File size: 73,315 Bytes
7493ebb f98df9d 7493ebb | 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 | from __future__ import annotations
import warnings
from dataclasses import dataclass
from typing import NamedTuple, Optional
import torch
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss
from torch.utils.checkpoint import checkpoint as torch_checkpoint
from transformers import AutoConfig, AutoModelForMaskedLM, PreTrainedModel
from transformers.modeling_outputs import MaskedLMOutput
from transformers.utils import ModelOutput
from .configuration_recursive import RecursiveMLMConfig
@dataclass
class IterationMetrics(ModelOutput):
"""Metrics for a single iteration of recursive refinement."""
accuracy: Optional[float] = None
entropy: Optional[float] = None
softmax_ce: Optional[float] = None
full_sequence_accuracy: Optional[float] = None
min_sequence_confidence: Optional[float] = None
@dataclass
class RecursiveMaskedLMOutput(MaskedLMOutput):
iteration_metrics: Optional[dict[int, IterationMetrics]] = None # Maps iteration index to metrics
next_soft_embeds: Optional[torch.Tensor] = None # For caching between training steps
all_logits: Optional[list[torch.Tensor]] = None # All T iterations' logits for trainer loss computation
# Flow matching state (for distillation — compact H-dim, not V-dim)
flow_noise_embed: Optional[torch.Tensor] = None # (num_masked, H) noise embedding
flow_t: Optional[torch.Tensor] = None # (num_masked,) per-token time levels
class SelfDistillationOutput(NamedTuple):
"""Output from self-distillation forward pass."""
loss: torch.Tensor # KL divergence loss (scalar, has grad)
teacher_logits: torch.Tensor # For metrics/debugging (detached)
student_logits: torch.Tensor # For metrics/debugging (has grad)
degradation_temperature: float # Mean per-token temperature sampled
teacher_entropy: float # Entropy of teacher distribution (for monitoring)
student_entropy: float # Entropy of student distribution (for monitoring)
agreement_rate: float # Fraction where teacher and student argmax agree
class RecursiveMaskedLM(PreTrainedModel):
"""
Wraps any HF MLM with recursive soft-token refinement.
At each step:
1. Normalize logits -> probs
2. Compute soft embeddings: probs @ embedding_weight + mask_embedding
3. Forward through MLM
4. Accumulate weighted loss
"""
config_class = RecursiveMLMConfig
base_model_prefix = "mlm"
supports_gradient_checkpointing = True
def __init__(self, config: RecursiveMLMConfig, base_model: Optional[PreTrainedModel] = None):
super().__init__(config)
if base_model is not None:
# Pre-trained model provided - assign directly WITHOUT calling post_init()
# to avoid reinitializing the pre-trained weights via _init_weights()
self.mlm = base_model
elif config.base_model_config is not None:
model_type = config.base_model_config.get("model_type", "")
if model_type == "llada":
from .configuration_llada import LLaDAConfig
from .modeling_llada import LLaDAModelLM
base_config = LLaDAConfig.from_dict(config.base_model_config)
self.mlm = LLaDAModelLM(base_config)
else:
base_config = AutoConfig.for_model(**config.base_model_config)
self.mlm = AutoModelForMaskedLM.from_config(base_config)
# Only call post_init() for freshly created models (needs weight init)
self.post_init()
else:
raise ValueError("Need either base_model or config.base_model_config")
@classmethod
def from_mlm_pretrained(
cls,
mlm_name_or_path: str,
num_recursions: int = 8,
normalization: str = "softmax",
loss_weight: str = "linear",
mask_token_id: Optional[int] = None,
temperature: float = 1.0,
gradient_steps: Optional[int] = None,
# === Convergence schedule parameters ===
schedule: str = "linear",
causal_strength: float = 1.0,
# === Effect parameters ===
temperature_max: float = 0.0,
entropy_target_max: float = 0.0,
entropy_floor_max: float = 0.0,
smear_sigma_max: float = 0.0,
noise_std_max: float = 0.0,
iteration_rope_dim_fraction: float = 0.0,
use_recursion_checkpointing: bool = True,
# === Soft embedding method ===
soft_embedding_method: str = "softmax",
soft_embedding_ema_step: float = 1.0,
# === Flow matching parameters ===
flow_matching_enabled: bool = False,
flow_matching_lambda: float = 0.5,
flow_matching_t_distribution: str = "logit_normal",
flow_matching_t_logit_mean: float = -0.4,
flow_matching_t_logit_std: float = 1.0,
flow_matching_t_min: float = 0.01,
flow_matching_t_max: float = 0.99,
flow_matching_mask_scale: bool = False,
**model_kwargs,
) -> "RecursiveMaskedLM":
"""Load a pretrained MLM and wrap it for recursive refinement."""
base_model = AutoModelForMaskedLM.from_pretrained(mlm_name_or_path, **model_kwargs)
return cls.from_base_model(
base_model,
num_recursions=num_recursions,
normalization=normalization,
loss_weight=loss_weight,
mask_token_id=mask_token_id,
temperature=temperature,
gradient_steps=gradient_steps,
schedule=schedule,
causal_strength=causal_strength,
temperature_max=temperature_max,
entropy_target_max=entropy_target_max,
entropy_floor_max=entropy_floor_max,
smear_sigma_max=smear_sigma_max,
noise_std_max=noise_std_max,
iteration_rope_dim_fraction=iteration_rope_dim_fraction,
use_recursion_checkpointing=use_recursion_checkpointing,
soft_embedding_method=soft_embedding_method,
soft_embedding_ema_step=soft_embedding_ema_step,
flow_matching_enabled=flow_matching_enabled,
flow_matching_lambda=flow_matching_lambda,
flow_matching_t_distribution=flow_matching_t_distribution,
flow_matching_t_logit_mean=flow_matching_t_logit_mean,
flow_matching_t_logit_std=flow_matching_t_logit_std,
flow_matching_t_min=flow_matching_t_min,
flow_matching_t_max=flow_matching_t_max,
flow_matching_mask_scale=flow_matching_mask_scale,
)
@classmethod
def from_base_model(
cls,
base_model: PreTrainedModel,
num_recursions: int = 8,
normalization: str = "softmax",
loss_weight: str = "linear",
mask_token_id: Optional[int] = None,
temperature: float = 1.0,
gradient_steps: Optional[int] = None,
# === Convergence schedule parameters ===
schedule: str = "linear",
causal_strength: float = 1.0,
# === Effect parameters ===
temperature_max: float = 0.0,
entropy_target_max: float = 0.0,
entropy_floor_max: float = 0.0,
smear_sigma_max: float = 0.0,
noise_std_max: float = 0.0,
iteration_rope_dim_fraction: float = 0.0,
use_recursion_checkpointing: bool = True,
# === Soft embedding method ===
soft_embedding_method: str = "softmax",
soft_embedding_ema_step: float = 1.0,
# === Flow matching parameters ===
flow_matching_enabled: bool = False,
flow_matching_lambda: float = 0.5,
flow_matching_t_distribution: str = "logit_normal",
flow_matching_t_logit_mean: float = -0.4,
flow_matching_t_logit_std: float = 1.0,
flow_matching_t_min: float = 0.01,
flow_matching_t_max: float = 0.99,
flow_matching_mask_scale: bool = False,
) -> "RecursiveMaskedLM":
"""Wrap an existing model for recursive refinement.
Use this for models not loadable via AutoModelForMaskedLM (e.g., LLaDA).
