3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
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.
13 # This implementation is equipped with RNN layers to replace the MHA
20 from torch.nn import functional as F
26 ######################################################################
28 # A BracketedSequence is a BxTx... tensor with a first and a nb time
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
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.
41 # Modules able to process brackets may implement a cache that is
42 # resetted when init_cache is True
45 class BracketedSequence:
46 def __init__(self, x, first=None, nb=None, init_cache=None):
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
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
57 return self.x[:, self.first : self.first + self.nb]
60 return self.first == 0 and self.nb == self.x.size(1)
63 ######################################################################
66 class CacheWrapper(nn.Module):
67 def __init__(self, *f):
69 self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
71 def forward(self, bs):
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
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())
81 return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
84 ##############################
87 class WithResidual(nn.Module):
88 def __init__(self, *f):
90 self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
92 def forward(self, bs):
93 return BracketedSequence(bs.x + self.f(bs).x, bs.first, bs.nb, bs.init_cache)
96 ##############################
99 class AddPositionalEncoding(nn.Module):
100 def __init__(self, len_max):
102 self.len_max = len_max
104 # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
106 def forward(self, bs):
108 t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
111 j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
116 t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
118 self.cache_y = bs.x.new(bs.x.size())
120 self.cache_y[:, bs.first : bs.first + bs.nb] = (
121 bs.slice() + self.pe[bs.first : bs.first + bs.nb]
124 return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
130 # X is /.../xTxD A is /.../xT Y_init is /.../xD
133 def pscan_dim(A, X, Y_init, dim=-2):
135 a, T, b = s[:dim].numel(), s[dim], s[dim + 1 :].numel()
137 A = A.reshape(a, T, *s[dim + 1 : -1])
138 X = X.reshape(a, T, *s[dim + 1 : -1], -1)
141 Y_init = X.new_zeros(a, *s[dim + 1 : -1], X.size(-1))
143 Y_init = Y_init.reshape(a, *s[dim + 1 : -1], -1)
145 Y = pscan.pscan(A, X, Y_init).reshape(s)
150 def pscan_shape(A, X, Y_init):
152 A = A.reshape(-1, s[-2])
153 X = X.reshape(-1, s[-2], s[-1])
156 Y_init = X.new_zeros(X.size(0), s[-1])
158 Y_init = Y_init.reshape(-1, s[-1])
160 Y = pscan.pscan(A, X, Y_init).reshape(s)
165 def nsum_shape(X, Y_init):
167 X = X.reshape(-1, s[-2], s[-1]) # ntd
169 Y = 0 if Y_init is None else Y_init.reshape(-1, s[-1])
172 for k in range(X.size(1)):
174 Y = Y / Y.norm(dim=-1, keepdim=True).clamp(min=1)
177 return torch.cat(result, dim=1).reshape(s)
180 ##############################
183 class DumbRec(nn.Module):
191 attention_dropout=0.0,
199 return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
201 self.nb_lines = nb_lines
202 self.attention_dropout = attention_dropout
204 self.k_star = randw(nb_lines, dim_qk)
206 self.w_qw = randw(nb_heads, dim_qk, dim_model)
207 self.w_qr = randw(nb_heads, dim_qk, dim_model)
208 # self.w_k = randw(nb_heads, dim_qk, dim_model)
209 self.w_v = randw(nb_heads, dim_v, dim_model)
210 self.w_o = randw(dim_v * nb_heads, dim_model)
212 def reset_inner_loss(self):
213 self.acc_attention = 0
216 def get_inner_loss(self):
217 warnings.warn("l2 regularization", RuntimeWarning)
218 return (self.acc_attention / self.acc_nb).pow(2).sum()
219 # return torch.tensor([0], device=self.w_qw.device)
221 def forward(self, bs):
222 x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
225 self.rec_v = x_q.new_zeros(
226 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
228 # self.rec_k = x_q.new_zeros(
229 # x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
231 self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
233 ######################################################################
236 k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
238 warnings.warn("rotating key barrel", RuntimeWarning)
239 k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
240 t_barrel = torch.arange(t0, t1, device=k_star.device)
241 t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
243 torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
245 k_star = k_star[l_barrel, t_barrel]
247 ######################################################################
