File size: 35,223 Bytes
a517ecd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 |
from __future__ import annotations
from typing import Callable
import math
from functools import partial
from itertools import zip_longest
from collections import namedtuple
import torch
from torch import nn, stack, cat, is_tensor, tensor, Tensor
import torch.nn.functional as F
from torch.nn import Linear, Module, Parameter, ParameterList, ParameterDict
from torch.func import functional_call, vmap, grad
from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten
from tensordict import TensorDict
from assoc_scan import AssocScan
from titans_pytorch.memory_models import(
MemoryMLP,
ResidualNorm
)
import einx
from einops import einsum, rearrange, repeat, reduce, pack, unpack
from einops.layers.torch import Rearrange, Reduce
"""
ein notation:
b - batch
h - heads
bh - batch and heads
n - sequence
d - feature dimension
c - intra-chunk
w - num memory network weight parameters
o - momentum orders
u - key / value updates - allowing a token to emit multiple key / values
"""
LinearNoBias = partial(Linear, bias = False)
# neural mem state related
NeuralMemState = namedtuple('NeuralMemState', [
'seq_index',
'weights',
'cache_store_segment',
'states',
'updates',
])
def mem_state_detach(
state: NeuralMemState
):
assert isinstance(state, NeuralMemState)
state = tree_map(lambda t: t.detach() if is_tensor(t) else t, tuple(state))
return NeuralMemState(*state)
# functions
def exists(v):
return v is not None
def default(*args):
for arg in args:
if exists(arg):
return arg
return None
def identity(t):
return t
def xnor(x, y):
return not (x ^ y)
def divisible_by(num, den):
return (num % den) == 0
def safe_cat(inputs, dim = -2):
inputs = tuple(filter(exists, inputs))
if len(inputs) == 0:
return None
elif len(inputs) == 1:
return inputs[0]
return cat(inputs, dim = dim)
def is_empty_tensor(t):
return t.numel() == 0
def dict_get_value_shapes(td):
return [v.shape for k, v in td.items()]
def rearrange_dict_values(td, pattern, **kwargs):
return td.apply(lambda t: rearrange(t, pattern, **kwargs))
def repeat_dict_values(td, pattern, **kwargs):
return td.apply(lambda t: repeat(t, pattern, **kwargs))
def pair(v):
return (v, v) if not isinstance(v, tuple) else v
def round_down_multiple(seq, mult):
return seq // mult * mult
def round_up_multiple(seq, mult):
return math.ceil(seq / mult) * mult
def pad_at_dim(t, pad, dim = -1, value = 0.):
dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
zeros = ((0, 0) * dims_from_right)
return F.pad(t, (*zeros, *pad), value = value)
def pack_one_with_inverse(t, pattern):
packed, packed_shape = pack([t], pattern)
def inverse(out, inv_pattern = None):
inv_pattern = default(inv_pattern, pattern)
return unpack(out, packed_shape, inv_pattern)[0]
return packed, inverse
def Sequential(*modules):
modules = [*filter(exists, modules)]
if len(modules) == 0:
return nn.Identity()
if len(modules) == 1:
return modules[0]
return nn.Sequential(*modules)
# softclamping gradients
def softclamp_max(t, max_value):
half_max_value = max_value / 2
return ((t / half_max_value).tanh() * half_max_value) + half_max_value
def softclamp_grad_norm(t, max_value, eps: float = 1e-6):
if is_empty_tensor(t):
return t
t, inverse = pack_one_with_inverse(t, 'bn *')
norm = t.norm(dim = -1, keepdim = True).clamp(min = eps)
clamped_norm = softclamp_max(norm, max_value)
t = t * (clamped_norm / norm)
return inverse(t)
# spectral norming the surprise update w/ newton schulz matrix iter
# Keller Jordan et al. from OSS w/ nanogpt, now being used for two works, Atlas and 'TTT done right'
def newtonschulz5(
t,
steps = 5,
eps = 1e-7,
coefs = (3.4445, -4.7750, 2.0315)
):
if t.ndim <= 3:
return t
shape = t.shape
should_transpose = shape[-2] > shape[-1]
if should_transpose:
t = t.transpose(-1, -2)
t, inv_pack = pack_one_with_inverse(t, '* i j')
t = t / t.norm(dim = (-1, -2), keepdim = True).clamp(min = eps)
a, b, c = coefs
for _ in range(steps):
A = t @ t.transpose(-1, -2)
B = b * A + c * A @ A
t = a * t + B @ t
if should_transpose:
t = t.transpose(-1, -2)
return inv_pack(t)
# multi head rmsnorm
class MultiheadRMSNorm(Module):
def __init__(self, dim, heads):
super().__init__()
self.rmsnorm = nn.RMSNorm(dim, elementwise_affine = False)
self.gamma = Parameter(torch.zeros(heads, 1, dim))
def forward(self, x):
return self.rmsnorm(x) * (self.gamma + 1.)
