File size: 106,325 Bytes
65f521e | 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 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 | """
2026.6.7
2026.6.9
5.5.0
1.7.0
__UNSLOTH_VERSIONING__
"""
# Unsloth auto generated code
# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
from torch import Tensor
import torch
import torch.nn as nn
from torch.nn import functional as F
from unsloth_zoo.temporary_patches.common import torch_compile
from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
from trl.trainer.sft_trainer import (Any, AutoProcessor, Callable, DataCollator, DataCollatorForLanguageModeling, DataCollatorForVisionLanguageModeling, Dataset, EvalPrediction, F, FLASH_ATTENTION_VARIANTS, IterableDataset, Path, PeftConfig, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, SFTConfig, SFTTrainer, TrainerCallback, TrainingArguments, Version, _BaseTrainer, _CHUNKED_LM_HEAD_CHUNK_SIZE, _chunk, _patch_chunked_ce_lm_head, apply_chat_template, clone_chat_template, contextlib, create_model_from_path, dataclass, defaultdict, dft_loss, get_act_offloading_ctx_manager, get_config_model_id, get_dataset_column_names, get_peft_model, get_training_chat_template, has_generation_markers, is_chat_template_stop_token_trained, is_conversational, is_peft_available, is_peft_model, logger, nn, os, pack_dataset, pad, peft, prepare_multimodal_messages, selective_log_softmax, torch, transformers, types, warnings, Any, AutoProcessor, Callable, DataCollator, DataCollatorForLanguageModeling, DataCollatorForVisionLanguageModeling, Dataset, EvalPrediction, F, FLASH_ATTENTION_VARIANTS, IterableDataset, PeftConfig, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, SFTConfig, SFTTrainer, TrainerCallback, TrainingArguments, Version, clone_chat_template, contextlib, create_model_from_path, defaultdict, dft_loss, get_act_offloading_ctx_manager, get_config_model_id, get_peft_model, get_training_chat_template, has_generation_markers, is_chat_template_stop_token_trained, is_conversational, is_peft_available, is_peft_model, logger, nn, os, pad, peft, torch, transformers, Callable, DataCollator, DataCollatorForLanguageModeling, Dataset, F, IterableDataset, apply_chat_template, is_conversational, os, pack_dataset, pad, transformers, warnings, F, PeftModel, PreTrainedModel, is_peft_available, logger, os, peft, torch, F, os)
import os
import math
import logging
from typing import *
from dataclasses import dataclass, field
from packaging.version import Version
import torch
import numpy as np
from contextlib import nullcontext
from torch.nn import functional as F
import inspect
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling as TransformersDataCollatorForLanguageModeling
from transformers.training_args import ParallelMode
from unsloth_zoo.device_type import DEVICE_TYPE, device_synchronize
# Wrap trainer with padding to right and enable training mode
import functools
from types import MethodType
try:
from unsloth_zoo.gradient_checkpointing import reset_unsloth_gradient_checkpointing_buffers
except:
def reset_unsloth_gradient_checkpointing_buffers(): pass
# Canonical reset lives in unsloth.models._utils so the SFT auto-packing wrapper and the plain
# Trainer loop can import the same helper; fall back to a no-op only if it can't be imported.
try:
from unsloth.models._utils import _unsloth_reset_stray_compile_cache
except Exception:
def _unsloth_reset_stray_compile_cache(self): pass
def prepare_for_training_mode(f):
@functools.wraps(f)
def wrapper(self, *args, **kwargs):
# Drop any torch.compile graph cache poisoned by a stray pre-train forward.
try:
_unsloth_reset_stray_compile_cache(self)
except Exception:
pass
# Finish the previous W&B run if this is a subsequent train() call.
# We do this at the START of train() (not the end) so that
# evaluate() / log() still work after train() completes.
# HF's WandbCallback.setup() will call wandb.init() for the new run.
# See: https://github.com/unslothai/unsloth/issues/3954
if getattr(self, '_unsloth_training_completed', False):
try:
import wandb
if wandb.run is not None:
wandb.finish()
# Reset HF's WandbCallback so it calls wandb.init() for the new run
for cb in self.callback_handler.callbacks:
if type(cb).__name__ == 'WandbCallback':
cb._initialized = False
break
except:
pass
# Enable training mode
_was_training = None
# Get gradient checkpointing setting from training arguments
use_gc = getattr(self.args, 'gradient_checkpointing', True)
if hasattr(self, 'model') and hasattr(self.model, "training"):
_was_training = self.model.training
if hasattr(self, 'model') and hasattr(self.model, "for_training"):
self.model.for_training(use_gradient_checkpointing=use_gc)
output = f(self, *args, **kwargs)
# Restore previous mode when possible
if hasattr(self, 'model') and hasattr(self.model, "for_inference"):
if _was_training is False:
self.model.for_inference()
elif _was_training is True and hasattr(self.model, "for_training"):
self.model.for_training(use_gradient_checkpointing=use_gc)
# Reset gradient checkpointing buffers to free memory while staying ready for next run
try:
reset_unsloth_gradient_checkpointing_buffers()
except:
pass
# Mark that training completed so the next train() call can
# finish this W&B run before starting a new one
self._unsloth_training_completed = True
return output
return wrapper
pass
torch_compile_options = {
"epilogue_fusion" : True,
"max_autotune" : False,
"shape_padding" : True,
"trace.enabled" : False,
"triton.cudagraphs" : False,
}
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def chunked_hidden_states_selective_log_softmax(
hidden_states: torch.Tensor,
lm_head: torch.Tensor,
index: torch.Tensor,
chunks: int = 4,
logit_scale_multiply: float = 0.0,
logit_scale_divide: float = 0.0,
logit_softcapping: float = 0.0,
temperature: float = 1.0,
) -> torch.Tensor:
# All Unsloth Zoo code licensed under AGPL3
flat_hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1])
flat_index = index.reshape(-1)
chunked_hidden_states = torch.chunk(flat_hidden_states, chunks=chunks, dim=0)
chunked_index = torch.chunk(flat_index, chunks=chunks, dim=0)
all_per_token_logps = []
for chunk_hidden_states, chunk_index in zip(chunked_hidden_states, chunked_index):
chunk_logits = chunk_hidden_states.to(lm_head.dtype) @ lm_head.t()
if logit_scale_multiply != 0.0:
chunk_logits = chunk_logits * logit_scale_multiply
if logit_scale_divide != 0.0:
chunk_logits = chunk_logits / logit_scale_divide
if logit_softcapping != 0.0:
chunk_logits = logit_softcapping * torch.tanh(chunk_logits / logit_softcapping)
chunk_logits = chunk_logits.to(torch.float32)
if temperature != 1.0:
chunk_logits = chunk_logits / temperature
selected_logits = torch.gather(chunk_logits, dim=-1, index=chunk_index.unsqueeze(-1)).squeeze(-1)
logsumexp_values = torch.logsumexp(chunk_logits, dim=-1)
per_token_logps = selected_logits - logsumexp_values
all_per_token_logps.append(per_token_logps)
all_per_token_logps = torch.concat(all_per_token_logps)
all_per_token_logps = all_per_token_logps.reshape((hidden_states.shape[0], hidden_states.shape[1]))
return all_per_token_logps
@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def chunked_selective_log_softmax(
logits,
index,
temperature: float = 1.0,
chunks: int = 4,
):
chunked_logits = torch.chunk(logits.reshape(-1, logits.shape[-1]), chunks = chunks, dim = 0)
chunked_index = torch.chunk(index.reshape(-1), chunks = chunks, dim = 0)
all_per_token_logps = []
# Per-chunk selective_log_softmax.
for chunk_logits, chunk_index in zip(chunked_logits, chunked_index):
chunk_logits = chunk_logits.to(torch.float32)
if temperature != 1.0:
chunk_logits = chunk_logits / temperature
selected_logits = torch.gather(chunk_logits, dim = -1, index = chunk_index.unsqueeze(-1)).squeeze(-1)
logsumexp_values = torch.logsumexp(chunk_logits, dim = -1)
per_token_logps = selected_logits - logsumexp_values
all_per_token_logps.append(per_token_logps)
pass
all_per_token_logps = torch.concat(all_per_token_logps)
all_per_token_logps = all_per_token_logps.reshape((logits.shape[0], logits.shape[1]))
return all_per_token_logps
def calculate_pad_tokens_in_prompt(
input_ids: torch.Tensor,
logits_to_keep: int,
pad_token_id: int
) -> torch.Tensor:
"""Count left-padded tokens per sequence, e.g. [pad, pad, pad, cat] -> 3."""
if logits_to_keep >= input_ids.shape[1]:
raise ValueError("logits_to_keep must be smaller than the sequence length.")
prompt_section = input_ids[:, :-logits_to_keep]
padding_mask = (prompt_section == pad_token_id)
pad_token_counts = padding_mask.sum(dim=1)
return pad_token_counts
def create_completion_attention_mask(
completion_input_ids: torch.Tensor,
left_pad_tokens_per_prompt: torch.Tensor,
max_left_pad: int,
pad_token_id: int
) -> torch.Tensor:
"""Build a completion mask that zeros leading prompt and trailing pad tokens.
For [p,p,p,c,c,c,pad,pad,pad] (p=sliced prompt, c=completion, pad=padding)
this returns [0,0,0,1,1,1,0,0,0].