Args:
base_model: The base MLM model to wrap
num_recursions: Number of recursive refinement steps
normalization: Normalization method for logits (softmax, stable_softmax)
loss_weight: Loss weighting scheme (last_1, last_2, linear, uniform)
mask_token_id: Token ID for [MASK]
temperature: Temperature for softmax normalization
gradient_steps: Number of final steps to backprop through
schedule: Convergence schedule type ("linear" or "causal")
causal_strength: How much faster early positions converge (causal only)
temperature_max: Max temperature boost for uncertain positions
entropy_target_max: Target entropy at progress=0 (two-sided, recommended)
entropy_floor_max: Min entropy floor (one-sided)
smear_sigma_max: Max Gaussian sigma for position smearing
noise_std_max: Max std of Gaussian noise on logits
iteration_rope_dim_fraction: Fraction of dims for iteration RoPE
use_recursion_checkpointing: Enable gradient checkpointing for iterations
soft_embedding_method: How to convert logits to soft embeddings
soft_embedding_ema_step: EMA step size (1.0 = no EMA, <1.0 = blend with previous)
flow_matching_enabled: Enable CFM-inspired flow matching framework
flow_matching_lambda: Weight of distillation KL loss relative to CE
flow_matching_t_distribution: Time sampling distribution ("logit_normal" or "uniform")
flow_matching_t_logit_mean: Mean of logit-normal distribution
flow_matching_t_logit_std: Std of logit-normal distribution
flow_matching_t_min: Minimum time value (clamp)
flow_matching_t_max: Maximum time value (clamp)
flow_matching_mask_scale: Scale mask_emb by (1-t) if True, binary if False
"""
config = RecursiveMLMConfig.from_base_model_config(
base_model.config,
num_recursions=num_recursions,
normalization=normalization,
loss_weight=loss_weight,
mask_token_id=mask_token_id,
temperature=temperature,
gradient_steps=gradient_steps,
schedule=schedule,
causal_strength=causal_strength,
temperature_max=temperature_max,
entropy_target_max=entropy_target_max,
entropy_floor_max=entropy_floor_max,
smear_sigma_max=smear_sigma_max,
noise_std_max=noise_std_max,
iteration_rope_dim_fraction=iteration_rope_dim_fraction,
use_recursion_checkpointing=use_recursion_checkpointing,
soft_embedding_method=soft_embedding_method,
soft_embedding_ema_step=soft_embedding_ema_step,
flow_matching_enabled=flow_matching_enabled,
flow_matching_lambda=flow_matching_lambda,
flow_matching_t_distribution=flow_matching_t_distribution,
flow_matching_t_logit_mean=flow_matching_t_logit_mean,
flow_matching_t_logit_std=flow_matching_t_logit_std,
flow_matching_t_min=flow_matching_t_min,
flow_matching_t_max=flow_matching_t_max,
flow_matching_mask_scale=flow_matching_mask_scale,
)
return cls(config, base_model=base_model)
@property
def embed_weight(self) -> torch.Tensor:
return self.mlm.get_input_embeddings().weight
def get_input_embeddings(self):
return self.mlm.get_input_embeddings()
def set_input_embeddings(self, value):
self.mlm.set_input_embeddings(value)
def get_output_embeddings(self):
return self.mlm.get_output_embeddings()
def set_output_embeddings(self, new_embeddings):
self.mlm.set_output_embeddings(new_embeddings)
def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
"""Enable gradient checkpointing with correct settings for recursion.
Forces use_reentrant=False which is required for:
- Nested checkpoint calls (base model + recursion checkpointing)
- Models with frozen parameters
- Complex gradient flows through soft embeddings
"""
if gradient_checkpointing_kwargs is None:
gradient_checkpointing_kwargs = {}
# Force use_reentrant=False for nested checkpointing compatibility
gradient_checkpointing_kwargs.setdefault("use_reentrant", False)
self.mlm.gradient_checkpointing_enable(gradient_checkpointing_kwargs)
def gradient_checkpointing_disable(self):
"""Disable gradient checkpointing in the underlying MLM."""
self.mlm.gradient_checkpointing_disable()
def _single_iteration_checkpointable(
self,
soft_embeds: torch.Tensor,
base_embeds: torch.Tensor,
mask_pos: torch.Tensor,
attention_mask: torch.Tensor,
embed_weight: torch.Tensor,
mask_emb: torch.Tensor,
temperature: torch.Tensor,
position_ids: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Single differentiable iteration for checkpointing.
This method performs one iteration of recursive refinement in a way that
maintains gradient flow and is compatible with torch.utils.checkpoint.
Args:
soft_embeds: (B, L, H) - current soft embeddings
base_embeds: (B, L, H) - original token embeddings
mask_pos: (B, L) bool - which positions are masked
attention_mask: (B, L) - attention mask for MLM
embed_weight: (V, H) - embedding weight matrix
mask_emb: (H,) - mask token embedding
temperature: scalar tensor - softmax temperature
Returns:
logits: (B, L, V) - output logits from this iteration
next_soft_embeds: (B, L, H) - soft embeddings for next iteration
"""
# Blend: use soft_embeds at masked positions, base_embeds elsewhere
inputs_embeds = torch.where(mask_pos.unsqueeze(-1), soft_embeds, base_embeds)
# Forward through base MLM
outputs = self.mlm(
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
position_ids=position_ids,
return_dict=True,
)
logits = outputs.logits
# Compute soft embeddings for next iteration (DIFFERENTIABLE - no detach!)
next_soft_embeds = base_embeds.clone()
if mask_pos.any():
masked_logits = logits[mask_pos] # (num_masked, V)
# Convert logits to mixing weights based on soft_embedding_method
if self.config.soft_embedding_method == "none":
# No normalization - use raw logits directly
weights = masked_logits # Differentiable!
elif self.config.soft_embedding_method == "l2_normalize":
# L2 normalize logits - removes softmax bottleneck for smoother gradients
weights = F.normalize(masked_logits, p=2, dim=-1) # Differentiable!
else:
# Default: softmax normalization
weights = F.softmax(masked_logits / temperature, dim=-1) # Differentiable!
soft_emb = weights @ embed_weight + mask_emb # Differentiable!
# Apply EMA blending with previous soft embeddings if enabled
ema_step = self.config.soft_embedding_ema_step
if ema_step < 1.0:
prev_soft_emb = soft_embeds[mask_pos] # Previous iteration's soft embeddings
soft_emb = (1.0 - ema_step) * prev_soft_emb + ema_step * soft_emb
next_soft_embeds[mask_pos] = soft_emb
return logits, next_soft_embeds
def _stable_softmax(self, logits: torch.Tensor, T: float = 1.0, dim: int = -1, eps: float = 1e-12) -> torch.Tensor:
"""Numerically stable softmax with temperature T > 0."""
z = logits / max(T, eps)
z = z - z.max(dim=dim, keepdim=True).values # subtract max
z = torch.exp(z) # safe since z <= 0
z_sum = z.sum(dim=dim, keepdim=True)
return z / z_sum.clamp(min=eps)
def normalize(self, logits: torch.Tensor) -> torch.Tensor:
"""Normalize logits -> mixing weights. Shape: (B, L, V) -> (B, L, V)"""
norm = self.config.normalization.lower()
T = self.config.temperature
V = logits.shape[-1]
if norm == "none":
return logits
if norm == "softmax":
return torch.softmax(logits / T, dim=-1)
if norm == "stable_softmax":
return self._stable_softmax(logits, T=T, dim=-1)
raise ValueError(f"Unknown normalization: {norm}")
def step_weight(self, t: int, T: int) -> float:
"""Loss weight for step t of T."""
lw = self.config.loss_weight
if lw == "linear":
return (t + 1) / T
if lw == "uniform":
return 1.0
if lw == "last_1":
return 1.0 if t == T - 1 else 0.0
if lw == "last_2":
return 1.0 if T - t <= 2 else 0.0
raise ValueError(f"Unknown loss_weight: {lw}")
# ==================== CONVERGENCE SCHEDULE SYSTEM ====================
#
# The core idea: control WHEN each position is allowed to converge.
#
# Schedule types:
# - "linear": All positions converge at the same rate
# - "causal": Early positions converge first, late positions last
#
# Effects (mechanisms to enforce the schedule):
# - temperature: Raise temperature for positions not yet allowed to converge
# - entropy_floor: Force minimum entropy
# - entropy_target: Force exact entropy via bisection (ARChitects-style)
# - smear: Spread probability across neighboring positions
# - noise: Add Gaussian noise to logits
#
# Each effect uses per-position "convergence progress" (0=uncertain, 1=can converge)
def _compute_convergence_progress(
self,
iteration: int,
total_iterations: int,
seq_length: int,
mask_positions: torch.Tensor,
schedule: str = "linear",
causal_strength: float = 1.0,
device: torch.device = None,
dtype: torch.dtype = None,
) -> torch.Tensor:
"""
Compute per-position convergence progress based on schedule.
