File size: 58,675 Bytes
73bcbf2 61f4597 73bcbf2 61f4597 73bcbf2 61f4597 73bcbf2 61f4597 73bcbf2 |
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 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 |
# pylint: disable=too-many-lines, too-many-arguments, too-many-positional-arguments, too-many-instance-attributes, too-many-locals
"""
This module implements the TPTT model with linear attention (LiZA) and LoRA support.
Author : Fabien FURFARO
TPTT : Transforming Pretrained Transformers into Titans (https://arxiv.org/abs/2506.17671)
"""
import logging
import math
import os
from pathlib import Path
import re
import shutil
from functools import partial
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
from einops import rearrange
from huggingface_hub import hf_hub_download, list_repo_files
from peft import LoraConfig, PeftModel, get_peft_model
from safetensors import safe_open
from safetensors.torch import save_file
from torch import nn
from torch.utils.checkpoint import checkpoint
from transformers import (
AutoConfig,
AutoModel,
AutoModelForCausalLM,
DynamicCache,
PreTrainedModel,
)
from transformers.configuration_utils import PretrainedConfig
from .configuration_tptt import TpttConfig
logger = logging.getLogger(__name__) # monitoring
class LCache:
"""Cache for storing intermediate states of linear attention layers."""
def __init__(self):
"""Stores per-layer intermediate states: {layer_idx: state_dict}"""
self.inputs_states: Dict[int, Dict[str, torch.Tensor]] = (
{}
) # recurrent states and qkv buffers
def __getitem__(self, layer_idx: int) -> Optional[Dict[str, torch.Tensor]]:
"""Retrieve cached state for a given layer, or None if not present"""
return self.inputs_states.get(layer_idx, None)
def update(self, layer_idx: int, **kwargs):
"""Detach all tensors to avoid retaining computation graphs"""
detached_kwargs = {
k: v.detach() if isinstance(v, torch.Tensor) else v
for k, v in kwargs.items()
}
# Update or create the state for the specified layer
if layer_idx in self.inputs_states:
self.inputs_states[layer_idx].update(detached_kwargs)
else:
self.inputs_states[layer_idx] = detached_kwargs
def reset(self):
"""Clear all cached states and reset the token counter"""
self.inputs_states.clear()
class CausalAvgPool1d(nn.Module):
"""Causal sliding window average (uniform, no shape loss along sequence)"""
def __init__(
self, output_size: int, offsets: tuple[int] = (0, 1, 2), mode: str = "replicate"
):
super().__init__()
self.offsets = offsets
self.mode = mode
self.pool = nn.AdaptiveAvgPool1d(output_size=output_size)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""x: [B, S, F] → [B, S, F → output_size]"""
x_ = x.transpose(1, 2) # [B, F, S]
idxs = torch.tensor(self.offsets, device=x.device)
ksize = idxs.max() - idxs.min() + 1
w = torch.zeros(ksize, device=x.device, dtype=x.dtype)
w[idxs - idxs.min()] = 1 / len(self.offsets) # Always uniform weights
kernel = w.repeat(x_.shape[1], 1).reshape(x_.shape[1], 1, ksize)
pad_left = -idxs.min().item()
pad_right = (ksize - 1) - pad_left
x_pad = F.pad(x_, (pad_left, pad_right), mode=self.mode)
y = F.conv1d(x_pad, kernel, groups=x_.shape[1]) # pylint: disable=not-callable
return self.pool(y.transpose(1, 2)) # [B, S, F → output_size]
class LinearAttention(nn.Module):
"""
Linear multi-head attention layer: [B, S, D] -> [B, S, D]
Projections + gating + efficient linear attention mechanism (TPTT compatible).
"""
def __init__(
self,
hidden_dim: int,
num_heads: int,
head_dim: Optional[int] = None,
num_key_value_heads: Optional[int] = None,
num_key_value_groups: Optional[int] = None,
bias: bool = True,
dropout: Optional[float] = None,
linear_precision: torch.dtype = torch.float32,
padding_side: str = "right",
shared_attn: bool = False, # shared attention
layer_idx: int = 0,
operator_mode: str = "delta_rule",
use_linear_checkpoint: bool = False,
recurrent_config: Optional[Dict[str, Any]] = None,
linear_cache: Optional[LCache] = None,
max_chunk_size: int = 64,
bidirectional: bool = False, # not used if causal
pooling_config: Optional[Dict[str, Any]] = None,
):
super().__init__()
if pooling_config is None:
pooling_config = {
"offsets": (0, 1, 2),
"mode": "replicate",
}
self.hidden_dim = hidden_dim
self.num_heads = num_heads
self.head_dim = head_dim or hidden_dim // num_heads
self.num_key_value_heads = num_key_value_heads or num_heads
self.num_key_value_groups = num_key_value_groups or (
num_heads // (num_key_value_heads or num_heads)
)
self.scaling = self.head_dim**-0.5
self.linear_precision = linear_precision
self.padding_side = padding_side
self.shared_attn = shared_attn
if not shared_attn:
self.q_proj = nn.Linear(hidden_dim, num_heads * self.head_dim, bias=bias)
self.k_proj = nn.Linear(
hidden_dim, self.num_key_value_heads * self.head_dim, bias=bias
)
self.v_proj = nn.Linear(
hidden_dim, self.num_key_value_heads * self.head_dim, bias=bias
)
self.out_proj = nn.Linear(num_heads * self.head_dim, hidden_dim, bias=bias)
self.dropout = nn.Dropout(dropout) if dropout is not None else None
self.linear_operator = LinearAttentionOp(
layer_idx=layer_idx,
operator_mode=operator_mode,
use_linear_checkpoint=use_linear_checkpoint,
recurrent_config=recurrent_config,
max_chunk_size=max_chunk_size,
linear_cache=linear_cache,
linear_precision=linear_precision,
)
self.bidirectional = bidirectional
# Causal average pooling for gating
self.pooling_config = pooling_config
self.pool_g = CausalAvgPool1d(
output_size=self.head_dim * self.num_key_value_heads, **pooling_config
)
def forward(
self,
x: Union[List[torch.Tensor], torch.Tensor],
attn_mask: Optional[torch.Tensor] = None,
out_proj: Optional[nn.Module] = None,
**kwargs: Any,
) -> torch.Tensor:
"""
Forward pass for linear attention. Input shape: [B, S, D], output [B, S, D].