"""
batch_size, completion_len = completion_input_ids.shape
device = completion_input_ids.device
num_tokens_to_mask = max_left_pad - left_pad_tokens_per_prompt
indices = torch.arange(completion_len, device=device).unsqueeze(0)
shift_mask = indices >= num_tokens_to_mask.unsqueeze(1)
non_padding_mask = (completion_input_ids != pad_token_id)
final_mask = shift_mask & non_padding_mask
return final_mask
def left_pack_padding(tensor: torch.Tensor, pad_id: int) -> torch.Tensor:
"""Move all padding tokens in each sequence to the right."""
mask = (tensor != pad_id)
# stable=True since the binary mask is unordered.
sorted_indices = torch.argsort(mask, dim=1, descending=True, stable=True)
packed_tensor = torch.gather(tensor, 1, sorted_indices)
return packed_tensor
def align_logprobs_with_mask(
logprob_tensor: torch.Tensor,
attention_mask: torch.Tensor,
pad_value: float = 0.0
) -> torch.Tensor:
"""Align a log probability tensor with a given attention mask."""
device = logprob_tensor.device
batch_size, logprob_seq_len = logprob_tensor.shape
mask_seq_len = attention_mask.shape[1]
padded_logprobs = torch.full(
attention_mask.shape,
fill_value=pad_value,
dtype=logprob_tensor.dtype,
device=device
)
left_pad_counts = torch.argmax(attention_mask, dim=1)
cols = torch.arange(logprob_seq_len, device=device)
dest_indices = left_pad_counts.unsqueeze(1) + cols
# Destination row indices, shape [batch_size, logprob_seq_len].
row_indices = torch.arange(batch_size, device=device).unsqueeze(1).expand_as(dest_indices)
# Keep only in-bounds destinations, then scatter via advanced indexing.
valid_mask = dest_indices < mask_seq_len
valid_rows = row_indices[valid_mask]
valid_cols = dest_indices[valid_mask]
valid_vals = logprob_tensor[valid_mask]
padded_logprobs[valid_rows, valid_cols] = valid_vals
return padded_logprobs
def align_completion_tool_mask(
tool_mask: torch.Tensor,
completion_mask: torch.Tensor,
) -> torch.Tensor:
"""Align a raw completion-length tool/env mask with Unsloth's repacked loss mask."""
if tool_mask is None:
return completion_mask
if tool_mask.shape[0] != completion_mask.shape[0]:
raise ValueError("tool_mask batch size must match completion_mask batch size.")
tool_mask = tool_mask.to(device=completion_mask.device)
if tool_mask.shape == completion_mask.shape:
aligned_tool_mask = tool_mask
else:
aligned_tool_mask = align_logprobs_with_mask(
tool_mask,
completion_mask,
pad_value=0,
)
return completion_mask * aligned_tool_mask.to(dtype=completion_mask.dtype)
def autotune_batch_and_chunks(
total_input_rows,
seq_len,
hidden_size,
vocab_size,
dtype_bytes=16,
multiplier=None
):
if multiplier is None:
final_m = max(4, seq_len // 4096)
else:
final_m = multiplier
if torch.cuda.is_available():
free_bytes, _ = torch.cuda.mem_get_info()
limit_gb = (free_bytes / (1024**3))*.80
elif hasattr(torch, "xpu") and torch.xpu.is_available():
# XPU: estimate free memory as total - reserved.
total_mem = torch.xpu.get_device_properties(0).total_memory
reserved_mem = torch.xpu.memory_reserved()
free_bytes = total_mem - reserved_mem
limit_gb = (free_bytes / (1024**3)) * 0.80
else:
# Fallback: assume 8GB available.
limit_gb = 8.0
bytes_to_gb = 1024**3
b_vals = torch.arange(total_input_rows, 0, -1, device='cpu', dtype=torch.float32)
hidden_gb = (b_vals * seq_len * hidden_size * dtype_bytes) / bytes_to_gb
base_logits = ((b_vals/total_input_rows) * b_vals * seq_len * vocab_size * dtype_bytes) / bytes_to_gb
logits_gb = base_logits / final_m
total_mem_gb = hidden_gb + logits_gb
valid_mask = total_mem_gb <= limit_gb
valid_indices = torch.nonzero(valid_mask, as_tuple=False)
if valid_indices.shape[0] == 0:
#This means your GPU will OOM
return 4, final_m
best_idx = valid_indices[0].item()
final_b = int(b_vals[best_idx].item())
return final_b, final_m
def sanitize_logprob(logprob):
"""Local port of trl.scripts.vllm_serve.sanitize_logprob.
Filters NaN logprobs from vLLM outputs."""
value = logprob.logprob
if math.isnan(value):
logging.getLogger(__name__).warning(
f"Generated NaN logprob, token logprob '{logprob}' will be ignored"
)
return None
return value
@dataclass
class UnslothSFTConfig(SFTConfig):
"""
Configuration class for the [`SFTTrainer`].
This class includes only the parameters that are specific to SFT training. For a full list of training arguments,
please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may
differ from those in [`~transformers.TrainingArguments`].
Using [`~transformers.HfArgumentParser`] we can turn this class into
[argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the
command line.
Parameters:
> Parameters that control the model
model_init_kwargs (`dict[str, Any]`, *optional*):
Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model`
argument of the [`SFTTrainer`] is provided as a string.
trust_remote_code (`bool`, *optional*, defaults to `False`):
Whether to allow loading models and tokenizers that ship custom Python code from the Hub. Forwarded to
[`~transformers.AutoModelForCausalLM.from_pretrained`] and
[`~transformers.AutoProcessor.from_pretrained`].
router_aux_loss_coef (`float`, *optional*, defaults to `0.001`):
Coefficient of the load-balancing auxiliary loss. Only has an effect when training a Mixture-of-Experts
(MoE) model; for other models it does nothing. The auxiliary loss is added to the training loss with this
weight. Set to `0.0` to disable it.
chat_template_path (`str`, *optional*):
If specified, sets the model's chat template. This can either be the path to a tokenizer (local directory
or Hugging Face Hub model) or a direct path to a Jinja template file. When using a Jinja file, you must
ensure that any special tokens referenced in the template are added to the tokenizer and that the model's
embedding layer is resized accordingly.
> Parameters that control the data preprocessing
dataset_text_field (`str`, *optional*, defaults to `"text"`):
Name of the column that contains text data in the dataset.
dataset_kwargs (`dict[str, Any]`, *optional*):
Dictionary of optional keyword arguments for the dataset preparation. The only supported key is
`skip_prepare_dataset`. When the model is a VLM, `skip_prepare_dataset` is automatically treated as `True`
regardless of the provided value, since preprocessing is done on the fly.
dataset_num_proc (`int`, *optional*):
Number of processes to use for processing the dataset.
eos_token (`str`, *optional*):
Token used to indicate the end of a turn or sequence. If `None`, it defaults to
`processing_class.eos_token`.
max_length (`int` or `None`, *optional*, defaults to `1024`):
Maximum length of the tokenized sequence. Sequences longer than `max_length` are truncated from the left
or right depending on `truncation_mode`. If `None`, no truncation is applied. When packing is enabled,
this value sets the sequence length.
truncation_mode (`str`, *optional*, defaults to `"keep_start"`):
Truncation mode to use when the sequence exceeds `max_length`. The only supported value is
`"keep_start"`. The `"keep_end"` value is deprecated and will be removed in v2.0.0.
shuffle_dataset (`bool`, *optional*, defaults to `False`):
Whether to shuffle the dataset.
packing (`bool`, *optional*, defaults to `False`):
Whether to group multiple sequences into fixed-length blocks to improve computational efficiency and reduce
padding. Uses `max_length` to define sequence length.
packing_strategy (`str`, *optional*, defaults to `"bfd"`):
Strategy for packing sequences. Can be `"bfd"` (best-fit decreasing, truncates overflow), `"bfd_split"`
(best-fit decreasing, splits overflow sequences), or `"wrapped"` (aggressive, cuts mid-sequence).
padding_free (`bool`, *optional*, defaults to `False`):
Whether to perform forward passes without padding by flattening all sequences in the batch into a single
continuous sequence. This reduces memory usage by eliminating padding overhead. Currently, this is only
supported with the FlashAttention 2 or 3, which can efficiently handle the flattened batch structure. When
packing is enabled with strategy `"bfd"`, padding-free is enabled, regardless of the value of this
parameter.
pad_to_multiple_of (`int`, *optional*):
If set, the sequences will be padded to a multiple of this value.
eval_packing (`bool`, *optional*):
Whether to pack the eval dataset. If `None`, uses the same value as `packing`.
> Parameters that control the training
completion_only_loss (`bool`, *optional*):
Whether to compute loss only on the completion part of the sequence. If set to `True`, loss is computed
only on the completion, which is supported only for [prompt-completion](#prompt-completion) datasets. If
`False`, loss is computed on the entire sequence. If `None` (default), the behavior depends on the dataset:
loss is computed on the completion for [prompt-completion](#prompt-completion) datasets, and on the full
sequence for [language modeling](#language-modeling) datasets.
assistant_only_loss (`bool`, *optional*, defaults to `False`):
Whether to compute loss only on the assistant part of the sequence. If set to `True`, loss is computed only
on the assistant responses, which is supported only for [conversational](#conversational) datasets. If
`False`, loss is computed on the entire sequence.
loss_type (`str`, *optional*, defaults to `"chunked_nll"`):
Type of loss to use. When left unset, it defaults to `"chunked_nll"`, except when `use_liger_kernel=True`,
in which case it defaults to `"nll"`. Possible values are:
- `"nll"`: standard negative log-likelihood.
- `"dft"`: Dynamic Fine-Tuning, as described in
[this paper](https://huggingface.co/papers/2508.05629).
- `"chunked_nll"`: same math as `"nll"`, but the `lm_head` projection is computed on non-ignored tokens
only (positions with `labels == -100` are dropped before the matmul) and the cross-entropy is processed
in chunks of tokens to reduce peak activation memory. Not compatible with `use_liger_kernel`.
activation_offloading (`bool`, *optional*, defaults to `False`):
Whether to offload the activations to the CPU.
> Deprecated parameters
pad_token:
<Deprecated version="1.1.0">
Parameter `pad_token` is deprecated and will be removed in version v2.0.0. Set `tokenizer.pad_token`
directly and pass it as `processing_class` to the trainer instead.
</Deprecated>
> [!NOTE]
> These parameters have default values different from [`~transformers.TrainingArguments`]:
> - `logging_steps`: Defaults to `10` instead of `500`.
> - `gradient_checkpointing`: Defaults to `True` instead of `False`.
> - `bf16`: Defaults to `True` if `fp16` is not set, instead of `False`.
> - `learning_rate`: Defaults to `2e-5` instead of `5e-5`.
