mygpt.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 # This is an implementation from scratch of a "GPT", that is a model
   9 # composed of several causal self-attention blocks. It is equipped
  10 # with a caching mechanism for keys and values to avoid a O(N^3) cost
  11 # for auto-regression.
  12
  13 import math, warnings
  14
  15 import torch, einops
  16
  17 from torch import nn
  18 from torch.nn import functional as F
  19
  20 import ffutils
  21
  22 # import memload
  23
  24 ######################################################################
  25
  26 # A BracketedSequence is a BxTx... tensor with a first and a nb time
  27 # steps to compute.
  28
  29 # Modules able to process it expect that they will have to process a
  30 # first bracket starting at t=0, followed by a succession of brackets
  31 # that move forward in time, do not overlap, and cover the axis T with
  32 # no holes.
  33 #
  34 # Although it is more general, for a classical prompt-conditioned
  35 # auto-regressive process it will be a first bracket starting at 0 and
  36 # of arbitrary length for the "prompt", followed by brackets of length
  37 # 1 for the successive tokens.
  38 #
  39 # Modules able to process brackets may implement a cache that is
  40 # resetted when the input bracket starts at t=0
  41
  42
  43 class BracketedSequence:
  44     def __init__(self, x, first=None, nb=None, init_cache=None):
  45         self.x = x
  46         assert (first is None and nb is None and init_cache is None) or (
  47             first is not None and nb is not None and init_cache is not None
  48         )
  49
  50         self.first = 0 if first is None else first
  51         self.nb = x.size(1) if nb is None else nb
  52         self.init_cache = True if init_cache is None else init_cache
  53
  54     def slice(self):
  55         return self.x[:, self.first : self.first + self.nb]
  56
  57     def complete(self):
  58         return self.first == 0 and self.nb == self.x.size(1)
  59
  60
  61 ######################################################################
  62
  63
  64 class CacheWrapper(nn.Module):
  65     def __init__(self, *f):
  66         super().__init__()
  67         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  68
  69     def forward(self, bs):
  70         if bs.init_cache:
  71             y = self.f(bs.slice())
  72             self.cache_y = y.new(*((y.size(0), bs.x.size(1)) + y.size()[2:]))
  73             self.cache_y[:, bs.first : bs.first + bs.nb] = y
  74         else:
  75             assert tuple(bs.x.size()[:2]) == tuple(self.cache_y.size()[:2])
  76             assert bs.first + bs.nb <= self.cache_y.size(1)
  77             self.cache_y[:, bs.first : bs.first + bs.nb] = self.f(bs.slice())
  78
  79         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
  80
  81
  82 ##############################
  83
  84
  85 class WithResidual(nn.Module):
  86     def __init__(self, *f):
  87         super().__init__()
  88         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  89
  90     def forward(self, bs):
  91         return BracketedSequence(bs.x + self.f(bs).x, bs.first, bs.nb, bs.init_cache)
  92
  93
  94 ##############################
  95
  96
  97 class AddPositionalEncoding(nn.Module):
  98     def __init__(self, len_max):
  99         super().__init__()
 100         self.len_max = len_max
 101
 102     # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
 103
 104     def forward(self, bs):
 105         if bs.init_cache:
 106             t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
 107                 :, None
 108             ]
 109             j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
 110                 None, :
 111             ]
 112             k = j % 2
 113             self.pe = torch.sin(
 114                 t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
 115             )
 116             self.cache_y = bs.x.new(bs.x.size())
 117
 118         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 119             bs.slice() + self.pe[bs.first : bs.first + bs.nb]
 120         )
 121
 122         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 123
 124
 125 import pscan
 126
 127
 128 # X is /.../xTxD   A is /.../xT   Y_init is /.../xD
 129
 130
 131 def pscan_dim(A, X, Y_init, dim=-2):
 132     s = X.size()
 133     a, T, b = s[:dim].numel(), s[dim], s[dim + 1 :].numel()
 134
 135     A = A.reshape(a, T, *s[dim + 1 : -1])
 136     X = X.reshape(a, T, *s[dim + 1 : -1], -1)
 137
 138     if Y_init is None:
 139         Y_init = X.new_zeros(a, *s[dim + 1 : -1], X.size(-1))
 140     else:
 141         Y_init = Y_init.reshape(a, *s[dim + 1 : -1], -1)
 142
 143     Y = pscan.pscan(A, X, Y_init).reshape(s)
 144
 145     return Y
 146
 147
 148 def pscan_shape(A, X, Y_init):
 149     s = X.size()
 150     A = A.reshape(-1, s[-2])
 151     X = X.reshape(-1, s[-2], s[-1])
 152
 153     if Y_init is None:
 154         Y_init = X.new_zeros(X.size(0), s[-1])
 155     else:
 156         Y_init = Y_init.reshape(-1, s[-1])