# chunk pooling
class AveragePool(Module):
def __init__(
self,
chunk_size
):
super().__init__()
self.chunk_size = chunk_size
def forward(
self,
x,
chunk_size = None
):
chunk_size = default(chunk_size, self.chunk_size)
return reduce(x, 'b (n c) d -> b n d', 'mean', c = chunk_size)
class AttentionPool(Module):
def __init__(
self,
dim,
chunk_size
):
"""
taken from Enformer https://www.nature.com/articles/s41592-021-01252-x , in turn taken from somewhere else
"""
super().__init__()
self.chunk_size = chunk_size
self.to_attn_logits = nn.Linear(dim, dim)
# default to average pool
nn.init.zeros_(self.to_attn_logits.weight)
nn.init.zeros_(self.to_attn_logits.bias)
def forward(
self,
x,
chunk_size = None
):
chunk_size = default(chunk_size, self.chunk_size)
x = rearrange(x, 'b (n c) d -> b n c d', c = chunk_size)
attn_logits = self.to_attn_logits(x)
attn = attn_logits.softmax(dim = -2)
return reduce(x * attn, 'b n c d -> b n d', 'sum')
# main neural memory
def default_adaptive_step_transform(adaptive_step, max_lr = 1e-2):
return adaptive_step.sigmoid() * max_lr
def default_loss_fn(pred, target):
return (pred - target).pow(2).mean(dim = -1)
class NeuralMemory(Module):
def __init__(
self,
dim,
chunk_size: int | tuple[int, int] = 1,
batch_size = None,
dim_head = None,
heads = 1,
model: Module | None = None,
store_memory_loss_fn: Callable = default_loss_fn,
adaptive_step_transform: Callable | None = None,
default_step_transform_max_lr = 1.,
per_parameter_lr_modulation = False, # allow outer network to control learning rate per weight matrix of memory network
max_mem_layer_modulation = 1., # max of 10.
per_head_learned_parameters = True,
attn_pool_chunks = False,
momentum = True,
momentum_order = 1,
learned_momentum_combine = False,
learned_combine_include_zeroth = False,
num_kv_per_token = 1, # whether a single token can do multiple updates to the memory model
qkv_receives_diff_views = False, # to address an issue raised by a phd student (who will be credited if experiments are green). basically the issue raised is that the memory MLP is only learning Wk @ Wv linear mapping and that may not be expressive enough. we will use hyper connections to allow the network to choose different previous layer inputs as keys / values and see if that does anything
pre_rmsnorm = True,
post_rmsnorm = False,
qk_rmsnorm = False,
max_grad_norm: float | None = None,
use_accelerated_scan = False,
activation: Module | None = None,
init_adaptive_step_bias = None,
init_momentum_bias = None,
init_decay_bias = None,
accept_weight_residual = False,
spectral_norm_surprises = False,
gated_transition = False,
mem_model_norm_add_residual = True, # by default, layernorm output and add residual as proposed in TTT paper, but could be removed
default_model_kwargs: dict = dict(
depth = 2,
expansion_factor = 4.