"""
vllm_sampling_params: Optional[Any] = field(
default = None,
metadata = {'help': 'vLLM SamplingParams'},
)
unsloth_num_chunks : Optional[int] = field(
default = -1,
metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
)
unsloth_logit_chunk_multiplier : Optional[int] = field(
default = None,
metadata = {'help': 'Multiplier for chunked logit computations.'},
)
unsloth_grpo_mini_batch : Optional[int] = field(
default = None,
metadata = {'help': 'Mini batch size for GRPO hidden state accumulation. Default is None unless user defines it.'},
)
max_seq_length : Optional[int] = field(
default = None,
metadata = {'help': 'Maximum sequence length to truncate to.'},
)
def __init__(
self,
output_dir = None,
per_device_train_batch_size = 4,
num_train_epochs = 3.0,
max_steps = -1,
learning_rate = 5e-05,
lr_scheduler_type = 'linear',
lr_scheduler_kwargs = None,
warmup_steps = 0.1,
optim = 'adamw_8bit',
optim_args = None,
weight_decay = 0.001,
adam_beta1 = 0.9,
adam_beta2 = 0.999,
adam_epsilon = 1e-08,
optim_target_modules = None,
gradient_accumulation_steps = 2,
average_tokens_across_devices = True,
max_grad_norm = 1.0,
label_smoothing_factor = 0.0,
bf16 = False,
fp16 = False,
bf16_full_eval = False,
fp16_full_eval = False,
tf32 = None,
gradient_checkpointing = True,
gradient_checkpointing_kwargs = None,
torch_compile = False,
torch_compile_backend = None,
torch_compile_mode = None,
use_liger_kernel = False,
liger_kernel_config = None,
use_cache = False,
neftune_noise_alpha = None,
torch_empty_cache_steps = 250,
auto_find_batch_size = False,
logging_strategy = 'steps',
logging_steps = 1,
logging_first_step = False,
log_on_each_node = True,
logging_nan_inf_filter = False,
include_num_input_tokens_seen = False,
log_level = 'passive',
log_level_replica = 'warning',
disable_tqdm = None,
report_to = 'none',
run_name = None,
project = 'huggingface',
trackio_space_id = 'trackio',
eval_strategy = 'no',
eval_steps = None,
eval_delay = 0,
per_device_eval_batch_size = 4,
prediction_loss_only = False,
eval_on_start = False,
eval_do_concat_batches = True,
eval_use_gather_object = False,
eval_accumulation_steps = 2,
batch_eval_metrics = False,
save_only_model = False,
save_strategy = 'steps',
save_steps = 500,
save_on_each_node = False,
save_total_limit = None,
enable_jit_checkpoint = False,
push_to_hub = False,
hub_token = None,
hub_private_repo = None,
hub_model_id = None,
hub_strategy = 'every_save',
hub_always_push = False,
hub_revision = None,
load_best_model_at_end = False,
metric_for_best_model = None,
greater_is_better = None,
ignore_data_skip = False,
restore_callback_states_from_checkpoint = False,
full_determinism = False,
seed = 3407,
data_seed = 3407,
use_cpu = False,
accelerator_config = None,
parallelism_config = None,
dataloader_drop_last = False,
dataloader_num_workers = 0,
dataloader_pin_memory = True,
dataloader_persistent_workers = False,
dataloader_prefetch_factor = None,
remove_unused_columns = True,
label_names = None,
train_sampling_strategy = 'random',
length_column_name = 'length',
ddp_find_unused_parameters = None,
ddp_bucket_cap_mb = None,
ddp_broadcast_buffers = None,
ddp_backend = None,
ddp_timeout = 1800,
fsdp = None,
fsdp_config = None,
deepspeed = None,
debug = '',
skip_memory_metrics = True,
do_train = False,
do_eval = False,
do_predict = False,
resume_from_checkpoint = None,
warmup_ratio = None,
logging_dir = None,
local_rank = -1,
model_init_kwargs = None,
router_aux_loss_coef = 0.001,
trust_remote_code = False,
chat_template_path = None,
dataset_text_field = 'text',
dataset_kwargs = None,
dataset_num_proc = None,
eos_token = None,
max_length = 1024,
truncation_mode = 'keep_start',
shuffle_dataset = False,
packing = False,
packing_strategy = 'bfd',
padding_free = None,
pad_to_multiple_of = None,
eval_packing = None,
completion_only_loss = None,
assistant_only_loss = False,
loss_type = None,
activation_offloading = False,
pad_token = None,
vllm_sampling_params = None,
unsloth_num_chunks = -1,
unsloth_logit_chunk_multiplier = None,
unsloth_grpo_mini_batch = None,
max_seq_length = None,
**kwargs,
):
if learning_rate < 1e-7: print(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
if learning_rate > 1: print(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
if num_train_epochs is None:
num_train_epochs = 3.0 # Default to 3 epochs if None, max_steps will override
if output_dir is None and save_strategy == 'steps' and save_steps == 500:
output_dir = 'unsloth_training_checkpoints'
save_strategy = 'no'
import multiprocessing as _mp
if dataset_num_proc is None:
if _mp.get_start_method() != 'fork':
dataset_num_proc = None
else:
import psutil
dataset_num_proc = min(max((psutil.cpu_count() or 1)+4, 2), 64)
memory_gb_left = psutil.virtual_memory().available / (1024**3)
if memory_gb_left <= 2: dataset_num_proc = 1
else: dataset_num_proc = min(dataset_num_proc, int(memory_gb_left))
if os.environ.get('UNSLOTH_ENABLE_FLEX_ATTENTION', '0') == '1':
from unsloth_zoo.flex_attention import HAS_FLEX_ATTENTION
if HAS_FLEX_ATTENTION and pad_to_multiple_of is None:
from unsloth_zoo.flex_attention import FLEX_ATTENTION_BLOCK_SIZE
pad_to_multiple_of = FLEX_ATTENTION_BLOCK_SIZE
super().__init__(
output_dir = output_dir,
per_device_train_batch_size = per_device_train_batch_size,
num_train_epochs = num_train_epochs,
max_steps = max_steps,
learning_rate = learning_rate,
lr_scheduler_type = lr_scheduler_type,
lr_scheduler_kwargs = lr_scheduler_kwargs,
warmup_steps = warmup_steps,
optim = optim,
optim_args = optim_args,
weight_decay = weight_decay,
adam_beta1 = adam_beta1,
adam_beta2 = adam_beta2,
adam_epsilon = adam_epsilon,
optim_target_modules = optim_target_modules,
gradient_accumulation_steps = gradient_accumulation_steps,
average_tokens_across_devices = average_tokens_across_devices,
max_grad_norm = max_grad_norm,
label_smoothing_factor = label_smoothing_factor,
bf16 = bf16,
fp16 = fp16,
bf16_full_eval = bf16_full_eval,
fp16_full_eval = fp16_full_eval,
tf32 = tf32,
gradient_checkpointing = gradient_checkpointing,
gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
torch_compile = torch_compile,
torch_compile_backend = torch_compile_backend,
torch_compile_mode = torch_compile_mode,
use_liger_kernel = use_liger_kernel,
liger_kernel_config = liger_kernel_config,
use_cache = use_cache,
neftune_noise_alpha = neftune_noise_alpha,
torch_empty_cache_steps = torch_empty_cache_steps,
auto_find_batch_size = auto_find_batch_size,
logging_strategy = logging_strategy,
logging_steps = logging_steps,
logging_first_step = logging_first_step,
log_on_each_node = log_on_each_node,
logging_nan_inf_filter = logging_nan_inf_filter,
include_num_input_tokens_seen = include_num_input_tokens_seen,
log_level = log_level,
log_level_replica = log_level_replica,
disable_tqdm = disable_tqdm,
report_to = report_to,
run_name = run_name,
project = project,
trackio_space_id = trackio_space_id,
eval_strategy = eval_strategy,
eval_steps = eval_steps,
eval_delay = eval_delay,
per_device_eval_batch_size = per_device_eval_batch_size,
prediction_loss_only = prediction_loss_only,
eval_on_start = eval_on_start,
eval_do_concat_batches = eval_do_concat_batches,
eval_use_gather_object = eval_use_gather_object,
eval_accumulation_steps = eval_accumulation_steps,
batch_eval_metrics = batch_eval_metrics,
save_only_model = save_only_model,
save_strategy = save_strategy,
save_steps = save_steps,
save_on_each_node = save_on_each_node,
save_total_limit = save_total_limit,
enable_jit_checkpoint = enable_jit_checkpoint,
push_to_hub = push_to_hub,
hub_token = hub_token,
hub_private_repo = hub_private_repo,
hub_model_id = hub_model_id,
hub_strategy = hub_strategy,
hub_always_push = hub_always_push,
hub_revision = hub_revision,
load_best_model_at_end = load_best_model_at_end,
metric_for_best_model = metric_for_best_model,
greater_is_better = greater_is_better,
ignore_data_skip = ignore_data_skip,
restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
full_determinism = full_determinism,
seed = seed,
data_seed = data_seed,
use_cpu = use_cpu,
accelerator_config = accelerator_config,
parallelism_config = parallelism_config,
dataloader_drop_last = dataloader_drop_last,
dataloader_num_workers = dataloader_num_workers,
dataloader_pin_memory = dataloader_pin_memory,
dataloader_persistent_workers = dataloader_persistent_workers,
dataloader_prefetch_factor = dataloader_prefetch_factor,
remove_unused_columns = remove_unused_columns,
label_names = label_names,
train_sampling_strategy = train_sampling_strategy,
length_column_name = length_column_name,
ddp_find_unused_parameters = ddp_find_unused_parameters,
ddp_bucket_cap_mb = ddp_bucket_cap_mb,
ddp_broadcast_buffers = ddp_broadcast_buffers,
ddp_backend = ddp_backend,
ddp_timeout = ddp_timeout,
fsdp = fsdp,
fsdp_config = fsdp_config,
deepspeed = deepspeed,
debug = debug,
skip_memory_metrics = skip_memory_metrics,
do_train = do_train,
do_eval = do_eval,
do_predict = do_predict,
resume_from_checkpoint = resume_from_checkpoint,
warmup_ratio = warmup_ratio,
logging_dir = logging_dir,
local_rank = local_rank,
model_init_kwargs = model_init_kwargs,
router_aux_loss_coef = router_aux_loss_coef,
trust_remote_code = trust_remote_code,
chat_template_path = chat_template_path,
dataset_text_field = dataset_text_field,
dataset_kwargs = dataset_kwargs,
dataset_num_proc = dataset_num_proc,
eos_token = eos_token,
max_length = max_length,
truncation_mode = truncation_mode,
shuffle_dataset = shuffle_dataset,
packing = packing,
packing_strategy = packing_strategy,
padding_free = padding_free,
pad_to_multiple_of = pad_to_multiple_of,
eval_packing = eval_packing,
completion_only_loss = completion_only_loss,
assistant_only_loss = assistant_only_loss,
loss_type = loss_type,
activation_offloading = activation_offloading,
pad_token = pad_token,**kwargs)
self.vllm_sampling_params = vllm_sampling_params
self.unsloth_num_chunks = unsloth_num_chunks
if unsloth_grpo_mini_batch is not None:
if self.generation_batch_size >= unsloth_grpo_mini_batch:
self.unsloth_grpo_mini_batch = unsloth_grpo_mini_batch
else:
raise ValueError(
f"Unsloth GRPO mini batch size needs to be less than or equal to the effective generation batch size, "
f"which is self.per_device_train_batch_size * gradient_accumulation_steps."