"""
if not self.shared_attn:
hidden_states = x[0] if isinstance(x, (list, tuple)) else x
# Projections
q = self.q_proj(hidden_states)
k = self.k_proj(hidden_states)
v = self.v_proj(hidden_states)
out_proj = self.out_proj
else:
# Shared attention <=> no projections here
q, k, v = x[0], x[1], x[2]
out_proj = self.out_proj if out_proj is None else out_proj
# get dtype and device
final_dtype, final_device = q.dtype, q.device
# Masking if needed
if attn_mask is not None:
v = apply_linear_attention_mask(attn_mask, v, self.padding_side)
# Forget and Write Gating for linear attn (abusive term)
f_g, w_g = self.pool_g(k), self.pool_g(v)
# Reshape for multi-head
q = rearrange(q, "b n (h d) -> b h n d", h=self.num_heads)
k = rearrange(k, "b n (h d) -> b h n d", h=self.num_key_value_heads)
v = rearrange(v, "b n (h d) -> b h n d", h=self.num_key_value_heads)
f_g = rearrange(f_g, "b n (h m) -> b h n m", h=self.num_key_value_heads)
w_g = rearrange(w_g, "b n (h m) -> b h n m", h=self.num_key_value_heads)
# Repeat for GQA
k = k.repeat_interleave(self.num_key_value_groups, dim=1)
v = v.repeat_interleave(self.num_key_value_groups, dim=1)
f_g = f_g.repeat_interleave(self.num_key_value_groups, dim=1)
w_g = w_g.repeat_interleave(self.num_key_value_groups, dim=1)
## DeltaNet-style: Silu activation and normalization
q = F.normalize(F.silu(q), p=2, dim=-1, eps=1e-6)
k = F.normalize(F.silu(k), p=2, dim=-1, eps=1e-6)
## linear stability part
v = ensure_stability(v * self.scaling, min_val=-1e4, max_val=1e4)
# Apply sigmoid to forget and write gates
f_g = torch.clamp(torch.sigmoid(f_g), min=1e-6, max=1 - 1e-6)
w_g = torch.clamp(torch.sigmoid(w_g), min=1e-6, max=1 - 1e-6)
# Convert to linear_precision (float32) for numerical stability and get model dtype
q, k, v, f_g, w_g = (
x.to(self.linear_precision).contiguous() for x in (q, k, v, f_g, w_g)
)
g = (f_g, w_g)
# Linear Attention Core, output: [B, H, S, d]
if self.bidirectional: # Work only with uncausal attention
# Forward direction
out_forward = self.linear_operator(q, k, v, g, **kwargs)
# Backward direction: flip the input sequence on the time dimension (dim=2)
kwargs_bwd = kwargs.copy()
kwargs_bwd["use_cache"] = False
out_backward = self.linear_operator(
torch.flip(q, dims=[2]),
torch.flip(k, dims=[2]),
torch.flip(v, dims=[2]),
tuple(torch.flip(t, dims=[2]) for t in g),
**kwargs_bwd,
)
# Flip the output back to restore proper order
out_backward = torch.flip(out_backward, dims=[2])
# Fusion: here, simple addition
out = out_forward + out_backward
else:
out = self.linear_operator(q, k, v, g, **kwargs)
# Merge heads and project: [B, H, S, d] -> [B, S, H*d] -> Out proj
out = rearrange(out, "b h s d -> b s (h d)")
# Normalize output (RMS norm). Note: bidirectional compatibility
out = out / out.pow(2).mean(dim=-1, keepdim=True).add(1e-6).sqrt()
# Ensure dtype and device consistency
out = out.to(dtype=final_dtype, device=final_device)
# Apply output projection
out = out_proj(out) # [B, S, D]
out = ensure_stability(out, min_val=-1e4, max_val=1e4)
# Apply dropout if specified
if self.dropout is not None:
out = self.dropout(out)
return out
class LiZAttention(nn.Module):
"""LiZA Linear Attention module, mixing linear and vanilla attention."""
def __init__(
self,
base_attn: nn.Module,
layer_idx: int,
base_config: PretrainedConfig, # Backbone Config
linear_cache: Optional[LCache] = None,
operator_mode: str = "delta_rule",
use_linear_checkpoint: bool = False,
recurrent_config: Optional[Dict[str, Any]] = None,
max_self_attn_length: Optional[int] = None, # unnecessary
base_scale_attn: bool = False,
mag_weight: float = 0.5,
cross_gate: bool = False,
max_chunk_size: int = 64,
linear_precision: Union[str, torch.dtype] = "float32",
padding_side: str = "right", # for tokenizer
disable_linear_attn: bool = False,
bidirectional: bool = False, # if True, use bidirectional attention
pooling_config: Optional[Dict[str, Any]] = None,
):
super().__init__()
if isinstance(linear_precision, str):
linear_precision = getattr(torch, linear_precision)
self.linear_precision = linear_precision
self.base_attn: nn.Module = base_attn
self.base_config = base_config
self.layer_idx = layer_idx
self.max_self_attn_length = max_self_attn_length
self.base_scale_attn = base_scale_attn
self.mag_weight = mag_weight
self.cross_gate = cross_gate
self.max_chunk_size = max_chunk_size
self.linear_precision = linear_precision
self.padding_side = padding_side
self.disable_linear_attn = disable_linear_attn
(
self.num_heads,
self.head_dim,
self.num_key_value_heads,
self.num_key_value_groups,
self.hidden_dim,
) = self._get_attention_parameters(base_attn, base_config)
self.scaling = self.head_dim**-0.5
self.linear_attn = LinearAttention(
layer_idx=layer_idx,
shared_attn=True,
operator_mode=operator_mode,
use_linear_checkpoint=use_linear_checkpoint,
recurrent_config=recurrent_config,
hidden_dim=self.hidden_dim,
num_heads=self.num_heads,
head_dim=self.head_dim,
num_key_value_heads=self.num_key_value_heads,
num_key_value_groups=self.num_key_value_groups,
linear_precision=linear_precision,
linear_cache=linear_cache,
max_chunk_size=max_chunk_size,
padding_side=padding_side,
bidirectional=bidirectional,
pooling_config=pooling_config,
)
def _get_attention_parameters(
self, base_attn: nn.Module, base_config: PretrainedConfig
) -> Tuple[Optional[int], Optional[int], Optional[int], Optional[int]]:
"""Retrieve the attention parameters from the base attention module."""
