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_flashback = 0.0
 487         self.proba_gate_dropout = 0.0
 488
 489         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
 490         self.b_G = nn.Parameter(
 491             torch.full(
 492                 (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
 493             )
 494         )
 495
 496         self.w_K = randw(nb_heads, dim_qk, dim_model)
 497         self.w_V = randw(nb_heads, dim_v, dim_model)
 498         self.w_Q = randw(nb_heads, dim_qk, dim_model)
 499         self.w_O = randw(dim_v * nb_heads, dim_model)
 500
 501         self.init_K_rec = randw(caterpillar_height, caterpillar_length, dim_qk)
 502         self.init_V_rec = randw(caterpillar_height, caterpillar_length, dim_v)
 503
 504     def reset_inner_loss(self):
 505         self.acc_attention = 0
 506         self.acc_nb = 0
 507
 508     def get_inner_loss(self):
 509         # warnings.warn("l2 regularization", RuntimeWarning)
 510         # return (self.acc_attention / self.acc_nb).pow(2).sum()
 511         return torch.tensor([0], device=self.w_Q.device)
 512
 513     def forward(self, bs):
 514         # Dimensions to make the source a bit clearer, that's needed
 515
 516         X, t0, t1 = bs.slice(), bs.first, bs.first + bs.nb
 517
 518         N = bs.x.size(0)
 519         T = bs.x.size(1)
 520         H = self.w_V.size(0)
 521         DV = self.w_V.size(1)
 522         DK = self.w_K.size(1)
 523         DM = self.w_O.size(1)
 524         CH = self.caterpillar_height
 525         CL = self.caterpillar_length
 526
 527         assert (
 528             t0 >= CL and (t1 - t0) % CL == 0
 529         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 530
 531         # We cache values to deal efficiently with auto-regression
 532
 533         if bs.init_cache:
 534             self.rec_V = X.new_zeros(N, CH, T, DV)
 535             self.rec_K = X.new_zeros(N, CH, T, DK)
 536             # We start the recurrent sequences with optimizable
 537             # initial values. No idea if it helps.
 538             self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
 539             self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
 540
 541             self.cache_Y = X.new_zeros(N, T, DM)
 542
 543         ######################################################################
 544         # Compute the recurrent state
 545
 546         # This is the Gating sequence that modulates the storing of
 547         # the new key and value in the CH pairs of the current
 548         # stack. There are CH independent gating values, which means
 549         # that the current K/V may be stored in multiple pairs of the
 550         # recurrent state, or not at all.
 551
 552         G = (
 553             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
 554         ).sigmoid()
 555
 556         # Clip the gating to avoid values greater than 1 when several
 557         # heads hit the same row
 558
 559         G = G / G.sum(1, keepdim=True).clamp(min=1)
 560
 561         if self.training and self.proba_gate_dropout > 0.0:
 562             warnings.warn("gate dropout", RuntimeWarning)
 563             epsilon = 0.5
 564
 565         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
 566         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
 567
 568         # We prepare the arguments for the parallel scan
 569
 570         A = 1 - G.sum(1)
 571         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
 572         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
 573
 574         # Initial recurrent state
 575
 576         init_rec_V = self.rec_V[:, :, t0 - CL : t0]
 577         init_rec_K = self.rec_K[:, :, t0 - CL : t0]
 578
 579         #################################################################
 580         # Associative scan
 581
 582         # Here there is a trick: Since the stack at position t is
 583         # computed by updating that at position t-CL, the parallel
 584         # scan operates with a period of CL. To do so we split the
 585         # sequence indexing in two axes, the second of size CL, and
 586         # run the parallel scan using the first as the sequence index.