)
self.unsloth_logit_chunk_multiplier = unsloth_logit_chunk_multiplier
self.max_seq_length = max_seq_length
# Unsloth: Remove use_reentrant=False forced by TRL 0.27.0+
if getattr(self, 'gradient_checkpointing_kwargs', None) is not None:
if 'use_reentrant' in self.gradient_checkpointing_kwargs:
del self.gradient_checkpointing_kwargs['use_reentrant']
pass
class _UnslothSFTTrainer(_BaseTrainer):
""""""
_tag_names = ["trl", "sft"]
_name = "SFT"
def __init__(
self,
model: "str | PreTrainedModel | PeftModel",
args: SFTConfig | TrainingArguments | None = None,
data_collator: DataCollator | None = None,
train_dataset: Dataset | IterableDataset | None = None,
eval_dataset: Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None = None,
processing_class: PreTrainedTokenizerBase | ProcessorMixin | None = None,
compute_loss_func: Callable | None = None,
compute_metrics: Callable[[EvalPrediction], dict] | None = None,
callbacks: list[TrainerCallback] | None = None,
optimizers: tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None] = (None, None),
optimizer_cls_and_kwargs: tuple[type[torch.optim.Optimizer], dict[str, Any]] | None = None,
preprocess_logits_for_metrics: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] | None = None,
peft_config: "PeftConfig | None" = None,
formatting_func: Callable[[dict], str] | None = None,
):
# Args
if args is None:
model_name = model if isinstance(model, str) else get_config_model_id(model.config)
model_name = model_name.split("/")[-1]
args = SFTConfig(f"{model_name}-SFT")
elif isinstance(args, TrainingArguments) and not isinstance(args, SFTConfig):
dict_args = args.to_dict()
dict_args["hub_token"] = args.hub_token # to_dict hides the hub_token
if Version(transformers.__version__) < Version("5.0.0"):
dict_args.pop("push_to_hub_token", None)
args = SFTConfig(**dict_args)
if train_dataset is None:
raise ValueError("`train_dataset` is required")
elif isinstance(train_dataset, IterableDataset):
# IterableDataset requires dispatch_batches=False because Accelerate's dispatch mode may try to concatenate
# batches from multiple processes, leading to mismatch errors.
if args.accelerator_config.dispatch_batches is True:
logger.warning(
"You are using an `IterableDataset` for training with `dispatch_batches=True`. `dispatch_batches` "
"is forced to `False` when using an `IterableDataset`. To remove this warning, unset "
"`dispatch_batches` in `SFTConfig` or set it to `False`."
)
args.accelerator_config.dispatch_batches = False
# Model
if isinstance(model, str):
model_init_kwargs = args.model_init_kwargs or {}
# Distributed training requires device_map=None ["auto" fails]
if args.distributed_state.distributed_type in ["MULTI_GPU", "DEEPSPEED"]:
model_init_kwargs["device_map"] = None
model_init_kwargs.setdefault("trust_remote_code", args.trust_remote_code)
model = create_model_from_path(model, **model_init_kwargs)
else:
if args.model_init_kwargs is not None:
logger.warning(
"You passed `model_init_kwargs` to the `SFTConfig`, but your model is already instantiated. "
"The `model_init_kwargs` will be ignored."
)
# Non-quantized models do not have the `is_loaded_in_{8,4}bit` attributes, whereas quantized models do
_is_quantized_model = getattr(model, "is_loaded_in_4bit", False) or getattr(model, "is_loaded_in_8bit", False)
# Processing class
if processing_class is None:
processing_class = AutoProcessor.from_pretrained(
get_config_model_id(model.config), trust_remote_code=args.trust_remote_code
)
# Handle pad token for processors or tokenizers
if isinstance(processing_class, ProcessorMixin):
self._tokenizer = processing_class.tokenizer
self._is_vlm = True
elif isinstance(processing_class, PreTrainedTokenizerBase):
self._tokenizer = processing_class
self._is_vlm = False
else:
raise TypeError("The `processing_class` must be either a `PreTrainedTokenizerBase` or a `ProcessorMixin`")
if args.eos_token is not None:
if args.eos_token not in self._tokenizer.get_vocab():
raise ValueError(
f"The specified `eos_token` ('{args.eos_token}') is not found in the vocabulary of the given "
f"`processing_class` ({processing_class.__class__.__name__}). Ensure that the `eos_token` exists "
"in the vocabulary before using it as an EOS token."
)
self._tokenizer.eos_token = args.eos_token
if args.chat_template_path is not None:
if os.path.isfile(args.chat_template_path) and args.chat_template_path.endswith((".jinja", ".j2")):
with open(args.chat_template_path, encoding="utf-8") as chat_template_file:
processing_class.chat_template = chat_template_file.read()
added_tokens = []
else:
model, processing_class, added_tokens = clone_chat_template(
model, processing_class, args.chat_template_path
)
else:
added_tokens = []
# Catch some wrong configurations related to VLMs
if self._is_vlm and args.packing:
raise ValueError(
"Packing is not supported for vision-language models. Please set `packing=False` in the SFTConfig."
)
if self._is_vlm and args.padding_free:
raise ValueError(
"Padding-free training is yet not supported for vision-language models. Please set "
"`padding_free=False` in the `SFTConfig`."
)
if self._is_vlm and args.assistant_only_loss:
raise ValueError(
"Assistant-only loss is not yet supported for vision-language models. Please set "
"`assistant_only_loss=False` in the `SFTConfig`."
)
if self._is_vlm and args.max_length is not None and args.truncation_mode == "keep_end":
raise ValueError(
"truncation_mode='keep_end' is not supported for vision-language models. Image tokens reside "
"inside the prompt portion of the sequence; depending on the example, keep_end may silently "
"drop them, causing pixel_values to be forwarded to the model with no corresponding visual "
"tokens in input_ids. Use truncation_mode='keep_start' (the default) or set max_length=None."
)
# PEFT
if False:
if not is_peft_available():
raise ImportError(
"You passed `peft_config` but the `peft` library is not installed. "
"Install it with `pip install trl[peft]`."
)
if not isinstance(peft_config, PeftConfig):
raise TypeError(
f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
f"got {type(peft_config).__name__}."
)
if is_peft_model(model):
raise ValueError(
"You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first merge "
"and unload the existing adapter, save the resulting base model, and then pass that base model along "
"with the new `peft_config` to the trainer."
)
if added_tokens:
# Ensure that the added tokens are trainable
if peft_config.trainable_token_indices is None:
peft_config.trainable_token_indices = {"embed_tokens": added_tokens}
elif "embed_tokens" not in peft_config.trainable_token_indices:
peft_config.trainable_token_indices["embed_tokens"] = added_tokens
else:
peft_config.trainable_token_indices["embed_tokens"].extend(added_tokens)
# Ensure that the lm_head is trainable
if peft_config.modules_to_save is None or "lm_head" not in peft_config.modules_to_save:
logger.warning(
"Cloning chat template added new tokens to the tokenizer, but 'lm_head' is not in PEFT's "
"`modules_to_save`. As a result, the model may not learn to generate outputs with these new "
"tokens, leading to degraded generation quality. To fix this, add "
"`modules_to_save=['lm_head']` to your PEFT configuration."
)
if peft_config.modules_to_save is None:
peft_config.modules_to_save = ["lm_head"]
else:
peft_config.modules_to_save.append("lm_head")
# Create PEFT model
# ZeRO-3 + PEFT for non-quantized models:
# - PEFT's default autocast_adapter_dtype=True upcasts LoRA adapter params to fp32 even when the base model is bf16.
# - ZeRO-3's _allgather_params_coalesced allocates output buffers using the dtype of the first persistent parameter,
# so mixed-dtype persistent_parameters [bf16 base + fp32 LoRA] cause a TypeError on the first optimizer step.
# - Passing autocast_adapter_dtype=False keeps adapter params in the base model dtype [bf16], fixing the mismatch.
# - This is safe: the fp32 upcast is a QLoRA-specific concern [low-bit quantized base models], not needed for
# non-quantized bf16 training.
# - See:
# - TRL issue: https://github.com/huggingface/trl/issues/6089
# - Upstream issue: https://github.com/deepspeedai/DeepSpeed/issues/8072
# - autocast_adapter_dtype was introduced in PEFT 0.12.0; before, no upcast existed: no need to pass the kwarg
get_peft_model_kwargs = {}
if (
args.deepspeed_plugin is not None
and args.deepspeed_plugin.zero_stage == 3
and not _is_quantized_model
and Version(peft.__version__) >= Version("0.12.0")
):
get_peft_model_kwargs["autocast_adapter_dtype"] = False
model = get_peft_model(model, peft_config, **get_peft_model_kwargs)
# PEFT + DeepSpeed ZeRO-3 requires reentrant checkpointing. For more details, see
# https://github.com/huggingface/trl/issues/2514#issuecomment-2692152703
if (
is_peft_model(model)
and args.deepspeed_plugin is not None
and args.deepspeed_plugin.zero_stage == 3
and args.gradient_checkpointing
):
args.gradient_checkpointing_kwargs = args.gradient_checkpointing_kwargs or {}
use_reentrant = args.gradient_checkpointing_kwargs.get("use_reentrant")
if use_reentrant is False:
logger.warning(
"You are using PEFT with DeepSpeed ZeRO-3 and gradient checkpointing with `use_reentrant=False`. "
"`use_reentrant` is forced to `True` in this configuration to ensure correct training. To remove "
"this warning, unset `use_reentrant` in `gradient_checkpointing_kwargs` or set it to `True`."