Args:
iteration: Current iteration (0-indexed)
total_iterations: Total number of iterations
seq_length: Full sequence length L
mask_positions: Position indices of masked tokens (num_masked,)
schedule: "linear" or "causal"
causal_strength: How much faster early positions converge (for causal schedule)
Returns:
progress: (num_masked,) tensor with values in [0, 1]
0 = position should be maximally uncertain
1 = position is allowed to fully converge
"""
base_progress = iteration / max(total_iterations - 1, 1)
if schedule == "linear":
return torch.full(
(mask_positions.shape[0],),
base_progress,
device=device,
dtype=dtype
)
elif schedule == "causal":
position_factor = mask_positions.float() / max(seq_length - 1, 1)
effective_progress = base_progress * (1.0 + causal_strength * (1.0 - position_factor))
return effective_progress.clamp(0.0, 1.0)
else:
raise ValueError(f"Unknown schedule: {schedule}")
def _apply_temperature_effect(
self,
logits: torch.Tensor,
progress: torch.Tensor,
temperature_max: float,
) -> torch.Tensor:
"""
Apply per-position temperature scaling based on convergence progress.
Low progress = high temperature (uncertain), high progress = temperature 1.0.
"""
if temperature_max <= 0:
return logits
temperature = 1.0 + temperature_max * (1.0 - progress)
temperature = temperature.unsqueeze(-1)
return logits / temperature
def _apply_entropy_floor_effect(
self,
probs: torch.Tensor,
progress: torch.Tensor,
entropy_floor_max: float,
) -> torch.Tensor:
"""
Ensure minimum entropy based on convergence progress.
Low progress = high entropy floor, high progress = no floor.
NOTE: This is a ONE-SIDED constraint (floor only).
"""
if entropy_floor_max <= 0:
return probs
entropy_floor = entropy_floor_max * (1.0 - progress)
log_probs = torch.log(probs + 1e-10)
current_entropy = -(probs * log_probs).sum(dim=-1)
below_floor = current_entropy < entropy_floor
if not below_floor.any():
return probs
logits = torch.log(probs + 1e-10)
target_ratio = entropy_floor / (current_entropy + 1e-10)
temperature = torch.ones_like(current_entropy)
temperature[below_floor] = target_ratio[below_floor].clamp(1.0, 10.0)
scaled_probs = torch.softmax(logits / temperature.unsqueeze(-1), dim=-1)
result = probs.clone()
result[below_floor] = scaled_probs[below_floor]
return result
def _find_temperature_for_target_entropy(
self,
logits: torch.Tensor,
target_entropy: torch.Tensor,
tol: float = 1e-3,
max_iter: int = 32,
T_low: float = 1e-6,
T_high_init: float = 1.0,
max_T: float = 100.0,
) -> torch.Tensor:
"""
Find per-position temperatures that achieve exactly the target entropy.
Uses bisection search, adapted from ARChitects' implementation.
Args:
logits: Raw logits (num_positions, V)
target_entropy: Target entropy per position (num_positions,) or scalar
tol: Entropy tolerance for convergence
max_iter: Maximum bisection iterations
T_low: Minimum temperature (near-greedy)
T_high_init: Initial upper bound for search
max_T: Maximum allowed temperature
Returns:
temperatures: (num_positions,) temperatures that achieve target entropy
"""
N, V = logits.shape
device, dtype = logits.device, logits.dtype
H_max = torch.log(torch.tensor(V, device=device, dtype=dtype))
if target_entropy.dim() == 0:
target = target_entropy.expand(N).clone()
else:
target = target_entropy.clone()
target = target.clamp(0.0, H_max)
def compute_entropy(logits_: torch.Tensor, temps: torch.Tensor) -> torch.Tensor:
temps = temps.unsqueeze(-1).clamp(min=T_low)
scaled = logits_ / temps
scaled = scaled - scaled.max(dim=-1, keepdim=True).values
probs = torch.softmax(scaled, dim=-1)
log_probs = torch.log(probs + 1e-12)
return -(probs * log_probs).sum(dim=-1)
lo = torch.full((N,), T_low, device=device, dtype=dtype)
hi = torch.full((N,), T_high_init, device=device, dtype=dtype)
H_lo = compute_entropy(logits, lo)
done_low = target <= (H_lo + tol)
H_hi = compute_entropy(logits, hi)
needs_expansion = (H_hi < target - tol) & ~done_low
for _ in range(100):
if not needs_expansion.any():
break
hi[needs_expansion] = (hi[needs_expansion] * 2.0).clamp(max=max_T)
H_hi[needs_expansion] = compute_entropy(
logits[needs_expansion], hi[needs_expansion]
)
needs_expansion = (H_hi < target - tol) & ~done_low & (hi < max_T - 1e-6)
can_bisect = ~done_low & (H_hi >= target - tol)
for _ in range(max_iter):
if not can_bisect.any():
break
mid = (lo + hi) / 2.0
H_mid = compute_entropy(logits, mid)
too_low = (H_mid < target) & can_bisect
lo[too_low] = mid[too_low]
hi[~too_low & can_bisect] = mid[~too_low & can_bisect]
converged = (hi - lo) <= tol * mid.clamp(min=1.0)
can_bisect = can_bisect & ~converged
temps = torch.zeros(N, device=device, dtype=dtype)
temps[done_low] = T_low
temps[~done_low] = (lo[~done_low] + hi[~done_low]) / 2.0
return temps
def _apply_target_entropy_effect(
self,
logits: torch.Tensor,
progress: torch.Tensor,
entropy_target_max: float,
entropy_target_min: float = 0.0,
) -> torch.Tensor:
"""
Adjust temperature to achieve EXACTLY the target entropy per position.
This is a TWO-SIDED constraint: both raises and lowers entropy as needed.
Args:
logits: Raw logits (num_masked, V)
progress: Per-position convergence progress (num_masked,)
entropy_target_max: Target entropy at progress=0
entropy_target_min: Target entropy at progress=1 (usually ~0)
Returns:
probs: Probabilities with entropy matching targets
"""
if entropy_target_max <= 0:
return torch.softmax(logits, dim=-1)
target_entropy = entropy_target_max * (1.0 - progress) + entropy_target_min * progress
temps = self._find_temperature_for_target_entropy(logits, target_entropy)
temps = temps.unsqueeze(-1).clamp(min=1e-6)
return torch.softmax(logits / temps, dim=-1)
def _apply_smear_effect(
self,
probs: torch.Tensor,
mask_pos: torch.Tensor,
progress_full: torch.Tensor,
smear_sigma_max: float,
) -> torch.Tensor:
"""
Apply positional smearing with per-position sigma based on progress.
Low progress = high smearing, high progress = no smearing.
Note: This operates on full (B, L, V) tensor because smearing mixes across positions.