 157
 158     Y = pscan.pscan(A, X, Y_init).reshape(s)
 159
 160     return Y
 161
 162
 163 def nsum_shape(X, Y_init):
 164     s = X.size()
 165     X = X.reshape(-1, s[-2], s[-1])  # ntd
 166
 167     Y = 0 if Y_init is None else Y_init.reshape(-1, s[-1])
 168     result = []
 169
 170     for k in range(X.size(1)):
 171         Y = Y + X[:, k]
 172         Y = Y / Y.norm(dim=-1, keepdim=True).clamp(min=1)
 173         result.append(Y)
 174
 175     return torch.cat(result, dim=1).reshape(s)
 176
 177
 178 ##############################
 179
 180
 181 class DumbRec(nn.Module):
 182     def __init__(
 183         self,
 184         dim_model,
 185         dim_qk,
 186         dim_v,
 187         nb_heads,
 188         nb_lines,
 189         attention_dropout=0.0,
 190         len_max=1e5,
 191     ):
 192         super().__init__()
 193
 194         def randw(*d):
 195             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 196
 197         self.nb_lines = nb_lines
 198         self.attention_dropout = attention_dropout
 199
 200         self.k_star = randw(nb_lines, dim_qk)
 201
 202         self.w_qw = randw(nb_heads, dim_qk, dim_model)
 203         self.w_qr = randw(nb_heads, dim_qk, dim_model)
 204         # self.w_k = randw(nb_heads, dim_qk, dim_model)
 205         self.w_v = randw(nb_heads, dim_v, dim_model)
 206         self.w_o = randw(dim_v * nb_heads, dim_model)
 207
 208     def reset_inner_loss(self):
 209         self.acc_attention = 0
 210         self.acc_nb = 0
 211
 212     def get_inner_loss(self):
 213         warnings.warn("l2 regularization", RuntimeWarning)
 214         return (self.acc_attention / self.acc_nb).pow(2).sum()
 215         # return torch.tensor([0], device=self.w_qw.device)
 216
 217     def forward(self, bs):
 218         x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
 219
 220         if bs.init_cache:
 221             self.rec_v = x_q.new_zeros(
 222                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
 223             )
 224             # self.rec_k = x_q.new_zeros(
 225             # x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
 226             # )
 227             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 228
 229         ######################################################################
 230         # Prepare the keys
 231
 232         k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
 233
 234         warnings.warn("rotating key barrel", RuntimeWarning)
 235         k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
 236         t_barrel = torch.arange(t0, t1, device=k_star.device)
 237         t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
 238         l_barrel = (
 239             torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
 240         ) % k_star.size(0)
 241         k_star = k_star[l_barrel, t_barrel]
 242
 243         ######################################################################
 244         # Compute the recurrent state
 245
 246         qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
 247
 248         v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
 249         # k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
 250
 251         aw = torch.einsum(
 252             "nhtd,ltd->nhlt",
 253             qw,
 254             k_star,
 255         ) / math.sqrt(self.w_qw.size(1))
 256
 257         aw = aw.softmax(dim=2)  # nhlt
 258
 259         if self.train:
 260             self.acc_attention += aw.sum(dim=(0, 1, 3))
 261             self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
 262
 263         aw = F.dropout(aw, self.attention_dropout, self.training)
 264
 265         A = 1 - aw.sum(dim=1)  # nlt
 266
 267         V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
 268         # K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
 269
 270         if t0 == 0:
 271             V0 = None
 272             # K0 = None
 273         else:
 274             V0 = self.rec_v[:, :, t0 - 1]
 275             # K0 = self.rec_k[:, :, t0 - 1]
 276
 277         self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
 278         # self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
 279
 280         ######################################################################
 281         # compute the readout
 282
 283         qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
 284
 285         ar = torch.einsum(
 286             "nhtd,ld->nhlt",
 287             qr,
 288             # self.rec_k[:, :, t0:t1],
 289             self.k_star,
 290         ) / math.sqrt(self.w_qr.size(1))
 291
 292         ar = ar.softmax(dim=2)  # nhlt
 293
 294         ar = F.dropout(ar, self.attention_dropout, self.training)
 295
 296         y = torch.einsum(
 297             "nhlt,nltd->nthd",
 298             ar,
 299             self.rec_v[:, :, t0:t1],
 300         ).flatten(2)
 301
 302         self.cache_y[:, t0:t1] = y @ self.w_o
 303
 304         return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
 305
 306
 307 ##############################