 587
 588         A = A.unflatten(2, (-1, CL))
 589         gated_V = gated_V.unflatten(2, (-1, CL))
 590         gated_K = gated_K.unflatten(2, (-1, CL))
 591
 592         next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
 593         next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
 594
 595         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
 596         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 597
 598         #################################################################
 599
 600         if self.training and self.proba_flashback > 0.0:
 601             warnings.warn("flash back", RuntimeWarning)
 602             # This piece of code makes the assumption that there is
 603             # nothing informative before t0, otherwise we'd have to
 604             # implement a cache for V and K too. This should not be
 605             # too much of a problem since this is used only during
 606             # train, where full sequence are available
 607
 608             n = torch.arange(N, device=X.device)[:, None, None, None]
 609             t = torch.arange(t0, t1, device=X.device)[None, None, :, None]
 610             dv = torch.arange(DV, device=X.device)[None, None, None, :]
 611             dk = torch.arange(DK, device=X.device)[None, None, None, :]
 612
 613             u = (
 614                 torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
 615             ) * CL
 616
 617             src_time = t - u - t0
 618             src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
 619
 620             mask = (
 621                 torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
 622             ).long()
 623
 624             self.rec_V[:, :, t0:t1] = (
 625                 mask * V[n, src_head, src_time, dv]
 626                 + (1 - mask) * self.rec_V[:, :, t0:t1]
 627             )
 628
 629             self.rec_K[:, :, t0:t1] = (
 630                 mask * K[n, src_head, src_time, dk]
 631                 + (1 - mask) * self.rec_K[:, :, t0:t1]
 632             )
 633
 634         ######################################################################
 635         # compute the readout
 636
 637         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 638
 639         # We build tensors NxHxTxFxL where N is the sample index, H
 640         # the head, T the time, F the row in the caterpillar, and L
 641         # the column in the caterpillar
 642
 643         windowed_V = moving_window(
 644             self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 645         )
 646
 647         windowed_K = moving_window(
 648             self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 649         )
 650
 651         # We have an attention score for each of the CHxCL values
 652
 653         ar = torch.einsum(
 654             "nhtd,nftld->nhtfl",
 655             Q,
 656             windowed_K,
 657         ) / math.sqrt(DK)
 658
 659         # softmax can operate only on one dimension, hence the
 660         # flattening
 661
 662         ar = ar.flatten(3).softmax(dim=3).view(ar.size())
 663
 664         ar = F.dropout(ar, self.attention_dropout, self.training)
 665
 666         # Compute the output for each head, flatten to concatenate
 667
 668         Y = torch.einsum(
 669             "nhtfl,nftld->nthd",
 670             ar,
 671             windowed_V,
 672         ).flatten(2)
 673
 674         # Compute the final output
 675
 676         self.cache_Y[:, t0:t1] = Y @ self.w_O
 677
 678         return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
 679
 680
 681 ##############################
 682
 683
 684 class QKVAttention(nn.Module):
 685     def __init__(
 686         self,
 687         dim_model,
 688         dim_qk,
 689         dim_v,
 690         nb_heads=1,
 691         causal=False,
 692         attention_dropout=0.0,
 693     ):
 694         super().__init__()
 695
 696         def randw(*d):
 697             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 698
 699         self.causal = causal
 700         self.attention_dropout = attention_dropout
 701         self.record_attention = False
 702
 703         self.w_q = randw(nb_heads, dim_qk, dim_model)
 704         self.w_k = randw(nb_heads, dim_qk, dim_model)
 705         self.w_v = randw(nb_heads, dim_v, dim_model)
 706         self.w_o = randw(dim_v * nb_heads, dim_model)
 707
 708     def forward(self, bs):
 709         x_q = bs.x
 710
 711         assert (
 712             self.causal or bs.complete()
 713         ), "Partial evaluation is only possible for causal models"
 714
 715         if bs.init_cache:
 716             self.cache_k = x_q.new_zeros(
 717                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 718             )
 719             self.cache_v = x_q.new_zeros(
 720                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 721             )
 722             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 723
 724         q = torch.einsum("ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_q)
 725
 726         self.cache_k[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 727             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_k
 728         )
 729         self.cache_v[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 730             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_v
 731         )
 732
 733         a = torch.einsum(
 734             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs.first + bs.nb]
 735         ) / math.sqrt(self.w_q.size(1))
 736
 737         if self.causal:
 738             if bs.init_cache:
 739                 self.cache_attzero = (
 740                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 741                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 742                 )
 743             a = a.masked_fill(
 744                 self.cache_attzero[
 745                     :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb
 746                 ],
 747                 float("-inf"),
 748             )
 749
 750         a = a.softmax(dim=3)
 751
 752         if self.record_attention:
 753             self.a = a
 754
 755         a = F.dropout(a, self.attention_dropout, self.training)
 756
 757         y = torch.einsum(
 758             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs.first + bs.nb]
 759         ).flatten(2)
 760
 761         self.cache_y[:, bs.first : bs.first + bs.nb] = y @ self.w_o
 762
 763         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 764
 765
 766 ##############################
 767
 768
 769 class MyGPT(nn.Module):
 770     def __init__(
 771         self,
 772         vocabulary_size,
 773         dim_model,
 774         dim_keys,
 775         dim_hidden,
 776         nb_heads,
 777         nb_blocks,
 778         nb_lines=None,
 779         caterpillar_height=None,
 780         causal=False,
 781         dropout=0.0,
 782         len_max=1e5,
 783         attention_layer="kvrec",
 784     ):
 785         super().__init__()
 786
 787         assert attention_layer in {
 788             "mha",
 789             "dumbrec",
 790             "kvrec",
 791             "caterpillar",
 792         }, f"Unknown attention operator {attention_layer}."