)
args.gradient_checkpointing_kwargs["use_reentrant"] = True
# When using gradient checkpointing with PEFT, we need to enable input gradients. transformers.Trainer normally
# handles this, but a bug currently prevents it; see https://github.com/huggingface/transformers/issues/42489
if is_peft_model(model) and args.gradient_checkpointing:
model.enable_input_require_grads()
# When using QLoRA, the PEFT adapter weights are converted to bf16 to follow the recommendations from the
# original paper [see https://huggingface.co/papers/2305.14314, paragraph 3]. Normally, this can be done by
# passing `autocast_adapter_dtype=False` to `get_peft_model`, but this option is not yet supported for
# quantized models. See: https://github.com/huggingface/peft/issues/2889
if _is_quantized_model:
for param in model.parameters():
if param.requires_grad:
param.data = param.data.to(torch.bfloat16)
# In Prompt Tuning a small set of trainable virtual tokens [continuous prompt embeddings] is prepended to the
# input. We store the number of these tokens so we can account for them correctly when calculating accuracy.
self.num_virtual_tokens = 0
if is_peft_model(model):
if model.active_adapter in model.peft_config:
peft_model_config = model.peft_config[model.active_adapter]
self.num_virtual_tokens = getattr(peft_model_config, "num_virtual_tokens", 0)
# Data collator
# BFD packing requires padding-free mode; otherwise, the collator outputs padded attention masks, causing
# FlashAttention to ignore position_ids and recompute them incorrectly from the padded attention mask.
self.padding_free = args.padding_free or (args.packing and args.packing_strategy in {"bfd", "bfd_split"})
use_flash_attention = model.config._attn_implementation in FLASH_ATTENTION_VARIANTS
if self.padding_free:
if data_collator is not None:
raise ValueError("Passing a custom data collator is not supported when using padding-free.")
if args.packing and args.packing_strategy == "wrapped":
logger.warning(
"You are passing `padding_free=True` with the 'wrapped' packing strategy, which is not "
"recommended. Please refer to the documentation to understand why this is not recommended."
)
if not use_flash_attention:
logger.warning(
"Padding-free training is enabled, but the attention implementation is not set to a supported "
"Flash Attention variant. Padding-free training flattens batches into a single sequence, and only "
"the following implementations are known to reliably support this: "
f"{', '.join(sorted(FLASH_ATTENTION_VARIANTS))}. Using other implementations may lead to "
"unexpected behavior. To ensure compatibility, set `attn_implementation` in the model "
"configuration to one of these supported options or verify that your attention mechanism can "
"handle flattened sequences."
)
# Decide whether to use completion-only loss: if not specified, then it is set to True if the dataset format
# is prompt-completion, and False if the dataset format is language modeling.
dataset_sample = next(iter(train_dataset))
if args.completion_only_loss is None:
self.completion_only_loss = "prompt" in dataset_sample and "completion" in dataset_sample
else:
self.completion_only_loss = args.completion_only_loss
self._is_vision_dataset = "image" in dataset_sample or "images" in dataset_sample
# Unsloth: override _is_vlm for VLM models that pass a bare tokenizer
if not self._is_vlm and self._is_vision_dataset:
_m = model
if hasattr(_m, "model"): _m = _m.model
if hasattr(getattr(_m, "config", None), "vision_config") or \
_m.__class__.__name__.endswith("ForConditionalGeneration"):
self._is_vlm = True
if self._is_vision_dataset and not self._is_vlm:
raise ValueError(
"The dataset appears to be vision-related (contains 'image' or 'images' keys), but the provided "
"model does not seem to be a vision-language model. Please check your model and dataset."
)
if data_collator is None and not self._is_vision_dataset:
# Get the pad token: if not provided, use the one from the processing class or the eos token
# if the processing class does not have a pad token.
pad_token = args.pad_token or self._tokenizer.pad_token or self._tokenizer.eos_token
if pad_token not in self._tokenizer.get_vocab():
raise ValueError(
f"The specified `pad_token` ('{pad_token}') is not found in the vocabulary of the given "
f"`processing_class` ({processing_class.__class__.__name__}). Ensure that the `pad_token` exists "
"in the vocabulary before using it as a padding token."
)
self._tokenizer.pad_token = pad_token
data_collator = DataCollatorForLanguageModeling(
pad_token_id=self._tokenizer.pad_token_id,
max_length=None if self.padding_free else args.max_length,
truncation_mode=args.truncation_mode,
padding_free=self.padding_free,
pad_to_multiple_of=args.pad_to_multiple_of,
)
elif data_collator is None and self._is_vision_dataset:
data_collator = DataCollatorForVisionLanguageModeling(
processor=processing_class,
max_length=args.max_length,
completion_only_loss=self.completion_only_loss,
pad_to_multiple_of=args.pad_to_multiple_of,
dataset_text_field=args.dataset_text_field,
)
if args.packing and args.packing_strategy in {"bfd", "bfd_split"} and not use_flash_attention:
logger.warning(
"You are using packing, but the attention implementation is not set to a supported Flash Attention "
"variant. Packing gathers multiple samples into a single sequence, and only the following "
f"implementations are known to reliably support this: {', '.join(sorted(FLASH_ATTENTION_VARIANTS))}. "
"Using other implementations may lead to cross-contamination between samples. To avoid this, either "
"disable packing by setting `packing=False`, or set `attn_implementation` in the model configuration "
"to one of these supported options."
)
if args.assistant_only_loss and not is_conversational(dataset_sample):
raise ValueError(
"You set `assistant_only_loss=True`, but the dataset is not conversational. This option is only "
"supported for conversational datasets."
)
# When assistant_only_loss is enabled, swap in a training chat template with {% generation %} markers
# if the current template doesn't already have them.
if args.assistant_only_loss and not has_generation_markers(processing_class.chat_template):
self.chat_template = get_training_chat_template(processing_class)
else:
self.chat_template = None
# A template can define generation markers and still attribute the assistant's end-of-turn token to the next
# message, leaving it out of the assistant mask so the model is never trained to stop.
if args.assistant_only_loss and not is_chat_template_stop_token_trained(
processing_class, chat_template=self.chat_template
):
logger.warning(
"The chat template does not include the assistant turn's end-of-turn token in the loss mask; "
"the model may not learn to stop."
)
# Dataset
if self.padding_free and not args.packing and args.max_length is not None and not self._is_vision_dataset:
raise ValueError(
"When `padding_free=True` without packing, `max_length` is not enforced. Either enable packing "
"(e.g., `packing=True, packing_strategy='bfd'`), provide already truncated inputs, or set "
"`max_length=None`."
)
# Skip dataset preparation if `skip_prepare_dataset=True` in `dataset_kwargs`, or if it's a VLM, where
# preprocessing [e.g., image-to-pixel conversion] is too costly and done on the fly instead.
self._skip_prepare_dataset = (
args.dataset_kwargs is not None
and args.dataset_kwargs.get("skip_prepare_dataset", False)
or self._is_vision_dataset
)
# Kept on the instance so that `evaluate` can preprocess freshly-passed eval datasets the same way.
self._formatting_func = formatting_func
eval_datasets = (
eval_dataset if isinstance(eval_dataset, dict) else {"eval": eval_dataset} if eval_dataset else {}
)
self._reject_skip_prepare_without_labels({"train": train_dataset, **eval_datasets}, data_collator)
if not self._skip_prepare_dataset:
if self.completion_only_loss and formatting_func:
raise ValueError(
"A formatting function was provided while `completion_only_loss=True`, which is incompatible. "
"Using a formatter converts the dataset to a language modeling type, conflicting with "
"completion-only loss. To resolve this, apply your formatting function before passing the "
"dataset, or disable `completion_only_loss` in `SFTConfig`."
)
self._unsloth_model_ref = model
train_dataset = self._prepare_dataset(
train_dataset, processing_class, args, args.packing, formatting_func, "train"
)
if eval_dataset is not None:
packing = args.packing if args.eval_packing is None else args.eval_packing
if isinstance(eval_dataset, dict):
eval_dataset = {
key: self._prepare_dataset(dataset, processing_class, args, packing, formatting_func, key)
for key, dataset in eval_dataset.items()
}
else:
eval_dataset = self._prepare_dataset(
eval_dataset, processing_class, args, packing, formatting_func, "eval"
)
# Loss function
if not args.use_liger_kernel: # liger supports dft loss by just passing use_token_scaling=True
if args.loss_type == "nll":
pass # use the default loss
elif args.loss_type == "dft":
if compute_loss_func is not None:
raise ValueError(
"You passed a `compute_loss_func` together with `loss_type='dft'` to the `SFTTrainer`. "
"When using `loss_type='dft'`, the loss function is internally set to the DFT loss, so "
"passing a `compute_loss_func` is not allowed."
)
compute_loss_func = dft_loss
elif args.loss_type == "chunked_nll":
# Same math as `"nll"` but the `lm_head` matmul is skipped on ignored tokens and the CE is computed in
# chunks of tokens. Implemented by patching the model's forward before `super[].__init__` so accelerate
# wraps the patched forward.
# For PEFT, patch the inner causal LM rather than the `PeftModel` wrapper. LoRA / IA³ /
# `modules_to_save` adapters live in the module tree, so they're hit even when we bypass
# `PeftModel.forward`. Prompt-learning variants need `PeftModel.forward` to run first [to inject
# virtual tokens], then it delegates into the patched inner forward.
target = model.get_base_model() if is_peft_model(model) else model
# The chunked path reads the output projection weight directly, which would silently drop the
# adapter delta [and starve its parameters of gradients] if the head itself is a PEFT tuner layer.
if is_peft_model(model):
from peft.tuners.tuners_utils import BaseTunerLayer
if isinstance(target.get_output_embeddings(), BaseTunerLayer):
raise ValueError(
"`loss_type='chunked_nll'` is not supported when `lm_head` is wrapped by a PEFT adapter "
"(e.g. `target_modules='all-linear'` or explicitly including `'lm_head'`). Either remove "
"`lm_head` from `target_modules`, or switch to `loss_type='nll'`. If this is a real use "
"case for you, please open an issue at https://github.com/huggingface/trl/issues."
)
_patch_chunked_ce_lm_head(target, chunk_size=_CHUNKED_LM_HEAD_CHUNK_SIZE, is_vlm=self._is_vlm)
else:
raise ValueError(
f"Invalid `loss_type` {args.loss_type} passed. Supported values are 'nll', 'dft', and "
"'chunked_nll'."