248 # Compute the recurrent state
250 qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
252 v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
253 # k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
259 ) / math.sqrt(self.w_qw.size(1))
261 aw = aw.softmax(dim=2) # nhlt
264 self.acc_attention += aw.sum(dim=(0, 1, 3))
265 self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
267 aw = F.dropout(aw, self.attention_dropout, self.training)
269 A = 1 - aw.sum(dim=1) # nlt
271 V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
272 # K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
278 V0 = self.rec_v[:, :, t0 - 1]
279 # K0 = self.rec_k[:, :, t0 - 1]
281 self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
282 # self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
284 ######################################################################
285 # compute the readout
287 qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
292 # self.rec_k[:, :, t0:t1],
294 ) / math.sqrt(self.w_qr.size(1))
296 ar = ar.softmax(dim=2) # nhlt
298 ar = F.dropout(ar, self.attention_dropout, self.training)
303 self.rec_v[:, :, t0:t1],
306 self.cache_y[:, t0:t1] = y @ self.w_o
308 return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
311 ##############################
314 class KVRec(nn.Module):
322 attention_dropout=0.0,
330 return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
332 self.nb_lines = nb_lines
333 self.attention_dropout = attention_dropout
335 self.k_star = randw(nb_lines, dim_qk)
337 self.w_qw = randw(nb_heads, dim_qk, dim_model)
338 self.w_qr = randw(nb_heads, dim_qk, dim_model)
339 self.w_k = randw(nb_heads, dim_qk, dim_model)
340 self.w_v = randw(nb_heads, dim_v, dim_model)
341 self.w_o = randw(dim_v * nb_heads, dim_model)
343 def reset_inner_loss(self):
344 self.acc_attention = 0
347 def get_inner_loss(self):
348 warnings.warn("l2 regularization", RuntimeWarning)
349 return (self.acc_attention / self.acc_nb).pow(2).sum()
350 # return torch.tensor([0], device=self.w_qw.device)
351 # warnings.warn("side regularization", RuntimeWarning)
353 # (0.5 / self.nb_lines - self.acc_attention / self.acc_nb).clamp(min=0).sum()
355 # return torch.tensor([0], device=self.w_qw.device)
357 def forward(self, bs):
358 x_q, t0, t1 = bs.x, bs.first, bs.first + bs.nb
361 self.rec_v = x_q.new_zeros(
362 x_q.size(0), self.nb_lines, x_q.size(1), self.w_v.size(1)
364 self.rec_k = x_q.new_zeros(
365 x_q.size(0), self.nb_lines, x_q.size(1), self.w_k.size(1)
367 self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
369 ######################################################################
372 k_star = self.k_star[:, None, :].expand(-1, t1 - t0, -1)
374 warnings.warn("rotating key barrel", RuntimeWarning)
375 k_star = self.k_star[:, None, :].expand(-1, x_q.size(1), -1)
376 t_barrel = torch.arange(t0, t1, device=k_star.device)
377 t_barrel = t_barrel[None, :].expand(k_star.size(0), t1 - t0)
379 torch.arange(k_star.size(0), device=k_star.device)[:, None] + t_barrel
381 k_star = k_star[l_barrel, t_barrel]
383 ######################################################################
384 # Compute the recurrent state
386 qw = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qw)
388 v = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_v)
389 k = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_k)
395 ) / math.sqrt(self.w_qw.size(1))
397 aw = aw.softmax(dim=2) # nhlt
400 # We want all the memory lines to be used similarly
401 self.acc_attention += aw.sum(dim=(0, 1, 3)) # Sum accross NxHx_xT
402 self.acc_nb += aw.size(0) * aw.size(1) * aw.size(3)
404 aw = F.dropout(aw, self.attention_dropout, self.training)
406 A = 1 - aw.sum(dim=1) # nlt
408 V = torch.einsum("nhlt,nhtd->nltd", aw, v).contiguous()
409 K = torch.einsum("nhlt,nhtd->nltd", aw, k).contiguous()
415 V0 = self.rec_v[:, :, t0 - 1]
416 K0 = self.rec_k[:, :, t0 - 1]
418 self.rec_v[:, :, t0:t1] = pscan_shape(A, V, V0)
419 self.rec_k[:, :, t0:t1] = pscan_shape(A, K, K0)
421 ######################################################################
422 # compute the readout
424 qr = torch.einsum("ntc,hdc->nhtd", x_q[:, t0:t1], self.w_qr)
429 self.rec_k[:, :, t0:t1],
430 ) / math.sqrt(self.w_qr.size(1))
432 ar = ar.softmax(dim=2) # nhlt
434 ar = F.dropout(ar, self.attention_dropout, self.training)
439 self.rec_v[:, :, t0:t1],
442 self.cache_y[:, t0:t1] = y @ self.w_o
444 return BracketedSequence(self.cache_y, t0, t1 - t0, bs.init_cache)
447 ##############################
450 # Returns a tensor with an additional index at rank win_dim, that move
451 # along the same dimension as dim, on a domain {0...win_size-1}, and
452 # dim is restricted on a domain reduced by win_size-1 values.
