", RuntimeWarning)
 509         # return (self.acc_attention / self.acc_nb).pow(2).sum()
 510         return torch.tensor([0], device=self.w_Q.device)
 511
 512     def forward(self, bs):
 513         # Dimensions to make the source a bit clearer, that's needed
 514
 515         X, t0, t1 = bs.slice(), bs.first, bs.first + bs.nb
 516
 517         N = bs.x.size(0)
 518         T = bs.x.size(1)
 519         H = self.w_V.size(0)
 520         DV = self.w_V.size(1)
 521         DK = self.w_K.size(1)
 522         DM = self.w_O.size(1)
 523         CH = self.caterpillar_height
 524         CL = self.caterpillar_length
 525
 526         assert (
 527             t0 >= CL and (t1 - t0) % CL == 0
 528         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 529
 530         # We cache values to deal efficiently with auto-regression
 531
 532         if bs.init_cache:
 533             self.rec_V = X.new_zeros(N, CH, T, DV)
 534             self.rec_K = X.new_zeros(N, CH, T, DK)
 535             # We start the recurrent sequences with optimizable
 536             # initial values. No idea if it helps.
 537             self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
 538             self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
 539
 540             self.cache_Y = X.new_zeros(N, T, DM)
 541
 542         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
 543         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
 544
 545         ######################################################################
 546         # Compute the recurrent state
 547
 548         # This is the Gating sequence that modulates the storing of
 549         # the new key and value in the CH pairs of the current
 550         # stack. There are CH independent gating values, which means
 551         # that the current K/V may be stored in multiple pairs of the
 552         # recurrent state, or not at all.
 553
 554         G = (
 555             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
 556         ).sigmoid()
 557
 558         # Clip the gating to avoid values greater than 1 when several
 559         # heads hit the same row
 560
 561         G = G / G.sum(1, keepdim=True).clamp(min=1)
 562
 563         # We prepare the arguments for the parallel scan
 564
 565         A = 1 - G.sum(1)
 566         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
 567         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
 568
 569         # We start from cached values, which matters in inference
 570
 571         init_rec_V = self.rec_V[:, :, t0 - CL : t0]
 572         init_rec_K = self.rec_K[:, :, t0 - CL : t0]
 573
 574         ######################################################################
 575
 576         if self.training and self.proba_gate_dropout > 0.0:
 577             warnings.warn("gate dropout", RuntimeWarning)
 578             epsilon = 0.5
 579
 580         #################################################################
 581         # Associative scan
 582
 583         # Here there is a trick: Since the stack at position t is
 584         # computed by updating that at position t-CL, the parallel
 585         # scan operates with a period of CL. To do so we split the
 586         # sequence indexing in two axes, the second of size CL, and
 587         # run the parallel scan using the first as the sequence index.
 588
 589         A = A.unflatten(2, (-1, CL))
 590         gated_V = gated_V.unflatten(2, (-1, CL))
 591         gated_K = gated_K.unflatten(2, (-1, CL))
 592
 593         next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
 594         next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
 595
 596         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
 597         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 598
 599         ######################################################################
 600         # compute the readout
 601
 602         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 603
 604         # We build tensors NxHxTxFxL where N is the sample index, H
 605         # the head, T the time, F the row in the caterpillar, and L
 606         # the column in the caterpillar
 607
 608         windowed_V = moving_window(
 609             self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 610         )
 611
 612         windowed_K = moving_window(
 613             self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 614         )
 615
 616         # We have an attention score for each of the CHxCL values