# first order base attention module and second order config
num_heads = (
getattr(base_attn, "num_heads", None)
or getattr(base_attn, "num_q_heads", None)
or getattr(base_config, "num_heads", None)
or getattr(base_config, "num_attention_heads", None)
)
head_dim = (
getattr(base_attn, "head_dim", None)
or getattr(base_attn, "attention_head_size", None)
or getattr(base_config, "head_dim", None)
or (
getattr(base_config, "hidden_size", None) // num_heads
if num_heads and getattr(base_config, "hidden_size", None)
else None
)
)
num_key_value_heads = (
getattr(base_attn, "num_kv_heads", None)
or getattr(base_attn, "num_k_heads", None)
or getattr(base_config, "num_key_value_heads", None)
or num_heads # fallback
)
num_key_value_groups = getattr(base_attn, "num_key_value_groups", None) or (
num_heads // num_key_value_heads if num_heads and num_key_value_heads else 1
)
hidden_dim = getattr(base_config, "hidden_size", None) or head_dim * num_heads
return (
num_heads,
head_dim,
num_key_value_heads,
num_key_value_groups,
hidden_dim,
)
def _apply_shared_projections(
self, hidden_states: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, nn.Module]:
base_attn = self.base_attn
if hasattr(base_attn, "q_proj"):
# LLama, OLMO and Mistral style
q = base_attn.q_proj(hidden_states)
k = base_attn.k_proj(hidden_states)
v = base_attn.v_proj(hidden_states)
out_proj = base_attn.o_proj
elif hasattr(base_attn, "qkv_proj"):
# OpenELM and GPT-Neo style : QKV fused, split on the last dimension
qkv = base_attn.qkv_proj(hidden_states)
q, k, v = split_qkv(base_attn, qkv)
out_proj = base_attn.out_proj
elif hasattr(base_attn, "c_attn") and hasattr(base_attn, "c_proj"):
# GPT-2 style
qkv = base_attn.c_attn(hidden_states)
q, k, v = qkv.chunk(3, dim=-1)
out_proj = base_attn.c_proj
elif all(hasattr(base_attn, n) for n in ["query", "key", "value"]):
# BERT - ViT
q = base_attn.query(hidden_states)
k = base_attn.key(hidden_states)
v = base_attn.value(hidden_states)
out_proj = getattr(base_attn, "dense", None) # ou output.dense
else:
raise ValueError("Unsupported attention module: cannot find projections.")
# Ensure stability
q = ensure_stability(q, min_val=-1e4, max_val=1e4)
k = ensure_stability(k, min_val=-1e4, max_val=1e4)
v = ensure_stability(v, min_val=-1e4, max_val=1e4)
return q, k, v, out_proj
def _process_self_attn(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor],
kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[DynamicCache], int]:
"""Process the self-attention part (with truncation)."""
if self.max_self_attn_length: # Not needed for SWA (nonparam memorize context)
hidden_states, attention_mask = truncate_attention_mask(
hidden_states, attention_mask, self.max_self_attn_length
)
if kwargs.get("position_embeddings", None) is not None:
cos, sin = kwargs["position_embeddings"]
cos = cos[:, -self.max_self_attn_length :]
sin = sin[:, -self.max_self_attn_length :]
kwargs["position_embeddings"] = (cos, sin)
if isinstance(kwargs.get("past_key_value", None), DynamicCache):
# cache management
if (
len(kwargs["past_key_value"]) > self.layer_idx
and self.layer_idx == 0
):
kwargs["past_key_value"].crop(self.max_self_attn_length - 1)
# Ensure attention mask is of the correct dtype and device
if attention_mask is not None:
attention_mask = attention_mask.to(
dtype=hidden_states.dtype, device=hidden_states.device
)
# Standard attention (mask and rotation is applied inside)
base_attn_outputs = self.base_attn(
hidden_states,
attention_mask=attention_mask,
**kwargs,
)
if isinstance(base_attn_outputs, tuple):
if len(base_attn_outputs) == 3:
o_base, attn_weights, present_key_value = base_attn_outputs
expected_attn_mode = 3
elif len(base_attn_outputs) == 2:
o_base, attn_weights = base_attn_outputs
present_key_value, expected_attn_mode = None, 2
else:
raise ValueError(
f"Unexpected number of outputs from base_attn: {len(base_attn_outputs)}"
)
else:
o_base = base_attn_outputs
attn_weights, present_key_value, expected_attn_mode = None, None, 1
# Ensure stability
o_base = ensure_stability(o_base, min_val=-1e4, max_val=1e4)
return o_base, attn_weights, present_key_value, expected_attn_mode
def _prepare_attn_mixin(
self,
o_lin: torch.Tensor,
o_base: torch.Tensor,
tensor_dtype: torch.dtype,
eps: float = 1e-5,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Prepare linear attn for mixing with self attn."""
# Force cast typing, shape : [b n (h d)]
o_lin = o_lin.to(tensor_dtype)
o_base = o_base.to(tensor_dtype)
# feature scaling
if self.base_scale_attn:
scaler = o_base.pow(2).mean(dim=-1, keepdim=True).add(eps).sqrt()
o_lin = scaler * o_lin
return o_lin, o_base
def _apply_mag(
self, linear_attention: torch.Tensor, softmax_attention: torch.Tensor
) -> torch.Tensor:
"""Apply the MAG strategy"""
# Left-Padding management
if linear_attention.shape[1] != softmax_attention.shape[1]:
left_trunc = min(linear_attention.shape[1], softmax_attention.shape[1])
linear_attention, softmax_attention = (
linear_attention[:, -left_trunc:],
softmax_attention[:, -left_trunc:],
)
# NAM : Neural Attention Mixer (with graph forcing)
mag_weight = torch.tensor(
self.mag_weight,
dtype=softmax_attention.dtype,
device=softmax_attention.device,
)
softmax_weighted = (1 - mag_weight) * softmax_attention
linear_weighted = mag_weight * linear_attention
if self.cross_gate:
output_attention = (
softmax_weighted + linear_weighted + softmax_weighted * linear_weighted
) # complex cross product (unlinear interaction)
else:
output_attention = softmax_weighted + linear_weighted # classic
if torch.allclose(softmax_weighted, output_attention):
logger.info(
"[LOG] layer : %s, softmax_weighted and output_attention are close.",
self.layer_idx,
)
# Final output
return ensure_stability(output_attention, min_val=-1e4, max_val=1e4)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
**kwargs,
) -> torch.Tensor:
"""Mix linear and self attention forward"""
device = hidden_states.device
tensor_dtype = hidden_states.dtype
self.base_attn.to(device)
if self.training:
kwargs.pop("past_key_value", None)
kwargs["use_cache"] = False
elif "use_cache" not in kwargs:
kwargs.pop("past_key_value", None)
kwargs["use_cache"] = False
kwargs.pop("position_ids", None) # obsolete
# Apply shared projections
q, k, v, out_proj = self._apply_shared_projections(hidden_states)
# Apply linear attention to hidden states
o_lin = self.linear_attn(
x=[q, k, v], attn_mask=attention_mask, out_proj=out_proj, **kwargs
)
# Process self attn with truncation
o_base, attn_weights, present_key_value, expected_attn_mode = (
self._process_self_attn(hidden_states, attention_mask, kwargs)
)
# Prepare output mixing
o_lin, o_base = self._prepare_attn_mixin(o_lin, o_base, tensor_dtype, eps=1e-5)
# Apply Memory as Gate in self-attention (with length management and ablation)
out = o_base if self.disable_linear_attn else self._apply_mag(o_lin, o_base)
# Return output following transformer convention
if expected_attn_mode == 3:
return out, attn_weights, present_key_value
if expected_attn_mode == 2:
return out, attn_weights
return out
def load_tptt_safetensors(
repo_or_path: str,
model: Union[PreTrainedModel, PeftModel],
subfolder: Optional[str] = None,
token: Optional[str] = None,
) -> Union[PreTrainedModel, PeftModel]:
"""Load Tptt safetensor from LoRA/PEFT weights and adapt keys if needed."""