)
):
super().__init__()
dim_head = default(dim_head, dim)
assert not (heads == 1 and dim_head != dim)
self.retrieve_chunk_size, self.store_chunk_size = pair(chunk_size)
# batch size
if exists(batch_size):
assert divisible_by(batch_size, self.store_chunk_size)
self.batch_size = batch_size
# associative scan
self.assoc_scan = AssocScan(use_accelerated = use_accelerated_scan)
# key values receiving different views
self.qkv_receives_diff_views = qkv_receives_diff_views
# norms
self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
self.store_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
self.multihead_rmsnorm = MultiheadRMSNorm(dim_head, heads) if post_rmsnorm else nn.Identity()
self.q_norm = MultiheadRMSNorm(dim_head, heads) if qk_rmsnorm else nn.Identity()
self.k_norm = MultiheadRMSNorm(dim_head, heads) if qk_rmsnorm else nn.Identity()
# maybe multi-headed
dim_inner = dim_head * heads
self.heads = heads
self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
self.split_kv_heads = Rearrange('b n (h u d) -> b h (n u) d', h = heads, u = num_kv_per_token)
self.merge_heads = Rearrange('b h n d -> b n (h d)')
self.combine_heads = LinearNoBias(dim_inner, dim) if heads > 1 else nn.Identity()
self.retrieve_gate = Sequential(
LinearNoBias(dim, heads),
Rearrange('b n h -> b h n 1'),
nn.Sigmoid()
) if heads > 1 else None
# memory model
if not exists(model):
model = MemoryMLP(dim_head, **default_model_kwargs)
# validate memory model
assert not exists(next(model.buffers(), None)), 'model cannot have buffers for now'
test_shape = (3, 2, dim_head)
with torch.no_grad():
try:
test_input = torch.randn(test_shape)
mem_model_output = model(test_input)
except:
raise RuntimeError(f'memory model unable to accept a tensor of shape {test_shape}')
assert mem_model_output.shape == test_shape, 'output of memory model needs to be same shape as input'
# the memory is the weights of the model
if mem_model_norm_add_residual:
model = ResidualNorm(dim = dim_head, model = model)
self.memory_model = model
mem_model_params = dict(model.named_parameters())
self.num_memory_parameter_tensors = len(mem_model_params)
self.memory_model_parameter_names = [*mem_model_params.keys()]
memory_model_parameters = [*mem_model_params.values()]
if per_head_learned_parameters:
memory_model_parameters = [repeat(p, '... -> h ...', h = heads) for p in memory_model_parameters]
self.init_weight_shape = [p.shape for p in memory_model_parameters]
self.memory_model_parameters = ParameterList(memory_model_parameters)
self.per_head_learned_parameters = per_head_learned_parameters
# the chunk size within the paper where adaptive step, momentum, weight decay are shared
self.chunk_size = chunk_size
# prepare function for per sample gradients from model above, using torch.func
def forward_and_loss(params, inputs, loss_weights, target):
pred = functional_call(self.memory_model, params, inputs)
loss = self.store_memory_loss_fn(pred, target) # simple mse loss in paper - eq (12) - |M(k) - v|²
weighted_loss = loss * loss_weights
return weighted_loss.sum(), loss
# two functions
grad_fn = grad(forward_and_loss, has_aux = True)
self.per_sample_grad_fn = vmap(grad_fn, in_dims = (0, 0, 0, 0))
# queries for retrieving from the model
self.to_queries = Sequential(LinearNoBias(dim, dim_inner), activation)
# keys and values for storing to the model
assert num_kv_per_token > 0
self.to_keys = Sequential(
LinearNoBias(dim, dim_inner * num_kv_per_token),
activation,
)
self.to_values = Sequential(