455 def moving_window(x, dim, win_dim, win_size):
456 size, stride = x.size(), x.stride()
457 size = size[:dim] + (size[dim] - win_size + 1,) + size[dim + 1 :]
458 size = size[:win_dim] + (win_size,) + size[win_dim:]
459 stride = stride[:win_dim] + (stride[dim],) + stride[win_dim:]
461 return x.as_strided(size=size, stride=stride)
464 ##############################
467 class Caterpillar(nn.Module):
476 attention_dropout=0.0,
483 warnings.warn("Caterpillar", RuntimeWarning)
485 def randw(*d, amplitude=None):
486 if amplitude is None:
487 amplitude = 1 / math.sqrt(d[-1])
488 return nn.Parameter(amplitude * torch.randn(*d))
490 self.caterpillar_length = caterpillar_length
491 self.caterpillar_height = caterpillar_height
492 self.attention_dropout = attention_dropout
494 ######################################################################
497 x = kwargs.get("gate_dropout")
499 self.proba_gate_dropout = 0.0
501 self.proba_gate_dropout = float(x)
503 logger(f"self.proba_gate_dropout {self.proba_gate_dropout}")
505 x = kwargs.get("default_bg")
507 default_bg = -math.log(caterpillar_height - 1)
509 default_bg = float(x)
511 logger(f"default_bg {default_bg}")
513 ######################################################################
515 self.w_G = randw(nb_heads, caterpillar_height, dim_model)
516 self.b_G = nn.Parameter(torch.full((nb_heads, caterpillar_height), default_bg))
518 self.w_K = randw(nb_heads, dim_qk, dim_model)
519 self.w_V = randw(nb_heads, dim_v, dim_model)
520 self.w_Q = randw(nb_heads, dim_qk, dim_model)
521 self.w_O = randw(dim_v * nb_heads, dim_model)
523 self.init_K_rec = randw(
528 self.init_V_rec = randw(
534 def reset_inner_loss(self):
535 self.acc_attention = 0
538 def get_inner_loss(self):
539 # warnings.warn("l2 regularization", RuntimeWarning)
540 # return (self.acc_attention / self.acc_nb).pow(2).sum()
541 return torch.tensor([0], device=self.w_Q.device)
543 def forward(self, bs):
544 # Dimensions to make the source a bit clearer, that's needed
546 X, t0, t1 = bs.slice(), bs.first, bs.first + bs.nb
551 DV = self.w_V.size(1)
552 DK = self.w_K.size(1)
553 DM = self.w_O.size(1)
554 R = self.caterpillar_height
555 L = self.caterpillar_length
558 t0 >= L and (t1 - t0) % L == 0
559 ), f"bs.first should be greater than caterpillar_length, and bs.nb should be a multiple of caterpillar_length"
561 # We cache values to deal efficiently with auto-regression
564 self.rec_V = X.new_zeros(N, R, T, DV)
565 self.rec_K = X.new_zeros(N, R, T, DK)
566 # We start the recurrent sequences with optimizable
567 # initial values. No idea if it helps.
