-
Notifications
You must be signed in to change notification settings - Fork 87
/
Copy pathmodeling_longllama.py
1467 lines (1232 loc) · 63.4 KB
/
modeling_longllama.py
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
# coding=utf-8
# Copyright 2023 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch LongLLaMA model."""
from dataclasses import dataclass
import math
from typing import List, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.activations import ACT2FN
from transformers.modeling_outputs import (
BaseModelOutputWithPast,
CausalLMOutputWithPast,
SequenceClassifierOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
replace_return_docstrings,
)
from .configuration_longllama import LongLlamaConfig
from .longllama_utils import mem_apply_update, LongLlamaMemCache, LongLlamaMemConfig
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "LongLlamaConfig"
@dataclass
class LongLlamaModelOutputWithPast(BaseModelOutputWithPast):
"""
Based on BaseModelOutputWithPast
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
mem_caches (`tuple(LongLlamaMemCache))`, *optional*, returned for layers with memory cache enabled):
For the layers without memory None is returned
"""
mem_caches: Optional[LongLlamaMemCache] = None
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
def _make_causal_mask(
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
):
"""
Make causal mask used for bi-directional self-attention.
"""
bsz, tgt_len = input_ids_shape
mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
mask_cond = torch.arange(mask.size(-1), device=device)
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
mask = mask.to(dtype)
if past_key_values_length > 0:
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
# Copied from transformers.models.bart.modeling_bart._expand_mask
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
"""
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
"""
bsz, src_len = mask.size()
tgt_len = tgt_len if tgt_len is not None else src_len
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
inverted_mask = 1.0 - expanded_mask
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->LongLlama
class LongLlamaRMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
"""
LongLlamaRMSNorm is equivalent to T5LayerNorm
"""
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
return self.weight * hidden_states.to(input_dtype)
# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->LongLlama
class LongLlamaRotaryEmbedding(torch.nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
super().__init__()
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# Build here to make `torch.jit.trace` work.
self._set_cos_sin_cache(
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
)
def _set_cos_sin_cache(self, seq_len, device, dtype):
self.max_seq_len_cached = seq_len
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
)
def rotate_half(x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
# Based on transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def rotate_one(x, cos, sin, position_ids):
if len(position_ids.shape) != 2 or x.shape[0] != position_ids.shape[0] or x.shape[-2] != position_ids.shape[1]:
raise ValueError(f"Position ids shoud have shape [bsz, seq_len] got {position_ids.shape}")
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
x_embed = (x * cos) + (rotate_half(x) * sin)
return x_embed
def rotate_as_if_first(x, rotary_emb):
# x: [bs, num_attention_heads, seq_len, head_size]
# apply rotary as if all elements were first in the sequence
cos, sin = rotary_emb(x, x.shape[-2])
return rotate_one(x, cos, sin, torch.zeros(x.shape[0], x.shape[-2], dtype=torch.long, device=cos.device))
# Based on an 4.30 transformers.models.llama.modeling_llama.LlamaMLP with Llama->LongLlama
class LongLlamaMLP(nn.Module):
def __init__(
self,
hidden_size: int,
intermediate_size: int,
hidden_act: str,
):
super().__init__()
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.act_fn = ACT2FN[hidden_act]
def forward(self, x):
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
# Modified transformers.models.llama.modeling_llama.LlamaAttention
class LongLlamaAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper with FoT modifications"""
def __init__(self, config: LongLlamaConfig, mem_config: Optional[LongLlamaMemConfig] = None):
super().__init__()
self.config = config
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.max_position_embeddings = config.max_position_embeddings
self.max_cache = self.max_position_embeddings
self.rope_theta = config.rope_theta
if (self.head_dim * self.num_heads) != self.hidden_size:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