LinearNoBias(dim, dim_inner * num_kv_per_token),
activation,
)
self.store_memory_loss_fn = store_memory_loss_fn
self.num_kv_per_token = num_kv_per_token
# `chunk_size` refers to chunk size used for storing to memory model weights
chunk_size = self.store_chunk_size
# whether to use averaging of chunks, or attention pooling
assert not (attn_pool_chunks and chunk_size == 1), '`attn_pool_chunks` cannot be set to True if `chunk_size` is set to 1'
if not attn_pool_chunks:
self.reduce_to_chunk_rep = AveragePool(chunk_size = chunk_size)
else:
self.reduce_to_chunk_rep = AttentionPool(dim, chunk_size = chunk_size)
# learned adaptive learning rate
self.to_adaptive_step = Sequential(
nn.Linear(dim, heads * num_kv_per_token),
Rearrange('b n (h u) -> (b h) (n u)', u = num_kv_per_token)
)
if not exists(adaptive_step_transform):
adaptive_step_transform = partial(default_adaptive_step_transform, max_lr = default_step_transform_max_lr)
self.adaptive_step_transform = adaptive_step_transform
# momentum related
self.to_momentum = Sequential(
nn.Linear(dim, heads * momentum_order),
Rearrange('b n (h o) -> o (b h) n 1', o = momentum_order)
) if momentum else None
self.momentum_order = momentum_order
self.to_learned_momentum_combine = None
if learned_momentum_combine:
assert momentum
assert momentum_order > 1, 'only second order momentum allowed for now, but may allow learned combination of zeroth'
if learned_combine_include_zeroth:
momentum_order += 1
self.to_learned_momentum_combine = Sequential(
nn.Linear(dim, heads * momentum_order),
Rearrange('b n (h o) -> o (b h) n', h = heads),
nn.Softmax(dim = 0),
)
self.learned_combine_include_zeroth = learned_combine_include_zeroth
# per layer learning rate modulation
self.to_layer_modulation = Sequential(
nn.Linear(dim, heads * self.num_memory_parameter_tensors),
Rearrange('b n (h w) -> w (b h) n', h = heads),
nn.Sigmoid()
) if per_parameter_lr_modulation else None
self.max_mem_layer_modulation = max_mem_layer_modulation
# learned weight residual
self.to_learned_weight_residual_mix = Sequential(
nn.Linear(dim, heads),
Rearrange('b n h -> b h n'),
nn.Sigmoid()
) if accept_weight_residual else None
# allow for softclamp the gradient norms for storing memories
self.max_grad_norm = max_grad_norm
# spectral norming the surprises before update, a la Muon from Jordan et al.
self.spectral_norm_surprises = spectral_norm_surprises
# weight decay factor
self.to_decay_factor = Sequential(
nn.Linear(dim, heads),
Rearrange('b n h -> (b h) n 1')
)
# learned transition, as seeing instability when decreasing neural mem batch size
# perhaps it can slowly learn to adjust from early residual to fully transitioning to new weights every batch size
self.transition_gate = nn.Parameter(tensor(-5.)) if gated_transition else None
# inits
if exists(init_adaptive_step_bias):
linear = self.to_adaptive_step[0]
nn.init.zeros_(linear.weight)
nn.init.constant_(linear.bias, init_adaptive_step_bias)
if exists(init_momentum_bias):
linear = self.to_momentum[0]
nn.init.zeros_(linear.weight)
nn.init.constant_(linear.bias, init_momentum_bias)
if exists(init_decay_bias):
linear = self.to_decay_factor[0]
nn.init.zeros_(linear.weight)
nn.init.constant_(linear.bias, init_decay_bias)
# maybe use accelerated scan
self.use_accelerated_scan = use_accelerated_scan
self.register_buffer('zero', torch.tensor(0.), persistent = False)
@property
def memory_model_parameter_dict(self):
return TensorDict(dict(zip(self.memory_model_parameter_names, self.memory_model_parameters)))