568 self.rec_V[:, :, t0 - L : t0] = self.init_V_rec[None, :, :, :]
569 self.rec_K[:, :, t0 - L : t0] = self.init_K_rec[None, :, :, :]
571 self.cache_Y = X.new_zeros(N, T, DM)
573 V = torch.einsum("ntc,hdc->nhtd", X, self.w_V)
574 K = torch.einsum("ntc,hdc->nhtd", X, self.w_K)
576 ######################################################################
577 # Compute the recurrent state
579 # This is the Gating sequence that modulates the storing of
580 # the new key and value in the R pairs of the current
581 # stack. There are R independent gating values, which means
582 # that the current K/V may be stored in multiple pairs of the
583 # recurrent state, or not at all.
586 torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
589 # warnings.warn("softmax gating", RuntimeWarning)
592 # torch.einsum("ntc,hrc->nhrt", X, self.w_G) + self.b_G[None, :, :, None]
595 ######################################################################
598 if self.training and self.proba_gate_dropout > 0.0:
599 # This is a better implementation of "flashbacks".
601 # G is NxHxExT where e is the caterpillar's row.
603 warnings.warn("gate dropout", RuntimeWarning)
606 torch.rand(G.size(), device=G.device) <= self.proba_gate_dropout
609 alpha = G / (1 - self.proba_gate_dropout)
611 G = alpha * (1 - kill)
613 ######################################################################
614 # Clip the gating to avoid values greater than 1 when several
615 # heads hit the same row
617 G = G / G.sum(1, keepdim=True).clamp(min=1)
619 ######################################################################
620 # Roll the gating indexes
622 # warnings.warn("rotating barrel", RuntimeWarning)
624 # r_barrel = torch.arange(R, device=G.device)[None, None, :, None]
625 # t_barrel = torch.arange(t1 - t0, device=G.device)[None, None, None, :]
626 # r_barrel = (r_barrel + (t_barrel + t0) // L) % R
627 # G = G.gather(dim=2, index=r_barrel.expand_as(G))
629 # We prepare the arguments for the parallel scan
633 # warnings.warn("harmonic recurrence", RuntimeWarning)
634 # har = torch.arange(t0, t1, device = G.device).float() + 1
635 # A = har / (har + 1)
638 gated_V = torch.einsum("nhrt,nhtd->nrtd", G, V)
639 gated_K = torch.einsum("nhrt,nhtd->nrtd", G, K)
641 # We start from cached values, which matters in inference
643 init_rec_V = self.rec_V[:, :, t0 - L : t0]
644 init_rec_K = self.rec_K[:, :, t0 - L : t0]
646 #################################################################
649 # Here there is a trick: Since the stack at position t is
650 # computed by updating that at position t-L, the parallel
651 # scan operates with a period of L. To do so we split the
652 # sequence indexing in two axes, the second of size L, and
653 # run the parallel scan using the first as the sequence index.
