File size: 70,373 Bytes
60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 4e6d38e 60b91bb 1739429 db3ebcb 1739429 60b91bb 1739429 db3ebcb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 db3ebcb 1739429 60b91bb 1739429 db3ebcb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1cbb7a1 60b91bb 1739429 1cbb7a1 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 60b91bb 1739429 | 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 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 | """
BitMar Model for Hugging Face Transformers
BitNet-quantized Vision-Language Episodic Memory Transformer
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import logging
import math
import os
import pickle
import gzip
from typing import Dict, List, Optional, Tuple, Union
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutput, BaseModelOutput
import time
logger = logging.getLogger(__name__)
class BitMarConfig(PretrainedConfig):
"""Configuration class for BitMar model"""
model_type = "bitmar"
def __init__(
self,
vocab_size: int = 50257,
text_encoder_dim: int = 128,
text_encoder_layers: int = 4,
text_encoder_heads: int = 4,
text_decoder_dim: int = 128,
text_decoder_layers: int = 4,
text_decoder_heads: int = 4,
vision_encoder_dim: int = 768,
vision_latent_size: int = 128,
vision_hidden_size: int = 64,
vision_compression_method: str = "learned_compression",
vision_spatial_pooling: bool = True,
vision_pool_size: int = 2,
fusion_hidden_size: int = 128,
fusion_num_heads: int = 4,
fusion_num_layers: int = 2,
memory_size: int = 32,
episode_dim: int = 128,
memory_alpha: float = 0.2,
direct_writing: bool = True,
memory_compression: bool = True,
max_seq_len: int = 256,
dropout: float = 0.15,
initializer_range: float = 0.02,
layer_norm_epsilon: float = 1e-5,
use_cache: bool = True,
tie_word_embeddings: bool = True,
pad_token_id: int = 50256,
bos_token_id: int = 50256,
eos_token_id: int = 50256,
**kwargs
):
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
**kwargs
)
self.vocab_size = vocab_size
self.text_encoder_dim = text_encoder_dim
self.text_encoder_layers = text_encoder_layers
self.text_encoder_heads = text_encoder_heads
self.text_decoder_dim = text_decoder_dim
self.text_decoder_layers = text_decoder_layers
self.text_decoder_heads = text_decoder_heads
self.vision_encoder_dim = vision_encoder_dim
self.vision_latent_size = vision_latent_size
self.vision_hidden_size = vision_hidden_size
self.vision_compression_method = vision_compression_method
self.vision_spatial_pooling = vision_spatial_pooling
self.vision_pool_size = vision_pool_size
self.fusion_hidden_size = fusion_hidden_size
self.fusion_num_heads = fusion_num_heads
self.fusion_num_layers = fusion_num_layers
self.memory_size = memory_size
self.episode_dim = episode_dim
self.memory_alpha = memory_alpha
self.direct_writing = direct_writing
self.memory_compression = memory_compression
self.max_seq_len = max_seq_len
self.dropout = dropout
self.initializer_range = initializer_range
self.layer_norm_epsilon = layer_norm_epsilon
self.use_cache = use_cache
self.tie_word_embeddings = tie_word_embeddings
class BitNetLinear(nn.Module):
"""1.58-bit Linear layer following BitNet b1.58 architecture - FIXED VERSION"""
def __init__(self, in_features: int, out_features: int, bias: bool = True):
super().__init__()
self.in_features = in_features
self.out_features = out_features
# Weight parameters (full precision for training)
self.weight = nn.Parameter(torch.randn(out_features, in_features))
self.bias = nn.Parameter(torch.zeros(out_features)) if bias else None
# FIXED
self.register_buffer('weight_scale', torch.tensor(1.0))
self.register_buffer('input_scale', torch.tensor(1.0))
def quantize_weights_1_58_bit(self, weight: torch.Tensor) -> torch.Tensor:
"""BitNet b1.58 weight quantization: {-1, 0, +1}"""
# Handle empty tensors
if weight.numel() == 0:
return weight
# Compute scaling factor with numerical stability
scale = weight.abs().mean()
# Handle case where all weights are zero
if scale < 1e-8:
scale = torch.tensor(1e-5, device=weight.device, dtype=weight.dtype)
self.weight_scale.data = scale.clamp(min=1e-5, max=1e3)
# Normalize weights with gradient clipping
weight_norm = torch.clamp(weight / self.weight_scale, min=-10.0, max=10.0)
# 1.58-bit quantization with threshold
threshold = 2.0 / 3.0 # Optimal threshold for ternary quantization
# Create ternary weights
quantized = torch.zeros_like(weight_norm)
quantized[weight_norm > threshold] = 1.0
quantized[weight_norm < -threshold] = -1.0
# Values between -threshold and threshold remain 0
return quantized
def quantize_activations_8bit(self, x: torch.Tensor) -> torch.Tensor:
"""8-bit activation quantization with numerical stability"""
# Handle empty tensors
if x.numel() == 0:
return x
# Clamp extreme values to prevent overflow
x_clamped = torch.clamp(x, min=-1e6, max=1e6)
# Handle scalar tensors
if x_clamped.numel() == 1:
return x_clamped
# Compute quantization parameters
x_min, x_max = x_clamped.min(), x_clamped.max()
# Prevent division by zero
range_val = x_max - x_min
if range_val < 1e-8:
return x_clamped
scale = range_val / 255.0
self.input_scale.data = scale.clamp(min=1e-8, max=1e3)
# Quantize to 8-bit
zero_point = (-x_min / scale).round().clamp(0, 255)
quantized = ((x_clamped / scale) + zero_point).round().clamp(0, 255)
# Dequantize
dequantized = scale * (quantized - zero_point)
return dequantized
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.training:
# Full precision training with straight-through estimator
# Forward pass with quantized weights but gradients flow through original weights
weight_q = self.quantize_weights_1_58_bit(self.weight)
weight_forward = weight_q * self.weight_scale
# Use original weight for gradient computation
weight_forward = weight_forward + (self.weight - self.weight.detach())
return F.linear(x, weight_forward, self.bias)
else:
# Inference with full quantization
weight_q = self.quantize_weights_1_58_bit(self.weight) * self.weight_scale
x_q = self.quantize_activations_8bit(x)
return F.linear(x_q, weight_q, self.bias)
class BitNetMLP(nn.Module):
"""BitNet MLP block with 1.58-bit quantization"""
def __init__(self, dim: int, hidden_dim: int, dropout: float = 0.1):
super().__init__()
self.fc1 = BitNetLinear(dim, hidden_dim)
self.fc2 = BitNetLinear(hidden_dim, dim)
self.activation = nn.GELU()
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm(dim)
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
x = self.fc1(x)
x = self.activation(x)
x = self.dropout(x)
x = self.fc2(x)
x = self.dropout(x)
return self.norm(x + residual)
class BitNetAttention(nn.Module):
"""Multi-head attention with BitNet quantization"""
def __init__(
self,
dim: int,
num_heads: int,
dropout: float = 0.1,
bias: bool = True
):
super().__init__()
assert dim % num_heads == 0
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim ** -0.5
# BitNet quantized projections
self.q_proj = BitNetLinear(dim, dim, bias=bias)
self.k_proj = BitNetLinear(dim, dim, bias=bias)
self.v_proj = BitNetLinear(dim, dim, bias=bias)
self.out_proj = BitNetLinear(dim, dim, bias=bias)
self.dropout = nn.Dropout(dropout)
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
mask: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
batch_size, seq_len = query.shape[:2]
# Validate input dimensions
if query.size(-1) != self.dim:
raise ValueError(f"Query dimension {query.size(-1)} doesn't match expected {self.dim}")
if key.size(-1) != self.dim:
raise ValueError(f"Key dimension {key.size(-1)} doesn't match expected {self.dim}")
if value.size(-1) != self.dim:
raise ValueError(f"Value dimension {value.size(-1)} doesn't match expected {self.dim}")
# Linear projections
q = self.q_proj(query)
k = self.k_proj(key)
v = self.v_proj(value)
# Get key/value sequence length (handle different shapes)
key_seq_len = key.size(1)
# Reshape for multi-head attention with proper dimension checking
q = q.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(batch_size, key_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(batch_size, key_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
# Attention computation
attention_scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
if mask is not None:
# Handle mask shape: expand to match attention scores shape
if mask.dim() == 2: # [batch_size, seq_len]
mask = mask.unsqueeze(1).unsqueeze(1) # [batch_size, 1, 1, seq_len]
elif mask.dim() == 3: # [batch_size, seq_len, seq_len]
mask = mask.unsqueeze(1) # [batch_size, 1, seq_len, seq_len]
# Expand mask to match attention scores shape [batch_size, num_heads, seq_len, key_seq_len]
if mask.size(-1) != key_seq_len:
# Adjust mask if needed
if mask.size(-1) == seq_len:
# Pad or trim mask to match key_seq_len
if key_seq_len > seq_len:
pad_size = key_seq_len - seq_len
mask = torch.cat([mask, torch.zeros(*mask.shape[:-1], pad_size, device=mask.device, dtype=mask.dtype)], dim=-1)
else:
mask = mask[..., :key_seq_len]
mask = mask.expand(batch_size, self.num_heads, seq_len, key_seq_len)
attention_scores.masked_fill_(mask == 0, float('-inf'))
attention_weights = F.softmax(attention_scores, dim=-1)
attention_weights = self.dropout(attention_weights)
# Apply attention to values
attended = torch.matmul(attention_weights, v)
# Reshape and project output
attended = attended.transpose(1, 2).contiguous().view(
batch_size, seq_len, self.dim
)
output = self.out_proj(attended)
return output, attention_weights.mean(dim=1) # Average across heads
class BitNetTransformerBlock(nn.Module):
"""BitNet Transformer block with quantized components"""
def __init__(
self,
dim: int,
num_heads: int,
mlp_ratio: float = 4.0,
dropout: float = 0.1
):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
self.attn = BitNetAttention(dim, num_heads, dropout)
self.norm2 = nn.LayerNorm(dim)
self.mlp = BitNetMLP(dim, int(dim * mlp_ratio), dropout)
def forward(
self,
x: torch.Tensor,
mask: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
# Self-attention with residual connection
normed_x = self.norm1(x)
attn_out, attn_weights = self.attn(normed_x, normed_x, normed_x, mask)
x = x + attn_out
# MLP with residual connection
x = x + self.mlp(self.norm2(x))
return x, attn_weights
class BitNetTextEncoder(nn.Module):
"""BitNet-based text encoder"""
def __init__(
self,
vocab_size: int,
dim: int,
num_layers: int,
num_heads: int,
max_seq_len: int = 512,
dropout: float = 0.1
):
super().__init__()
self.dim = dim
self.max_seq_len = max_seq_len
# Token embeddings (kept full precision)
self.token_embedding = nn.Embedding(vocab_size, dim)
self.position_embedding = nn.Embedding(max_seq_len, dim)
# BitNet transformer layers
self.layers = nn.ModuleList([
BitNetTransformerBlock(dim, num_heads, dropout=dropout)
for _ in range(num_layers)
])
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm(dim)
# Initialize embeddings
nn.init.normal_(self.token_embedding.weight, std=0.02)
nn.init.normal_(self.position_embedding.weight, std=0.02)
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
batch_size, seq_len = input_ids.shape
# Embeddings
positions = torch.arange(seq_len, device=input_ids.device).unsqueeze(0)
x = self.token_embedding(input_ids) + \
self.position_embedding(positions)
x = self.dropout(x)
# Transform through BitNet layers
attention_patterns = []
for layer in self.layers:
# Convert attention mask to the right format for the layer
layer_mask = None
if attention_mask is not None:
# Create a mask where 1 means attend, 0 means don't attend
layer_mask = attention_mask.unsqueeze(
1).unsqueeze(2) # [batch_size, 1, 1, seq_len]
x, attn_weights = layer(x, layer_mask)
attention_patterns.append(attn_weights)
x = self.norm(x)
return x, attention_patterns
class BitNetTextDecoder(nn.Module):
"""BitNet-based text decoder with causal masking"""
def __init__(
self,
vocab_size: int,
dim: int,
num_layers: int,
num_heads: int,
max_seq_len: int = 512,
dropout: float = 0.1
):
super().__init__()
self.dim = dim
self.max_seq_len = max_seq_len
# Token embeddings
self.token_embedding = nn.Embedding(vocab_size, dim)
self.position_embedding = nn.Embedding(max_seq_len, dim)
# BitNet transformer layers
self.layers = nn.ModuleList([
BitNetTransformerBlock(dim, num_heads, dropout=dropout)
for _ in range(num_layers)
])
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm(dim)
# Output projection to vocabulary
self.lm_head = BitNetLinear(dim, vocab_size, bias=False)
# Initialize embeddings
nn.init.normal_(self.token_embedding.weight, std=0.02)
nn.init.normal_(self.position_embedding.weight, std=0.02)
# Register causal mask
self.register_buffer(
'causal_mask',
torch.tril(torch.ones(max_seq_len, max_seq_len)
).unsqueeze(0).unsqueeze(0)
)
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None
) -> Dict[str, torch.Tensor]:
if input_ids is not None:
batch_size, seq_len = input_ids.shape
positions = torch.arange(
seq_len, device=input_ids.device).unsqueeze(0)
x = self.token_embedding(input_ids) + \
self.position_embedding(positions)
elif inputs_embeds is not None:
batch_size, seq_len = inputs_embeds.shape[:2]
positions = torch.arange(
seq_len, device=inputs_embeds.device).unsqueeze(0)
x = inputs_embeds + self.position_embedding(positions)
else:
raise ValueError(
"Either input_ids or inputs_embeds must be provided")
x = self.dropout(x)
# Create causal mask
causal_mask = self.causal_mask[:, :, :seq_len, :seq_len]
if attention_mask is not None:
# Combine causal mask with padding mask
mask = attention_mask.unsqueeze(1).unsqueeze(2) * causal_mask
else:
mask = causal_mask
# Transform through BitNet layers
attention_patterns = []
for layer in self.layers:
x, attn_weights = layer(x, mask)
attention_patterns.append(attn_weights)
x = self.norm(x)
logits = self.lm_head(x)
loss = None
if labels is not None:
# Shift labels for causal LM
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100
)
return {
'logits': logits,
'loss': loss,
'attention_patterns': attention_patterns
}
class EpisodicMemory(nn.Module):
"""Episodic Memory mechanism inspired by Larimar with performance optimizations and external storage support"""
def __init__(
self,
memory_size: int,
episode_dim: int,
alpha: float = 0.1,
direct_writing: bool = True,
observation_noise_std: float = 1e-6,
external_storage: bool = False,
memory_storage_path: str = None,
compression_enabled: bool = True,
lazy_loading: bool = False
):
super().__init__()
self.memory_size = memory_size
self.episode_dim = episode_dim
self.alpha = alpha
self.direct_writing = direct_writing
self.observation_noise_std = observation_noise_std
# External storage configuration
self.external_storage = external_storage
self.memory_storage_path = memory_storage_path
self.compression_enabled = compression_enabled
self.lazy_loading = lazy_loading
self._memory_loaded = False
self._memory_version = 1
# Memory storage with improved initialization
if external_storage and lazy_loading:
# For lazy loading, we'll initialize empty and load when needed
self._memory_data = None
self._metadata = None
else:
# Standard initialization for compatibility
self.register_buffer('memory', torch.randn(memory_size, episode_dim) * 0.02)
self.register_buffer('memory_age', torch.zeros(memory_size))
self.register_buffer('memory_usage', torch.zeros(memory_size))
# Always initialize these for proper functioning
self.register_buffer('memory_quality', torch.zeros(memory_size))
self.register_buffer('memory_importance', torch.ones(memory_size))
self.register_buffer('memory_mean', torch.zeros(episode_dim))
self.register_buffer('memory_std', torch.ones(episode_dim))
self.register_buffer('update_count', torch.tensor(0))
# Enhanced memory access networks with residual connections
self.query_net = nn.Sequential(
BitNetLinear(episode_dim, episode_dim),
nn.LayerNorm(episode_dim),
nn.GELU(),
BitNetLinear(episode_dim, episode_dim)
)
self.key_net = nn.Sequential(
BitNetLinear(episode_dim, episode_dim),
nn.LayerNorm(episode_dim),
nn.GELU(),
BitNetLinear(episode_dim, episode_dim)
)
self.value_net = nn.Sequential(
BitNetLinear(episode_dim, episode_dim),
nn.LayerNorm(episode_dim),
nn.GELU(),
BitNetLinear(episode_dim, episode_dim)
)
# Add temperature parameter for attention sharpening
self.register_parameter('attention_temperature', nn.Parameter(torch.tensor(1.0)))
# Memory consolidation network for better episode encoding
self.consolidation_net = nn.Sequential(
BitNetLinear(episode_dim, episode_dim * 2),
nn.LayerNorm(episode_dim * 2),
nn.GELU(),
nn.Dropout(0.1),
BitNetLinear(episode_dim * 2, episode_dim),
nn.LayerNorm(episode_dim)
)
def _ensure_memory_loaded(self):
"""Ensure memory is loaded into device memory"""
if self.external_storage and self.lazy_loading and not self._memory_loaded:
self.load_external_memory()
elif not hasattr(self, 'memory'):
# Initialize if not present (compatibility mode)
self.register_buffer('memory', torch.randn(self.memory_size, self.episode_dim) * 0.02)
self.register_buffer('memory_age', torch.zeros(self.memory_size))
self.register_buffer('memory_usage', torch.zeros(self.memory_size))
def save_external_memory(self, path: str = None, compress: bool = None) -> str:
"""Save episodic memory to external storage"""
import os
import json
from pathlib import Path
# Use provided path or default
save_path = path or self.memory_storage_path or "episodic_memory.pt"
save_path = Path(save_path)
save_path.parent.mkdir(parents=True, exist_ok=True)
# Use provided compression setting or default
use_compression = compress if compress is not None else self.compression_enabled
# Prepare memory data
memory_data = {
'memory': self.memory.cpu() if hasattr(self, 'memory') else torch.randn(self.memory_size, self.episode_dim) * 0.02,
'memory_age': self.memory_age.cpu() if hasattr(self, 'memory_age') else torch.zeros(self.memory_size),
'memory_usage': self.memory_usage.cpu() if hasattr(self, 'memory_usage') else torch.zeros(self.memory_size),
'memory_quality': self.memory_quality.cpu(),
'memory_importance': self.memory_importance.cpu(),
'memory_mean': self.memory_mean.cpu(),
'memory_std': self.memory_std.cpu(),
'update_count': self.update_count.cpu(),
'version': self._memory_version,
'metadata': {
'memory_size': self.memory_size,
'episode_dim': self.episode_dim,
'alpha': self.alpha,
'creation_timestamp': torch.tensor(time.time()),
'compression_enabled': use_compression
}
}
# Apply compression if enabled
if use_compression:
# Quantize memory to reduce storage size
memory_data['memory'] = self._compress_memory_tensor(memory_data['memory'])
memory_data['compressed'] = True
else:
memory_data['compressed'] = False
# Save to file
torch.save(memory_data, save_path)
# Also save metadata separately for quick access
metadata_path = save_path.with_suffix('.json')
with open(metadata_path, 'w') as f:
json.dump({
'memory_size': self.memory_size,
'episode_dim': self.episode_dim,
'version': self._memory_version,
'compressed': use_compression,
'file_size_mb': save_path.stat().st_size / (1024 * 1024),
'creation_timestamp': time.time()
}, f, indent=2)
logger.info(f"💾 Episodic memory saved to: {save_path}")
logger.info(f"📊 Memory size: {save_path.stat().st_size / 1024:.1f} KB")
return str(save_path)
def load_external_memory(self, path: str = None, device: str = None) -> bool:
"""Load episodic memory from external storage"""
import json
from pathlib import Path
# Use provided path or default
load_path = path or self.memory_storage_path or "episodic_memory.pt"
load_path = Path(load_path)
if not load_path.exists():
logger.warning(f"⚠️ External memory file not found: {load_path}")
return False
try:
# Load memory data
memory_data = torch.load(load_path, map_location='cpu')
# Validate compatibility
if memory_data['metadata']['memory_size'] != self.memory_size:
logger.error(f"❌ Memory size mismatch: expected {self.memory_size}, got {memory_data['metadata']['memory_size']}")
return False
if memory_data['metadata']['episode_dim'] != self.episode_dim:
logger.error(f"❌ Episode dimension mismatch: expected {self.episode_dim}, got {memory_data['metadata']['episode_dim']}")
return False
# Set device
device = device or next(self.parameters()).device
# Decompress if needed
if memory_data.get('compressed', False):
memory_tensor = self._decompress_memory_tensor(memory_data['memory'])
else:
memory_tensor = memory_data['memory']
# Load memory tensors
if hasattr(self, 'memory'):
self.memory.copy_(memory_tensor.to(device))
self.memory_age.copy_(memory_data['memory_age'].to(device))
self.memory_usage.copy_(memory_data['memory_usage'].to(device))
else:
# Register buffers if not present (lazy loading case)
self.register_buffer('memory', memory_tensor.to(device))
self.register_buffer('memory_age', memory_data['memory_age'].to(device))
self.register_buffer('memory_usage', memory_data['memory_usage'].to(device))
self.memory_quality.copy_(memory_data['memory_quality'].to(device))
self.memory_importance.copy_(memory_data['memory_importance'].to(device))
self.memory_mean.copy_(memory_data['memory_mean'].to(device))
self.memory_std.copy_(memory_data['memory_std'].to(device))
self.update_count.copy_(memory_data['update_count'].to(device))
self._memory_version = memory_data.get('version', 1)
self._memory_loaded = True
logger.info(f"✅ Episodic memory loaded from: {load_path}")
logger.info(f"📊 Memory version: {self._memory_version}")
return True
except Exception as e:
logger.error(f"❌ Failed to load external memory: {e}")
return False
def _compress_memory_tensor(self, tensor: torch.Tensor) -> torch.Tensor:
"""Compress memory tensor for storage"""
# Quantize to int8 to reduce storage size
tensor_min = tensor.min()
tensor_max = tensor.max()
# Avoid division by zero
tensor_range = tensor_max - tensor_min
if tensor_range < 1e-8:
return tensor
# Quantize to int8 range
quantized = ((tensor - tensor_min) / tensor_range * 255).round().clamp(0, 255).to(torch.uint8)
# Store quantization parameters
return {
'data': quantized,
'min': tensor_min,
'max': tensor_max,
'original_shape': tensor.shape
}
def _decompress_memory_tensor(self, compressed_data) -> torch.Tensor:
"""Decompress memory tensor"""
if isinstance(compressed_data, dict):
quantized = compressed_data['data'].float()
tensor_min = compressed_data['min']
tensor_max = compressed_data['max']
# Dequantize
tensor_range = tensor_max - tensor_min
dequantized = (quantized / 255.0) * tensor_range + tensor_min
return dequantized.view(compressed_data['original_shape'])
else:
# Not compressed, return as-is
return compressed_data
def _update_memory_statistics(self, episodes: torch.Tensor):
"""Update running statistics for memory normalization"""
with torch.no_grad():
batch_mean = episodes.mean(dim=0)
batch_var = episodes.var(dim=0, unbiased=False)
# Exponential moving average
momentum = 0.1
self.memory_mean = (1 - momentum) * self.memory_mean + momentum * batch_mean
self.memory_std = torch.sqrt((1 - momentum) * self.memory_std**2 + momentum * batch_var)
self.update_count += 1
def _normalize_episodes(self, episodes: torch.Tensor) -> torch.Tensor:
"""Normalize episodes using running statistics"""
if self.update_count > 10: # Only normalize after some updates
return (episodes - self.memory_mean) / (self.memory_std + 1e-8)
return episodes
def _compute_episode_quality(self, episode: torch.Tensor, retrieved: torch.Tensor) -> torch.Tensor:
"""Compute quality score for memory episodes"""
# Quality based on diversity and relevance
similarity_to_memory = torch.cosine_similarity(
episode.unsqueeze(1), self.memory.unsqueeze(0), dim=-1
).max(dim=1)[0]
# Encourage diversity - lower similarity = higher quality
diversity_score = 1.0 - similarity_to_memory
# Relevance score based on retrieval quality
retrieval_quality = torch.cosine_similarity(episode, retrieved, dim=-1)
# Combined quality score
return 0.7 * diversity_score + 0.3 * retrieval_quality
def write_memory(self, episode: torch.Tensor) -> torch.Tensor:
"""Optimized memory writing with intelligent slot selection"""
batch_size = episode.size(0)
# Apply consolidation to improve episode representation
consolidated_episode = self.consolidation_net(episode) + episode # Residual connection
# Update statistics
self._update_memory_statistics(consolidated_episode)
# Normalize episodes
normalized_episode = self._normalize_episodes(consolidated_episode)
if self.direct_writing:
# Enhanced slot selection combining age, usage, and quality
if batch_size <= self.memory_size:
# Compute composite scores for slot selection
age_scores = -self.memory_age # Prefer older slots
usage_scores = -self.memory_usage # Prefer less used slots
quality_scores = -self.memory_quality # Prefer lower quality slots
importance_scores = -self.memory_importance # Prefer less important slots
# Weighted combination
composite_scores = (
0.4 * age_scores +
0.3 * usage_scores +
0.2 * quality_scores +
0.1 * importance_scores
)
_, best_indices = composite_scores.topk(batch_size, largest=True)
# Update memory slots with momentum-based updates
momentum = self.alpha
self.memory[best_indices] = (
(1 - momentum) * self.memory[best_indices] +
momentum * normalized_episode.detach()
)
# Update metadata
self.memory_age[best_indices] = self.memory_age.max() + 1
self.memory_usage[best_indices] += 1
# Update quality scores (will be computed during read)
with torch.no_grad():
# Temporary quality estimation based on internal consistency
temp_quality = torch.norm(normalized_episode, dim=-1)
self.memory_quality[best_indices] = temp_quality.detach()
else:
# Handle large batches efficiently
for i in range(0, batch_size, self.memory_size):
end_idx = min(i + self.memory_size, batch_size)
chunk_size = end_idx - i
# Apply same logic for chunks
age_scores = -self.memory_age
usage_scores = -self.memory_usage
quality_scores = -self.memory_quality
importance_scores = -self.memory_importance
composite_scores = (
0.4 * age_scores +
0.3 * usage_scores +
0.2 * quality_scores +
0.1 * importance_scores
)
_, chunk_indices = composite_scores.topk(chunk_size, largest=True)
momentum = self.alpha
self.memory[chunk_indices] = (
(1 - momentum) * self.memory[chunk_indices] +
momentum * normalized_episode[i:end_idx].detach()
)
self.memory_age[chunk_indices] = self.memory_age.max() + 1 + i
self.memory_usage[chunk_indices] += 1
temp_quality = torch.norm(normalized_episode[i:end_idx], dim=-1)