"""
if smear_sigma_max <= 0:
return probs
B, L, V = probs.shape
sigma_per_pos = smear_sigma_max * (1.0 - progress_full)
avg_sigma = sigma_per_pos[mask_pos].mean().item()
if avg_sigma < 0.1:
return probs
positions = torch.arange(L, device=probs.device, dtype=probs.dtype)
diff = positions.unsqueeze(0) - positions.unsqueeze(1)
kernel = torch.exp(-0.5 * (diff / avg_sigma) ** 2)
kernel = kernel / kernel.sum(dim=1, keepdim=True)
smeared = torch.einsum('ij,bjv->biv', kernel, probs)
smeared = smeared / smeared.sum(dim=-1, keepdim=True).clamp(min=1e-10)
blend = progress_full.unsqueeze(-1)
result = blend * probs + (1 - blend) * smeared
output = probs.clone()
output[mask_pos] = result[mask_pos]
return output
def _apply_noise_effect(
self,
logits: torch.Tensor,
progress: torch.Tensor,
noise_std_max: float,
) -> torch.Tensor:
"""
Add Gaussian noise to logits based on convergence progress.
Low progress = high noise, high progress = no noise.
"""
if noise_std_max <= 0:
return logits
noise_std = noise_std_max * (1.0 - progress)
noise_std = noise_std.unsqueeze(-1)
noise = torch.randn_like(logits) * noise_std
return logits + noise
def _apply_iteration_rope(
self,
embeds: torch.Tensor,
iteration: int,
total_iterations: int,
dim_fraction: float = 0.25,
base: float = 10000.0,
) -> torch.Tensor:
"""
Apply rotary embedding based on iteration progress.
Uses a subset of dimensions to avoid interfering with position RoPE.
"""
if dim_fraction <= 0:
return embeds
H = embeds.shape[-1]
rot_dim = int(H * dim_fraction)
rot_dim = rot_dim - (rot_dim % 2)
if rot_dim < 2:
return embeds
progress = iteration / max(total_iterations - 1, 1)
inv_freq = 1.0 / (base ** (torch.arange(0, rot_dim, 2, device=embeds.device, dtype=embeds.dtype) / rot_dim))
angles = progress * inv_freq * 3.14159
cos, sin = torch.cos(angles), torch.sin(angles)
if embeds.dim() == 2:
cos, sin = cos.unsqueeze(0), sin.unsqueeze(0)
elif embeds.dim() == 3:
cos = cos.unsqueeze(0).unsqueeze(0)
sin = sin.unsqueeze(0).unsqueeze(0)
embeds_out = embeds.clone()
x1, x2 = embeds[..., -rot_dim::2], embeds[..., -rot_dim+1::2]
embeds_out[..., -rot_dim::2] = x1 * cos - x2 * sin
embeds_out[..., -rot_dim+1::2] = x1 * sin + x2 * cos
return embeds_out
# ==================== FLOW MATCHING ====================
def _sample_flow_matching_t(self, num_tokens: int, device: torch.device) -> torch.Tensor:
"""Sample per-token time levels for flow matching.
Returns:
t: (num_tokens,) tensor of time levels in [t_min, t_max]
"""
dist = self.config.flow_matching_t_distribution
if dist == "logit_normal":
z = torch.randn(num_tokens, device=device)
z = z * self.config.flow_matching_t_logit_std + self.config.flow_matching_t_logit_mean
t = torch.sigmoid(z)
elif dist == "uniform":
t = torch.empty(num_tokens, device=device).uniform_(0, 1)
else:
raise ValueError(f"Unknown flow_matching_t_distribution: {dist}")
return t.clamp(self.config.flow_matching_t_min, self.config.flow_matching_t_max)
def compute_flow_matching_distillation_loss(
self,
input_ids: torch.Tensor,
teacher_logits: torch.Tensor,
labels: torch.Tensor,
flow_noise_embed: torch.Tensor,
flow_t: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
) -> SelfDistillationOutput:
"""
CFM flow matching distillation: teacher sees state at time t, student sees
noisier state at time s < t on the same interpolation path.
Both should predict the same endpoint (target token). The student must
learn to refine from noisier inputs by matching the teacher's predictions.
Args:
input_ids: Input with [MASK] tokens at positions to predict
teacher_logits: Logits from the forward pass (will be detached)
labels: Target tokens at masked positions (-100 elsewhere)
flow_noise_embed: (num_masked, H) noise embeddings from forward
flow_t: (num_masked,) per-token time levels from forward
attention_mask: Standard attention mask
position_ids: Position IDs (if needed by base model)
Returns:
SelfDistillationOutput with loss, logits, time gap, and diagnostics
"""
mask_id = self.config.mask_token_id
mask_pos = (input_ids == mask_id) # (B, L)
device = input_ids.device
num_masked = mask_pos.sum().item()
if num_masked == 0:
zero = torch.tensor(0.0, device=device, requires_grad=True)
dummy = torch.zeros(1, device=device)
return SelfDistillationOutput(zero, dummy, dummy, 0.0, 0.0, 0.0, 1.0)
teacher_logits = teacher_logits.detach()
embed_weight = self.embed_weight
mask_emb = embed_weight[mask_id] # (H,)
base_embeds = self.get_input_embeddings()(input_ids) # (B, L, H)
# Target embeddings from labels
target_ids = labels[mask_pos] # (num_masked,)
target_embed = embed_weight[target_ids] # (num_masked, H)
# Sample student time s ~ U(0, t) per token
s_per_token = flow_t * torch.rand(num_masked, device=device) # (num_masked,)
# Student state: same noise, earlier time (noisier)
s_col = s_per_token.unsqueeze(-1).to(base_embeds.dtype) # (num_masked, 1)
student_interp = (1 - s_col) * flow_noise_embed + s_col * target_embed
if self.config.flow_matching_mask_scale:
student_masked_embeds = student_interp + (1 - s_col) * mask_emb
else:
student_masked_embeds = student_interp + mask_emb
# Build full student input (detached — gradient only flows through student's forward)
student_embeds = base_embeds.detach().clone()
student_embeds[mask_pos] = student_masked_embeds.detach()
student_inputs = torch.where(
mask_pos.unsqueeze(-1), student_embeds, base_embeds.detach()
)
if attention_mask is None:
attention_mask = torch.ones_like(input_ids, dtype=base_embeds.dtype)
student_out = self.mlm(
inputs_embeds=student_inputs,
attention_mask=attention_mask,
position_ids=position_ids,
return_dict=True,
)
student_logits = student_out.logits # (B, L, V) — has gradient
# KL divergence loss on masked positions
t_logits = teacher_logits[mask_pos] # (num_masked, V)
s_logits = student_logits[mask_pos] # (num_masked, V)
teacher_probs = F.softmax(t_logits, dim=-1)
student_log_probs = F.log_softmax(s_logits, dim=-1)
kl_loss = F.kl_div(
student_log_probs,
teacher_probs,
reduction="batchmean",
)
# Diagnostic metrics
with torch.no_grad():
teacher_log_probs = torch.log(teacher_probs + 1e-10)
teacher_entropy = -(teacher_probs * teacher_log_probs).sum(dim=-1).mean().item()
student_probs = F.softmax(s_logits.detach(), dim=-1)
student_log_probs_det = torch.log(student_probs + 1e-10)
student_entropy = -(student_probs * student_log_probs_det).sum(dim=-1).mean().item()
agreement = (t_logits.argmax(dim=-1) == s_logits.detach().argmax(dim=-1)).float().mean().item()
mean_time_gap = (flow_t - s_per_token).mean().item()
return SelfDistillationOutput(
loss=kl_loss,
teacher_logits=teacher_logits,
student_logits=student_logits,
degradation_temperature=mean_time_gap,
teacher_entropy=teacher_entropy,
student_entropy=student_entropy,
agreement_rate=agreement,
)
# ==================== SELF-DISTILLATION (legacy) ====================
def compute_self_distillation_loss(
self,
input_ids: torch.Tensor,
teacher_logits: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
temperature_min: Optional[float] = None,
temperature_max: Optional[float] = None,
temperature_distribution: Optional[str] = None,
) -> SelfDistillationOutput:
"""
CFM-style self-distillation: model's predictions should be consistent
across different levels of input degradation.