 308
 309
 310 class KVRec(nn.Module):
 311     def __init__(
 312         self,
 313         dim_model,
 314         dim_qk,
 315         dim_v,
 316         nb_heads,
 317         nb_lines,
 318         attention_dropout=0.0,
 319         len_max=1e5,
 320     ):
 321         super().__init__()
 322
 323         def randw(*d):
 324             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 325
 326         self.nb_lines = nb_lines
 327         self.attention_dropout = attention_dropout
 328
 329         self.k_star = randw(nb_lines, dim_qk)
 330
 331         self.w_qw = randw(nb_heads, dim_qk, dim_model)
 332         self.w_qr = randw(nb_heads, dim_qk, dim_model)
 333         self.w_k = randw(nb_heads, dim_qk, dim_model)
 334         self.w_v = randw(nb_heads, dim_v, dim_model)
 335         self.w_o = randw(dim_v * nb_heads, dim_model)
 336
 337     def reset_inner_loss(self):
 338         self.acc_attention = 0
 339         self.acc_nb = 0
 340
 341     def get_inner_loss(self):
 342         warnings.warn("l2 regularization", RuntimeWarning)
 343         return (self.acc_attention / self.acc_nb).pow(2).sum()
 344         # return torch.tensor([0], device=self.w_qw.device)
 345         # warnings.warn("side regularization", RuntimeWarning)
 346         # return (
 347         # (0.5 / self.nb_lines - self.acc_attention / self.acc_nb).clamp(min=0).sum()
 348         # )
 349         # return torch.tensor([0], device=self.w_qw.device)
 350
 351     def forward(self, bs):
 352         x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
 353
 354         if bs.init_cache:
 355             self.rec_v = x_q.new_zeros(
 356                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
 357             )
 358             self.rec_k = x_q.new_zeros(
 359                 x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
 360             )
 361             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 362
 363         ######################################################################
 364         # Prepare the keys
 365
 366         k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
 367
 368         warnings.warn("rotating key barrel", RuntimeWarning)
 369         k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
 370         t_barrel = torch.arange(t0, t1, device=k_star.device)
 371         t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
 372         l_barrel = (
 373             torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
 374         ) % k_star.size(0)
 375         k_star = k_star[l_barrel, t_barrel]
 376
 377         ######################################################################
 378         # Compute the recurrent state
 379
 380         qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
 381
 382         v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
 383         k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
 384
 385         aw = torch.einsum(
 386             "nhtd,ltd->nhlt",
 387             qw,
 388             k_star,
 389         ) / math.sqrt(self.w_qw.size(1))
 390
 391         aw = aw.softmax(dim=2)  # nhlt
 392
 393         if self.train:
 394             # We want all the memory lines to be used similarly
 395             self.acc_attention += aw.sum(dim=(0, 1, 3))  # Sum accross NxHx_xT
 396             self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
 397
 398         aw = F.dropout(aw, self.attention_dropout, self.training)
 399
 400         A = 1 - aw.sum(dim=1)  # nlt
 401
 402         V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
 403         K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
 404
 405         if t0 == 0:
 406             V0 = None
 407             K0 = None
 408         else:
 409             V0 = self.rec_v[:, :, t0 - 1]
 410             K0 = self.rec_k[:, :, t0 - 1]
 411
 412         self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
 413         self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
 414
 415         ######################################################################
 416         # compute the readout
 417
 418         qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
 419
 420         ar = torch.einsum(
 421             "nhtd,nltd->nhlt",
 422             qr,
 423             self.rec_k[:, :, t0:t1],
 424         ) / math.sqrt(self.w_qr.size(1))
 425
 426         ar = ar.softmax(dim=2)  # nhlt
 427
 428         ar = F.dropout(ar, self.attention_dropout, self.training)
 429
 430         y = torch.einsum(
 431             "nhlt,nltd->nthd",
 432             ar,
 433             self.rec_v[:, :, t0:t1],
 434         ).flatten(2)
 435
 436         self.cache_y[:, t0:t1] = y @ self.w_o
 437
 438         return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
 439
 440
 441 ##############################
 442
 443
 444 # Returns a tensor with an additional index at rank win_dim, that move
 445 # along the same dimension as dim, on a domain {0...win_size-1}, and
 446 # dim is restricted on a domain reduced by win_size-1 values.