655 A = A.unflatten(2, (-1, L))
656 gated_V = gated_V.unflatten(2, (-1, L))
657 gated_K = gated_K.unflatten(2, (-1, L))
659 next_V = pscan_dim(A, gated_V, init_rec_V, dim=2)
660 next_K = pscan_dim(A, gated_K, init_rec_K, dim=2)
662 self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
663 self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
665 ######################################################################
666 # compute the readout
668 Q = torch.einsum("ntc,hdc->nhtd", X, self.w_Q)
670 # We build tensors NxHxTxFxL where N is the sample index, H
671 # the head, T the time, F the row in the caterpillar, and L
672 # the column in the caterpillar
674 windowed_V = moving_window(
675 self.rec_V[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
678 windowed_K = moving_window(
679 self.rec_K[:, :, t0 - L + 1 : t1], dim=2, win_dim=3, win_size=L
682 # We have an attention score for each of the RxL values
690 # softmax can operate only on one dimension, hence the
693 ar = ar.flatten(3).softmax(dim=3).view(ar.size())
695 ar = F.dropout(ar, self.attention_dropout, self.training)
697 # Compute the output for each head, flatten to concatenate
705 # Compute the final output
707 self.cache_Y[:, t0:t1] = Y @ self.w_O
709 return BracketedSequence(self.cache_Y, t0, t1 - t0, bs.init_cache)
712 ##############################
715 class QKVAttention(nn.Module):
723 attention_dropout=0.0,
730 return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
733 self.attention_dropout = attention_dropout
734 self.record_attention = False
736 self.w_q = randw(nb_heads, dim_qk, dim_model)
737 self.w_k = randw(nb_heads, dim_qk, dim_model)
738 self.w_v = randw(nb_heads, dim_v, dim_model)
739 self.w_o = randw(dim_v * nb_heads, dim_model)
741 def forward(self, bs):
745 self.causal or bs.complete()
746 ), "Partial evaluation is only possible for causal models"
749 self.cache_k = x_q.new_zeros(
750 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
752 self.cache_v = x_q.new_zeros(
753 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
755 self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
757 q = torch.einsum("ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_q)
759 self.cache_k[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
760 "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_k
762 self.cache_v[:, :, bs.first : bs.first + bs.nb] = torch.einsum(
763 "ntc,hdc->nhtd", x_q[:, bs.first : bs.first + bs.nb], self.w_v
767 "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs.first + bs.nb]
768 ) / math.sqrt(self.w_q.size(1))
772 self.cache_attzero = (
773 torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
774 < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
778 :, :, bs.first : bs.first + bs.nb, : bs.first + bs.nb
785 if self.record_attention:
788 a = F.dropout(a, self.attention_dropout, self.training)
791 "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs.first + bs.nb]
794 self.cache_y[:, bs.first : bs.first + bs.nb] = y @ self.w_o
796 return BracketedSequence(self.cache_y, bs.first, bs.nb, bs.init_cache)
799 ##############################
802 class MyGPT(nn.Module):
812 caterpillar_height=None,
816 attention_layer="kvrec",
822 assert attention_layer in {
827 }, f"Unknown attention operator {attention_layer}."
829 if attention_layer == "caterpillar":
830 assert nb_lines % caterpillar_height == 0
831 self.caterpillar_length = nb_lines // caterpillar_height
832 self.caterpillar_height = caterpillar_height
834 self.caterpillar_length = -1
835 self.caterpillar_height = -1
837 assert dim_model % nb_heads == 0
839 self.embedding = nn.Sequential(
840 CacheWrapper(nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)),
841 AddPositionalEncoding(len_max),
847 if attention_layer == "mha":
851 dim_v=dim_model // nb_heads,
854 attention_dropout=dropout,
858 elif attention_layer == "dumbrec":
862 dim_v=dim_model // nb_heads,
865 attention_dropout=dropout,
869 elif attention_layer == "kvrec":
873 dim_v=dim_model // nb_heads,
876 attention_dropout=dropout,
880 elif attention_layer == "caterpillar":
884 dim_v=dim_model // nb_heads,
886 caterpillar_length=self.caterpillar_length,
887 caterpillar_height=self.caterpillar_height,
888 attention_dropout=dropout,
893 raise ValueError(f"Unknown attention type {attention_layer}.")