self.memory_quality[chunk_indices] = temp_quality.detach()
return consolidated_episode
def read_memory(self, query: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""Optimized memory reading with enhanced attention"""
batch_size = query.size(0)
# Validate query dimensions
if query.size(-1) != self.episode_dim:
raise ValueError(f"Query dimension {query.size(-1)} doesn't match memory episode_dim {self.episode_dim}")
# Normalize query
normalized_query = self._normalize_episodes(query)
# Enhanced query, key, value computation with residual connections
q = self.query_net(normalized_query) + normalized_query # Residual
k = self.key_net(self.memory) + self.memory # Residual
v = self.value_net(self.memory) + self.memory # Residual
# Scaled dot-product attention with learnable temperature
attention_scores = torch.matmul(q, k.transpose(0, 1)) / (
math.sqrt(self.episode_dim) * self.attention_temperature.clamp(min=0.1, max=10.0)
)
# Add importance weighting to attention scores
importance_weights = self.memory_importance.unsqueeze(0).expand(batch_size, -1)
attention_scores = attention_scores + torch.log(importance_weights + 1e-8)
# Apply attention with improved stability
attention_weights = F.softmax(attention_scores, dim=-1)
# Add attention dropout for regularization during training
if self.training:
attention_weights = F.dropout(attention_weights, p=0.1)
# Weighted memory retrieval
retrieved = torch.matmul(attention_weights, v)
# Update memory access statistics and importance
with torch.no_grad():
access_counts = attention_weights.sum(0)
self.memory_usage += access_counts
# Update importance based on usage frequency
self.memory_importance = 0.9 * self.memory_importance + 0.1 * (access_counts + 1e-8)
# Update quality scores based on retrieval effectiveness
if hasattr(self, '_last_query_quality'):
quality_update = self._compute_episode_quality(query, retrieved)
# Update quality for attended slots
attended_indices = attention_weights.max(0)[1] # Most attended slots
self.memory_quality[attended_indices] = (
0.8 * self.memory_quality[attended_indices] +
0.2 * quality_update.mean()
)
return retrieved, attention_weights
def forward(self, episode: torch.Tensor, mode: str = "read_write") -> Tuple[torch.Tensor, torch.Tensor]:
"""Enhanced forward pass with memory consolidation"""
if mode == "write":
return self.write_memory(episode), None
elif mode == "read":
return self.read_memory(episode)
else: # read_write
# Write episode to memory with consolidation
consolidated_episode = self.write_memory(episode)
# Read from memory using consolidated episode as query
retrieved, attention_weights = self.read_memory(consolidated_episode)
# Memory-augmented output combining input and retrieved memory
output = 0.7 * consolidated_episode + 0.3 * retrieved
return output, attention_weights
def get_memory_statistics(self) -> Dict[str, torch.Tensor]:
"""Get comprehensive memory statistics for monitoring"""
return {
'memory_usage_distribution': self.memory_usage,
'memory_age_distribution': self.memory_age,
'memory_quality_scores': self.memory_quality,
'memory_importance': self.memory_importance,
'attention_temperature': self.attention_temperature,
'memory_utilization': (self.memory_usage > 0).float().mean(),
'memory_diversity': torch.std(self.memory, dim=0).mean(),
'update_count': self.update_count
}
def consolidate_memory(self):
"""Explicit memory consolidation for improved organization"""
with torch.no_grad():
# Sort memory by importance and quality
importance_quality_score = 0.6 * self.memory_importance + 0.4 * self.memory_quality
sorted_indices = torch.argsort(importance_quality_score, descending=True)
# Reorganize memory to group similar episodes
sorted_memory = self.memory[sorted_indices]
self.memory.copy_(sorted_memory)
# Update corresponding metadata
self.memory_age[:] = self.memory_age[sorted_indices]
self.memory_usage[:] = self.memory_usage[sorted_indices]
self.memory_quality[:] = self.memory_quality[sorted_indices]
self.memory_importance[:] = self.memory_importance[sorted_indices]
def get_memory_info(self) -> Dict:
"""Get comprehensive memory information"""
info = {
'memory_size': self.memory_size,
'episode_dim': self.episode_dim,
'external_storage': self.external_storage,
'compression_enabled': self.compression_enabled,
'lazy_loading': self.lazy_loading,
'memory_loaded': self._memory_loaded if self.external_storage else True,
'version': self._memory_version,
'storage_path': self.memory_storage_path
}
if hasattr(self, 'memory'):
info.update({
'memory_utilization': (self.memory_usage > 0).float().mean().item(),
'memory_diversity': torch.std(self.memory, dim=0).mean().item(),
'update_count': self.update_count.item(),
'memory_device': str(self.memory.device)
})
return info
def create_memory_snapshot(self, snapshot_name: str = None) -> str:
"""Create a named snapshot of the current memory state"""
import time
from pathlib import Path
timestamp = int(time.time())
snapshot_name = snapshot_name or f"memory_snapshot_{timestamp}"
# Create snapshots directory
snapshots_dir = Path("memory_snapshots")
snapshots_dir.mkdir(exist_ok=True)
snapshot_path = snapshots_dir / f"{snapshot_name}.pt"
# Save current memory state
saved_path = self.save_external_memory(str(snapshot_path), compress=True)
logger.info(f"📸 Memory snapshot created: {saved_path}")
return saved_path
def load_memory_snapshot(self, snapshot_name: str) -> bool:
"""Load a named memory snapshot"""
from pathlib import Path
snapshots_dir = Path("memory_snapshots")
snapshot_path = snapshots_dir / f"{snapshot_name}.pt"
if not snapshot_path.exists():
logger.warning(f"⚠️ Snapshot not found: {snapshot_path}")
return False
success = self.load_external_memory(str(snapshot_path))
if success:
logger.info(f"📸 Memory snapshot loaded: {snapshot_name}")
return success
def enable_external_storage(self, storage_path: str = None, compress: bool = True, lazy: bool = False):
"""Enable external storage mode for edge deployment"""
self.external_storage = True
self.memory_storage_path = storage_path or "episodic_memory.pt"
self.compression_enabled = compress
self.lazy_loading = lazy
logger.info(f"🔄 External storage enabled: {self.memory_storage_path}")
logger.info(f" Compression: {compress}, Lazy loading: {lazy}")
def disable_external_storage(self):
"""Disable external storage and return to integrated mode"""
# Ensure memory is loaded before disabling external storage
self._ensure_memory_loaded()
self.external_storage = False
self.lazy_loading = False
self._memory_loaded = True
logger.info("🔄 External storage disabled, using integrated mode")