 447
 448
 449 def moving_window(x, dim, win_dim, win_size):
 450     size, stride = x.size(), x.stride()
 451     size = size[:dim] + (size[dim] - win_size + 1,) + size[dim + 1 :]
 452     size = size[:win_dim] + (win_size,) + size[win_dim:]
 453     stride = stride[:win_dim] + (stride[dim],) + stride[win_dim:]
 454
 455     return x.as_strided(size=size, stride=stride)
 456
 457
 458 ##############################
 459
 460
 461 class Caterpillar(nn.Module):
 462     def __init__(
 463         self,
 464         dim_model,
 465         dim_qk,
 466         dim_v,
 467         nb_heads,
 468         caterpillar_length,
 469         caterpillar_height,
 470         attention_dropout=0.0,
 471         len_max=1e5,
 472     ):
 473         super().__init__()
 474
 475         warnings.warn("Caterpillar", RuntimeWarning)
 476
 477         def randw(*d):
 478             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 479
 480         self.caterpillar_length = caterpillar_length
 481         self.caterpillar_height = caterpillar_height
 482         self.attention_dropout = attention_dropout
 483
 484         warnings.warn("flash back", RuntimeWarning)
 485         self.proba_flashback = 0.1
 486
 487         self.w_G = randw(nb_heads, caterpillar_height, dim_model)
 488         self.b_G = nn.Parameter(
 489             torch.full(
 490                 (nb_heads, caterpillar_height), -math.log(caterpillar_height - 1)
 491             )
 492         )
 493
 494         self.w_K = randw(nb_heads, dim_qk, dim_model)
 495         self.w_V = randw(nb_heads, dim_v, dim_model)
 496         self.w_Q = randw(nb_heads, dim_qk, dim_model)
 497         self.w_O = randw(dim_v * nb_heads, dim_model)
 498
 499         self.init_K_rec = randw(caterpillar_height, caterpillar_length, dim_qk)
 500         self.init_V_rec = randw(caterpillar_height, caterpillar_length, dim_v)
 501
 502     def reset_inner_loss(self):
 503         self.acc_attention = 0
 504         self.acc_nb = 0
 505
 506     def get_inner_loss(self):
 507         # warnings.warn("l2 regularization", RuntimeWarning)
 508         # return (self.acc_attention / self.acc_nb).pow(2).sum()
 509         return torch.tensor([0], device=self.w_Q.device)
 510
 511     def forward(self, bs):
 512         # Dimensions to make the source a bit clearer, that's needed
 513
 514         X, t0, t1 = bs.slice(), bs.first, bs.first + bs.nb
 515
 516         N = bs.x.size(0)
 517         T = bs.x.size(1)
 518         H = self.w_V.size(0)
 519         DV = self.w_V.size(1)
 520         DK = self.w_K.size(1)
 521         DM = self.w_O.size(1)
 522         CH = self.caterpillar_height
 523         CL = self.caterpillar_length
 524
 525         assert (
 526             t0 >= CL and (t1 - t0) % CL == 0
 527         ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
 528
 529         # We cache values to deal efficiently with auto-regression
 530
 531         if bs.init_cache:
 532             self.rec_V = X.new_zeros(N, CH, T, DV)
 533             self.rec_K = X.new_zeros(N, CH, T, DK)
 534             # We start the recurrent sequences with optimizable
 535             # initial values. No idea if it helps.