self._init_rope()
self.mem_config = mem_config
def _init_rope(self):
assert self.config.rope_scaling is None
self.rotary_emb = LongLlamaRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: bool = False,
use_cache: bool = False,
mem_cache: Optional[LongLlamaMemCache] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if attention_mask is None:
tgt_seq_len = hidden_states.shape[-2]
if past_key_value is not None:
src_seq_len = past_key_value[0].shape[-2] + tgt_seq_len
else:
src_seq_len = tgt_seq_len
attention_mask = torch.zeros(
hidden_states.shape[0],
1,
tgt_seq_len,
src_seq_len,
device=hidden_states.device,
dtype=hidden_states.dtype,
)
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
position_ids = position_ids[:, None, :, None]
if position_ids.shape != (key_states.shape[0], 1, key_states.shape[-2], 1):
raise ValueError("position_ids should match batch and seq_len of the input")
mem_no_local_cache = self.mem_config is not None and past_key_value is None and (not use_cache)
mem_and_local_cache = self.mem_config is not None and use_cache
# positonal embeddings can be disabled for memory layers
use_positionals = self.mem_config is None or self.mem_config.positionals
if mem_no_local_cache:
# the whole context window will be moved to memory cache after the attention
if use_positionals:
# positionally embedd memory content as first token in the sequence
rfst_key_states = rotate_as_if_first(key_states, self.rotary_emb)
else:
rfst_key_states = key_states
# attention_mask [bsz, 1, tgt_seq_len, src_seq_len]
# we base the mask on the last token in the context window
mem_update = LongLlamaMemCache(
keys=rfst_key_states.to(self.mem_config.cache_dtype),
values=value_states.to(self.mem_config.cache_dtype),
masks=attention_mask[..., -1, :, None],
)
if past_key_value is not None:
past_local_cache_size = past_key_value[0].shape[-2]
key_states = torch.cat([past_key_value[0], key_states], dim=-2)
value_states = torch.cat([past_key_value[1], value_states], dim=-2)
# FoT additionally stores position_ids to support long inputs
position_ids = torch.cat([past_key_value[2], position_ids], dim=-2)
if attention_mask.shape[-1] != key_states.shape[-2] and attention_mask.shape[-2] != query_states.shape[-2]:
raise ValueError("attention_mask should be provided for all key_states in local context")
# local cache is maintained so that it is <= self.max_cache
# remaining elements are either dropped or go to memory cache
if key_states.shape[-2] > self.max_cache:
num_elems_to_drop = past_local_cache_size
if mem_and_local_cache:
drop_keys = key_states[:, :, :num_elems_to_drop, :]
drop_values = value_states[:, :, :num_elems_to_drop, :]
# as memory mask use the masking of the last key in context
# attention_mask [bsz, 1, tgt_seq_len, src_seq_len]
drop_masks = attention_mask[..., -1, :, None]
drop_masks = drop_masks[:, :, :num_elems_to_drop, :]
if use_positionals:
rfst_drop_keys = rotate_as_if_first(drop_keys, self.rotary_emb)
else:
rfst_drop_keys = drop_keys
mem_update = LongLlamaMemCache(
keys=rfst_drop_keys.to(self.mem_config.cache_dtype),
values=drop_values.to(self.mem_config.cache_dtype),
masks=drop_masks,
)
if mem_cache is None:
mem_cache = mem_update
else:
mem_cache = mem_apply_update(
prev_mem_cache=mem_cache, new_mem_content=mem_update, mem_config=self.mem_config
)
key_states = key_states[:, :, num_elems_to_drop:, :]
value_states = value_states[:, :, num_elems_to_drop:, :]
position_ids = position_ids[:, :, num_elems_to_drop:, :]
attention_mask = attention_mask[..., num_elems_to_drop:]
# FoT additionally stores position_ids to support long inputs
past_key_value = (key_states, value_states, position_ids) if use_cache else None
kv_seq_len = key_states.shape[-2]
if use_positionals:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
rel_pos_ids = position_ids - torch.min(position_ids, dim=-2, keepdim=True)[0]
rel_pos_ids = rel_pos_ids.squeeze(3).squeeze(1)
query_states = rotate_one(query_states, cos, sin, rel_pos_ids[:, -query_states.shape[-2] :])
key_states = rotate_one(key_states, cos, sin, rel_pos_ids)
if self.mem_config is not None and self.mem_config.attention_grouping is not None:
attn_grouping_h, attn_grouping_q = self.mem_config.attention_grouping
if attn_grouping_h <= 0 or attn_grouping_q <= 0:
raise ValueError("Attention grouping should be positive")
else:
attn_grouping_h, attn_grouping_q = self.num_heads, q_len
attn_output_h = []
for beg_h in range(0, self.num_heads, attn_grouping_h):
end_h = min(beg_h + attn_grouping_h, self.num_heads)
attn_output_q = []
for beg_q in range(0, q_len, attn_grouping_q):
end_q = min(beg_q + attn_grouping_q, q_len)
if self.config.torch_attention:
if mem_cache is not None:
attn_keys = torch.concat(
[key_states[:, beg_h:end_h], mem_cache.keys[:, beg_h:end_h].to(key_states.dtype)], dim=-2
)
attn_values = torch.concat(
[value_states[:, beg_h:end_h], mem_cache.values[:, beg_h:end_h].to(value_states.dtype)],
dim=-2,
)
mem_mask = mem_cache.masks.squeeze(-1).unsqueeze(-2)
assert len(mem_mask.shape) == 4
assert mem_mask.shape[2] == 1
assert mem_mask.shape[3] == mem_cache.keys.shape[-2]
mem_mask = torch.broadcast_to(
mem_mask, (mem_mask.shape[0], mem_mask.shape[1], end_q - beg_q, mem_mask.shape[3])
)
attn_mask = torch.concat([attention_mask[:, :, beg_q:end_q], mem_mask], dim=-1)
assert attn_mask.shape[-1] == attn_keys.shape[-2]
else:
attn_keys = key_states[:, beg_h:end_h]
attn_values = value_states[:, beg_h:end_h]
attn_mask = attention_mask[:, :, beg_q:end_q]
attn_queries = query_states[:, beg_h:end_h, beg_q:end_q]
attn_output = torch.nn.functional.scaled_dot_product_attention(
query=attn_queries, key=attn_keys, value=attn_values, attn_mask=attn_mask
)
attn_output_q.append(attn_output)
else:
attn_weights = torch.matmul(
query_states[:, beg_h:end_h, beg_q:end_q], key_states[:, beg_h:end_h].transpose(2, 3)
) / math.sqrt(self.head_dim)
if attn_weights.size() != (bsz, end_h - beg_h, end_q - beg_q, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, end_h - beg_h, end_q - beg_q, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask[:, :, beg_q:end_q]
min_value = (
torch.finfo(attn_weights.dtype).min
if -1000000.0 < torch.finfo(attn_weights.dtype).min
else -1000000.0
)
attn_weights = torch.max(
attn_weights, torch.tensor(min_value, device=attn_weights.device, dtype=attn_weights.dtype)
)
if mem_cache is not None:
mem_mask = mem_cache.masks.squeeze(-1).unsqueeze(-2)
mem_attn_weights = torch.matmul(
query_states[:, beg_h:end_h, beg_q:end_q],
mem_cache.keys[:, beg_h:end_h].transpose(2, 3).to(key_states.dtype),
) / math.sqrt(self.head_dim)
assert mem_mask.shape[2] == 1
mem_attn_weights = mem_attn_weights + mem_mask
min_value = (
torch.finfo(mem_attn_weights.dtype).min
if -1000000.0 < torch.finfo(mem_attn_weights.dtype).min
else -1000000.0
)
mem_attn_weights = torch.max(
mem_attn_weights,
torch.tensor(min_value, device=mem_attn_weights.device, dtype=mem_attn_weights.dtype),
)
attn_weights = torch.concat([attn_weights, mem_attn_weights], dim=-1)
combined_value_states = torch.concat(
[value_states[:, beg_h:end_h], mem_cache.values[:, beg_h:end_h].to(value_states.dtype)],
dim=-2,
)
else:
combined_value_states = value_states[:, beg_h:end_h]
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
query_states.dtype
)
attn_output = torch.matmul(attn_weights, combined_value_states)
assert attn_output.shape[-2] == end_q - beg_q
attn_output_q.append(attn_output)
attn_output_h.append(torch.concat(attn_output_q, dim=-2))
attn_output = torch.concat(attn_output_h, dim=-3)
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
if mem_no_local_cache:
if mem_cache is not None:
mem_cache = mem_apply_update(
prev_mem_cache=mem_cache, new_mem_content=mem_update, mem_config=self.mem_config
)
else:
mem_cache = mem_update
return attn_output, attn_weights, past_key_value, mem_cache
# Modified transformers.models.llama.modeling_llama.LlamaDecoderLayer
class LongLlamaDecoderLayer(nn.Module):
def __init__(self, config: LongLlamaConfig, mem_config: Optional[LongLlamaMemConfig] = None):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = LongLlamaAttention(config=config, mem_config=mem_config)
self.mlp = LongLlamaMLP(
hidden_size=self.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
)
self.input_layernorm = LongLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = LongLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
mem_cache: Optional[LongLlamaMemCache] = None,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
along with information about positions
mem_cache (`LongLlamaMemCache`, *optional*): memory cache for specific layers
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
hidden_states, self_attn_weights, present_key_value, mem_cache = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
mem_cache=mem_cache,
)
hidden_states = residual + hidden_states
# Fully Connected
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights,)
if use_cache:
outputs += (present_key_value,)
return outputs + (mem_cache,)
LONGLLAMA_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`LongLlamaConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
LONGLLAMA_MEML_DOCSTRING = r"""
mem_layers ([`int`], *optional*):
Indices of layers to be augmented with memory, if None then parameters from config will be used
mem_dtype (`str`, *optional*):
Keys and values will be casted to this type for storage.