 617
 618         ar = torch.einsum(
 619             "nhtd,nftld->nhtfl",
 620             Q,
 621             windowed_K,
 622         ) / math.sqrt(DK)
 623
 624         # softmax can operate only on one dimension, hence the
 625         # flattening
 626
 627         ar = ar.flatten(3).softmax(dim=3).view(ar.size())
 628
 629         ar = F.dropout(ar, self.attention_dropout, self.training)
 630
 631         # Compute the output for each head, flatten to concatenate
 632
 633         Y = torch.einsum(
 634             "nhtfl,nftld->nthd",
 635             ar,
 636             windowed_V,
 637         ).flatten(2)
 638
 639         # Compute the final output
 640
 641         self.cache_Y[:, t0:t1] = Y @ self.w_O
 642
 643         return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
 644
 645
 646 ##############################
 647
 648
 649 class QKVAttention(nn.Module):
 650     def __init__(
 651         self,
 652         dim_model,
 653         dim_qk,
 654         dim_v,
 655         nb_heads=1,
 656         causal=False,
 657         attention_dropout=0.0,
 658     ):
 659         super().__init__()
 660
 661         def randw(*d):
 662             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 663
 664         self.causal = causal
 665         self.attention_dropout = attention_dropout
 666         self.record_attention = False
 667
 668         self.w_q = randw(nb_heads, dim_qk, dim_model)
 669         self.w_k = randw(nb_heads, dim_qk, dim_model)
 670         self.w_v = randw(nb_heads, dim_v, dim_model)
 671         self.w_o = randw(dim_v * nb_heads, dim_model)
 672
 673     def forward(self, bs):
 674         x_q = bs.x
 675
 676         assert (
 677             self.causal or bs.complete()
 678         ), "Partial evaluation is only possible for causal models"
 679
 680         if bs.init_cache:
 681             self.cache_k = x_q.new_zeros(
 682                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 683             )
 684             self.cache_v = x_q.new_zeros(
 685                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 686             )
 687             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 688
 689         q = torch.einsum("ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_q)
 690
 691         self.cache_k[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 692             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_k
 693         )
 694         self.cache_v[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 695             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_v
 696         )
 697
 698         a = torch.einsum(
 699             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs.first + bs.nb]
 700         ) / math.sqrt(self.w_q.size(1))
 701
 702         if self.causal:
 703             if bs.init_cache:
 704                 self.cache_attzero = (
 705                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 706                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 707                 )
 708             a = a.masked_fill(
 709                 self.cache_attzero[
 710                     :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb
 711                 ],
 712                 float("-inf"),
 713             )
 714
 715         a = a.softmax(dim=3)
 716
 717         if self.record_attention:
 718             self.a = a
 719
 720         a = F.dropout(a, self.attention_dropout, self.training)
 721
 722         y = torch.einsum(
 723             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs.first + bs.nb]
 724         ).flatten(2)
 725
 726         self.cache_y[:, bs.first : bs.first + bs.nb] = y @ self.w_o
 727
 728         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 729
 730
 731 ##############################
 732
 733
 734 class MyGPT(nn.Module):
 735     def __init__(
 736         self,
 737         vocabulary_size,
 738         dim_model,
 739         dim_keys,
 740         dim_hidden,
 741         nb_heads,
 742         nb_blocks,
 743         nb_lines=None,
 744         caterpillar_height=None,
 745         causal=False,
 746         dropout=0.0,
 747         len_max=1e5,
 748         attention_layer="kvrec",
 749     ):
 750         super().__init__()
 751
 752         assert attention_layer in {
 753             "mha",
 754             "dumbrec",
 755             "kvrec",
 756             "caterpillar",
 757         }, f"Unknown attention operator {attention_layer}."