def init_weights(
self,
batch,
):
if self.per_head_learned_parameters:
weights = repeat_dict_values(self.memory_model_parameter_dict, 'h ... -> (b h) ...', b = batch)
else:
weights = repeat_dict_values(self.memory_model_parameter_dict, '... -> bh ...', bh = batch * self.heads)
return weights
def init_momentum(
self,
batch,
):
zeros = self.memory_model_parameter_dict.clone().zero_()
if self.per_head_learned_parameters:
zeros = repeat_dict_values(zeros, 'h ... -> o (b h) ...', b = batch, o = self.momentum_order)
else:
zeros = repeat_dict_values(zeros, '... -> o bh ...', bh = batch * self.heads, o = self.momentum_order)
return zeros
def store_memories(
self,
seq,
weights: dict[str, Tensor] | None = None,
past_state: tuple[dict[str, Tensor], dict[str, Tensor]] | None = None,
seq_index = 0,
prev_weights = None,
mask: Tensor | None = None,
return_surprises = True
):
if self.qkv_receives_diff_views:
_, batch, seq_len = seq.shape[:3]
else:
batch, seq_len = seq.shape[:2]
# shapes and variables
heads, chunk_size, num_updates = self.heads, self.store_chunk_size, self.num_kv_per_token
# curtail sequence by multiple of the chunk size
# only a complete chunk of the sequence provides the memory for the next chunk
round_down_seq_len = round_down_multiple(seq_len, chunk_size)
num_chunks = round_down_seq_len // chunk_size
seq, remainder = seq[..., :round_down_seq_len, :], seq[..., round_down_seq_len:, :]
next_seq_len_index = seq_index + round_down_seq_len
# init weights if needed
# weights of the memory network
if not exists(weights):
weights = self.init_weights(batch)
weights = TensorDict(weights)
# allow for neural memory of a previous layer to influence surprise of current layer
weights_for_surprise = repeat_dict_values(weights, 'b ... -> b n ...', n = num_chunks)
# initial norm
seq = self.store_norm(seq)
# handle keys and values coming from different sequences from hyper connection
values_seq = seq
if self.qkv_receives_diff_views:
seq, values_seq = seq
# derive learned hparams for optimization of memory network
adaptive_lr = self.to_adaptive_step(seq)
adaptive_lr = self.adaptive_step_transform(adaptive_lr)
chunked_seq = self.reduce_to_chunk_rep(seq, chunk_size = chunk_size)
decay_factor = self.to_decay_factor(chunked_seq).sigmoid()
need_layer_lr_mod = exists(self.to_layer_modulation) and num_chunks > 0
has_momentum = exists(self.to_momentum)
if has_momentum:
adaptive_momentum = self.to_momentum(chunked_seq).sigmoid()
learned_combine = exists(self.to_learned_momentum_combine)
if learned_combine:
combine_momentums = self.to_learned_momentum_combine(chunked_seq)
if need_layer_lr_mod:
layer_lr_mod = self.to_layer_modulation(chunked_seq) * self.max_mem_layer_modulation
# keys and values
keys = self.to_keys(seq)
values = self.to_values(values_seq)
# maybe multi head
keys, values = map(self.split_kv_heads, (keys, values))
# maybe keys rmsnorm
keys = self.k_norm(keys)
# take care of chunking
keys, values = tuple(rearrange(t, 'b h (n c u) d -> (b h n) (c u) d', c = chunk_size, u = num_updates) for t in (keys, values))
# adaptive lr
adaptive_lr = rearrange(adaptive_lr, 'b (n c u) -> (b n) (c u)', c = chunk_size, u = num_updates)
# optionally a storing memories mask can be passed in. if False, will set the learning rate to 0. for those positions
if exists(mask):
mask = mask[..., :round_down_seq_len]
mask = repeat(mask, 'b (n c) -> (b h n) (c u)', h = heads, u = num_updates, c = chunk_size)
adaptive_lr = torch.where(mask, adaptive_lr, 0.)