 536             self.rec_V[:, :, t0 - CL : t0] = self.init_V_rec[None, :, :, :]
 537             self.rec_K[:, :, t0 - CL : t0] = self.init_K_rec[None, :, :, :]
 538
 539             self.cache_Y = X.new_zeros(N, T, DM)
 540
 541         ######################################################################
 542         # Compute the recurrent state
 543
 544         # This is the Gating sequence that modulates the storing of
 545         # the new key and value in the CH pairs of the current
 546         # stack. The CH gating values are independent, which means
 547         # that the current K/V could be stored in multiple pairs of the
 548         # recurrent state, or not at all.
 549
 550         G = (
 551             torch.einsum("ntc,hec->nhet", X, self.w_G) + self.b_G[None, :, :, None]
 552         ).sigmoid()
 553
 554         # That bas a bad idea
 555         # G = F.dropout(G, self.attention_dropout, self.training)
 556
 557         V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
 558         K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
 559
 560         # We prepare the arguments for the parallel scan
 561
 562         A = 1 - G.sum(1)
 563         gated_V = torch.einsum("nhet,nhtd->netd", G, V)
 564         gated_K = torch.einsum("nhet,nhtd->netd", G, K)
 565
 566         init_rec_V = self.rec_V[:, :, t0 - CL : t0]
 567         init_rec_K = self.rec_K[:, :, t0 - CL : t0]
 568
 569         # Here there is a trick: Since the stack at time t is computed
 570         # by updating that at time t-L, the parallel scan operates
 571         # with a period of L. To do so we split the time indexing in
 572         # two axes, the second of size CL, and run the parallel scan
 573         # using the other as the sequence index.
 574
 575         A = A.unflatten(2, (-1, CL))
 576         gated_V = gated_V.unflatten(2, (-1, CL))
 577         gated_K = gated_K.unflatten(2, (-1, CL))
 578
 579         next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
 580         next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
 581
 582         # Put back the sequence index
 583
 584         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
 585         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 586
 587         if self.training and self.proba_flashback > 0.0:
 588             # This piece of code makes the assumption that there is
 589             # nothing informative before t0, otherwise we'd have to
 590             # implement a cache for V and K too. This should not be
 591             # too much of a problem since this is used only during
 592             # train, where full sequence are available
 593
 594             n = torch.arange(N, device=X.device)[:, None, None, None]
 595             t = torch.arange(t0, t1, device=X.device)[None, None, :, None]
 596             dv = torch.arange(DV, device=X.device)[None, None, None, :]
 597             dk = torch.arange(DK, device=X.device)[None, None, None, :]
 598
 599             u = (
 600                 torch.rand(N, CH, t1 - t0, 1, device=X.device).mul(t).long() // CL
 601             ) * CL
 602
 603             src_time = t - u - t0
 604             src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
 605
 606             mask_V = (
 607                 torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
 608             ).long()
 609             self.rec_V[:, :, t0:t1] = (
 610                 mask_V * V[n, src_head, src_time, dv]
 611                 + (1 - mask_V) * self.rec_V[:, :, t0:t1]
 612             )
 613
 614             mask_K = (
 615                 torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback
 616             ).long()
 617             self.rec_K[:, :, t0:t1] = (
 618                 mask_K * K[n, src_head, src_time, dk]
 619                 + (1 - mask_K) * self.rec_K[:, :, t0:t1]
 620             )
 621
 622         ######################################################################
 623         # compute the readout
 624
 625         Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
 626
 627         # We build tensors NxHxTxFxL where N is the sample index, H
 628         # the head, T the time, F the row in the caterpillar, and L
 629         # the column in the caterpillar
 630
 631         windowed_V = moving_window(
 632             self.rec_V[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 633         )
 634
 635         windowed_K = moving_window(
 636             self.rec_K[:, :, t0 - CL + 1 : t1], dim=2, win_dim=3, win_size=CL
 637         )
 638
 639         # We have an attention score for each of the CHxCL values
 640
 641         ar = torch.einsum(
 642             "nhtd,nftld->nhtfl",