 758
 759         if attention_layer == "caterpillar":
 760             assert nb_lines % caterpillar_height == 0
 761             self.caterpillar_length = nb_lines // caterpillar_height
 762             self.caterpillar_height = caterpillar_height
 763         else:
 764             self.caterpillar_length = -1
 765             self.caterpillar_height = -1
 766
 767         assert dim_model % nb_heads == 0
 768
 769         self.embedding = nn.Sequential(
 770             CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
 771             AddPositionalEncoding(len_max),
 772         )
 773
 774         trunk_blocks = []
 775
 776         def attlayer():
 777             if attention_layer == "mha":
 778                 return QKVAttention(
 779                     dim_model=dim_model,
 780                     dim_qk=dim_keys,
 781                     dim_v=dim_model // nb_heads,
 782                     nb_heads=nb_heads,
 783                     causal=causal,
 784                     attention_dropout=dropout,
 785                 )
 786             elif attention_layer == "dumbrec":
 787                 return DumbRec(
 788                     dim_model=dim_model,
 789                     dim_qk=dim_keys,
 790                     dim_v=dim_model // nb_heads,
 791                     nb_heads=nb_heads,
 792                     nb_lines=nb_lines,
 793                     attention_dropout=dropout,
 794                 )
 795             elif attention_layer == "kvrec":
 796                 return KVRec(
 797                     dim_model=dim_model,
 798                     dim_qk=dim_keys,
 799                     dim_v=dim_model // nb_heads,
 800                     nb_heads=nb_heads,
 801                     nb_lines=nb_lines,
 802                     attention_dropout=dropout,
 803                 )
 804             elif attention_layer == "caterpillar":
 805                 return Caterpillar(
 806                     dim_model=dim_model,
 807                     dim_qk=dim_keys,
 808                     dim_v=dim_model // nb_heads,
 809                     nb_heads=nb_heads,
 810                     caterpillar_length=self.caterpillar_length,
 811                     caterpillar_height=self.caterpillar_height,
 812                     attention_dropout=dropout,
 813                 )
 814             else:
 815                 raise ValueError(f"Unknown attention type {attention_layer}.")
 816
 817         for b in range(nb_blocks):
 818             trunk_blocks += [
 819                 WithResidual(
 820                     CacheWrapper(nn.LayerNorm((dim_model,))),
 821                     attlayer(),
 822                 ),
 823                 WithResidual(
 824                     CacheWrapper(
 825                         nn.LayerNorm((dim_model,)),
 826                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 827                         nn.ReLU(),
 828                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 829                         nn.Dropout(dropout),
 830                     ),
 831                 ),
 832             ]
 833
 834         self.trunk = nn.Sequential(*trunk_blocks)
 835
 836         self.readout = CacheWrapper(
 837             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 838         )
 839
 840         with torch.no_grad():
 841             for m in self.modules():
 842                 if isinstance(m, nn.Embedding):
 843                     m.weight.normal_(mean=0, std=2e-2)
 844                 elif isinstance(m, nn.LayerNorm):
 845                     m.bias.zero_()
 846                     m.weight.fill_(1.0)
 847
 848         self.reset_inner_loss()
 849
 850     def forward(self, bs):
 851         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb, bs.init_cache)
 852
 853         # To make the code simpler in the Caterpillar layer, we pad
 854         # here. It's unclear if/how much it hurts computationaly by
 855         # increasing the sequence length for the other layers
 856
 857         if self.caterpillar_length > 0:
 858             original_nb = bs.nb
 859             if bs.nb % self.caterpillar_length > 0:
 860                 bs.nb += self.caterpillar_length - bs.nb % self.caterpillar_length
 861
 862             bs = BracketedSequence(
 863                 F.pad(bs.x, (self.caterpillar_length, self.caterpillar_length)),
 864                 bs.first + self.caterpillar_length,
 865                 bs.nb,
 866                 bs.init_cache,
 867             )
 868
 869         bs = self.embedding(bs)
 870         bs = self.trunk(bs)
 871         bs = self.readout(bs)
 872
 873         if self.caterpillar_length > 0:
 874             bs = BracketedSequence(
 875                 F.pad(bs.x, (0, 0, -self.caterpillar_length, -self.caterpillar_length)),
 876                 bs.first - self.caterpillar_length,
 877                 original_nb,
 878                 bs.init_cache,
 879             )
 880
 881         return bs
 882
 883     # ar_mask is a tensor with 0s and 1s, of same shape as input, with
 884     # 1s where tokens should be generated. The others are kept
 885     # unchanged.