# maybe add previous layer weight
assert xnor(exists(self.to_learned_weight_residual_mix), exists(prev_weights))
if exists(prev_weights):
start_index = math.ceil(seq_index / chunk_size)
end_index = start_index + num_chunks
prev_weights = prev_weights.apply(lambda t: t[:, start_index:end_index])
if exists(self.to_learned_weight_residual_mix) and num_chunks > 0:
mix = self.to_learned_weight_residual_mix(chunked_seq)
mix = rearrange(mix, 'b h n -> (b h) n')
prev_weights = prev_weights.apply(lambda t: einx.multiply('bh n, bh n ... -> bh n ...', mix, t))
weights_for_surprise = weights_for_surprise + prev_weights
# flatten batch and time if surprise depends on previous layer memory model
weights_for_surprise = rearrange_dict_values(weights_for_surprise, 'b n ... -> (b n) ...')
# get grads and extra auxiliary loss (for backwarding through qkv projection in base neural memory module)
grads, unweighted_mem_model_loss = self.per_sample_grad_fn(dict(weights_for_surprise), keys, adaptive_lr, values)
grads = TensorDict(grads)
# surprises
adaptive_lr = rearrange(adaptive_lr, '(b h n) c -> b h (n c)', b = batch, h = heads)
unweighted_mem_model_loss = rearrange(unweighted_mem_model_loss, '(b h n) c -> b h (n c)', b = batch, h = heads)
# maybe softclamp grad norm
if exists(self.max_grad_norm):
grads = grads.apply(lambda t: softclamp_grad_norm(t, self.max_grad_norm))
# restore batch and sequence dimension
grads = rearrange_dict_values(grads, '(b n) ... -> b n ...', b = batch * heads)
# maybe per layer modulation
if need_layer_lr_mod:
grads = TensorDict({name: einx.multiply('b h, b h ... -> b h ...', layer_lr_mod, t) for layer_lr_mod, (name, t) in zip(layer_lr_mod, grads.items())})
# negative gradients, adaptive lr already applied as loss weight
surprises = grads.mul(-1)
# past states
if not exists(past_state):
# minibatch_init_weight corresponds to W0 in figure 7 of TTT paper
minibatch_init_weight = weights
init_momentum = self.init_momentum(batch)
past_state = (minibatch_init_weight, init_momentum)
past_last_update, past_last_momentum = past_state
# early return if sequence length less than chunk size
if num_chunks == 0:
updates = rearrange_dict_values(weights, 'bh ... -> bh 1 ...')
next_store_state = NeuralMemState(next_seq_len_index, weights, remainder, past_state, updates)
output = (updates, next_store_state)
if not return_surprises:
return output
return (*output, (unweighted_mem_model_loss, adaptive_lr))
# momentum + weight decay - momentum is the new contribution, as most linear RNNs have learned forgetting gates
updates = TensorDict()
next_last_update = TensorDict()
next_last_momentum = TensorDict()
for (param_name, surprise), (_, last_update) in zip(surprises.items(), past_last_update.items()):
update = surprise
# derive momentum with associative scan - eq (10)
if has_momentum:
momentum = surprise
momentums = [] # stores all momentum orders starting with first, to generalize to Nth order momentum
last_momentum = past_last_momentum[param_name]
# go from first order momentum all the way to the Nth
for one_adaptive_momentum, one_last_momentum in zip_longest(adaptive_momentum, last_momentum):
momentum = self.assoc_scan(one_adaptive_momentum, momentum, prev = one_last_momentum) # momentum is S / surprise in the paper
momentums.append(momentum)
momentums = stack(momentums)
next_last_momentum[param_name] = momentums[:, :, -1] # momentums shape is Float['o bh n 1']
if learned_combine and self.learned_combine_include_zeroth:
# add the original surprise if learned combination of momentums
momentums = cat((rearrange(surprise, '... -> 1 ...'), momentums), dim = 0)
if not learned_combine:
update = momentums[-1]
else:
update = einsum(combine_momentums, momentums, 'o b n, o b n ... -> b n ...')