 643             Q,
 644             windowed_K,
 645         ) / math.sqrt(DK)
 646
 647         # softmax can operate only on one dimension, hence the
 648         # flattening
 649
 650         ar = ar.flatten(3).softmax(dim=3).view(ar.size())
 651
 652         ar = F.dropout(ar, self.attention_dropout, self.training)
 653
 654         # Compute the output for each head, flatten to concatenate
 655
 656         Y = torch.einsum(
 657             "nhtfl,nftld->nthd",
 658             ar,
 659             windowed_V,
 660         ).flatten(2)
 661
 662         # Compute the final output
 663
 664         self.cache_Y[:, t0:t1] = Y @ self.w_O
 665
 666         return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
 667
 668
 669 ##############################
 670
 671
 672 class QKVAttention(nn.Module):
 673     def __init__(
 674         self,
 675         dim_model,
 676         dim_qk,
 677         dim_v,
 678         nb_heads=1,
 679         causal=False,
 680         attention_dropout=0.0,
 681     ):
 682         super().__init__()
 683
 684         def randw(*d):
 685             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 686
 687         self.causal = causal
 688         self.attention_dropout = attention_dropout
 689         self.record_attention = False
 690
 691         self.w_q = randw(nb_heads, dim_qk, dim_model)
 692         self.w_k = randw(nb_heads, dim_qk, dim_model)
 693         self.w_v = randw(nb_heads, dim_v, dim_model)
 694         self.w_o = randw(dim_v * nb_heads, dim_model)
 695
 696     def forward(self, bs):
 697         x_q = bs.x
 698
 699         assert (
 700             self.causal or bs.complete()
 701         ), "Partial evaluation is only possible for causal models"
 702
 703         if bs.init_cache:
 704             self.cache_k = x_q.new_zeros(
 705                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 706             )
 707             self.cache_v = x_q.new_zeros(
 708                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 709             )
 710             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 711
 712         q = torch.einsum("ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_q)
 713
 714         self.cache_k[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 715             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_k
 716         )
 717         self.cache_v[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
 718             "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_v
 719         )
 720
 721         a = torch.einsum(
 722             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs.first + bs.nb]
 723         ) / math.sqrt(self.w_q.size(1))
 724
 725         if self.causal:
 726             if bs.init_cache:
 727                 self.cache_attzero = (
 728                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 729                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 730                 )
 731             a = a.masked_fill(
 732                 self.cache_attzero[
 733                     :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb
 734                 ],
 735                 float("-inf"),
 736             )
 737
 738         a = a.softmax(dim=3)
 739
 740         if self.record_attention:
 741             self.a = a
 742
 743         a = F.dropout(a, self.attention_dropout, self.training)
 744
 745         y = torch.einsum(
 746             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs.first + bs.nb]
 747         ).flatten(2)
 748
 749         self.cache_y[:, bs.first : bs.first + bs.nb] = y @ self.w_o
 750
 751         return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
 752
 753
 754 ##############################
 755
 756
 757 class MyGPT(nn.Module):
 758     def __init__(
 759         self,
 760         vocabulary_size,
 761         dim_model,
 762         dim_keys,
 763         dim_hidden,
 764         nb_heads,
 765         nb_blocks,
 766         nb_lines=None,
 767         caterpillar_height=None,
 768         dim_rec_v=-1,
 769         causal=False,
 770         dropout=0.0,
 771         len_max=1e5,
 772         attention_layer="kvrec",
 773     ):
 774         super().__init__()
 775
 776         assert attention_layer in {"mha", "dumbrec", "kvrec", "caterpillar"}
 777
 778         if attention_layer == "caterpillar":
 779             assert nb_lines % caterpillar_height == 0
 780             self.caterpillar_length = nb_lines // caterpillar_height
 781             self.caterpillar_height = caterpillar_height
 782         else:
 783             self.caterpillar_length = -1
 784             self.caterpillar_height = -1
 785
 786         assert dim_model % nb_heads == 0