 886
 887     def masked_inplace_autoregression(
 888         self,
 889         input_src,
 890         ar_mask_src,
 891         forbidden_tokens=None,
 892         deterministic_synthesis=False,
 893     ):
 894         input = input_src.to(self.readout.f.weight.device)
 895         ar_mask = ar_mask_src.to(self.readout.f.weight.device)
 896         to_generate = (ar_mask.sum(0) > 0).nonzero()
 897         if to_generate.min() > 0:
 898             self(
 899                 BracketedSequence(input, 0, to_generate.min(), True)
 900             )  # Needed to initialize the model's cache
 901         for s in range(to_generate.min(), to_generate.max() + 1):
 902             output = self(BracketedSequence(input, s, 1, s == 0)).x
 903             logits = output[:, s]
 904             if forbidden_tokens is not None:
 905                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 906             if deterministic_synthesis:
 907                 t_next = logits.argmax(1)
 908             else:
 909                 dist = torch.distributions.categorical.Categorical(logits=logits)
 910                 t_next = dist.sample()
 911             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 912
 913         input_src.copy_(input)
 914
 915     def reset_inner_loss(self):
 916         for m in self.modules():
 917             if m is not self and hasattr(m, "reset_inner_loss"):
 918                 m.reset_inner_loss()
 919
 920     def get_inner_loss(self):
 921         l = torch.tensor([0.0], device=self.readout.f.weight.device)
 922         for m in self.modules():
 923             if m is not self and hasattr(m, "get_inner_loss"):
 924                 l += m.get_inner_loss()
 925         return l
 926
 927     def record_attention(self, v=True):
 928         for m in self.modules():
 929             if isinstance(m, QKVAttention):
 930                 m.record_attention = v
 931
 932     def retrieve_attention(self):
 933         a = []
 934         for m in self.modules():
 935             if isinstance(m, QKVAttention):
 936                 a.append(m.a)
 937         return a
 938
 939
 940 ######################################################################
 941
 942 if __name__ == "__main__":
 943     print("Basic check.")
 944
 945     m = Caterpillar(
 946         dim_model=4,
 947         dim_qk=3,
 948         dim_v=7,
 949         nb_heads=1,
 950         caterpillar_length=7,
 951         caterpillar_height=3,
 952         attention_dropout=0.0,
 953     )
 954
 955     m.reset_inner_loss()
 956     x = torch.randn(1, 21 + 2 * 7, 4)
 957     y1 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 958     y2 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 959     y3a = m(BracketedSequence(x, first=7, nb=14, init_cache=True)).x[:, 7:21]
 960     y3b = m(BracketedSequence(x, first=21, nb=7, init_cache=False)).x[:, 21:28]
 961     print((y1 - y2).abs().max())
 962     print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
 963     exit(0)
 964
 965     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 966
 967     vocabulary_size = 128
 968     x = torch.randint(vocabulary_size, (6, 1024))
 969
 970     model = MyGPT(
 971         vocabulary_size=vocabulary_size,
 972         dim_model=512,
 973         dim_keys=64,
 974         dim_hidden=2048,
 975         nb_heads=8,
 976         nb_lines=128,
 977         nb_blocks=12,
 978         dropout=0.1,
 979         causal=True,
 980     )
 981
 982     x = x.to(device)
 983     model.to(device)
 984
 985     import time, sys
 986
 987     # import torchvision.models as models
 988     # from torch.profiler import profile, record_function, ProfilerActivity
 989
 990     # with profile(activities=[ProfilerActivity.CPU,  ProfilerActivity.CUDA], profile_memory=True, record_shapes=True) as prof:
 991     # with record_function("model_inference"):
 992
 993     model.eval()
 994     for i in range(3):
 995         start_time = time.perf_counter()
 996         for k in range(10):
 997             model(BracketedSequence(x))
 998         duration = time.perf_counter() - start_time
 999         print(duration)
1000         sys.stdout.flush()
1001
1002     # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
1003     # print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
1004
1005     # print("##############################################################")
1006     # y2 = torch.randn_like(y1)
1007     # for s in range(x.size(1)):
1008     # z = model(BracketedSequence(x, s, 1))
1009     # y2[:, s : s + 1] = z.slice()
1010
1011     # print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
1012
1013 ######################################################################