# sharding not supported yet (e.g. : -00001-of-00005.safetensors, ...)
fname = "adapter_model.safetensors"
# subfolder management
if subfolder:
repo_or_path_norm = os.path.normpath(repo_or_path)
subfolder_norm = os.path.normpath(subfolder)
if not repo_or_path_norm.endswith(subfolder_norm):
fname = f"{subfolder}/{fname}" if subfolder else fname
# Find file path
if os.path.isdir(repo_or_path):
path = os.path.join(repo_or_path, fname)
if not os.path.exists(path):
return model
else:
if fname not in list_repo_files(repo_or_path, token=token):
return model
path = hf_hub_download(repo_or_path, fname, token=token)
# Load weights from safetensors
with safe_open(path, framework="pt") as f:
state_dict = {k: f.get_tensor(k) for k in f.keys()}
# Adapt LoRA/Specific keys if needed (add .default if expected by the model)
def adapt_keys(sd, model):
model_keys = list(model.state_dict().keys())
if any(k.startswith("tptt_model.base_model.") for k in model_keys):
prefix = "tptt_model.base_model."
elif any(k.startswith("base_model.") for k in model_keys):
prefix = "base_model."
else:
prefix = ""
has_base_attn = any(".base_attn." in k for k in model_keys)
def adapt_key(k):
k_ = k if k.startswith(prefix) else prefix + k
# first, verify and modify base_attn (LiZA)
if ".base_attn." in k_ and not has_base_attn:
k_ = k_.replace(".base_attn.", ".")
# change LoRA if needed
if (
k_.endswith("lora_A.weight") or k_.endswith("lora_B.weight")
) and k_.replace(".weight", ".default.weight") in model_keys:
k_ = k_.replace(".weight", ".default.weight")
return k_
return {adapt_key(k): v for k, v in sd.items()}
state_dict = adapt_keys(state_dict, model)
# Cast tensors to the expected dtype of the model parameters
model_state_dict = model.state_dict()
for k, v in state_dict.items():
if k in model_state_dict:
expected_dtype = model_state_dict[k].dtype
if v.dtype != expected_dtype:
state_dict[k] = v.to(expected_dtype)
logger.info("Input LoRA/Specific keys: %s", [k for k in state_dict.keys()])
# Load into model
missing, unexpected = model.load_state_dict(state_dict, strict=False, assign=True)
missing_lora = [k for k in missing if "lora" in k]
if missing_lora:
logger.warning("Missing keys: %s", missing_lora)
if unexpected:
logger.warning("Unexpected keys: %s", unexpected)
return model
def get_tptt_model( # pylint: disable=too-many-arguments, too-many-positional-arguments
model: nn.Module,
base_config: PretrainedConfig, # ou LlamaConfig, MistralConfig, etc.
linear_cache: Optional[LCache] = None,
liza_attention: nn.Module = LiZAttention,
target_modules_names: Optional[list[str]] = None,
operator_mode: str = "delta_rule",
use_linear_checkpoint: bool = False,
recurrent_config: Optional[Dict[str, Any]] = None,
base_scale_attn: bool = False,
mag_weight: float = 0.5,
cross_gate: bool = False,
max_chunk_size: int = 64,
linear_precision: torch.dtype = torch.float32,
max_self_attn_length: Optional[int] = None, # unnecessary
padding_side: str = "right", # for tokenizer
bidirectional: bool = False, # if True, use bidirectional attention
pooling_config: Optional[Dict[str, Any]] = None,
**kwargs, # quickfix unexpected arguments
) -> Tuple[PreTrainedModel, LCache]:
"""Replace target modules in a model with LiZAttention."""
if target_modules_names is None:
target_modules_names = ["attn", "self_attn", "attention"]
# Find target modules by suffix (e.g., "attn", "attention")
target_modules_names = [
name
for name, _ in model.named_modules()
if any(name.endswith(suffix) for suffix in target_modules_names)
and not any(f".{suffix}." in name for suffix in target_modules_names)
]
if not target_modules_names:
raise ValueError(
f"Target modules '{target_modules_names}' not found in the model."
)
# Prepare recurrent config
linear_cache = linear_cache or LCache()
# Inject LiZAttention into the model
for name, _ in model.named_modules():
if name in target_modules_names:
parent = model
*path, last = name.split(".")
for p in path:
parent = getattr(parent, p)
layer_idx = extract_layer_idx(name)
setattr(
parent,
last,
liza_attention(
getattr(parent, last),
layer_idx=layer_idx,
base_config=base_config,
linear_cache=linear_cache,
operator_mode=operator_mode,
use_linear_checkpoint=use_linear_checkpoint,
recurrent_config=recurrent_config,
max_self_attn_length=max_self_attn_length,
base_scale_attn=base_scale_attn,
mag_weight=mag_weight,
cross_gate=cross_gate,
max_chunk_size=max_chunk_size,
linear_precision=linear_precision,
padding_side=padding_side,
bidirectional=bidirectional,
pooling_config=pooling_config,
),
)
return model, linear_cache
def save_tptt_safetensors(model, path: str, name: str = "adapter_model.safetensors"):
"""Save trainable LoRA/Specific weights and adapting key names"""
# 1. Get the full state_dict
all_sd = model.state_dict()
# 2. Identify trainable parameter names (usually only LoRA/PEFT adapters)
trainable_keys = [
name for name, param in model.named_parameters() if param.requires_grad
] # Also, you can manually select specific keys in model after load
# 3. Filter and adapt the keys (Remove custom model encapsulation info)
to_save = {
k.replace("tptt_model.", "").replace("base_model.", ""): all_sd[k]
for k in trainable_keys
}
# 4. Save the filtered adapters to a safetensors file
if to_save:
os.makedirs(os.path.dirname(path), exist_ok=True)
# sharding not supported yet (e.g. : -00001-of-00005.safetensors, ...)
save_file(to_save, os.path.join(path, name))
class TpttModel(PreTrainedModel):
"""
TPTT model wrapper with linear attention (LiZA) and LoRA support.
Handles only architecture and weights.
"""
config_class = TpttConfig
def __init__(
self,
config: TpttConfig,
**kwargs,
):
"""
Initialize TpttModel with a given config and backbone.