 793
 794         if attention_layer == "caterpillar":
 795             assert nb_lines % caterpillar_height == 0
 796             self.caterpillar_length = nb_lines // caterpillar_height
 797             self.caterpillar_height = caterpillar_height
 798         else:
 799             self.caterpillar_length = -1
 800             self.caterpillar_height = -1
 801
 802         assert dim_model % nb_heads == 0
 803
 804         self.embedding = nn.Sequential(
 805             CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
 806             AddPositionalEncoding(len_max),
 807         )
 808
 809         trunk_blocks = []
 810
 811         def attlayer():
 812             if attention_layer == "mha":
 813                 return QKVAttention(
 814                     dim_model=dim_model,
 815                     dim_qk=dim_keys,
 816                     dim_v=dim_model // nb_heads,
 817                     nb_heads=nb_heads,
 818                     causal=causal,
 819                     attention_dropout=dropout,
 820                 )
 821             elif attention_layer == "dumbrec":
 822                 return DumbRec(
 823                     dim_model=dim_model,
 824                     dim_qk=dim_keys,
 825                     dim_v=dim_model // nb_heads,
 826                     nb_heads=nb_heads,
 827                     nb_lines=nb_lines,
 828                     attention_dropout=dropout,
 829                 )
 830             elif attention_layer == "kvrec":
 831                 return KVRec(
 832                     dim_model=dim_model,
 833                     dim_qk=dim_keys,
 834                     dim_v=dim_model // nb_heads,
 835                     nb_heads=nb_heads,
 836                     nb_lines=nb_lines,
 837                     attention_dropout=dropout,
 838                 )
 839             elif attention_layer == "caterpillar":
 840                 return Caterpillar(
 841                     dim_model=dim_model,
 842                     dim_qk=dim_keys,
 843                     dim_v=dim_model // nb_heads,
 844                     nb_heads=nb_heads,
 845                     caterpillar_length=self.caterpillar_length,
 846                     caterpillar_height=self.caterpillar_height,
 847                     attention_dropout=dropout,
 848                 )
 849             else:
 850                 raise ValueError(f"Unknown attention type {attention_layer}.")
 851
 852         for b in range(nb_blocks):
 853             trunk_blocks += [
 854                 WithResidual(
 855                     CacheWrapper(nn.LayerNorm((dim_model,))),
 856                     attlayer(),
 857                 ),
 858                 WithResidual(
 859                     CacheWrapper(
 860                         nn.LayerNorm((dim_model,)),
 861                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 862                         nn.ReLU(),
 863                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 864                         nn.Dropout(dropout),
 865                     ),
 866                 ),
 867             ]
 868
 869         self.trunk = nn.Sequential(*trunk_blocks)
 870
 871         self.readout = CacheWrapper(
 872             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 873         )
 874
 875         with torch.no_grad():
 876             for m in self.modules():
 877                 if isinstance(m, nn.Embedding):
 878                     m.weight.normal_(mean=0, std=2e-2)
 879                 elif isinstance(m, nn.LayerNorm):
 880                     m.bias.zero_()
 881                     m.weight.fill_(1.0)
 882
 883         self.reset_inner_loss()
 884
 885     def forward(self, bs):
 886         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb, bs.init_cache)
 887
 888         # To make the code simpler in the Caterpillar layer, we pad
 889         # here. It's unclear if/how much it hurts computationaly by
 890         # increasing the sequence length for the other layers
 891
 892         if self.caterpillar_length > 0:
 893             original_nb = bs.nb
 894             if bs.nb % self.caterpillar_length > 0:
 895                 bs.nb += self.caterpillar_length - bs.nb % self.caterpillar_length
 896
 897             bs = BracketedSequence(
 898                 F.pad(bs.x, (self.caterpillar_length, self.caterpillar_length)),
 899                 bs.first + self.caterpillar_length,
 900                 bs.nb,
 901                 bs.init_cache,
 902             )
 903
 904         bs = self.embedding(bs)
 905         bs = self.trunk(bs)
 906         bs = self.readout(bs)
 907
 908         if self.caterpillar_length > 0:
 909             bs = BracketedSequence(
 910                 F.pad(bs.x, (0, 0, -self.caterpillar_length, -self.caterpillar_length)),
 911                 bs.first - self.caterpillar_length,
 912                 original_nb,
 913                 bs.init_cache,
 914             )
 915
 916         return bs
 917
 918     # ar_mask is a tensor with 0s and 1s, of same shape as input, with
 919     # 1s where tokens should be generated. The others are kept
 920     # unchanged.