)
elif args.loss_type == "chunked_nll":
raise ValueError("`loss_type='chunked_nll'` is not compatible with `use_liger_kernel=True`.")
# Transformers explicitly set use_reentrant=True in the past to silence a PyTorch warning, but the default was
# never updated once PyTorch switched to recommending use_reentrant=False. Until that change lands upstream
# [see https://github.com/huggingface/transformers/pull/43203] and is released [most likely in 5.0.0], we
# default to the recommended non-reentrant behavior here, while preserving any user-provided value.
if args.gradient_checkpointing and Version(transformers.__version__) < Version("5.0.0"):
args.gradient_checkpointing_kwargs = args.gradient_checkpointing_kwargs or {}
args.gradient_checkpointing_kwargs.setdefault("use_reentrant", False)
super().__init__(
model=model,
args=args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
processing_class=processing_class,
compute_loss_func=compute_loss_func,
compute_metrics=compute_metrics,
callbacks=callbacks,
optimizers=optimizers,
optimizer_cls_and_kwargs=optimizer_cls_and_kwargs,
preprocess_logits_for_metrics=preprocess_logits_for_metrics,
)
# Initialize activation offloading context
if self.args.activation_offloading:
self.maybe_activation_offload_context = get_act_offloading_ctx_manager(model=self.model)
else:
self.maybe_activation_offload_context = contextlib.nullcontext()
# MoE load-balancing auxiliary loss, applied to Mixture-of-Experts models [no effect otherwise]
text_config = model.config.get_text_config()
is_moe = getattr(text_config, "output_router_logits", None) is not None
self.aux_loss_enabled = is_moe and self.args.router_aux_loss_coef != 0.0
if is_moe:
# The native and chunked forwards add the aux loss from the model config, so keep the config in sync with
# the coef: enable it [and propagate the coef] when non-zero, disable it otherwise. This overrides any
# `output_router_logits` the model was loaded with, so `router_aux_loss_coef=0.0` reliably turns it off.
text_config.output_router_logits = self.aux_loss_enabled
text_config.router_aux_loss_coef = self.args.router_aux_loss_coef
# Initialize the metrics
self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
self._total_train_tokens = 0
# Add tags to the model
self.model.add_model_tags(self._tag_names)
def _tokenize(
self,
processing_class: PreTrainedTokenizerBase | ProcessorMixin,
input: str | list,
**kwargs,
) -> dict[str, list]:
"""Tokenize a single example for dataset preprocessing.
Dispatches to `apply_chat_template` for conversational input (list of message dicts) and to `__call__` for
non-conversational input (str). For VLMs, normalizes the batch dimension that processors emit even for single
examples.
Args:
processing_class ([`~transformers.PreTrainedTokenizerBase`] or [`~transformers.ProcessorMixin`]):
The tokenizer or processor to use.
input (`str` or `list`):
A string for non-conversational input, or a list of message dicts for conversational input.
**kwargs:
Forwarded to `apply_chat_template` (e.g. `add_generation_prompt`, `return_assistant_tokens_mask`).
Returns:
`dict` with at least an `"input_ids"` key mapping to a flat `list[int]`.
"""
if isinstance(input, list): # conversational: list of message dicts
if self._is_vlm:
input = prepare_multimodal_messages(input)
result = processing_class.apply_chat_template(
input, tokenize=True, return_dict=True, chat_template=self.chat_template, **kwargs
)
else: # non-conversational: plain text string
result = processing_class(text=input)
# VLMs emit a batch dimension even for single examples; unwrap it
if self._is_vlm:
return {k: v[0] for k, v in result.items()}
return result
def _prepare_dataset(
self,
dataset: Union[Dataset, IterableDataset],
processing_class,
args,
packing: bool,
formatting_func: Optional[Callable[[dict], str]],
dataset_name: str,
) -> Union[Dataset, IterableDataset]:
# All Unsloth Zoo code licensed under LGPLv3
try:
if isinstance(dataset, ConstantLengthDataset): return dataset
except:
pass
map_kwargs = {}
use_desc = isinstance(dataset, Dataset)
is_vlm = hasattr(processing_class, "tokenizer")
tokenizer = processing_class
if is_vlm: tokenizer = processing_class.tokenizer
# Detect whether the model's module needs token_type_ids when training
import sys as _sys
_needs_token_type_ids = False
# Split to avoid compiler substring match on masking_utils names
_ccm = 'create_' + 'causal_mask_mapping'
_model = getattr(self, '_unsloth_model_ref', None) or getattr(self, 'model', None)
if _model is not None:
for _m in (_model, getattr(_model, 'model', None)):
if _m is None: continue
_mod = _sys.modules.get(type(_m).__module__)
if _mod is not None and hasattr(_mod, _ccm):
_needs_token_type_ids = True
break
if not _needs_token_type_ids:
# Fallback: model not yet available, check processor class MRO
for _base in type(processing_class).__mro__:
_base_mod = getattr(_base, '__module__', '')
if 'transformers.models.' in _base_mod:
_modeling_mod = _base_mod.replace('.processing_', '.modeling_')
_mod = _sys.modules.get(_modeling_mod)
if _mod is not None and hasattr(_mod, _ccm):
_needs_token_type_ids = True
break
if _needs_token_type_ids and hasattr(args, 'remove_unused_columns'):
args.remove_unused_columns = False
# Get max length
max_seq_length = getattr(args, "max_length", 0)
if max_seq_length == 0: max_seq_length = getattr(args, "max_seq_length", 0)
if max_seq_length == 0: max_seq_length = getattr(self, "max_seq_length", 0)
if max_seq_length == 0: max_seq_length = getattr(self, "max_seq", 0)
if max_seq_length == 0: raise RuntimeError("Unsloth: max_seq_length is 0! Please specify one!")
dataset_text_field = getattr(args, "dataset_text_field", "text")
do_truncation = max_seq_length != 0
do_formatting_func = False
do_tokenize = True
do_prompt_completion = False
# Get correct column names
column_names = set(next(iter(dataset)).keys())
used_column_names = ["input_ids"]
if "attention_mask" in column_names:
used_column_names.append("attention_mask")
if _needs_token_type_ids:
used_column_names.append("token_type_ids")
# Skip tokenization if already tokenized; just set the data collator
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling
if "labels" in column_names:
# Most likely forgot data collator!
if is_vlm and not hasattr(tokenizer, "pad"):
raise RuntimeError(f"Unsloth: {processing_class.__class__} does not have .pad!")
self.data_collator = DataCollatorForSeq2Seq(tokenizer)
used_column_names.append("labels")
do_tokenize = False
elif "input_ids" in column_names:
if is_vlm and not hasattr(tokenizer, "pad"):
raise RuntimeError(f"Unsloth: {processing_class.__class__} does not have .pad!")
self.data_collator = DataCollatorForLanguageModeling(tokenizer, mlm = False)
do_tokenize = False
elif "prompt" in column_names and "completion" in column_names:
# Prompt/completion dataset (used with completion_only_loss).
# TRL's __init__ already set self.data_collator for completion_only_loss
# before calling us -- we must NOT overwrite it here.
do_prompt_completion = True
used_column_names.append("completion_mask")
elif dataset_text_field not in column_names:
do_formatting_func = True
if formatting_func is None:
raise RuntimeError("Unsloth: You must specify a `formatting_func`")
pass
if do_tokenize:
if do_formatting_func:
test_text = formatting_func(next(iter(dataset)))
if not isinstance(test_text, list):
raise ValueError(
"Unsloth: The `formatting_func` should return a list of processed strings."
)
test_text = test_text[0]
elif do_prompt_completion:
_first_ex = next(iter(dataset))
try:
from trl import is_conversational as _sft_is_conversational
except ImportError:
def _sft_is_conversational(example):
for key in ("prompt", "completion", "messages"):
val = example.get(key)
if isinstance(val, list) and val and isinstance(val[0], dict):
if "role" in val[0] and "content" in val[0]:
return True
return False
_is_conv = _sft_is_conversational(_first_ex)
if not _is_conv:
test_text = _first_ex["prompt"]
else:
test_text = None # chat template handles BOS
else:
test_text = next(iter(dataset))[dataset_text_field][0]
chat_template = getattr(processing_class, 'chat_template', '')
if chat_template == '' and is_vlm:
chat_template = getattr(tokenizer, 'chat_template', '')
if chat_template is None:
chat_template = ''
# Detect double BOS so we can drop the duplicate
add_special_tokens = True
bos_token_1 = getattr(processing_class, 'bos_token', None)
bos_token_2 = getattr(tokenizer, 'bos_token', None)
bos_token = bos_token_1 or bos_token_2
if bos_token is not None:
if (test_text is not None and test_text.startswith(bos_token)) or bos_token in chat_template:
add_special_tokens = False
print("Unsloth: We found double BOS tokens - we shall remove one automatically.")