Injects LiZA attention modules into the backbone.
"""
super().__init__(config, **kwargs)
repo_or_path = getattr(config, "_base_path", None) or config._name_or_path
# 1. Load backbone (with subfolder management) :
kwargs_bb = kwargs.copy()
if config.base_model_subfolder is not None:
kwargs_bb["subfolder"] = config.base_model_subfolder
else:
kwargs_bb.pop("subfolder", None)
if config.model_task == "causal_lm":
tptt_model = AutoModelForCausalLM.from_pretrained(
config.base_model_name, **kwargs_bb
)
else:
tptt_model = AutoModel.from_pretrained(config.base_model_name, **kwargs_bb)
# 2. Inject LiZA attention
self.linear_cache = LCache()
tptt_model, self.linear_cache = get_tptt_model(
tptt_model, config, self.linear_cache, **config.to_dict()
)
# 3. Apply LoRA/Specific if present and configured
if config.lora_config is not None:
lora_config_obj = LoraConfig(**config.lora_config)
tptt_model = get_peft_model(tptt_model, lora_config_obj)
else:
# Doesn't work if quantization is applied !
tptt_model = set_trainable_parameters(tptt_model)
# 4. Load safetensor if tptt/peft adaptor in repo
if repo_or_path:
tptt_model = load_tptt_safetensors(
repo_or_path,
tptt_model,
subfolder=kwargs.get("subfolder", None),
token=kwargs.get("token", None),
)
self.tptt_model = tptt_model
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
labels: Optional[torch.LongTensor] = None,
**kwargs,
):
"""Forward pass. All arguments are passed to the underlying base model."""
if self.training:
kwargs["use_cache"] = False
kwargs.pop("num_items_in_batch", None)
elif "use_cache" not in kwargs: # evaluation
kwargs.pop("num_items_in_batch", None)
kwargs["use_cache"] = False
return self.tptt_model(
input_ids=input_ids, attention_mask=attention_mask, labels=labels, **kwargs
)
def generate(self, *args, **kwargs):
"""Delegate the generate call to the backbone model, which supports generation"""
return self.tptt_model.generate(*args, **kwargs)
def save_pretrained(self, path: str, **kwargs):
"""Save model weights, config, and source code to the given path."""
# 0. Save complete tptt config (with or without LoRA)
super().save_pretrained(path, **kwargs) # pylint: disable=no-member
self._adjust_save_strategy(path, **kwargs)
# 1. Save true weights and adapte keys
save_tptt_safetensors(self, path)
# 2. Copy Python files for trust_remote_code
self._copy_source_files(path, **kwargs)
def _adjust_save_strategy(self, path: str, **kwargs):
"""Re-adapt/remove the weight safetensor and saved adapter config"""
if isinstance(self.tptt_model, PeftModel):
self.tptt_model.save_pretrained(path, **kwargs)
safetensor_path = os.path.join(path, "model.safetensors")
if os.path.exists(safetensor_path):
os.remove(safetensor_path)
adapter_path = os.path.join(path, "adapter_config.json")
if os.path.exists(adapter_path):
os.remove(adapter_path)
def _copy_source_files(self, target_path: str, **kwargs):
"""Copy all .py files from package directory for trust_remote_code."""
src_dir = os.path.dirname(os.path.abspath(__file__))
dst_dir = (
f"./{str(Path(target_path).parts[0])}"
if kwargs.get("subfolder", False)
else target_path
)
for fname in os.listdir(src_dir):
if fname.endswith(".py"):
src = os.path.join(src_dir, fname)
dst = os.path.join(dst_dir, fname)
shutil.copy2(src, dst)
def retie_lm_after_load(self, **kwargs):
"""Re-link lm_head after loading external weights."""
embed_lm = find_embedding_lm(self.tptt_model)
if embed_lm is not None and hasattr(self.tptt_model, "lm_head"):
if self.tptt_model.lm_head is None: # ensure lm_head exists
self.tptt_model.lm_head = nn.Linear(
embed_lm.weight.shape[1], embed_lm.weight.shape[0], bias=False
)
if kwargs.get("tie_word_embeddings", True):
self.tptt_model.lm_head.weight = embed_lm.weight # share weights
logger.info("Weights of lm_head have been shared with embedding.")
else:
self.tptt_model.lm_head.weight = nn.Parameter(embed_lm.weight.clone())
logger.info("Weights of lm_head have been cloned from the embedding.")
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path=None, *model_args, **kwargs):
"""Custom from_pretrained that accepts the standard positional argument"""
config = kwargs.pop("config", None)
repo_or_path = (
pretrained_model_name_or_path
or kwargs.pop("pretrained_model_name_or_path", None)
or kwargs.pop("repo_or_path", None)
or (getattr(config, "_base_path", None) if config else None)
or (getattr(config, "_name_or_path", None) if config else None)
)
if config is None and repo_or_path is not None:
config = AutoConfig.from_pretrained(repo_or_path, **kwargs)
model = cls(config, *model_args, **kwargs)
model.retie_lm_after_load(**kwargs)
return model
TpttModel.register_for_auto_class("AutoModelForCausalLM")
class LinearAttentionOp(nn.Module):
"""Base class for linear attention operators."""
def __init__(
self,
layer_idx: int,
operator_mode: str = "delta_rule",
use_linear_checkpoint: bool = False,
recurrent_config: Optional[dict] = None,
max_chunk_size: int = 64,
linear_cache: Optional[LCache] = None,
linear_precision: torch.dtype = torch.float32,
):
super().__init__()
self.layer_idx = layer_idx
if recurrent_config is None:
operator_mode = "delta_rule" # force default operator mode if no config
recurrent_config = {
"order": 1,
"gate_type": "k",
"linear": True,
"trick": "derivative",
}
self.operator_mode = operator_mode
self.use_linear_checkpoint = use_linear_checkpoint
self.order = recurrent_config["order"]
self.gate_type = recurrent_config["gate_type"]
self.linear = recurrent_config["linear"]
self.trick = recurrent_config["trick"]
self.max_chunk_size = max_chunk_size
self.linear_cache = linear_cache or LCache()
self.linear_precision = linear_precision
def compute_gate(self, beta: Tuple[torch.Tensor]) -> torch.Tensor:
"""
Compute the gating tensor according to the gate_type.
"""
if self.gate_type == "k":
return torch.clamp(beta[0], min=1e-6, max=1 - 1e-6)
if self.gate_type == "v":
return torch.clamp(beta[1], min=1e-6, max=1 - 1e-6)
if self.gate_type == "kv":
return torch.clamp(beta[0] * beta[1], min=1e-6, max=1 - 1e-6)
raise ValueError(f"Unsupported gate_type: {self.gate_type}")
def get_cache(self, use_cache: bool) -> Tuple[
Optional[torch.Tensor],
Optional[Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]],
]:
"""
Retrieve recurrent state and qkv buffers from the cache.