 921
 922     def masked_inplace_autoregression(
 923         self,
 924         input_src,
 925         ar_mask_src,
 926         forbidden_tokens=None,
 927         deterministic_synthesis=False,
 928     ):
 929         input = input_src.to(self.readout.f.weight.device)
 930         ar_mask = ar_mask_src.to(self.readout.f.weight.device)
 931         to_generate = (ar_mask.sum(0) > 0).nonzero()
 932         if to_generate.min() > 0:
 933             self(
 934                 BracketedSequence(input, 0, to_generate.min(), True)
 935             )  # Needed to initialize the model's cache
 936         for s in range(to_generate.min(), to_generate.max() + 1):
 937             output = self(BracketedSequence(input, s, 1, s == 0)).x
 938             logits = output[:, s]
 939             if forbidden_tokens is not None:
 940                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 941             if deterministic_synthesis:
 942                 t_next = logits.argmax(1)
 943             else:
 944                 dist = torch.distributions.categorical.Categorical(logits=logits)
 945                 t_next = dist.sample()
 946             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 947
 948         input_src.copy_(input)
 949
 950     def reset_inner_loss(self):
 951         for m in self.modules():
 952             if m is not self and hasattr(m, "reset_inner_loss"):
 953                 m.reset_inner_loss()
 954
 955     def get_inner_loss(self):
 956         l = torch.tensor([0.0], device=self.readout.f.weight.device)
 957         for m in self.modules():
 958             if m is not self and hasattr(m, "get_inner_loss"):
 959                 l += m.get_inner_loss()
 960         return l
 961
 962     def record_attention(self, v=True):
 963         for m in self.modules():
 964             if isinstance(m, QKVAttention):
 965                 m.record_attention = v
 966
 967     def retrieve_attention(self):
 968         a = []
 969         for m in self.modules():
 970             if isinstance(m, QKVAttention):
 971                 a.append(m.a)
 972         return a
 973
 974
 975 ######################################################################
 976
 977 if __name__ == "__main__":
 978     print("Basic check.")
 979
 980     m = Caterpillar(
 981         dim_model=4,
 982         dim_qk=3,
 983         dim_v=7,
 984         nb_heads=1,
 985         caterpillar_length=7,
 986         caterpillar_height=3,
 987         attention_dropout=0.0,
 988     )
 989
 990     m.reset_inner_loss()
 991     x = torch.randn(1, 21 + 2 * 7, 4)
 992     y1 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 993     y2 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 994     y3a = m(BracketedSequence(x, first=7, nb=14, init_cache=True)).x[:, 7:21]
 995     y3b = m(BracketedSequence(x, first=21, nb=7, init_cache=False)).x[:, 21:28]
 996     print((y1 - y2).abs().max())
 997     print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
 998     exit(0)
 999
1000     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
1001
1002     vocabulary_size = 128
1003     x = torch.randint(vocabulary_size, (6, 1024))
1004
1005     model = MyGPT(
1006         vocabulary_size=vocabulary_size,
1007         dim_model=512,
1008         dim_keys=64,
1009         dim_hidden=2048,
1010         nb_heads=8,
1011         nb_lines=128,
1012         nb_blocks=12,
1013         dropout=0.1,
1014         causal=True,
1015     )
1016
1017     x = x.to(device)
1018     model.to(device)
1019
1020     import time, sys
1021
1022     # import torchvision.models as models
1023     # from torch.profiler import profile, record_function, ProfilerActivity
1024
1025     # with profile(activities=[ProfilerActivity.CPU,  ProfilerActivity.CUDA], profile_memory=True, record_shapes=True) as prof:
1026     # with record_function("model_inference"):
1027
1028     model.eval()
1029     for i in range(3):
1030         start_time = time.perf_counter()
1031         for k in range(10):
1032             model(BracketedSequence(x))
1033         duration = time.perf_counter() - start_time
1034         print(duration)
1035         sys.stdout.flush()
1036
1037     # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
1038     # print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
1039
1040     # print("##############################################################")
1041     # y2 = torch.randn_like(y1)
1042     # for s in range(x.size(1)):
1043     # z = model(BracketedSequence(x, s, 1))
1044     # y2[:, s : s + 1] = z.slice()
1045
1046     # print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
1047
1048 ######################################################################