# ...existing code for other methods...
class CrossModalFusion(nn.Module):
"""Cross-modal fusion module for text and vision features"""
def __init__(
self,
text_dim: int,
vision_dim: int,
hidden_dim: int,
num_heads: int = 8,
num_layers: int = 2
):
super().__init__()
self.text_dim = text_dim
self.vision_dim = vision_dim
self.hidden_dim = hidden_dim
# Projection layers
self.text_proj = BitNetLinear(text_dim, hidden_dim)
self.vision_proj = BitNetLinear(vision_dim, hidden_dim)
# Cross-attention layers
self.cross_attention_layers = nn.ModuleList([
BitNetAttention(
dim=hidden_dim,
num_heads=num_heads
) for _ in range(num_layers)
])
# Layer normalization
self.layer_norms = nn.ModuleList([
nn.LayerNorm(hidden_dim) for _ in range(num_layers)
])
# Output projection
self.output_proj = BitNetLinear(hidden_dim, hidden_dim)
def forward(
self,
text_features: torch.Tensor,
vision_features: torch.Tensor
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
"""
Args:
text_features: [batch_size, seq_len, text_dim]
vision_features: [batch_size, vision_dim]
Returns:
fused_features: [batch_size, seq_len, hidden_dim]
attention_weights: Dict of attention patterns
"""
batch_size, seq_len = text_features.shape[:2]
# Validate input dimensions
if text_features.size(-1) != self.text_dim:
raise ValueError(f"Text features dimension {text_features.size(-1)} doesn't match expected {self.text_dim}")
if vision_features.size(-1) != self.vision_dim:
raise ValueError(f"Vision features dimension {vision_features.size(-1)} doesn't match expected {self.vision_dim}")
# Project to common dimension
# [batch_size, seq_len, hidden_dim]
text_proj = self.text_proj(text_features)
vision_proj = self.vision_proj(vision_features).unsqueeze(1) # [batch_size, 1, hidden_dim]
# Cross-attention fusion
fused = text_proj
attention_weights = {}
for i, (attn_layer, norm_layer) in enumerate(zip(self.cross_attention_layers, self.layer_norms)):
# Text-to-vision cross-attention
attn_output, attn_weights = attn_layer(
query=fused,
key=vision_proj,
value=vision_proj
)
# Residual connection and normalization
fused = norm_layer(fused + attn_output)
attention_weights[f'layer_{i}'] = attn_weights
# Output projection
output = self.output_proj(fused)
return output, attention_weights
class VisionEncoder(nn.Module):
"""Quantized Vision Encoder for DiNOv2 features"""
def __init__(
self,
input_dim: int = 768,
hidden_dim: int = 512,
output_dim: int = 768,
num_layers: int = 2
):
super().__init__()
# Quantized layers
self.layers = nn.ModuleList([
BitNetLinear(input_dim if i == 0 else hidden_dim, hidden_dim)
for i in range(num_layers)
])
# Output projection
self.output_proj = BitNetLinear(hidden_dim, output_dim)
# Activation and normalization
self.activation = nn.GELU()
self.layer_norms = nn.ModuleList([
nn.LayerNorm(hidden_dim) for _ in range(num_layers)
])
self.dropout = nn.Dropout(0.1)
def forward(self, vision_features: torch.Tensor) -> torch.Tensor:
"""
Args:
vision_features: [batch_size, input_dim] - DiNOv2 features
Returns:
encoded_features: [batch_size, output_dim]
"""
# Handle potential extra dimensions
if vision_features.dim() > 2:
# Flatten any extra dimensions except batch
original_shape = vision_features.shape
vision_features = vision_features.view(original_shape[0], -1)
# Ensure we have the expected input dimension
if vision_features.size(-1) != self.layers[0].in_features:
# Take only the first input_dim features if we have more
if vision_features.size(-1) > self.layers[0].in_features:
vision_features = vision_features[:, :self.layers[0].in_features]
else:
raise ValueError(f"Vision features dimension {vision_features.size(-1)} is smaller than expected {self.layers[0].in_features}")
x = vision_features
for layer, norm in zip(self.layers, self.layer_norms):
x = layer(x)
x = norm(x)
x = self.activation(x)
x = self.dropout(x)
# Output projection
output = self.output_proj(x)
return output
class BitMarModel(PreTrainedModel):
"""
BitMar: BitNet-quantized Vision-Language Episodic Memory Transformer
Compatible with Hugging Face Transformers
"""
config_class = BitMarConfig
base_model_prefix = "bitmar"
supports_gradient_checkpointing = True
_no_split_modules = ["BitNetTransformerBlock", "EpisodicMemory"]
def __init__(self, config: BitMarConfig):
super().__init__(config)
self.config = config
# Loss balancing parameters
self.cross_modal_loss_weight = getattr(config, 'cross_modal_loss_weight', 0.1)
self.text_loss_weight = getattr(config, 'text_loss_weight', 1.0)
self.vision_loss_weight = getattr(config, 'vision_loss_weight', 0.1)
self.memory_loss_weight = getattr(config, 'memory_loss_weight', 0.05)
# Dynamic loss scaling
self.adaptive_loss_scaling = getattr(config, 'adaptive_loss_scaling', True)
self.loss_scale_temperature = getattr(config, 'loss_scale_temperature', 0.07)
# Encoder freezing parameters
self.freeze_text_encoder_steps = getattr(config, 'freeze_text_encoder_steps', 0)
self.freeze_vision_encoder_steps = getattr(config, 'freeze_vision_encoder_steps', 0)
self.current_step = 0
# BitNet text encoder/decoder
self.text_encoder = BitNetTextEncoder(
vocab_size=config.vocab_size,
dim=config.text_encoder_dim,
num_layers=config.text_encoder_layers,
num_heads=config.text_encoder_heads,
max_seq_len=config.max_seq_len,
dropout=config.dropout
)
self.text_decoder = BitNetTextDecoder(
vocab_size=config.vocab_size,
dim=config.text_decoder_dim,
num_layers=config.text_decoder_layers,
num_heads=config.text_decoder_heads,
max_seq_len=config.max_seq_len,
dropout=config.dropout
)
# Vision processing with BitNet quantization
self.vision_encoder = VisionEncoder(
input_dim=config.vision_encoder_dim,
hidden_dim=config.vision_hidden_size,
output_dim=config.vision_latent_size
)
# Cross-modal fusion with BitNet
self.fusion = CrossModalFusion(
text_dim=config.text_encoder_dim,
vision_dim=config.vision_latent_size,
hidden_dim=config.fusion_hidden_size,
num_heads=config.fusion_num_heads,
num_layers=config.fusion_num_layers
)
# Episodic memory with BitNet quantization
self.memory = EpisodicMemory(
memory_size=config.memory_size,
episode_dim=config.episode_dim,
alpha=config.memory_alpha,
direct_writing=config.direct_writing
)
# Additional BitNet projection layers
self.text_to_episode = BitNetLinear(
config.text_encoder_dim,
config.episode_dim
)
self.vision_to_episode = BitNetLinear(
config.vision_latent_size,
config.episode_dim
)
self.memory_to_decoder = BitNetLinear(
config.episode_dim,
config.fusion_hidden_size
)
# Projection to decoder dimension
self.decoder_input_proj = BitNetLinear(
config.fusion_hidden_size,
config.text_decoder_dim
)
# Initialize tokenizer (for compatibility)
try:
from transformers import AutoTokenizer
self.tokenizer = AutoTokenizer.from_pretrained('gpt2')
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
except:
self.tokenizer = None
self.post_init()
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, (nn.Linear, BitNetLinear)):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if hasattr(module, 'bias') and module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