895 for b in range(nb_blocks):
898 CacheWrapper(nn.LayerNorm((dim_model,))),
903 nn.LayerNorm((dim_model,)),
904 nn.Linear(in_features=dim_model, out_features=dim_hidden),
906 nn.Linear(in_features=dim_hidden, out_features=dim_model),
912 self.trunk = nn.Sequential(*trunk_blocks)
914 self.readout = CacheWrapper(
915 nn.Linear(in_features=dim_model, out_features=vocabulary_size)
918 with torch.no_grad():
919 for m in self.modules():
920 if isinstance(m, nn.Embedding):
921 m.weight.normal_(mean=0, std=2e-2)
922 elif isinstance(m, nn.LayerNorm):
926 self.reset_inner_loss()
928 def forward(self, bs):
929 bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb, bs.init_cache)
931 # To make the code simpler in the Caterpillar layer, we pad
932 # here. It's unclear if/how much it hurts computationaly by
933 # increasing the sequence length for the other layers
935 if self.caterpillar_length > 0:
937 if bs.nb % self.caterpillar_length > 0:
938 bs.nb += self.caterpillar_length - bs.nb % self.caterpillar_length
940 bs = BracketedSequence(
941 F.pad(bs.x, (self.caterpillar_length, self.caterpillar_length)),
942 bs.first + self.caterpillar_length,
947 bs = self.embedding(bs)
949 bs = self.readout(bs)
951 if self.caterpillar_length > 0:
952 bs = BracketedSequence(
953 F.pad(bs.x, (0, 0, -self.caterpillar_length, -self.caterpillar_length)),
954 bs.first - self.caterpillar_length,
961 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
962 # 1s where tokens should be generated. The others are kept
965 def masked_inplace_autoregression(
969 forbidden_tokens=None,
970 deterministic_synthesis=False,
972 input = input_src.to(self.readout.f.weight.device)
973 ar_mask = ar_mask_src.to(self.readout.f.weight.device)
974 to_generate = (ar_mask.sum(0) > 0).nonzero()
975 if to_generate.min() > 0:
977 BracketedSequence(input, 0, to_generate.min(), True)
978 ) # Needed to initialize the model's cache
979 for s in range(to_generate.min(), to_generate.max() + 1):
980 output = self(BracketedSequence(input, s, 1, s == 0)).x
981 logits = output[:, s]
982 if forbidden_tokens is not None:
983 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
984 if deterministic_synthesis:
985 t_next = logits.argmax(1)
987 dist = torch.distributions.categorical.Categorical(logits=logits)
988 t_next = dist.sample()
989 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
991 input_src.copy_(input)
993 def reset_inner_loss(self):
994 for m in self.modules():
995 if m is not self and hasattr(m, "reset_inner_loss"):
998 def get_inner_loss(self):
999 l = torch.tensor([0.0], device=self.readout.f.weight.device)
1000 for m in self.modules():
1001 if m is not self and hasattr(m, "get_inner_loss"):
1002 l += m.get_inner_loss()
1005 def record_attention(self, v=True):
1006 for m in self.modules():
1007 if isinstance(m, QKVAttention):
1008 m.record_attention = v
1010 def retrieve_attention(self):
1012 for m in self.modules():
1013 if isinstance(m, QKVAttention):
1018 ######################################################################
1020 if __name__ == "__main__":
1021 print("Basic check.")
1028 caterpillar_length=7,
1029 caterpillar_height=3,
1030 attention_dropout=0.0,
1033 m.reset_inner_loss()
1034 x = torch.randn(1, 21 + 2 * 7, 4)
1035 y1 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
1036 y2 = m(BracketedSequence(x, first=7, nb=21, init_cache=True)).x[:, 7:28]
1037 y3a = m(BracketedSequence(x, first=7, nb=14, init_cache=True)).x[:, 7:21]
1038 y3b = m(BracketedSequence(x, first=21, nb=7, init_cache=False)).x[:, 21:28]
1039 print((y1 - y2).abs().max())
1040 print((y1 - torch.cat([y3a, y3b], dim=1)).abs().max())
1043 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
1045 vocabulary_size = 128
1046 x = torch.randint(vocabulary_size, (6, 1024))
1049 vocabulary_size=vocabulary_size,
1065 # import torchvision.models as models
1066 # from torch.profiler import profile, record_function, ProfilerActivity
1068 # with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA], profile_memory=True, record_shapes=True) as prof:
1069 # with record_function("model_inference"):
1073 start_time = time.perf_counter()
1075 model(BracketedSequence(x))
1076 duration = time.perf_counter() - start_time
1080 # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=10))
1081 # print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
1083 # print("##############################################################")
1084 # y2 = torch.randn_like(y1)
1085 # for s in range(x.size(1)):
1086 # z = model(BracketedSequence(x, s, 1))
1087 # y2[:, s : s + 1] = z.slice()
1089 # print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
1091 ######################################################################