Process:
1. Take teacher logits (from standard forward pass, DETACHED)
2. Degrade: per-token random temperature → softer soft embeddings
3. Student: forward pass from degraded embeddings → logits (has grad)
4. Loss: KL(teacher || student) on masked positions
Each masked token gets its own independently sampled degradation
temperature, creating varied difficulty across the sequence.
Args:
input_ids: Input with [MASK] tokens at positions to predict
teacher_logits: Pre-computed teacher logits (will be detached).
Typically outputs.all_logits[0] or outputs.logits from standard forward.
attention_mask: Standard attention mask
position_ids: Position IDs (if needed by base model)
temperature_min: Min degradation temperature (default: config value)
temperature_max: Max degradation temperature (default: config value)
temperature_distribution: How to sample T (default: config value)
Returns:
SelfDistillationOutput with loss, logits, temperature, and diagnostics
"""
# Resolve defaults from config
temperature_min = temperature_min if temperature_min is not None else self.config.self_distillation_temperature_min
temperature_max = temperature_max if temperature_max is not None else self.config.self_distillation_temperature_max
temperature_distribution = temperature_distribution if temperature_distribution is not None else self.config.self_distillation_temperature_distribution
mask_id = self.config.mask_token_id
mask_pos = (input_ids == mask_id) # (B, L)
device = input_ids.device
num_masked = mask_pos.sum().item()
# Handle degenerate case: no masked positions
if num_masked == 0:
zero = torch.tensor(0.0, device=device, requires_grad=True)
dummy = torch.zeros(1, device=device)
return SelfDistillationOutput(zero, dummy, dummy, 1.0, 0.0, 0.0, 1.0)
# Ensure teacher logits are detached
teacher_logits = teacher_logits.detach()
embed_weight = self.embed_weight
mask_emb = embed_weight[mask_id] # (H,)
base_embeds = self.get_input_embeddings()(input_ids) # (B, L, H)
# ===== STEP 1: Sample per-token degradation temperatures =====
# Each masked position gets its own temperature independently
if temperature_distribution == "log_uniform":
log_min = torch.tensor(temperature_min, device=device).log()
log_max = torch.tensor(temperature_max, device=device).log()
log_T = torch.empty(num_masked, device=device).uniform_(log_min.item(), log_max.item())
T_per_token = log_T.exp() # (num_masked,)
elif temperature_distribution == "uniform":
T_per_token = torch.empty(num_masked, device=device).uniform_(
temperature_min, temperature_max
) # (num_masked,)
else:
raise ValueError(f"Unknown temperature distribution: {temperature_distribution}")
T_mean = T_per_token.mean().item()
# ===== STEP 2: Create degraded soft embeddings =====
# Per-token temperature scaling: each position gets its own T
masked_teacher_logits = teacher_logits[mask_pos] # (num_masked, V)
degraded_probs = F.softmax(masked_teacher_logits / T_per_token.unsqueeze(-1), dim=-1).to(embed_weight.dtype)
degraded_soft = degraded_probs @ embed_weight + mask_emb
degraded_soft_embeds = base_embeds.clone()
degraded_soft_embeds[mask_pos] = degraded_soft
degraded_soft_embeds = degraded_soft_embeds.detach()
# ===== STEP 3: Student forward from degraded input =====
student_inputs = torch.where(
mask_pos.unsqueeze(-1), degraded_soft_embeds, base_embeds.detach()
)
if attention_mask is None:
attention_mask = torch.ones_like(input_ids, dtype=base_embeds.dtype)
student_out = self.mlm(
inputs_embeds=student_inputs,
attention_mask=attention_mask,
position_ids=position_ids,
return_dict=True,
)
student_logits = student_out.logits # (B, L, V) — has gradient!
# ===== STEP 4: KL divergence loss on masked positions =====
t_logits = teacher_logits[mask_pos] # (num_masked, V)
s_logits = student_logits[mask_pos] # (num_masked, V)
teacher_probs = F.softmax(t_logits, dim=-1)
student_log_probs = F.log_softmax(s_logits, dim=-1)
# KL(teacher || student) = sum teacher * (log_teacher - log_student)
kl_loss = F.kl_div(
student_log_probs,
teacher_probs,
reduction="batchmean",
)
# ===== STEP 5: Compute diagnostic metrics =====
with torch.no_grad():
teacher_log_probs = torch.log(teacher_probs + 1e-10)
teacher_entropy = -(teacher_probs * teacher_log_probs).sum(dim=-1).mean().item()
student_probs = F.softmax(s_logits.detach(), dim=-1)
student_log_probs_det = torch.log(student_probs + 1e-10)
student_entropy = -(student_probs * student_log_probs_det).sum(dim=-1).mean().item()
agreement = (t_logits.argmax(dim=-1) == s_logits.detach().argmax(dim=-1)).float().mean().item()
return SelfDistillationOutput(
loss=kl_loss,
teacher_logits=teacher_logits,
student_logits=student_logits,
degradation_temperature=T_mean,
teacher_entropy=teacher_entropy,
student_entropy=student_entropy,
agreement_rate=agreement,
)
# ==================== MAIN SOFT EMBEDDING COMPUTATION ====================
@torch.no_grad()
def _compute_next_soft_embeds(
self,
logits: torch.Tensor,
mask_pos: torch.Tensor,
base_embeds: torch.Tensor,
prev_soft_embeds: Optional[torch.Tensor] = None,
iteration: int = 0,
total_iterations: int = 1,
# === Schedule parameters (default to config values) ===
schedule: Optional[str] = None,
causal_strength: Optional[float] = None,
# === Effect parameters (default to config values) ===
temperature_max: Optional[float] = None,
entropy_target_max: Optional[float] = None,
entropy_floor_max: Optional[float] = None,
smear_sigma_max: Optional[float] = None,
noise_std_max: Optional[float] = None,
iteration_rope_dim_fraction: Optional[float] = None,
) -> torch.Tensor:
"""
Compute soft embeddings from logits for the next iteration.
This function implements a unified "convergence schedule" system that controls
when each position is allowed to converge to a confident prediction.
Schedule Types:
"linear": All positions converge at the same rate (iteration-based only)
"causal": Early positions converge first, late positions last
Effects (mechanisms to enforce the schedule):
temperature_max: High temperature = more uniform distribution (one-sided)
entropy_target_max: Force EXACT entropy via bisection search (two-sided, recommended)
entropy_floor_max: Force MINIMUM entropy (one-sided, only prevents too confident)
smear_sigma_max: Spread probability across neighboring positions
noise_std_max: Add Gaussian noise to logits
All parameters default to their config values if not specified.
Args:
logits: Output logits from current iteration (B, L, V)
mask_pos: Boolean mask indicating which positions are masked (B, L)
base_embeds: Base token embeddings for non-masked positions (B, L, H)
iteration: Current iteration index (0-indexed)
total_iterations: Total number of iterations
Returns:
Soft embeddings for next iteration (B, L, H)
"""
# Use config values as defaults
schedule = schedule if schedule is not None else self.config.schedule
causal_strength = causal_strength if causal_strength is not None else self.config.causal_strength
temperature_max = temperature_max if temperature_max is not None else self.config.temperature_max
entropy_target_max = entropy_target_max if entropy_target_max is not None else self.config.entropy_target_max
entropy_floor_max = entropy_floor_max if entropy_floor_max is not None else self.config.entropy_floor_max
smear_sigma_max = smear_sigma_max if smear_sigma_max is not None else self.config.smear_sigma_max
noise_std_max = noise_std_max if noise_std_max is not None else self.config.noise_std_max
iteration_rope_dim_fraction = iteration_rope_dim_fraction if iteration_rope_dim_fraction is not None else self.config.iteration_rope_dim_fraction
soft_embeds = base_embeds.clone()
if not mask_pos.any():
return soft_embeds.detach()
B, L, V = logits.shape
device, dtype = logits.device, logits.dtype
# Check if any effects are enabled
has_effects = (
temperature_max > 0 or
entropy_target_max > 0 or
entropy_floor_max > 0 or
smear_sigma_max > 0 or
noise_std_max > 0 or
iteration_rope_dim_fraction > 0
)
if not has_effects:
# Simple path: no convergence schedule effects
masked_logits = logits[mask_pos]
embed_weight = self.embed_weight
# Convert logits to mixing weights based on soft_embedding_method
if self.config.soft_embedding_method == "none":
weights = masked_logits
elif self.config.soft_embedding_method == "l2_normalize":
weights = F.normalize(masked_logits, p=2, dim=-1)
else:
weights = self.normalize(masked_logits)
masked_soft = weights @ embed_weight
mask_emb = embed_weight[self.config.mask_token_id]
masked_soft = masked_soft + mask_emb
# Apply EMA blending with previous soft embeddings if enabled
ema_step = self.config.soft_embedding_ema_step
if ema_step < 1.0 and prev_soft_embeds is not None:
prev_masked_soft = prev_soft_embeds[mask_pos]
masked_soft = (1.0 - ema_step) * prev_masked_soft + ema_step * masked_soft
soft_embeds[mask_pos] = masked_soft
return soft_embeds.detach()
# ========== STEP 1: Compute per-position convergence progress ==========
batch_indices, position_indices = torch.where(mask_pos)
progress = self._compute_convergence_progress(
iteration=iteration,
total_iterations=total_iterations,
seq_length=L,
mask_positions=position_indices,
schedule=schedule,
causal_strength=causal_strength,
device=device,
dtype=dtype,
)
# Compute full (B, L) progress for smearing if needed
if smear_sigma_max > 0:
all_positions = torch.arange(L, device=device, dtype=dtype)
progress_full = self._compute_convergence_progress(
iteration=iteration,
total_iterations=total_iterations,
seq_length=L,
mask_positions=all_positions,
schedule=schedule,
causal_strength=causal_strength,
device=device,
dtype=dtype,
)
progress_full = progress_full.unsqueeze(0).expand(B, -1)
# ========== STEP 2: Apply smearing (needs full tensor) ==========
full_probs = self.normalize(logits)
if smear_sigma_max > 0:
full_probs = self._apply_smear_effect(
full_probs, mask_pos, progress_full, smear_sigma_max
)
# ========== STEP 3: Extract masked positions ==========
masked_logits = logits[mask_pos]
masked_probs = full_probs[mask_pos]
# ========== STEP 4: Apply temperature effect (on logits) ==========
if temperature_max > 0 and entropy_target_max <= 0:
masked_logits = self._apply_temperature_effect(
masked_logits, progress, temperature_max
)
masked_probs = torch.softmax(masked_logits, dim=-1)
# ========== STEP 5: Apply noise effect (on logits) ==========
if noise_std_max > 0:
masked_logits_noisy = self._apply_noise_effect(
torch.log(masked_probs + 1e-10), progress, noise_std_max
)
masked_probs = torch.softmax(masked_logits_noisy, dim=-1)
# ========== STEP 6: Apply entropy control ==========
if entropy_target_max > 0:
masked_probs = self._apply_target_entropy_effect(
masked_logits, progress, entropy_target_max
)
elif entropy_floor_max > 0:
masked_probs = self._apply_entropy_floor_effect(
masked_probs, progress, entropy_floor_max
)
# ========== STEP 7: Compute soft embeddings ==========
embed_weight = self.embed_weight
# Convert to mixing weights based on soft_embedding_method
if self.config.soft_embedding_method == "none":
# No normalization - use raw logits directly
weights = masked_logits
elif self.config.soft_embedding_method == "l2_normalize":
# L2 normalize bypasses all the softmax-based effects above
weights = F.normalize(masked_logits, p=2, dim=-1)
else:
weights = masked_probs
masked_soft = weights @ embed_weight
mask_emb = embed_weight[self.config.mask_token_id]
masked_soft = masked_soft + mask_emb
# ========== STEP 8: Apply iteration RoPE ==========
if iteration_rope_dim_fraction > 0:
masked_soft = self._apply_iteration_rope(
masked_soft, iteration, total_iterations, iteration_rope_dim_fraction
)
# ========== STEP 8.5: Apply EMA blending ==========
ema_step = self.config.soft_embedding_ema_step
if ema_step < 1.0 and prev_soft_embeds is not None:
prev_masked_soft = prev_soft_embeds[mask_pos]
masked_soft = (1.0 - ema_step) * prev_masked_soft + ema_step * masked_soft
# ========== STEP 9: Place back and return ==========
soft_embeds[mask_pos] = masked_soft
return soft_embeds.detach()
@torch.no_grad()
def _compute_iteration_metrics(
self, logits: torch.Tensor, labels: torch.Tensor
) -> IterationMetrics:
"""
Compute token-level AND sequence-level metrics for a single iteration.
Returns scalars only - no large tensor storage.
Token-level metrics:
- accuracy: fraction of correct token predictions
- entropy: average entropy per token
- softmax_ce: cross-entropy loss per token
Sequence-level metrics:
- full_sequence_accuracy: fraction of sequences where ALL tokens are correct
- min_sequence_confidence: mean of minimum top-1 confidence per sequence
"""
B = logits.shape[0]
# Move to CPU to avoid GPU OOM - metrics are for monitoring only
logits = logits.detach().cpu().float() # float32 is sufficient for metrics
target_labels = labels.detach().cpu().contiguous()
mask = target_labels != -100
if mask.sum() == 0:
return IterationMetrics(
accuracy=0.0,
entropy=0.0,
softmax_ce=0.0,
full_sequence_accuracy=0.0,
min_sequence_confidence=0.0,
)
logits = logits.contiguous()
predictions = logits.argmax(dim=-1)
correct = (predictions == target_labels) & mask
# ===== TOKEN-LEVEL METRICS =====
# Token accuracy
accuracy = (correct.sum() / mask.sum()).item()
# Extract valid tokens for entropy/CE
valid_logits = logits[mask]
valid_labels = target_labels[mask]
# Entropy (using log_softmax for numerical stability)
log_probs = torch.nn.functional.log_softmax(valid_logits, dim=-1)
probs = torch.exp(log_probs)
entropy = -(probs * log_probs).sum(dim=-1).mean().item()
# Cross-entropy
softmax_ce = torch.nn.functional.cross_entropy(
valid_logits, valid_labels, reduction="mean"
).item()
# ===== SEQUENCE-LEVEL METRICS =====
# Check which sequences have valid tokens
sequences_with_tokens = mask.any(dim=1) # (B,)
num_valid_sequences = sequences_with_tokens.sum().item()
if num_valid_sequences == 0:
return IterationMetrics(
accuracy=accuracy,
entropy=entropy,
softmax_ce=softmax_ce,
full_sequence_accuracy=0.0,
min_sequence_confidence=0.0,
)
# Full sequence accuracy: all tokens in sequence must be correct
num_correct_per_seq = correct.sum(dim=1) # (B,)
num_tokens_per_seq = mask.sum(dim=1) # (B,)
all_correct = (num_correct_per_seq == num_tokens_per_seq) & sequences_with_tokens
full_seq_accuracy = (all_correct.sum() / num_valid_sequences).item()
# Min sequence confidence: minimum top-1 probability within each sequence
probs_full = torch.softmax(logits, dim=-1) # (B, L, V) - already float32
top1_confidence = probs_full.max(dim=-1).values # (B, L)
min_confidences = []
for i in range(B):
if sequences_with_tokens[i]:
seq_confidences = top1_confidence[i][mask[i]] # (num_tokens_in_seq,)
min_confidences.append(seq_confidences.min().item())
min_seq_conf = sum(min_confidences) / len(min_confidences) if min_confidences else 0.0
return IterationMetrics(
accuracy=accuracy,
entropy=entropy,
softmax_ce=softmax_ce,
full_sequence_accuracy=full_seq_accuracy,
min_sequence_confidence=min_seq_conf,
)
def _single_iteration(
self,
t: int,
T: int,
soft_embeds: torch.Tensor,
base_embeds: torch.Tensor,
mask_pos: torch.Tensor,
attention_mask: Optional[torch.Tensor],
labels: Optional[torch.Tensor],
compute_metrics: bool,
position_ids: Optional[torch.Tensor] = None,
**kwargs,
) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[IterationMetrics]]:
"""
Execute a single iteration of recursive refinement.
Args:
t: Current iteration index (0 to T-1)
T: Total number of iterations
soft_embeds: Soft embeddings for mask positions
base_embeds: Base token embeddings from input_ids
mask_pos: Boolean mask of [MASK] positions (B, L)
attention_mask: Attention mask for MLM
labels: Target labels for loss computation
compute_metrics: Whether to compute iteration metrics
Returns:
logits: Output logits from MLM (B, L, V)
weighted_loss: Loss weighted by step_weight(t, T), or None if no labels
metrics: IterationMetrics, or None if not requested
"""
# Blend soft embeddings (at mask positions) with base embeddings (at non-mask positions)
inputs_embeds = torch.where(mask_pos.unsqueeze(-1), soft_embeds, base_embeds)
# Forward through base MLM
outputs = self.mlm(
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
position_ids=position_ids,
labels=labels,
return_dict=True,
**kwargs,
)
# Compute weighted loss for this iteration
weighted_loss = outputs.loss
if labels is not None:
if weighted_loss is None:
# Base model doesn't compute loss (e.g., LLaDA) - compute it ourselves
# Only compute loss on MASKED positions (MDLM training)
masked_logits = outputs.logits[mask_pos] # (num_masked, V)
masked_labels = labels[mask_pos] # (num_masked,)
loss_fct = CrossEntropyLoss() # -100 index = padding token
weighted_loss = loss_fct(masked_logits, masked_labels)
weighted_loss *= self.step_weight(t, T)
# Compute iteration metrics if requested
metrics = None
if compute_metrics and labels is not None:
metrics = self._compute_iteration_metrics(outputs.logits, labels)
return outputs.logits, weighted_loss, metrics
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
num_recursions: Optional[int] = None,
compute_iteration_metrics: bool = False,
use_recursion_checkpointing: Optional[bool] = None,
# Parameters for single-iteration training mode (DEPRECATED)
prev_soft_embeds: Optional[torch.Tensor] = None,
run_set_iteration: Optional[int] = None,
# === Convergence schedule parameters (None = use config defaults) ===
schedule: Optional[str] = None,
causal_strength: Optional[float] = None,
# === Effect parameters (None = use config defaults) ===
temperature_max: Optional[float] = None,
entropy_target_max: Optional[float] = None,
entropy_floor_max: Optional[float] = None,
smear_sigma_max: Optional[float] = None,
noise_std_max: Optional[float] = None,
iteration_rope_dim_fraction: Optional[float] = None,
**kwargs,
) -> RecursiveMaskedLMOutput:
"""
Forward with recursive refinement.
Supports three modes:
1. Checkpointed mode (default): Run all T recursions with gradient checkpointing.
Gradients flow through the entire chain; activations recomputed during backward.
2. Non-checkpointed mode (use_recursion_checkpointing=False): Store all activations.
Faster backward but higher memory.
3. Single-iteration mode (DEPRECATED - run_set_iteration is not None): Run only one
iteration. Use use_recursion_checkpointing=True instead.
Loss Weighting (config.loss_weight):
"last_1": Only final iteration loss (enables learning convergence behavior)
"last_2": Last 2 iterations
"linear": All iterations, linearly weighted (default)
"uniform": All iterations, uniformly weighted
Recursion Checkpointing:
use_recursion_checkpointing: Enable gradient checkpointing for iterations.
True = checkpoint each iteration, recompute during backward (default).
False = store all activations (higher memory, faster backward).
Convergence Schedule Parameters:
All schedule/effect parameters default to their config values if not specified.
Pass explicit values to override config for this forward pass.
schedule: "linear" or "causal" - controls when positions can converge
causal_strength: How much faster early positions converge (causal only)
temperature_max: Max temperature boost for uncertain positions
entropy_target_max: Target entropy at progress=0 (two-sided, recommended)
entropy_floor_max: Min entropy floor (one-sided)
smear_sigma_max: Max Gaussian sigma for position smearing
noise_std_max: Max std of Gaussian noise on logits
iteration_rope_dim_fraction: Fraction of dims for iteration RoPE
"""
B, L = input_ids.shape
V = self.embed_weight.shape[0]
mask_id = self.config.mask_token_id
if mask_id is None:
raise ValueError("mask_token_id must be set")
# Resolve config default for recursion checkpointing
use_recursion_checkpointing = (
use_recursion_checkpointing
if use_recursion_checkpointing is not None
else self.config.use_recursion_checkpointing
)
mask_pos = (input_ids == mask_id) # (B, L)
base_embeds = self.get_input_embeddings()(input_ids) # (B, L, H)
T = num_recursions or self.config.num_recursions
weight_sum = sum(self.step_weight(i, T) for i in range(T))
# Bundle schedule kwargs to pass to _compute_next_soft_embeds
schedule_kwargs = dict(
schedule=schedule,
causal_strength=causal_strength,
temperature_max=temperature_max,
entropy_target_max=entropy_target_max,
entropy_floor_max=entropy_floor_max,
smear_sigma_max=smear_sigma_max,
noise_std_max=noise_std_max,
iteration_rope_dim_fraction=iteration_rope_dim_fraction,
)
# ===== SINGLE ITERATION MODE (DEPRECATED) =====
if run_set_iteration is not None:
warnings.warn(
"run_set_iteration is deprecated. Use use_recursion_checkpointing=True instead, "
"which provides proper gradient flow through all iterations.",
DeprecationWarning,
stacklevel=2,
)
t = run_set_iteration
# Get soft embeddings for this iteration
if t == 0:
# t=0: Uniform prior = average embedding (equivalent to softmax(zeros) @ embed_weight)
# We compute this efficiently via embed_weight.mean() rather than creating large zero tensors
soft_embeds = base_embeds.clone()
if mask_pos.any():
avg_embed = self.embed_weight.mean(dim=0) # (H,) - mean over all V tokens
mask_emb = self.embed_weight[mask_id]
soft_embeds[mask_pos] = avg_embed + mask_emb
else:
if prev_soft_embeds is None:
raise ValueError(f"prev_soft_embeds must be provided for iteration {t}")
soft_embeds = prev_soft_embeds
logits, weighted_loss, metrics = self._single_iteration(
t, T, soft_embeds, base_embeds, mask_pos,
attention_mask, labels, compute_iteration_metrics,
position_ids=position_ids, **kwargs
)
# Normalize loss by total weight sum
loss = weighted_loss / weight_sum if weighted_loss is not None else None
# Compute soft embeddings for next iteration (if not last)
next_soft_embeds = None
if t < T - 1:
next_soft_embeds = self._compute_next_soft_embeds(
logits, mask_pos, base_embeds,
iteration=t,
total_iterations=T,
**schedule_kwargs,
)
return RecursiveMaskedLMOutput(
loss=loss,
logits=logits,
next_soft_embeds=next_soft_embeds,
iteration_metrics={t: metrics} if metrics is not None else None,
)
# ===== CHECKPOINTED MODE (gradient flow through all iterations) =====
embed_weight = self.embed_weight
mask_emb = embed_weight[mask_id] # (H,)
# Temperature must be a tensor for checkpointing (checkpoint requires tensor inputs)
temperature = torch.tensor(
self.config.temperature,
device=input_ids.device,
dtype=base_embeds.dtype,
)
# Ensure attention_mask is a tensor (required for checkpointing)
if attention_mask is None:
attention_mask = torch.ones(B, L, device=input_ids.device, dtype=base_embeds.dtype)
# Initialize soft embeddings for masked positions
soft_embeds = base_embeds.clone()
flow_noise_embed = None
flow_t_per_token = None
if self.config.flow_matching_enabled and self.training and labels is not None and mask_pos.any():
# Flow matching: interpolate between random noise and target on the simplex
num_masked = mask_pos.sum().item()
V = embed_weight.shape[0]
device = input_ids.device
# Sample per-token time levels (logit-normal by default)
flow_t_per_token = self._sample_flow_matching_t(num_masked, device)
# Random noise embedding: sample on simplex, project to H-dim
z = torch.randn(num_masked, V, device=device, dtype=base_embeds.dtype)
p_noise = F.softmax(z * self.config.flow_matching_noise_scale, dim=-1).to(base_embeds.dtype)
flow_noise_embed = p_noise @ embed_weight # (num_masked, H) — compact
# Target embedding from labels
target_ids = labels[mask_pos] # original token IDs at masked positions
target_embed = embed_weight[target_ids] # (num_masked, H)
# Interpolate in embedding space
t_col = flow_t_per_token.unsqueeze(-1).to(base_embeds.dtype) # (num_masked, 1)
interp_embed = (1 - t_col) * flow_noise_embed + t_col * target_embed
# Add mask signal (binary or scaled)
if self.config.flow_matching_mask_scale:
soft_embeds[mask_pos] = interp_embed + (1 - t_col) * mask_emb
else:
soft_embeds[mask_pos] = interp_embed + mask_emb
elif mask_pos.any():
# Standard uniform prior (average embedding + mask signal)
avg_embed = embed_weight.mean(dim=0) # (H,)
soft_embeds[mask_pos] = avg_embed + mask_emb
iteration_metrics = {} if compute_iteration_metrics and labels is not None else None
# Main recursion loop with optional checkpointing
all_logits = []
for t in range(T):
if self.training and use_recursion_checkpointing:
# Use checkpointing: activations recomputed during backward
# This maintains gradient flow while saving memory
logits, soft_embeds = torch_checkpoint(
self._single_iteration_checkpointable,
soft_embeds,
base_embeds,
mask_pos,
attention_mask,
embed_weight,
mask_emb,
temperature,
position_ids,
use_reentrant=False, # Critical for nested checkpointing!