pass
def _tokenize(example):
return tokenizer(
example[dataset_text_field] if not do_formatting_func else formatting_func(example),
truncation = do_truncation,
max_length = max_seq_length,
return_token_type_ids = _needs_token_type_ids,
add_special_tokens = add_special_tokens,
)
pass
if not isinstance(dataset, IterableDataset):
import multiprocessing as _mp
dataset_num_proc = getattr(args, "dataset_num_proc", None)
if dataset_num_proc is None:
if _mp.get_start_method() != 'fork':
dataset_num_proc = None
else:
import psutil
dataset_num_proc = min(max((psutil.cpu_count() or 1)+4, 2), 64)
memory_gb_left = psutil.virtual_memory().available / (1024**3)
if memory_gb_left <= 2:
dataset_num_proc = 1
else:
dataset_num_proc = min(dataset_num_proc, int(memory_gb_left))
map_kwargs["num_proc"] = dataset_num_proc
else:
map_kwargs["batch_size"] = dataset._ex_iterable.batch_size
if do_prompt_completion:
_eos_token = getattr(tokenizer, 'eos_token', None)
def _tokenize_pc(example):
if _is_conv:
prompt_ids = processing_class.apply_chat_template(
example["prompt"], tokenize=True,
add_generation_prompt=True, return_dict=False,
tools=example.get("tools"),
**(example.get("chat_template_kwargs") or {}),
)
if prompt_ids and isinstance(prompt_ids[0], list):
prompt_ids = prompt_ids[0]
pc_processed = processing_class.apply_chat_template(
example["prompt"] + example["completion"],
return_dict=True, tokenize=True,
tools=example.get("tools"),
**(example.get("chat_template_kwargs") or {}),
)
if isinstance(pc_processed.get("input_ids", [None])[0], list):
pc_processed = {k: v[0] for k, v in pc_processed.items()}
pc_ids = pc_processed["input_ids"]
else:
_completion = example["completion"]
if _eos_token and not _completion.endswith(_eos_token):
_completion = _completion + _eos_token
prompt_ids = tokenizer(
example["prompt"], add_special_tokens=add_special_tokens,
)["input_ids"]
pc_ids = tokenizer(
example["prompt"] + _completion,
add_special_tokens=add_special_tokens,
)["input_ids"]
if do_truncation and max_seq_length > 0:
pc_ids = pc_ids[:max_seq_length]
n_prompt = min(len(prompt_ids), len(pc_ids))
completion_mask = [0] * n_prompt + [1] * (len(pc_ids) - n_prompt)
result = {"input_ids": pc_ids, "completion_mask": completion_mask}
if _needs_token_type_ids:
result["token_type_ids"] = [0] * len(pc_ids)
return result
if use_desc:
map_kwargs["desc"] = 'Unsloth: Tokenizing ["prompt"+"completion"]'
import warnings as _w
with _w.catch_warnings():
_w.filterwarnings("ignore", message=".*couldn't be hashed properly.*")
dataset = dataset.map(
_tokenize_pc, batched=False,
remove_columns=list(column_names), **map_kwargs,
)
else:
if use_desc: map_kwargs["desc"] = f'Unsloth: Tokenizing ["{dataset_text_field}"]'
import warnings as _w
with _w.catch_warnings():
_w.filterwarnings("ignore", message=".*couldn't be hashed properly.*")
dataset = dataset.map(_tokenize, batched = True, remove_columns = list(column_names), **map_kwargs)
# VLMs need .pad; switch the data collator
if is_vlm and not hasattr(processing_class, "pad") and not do_prompt_completion:
data_collator = DataCollatorForLanguageModeling(tokenizer, mlm = False)
self.data_collator = data_collator
pass
pass
if packing:
# Use TRL's pack_dataset if available
try:
pack_dataset
except:
print("Unsloth: Hugging Face's packing is currently buggy - we're disabling it for now!")
return dataset
if max_seq_length == 0:
raise ValueError("When packing is enabled, `max_seq_length` can't be `None`.")
if use_desc: map_kwargs["desc"] = f"Unsloth: Packing {dataset_name} dataset"
dataset = pack_dataset(
dataset.select_columns(used_column_names),
max_seq_length,
getattr(args, "packing_strategy", "bfd"),
map_kwargs,
)
pass
return dataset
def _set_signature_columns_if_needed(self):
# If `self.args.remove_unused_columns` is True, non-signature columns are removed.
# By default, this method sets `self._signature_columns` to the model's expected inputs (usually, "input_ids",
# "attention_mask" and "labels"). Dataset preparation also produces a "seq_lengths" column (for packing /
# padding-free), so we override the default signature columns to keep it alongside the model inputs.
if self._signature_columns is None:
if self._is_vision_dataset:
self._signature_columns = ["messages", "prompt", "completion", "image", "images"]
else:
self._signature_columns = ["input_ids", "labels", "seq_lengths"]
def _reject_skip_prepare_without_labels(self, datasets: dict[str, Dataset], data_collator) -> None:
# This guard may look defensive, but it covers a behavior change introduced when label building moved from
# the collator to dataset preparation: the collator used to consume the mask columns directly, so a
# skipped-preparation dataset carrying masks trained correctly. Now labels are built during preparation, which
# is skipped here, and the collator ignores the mask columns. Without a "labels" column, such a dataset would
# silently optimize the loss over the full sequence, so we fail loudly instead. Checked both at init and in
# `evaluate`, since a dataset passed directly to `evaluate` also skips preparation.
if not (
self._skip_prepare_dataset
and not self._is_vision_dataset
and isinstance(data_collator, DataCollatorForLanguageModeling)
):
return
for name, dataset in datasets.items():
cols = get_dataset_column_names(dataset)
if "labels" not in cols and ("completion_mask" in cols or "assistant_masks" in cols):
raise ValueError(
f"The {name} dataset has mask columns but no 'labels', and `skip_prepare_dataset=True` skips "
"label building, so it would train on the full sequence. Add a 'labels' column (-100 for "
"non-loss tokens) or drop `skip_prepare_dataset`."
)
def evaluate(
self,
eval_dataset: Dataset | dict[str, Dataset] | None = None,
ignore_keys: list[str] | None = None,
metric_key_prefix: str = "eval",
) -> dict[str, float]:
# When a dataset is passed directly to `evaluate` (e.g. a held-out test set), preprocess it the same way
# `__init__` does, so that `evaluate` accepts the same dataset types as the trainer (language modeling,
# prompt-completion, etc.). `_prepare_dataset` is idempotent: it skips datasets that are already tokenized. A
# `str` selects a dataset that was already prepared at init time, so it's left untouched.
if not self._skip_prepare_dataset and eval_dataset is not None and not isinstance(eval_dataset, str):
packing = self.args.packing if self.args.eval_packing is None else self.args.eval_packing
if isinstance(eval_dataset, dict):
eval_dataset = {
key: self._prepare_dataset(
dataset, self.processing_class, self.args, packing, self._formatting_func, key
)
for key, dataset in eval_dataset.items()
}
else:
eval_dataset = self._prepare_dataset(
eval_dataset, self.processing_class, self.args, packing, self._formatting_func, "eval"
)
eval_datasets = (
eval_dataset
if isinstance(eval_dataset, dict)
else {"eval": eval_dataset}
if eval_dataset is not None and not isinstance(eval_dataset, str)
else {}
)
self._reject_skip_prepare_without_labels(eval_datasets, self.data_collator)
return super().evaluate(
eval_dataset=eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix
)
def compute_loss(
self,
model,
inputs,
return_outputs = False,
num_items_in_batch = None,
):
outputs = super().compute_loss(
model,
inputs,
return_outputs = return_outputs,
num_items_in_batch = num_items_in_batch,
)
return outputs
def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys=None):
# Preserve the eval loop intent so compute_loss can decide whether logits are needed.
inputs["_prediction_loss_only"] = prediction_loss_only
return super().prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
# Override training step to add activation offloading context.
def training_step(self, *args, **kwargs):
with self.maybe_activation_offload_context:
return super().training_step(*args, **kwargs)
def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
mode = "train" if self.model.training else "eval"
metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()} # average the metrics
# This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
# start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
if mode == "eval":
metrics = {f"eval_{key}": val for key, val in metrics.items()}
logs.update(metrics)
super().log(logs, start_time)
self._metrics[mode].clear()
# Ensure the model card is saved along with the checkpoint
def _save_checkpoint(self, model, trial):
if self.args.hub_model_id is None:
model_name = Path(self.args.output_dir).name
else:
model_name = self.args.hub_model_id.split("/")[-1]
self.create_model_card(model_name=model_name)
super()._save_checkpoint(model, trial)
class UnslothSFTTrainer(_UnslothSFTTrainer):
"""
Trainer for Supervised Fine-Tuning (SFT) method.
This class is a wrapper around the [`~transformers.Trainer`] class and inherits all of its attributes and methods.
Example:
```python
>>> from trl import SFTTrainer
>>> from datasets import load_dataset
>>> dataset = load_dataset("roneneldan/TinyStories", split="train[:1%]")
>>> trainer = SFTTrainer(
... model="Qwen/Qwen2.5-0.5B-Instruct",
... train_dataset=dataset,
... )
>>> trainer.train()
```
Args:
model (`str` or [`~transformers.PreTrainedModel`] or [`~peft.PeftModel`]):
Model to be trained. Can be either:
- A string, being the *model id* of a pretrained model hosted inside a model repo on huggingface.co, or a
path to a *directory* containing model weights saved using
[`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
using `<ModelArchitecture>.from_pretrained` (where `<ModelArchitecture>` is derived from the model
config) with the keyword arguments in `args.model_init_kwargs`.
- A [`~transformers.PreTrainedModel`] object. Only causal language models are supported.
- A [`~peft.PeftModel`] object. Only causal language models are supported.
args ([`SFTConfig`], *optional*):
Configuration for this trainer. If `None`, a default configuration is used.
data_collator ([`~transformers.DataCollator`], *optional*):
Function to use to form a batch from a list of elements of the processed `train_dataset` or `eval_dataset`.
Will default to [`~trainer.sft_trainer.DataCollatorForLanguageModeling`] if the model is a language model
and [`~trainer.sft_trainer.DataCollatorForVisionLanguageModeling`] if the model is a vision-language model.
Custom collators must truncate sequences before padding; the trainer does not apply post-collation
truncation.
train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]):
Dataset to use for training. This trainer supports both [language modeling](#language-modeling) type and
[prompt-completion](#prompt-completion) type. The format of the samples can be either:
- [Standard](dataset_formats#standard): Each sample contains plain text.
- [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role
and content).
The trainer also supports processed datasets (tokenized) as long as they contain an `input_ids` field.
eval_dataset ([`~datasets.Dataset`], [`~datasets.IterableDataset`] or `dict[str, Dataset | IterableDataset]`):
Dataset to use for evaluation. It must meet the same requirements as `train_dataset`.
processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.ProcessorMixin`], *optional*):
Processing class used to process the data. If `None`, the processing class is loaded from the model's name
with [`~transformers.AutoProcessor.from_pretrained`]. A padding token, `tokenizer.pad_token`, must be set.
If the processing class has not set a padding token, `tokenizer.eos_token` will be used as the default.
compute_loss_func (`Callable`, *optional*):
A function that accepts the raw model outputs, labels, and the number of items in the entire accumulated
batch (batch_size * gradient_accumulation_steps) and returns the loss. For example, see the default [loss
function](https://github.com/huggingface/transformers/blob/052e652d6d53c2b26ffde87e039b723949a53493/src/transformers/trainer.py#L3618)
used by [`Trainer`].
compute_metrics (`Callable[[EvalPrediction], dict]`, *optional*):
The function that will be used to compute metrics at evaluation. Must take a
[`~transformers.EvalPrediction`] and return a dictionary string to metric values. When passing
[`SFTConfig`] with `batch_eval_metrics` set to `True`, your `compute_metrics` function must take a boolean
`compute_result` argument. This will be triggered after the last eval batch to signal that the function
needs to calculate and return the global summary statistics rather than accumulating the batch-level
statistics.
callbacks (list of [`~transformers.TrainerCallback`], *optional*):
List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed
in [here](https://huggingface.co/docs/transformers/main_classes/callback).