# maybe spectral norm surprises
if self.spectral_norm_surprises:
update = newtonschulz5(update)
# use associative scan again for learned forgetting (weight decay) - eq (13)
update = self.assoc_scan(1. - decay_factor, update, prev = last_update, remove_prev = False)
updates[param_name] = update
next_last_update[param_name] = update[:, -1]
# determine next state for the storing of memories
next_state = (next_last_update, next_last_momentum)
next_store_state = NeuralMemState(next_seq_len_index, weights, remainder, next_state, updates)
# return updates to neural memory at all chunked timesteps + neural mem cache / state to be fed back
if not return_surprises:
return updates, next_store_state
return updates, next_store_state, (unweighted_mem_model_loss, adaptive_lr)
def retrieve_memories(
self,
seq,
weights: dict[str, Tensor],
):
chunk_size = self.retrieve_chunk_size
weights_have_expanded_shape = dict_get_value_shapes(weights) != self.init_weight_shape
batch, seq_len = seq.shape[:2]
# auto infer single token decoding, if there are only 1 set of weights and 1 token
is_one_token = seq_len == 1
is_one_weight = (not weights_have_expanded_shape) or next(iter(weights.values())).shape[1] == 1
is_single_token_decode = is_one_token and is_one_weight
if is_single_token_decode:
chunk_size = 1
# padding related, for chunked processing
need_pad = chunk_size > 1 or not is_one_weight
if need_pad:
seq = pad_at_dim(seq, (1, 0), dim = 1)
seq_len_plus_one = seq.shape[-2]
next_seq_len = round_up_multiple(seq_len_plus_one, chunk_size)
padding = next_seq_len - seq_len_plus_one
seq = pad_at_dim(seq, (0, padding), dim = 1)
# the parameters of the memory model stores the memories of the key / values
# when the MLP has only 1 weight matrix, it is equivalent to `kv` fast weight memories from linear attention literature (recall fetching of memories is q @ (kv)) / schmidhuber's paper
weights = TensorDict(weights)
# pre norm
seq = self.retrieve_norm(seq)
# sequence Float['b n d'] to queries
queries = self.to_queries(seq)
# maybe multihead
queries = self.split_heads(queries)
# maybe qk rmsnorm
queries = self.q_norm(queries)
# fetch values from memory model
if weights_have_expanded_shape:
weights = rearrange_dict_values(weights, 'b n ... -> (b n) ...')
queries = rearrange(queries, 'b h (n c) d -> (b h n) c d', c = chunk_size)
# forward functional call
values = functional_call(self.memory_model, dict(weights), queries)
# reconstitute batch dimension
values = rearrange(values, '(b h n) c d -> b h (n c) d', b = batch, h = self.heads)
values = self.multihead_rmsnorm(values)
# maybe gate
if exists(self.retrieve_gate):
values = values * self.retrieve_gate(seq)
# maybe merge heads and combine
values = self.merge_heads(values)
values = self.combine_heads(values)
# restore, pad with empty memory embed
if need_pad:
values = values[:, 1:]
return values[:, :seq_len]
def forward(
self,
seq,
store_seq = None,
state: NeuralMemState | None = None,
detach_mem_state = False,
prev_weights = None,
store_mask: Tensor | None = None,
return_surprises = False,
ttt_batch_size: int | None = None
):
is_multi_input = self.qkv_receives_diff_views
# handle single token
if seq.ndim == 2 or (is_multi_input and seq.ndim == 3):
seq = rearrange(seq, '... b d -> ... b 1 d')
is_single_token = seq.shape[-2] == 1
# if different views for qkv, then
if is_multi_input:
retrieve_seq, seq = seq[0], seq[1:]
else:
retrieve_seq = seq
# handle previous state init
if not exists(state):
state = (0, None, None, None, None)
seq_index, weights, cache_store_seq, past_state, updates = state
# store
store_seq = default(store_seq, seq)
# take care of cache
if exists(cache_store_seq):
store_seq = safe_cat((cache_store_seq, store_seq))
if exists(store_mask):
cache_len = cache_store_seq.shape[-2]