"""
if not use_cache:
return None, None
last_state = self.linear_cache[self.layer_idx]
if last_state is not None:
recurrent_state = last_state.get("recurrent_state", None)
qkv_buffers = last_state.get("qkv", None)
else:
recurrent_state = None
qkv_buffers = None
return recurrent_state, qkv_buffers
def save_cache(
self,
use_cache: bool,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
gate: torch.Tensor,
state: torch.Tensor,
) -> None:
"""
Save the recurrent state and qkv buffers to the cache.
"""
if not use_cache:
return
if self.order > 1:
qkv_buffers = (
q[:, :, -(self.order - 1) :, :],
k[:, :, -(self.order - 1) :, :],
v[:, :, -(self.order - 1) :, :],
gate[:, :, -(self.order - 1) :, :],
)
else:
qkv_buffers = None
self.linear_cache.update(self.layer_idx, recurrent_state=state, qkv=qkv_buffers)
def forward(
self,
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
beta: Union[Tuple[torch.Tensor], torch.Tensor],
**kwargs,
) -> torch.Tensor:
"""
Forward pass for the attention operator.
"""
# Ensure linear_precision for numerical stability (float32)
q, k, v = [x.to(self.linear_precision) for x in (q, k, v)]
if isinstance(beta, (tuple, list)):
beta = tuple(b.to(self.linear_precision) for b in beta)
else:
beta = beta.to(self.linear_precision)
gate = self.compute_gate(beta)
# Retrieve cache if needed
use_cache = kwargs.get("use_cache", False)
use_checkpoint = not (use_cache) and self.use_linear_checkpoint
recurrent_state, qkvb = self.get_cache(use_cache)
if qkvb is not None and qkvb[0].shape == q.shape:
q = torch.cat([qkvb[0].to(q.device), q], dim=2).to(self.linear_precision)
k = torch.cat([qkvb[1].to(q.device), k], dim=2).to(self.linear_precision)
v = torch.cat([qkvb[2].to(q.device), v], dim=2).to(self.linear_precision)
gate = torch.cat([qkvb[3].to(q.device), gate], dim=2).to(
self.linear_precision
)
output, state = self.chunk_delta_product_forward(
q,
k,
v,
gate,
self.max_chunk_size,
n=self.order,
trick=self.trick,
linear=self.linear,
initial_state=recurrent_state,
use_checkpoint=use_checkpoint,
linear_precision=self.linear_precision,
)
# Save cache if needed
self.save_cache(use_cache, q, k, v, gate, state)
return output
@staticmethod
def chunk_delta_product_forward(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
beta_gate: torch.Tensor,
chunk_size: int,
n: int = 1,
trick: str = "derivative",
linear: bool = True,
initial_state: Optional[torch.Tensor] = None,
use_checkpoint: bool = True,
linear_precision: torch.dtype = torch.float32,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Chunkwise parallel implementation https://arxiv.org/abs/2406.06484
For each chunk, processes chunk_size * n_orders steps (virtual tokens) in order.
"""
# --- Main chunk_delta_product_forward logic ---
batch_size, num_heads, seq_len, head_dim = query.shape
chunk_size = get_valid_chunk_size(seq_len, chunk_size)
num_chunks = seq_len // chunk_size
query_n = query if n == 1 else expand_virtual_tokens(query, n, trick)
key_n = key if n == 1 else expand_virtual_tokens(key, n, trick)
value_n = value if n == 1 else expand_virtual_tokens(value, n, trick)
beta_n = beta_gate if n == 1 else expand_virtual_tokens(beta_gate, n, trick)
q_chunks = chunk_sequence(query_n, num_chunks, chunk_size * n)
k_chunks = chunk_sequence(key_n, num_chunks, chunk_size * n)
v_chunks = chunk_sequence(value_n, num_chunks, chunk_size * n)
beta_chunks = chunk_sequence(beta_n, num_chunks, chunk_size * n)
k_beta = k_chunks * beta_chunks
v_beta = v_chunks * beta_chunks
householder = -(k_beta @ k_chunks.transpose(-2, -1)).tril(-1)
householder = ensure_stability(householder, min_val=-1e4, max_val=1e4)
# size : N = chunk_size * n
inv_hh = fast_invert_matrix(householder, dtype=linear_precision) # [(...),N,N]
w = ensure_stability(torch.matmul(inv_hh, k_beta), min_val=-1e4, max_val=1e4)
u = ensure_stability(torch.matmul(inv_hh, v_beta), min_val=-1e4, max_val=1e4)
state_shape = (batch_size, num_heads, n, head_dim, head_dim)
if initial_state is not None and initial_state.shape == state_shape:
state = initial_state.to(device=query.device, dtype=linear_precision)
else:
state = torch.full(
state_shape,
fill_value=1e-6, # stability if unlinear activation
device=query.device,
dtype=linear_precision,
)
output, final_state = sequential_delta_product_scan(
q_chunks.to(dtype=linear_precision),
w.to(dtype=linear_precision),
u.to(dtype=linear_precision),
n,
linear,
chunk_size,
state.to(dtype=linear_precision),
linear_precision=linear_precision,
use_checkpoint=use_checkpoint,
)
idx_last_order = torch.arange(chunk_size, device=output.device) * n + (n - 1)
output = output[:, :, :, idx_last_order, :] # [B, H, num_chunks, chunk_size, D]
output = output.reshape(batch_size, num_heads, seq_len, head_dim)
return output.to(dtype=linear_precision), final_state.to(dtype=linear_precision)
def sequential_delta_product_scan(
q_chunks: torch.Tensor,
w: torch.Tensor,
u: torch.Tensor,
n_orders: int,
linear_activation: bool,
current_chunk_size: int,
initial_recurrent_state: torch.Tensor,
linear_precision: torch.dtype,
use_checkpoint: bool,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
DeltaProduct implementation https://arxiv.org/abs/2502.10297
Implements the per-token Householder state updates.
"""
batch, head, num_chunks_inner, chunk_n_total, dim = q_chunks.shape
output_inner = torch.empty_like(q_chunks)
# initial_recurrent_state is H_{last_token_of_prev_chunk, n-1} ([B, H, D, D])
h_0_base = initial_recurrent_state[:, :, -1, :, :].clone()
def process_one_chunk(
q_chunk_params: torch.Tensor,
w_chunk_params: torch.Tensor,
u_chunk_params: torch.Tensor,
h_0_base: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Process a single chunk (with per-token state for n_orders > 1).
"""
o_intra_current_chunk = torch.zeros(
batch,
head,
chunk_n_total,
dim,
device=q_chunk_params.device,
dtype=linear_precision,
)
o_inter_current_chunk = torch.zeros_like(o_intra_current_chunk)
current_accumulated_state_per_token = (
h_0_base.unsqueeze(2).expand(-1, -1, current_chunk_size, -1, -1).clone()
) # [B, H, current_chunk_size, D, D]
for step in range(n_orders):
idx_virtual_tokens = (
torch.arange(current_chunk_size, device=q_chunk_params.device)
* n_orders
+ step
)
q_s = q_chunk_params[:, :, idx_virtual_tokens, :]
w_s = w_chunk_params[:, :, idx_virtual_tokens, :]
u_s = u_chunk_params[:, :, idx_virtual_tokens, :]
state_input_for_this_step = current_accumulated_state_per_token
## BLAS/cuBLAS einsum "bhcd,bhcdd->bhcd"
k_trans_h_old = (
torch.matmul(
w_s.unsqueeze(-2),
state_input_for_this_step,
)
.squeeze(-2)
.to(dtype=linear_precision)
)
u_val = u_s - k_trans_h_old
o_inter_current_chunk[:, :, idx_virtual_tokens, :] = (
torch.matmul(q_s.unsqueeze(-2), state_input_for_this_step)
.squeeze(-2)
.to(dtype=linear_precision)
)
## BLAS/cuBLAS einsum "bhcd,bhcd->bhcd"
o_intra_current_chunk[:, :, idx_virtual_tokens, :] = (q_s * u_val).to(
dtype=linear_precision
)
outer_product_term = torch.matmul(w_s.unsqueeze(-1), u_val.unsqueeze(-2))
new_state_i_per_token = state_input_for_this_step + outer_product_term
current_accumulated_state_per_token = new_state_i_per_token.to(
dtype=linear_precision
)
# Return all needed for next chunk
return (
o_intra_current_chunk,
o_inter_current_chunk,
current_accumulated_state_per_token[:, :, -1, :, :], # new h_0_base
)
for chunk_idx_inner in range(num_chunks_inner):
q_chunk_params = q_chunks[:, :, chunk_idx_inner]
w_chunk_params = w[:, :, chunk_idx_inner]
u_chunk_params = u[:, :, chunk_idx_inner]
# Checkpointed call if training
call = (
partial(checkpoint, use_reentrant=False)
if use_checkpoint
else lambda f, *a: f(*a)
)
o_intra, o_inter, h_0_base = call(
process_one_chunk,
q_chunk_params,
w_chunk_params,
u_chunk_params,
h_0_base,
)
if not linear_activation: # unlinear activation between chunks
h_0_base = unlinear_activation(h_0_base).to(dtype=linear_precision)
output_inner[:, :, chunk_idx_inner] = o_intra + o_inter
return output_inner, h_0_base
def unlinear_activation(x: torch.Tensor, scale: float = 2.0) -> torch.Tensor:
"""Unlinear activation between chunk"""
x_n = x.norm(p=2, dim=-1, keepdim=True) + 1e-6
x_gelu = F.gelu(scale * x / x_n, approximate="tanh") # pylint: disable=not-callable
return (x / scale) * x_gelu
def chunk_sequence(x: torch.Tensor, num_chunks: int, chunk_size: int) -> torch.Tensor:
"""Splits [B, H, S, D] to [B, H, num_chunks, chunk_size, D]"""
batch_size, num_heads, _, head_dim = x.shape
return x.reshape(batch_size, num_heads, num_chunks, chunk_size, head_dim)
def expand_virtual_tokens(
x: torch.Tensor, n: int, mode: str = "derivative"
) -> torch.Tensor:
"""Expand tokens into 'n' virtual tokens using the selected trick."""
batch_size, num_heads, seq_len, head_dim = x.shape
device, dtype = x.device, x.dtype
def derivative_expand(x: torch.Tensor) -> torch.Tensor:
"""Expand tokens using the derivative trick."""
x_pad = torch.cat(
[
torch.zeros(
batch_size, num_heads, n - 1, head_dim, device=device, dtype=dtype
),
x,
],
dim=2,
)
coeffs = torch.tensor(
[(-1) ** k * math.comb(n - 1, k) for k in range(n)],
device=device,
dtype=dtype,
)
coeffs /= coeffs.norm(p=1)
return (
(x_pad.unfold(2, n, 1) * coeffs.view(1, 1, 1, 1, n))
.flip(-1)
.permute(0, 1, 2, 4, 3)
.reshape(batch_size, num_heads, seq_len * n, head_dim)
)
def rotative_expand(x: torch.Tensor) -> torch.Tensor:
"""Expand tokens using the rotative trick."""
d_parity = head_dim // 2
angles = torch.arange(n, device=device, dtype=dtype) * (2 * math.pi / n)
cos = torch.cos(angles).view(1, 1, 1, n, 1)
sin = torch.sin(angles).view(1, 1, 1, n, 1)
if head_dim % 2:
x_pairs = x[..., :-1].view(batch_size, num_heads, seq_len, d_parity, 2)
else:
x_pairs = x.view(batch_size, num_heads, seq_len, d_parity, 2)
x_pairs = x_pairs.unsqueeze(3).expand(
batch_size, num_heads, seq_len, n, d_parity, 2
)
x0, x1 = x_pairs[..., 0], x_pairs[..., 1]
x0r = x0 * cos - x1 * sin
x1r = x0 * sin + x1 * cos
rot = torch.stack([x0r, x1r], -1).reshape(
batch_size, num_heads, seq_len, n, d_parity * 2
)
if head_dim % 2:
last = (
x[..., -1]
.unsqueeze(-1)
.unsqueeze(3)
.expand(batch_size, num_heads, seq_len, n, 1)
)
rot = torch.cat([rot, last], -1)
return rot.reshape(batch_size, num_heads, seq_len * n, head_dim)
if mode == "derivative":
return derivative_expand(x)
if mode == "rotative":
return rotative_expand(x)
if mode == "combined":
return (derivative_expand(x) + rotative_expand(x)) / 2
raise ValueError(f"Unknown mode: {mode}")
def extract_layer_idx(module_name: str) -> int:
"""Extract the layer index from a module name string."""
match = re.search(r"\.(\d+)\.", module_name)
if match:
return int(match.group(1))
return -1
def find_embedding_lm(module: nn.Module) -> Optional[nn.Module]:
"""Find the embedding weight in a model module."""
for _, child in module.named_modules():
if hasattr(child, "embed_tokens") and hasattr(child.embed_tokens, "weight"):
return child.embed_tokens
if hasattr(child, "token_embeddings") and hasattr(
child.token_embeddings, "weight"
):
return child.token_embeddings
return None
def set_trainable_parameters(
model: PreTrainedModel, trainable_patterns: List[str] = None
) -> PreTrainedModel:
"""Freeze model parameters except trainable_patterns."""
if trainable_patterns is None:
trainable_patterns = [
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"qkv_proj",
"out_proj",
"c_attn",
"c_proj",
"query",
"key",
"value",
]
for name, param in model.named_parameters():
param.requires_grad = any(pattern in name for pattern in trainable_patterns)
trainable_layers = [n for n, p in model.named_parameters() if p.requires_grad]
logger.info("Trainable parameters after freeze: %s", trainable_layers)
return model
def ensure_stability(
tensor: torch.Tensor, min_val: float = -1e4, max_val: float = 1e4
) -> torch.Tensor:
"""stability forcing"""
dtype = tensor.dtype
center = (max_val + min_val) / 2
tensor = torch.clamp(tensor, min=min_val, max=max_val)
tensor = torch.nan_to_num(tensor, nan=center, posinf=max_val, neginf=min_val)
return tensor.to(dtype=dtype)
def apply_linear_attention_mask(
attention_mask: torch.Tensor, v: torch.Tensor, padding_side: str = "right"
) -> torch.Tensor:
"""Extract if padding --> [B,S]"""
if attention_mask.dim() == 4 and attention_mask.shape[1] == 1:
mask = attention_mask.diagonal(dim1=-2, dim2=-1).squeeze(1)
else:
mask = attention_mask.squeeze(
dim=tuple(
i
for i in range(1, attention_mask.dim())
if attention_mask.shape[i] == 1
)
)
# Ensure cast to the same dtype as v and convert to binary mask
if not (
mask.dtype == torch.bool
or (
mask.dtype in [torch.uint8, torch.int32, torch.int64]
and mask.max() <= 1
and mask.min() >= 0
)
):
mask = (mask >= 0).to(v.dtype) # [-inf, 0, 0, -inf] --> [0, 1, 1, 0]
else:
mask = mask.to(v.dtype)
# mask is [batch, seq] --> Broadcast to v [batch, seq, (...)]
if padding_side == "left":
mask = mask[:, -v.shape[-2] :][(...,) + (None,) * (v.dim() - 2)]
else: # right padding
mask = mask[:, : v.shape[-2]][(...,) + (None,) * (v.dim() - 2)]
return v * mask
def truncate_attention_mask(
hidden_states: torch.Tensor, attention_mask: torch.Tensor, max_length: int
) -> tuple[torch.Tensor, torch.Tensor]:
"""Truncate hidden_states and attention_mask to the last window of size max_length"""
seq_dim = 1 # convention: (batch, seq, ...)
seq_len = hidden_states.shape[seq_dim]
if seq_len > max_length:
hidden_states = hidden_states.narrow(seq_dim, seq_len - max_length, max_length)
if attention_mask is not None:
# mask [batch, seq]
if attention_mask.dim() == 2:
attention_mask = attention_mask[:, -max_length:]
# mask [batch, seq, seq]
elif attention_mask.dim() == 3:
attention_mask = attention_mask[:, -max_length:, -max_length:]
# mask [batch, 1, seq, seq]
elif attention_mask.dim() == 4 and attention_mask.shape[1] == 1:
attention_mask = attention_mask[:, :, -max_length:, -max_length:]
else:
raise ValueError(
"No dimension in attention_mask matches sequence length of hidden_states."
)
return hidden_states, attention_mask
def fast_invert_matrix(
tri_tensor: torch.Tensor, dtype: torch.dtype = torch.float32
) -> torch.Tensor:
"""Equivalent to vectorized forward substitution applied to the identity matrix."""
tri_tensor = tri_tensor.to(dtype=dtype).clone()
chunk_size = tri_tensor.shape[-1]
for i in range(1, chunk_size):
tri_tensor[..., i, :i] = tri_tensor[..., i, :i] + (
tri_tensor[..., i, :, None].clone() * tri_tensor[..., :, :i].clone()
).sum(-2)
tri_tensor = tri_tensor + torch.eye(
chunk_size, dtype=dtype, device=tri_tensor.device
)
return tri_tensor.to(dtype=dtype)
def get_valid_chunk_size(total_l: int, chunk_size: int) -> int:
"""Return the largest chunk_size <= chunk_size that divides total_l."""
for c in range(min(chunk_size, total_l), 0, -1):
if total_l % c == 0:
return c
return 1
## RARELY
def split_qkv(
base_attn: nn.Module, qkv: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Split the QKV tensor into separate Q, K, and V tensors."""
num_q_heads = getattr(base_attn, "num_q_heads", None)
num_k_heads = getattr(base_attn, "num_k_heads", None)
num_v_heads = getattr(base_attn, "num_v_heads", None)
head_dim = getattr(base_attn, "head_dim", None)
if num_q_heads is None or num_k_heads is None or num_v_heads is None:
raise ValueError(
"Base attention must have num_q_heads, num_k_heads, and num_v_heads defined."
)
q_len = num_q_heads * head_dim
k_len = num_k_heads * head_dim
v_len = num_v_heads * head_dim
q, k, v = torch.split(qkv, [q_len, k_len, v_len], dim=-1)
return q, k, v
## OPTIONAL
def match_dim(x: torch.Tensor, dim: int, target_size: int) -> torch.Tensor:
"""Match the size of tensor x along dimension dim to target_size by interpolation"""
src_size = x.shape[dim]
if src_size == target_size:
return x
x = torch.moveaxis(x, dim, -1)
shape = x.shape
if src_size < target_size:
x = x.reshape(-1, 1, src_size)
x = F.interpolate(x, size=target_size, mode="linear", align_corners=False)
x = x.reshape(*shape[:-1], target_size)
else:
eye = torch.eye(target_size, src_size, device=x.device, dtype=x.dtype)
x = F.linear(x, eye) # pylint: disable=not-callable
x = torch.moveaxis(x, -1, dim)
return x
def soft_clamp(
x: torch.Tensor, min_val: float = 1e-6, max_val: float = 1 - 1e-6
) -> torch.Tensor:
"""Differentiable clamping for stability"""
dtype = x.dtype
scale = (max_val - min_val) / 2
center = (max_val + min_val) / 2
return (torch.tanh((x - center) / scale) * scale + center).to(dtype=dtype)
def describe(x: torch.Tensor, name="tensor") -> None:
"""Prints the shape, min, max, mean, and std of a tensor."""
stats = (x.min(), x.max(), x.mean(), x.std())
print(
f"{name} shape: {tuple(x.shape)}, "
+ f"min: {stats[0]:.4g}, max: {stats[1]:.4g}, "
+ f"mean: {stats[2]:.4g}, std: {stats[3]:.4g}, "
+ f"dtype: {x.dtype}, device: {x.device}"
)
|