 787
 788         self.embedding = nn.Sequential(
 789             CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
 790             AddPositionalEncoding(len_max),
 791         )
 792
 793         trunk_blocks = []
 794
 795         def attlayer():
 796             if attention_layer == "mha":
 797                 return QKVAttention(
 798                     dim_model=dim_model,
 799                     dim_qk=dim_keys,
 800                     dim_v=dim_model // nb_heads,
 801                     nb_heads=nb_heads,
 802                     causal=causal,
 803                     attention_dropout=dropout,
 804                 )
 805             elif attention_layer == "dumbrec":
 806                 return DumbRec(
 807                     dim_model=dim_model,
 808                     dim_qk=dim_keys,
 809                     dim_v=dim_rec_v,
 810                     nb_heads=nb_heads,
 811                     nb_lines=nb_lines,
 812                     attention_dropout=dropout,
 813                 )
 814             elif attention_layer == "kvrec":
 815                 return KVRec(
 816                     dim_model=dim_model,
 817                     dim_qk=dim_keys,
 818                     dim_v=dim_rec_v,
 819                     nb_heads=nb_heads,
 820                     nb_lines=nb_lines,
 821                     attention_dropout=dropout,
 822                 )
 823             elif attention_layer == "caterpillar":
 824                 return Caterpillar(
 825                     dim_model=dim_model,
 826                     dim_qk=dim_keys,
 827                     dim_v=dim_rec_v,
 828                     nb_heads=nb_heads,
 829                     caterpillar_length=self.caterpillar_length,
 830                     caterpillar_height=self.caterpillar_height,
 831                     attention_dropout=dropout,
 832                 )
 833             else:
 834                 raise ValueError(f"Unknown attention type {attention_layer}.")
 835
 836         for b in range(nb_blocks):
 837             trunk_blocks += [
 838                 WithResidual(
 839                     CacheWrapper(nn.LayerNorm((dim_model,))),
 840                     attlayer(),
 841                 ),
 842                 WithResidual(
 843                     CacheWrapper(
 844                         nn.LayerNorm((dim_model,)),
 845                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 846                         nn.ReLU(),
 847                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 848                         nn.Dropout(dropout),
 849                     ),
 850                 ),
 851             ]
 852
 853         self.trunk = nn.Sequential(*trunk_blocks)
 854
 855         self.readout = CacheWrapper(
 856             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 857         )
 858
 859         with torch.no_grad():
 860             for m in self.modules():
 861                 if isinstance(m, nn.Embedding):
 862                     m.weight.normal_(mean=0, std=2e-2)
 863                 elif isinstance(m, nn.LayerNorm):
 864                     m.bias.zero_()
 865                     m.weight.fill_(1.0)
 866
 867         self.reset_inner_loss()
 868
 869     def forward(self, bs):
 870         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb, bs.init_cache)
 871
 872         # To make the code simpler in the Caterpillar layer, we pad
 873         # here. It's unclear if/how much it hurts computationaly by
 874         # increasing the sequence length for the other layers
 875
 876         if self.caterpillar_length > 0:
 877             original_nb = bs.nb
 878             if bs.nb % self.caterpillar_length > 0:
 879                 bs.nb += self.caterpillar_length - bs.nb % self.caterpillar_length
 880
 881             bs = BracketedSequence(
 882                 F.pad(bs.x, (self.caterpillar_length, self.caterpillar_length)),
 883                 bs.first + self.caterpillar_length,
 884                 bs.nb,
 885                 bs.init_cache,
 886             )
 887
 888         bs = self.embedding(bs)
 889         bs = self.trunk(bs)
 890         bs = self.readout(bs)
 891
 892         if self.caterpillar_length > 0:
 893             bs = BracketedSequence(
 894                 F.pad(bs.x, (0, 0, -self.caterpillar_length, -self.caterpillar_length)),
 895                 bs.first - self.caterpillar_length,
 896                 original_nb,
 897                 bs.init_cache,
 898             )
 899
 900         return bs
 901
 902     # ar_mask is a tensor with 0s and 1s, of same shape as input, with
 903     # 1s where tokens should be generated. The others are kept
 904     # unchanged.
 905
 906     def masked_inplace_autoregression(
 907         self,
 908         input_src,
 909         ar_mask_src,
 910         forbidden_tokens=None,
 911         deterministic_synthesis=False,
 912     ):
 913         input = input_src.to(self.readout.f.weight.device)
 914         ar_mask = ar_mask_src.to(self.readout.f.weight.device)
 915         to_generate = (ar_mask.sum(0) > 0).nonzero()
 916         if to_generate.min() > 0:
 917             self(
 918                 BracketedSequence(input, 0, to_generate.min(), True)
 919             )  # Needed to initialize the model's cache
 920         for s in range(to_generate.min(), to_generate.max() + 1):
 921             output = self(BracketedSequence(input, s, 1, s == 0)).x
 922             logits = output[:, s]
 923             if forbidden_tokens is not None:
 924                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 925             if deterministic_synthesis:
 926                 t_next = logits.argmax(1)
 927             else:
 928                 dist = torch.distributions.categorical.Categorical(logits=logits)
 929                 t_next = dist.sample()
 930             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 931
 932         input_src.copy_(input)
 933
 934     def reset_inner_loss(self):
 935         for m in self.modules():
 936             if m is not self and hasattr(m, "reset_inner_loss"):
 937                 m.reset_inner_loss()
 938
 939     def get_inner_loss(self):
 940         l = torch.tensor([0.0], device=self.readout.f.weight.device)
 941         for m in self.modules():
 942             if m is not self and hasattr(m, "get_inner_loss"):
 943                 l += m.get_inner_loss()
 944         return l
 945
 946     def record_attention(self, v=True):
 947         for m in self.modules():
 948             if isinstance(m, QKVAttention):
 949                 m.record_attention = v
 950
 951     def retrieve_attention(self):
 952         a = []
 953         for m in self.modules():
 954             if isinstance(m, QKVAttention):
 955                 a.append(m.a)
 956         return a
 957
 958
 959 ######################################################################
 960
 961 if __name__ == "__main__":
 962     print("Basic check.")
 963
 964     m = Caterpillar(
 965         dim_model=4,
 966         dim_qk=3,
 967         dim_v=7,
 968         nb_heads=1,
 969         caterpillar_length=7,
 970         caterpillar_height=3,
 971         attention_dropout=0.0,
 972     )
 973
 974     m.reset_inner_loss()
 975     x = torch.randn(1, 21 + 2 * 7, 4)
 976     y1 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 977     y2 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
 978     y3a = m(BracketedSequence(x, first=7, nb=14, init_cache=True)).x[:, 7:21]
 979     y3b = m(BracketedSequence(x, first=21, nb=7, init_cache=False)).x[:, 21:28]
 980     print((y1 - y2).abs().max())
 981     print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
 982     exit(0)
 983
 984     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 985
 986     vocabulary_size = 128
 987     x = torch.randint(vocabulary_size, (6, 1024))
 988
 989     model = MyGPT(
 990         vocabulary_size=vocabulary_size,
 991         dim_model=512,
 992         dim_keys=64,
 993         dim_hidden=2048,
 994         nb_heads=8,
 995         nb_lines=128,
 996         nb_blocks=12,
 997         dropout=0.1,
 998         causal=True,
 999     )
1000
1001     x = x.to(device)
1002     model.to(device)
1003
1004     import time, sys
1005
1006     # import torchvision.models as models
1007     # from torch.profiler import profile, record_function, ProfilerActivity
1008
1009     # with profile(activities=[ProfilerActivity.CPU,  ProfilerActivity.CUDA], profile_memory=True, record_shapes=True) as prof:
1010     # with record_function("model_inference"):
1011
1012     model.eval()
1013     for i in range(3):
1014         start_time = time.perf_counter()
1015         for k in range(10):
1016             model(BracketedSequence(x))
1017         duration = time.perf_counter() - start_time
1018         print(duration)
1019         sys.stdout.flush()
1020
1021     # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
1022     # print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
1023
1024     # print("##############################################################")
1025     # y2 = torch.randn_like(y1)
1026     # for s in range(x.size(1)):
1027     # z = model(BracketedSequence(x, s, 1))
1028     # y2[:, s : s + 1] = z.slice()
1029
1030     # print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
1031
1032 ######################################################################