)
else:
# No checkpointing: store all activations (inference or explicit disable)
logits, soft_embeds = self._single_iteration_checkpointable(
soft_embeds,
base_embeds,
mask_pos,
attention_mask,
embed_weight,
mask_emb,
temperature,
position_ids,
)
all_logits.append(logits)
# Compute iteration metrics if requested (no grad needed)
if iteration_metrics is not None and labels is not None:
with torch.no_grad():
iteration_metrics[t] = self._compute_iteration_metrics(logits, labels)
# Return all logits for trainer to compute loss with proper normalization
# Trainer handles: timestep-based weighting, iteration weighting, batch/sequence/token normalization
return RecursiveMaskedLMOutput(
loss=None, # Let trainer compute loss
logits=logits, # Final logits for inference/metrics
all_logits=all_logits if self.training else None, # Only needed during training
iteration_metrics=iteration_metrics or None,
flow_noise_embed=flow_noise_embed, # For flow matching distillation
flow_t=flow_t_per_token, # For flow matching distillation
)
@torch.no_grad()
def _generate_flow_map(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor],
position_ids: Optional[torch.Tensor],
num_steps: int,
) -> torch.Tensor:
"""Fill in mask positions using the CFM flow map update rule.
Starts from a random point on the probability simplex and iteratively
moves toward the model's predictions using the flow map step rule.
Args:
input_ids: Input with [MASK] tokens at positions to fill
attention_mask: Attention mask
position_ids: Position IDs
num_steps: Number of flow map steps (finer = better, 1 step = greedy)
Returns:
Tensor with [MASK] positions filled with predicted tokens
"""
mask_pos = (input_ids == self.config.mask_token_id)
num_masked = mask_pos.sum().item()
if num_masked == 0:
return input_ids.clone()
device = input_ids.device
V = self.embed_weight.shape[0]
embed_weight = self.embed_weight
mask_emb = embed_weight[self.config.mask_token_id]
base_embeds = self.get_input_embeddings()(input_ids)
# Start from random simplex point
noise_scale = self.config.flow_matching_noise_scale
p = F.softmax(torch.randn(num_masked, V, device=device, dtype=base_embeds.dtype) * noise_scale, dim=-1).to(base_embeds.dtype)
times = torch.linspace(0, 1, num_steps + 1, device=device)
for i in range(num_steps):
t_now = times[i]
t_next = times[i + 1]
step_size = (t_next - t_now) / (1 - t_now)
# Mask signal (binary or scaled)
if self.config.flow_matching_mask_scale:
mask_signal = (1 - t_now) * mask_emb
else:
mask_signal = mask_emb
# Project current state to embedding space
embed = p @ embed_weight + mask_signal
soft_embeds = base_embeds.clone()
soft_embeds[mask_pos] = embed
inputs_embeds = torch.where(mask_pos.unsqueeze(-1), soft_embeds, base_embeds)
outputs = self.mlm(
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
position_ids=position_ids,
return_dict=True,
)
pi = F.softmax(outputs.logits[mask_pos], dim=-1).to(p.dtype)
# Flow map update: move toward model's prediction
p = p + step_size * (pi - p)
# Fix floating point drift off the simplex
p = p.clamp(min=0)
p = p / p.sum(dim=-1, keepdim=True)
result = input_ids.clone()
result[mask_pos] = p.argmax(dim=-1)
return result
@torch.no_grad()
def generate(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
num_recursions: Optional[int] = None,
# === Convergence schedule parameters (None = use config defaults) ===
schedule: Optional[str] = None,
causal_strength: Optional[float] = None,
# === Effect parameters (None = use config defaults) ===
temperature_max: Optional[float] = None,
entropy_target_max: Optional[float] = None,
entropy_floor_max: Optional[float] = None,
smear_sigma_max: Optional[float] = None,
noise_std_max: Optional[float] = None,
iteration_rope_dim_fraction: Optional[float] = None,
) -> torch.Tensor:
"""Fill in mask positions via iterative refinement.
When flow_matching_enabled, uses the CFM flow map update rule.
Otherwise, uses standard recursive soft-token refinement.
Args:
input_ids: Input token IDs with [MASK] tokens at positions to fill
attention_mask: Attention mask
num_recursions: Override number of recursions/steps (default: config value)
schedule: "linear" or "causal" convergence schedule
causal_strength: How much faster early positions converge (causal only)
temperature_max: Max temperature boost for uncertain positions
entropy_target_max: Target entropy at progress=0 (two-sided)
entropy_floor_max: Min entropy floor (one-sided)
smear_sigma_max: Max Gaussian sigma for position smearing
noise_std_max: Max std of Gaussian noise on logits
iteration_rope_dim_fraction: Fraction of dims for iteration RoPE
Returns:
Tensor with [MASK] positions filled with predicted tokens
"""
num_steps = num_recursions or self.config.num_recursions
if self.config.flow_matching_enabled:
return self._generate_flow_map(
input_ids, attention_mask, position_ids, num_steps
)
out = self.forward(
input_ids,
attention_mask,
position_ids=position_ids,
num_recursions=num_steps,
schedule=schedule,
causal_strength=causal_strength,
temperature_max=temperature_max,
entropy_target_max=entropy_target_max,
entropy_floor_max=entropy_floor_max,
smear_sigma_max=smear_sigma_max,
noise_std_max=noise_std_max,
iteration_rope_dim_fraction=iteration_rope_dim_fraction,
)
result = input_ids.clone()
mask_pos = (input_ids == self.config.mask_token_id)
result[mask_pos] = out.logits.argmax(dim=-1)[mask_pos]
return result
|