if hasattr(module, 'bias') and module.bias is not None:
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def encode_text(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
"""Encode text using BitNet encoder"""
text_features, attention_patterns = self.text_encoder(
input_ids=input_ids, attention_mask=attention_mask)
return text_features, attention_patterns
def encode_vision(self, vision_features: torch.Tensor) -> torch.Tensor:
"""Encode vision features using quantized vision encoder"""
vision_latent = self.vision_encoder(vision_features)
return vision_latent
def create_episode(
self,
text_features: torch.Tensor,
vision_latent: torch.Tensor,
attention_weights: Dict[str, torch.Tensor]
) -> torch.Tensor:
"""Create multimodal episode for memory storage"""
# Pool text features (mean pooling)
text_pooled = text_features.mean(dim=1)
# Project both text and vision to episode dimension
text_projected = self.text_to_episode(text_pooled)
vision_projected = self.vision_to_episode(vision_latent)
# Combine text and vision features
episode = text_projected + vision_projected
return episode
def create_episode_mixed(
self,
text_features: torch.Tensor,
vision_latent: torch.Tensor,
attention_weights: Dict[str, torch.Tensor],
has_vision: torch.Tensor
) -> torch.Tensor:
"""Create episodes with different handling for vision vs text-only samples"""
batch_size = text_features.size(0)
# Pool text features
text_pooled = text_features.mean(dim=1)
# Project to episode dimension
text_episode = self.text_to_episode(text_pooled)
vision_episode = self.vision_to_episode(vision_latent)
# For text-only samples, use only text features
# For multimodal samples, combine text and vision
episode = torch.zeros_like(text_episode)
# Text-only samples (has_vision == False)
text_only_mask = ~has_vision
if text_only_mask.any():
episode[text_only_mask] = text_episode[text_only_mask]
# Multimodal samples (has_vision == True)
multimodal_mask = has_vision
if multimodal_mask.any():
# Combine text and vision for multimodal samples
combined = text_episode[multimodal_mask] + vision_episode[multimodal_mask]
episode[multimodal_mask] = combined
return episode
def compute_cross_modal_contrastive_loss(
self,
text_features: torch.Tensor,
vision_features: torch.Tensor,
temperature: float = 0.07
) -> torch.Tensor:
"""Compute cross-modal contrastive loss similar to CLIP"""
batch_size = text_features.shape[0]
# Handle dimension mismatch between text and vision features
text_dim = text_features.shape[-1]
vision_dim = vision_features.shape[-1]
if text_dim != vision_dim:
# Project to smaller dimension to maintain compatibility
target_dim = min(text_dim, vision_dim)
if text_dim > vision_dim:
# Project text features to vision dimension
text_features = text_features[:, :target_dim]
else:
# Project vision features to text dimension
vision_features = vision_features[:, :target_dim]
# Normalize features
text_features = F.normalize(text_features, dim=-1)
vision_features = F.normalize(vision_features, dim=-1)
# Compute similarity matrix
logits = torch.matmul(text_features, vision_features.T) / temperature
# Create labels (diagonal should be positive pairs)
labels = torch.arange(batch_size, device=logits.device)
# Compute cross-entropy loss for both directions
text_to_vision_loss = F.cross_entropy(logits, labels)
vision_to_text_loss = F.cross_entropy(logits.T, labels)
return (text_to_vision_loss + vision_to_text_loss) / 2
def compute_vision_reconstruction_loss(
self,
original_vision: torch.Tensor,
reconstructed_vision: torch.Tensor
) -> torch.Tensor:
"""Compute vision reconstruction loss to prevent vision encoder collapse"""
return F.mse_loss(reconstructed_vision, original_vision)
def compute_memory_consistency_loss(
self,
episode: torch.Tensor,
retrieved_memory: torch.Tensor
) -> torch.Tensor:
"""Compute memory consistency loss to encourage meaningful memory usage"""
# L2 regularization on memory difference
memory_diff = episode - retrieved_memory
return torch.mean(torch.norm(memory_diff, dim=-1))
def compute_balanced_loss(
self,
decoder_loss: torch.Tensor,
cross_modal_loss: torch.Tensor,
vision_loss: Optional[torch.Tensor] = None,
memory_loss: Optional[torch.Tensor] = None,
step: int = 0,
adaptive_controller=None
) -> Dict[str, torch.Tensor]:
"""Compute balanced multi-objective loss with adaptive scaling"""
losses = {'decoder_loss': decoder_loss, 'cross_modal_loss': cross_modal_loss}
if vision_loss is not None:
losses['vision_loss'] = vision_loss
if memory_loss is not None:
losses['memory_loss'] = memory_loss
if self.adaptive_loss_scaling:
# Adaptive scaling based on loss magnitudes
with torch.no_grad():
# Compute relative loss scales
decoder_scale = decoder_loss.detach()
cross_modal_scale = cross_modal_loss.detach()
# Prevent division by zero
if decoder_scale > 1e-8:
adaptive_cross_modal_weight = (decoder_scale / cross_modal_scale.clamp(min=1e-8)) * self.cross_modal_loss_weight
else:
adaptive_cross_modal_weight = self.cross_modal_loss_weight
# Clamp adaptive weights
adaptive_cross_modal_weight = torch.clamp(adaptive_cross_modal_weight, 0.01, 1.0)
else:
adaptive_cross_modal_weight = self.cross_modal_loss_weight
# Apply loss scheduling (increase cross-modal importance over time)
cross_modal_schedule = min(1.0, step / 50000) # Ramp up over 50k steps
scheduled_cross_modal_weight = adaptive_cross_modal_weight * cross_modal_schedule
# Compute weighted total loss
total_loss = (
self.text_loss_weight * decoder_loss +
scheduled_cross_modal_weight * cross_modal_loss
)
if vision_loss is not None:
total_loss += self.vision_loss_weight * vision_loss
if memory_loss is not None:
total_loss += self.memory_loss_weight * memory_loss
losses.update({
'total_loss': total_loss,
'cross_modal_weight': scheduled_cross_modal_weight,
'adaptive_weight': adaptive_cross_modal_weight if self.adaptive_loss_scaling else torch.tensor(0.0)
})
return losses
def apply_encoder_freezing(self, step: int):
"""Apply temporary encoder freezing based on training step"""
self.current_step = step
# Freeze text encoder if within freezing window
freeze_text = step < self.freeze_text_encoder_steps
for param in self.text_encoder.parameters():
param.requires_grad = not freeze_text
# Freeze vision encoder if within freezing window
freeze_vision = step < self.freeze_vision_encoder_steps
for param in self.vision_encoder.parameters():
param.requires_grad = not freeze_vision
return {
'text_encoder_frozen': freeze_text,
'vision_encoder_frozen': freeze_vision
}
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
vision_features: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
mode: str = "train",
step: int = 0,
has_vision: Optional[torch.Tensor] = None,
**kwargs
) -> Union[Tuple, CausalLMOutput]:
"""
Forward pass through BitMar model with mixed vision/text batch support
Args:
has_vision: Boolean tensor [batch_size] indicating which samples have real vision features
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# CRITICAL FIX: Ensure input_ids are integers
if input_ids.dtype != torch.long:
input_ids = input_ids.long()
# CRITICAL FIX: Ensure labels are integers if provided
if labels is not None and labels.dtype != torch.long:
labels = labels.long()
if input_ids is None:
raise ValueError("input_ids must be provided")
batch_size, seq_len = input_ids.shape
# Handle missing attention mask
if attention_mask is None:
attention_mask = torch.ones_like(input_ids, dtype=torch.float)
# Ensure attention_mask is float
if attention_mask.dtype != torch.float:
attention_mask = attention_mask.float()
# Handle missing vision features
if vision_features is None:
vision_features = torch.zeros(batch_size, self.config.vision_encoder_dim,
device=input_ids.device, dtype=torch.float32)
# Validate input tensor dimensions
expected_vision_dim = self.config.vision_encoder_dim
if vision_features.dim() != 2 or vision_features.size(-1) != expected_vision_dim:
if vision_features.dim() > 2:
vision_features = vision_features.view(batch_size, -1)
if vision_features.size(-1) != expected_vision_dim:
# Pad or trim to expected dimension
if vision_features.size(-1) > expected_vision_dim:
vision_features = vision_features[:, :expected_vision_dim]
else:
padding = expected_vision_dim - vision_features.size(-1)
vision_features = F.pad(vision_features, (0, padding))
# Default has_vision to all True if not provided (backward compatibility)
if has_vision is None:
has_vision = torch.ones(batch_size, dtype=torch.bool, device=input_ids.device)
# Apply encoder freezing
freezing_status = {}
if mode == "train":
freezing_status = self.apply_encoder_freezing(step)
# Encode text (always available)
text_features, text_attention = self.encode_text(input_ids, attention_mask)
# Encode vision (with masking for text-only samples)
vision_latent = self.encode_vision(vision_features)
# Mask vision features for text-only samples
vision_mask = has_vision.float().unsqueeze(-1)
vision_latent_masked = vision_latent * vision_mask
# Cross-modal fusion
fused_features, cross_attention = self.fusion(text_features, vision_latent_masked)
# Create episodes
if has_vision.any() and (~has_vision).any():
# Mixed batch - use mixed episode creation
episode = self.create_episode_mixed(
text_features, vision_latent_masked, cross_attention, has_vision
)
else:
# Uniform batch - use standard episode creation
episode = self.create_episode(
text_features, vision_latent_masked, cross_attention
)
# Episodic memory interaction
if mode == "train":
retrieved_memory, memory_attention = self.memory(episode, mode="read_write")
else:
retrieved_memory, memory_attention = self.memory(episode, mode="read")
# Prepare decoder input
memory_context = self.memory_to_decoder(retrieved_memory)
memory_context_expanded = memory_context.unsqueeze(1).expand(-1, seq_len, -1)
fused_with_memory = fused_features + memory_context_expanded
decoder_input = self.decoder_input_proj(fused_with_memory)
# Generate text using BitNet decoder
decoder_outputs = self.text_decoder(
inputs_embeds=decoder_input,
attention_mask=attention_mask,
labels=labels
)
# Compute losses if in training mode
final_loss = None
loss_dict = {}
if mode == "train" and labels is not None:
# Primary decoder loss
decoder_loss = decoder_outputs['loss']
# Cross-modal contrastive loss (only for samples with vision)
cross_modal_loss = torch.tensor(0.0, device=input_ids.device)
if has_vision.any():
vision_indices = has_vision.nonzero(as_tuple=True)[0]
if len(vision_indices) > 0:
text_pooled = text_features[vision_indices].mean(dim=1)
vision_for_loss = vision_latent[vision_indices]
cross_modal_loss = self.compute_cross_modal_contrastive_loss(
text_pooled, vision_for_loss, temperature=self.loss_scale_temperature
)
# Optional additional losses
vision_loss = None
memory_loss = self.compute_memory_consistency_loss(episode, retrieved_memory)
# Compute balanced loss
loss_dict = self.compute_balanced_loss(
decoder_loss, cross_modal_loss, vision_loss, memory_loss, step
)
final_loss = loss_dict['total_loss']
elif decoder_outputs.get('loss') is not None:
final_loss = decoder_outputs['loss']
# Prepare outputs
if return_dict:
output = CausalLMOutput(
loss=final_loss,
logits=decoder_outputs['logits'],
hidden_states=fused_features if output_hidden_states else None,
attentions=text_attention if output_attentions else None,
)
# Add additional outputs for analysis
if mode == "train":
for key, value in loss_dict.items():
setattr(output, key, value)
for key, value in freezing_status.items():
setattr(output, key, value)
return output
else:
outputs = (decoder_outputs['logits'],)
if final_loss is not None:
outputs = (final_loss,) + outputs
if output_hidden_states:
outputs = outputs + (fused_features,)
if output_attentions:
outputs = outputs + (text_attention,)
return outputs
def generate(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
vision_features: Optional[torch.Tensor] = None,
max_length: int = 100,
temperature: float = 0.7,
top_p: float = 0.9,
do_sample: bool = True,
**kwargs
) -> torch.LongTensor:
"""Generate text given input text and vision features"""
self.eval()
batch_size = input_ids.size(0)
device = input_ids.device
# Handle missing vision features
if vision_features is None:
vision_features = torch.zeros(batch_size, self.config.vision_encoder_dim,
device=device, dtype=torch.float32)
# Handle attention mask
if attention_mask is None:
attention_mask = torch.ones_like(input_ids)
generated_ids = input_ids.clone()
current_attention_mask = attention_mask.clone()
with torch.no_grad():
for _ in range(max_length - input_ids.size(1)):
# Get model outputs
outputs = self.forward(
input_ids=generated_ids,
attention_mask=current_attention_mask,
vision_features=vision_features,
mode="inference",
return_dict=True
)
# Get next token logits
next_token_logits = outputs.logits[:, -1, :] / temperature
if do_sample:
# Apply top-p sampling
if top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above the threshold
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
next_token_logits[indices_to_remove] = float('-inf')
# Sample from the filtered distribution
probs = F.softmax(next_token_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
else:
# Greedy decoding
next_token = next_token_logits.argmax(dim=-1, keepdim=True)
# Append to generated sequence
generated_ids = torch.cat([generated_ids, next_token], dim=-1)
# Update attention mask
current_attention_mask = torch.cat([
current_attention_mask,
torch.ones(batch_size, 1, device=device)
], dim=-1)
# Stop if EOS token is generated
if (next_token == self.config.eos_token_id).all():
break
return generated_ids
def prepare_inputs_for_generation(
self,
input_ids,
past_key_values=None,
attention_mask=None,
vision_features=None,
**kwargs
):
"""Prepare inputs for generation"""
return {
"input_ids": input_ids,
"attention_mask": attention_mask,
"vision_features": vision_features,
"use_cache": kwargs.get("use_cache", True),
}
# Register the model with transformers
from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
AutoConfig.register("bitmar", BitMarConfig)
AutoModel.register(BitMarConfig, BitMarModel)
AutoModelForCausalLM.register(BitMarConfig, BitMarModel) |