if cache_len > 0:
cache_mask = torch.ones(
(*store_mask.shape[:-1], cache_len),
device=store_mask.device,
dtype=store_mask.dtype,
)
store_mask = torch.cat((cache_mask, store_mask), dim=-1)
if exists(store_mask) and store_mask.shape[-1] != store_seq.shape[-2]:
diff = store_seq.shape[-2] - store_mask.shape[-1]
if diff > 0:
store_seq = store_seq[..., diff:, :]
elif diff < 0:
store_mask = store_mask[..., (-diff):]
# compute split sizes of sequence
# for now manually update weights to last update at the correct boundaries
store_seq_len, chunk_size, batch_size = store_seq.shape[-2], self.chunk_size, default(ttt_batch_size, self.batch_size)
need_update_weights = exists(batch_size)
# determine split sizes and when to update
if need_update_weights:
update_after_final_store = divisible_by(seq_index + store_seq_len, batch_size)
seq_range = torch.arange(store_seq_len) + seq_index + 1
batch_boundary = divisible_by(seq_range, batch_size)
indices = seq_range[batch_boundary] - seq_index
indices = F.pad(indices, (1, 0), value = 0)
if indices[-1] != store_seq_len:
indices = F.pad(indices, (0, 1), value = store_seq_len)
split_sizes = (indices[1:] - indices[:-1]).tolist()
assert sum(split_sizes) == store_seq_len
else:
split_sizes = (store_seq_len,)
update_after_final_store = False
# accumulate updates
updates = None
def accum_updates(past_updates, future_updates):
if not exists(past_updates):
return future_updates
return TensorDict({param_name: cat((past_update[:, :-1], future_update), dim = 1) for (param_name, past_update), (_, future_update) in zip(past_updates.items(), future_updates.items())})
# loop through chunks of store sequences
store_seqs = store_seq.split(split_sizes, dim = -2)
if exists(store_mask):
store_masks = store_mask.split(split_sizes, dim = -1)
else:
store_masks = (None,) * len(split_sizes)
# whether to allow network to slowly adjust from initial weight throughout (residual path) to fully updating weights every batch
surprises = (None, None)
gate = None
if exists(self.transition_gate):
gate = self.transition_gate.sigmoid()
for ind, (store_seq_chunk, maybe_store_mask) in enumerate(zip(store_seqs, store_masks)):
is_last = ind == (len(store_seqs) - 1)
# store
next_updates, next_neural_mem_state, chunk_surprises = self.store_memories(
store_seq_chunk,
weights,
seq_index = seq_index,
past_state = past_state,
prev_weights = prev_weights,
mask = maybe_store_mask,
return_surprises = True
)
weights = next_neural_mem_state.weights
seq_index = next_neural_mem_state.seq_index
past_state = next_neural_mem_state.states
updates = accum_updates(updates, next_updates)
surprises = tuple(safe_cat(args, dim = -1) for args in zip(surprises, chunk_surprises))
if is_last and not update_after_final_store:
continue
# update weights once batch size is fulfilled
last_update, last_momentum = past_state
if exists(gate):
last_update = TensorDict({param_name: one_weight.lerp(one_last_update, gate) for (param_name, one_weight), (_, one_last_update) in zip(weights.items(), last_update.items())})
past_state = (last_update, last_momentum)
# set weights to the last updated weights for the last minibatch
weights = last_update
next_neural_mem_state = next_neural_mem_state._replace(
weights = weights,
states = past_state,
)
next_neural_mem_state = next_neural_mem_state._replace(updates = updates)
# retrieve
if is_single_token:
last_update, _ = next_neural_mem_state.states
updates = rearrange_dict_values(last_update, 'b ... -> b 1 ...')
retrieved = self.retrieve_memories(
retrieve_seq,
updates
)
# maybe detach
if detach_mem_state:
next_neural_mem_state = mem_state_detach(next_neural_mem_state)
# returning
if not return_surprises:
return retrieved, next_neural_mem_state
return retrieved, next_neural_mem_state, surprises
|