If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`]
method.
optimizers (`tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None]`, *optional*, defaults to `(None, None)`):
A tuple containing the optimizer and the scheduler to use. Will default to an instance of `AdamW` on your
model and a scheduler given by [`~transformers.get_linear_schedule_with_warmup`] controlled by `args`.
optimizer_cls_and_kwargs (`tuple[Type[torch.optim.Optimizer], Dict[str, Any]]`, *optional*):
A tuple containing the optimizer class and keyword arguments to use. Overrides `optim` and `optim_args` in
`args`. Incompatible with the `optimizers` argument.
Unlike `optimizers`, this argument avoids the need to place model parameters on the correct devices before
initializing the Trainer.
preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`, *optional*):
A function that preprocess the logits right before caching them at each evaluation step. Must take two
tensors, the logits and the labels, and return the logits once processed as desired. The modifications made
by this function will be reflected in the predictions received by `compute_metrics`.
Note that the labels (second parameter) will be `None` if the dataset does not have them.
peft_config ([`~peft.PeftConfig`], *optional*):
PEFT configuration used to wrap the model. If `None`, the model is not wrapped.
formatting_func (`Callable`, *optional*):
Formatting function applied to the dataset before tokenization. Applying the formatting function explicitly
converts the dataset into a [language modeling](#language-modeling) type.
"""
def __init__(
self,
model,
args = None,
data_collator = None,
train_dataset = None,
eval_dataset = None,
processing_class = None,
compute_loss_func = None,
compute_metrics = None,
callbacks = None,
optimizer_cls_and_kwargs = None,
preprocess_logits_for_metrics = None,
peft_config = None,
formatting_func = None,
**kwargs
):
if args is None: args = UnslothSFTConfig()
use_bf16 = getattr(args, 'bf16', False)
if type(use_bf16) is not bool: use_bf16 = False
use_fp16 = getattr(args, 'fp16', False)
if type(use_fp16) is not bool: use_fp16 = False
force_float32 = False
full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'
if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):
print('Unsloth: Switching to float32 training since model cannot work with float16')
force_float32 = True
mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
if dtype is None: dtype = model.get_input_embeddings().weight.dtype
from unsloth_zoo.utils import _get_dtype
dtype = _get_dtype(dtype)
float16 = dtype == torch.float16
if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
if force_float32:
# Forced float32 training
args.fp16 = False
args.bf16 = False
os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
# args.mixed_precision is a new argument which needs to be set now
elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
# Mixed precision training
args.fp16 = float16
args.bf16 = not float16
os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'
# args.mixed_precision is a new argument which needs to be set now
elif mixed_precision_dtype == 'bfloat16':
# Both False since bfloat16 full finetuning doesn't do any autocasting.
args.fp16 = False
args.bf16 = False
os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
# args.mixed_precision is a new argument which needs to be set now
if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
args.eval_strategy = 'steps'
if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
ga_steps = getattr(args, 'gradient_accumulation_steps', None)
if ga_steps is not None and ga_steps > 1:
from transformers import __version__ as transformers_version
if Version(transformers_version) <= Version('4.45.2'):
print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
'`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
if getattr(args, 'eval_strategy', 'no') != 'no':
eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
fp16_full_eval = getattr(args, 'fp16_full_eval', False)
if type(fp16_full_eval) is not bool: fp16_full_eval = False
bf16_full_eval = getattr(args, 'bf16_full_eval', False)
if type(bf16_full_eval) is not bool: bf16_full_eval = False
if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
if force_float32:
args.bf16_full_eval = False
args.fp16_full_eval = False
elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
args.bf16_full_eval = True
args.fp16_full_eval = False
elif not bf16_full_eval and not fp16_full_eval:
args.bf16_full_eval = args.bf16
args.fp16_full_eval = args.fp16
_output_logits = False
if locals().get('compute_metrics', None) is not None: _output_logits = True
if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
if _output_logits:
os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
if model is not None:
_warnings_issued = getattr(model, 'warnings_issued', None)
if _warnings_issued is None:
model.warnings_issued = {}
elif not isinstance(_warnings_issued, dict):
try:
model.warnings_issued = dict(_warnings_issued)
except Exception:
model.warnings_issued = {}
if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
pass
else:
model_max_seq_length = getattr(model, 'max_seq_length', None)
args_max_seq_length = getattr(args, 'max_seq_length', None)
if args_max_seq_length is None and model_max_seq_length is not None:
max_seq_length = model.max_seq_length
if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
elif args_max_seq_length is not None and model_max_seq_length is not None:
if args_max_seq_length > model_max_seq_length:
print('Unsloth: You set `max_seq_length` as ' + str(args_max_seq_length) + ' but '
'the maximum the model supports is ' + str(model_max_seq_length) + '. We shall reduce it.')
args.max_seq_length = model_max_seq_length
if 'max_length' not in locals() and not hasattr(args, 'max_length'):
pass
else:
if hasattr(args, 'max_seq_length') and args.max_seq_length is not None and args.max_seq_length > 0:
if hasattr(args, 'max_length'):
args.max_length = args.max_seq_length
max_length = args.max_length
else:
model_max_length = getattr(model, 'max_seq_length', None)
if model_max_length is None: model_max_length = getattr(model, 'max_length', None)
if model_max_length is not None:
args.max_length = model_max_length
max_length = args.max_length
elif hasattr(args, 'max_length') and args.max_length is not None:
max_length = args.max_length
# if we are here, then we are in a weird case where max_length is set but max_seq_length is not set
setattr(model, 'max_seq_length', max_length)
else:
print('Unsloth: We did not find `max_seq_length` or `max_length` in the model or args. We will set it to 1024.')
args.max_length = 1024
if model is not None and hasattr(model, 'for_training'):
model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
if 'processing_class' in locals():
if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
__tokenizer = processing_class if 'processing_class' in locals() else tokenizer
from unsloth_zoo.vision_utils import UnslothVisionDataCollator
if not isinstance(data_collator, UnslothVisionDataCollator):
if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
data_collator = TransformersDataCollatorForLanguageModeling(
__tokenizer,
mlm = False,
mlm_probability = 0.0,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
data_collator = DataCollatorForSeq2Seq(
__tokenizer,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
else:
if hasattr(args, 'remove_unused_columns'): args.remove_unused_columns = False
if hasattr(args, 'dataset_text_field'): args.dataset_text_field = ''
if hasattr(args, 'dataset_kwargs'): args.dataset_kwargs = {'skip_prepare_dataset': True}
if not isinstance(data_collator, UnslothVisionDataCollator):
if not hasattr(__tokenizer, 'pad') and hasattr(__tokenizer, 'tokenizer'):
if isinstance(data_collator, DataCollatorForSeq2Seq):
data_collator = DataCollatorForSeq2Seq(
__tokenizer.tokenizer,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling):
data_collator = TransformersDataCollatorForLanguageModeling(
__tokenizer.tokenizer,
mlm = False,
mlm_probability = 0.0,
pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
)
other_metrics = []
from unsloth_zoo.logging_utils import PatchRLStatistics
PatchRLStatistics('sft_trainer', other_metrics)
IGNORED_TOKENIZER_NAMES = os.environ.get('UNSLOTH_IGNORED_TOKENIZER_NAMES', '').split('\n')
from unsloth_zoo.tokenizer_utils import fix_untrained_tokens
from unsloth_zoo.training_utils import fix_zero_training_loss
if 'tokenizer' not in locals(): tokenizer = processing_class
fix_untrained_tokens(model, tokenizer, train_dataset, IGNORED_TOKENIZER_NAMES, eps = 1e-16)
fix_zero_training_loss(model, tokenizer, train_dataset)
# [TODO] Fix up DataParallel multiplying batch sizes
# [TODO] DDP works, but DP seems to not work? [TODO]
if getattr(args, "parallel_mode", None) == ParallelMode.NOT_DISTRIBUTED and args.n_gpu > 1:
if getattr(args, "_n_gpu", 1) != 1:
args._n_gpu = 1
if "model" in locals() and hasattr(model, "for_training"):
model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
super().__init__(
model = model,
args = args,
data_collator = data_collator,
train_dataset = train_dataset,
eval_dataset = eval_dataset,
processing_class = processing_class,
compute_loss_func = compute_loss_func,
compute_metrics = compute_metrics,
callbacks = callbacks,
optimizer_cls_and_kwargs = optimizer_cls_and_kwargs,
preprocess_logits_for_metrics = preprocess_logits_for_metrics,
peft_config = peft_config,
formatting_func = formatting_func,**kwargs)
if "model" in locals() and hasattr(model, "for_inference"):
model.for_inference()
if hasattr(self, 'neftune_hook_handle'):
self.neftune_hook_handle.remove()
if hasattr(self, 'neftune_hook_handle'): del self.neftune_hook_handle
if getattr(args, 'neftune_noise_alpha', None) is not None:
model.get_input_embeddings().neftune_noise_alpha = self.neftune_noise_alpha
pass
if hasattr(self, 'accelerator'):
scaler = self.accelerator.scaler
current_model = model
while hasattr(current_model, 'model'):
current_model.accelerator_scaler = scaler
current_model = current_model.model
current_model.accelerator_scaler = scaler
pass
if hasattr(self, 'train'):
self.train = MethodType(prepare_for_training_mode(self.__class__.train), self)
pass
if hasattr(self, 'llm') and self.llm is not None and hasattr(self.llm, 'get_tokenizer'):
_vllm_tok = self.llm.get_tokenizer()
_pc = getattr(self, 'processing_class', None) or getattr(self, 'tokenizer', None)
if _vllm_tok is not None and _pc is not None and getattr(_pc, 'chat_template', None) is not None and getattr(_vllm_tok, 'chat_template', None) is None:
_vllm_tok.chat_template = _pc.chat_template
pass
pass
if hasattr(logger, "addFilter"):
import logging
class HideLoggingMessage(logging.Filter):
def __init__(self, text): self.text = text
def filter(self, x): return not (self.text in x.getMessage())
pass
logger.addFilter(HideLoggingMessage("`use_cache=True`"))
|