"""
@add_start_docstrings(
"The bare LongLLaMA Model outputting raw hidden-states without any specific head on top.",
LONGLLAMA_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaPreTrainedModel with Llama->LongLlama
class LongLlamaPreTrainedModel(PreTrainedModel):
config_class = LongLlamaConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["LongLlamaDecoderLayer"]
_skip_keys_device_placement = "past_key_values"
_keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
def _init_weights(self, module):
std = self.config.initializer_range
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, LongLlamaModel):
module.gradient_checkpointing = value
LONGLLAMA_COMMON_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
information on the default strategy.
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings.
[What are position IDs?](../glossary#position-ids)
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`
or memory cache is enabled):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 1 additional tensor of shape
`(batch_size, 1, sequence_length, 1)`. For memory enriched layers it also contains content of memory cache.
It is padded with empty tensors so when returned it alwyas has 6 elements.
Contains pre-computed hidden-states (key and values in the self-attention blocks)
that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail. This is NOT supported in LongLlamaForCausalLM and LongLlamaForSequenceClassification
due to the specific input processing.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
LONGLLAMA_MODEL_INPUTS_DOCSTRING = r"""
mem_caches (`tuple(LongLlamaMemCache)`, *optional*)
Memory caches for specified layers, None for others
"""
LONGLLAMA_ADD_INPUTS_DOCSTRING = r"""
last_context_length (`int`, *optional*)
Useful for generation, specifies number of tokens that won't be loaded to memory and
will be left for generation cache
"""
def _prepare_pos_ids(past_key_values, batch_size, input_length, device):
if past_key_values is not None:
# take previous max pos_id + 1
if past_key_values[0][2].shape[0] != batch_size:
raise ValueError(
f"first dimension of past_key_values should match batch size: {batch_size}"
f"but got {past_key_values[0][2].shape[0]}"
)
next_pos = torch.max(past_key_values[0][2].view(batch_size, -1), dim=-1)[0] + 1
next_pos = next_pos.view(batch_size, 1)
else:
next_pos = torch.zeros(batch_size, 1, device=device, dtype=torch.long)
position_ids = torch.arange(0, input_length, dtype=torch.long, device=device).view(1, input_length)
position_ids = position_ids + next_pos
return position_ids
@add_start_docstrings(
"The bare LongLLaMA Model outputting raw hidden-states without any specific head on top.",
LONGLLAMA_START_DOCSTRING,
LONGLLAMA_MEML_DOCSTRING,
)
# Modified transformers.models.llama.modeling_llama.LlamaModel
class LongLlamaModel(LongLlamaPreTrainedModel):
"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LongLlamaDecoderLayer`]
Args:
config: LlamaConfig
"""
def __init__(self, config: LongLlamaConfig):
super().__init__(config)
self.mem_layers = config.mem_layers
self.mem_config = LongLlamaMemConfig(
positionals=config.mem_positionals,
cache_dtype=getattr(torch, config.mem_dtype),
attention_grouping=config.mem_attention_grouping,
)
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
for mem_layer_id in self.mem_layers:
if mem_layer_id < 0 or mem_layer_id >= config.num_hidden_layers:
raise ValueError(
f"Memory layer ids should be between 0 and {config.num_hidden_layers}, got {mem_layer_id}"
)
layers = []
for layer_id in range(config.num_hidden_layers):
if layer_id in self.mem_layers:
layer = LongLlamaDecoderLayer(config, mem_config=self.mem_config)
else:
layer = LongLlamaDecoderLayer(config, mem_config=None)
layers.append(layer)
self.layers = nn.ModuleList(layers)
self.norm = LongLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
# create causal mask
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
combined_attention_mask = None
if input_shape[-1] > 1:
combined_attention_mask = _make_causal_mask(
input_shape,
inputs_embeds.dtype,
device=inputs_embeds.device,
past_key_values_length=past_key_values_length,
)
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
inputs_embeds.device
)
combined_attention_mask = (
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
)
return combined_attention_mask
@add_start_docstrings_to_model_forward(LONGLLAMA_COMMON_INPUTS_DOCSTRING, LONGLLAMA_MODEL_INPUTS_DOCSTRING)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
mem_caches: Optional[Tuple[Optional[LongLlamaMemCache]]] = None,
) -> Union[Tuple, LongLlamaModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# retrieve input_ids and inputs_embeds
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
elif input_ids is not None:
batch_size, seq_length = input_ids.shape
elif inputs_embeds is not None:
batch_size, seq_length, _ = inputs_embeds.shape
else:
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
seq_length_with_past = seq_length
past_key_values_length = 0
if past_key_values is not None:
past_key_values_length = past_key_values[0][0].shape[-2]
seq_length_with_past = seq_length_with_past + past_key_values_length
if position_ids is None:
device = input_ids.device if input_ids is not None else inputs_embeds.device
position_ids = _prepare_pos_ids(past_key_values, batch_size, seq_length, device)
else:
position_ids = position_ids.view(-1, seq_length).long()
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
# embed positions
if attention_mask is None:
attention_mask = torch.ones(
(batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
)
attention_mask = self._prepare_decoder_attention_mask(
attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
)
hidden_states = inputs_embeds
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
next_decoder_cache = ()
next_mem_caches = ()
for idx, decoder_layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
past_key_value = past_key_values[idx] if past_key_values is not None else None
mem_cache = mem_caches[idx] if mem_caches else None
if mem_cache is not None and idx not in self.mem_layers:
raise ValueError("Memory cache provided for a non-memory leayer")
if (
self.gradient_checkpointing
and self.training
and mem_cache is None
and idx % self.config.gradient_checkpoint_every_ith == 0
):
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, output_attentions, None, mem_cache=None)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(decoder_layer),
hidden_states,
attention_mask,
position_ids,
None,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
mem_cache=mem_cache,
)
new_mem_cache = layer_outputs[-1]
layer_outputs = layer_outputs[:-1]
next_mem_caches += (new_mem_cache,)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
else:
next_decoder_cache += (None,)
if output_attentions:
all_self_attns += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
mem_cache_returned = False
for mem_cache in next_mem_caches:
if mem_cache is not None:
mem_cache_returned = True
next_mem_caches = next_mem_caches if mem_cache_returned else None
if not return_dict:
return tuple(
v
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, next_mem_caches]
if v is not None
)
return LongLlamaModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
mem_caches=next_mem_caches,
)
def _handle_output_of_past_key_values(outputs):
# merges local caches and memory caches into one single tuple of past_key_values
# in order to support generation
batch_size = outputs.last_hidden_state.shape[0]
if outputs.past_key_values is None and outputs.mem_caches is None:
return None
if outputs.past_key_values is None:
out_past_key_values = (None,) * len(outputs.mem_caches)
else:
out_past_key_values = outputs.past_key_values
if outputs.mem_caches is None:
out_mem_caches = (None,) * len(outputs.past_key_values)
else:
out_mem_caches = outputs.mem_caches
device = outputs.last_hidden_state.device
past_key_values = ()
for local_cache, mem_cache in zip(out_past_key_values, out_mem_caches):
layer = ()
if local_cache is not None:
assert len(local_cache) == 3
layer += local_cache
else:
layer += (torch.empty(batch_size, 0, 0, 0, device=device),) * 3
if mem_cache is not None:
layer += (mem_cache.keys, mem_cache.values, mem_cache.masks)
else:
layer += (torch.empty(batch_size, 0, 0, 0, device=device),) * 3
assert len(layer) == 6
past_key_values += (layer,)
return past_key_values
def _split_past_key_values(past_key_values):
# splits past_key_values to local cache and memory cache
local_cache_preset = False
mem_caches_present = False
if past_key_values is not None:
local_caches = ()
mem_caches = ()
for layer in past_key_values:
if len(layer) != 6:
raise ValueError(
"Expected elements of past_key_values to contain 6 elements."
"First 3 describing local cache and last 3 describing memory cache."
f"Instead got {len(layer)} elements"
)
else:
lk, lv, li, memk, memv, memm = layer
if lk.shape[-2] != 0:
local_cache_preset = True
local_caches += ((lk, lv, li),)
else:
local_caches += (None,)
if memk.shape[-2] != 0:
mem_caches_present = True
mem_caches += (LongLlamaMemCache(keys=memk, values=memv, masks=memm),)
else: