Spaces:
Sleeping
Sleeping
File size: 100,155 Bytes
2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 004ec91 2d9850d 3f5e53d 2d9850d 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 0afbf5b 08dcd28 3f5e53d 08dcd28 004ec91 08dcd28 004ec91 08dcd28 3f5e53d 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 f275ddc 004ec91 f275ddc 08dcd28 0afbf5b 507d48b 08dcd28 11bfff3 08dcd28 0afbf5b eddad46 08dcd28 f275ddc 08dcd28 eddad46 004ec91 a17fb4e 08dcd28 0afbf5b a17fb4e 08dcd28 dbbf449 a17fb4e 08dcd28 eddad46 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d eddad46 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 3f5e53d 004ec91 08dcd28 004ec91 08dcd28 004ec91 eddad46 004ec91 08dcd28 004ec91 08dcd28 ee49532 08dcd28 004ec91 08dcd28 6eaf3e3 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 08dcd28 004ec91 a17fb4e b4b603e 004ec91 08dcd28 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 9db1e30 dbbf449 9db1e30 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 004ec91 2e22656 004ec91 f275ddc 004ec91 f275ddc 004ec91 f275ddc 0afbf5b f275ddc 004ec91 f275ddc 0afbf5b f275ddc 004ec91 0afbf5b 004ec91 dbbf449 004ec91 dbbf449 004ec91 dbbf449 9bcf95f dbbf449 0afbf5b dbbf449 004ec91 0afbf5b 004ec91 dbbf449 004ec91 dbbf449 004ec91 0afbf5b dbbf449 004ec91 0afbf5b dbbf449 004ec91 0afbf5b 004ec91 a49c754 004ec91 a49c754 004ec91 ee49532 004ec91 dbbf449 004ec91 dbbf449 004ec91 dbbf449 004ec91 dbbf449 004ec91 ee49532 004ec91 ee49532 90afe7d ee49532 90afe7d ee49532 dbbf449 ee49532 2d9850d ee49532 004ec91 dbbf449 004ec91 dbbf449 004ec91 dbbf449 004ec91 eddad46 6ac4313 004ec91 dbbf449 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 eddad46 004ec91 dbbf449 004ec91 | 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 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 | # app.py
# =========================================================================
# 0. Bootstrap — pip upgrades antes de qualquer import sensível
# =========================================================================
import os
import sys
import subprocess
subprocess.check_call([
sys.executable, "-m", "pip", "install", "--quiet", "--upgrade",
"urllib3<2.0", "charset_normalizer<3.4"
])
BASE_PATH = os.path.dirname(os.path.abspath(__file__))
FINAL_OUTPUT_DIR = os.path.join(BASE_PATH, "trained_model_output")
# Suprimir warnings antes de qualquer outro import
import warnings
warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)
try:
from requests.packages.urllib3.exceptions import DependencyWarning
warnings.simplefilter("ignore", DependencyWarning)
except ImportError:
pass
# Silenciar oneDNN / TF (mesmo sem TensorFlow instalado)
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "0"
# =========================================================================
# 1. Stdlib
# =========================================================================
import glob
import inspect
import json
import logging
import math
import platform
import statistics
import threading
import time
import traceback
from contextlib import contextmanager, suppress
from datetime import datetime, timedelta
from threading import Thread
from typing import Any, Dict, Union
from warnings import warn as log_warning
# =========================================================================
# 2. Third-party — sistema / hardware
# =========================================================================
import chardet
import cpuinfo
import ctypes as ct
import psutil
# =========================================================================
# 3. PyTorch
# =========================================================================
import torch
import torch.nn as nn
# =========================================================================
# 4. Flask
# =========================================================================
from flask import Flask, jsonify, render_template, request, send_file, send_from_directory
# =========================================================================
# 5. HuggingFace / PEFT
# =========================================================================
from datasets import Dataset, load_from_disk
from peft import LoraConfig, PeftModel, TaskType, get_peft_model
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
DataCollatorForLanguageModeling,
Trainer,
TrainerCallback,
TrainerControl,
TrainerState,
TrainingArguments,
)
# =========================================================================
# 6. Local
# =========================================================================
from config_manager import (
_load_constants_from_file,
map_backend_to_frontend,
update_python_constants,
)
# =========================================================================
# 7. Logging
# =========================================================================
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s - %(levelname)s - %(message)s"
)
#---------------------
app = Flask(__name__)
all_data = []
#--------------------
print("-" * 80)
#-----------------------------------------------------------
def _ts():
return datetime.now().strftime("%H:%M:%S")
# -------------------------
# Deteção P/E cross-platform
# -------------------------
def detect_core_types():
"""
Tenta:
1) Windows: EfficiencyClass (Windows 11 expõe isto; em máquinas homogéneas virá tudo Efficiency=0).
2) Linux: sysfs core_type (1=Atom/E, 2=Core/P; 0=Unknown).
3) Fallback: assume homogéneo → todos P, E vazio.
Retorna (P_IDS, E_IDS, meta_dict).
"""
logical = psutil.cpu_count(logical=True) or os.cpu_count() or 1
# 1) Windows EfficiencyClass
if platform.system() == "Windows":
try:
RELATION_PROCESSOR_CORE = 0
class GROUP_AFFINITY(ct.Structure):
_fields_ = [("Mask", ct.c_ulonglong),
("Group", ct.c_ushort),
("Reserved", ct.c_ushort * 3)]
class PROCESSOR_RELATIONSHIP(ct.Structure):
_fields_ = [("Flags", ct.c_ubyte),
("EfficiencyClass", ct.c_ubyte),
("Reserved", ct.c_ubyte * 20),
("GroupCount", ct.c_ushort)]
class SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX_HEADER(ct.Structure):
_fields_ = [("Relationship", ct.c_int),
("Size", ct.c_ulong)]
GetLPIEx = ct.windll.kernel32.GetLogicalProcessorInformationEx
GetLPIEx.restype = ct.c_bool
GetLPIEx.argtypes = [ct.c_int, ct.c_void_p, ct.POINTER(ct.c_ulong)]
buf_size = ct.c_ulong(0)
GetLPIEx(RELATION_PROCESSOR_CORE, None, ct.byref(buf_size))
if buf_size.value:
buf = (ct.c_byte * buf_size.value)()
if GetLPIEx(RELATION_PROCESSOR_CORE, ct.byref(buf), ct.byref(buf_size)):
eff_by_logical = {}
offset = 0
single_group = (logical <= 64)
while offset < buf_size.value:
header = SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX_HEADER.from_buffer(buf, offset)
size = header.Size
if header.Relationship == RELATION_PROCESSOR_CORE:
pr = PROCESSOR_RELATIONSHIP.from_buffer(buf, offset + ct.sizeof(header))
eff = pr.EfficiencyClass # 0 = mais performante (P), >0 = mais eficiente (E)
ga_offset = offset + ct.sizeof(header) + ct.sizeof(PROCESSOR_RELATIONSHIP)
for i in range(pr.GroupCount):
ga = GROUP_AFFINITY.from_buffer(buf, ga_offset + i * ct.sizeof(GROUP_AFFINITY))
mask = ga.Mask
if single_group:
for bit in range(64):
if (mask >> bit) & 1:
eff_by_logical[bit] = eff
else:
# Multi-grupo (>64 lógicos): mapping global complexo, cai para fallback.
pass
offset += size
if eff_by_logical:
p_ids = sorted([i for i in range(logical) if eff_by_logical.get(i, 0) == 0])
e_ids = sorted([i for i in range(logical) if eff_by_logical.get(i, 0) > 0])
return p_ids, e_ids, {"method": "windows_efficiencyclass", "notes": []}
except Exception:
pass # continua para Linux/fallback
# 2) Linux sysfs core_type
if platform.system() == "Linux":
try:
p_ids, e_ids, unknown = [], [], []
for cpu in range(logical):
path = f"/sys/devices/system/cpu/cpu{cpu}/topology/core_type"
try:
with open(path) as f:
val = f.read().strip()
except FileNotFoundError:
val = None
if val is None:
p_ids = e_ids = []
break
try:
t = int(val)
except ValueError:
t = 0
# Kernel moderno: 1=Atom(E), 2=Core(P), 0=Unknown
if t == 2:
p_ids.append(cpu)
elif t == 1:
e_ids.append(cpu)
else:
unknown.append(cpu)
if p_ids or e_ids:
notes = []
if unknown:
notes.append(f"{len(unknown)} CPUs com core_type=Unknown (tratados como P).")
p_ids = sorted(p_ids + unknown) # conservador: desconhecidos como P
return sorted(p_ids), sorted(e_ids), {"method": "linux_core_type", "notes": notes}
except Exception:
pass
# 3) Fallback: homogéneo
return list(range(logical)), [], {"method": "homogeneous_fallback", "notes": []}
# -------------------------
# Mapa de siblings (HT) por core físico
# -------------------------
def get_core_siblings():
"""
Retorna lista de listas: cada sublista são os logical IDs que pertencem ao mesmo core físico.
"""
logical = psutil.cpu_count(logical=True) or 1
if platform.system() == "Windows":
try:
RELATION_PROCESSOR_CORE = 0
class GROUP_AFFINITY(ct.Structure):
_fields_ = [("Mask", ct.c_ulonglong),
("Group", ct.c_ushort),
("Reserved", ct.c_ushort * 3)]
class PROCESSOR_RELATIONSHIP(ct.Structure):
_fields_ = [("Flags", ct.c_ubyte),
("EfficiencyClass", ct.c_ubyte),
("Reserved", ct.c_ubyte * 20),
("GroupCount", ct.c_ushort)]
class SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX_HEADER(ct.Structure):
_fields_ = [("Relationship", ct.c_int),
("Size", ct.c_ulong)]
GetLPIEx = ct.windll.kernel32.GetLogicalProcessorInformationEx
GetLPIEx.restype = ct.c_bool
GetLPIEx.argtypes = [ct.c_int, ct.c_void_p, ct.POINTER(ct.c_ulong)]
buf_size = ct.c_ulong(0)
GetLPIEx(RELATION_PROCESSOR_CORE, None, ct.byref(buf_size))
if not buf_size.value:
return []
buf = (ct.c_byte * buf_size.value)()
if not GetLPIEx(RELATION_PROCESSOR_CORE, ct.byref(buf), ct.byref(buf_size)):
return []
siblings = []
offset = 0
single_group = (logical <= 64)
while offset < buf_size.value:
header = SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX_HEADER.from_buffer(buf, offset)
size = header.Size
if header.Relationship == RELATION_PROCESSOR_CORE:
pr = PROCESSOR_RELATIONSHIP.from_buffer(buf, offset + ct.sizeof(header))
ga_offset = offset + ct.sizeof(header) + ct.sizeof(PROCESSOR_RELATIONSHIP)
core_logicals = []
for i in range(pr.GroupCount):
ga = GROUP_AFFINITY.from_buffer(buf, ga_offset + i * ct.sizeof(GROUP_AFFINITY))
mask = ga.Mask
for bit in range(64):
if (mask >> bit) & 1:
core_logicals.append(bit) # single/multi-grupo: aproximação razoável
if core_logicals:
siblings.append(sorted(set(core_logicals)))
offset += size
return sorted(siblings, key=lambda s: min(s) if s else 1e9)
except Exception:
return []
elif platform.system() == "Linux":
try:
sibs = []
for cpu in range(logical):
path = f"/sys/devices/system/cpu/cpu{cpu}/topology/thread_siblings_list"
try:
with open(path) as f:
txt = f.read().strip()
except FileNotFoundError:
return []
items = []
for part in txt.split(","):
if "-" in part:
a, b = part.split("-")
items.extend(range(int(a), int(b) + 1))
else:
items.append(int(part))
sibs.append(sorted(set(items)))
# Deduplica sublistas iguais
seen = set()
uniq = []
for s in sibs:
t = tuple(s)
if t not in seen:
seen.add(t)
uniq.append(s)
return sorted(uniq, key=lambda s: min(s))
except Exception:
return []
return []
def order_by_physical_first(candidates, siblings_map):
"""
Reordena 'candidates' para usar primeiro 1 logical por core físico (evita siblings logo de início).
Se 'siblings_map' estiver vazio, retorna candidatos ordenados naturalmente.
"""
if not siblings_map:
return sorted(candidates)
cand_set = set(candidates)
first_pass = [next((x for x in g if x in cand_set), None) for g in siblings_map]
first_pass = [x for x in first_pass if x is not None]
others = [x for g in siblings_map for x in g if x in cand_set and x not in first_pass]
leftovers = [x for x in sorted(candidates) if x not in first_pass and x not in others]
return first_pass + others + leftovers
# =========================================================================
# 1. Configuração base de hardware, Otimização e Hiperparâmetros
# =========================================================================
# --- 1.1 Hardware ---
LOGICAL_CPUS = psutil.cpu_count(logical=True) or os.cpu_count() or 16
PHYSICAL_CPUS = psutil.cpu_count(logical=False) or max(1, LOGICAL_CPUS // 2)
TOTAL_RAM_GB = psutil.virtual_memory().total / (1024 ** 3)
# ===============================================================
# Ajuste de performance (auto/manual)
# ===============================================================
OMP_THREADS_UTILIZATION = 0.95
DETECTION_PERFORMANCE = "auto" # "auto" | "manual"
CORES_UTILIZATION = 1.0
def compute_cores(cpu_count: int, utilization: float) -> int:
"""Aplica uma utilização fracional a um contagem de CPUs com clamp [1, cpu_count]."""
return max(1, min(int(math.floor(cpu_count * utilization)), cpu_count))
# NOTA: compute_effective_cores e compute_omp_threads faziam exactamente o mesmo;
# ambos substituídos por compute_cores acima. Usos em baixo são drop-in equivalentes.
EFFECTIVE_LOGICAL_CPUS = compute_cores(LOGICAL_CPUS, CORES_UTILIZATION)
# --- Detectar núcleos P/E e filtrar para o limite efectivo ---
_DET_P_full, _DET_E_full, _META = detect_core_types()
_DET_P = [i for i in _DET_P_full if i < EFFECTIVE_LOGICAL_CPUS]
_DET_E = [i for i in _DET_E_full if i < EFFECTIVE_LOGICAL_CPUS]
# ===============================================================
# Meta-informação de arranque (modo auto/manual)
# ===============================================================
OMP_THREADS = compute_cores(EFFECTIVE_LOGICAL_CPUS, OMP_THREADS_UTILIZATION)
if DETECTION_PERFORMANCE.lower() == "auto":
total_det = len(_DET_P) + len(_DET_E)
if total_det == 0:
_AUTO_META = "auto (fallback homogéneo)"
elif len(_DET_P) / total_det >= 0.5:
_AUTO_META = f"auto (balanceado P/E → OMP={OMP_THREADS})"
else:
_AUTO_META = f"auto (balanceado E-heavy → OMP={OMP_THREADS})"
else:
_AUTO_META = "manual"
print(f"[INFO] [{_ts()}] Modo de desempenho: {DETECTION_PERFORMANCE} ({_AUTO_META})")
# --- 1.2 Alocação de cores para tensores (OMP) e DataLoader ---
siblings = get_core_siblings()
ordered_P = order_by_physical_first(_DET_P, siblings)
ordered_E = order_by_physical_first(_DET_E, siblings)
needed = OMP_THREADS
P_CORE_IDS = []
if ordered_P:
take = ordered_P[:min(needed, len(ordered_P))]
P_CORE_IDS.extend(take)
needed -= len(take)
if needed > 0 and ordered_E:
take = [i for i in ordered_E if i not in P_CORE_IDS][:needed]
P_CORE_IDS.extend(take)
needed -= len(take)
# Fallback homogéneo
if not P_CORE_IDS:
P_CORE_IDS = list(range(min(OMP_THREADS, EFFECTIVE_LOGICAL_CPUS)))
# Secundários (DataLoader): E sobrantes → P sobrantes → restantes
used = set(P_CORE_IDS)
available = set(range(EFFECTIVE_LOGICAL_CPUS))
secondary = (
[i for i in _DET_E if i not in used and i in available] +
[i for i in _DET_P if i not in used and i in available] +
[i for i in available if i not in used and i not in _DET_P and i not in _DET_E]
)
_seen = set()
REMAINING_CORE_IDS = [x for x in secondary if not (x in _seen or _seen.add(x))]
DATALOADER_WORKERS = min(len(REMAINING_CORE_IDS), max(0, EFFECTIVE_LOGICAL_CPUS - len(P_CORE_IDS)))
# Variáveis de ambiente
os.environ["OMP_NUM_THREADS"] = str(OMP_THREADS)
os.environ["MKL_NUM_THREADS"] = str(OMP_THREADS)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
def set_affinity(core_ids):
try:
if platform.system() == "Linux":
os.sched_setaffinity(0, set(core_ids))
elif platform.system() == "Windows":
psutil.Process().cpu_affinity(core_ids)
except Exception as e:
print(f"[WARN] [{_ts()}] Não foi possível definir afinidade: {e}")
final_affinity_cores = sorted(i for i in set(P_CORE_IDS + REMAINING_CORE_IDS) if i < EFFECTIVE_LOGICAL_CPUS)
set_affinity(final_affinity_cores)
# ---------------------------------------------------------------
# Logs de arranque (apenas uma vez, evita duplicação no reload do Flask)
# ---------------------------------------------------------------
if os.environ.get("WERKZEUG_RUN_MAIN") != "true" and not globals().get("_LOGS_PRINTED", False):
_LOGS_PRINTED = True
has_e = bool(_DET_E)
print(f"[INFO] [{_ts()}] Método de deteção: {_META.get('method')}")
for note in _META.get("notes", []):
print(f"[INFO] [{_ts()}] Nota: {note}")
print(f"[INFO] [{_ts()}] Detetados (filtrados) → P: {len(_DET_P)} | E: {len(_DET_E)}")
# Tensores
P_from_P = [i for i in P_CORE_IDS if i in _DET_P]
P_from_E = [i for i in P_CORE_IDS if i in _DET_E]
if has_e:
e_str = f", +E: {len(P_from_E)} {sorted(P_from_E)}" if P_from_E else ""
print(f"[INFO] [{_ts()}] Tensores (OMP={OMP_THREADS}) → P: {len(P_from_P)} {sorted(P_from_P)}{e_str}")
else:
print(f"[INFO] [{_ts()}] Tensores (OMP={OMP_THREADS}) [homogéneo] → {len(P_CORE_IDS)} {sorted(P_CORE_IDS)}")
# DataLoader
SEC_E = [i for i in REMAINING_CORE_IDS if i in _DET_E]
SEC_P = [i for i in REMAINING_CORE_IDS if i in _DET_P]
SEC_OTHER = [i for i in REMAINING_CORE_IDS if i not in _DET_P and i not in _DET_E]
if has_e:
if SEC_E: print(f"[INFO] [{_ts()}] Secundário/DataLoader → E: {len(SEC_E)} {sorted(SEC_E)}")
if SEC_P: print(f"[INFO] [{_ts()}] Secundário/DataLoader → P: {len(SEC_P)} {sorted(SEC_P)}")
if SEC_OTHER: print(f"[INFO] [{_ts()}] Secundário/DataLoader → Outros: {len(SEC_OTHER)} {sorted(SEC_OTHER)}")
else:
print(f"[INFO] [{_ts()}] Cores restantes (homogéneo) → {len(REMAINING_CORE_IDS)} {sorted(REMAINING_CORE_IDS)}")
print("-" * 80)
print(f"[INFO] [{_ts()}] LOGICAL_CPUS: {LOGICAL_CPUS} | PHYSICAL_CPUS: {PHYSICAL_CPUS} | TOTAL_RAM_GB: {TOTAL_RAM_GB:.2f}")
print(f"[INFO] [{_ts()}] Cores Utilization: {CORES_UTILIZATION:.2%} → efetivos: {EFFECTIVE_LOGICAL_CPUS}")
print(f"[INFO] [{_ts()}] OMP Threads Utilization: {OMP_THREADS_UTILIZATION:.2%}")
print(f"[INFO] [{_ts()}] OMP Threads: {OMP_THREADS} | Dataloader Workers: {DATALOADER_WORKERS}")
if DATALOADER_WORKERS == 0:
print(f"[INFO] [{_ts()}] Sem cores sobrantes para DataLoader (todos {EFFECTIVE_LOGICAL_CPUS} alocados a OMP).")
total_allocated = len(set(P_CORE_IDS + REMAINING_CORE_IDS))
print(f"[INFO] [{_ts()}] Total alocados (OMP + DataLoader): {total_allocated} (de {EFFECTIVE_LOGICAL_CPUS} efetivos)")
print("-" * 80)
print(f"[INFO] [{_ts()}] Servidor pronto.\n")
print("-" * 80)
#-----------------------------------------------------------------------------------------------------------------------------
# =========================================================================
# 1.3 Parâmetros base do treino
# =========================================================================
BASE_BATCH_SIZE = 4
INITIAL_ACCUMULATION_MIN_STEPS = 2
INITIAL_ACCUMULATION_MAX_STEPS = 4
BASE_EVAL_SIZE = 10
BASE_LEARNING_RATE = 1e-4
LR_SCHEDULER_TYPE = "constant_with_warmup"
WARMUP_RATIO = 0.03
LOGGIN_STEPS = 10
SAVE_STRATEGY = "steps"
SAVE_STEPS = 500
EVAL_STEPS = 10
WEIGHT_DECAY = 0.1
OPTIM = "adamw_torch"
# =========================================================================
# 1.4 – 1.6 RAM e MAX_LEN
# =========================================================================
TARGET_RAM_UTILIZATION = 0.95 # Limite máximo de uso da RAM total
# Alias mantido por compatibilidade com ajustar_accumulation_steps() e outros usos
TARGET_ACCUMULATION_RAM_UTILIZATION = TARGET_RAM_UTILIZATION
# Custo empírico de RAM por batch (ajustar conforme modelo)
ESTIMATED_BATCH_GB = 0.35 # 0.4 para LLaMA
BASE_MAX_LEN = 256
TARGET_MAX_LEN_UTILIZATION = 0.75
MAX_LEN_INCREMENT = 256
MAX_LEN_CAP = 256
DEFAULT_FALLBACK_CAP = 8192
TOKENIZER_SENTINEL_CAP = 10_000_000
# Custo empírico de RAM para o dataset (ajustar conforme modelo)
ESTIMATED_BASE_DATASET_RAM_GB = 1.50
COST_PER_INCREMENT_GB = 0.45 # 0.45 para LLaMA
# =========================================================================
# 1.7 DynamicAccumulationCallback
# =========================================================================
DYNAMIC_ACCUMULATION_MAX_STEPS = 256
DYNAMIC_ACCUMULATION_TARGET_UTIL = 0.99 # mantido para referência futura
DYNAMIC_ACCUMULATION_HIGH_RAM_LIMIT = 98.5
DYNAMIC_ACCUMULATION_LOW_RAM_LIMIT = 45.0
# =========================================================================
# 1.8 Ganchos opcionais para PyTorch
# =========================================================================
from contextlib import suppress
with suppress(ImportError):
import torch
if hasattr(torch, "set_num_threads"):
torch.set_num_threads(max(1, OMP_THREADS))
if hasattr(torch, "set_num_interop_threads"):
# Interop baixo ajuda a estabilidade — normalmente 1 ou 2 é suficiente
torch.set_num_interop_threads(max(1, min(2, DATALOADER_WORKERS)))
# =========================================================================
# 1.9 Outros parâmetros
# =========================================================================
OPTIMIZER_REFRESH_INTERVAL = 600 # segundos
# =========================================================================
# 2️.Configuração Inicial e Cálculos
# =========================================================================
# --- Funções de Ajuste Dinâmico ---
def ajustar_accumulation_steps(
base_batch: int = BASE_BATCH_SIZE,
target_utilization: float = TARGET_ACCUMULATION_RAM_UTILIZATION,
estimated_batch_gb: float = ESTIMATED_BATCH_GB,
min_steps: int = INITIAL_ACCUMULATION_MIN_STEPS,
max_steps: int = INITIAL_ACCUMULATION_MAX_STEPS
) -> int:
"""
Ajusta gradient_accumulation_steps com base na RAM disponível (somente no arranque).
Respeita limites próprios e separados da fase dinâmica.
"""
mem = psutil.virtual_memory()
free_gb = mem.available / (1024 ** 3)
available_ram_for_batch_gb = TOTAL_RAM_GB * target_utilization
max_batches_fit = int(available_ram_for_batch_gb / estimated_batch_gb)
steps = max(min_steps, min(max_batches_fit // base_batch, max_steps))
logging.info(f"Memória usada: {mem.percent:.1f}% ({free_gb:.1f} GB livres)")
logging.info(f"Ajustando gradient_accumulation_steps (inicial) → {steps}")
return steps
#----------------------------------------------------------
def ajustar_max_len(
tokenizer_model_max_length: int,
target_utilization: float = TARGET_MAX_LEN_UTILIZATION,
base_max_len: int = BASE_MAX_LEN,
increment: int = MAX_LEN_INCREMENT,
max_cap: int = MAX_LEN_CAP,
estimated_base_dataset_ram_gb: float = ESTIMATED_BASE_DATASET_RAM_GB,
cost_per_increment_gb: float = COST_PER_INCREMENT_GB
) -> int:
"""
Ajusta automaticamente MAX_LEN conforme RAM disponível e limite do modelo.
"""
# RAM que pode ser usada para o dataset tokenizado
target_dataset_ram_gb = TOTAL_RAM_GB * target_utilization
current_max_len = base_max_len
if target_dataset_ram_gb < estimated_base_dataset_ram_gb:
log_warning(f"RAM disponível ({target_dataset_ram_gb:.1f} GB) é menor que a RAM estimada para MAX_LEN base ({estimated_base_dataset_ram_gb:.1f} GB). Usando MAX_LEN={base_max_len}.")
return min(base_max_len, tokenizer_model_max_length)
# Quantidade de RAM que podemos usar para expandir o MAX_LEN além do custo base
expandable_ram_gb = target_dataset_ram_gb - estimated_base_dataset_ram_gb
if expandable_ram_gb > 0:
num_increments = int(expandable_ram_gb / cost_per_increment_gb)
current_max_len += num_increments * increment
# Limitar pelo MAX_LEN_CAP e pelo limite do modelo
final_max_len = min(current_max_len, max_cap, tokenizer_model_max_length)
# logging.info(f"RAM total: {TOTAL_RAM_GB:.1f} GB. RAM alvo para dataset: {target_dataset_ram_gb:.1f} GB.")
# logging.info(f"Ajustando MAX_LEN dinamicamente -> {final_max_len} (limite do modelo: {tokenizer_model_max_length})")
logging.info(
f"RAM total: {TOTAL_RAM_GB:.1f} GB. RAM alvo para dataset: {target_dataset_ram_gb:.1f} GB.")
logging.info(
f"Ajustando MAX_LEN dinamicamente -> {final_max_len} "
f"(limite efetivo do modelo/tokenizer: {tokenizer_model_max_length})"
)
return final_max_len
# ==============================
# 3.Optimizer Refresh Callback
# ==============================
class OptimizerRefreshCallback(TrainerCallback):
def __init__(self, refresh_interval_sec: int = OPTIMIZER_REFRESH_INTERVAL):
self.refresh_interval = refresh_interval_sec
self.last_refresh_time = time.time()
def on_step_end(self, args, state, control, **kwargs):
trainer = kwargs.get("trainer", None)
if trainer is None:
return control
now = time.time()
elapsed = now - self.last_refresh_time
if elapsed >= self.refresh_interval:
logging.info(f"🔄 Refresh do otimizador após {elapsed:.1f}s (passo {state.global_step})")
# Faz o refresh
if hasattr(trainer, "scaler") and trainer.scaler is not None:
trainer.scaler.step(trainer.optimizer)
trainer.scaler.update()
else:
trainer.optimizer.step()
trainer.optimizer.zero_grad()
self.last_refresh_time = now
return control
# ========================================================
# 4.CALLBACK DINÂMICO PARA AJUSTE DE GRADIENT ACCUMULATION
# ========================================================
class DynamicAccumulationCallback(TrainerCallback):
def __init__(
self,
base_batch: int = BASE_BATCH_SIZE,
max_steps: int = DYNAMIC_ACCUMULATION_MAX_STEPS,
high_ram_limit: int = DYNAMIC_ACCUMULATION_HIGH_RAM_LIMIT,
low_ram_limit: int = DYNAMIC_ACCUMULATION_LOW_RAM_LIMIT,
step_interval: int = 5, # nº de steps entre verificações
time_interval: int = 240 # intervalo mínimo em segundos (10 min)
):
self.base_batch = base_batch
self.max_steps = max_steps
self.high_ram_limit = high_ram_limit
self.low_ram_limit = low_ram_limit
self.step_interval = step_interval
self.time_interval = time_interval
# controle interno
self.last_check_step = 0
self.last_check_time = time.time()
def on_step_end(self, args, state, control, **kwargs):
global CURRENT_ACCUM_STEPS
current_step = state.global_step
now = time.time()
# --- 1️⃣ Verifica se há override manual ativo ---
if CURRENT_ACCUM_STEPS is not None:
adjusted_steps = max(1, min(self.max_steps, CURRENT_ACCUM_STEPS))
args.gradient_accumulation_steps = adjusted_steps
effective_batch = adjusted_steps * self.base_batch
train_progress["current_accum_steps"] = adjusted_steps
logging.info(
f"🧭 Ajuste manual aplicado → gradient_accumulation_steps = {adjusted_steps} "
f"(Effective Batch: {effective_batch})"
)
CURRENT_ACCUM_STEPS = None
return control
# --- 2️⃣ Verifica se já passou o intervalo mínimo ---
step_diff = current_step - self.last_check_step
time_diff = now - self.last_check_time
if step_diff < self.step_interval and time_diff < self.time_interval:
# ainda não é hora de checar novamente
return control
# Atualiza controle de tempo e step
self.last_check_step = current_step
self.last_check_time = now
# --- 3️⃣ Ajuste automático baseado na RAM ---
mem = psutil.virtual_memory()
used = mem.percent
current_steps = args.gradient_accumulation_steps
# RAM alta → reduzir steps
if used > self.high_ram_limit and current_steps > 1:
new_steps = max(1, math.floor(current_steps / 1.6))
args.gradient_accumulation_steps = new_steps
new_effective_batch = new_steps * self.base_batch # ✅ novo valor após ajuste
train_progress["current_accum_steps"] = new_steps
logging.warning(
f"⚠️ RAM alta ({used:.1f}%) → diminuindo accumulation_steps "
f"de {current_steps} para {new_steps} "
f"(Effective Batch: {new_effective_batch})"
)
# RAM ociosa → aumentar steps
elif used < self.low_ram_limit and current_steps < self.max_steps:
new_steps = current_steps + 1
args.gradient_accumulation_steps = new_steps
new_effective_batch = new_steps * self.base_batch # ✅ recalculado após aumento
train_progress["current_accum_steps"] = new_steps
logging.info(
f"💡 RAM ociosa ({used:.1f}%) → aumentando accumulation_steps "
f"de {current_steps} para {new_steps} "
f"(Effective Batch: {new_effective_batch})"
)
return control
# ================================================================
# Sistema de Logs Coloridos
# ================================================================
class Colors:
HEADER = "\033[95m"
OKBLUE = "\033[94m"
OKCYAN = "\033[96m"
OKGREEN = "\033[92m"
WARNING = "\033[93m"
FAIL = "\033[91m"
ENDC = "\033[0m"
BOLD = "\033[1m"
UNDERLINE = "\033[4m"
def _timestamp():
return datetime.now().strftime("%H:%M:%S")
def log_info(msg):
ts = _timestamp()
print(f"{Colors.OKCYAN}[INFO]{Colors.ENDC} [{ts}] {msg}")
training_logs.append(f"[INFO] [{ts}] {msg}")
def log_success(msg):
ts = _timestamp()
print(f"{Colors.OKGREEN}[SUCESSO]{Colors.ENDC} [{ts}] {msg}")
training_logs.append(f"[SUCESSO] [{ts}] {msg}")
def log_warning(msg):
ts = _timestamp()
print(f"{Colors.WARNING}[AVISO]{Colors.ENDC} [{ts}] {msg}")
training_logs.append(f"[AVISO] [{ts}] {msg}")
def log_error(msg):
ts = _timestamp()
print(f"{Colors.FAIL}[ERRO]{Colors.ENDC} [{ts}] {msg}")
training_logs.append(f"[ERRO] [{ts}] {msg}")
def log_step(msg, step=None, percent=None):
ts = _timestamp()
progress = f" | Passo {step}" if step is not None else ""
pct = f" ({percent:.1f}%)" if percent is not None else ""
print(f"{Colors.OKBLUE}[{ts}]{Colors.ENDC} {msg}{progress}{pct}")
training_logs.append(f"[{ts}] {msg}{progress}{pct}")
# ================================================================
# Estado partilhado entre callbacks
# (anteriormente espalhado em múltiplas globais)
# ================================================================
_step_state = {
"last_step_time": time.time(), # usado por SimpleStepTimerCallback
"last_log_time": time.time(), # usado por LogCallback
"last_logged_step": -1, # usado por LogCallback (evita log duplicado)
}
# ================================================================
# 1. SimpleStepTimerCallback
# Loga o tempo de cada passo individual em segundos.
# ================================================================
class SimpleStepTimerCallback(TrainerCallback):
def on_step_begin(self, args, state, control, **kwargs):
if state.is_local_process_zero:
_step_state["last_step_time"] = time.time()
def on_step_end(self, args, state, control, **kwargs):
if not state.is_local_process_zero or state.global_step == 0:
return
elapsed = time.time() - _step_state["last_step_time"]
msg = (
f"PASSO: - - | {state.global_step}/{state.max_steps} |"
f" Tempo: | {elapsed:.3f}s |"
)
logging.info(f">> {msg}")
log_info(msg) # regista em training_logs com timestamp
# ================================================================
# 2. LogCallback
# Loga métricas de treino/avaliação e actualiza train_progress.
# ================================================================
class LogCallback(TrainerCallback):
def on_log(self, args, state, control, logs=None, **kwargs):
global train_progress
if not state.is_local_process_zero or not logs:
return
step = state.global_step
total_steps = state.max_steps
# Evita repetir o log para o mesmo step
if step == _step_state["last_logged_step"]:
return
_step_state["last_logged_step"] = step
now = time.time()
elapsed = now - _step_state["last_log_time"]
_step_state["last_log_time"] = now
loss = logs.get("loss")
eval_loss = logs.get("eval_loss")
lr = logs.get("learning_rate")
accum = train_progress.get("current_accum_steps",
args.gradient_accumulation_steps)
mem = psutil.virtual_memory()
if loss is not None and total_steps > 0:
percent = step / total_steps * 100
msg = (
f"Treino: step {step}/{total_steps} | Perda: {loss:.4f} | "
f"Tempo/step: {elapsed:.2f}s | Memória usada: {mem.percent:.1f}% | "
f"LR: {lr:.2e} | Accum: {accum} |"
)
log_step(msg, step=step, percent=percent)
train_progress.update({
"current": step,
"total": total_steps,
"percent": int(percent),
"status": "training",
"message": msg,
})
elif eval_loss is not None:
msg = f"Avaliação: step {step}/{total_steps} | Eval Loss: {eval_loss:.4f}"
try:
perplexity = math.exp(eval_loss) if eval_loss < 20 else float("inf")
msg += f" | Perplexidade: {perplexity:.2f}"
except Exception:
pass
log_info(msg)
train_progress["message"] = msg
def on_epoch_end(self, args, state, control, **kwargs):
global epoch_losses
if not state.is_local_process_zero:
return
epoch = int(state.epoch)
# Último log desta época com loss de treino (sem eval_loss)
train_log = next(
(l for l in reversed(state.log_history)
if "loss" in l and "epoch" in l
and int(l["epoch"]) == epoch
and "eval_loss" not in l),
None
)
# Último log desta época com eval_loss
eval_log = next(
(l for l in reversed(state.log_history)
if "eval_loss" in l and "epoch" in l
and int(l["epoch"]) == epoch),
None
)
parts = []
if train_log:
loss = train_log["loss"]
parts.append(f"Perda Média de Treino: {loss:.4f}")
epoch_losses.append({"epoch": epoch, "loss": loss})
if eval_log:
eval_loss = eval_log["eval_loss"]
parts.append(f"Perda de Avaliação: {eval_loss:.4f}")
suffix = " | ".join(parts) if parts else ""
log_success(f"Fim da Época {epoch}" + (f" | {suffix}" if suffix else ""))
# ================================================================
# 3. DynamicAccumulationCallback
# Ajuste dinâmico de gradient_accumulation_steps via RAM.
# ================================================================
class DynamicAccumulationCallback(TrainerCallback):
def __init__(
self,
base_batch: int = BASE_BATCH_SIZE,
max_steps: int = DYNAMIC_ACCUMULATION_MAX_STEPS,
high_ram_limit: int = DYNAMIC_ACCUMULATION_HIGH_RAM_LIMIT,
low_ram_limit: int = DYNAMIC_ACCUMULATION_LOW_RAM_LIMIT,
step_interval: int = 5, # nº de steps entre verificações automáticas
time_interval: int = 240, # intervalo mínimo em segundos
):
self.base_batch = base_batch
self.max_steps = max_steps
self.high_ram_limit = high_ram_limit
self.low_ram_limit = low_ram_limit
self.step_interval = step_interval
self.time_interval = time_interval
self._last_check_step = 0
self._last_check_time = time.time()
def on_step_end(self, args, state, control, **kwargs):
global CURRENT_ACCUM_STEPS
# 1. Override manual activo → aplica e limpa
if CURRENT_ACCUM_STEPS is not None:
new = max(1, min(self.max_steps, CURRENT_ACCUM_STEPS))
args.gradient_accumulation_steps = new
train_progress["current_accum_steps"] = new
logging.info(
f"🧭 Ajuste manual aplicado → gradient_accumulation_steps = {new} "
f"(Effective Batch: {new * self.base_batch})"
)
CURRENT_ACCUM_STEPS = None
return control
# 2. Ainda não é hora de verificar?
step_diff = state.global_step - self._last_check_step
time_diff = time.time() - self._last_check_time
if step_diff < self.step_interval and time_diff < self.time_interval:
return control
self._last_check_step = state.global_step
self._last_check_time = time.time()
# 3. Ajuste automático baseado na RAM
used = psutil.virtual_memory().percent
current = args.gradient_accumulation_steps
if used > self.high_ram_limit and current > 1:
new = max(1, math.floor(current / 1.6))
args.gradient_accumulation_steps = new
train_progress["current_accum_steps"] = new
logging.warning(
f"⚠️ RAM alta ({used:.1f}%) → diminuindo accumulation_steps "
f"de {current} para {new} "
f"(Effective Batch: {new * self.base_batch})"
)
elif used < self.low_ram_limit and current < self.max_steps:
new = current + 1
args.gradient_accumulation_steps = new
train_progress["current_accum_steps"] = new
logging.info(
f"💡 RAM ociosa ({used:.1f}%) → aumentando accumulation_steps "
f"de {current} para {new} "
f"(Effective Batch: {new * self.base_batch})"
)
return control
# =========================================================================
# Variáveis Globais
# =========================================================================
model = None
tokenizer = None
chat_model = None
chat_tokenizer = None
gguf_logs = []
gguf_status = {"status": "idle", "message": ""}
GGUF_OUT_DIR = "gguf_output"
train_progress = {
"current": 0,
"total": 1,
"percent": 0,
"status": "not started",
"message": "Aguardando início do treino.",
}
training_logs = [] # linhas de log acumuladas durante o treino
epoch_losses = []
all_data = [] # dados do último treino (permite continuar)
dataset_logs = []
dataset_status = {"status": "idle", "records": 0, "message": ""}
DATASET_OUT_DIR = "dataset_output" # pasta onde o dataset é gravado
CURRENT_ACCUM_STEPS = None # None = sem override manual activo
LAST_ACCUM_ORIGIN = None # "manual" | "auto" | None
TOTAL_TRAIN_STEPS = None # preenchido após tokenização/Trainer
BASE_BATCH_SIZE_EFFECTIVE = 4 # usado em /api/train_status para Effective Batch
# =========================================================================
# Cálculos iniciais de acumulação
# =========================================================================
ACCUMULATION_STEPS = ajustar_accumulation_steps()
EFFECTIVE_BATCH_SIZE = BASE_BATCH_SIZE * ACCUMULATION_STEPS
LAST_ACCUM_ORIGIN = "auto"
BASE_BATCH_SIZE_EFFECTIVE = EFFECTIVE_BATCH_SIZE
# Log após o cálculo — agora ACCUMULATION_STEPS tem valor real
training_logs.append(f"[INFO] accumulation inicial = {ACCUMULATION_STEPS} (auto)")
print(f"torch.get_num_threads(): {torch.get_num_threads()}")
print(f"torch.get_num_interop_threads(): {torch.get_num_interop_threads()}")
#----------------------------
# Função principal de treino
#----------------------------
def train_model_lora(
file_data, epochs, model_path, output_dir, accumulation_steps,
dataloader_workers, mode, lora_adapter_to_load=None,
initial_epochs_completed=0, train_mode="new_train"
):
"""
Função principal de treino LoRA (Low-Rank Adaptation).
CPU-only: XPU/CUDA removidos por decisão de arquitectura.
"""
# ---------------- VARIÁVEIS GLOBAIS ----------------
global training_logs, train_progress, epoch_losses, all_data
training_logs.clear()
all_data = file_data
# ---------------- ESTADO INICIAL ----------------
initial_status_message = (
"Retomando treino a partir de adaptador LoRA salvo..."
if lora_adapter_to_load else
"Iniciando processo de treino..."
)
train_progress.update({
"current": 0,
"total": 1,
"percent": 0,
"status": "starting",
"message": initial_status_message,
})
log_info(initial_status_message)
log_info(f"OMP Threads: {OMP_THREADS}, Dataloader Workers: {DATALOADER_WORKERS}")
log_info(f"Batch base: {BASE_BATCH_SIZE}, Steps iniciais: {ACCUMULATION_STEPS}")
log_info(f"Batch efetivo inicial: {EFFECTIVE_BATCH_SIZE}")
# ---------------- AUXILIAR: detetar camadas LoRA ----------------
def guess_lora_targets(model):
names = {name for name, _ in model.named_modules()}
if any("q_proj" in n and "v_proj" in n for n in names):
return ["q_proj", "v_proj"]
if any("query_key_value" in n for n in names):
return ["query_key_value"]
if any("Wqkv" in n for n in names):
return ["Wqkv"]
if any("c_attn" in n for n in names):
return ["c_attn"]
log_warning("Não foi possível detetar os target_modules. Usando ['q_proj', 'v_proj'] por defeito.")
return ["q_proj", "v_proj"]
# ---------------- BLOCO 1: Carregar Modelo e Tokenizer ----------------
try:
log_step("A carregar modelo e tokenizer base...")
os.makedirs("offload", exist_ok=True)
global tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
log_info(
f"Tokens antes do ajuste: PAD={tokenizer.pad_token}, "
f"EOS={tokenizer.eos_token}, BOS={tokenizer.bos_token}"
)
if tokenizer.pad_token is None:
tokenizer.add_special_tokens({'pad_token': '<PAD>'})
if tokenizer.eos_token is None:
tokenizer.add_special_tokens({'eos_token': '</s>'})
if tokenizer.bos_token is None:
tokenizer.add_special_tokens({'bos_token': '<s>'})
if tokenizer.unk_token is None:
tokenizer.add_special_tokens({'unk_token': '<unk>'})
log_info(
f"Tokens após ajuste: PAD={tokenizer.pad_token}, "
f"EOS={tokenizer.eos_token}, BOS={tokenizer.bos_token}, "
f"UNK={tokenizer.unk_token}"
)
# CPU-only: bf16 é ~2× mais rápido que float32 em CPU moderno
model = AutoModelForCausalLM.from_pretrained(
model_path,
device_map="cpu",
offload_folder="offload",
dtype=torch.bfloat16,
low_cpu_mem_usage=True,
)
# model.resize_token_embeddings(len(tokenizer))
model.resize_token_embeddings(len(tokenizer), mean_resizing=False)
model.config.use_cache = False
model.gradient_checkpointing_enable()
log_info("Modelo e tokenizer carregados (CPU / bfloat16).")
# --- Variáveis de controlo de fluxo (usadas mais abaixo) ---
resume_from_trainer_checkpoint = None
lora_adapter_path_final = os.path.join(output_dir, "lora_model")
lora_adapter_to_load_for_training = None
lora_model_loaded_for_initial_eval = False
#----------------------------------------------------------------------------------
# --- Lógica de 4 opções ---
# Prioridade:
#1. Checkpoint do Trainer
#2. LoRA Final Salvo
#3. LoRA passado como parâmetro (para continuar)
#4. Novo LoRA
# 1. Tentar encontrar o checkpoint mais recente do Trainer
checkpoint_dirs = [d for d in os.listdir(output_dir) if d.startswith("checkpoint-")]
if checkpoint_dirs:
checkpoint_dirs.sort(key=lambda x: int(x.split('-')[1]), reverse=True)
latest_candidate_checkpoint_dir = os.path.join(output_dir, checkpoint_dirs[0])
trainer_state_exists = os.path.exists(os.path.join(latest_candidate_checkpoint_dir, "trainer_state.json"))
lora_adapter_exists_in_checkpoint = (
os.path.exists(os.path.join(latest_candidate_checkpoint_dir, "adapter_model.safetensors")) or
os.path.exists(os.path.join(latest_candidate_checkpoint_dir, "adapter_model.bin"))
)
if trainer_state_exists and lora_adapter_exists_in_checkpoint:
resume_from_trainer_checkpoint = latest_candidate_checkpoint_dir
lora_adapter_to_load_for_training = latest_candidate_checkpoint_dir # Usar este LoRA para o treino
log_info(f"Opção 1: Encontrado checkpoint COMPLETO do Trainer em: {resume_from_trainer_checkpoint}. O treino será retomado daqui.")
# Extrair initial_epochs_completed do trainer_state.json
trainer_state_path = os.path.join(resume_from_trainer_checkpoint, "trainer_state.json")
if os.path.exists(trainer_state_path):
with open(trainer_state_path, 'r') as f:
trainer_state = json.load(f)
initial_epochs_completed = int(trainer_state.get('epoch', 0))
log_info(f"Checkpoint do Trainer encontrado. Épocas já completadas: {initial_epochs_completed}")
else:
log_warning(f"trainer_state.json não encontrado em {resume_from_trainer_checkpoint}. Assumindo 0 épocas completadas.")
else:
log_info(f"Diretório '{latest_candidate_checkpoint_dir}' encontrado, mas NÃO é um checkpoint COMPLETO do Trainer (faltam trainer_state.json ou arquivos do adaptador LoRA).")
#------------------------------------------------------------------------------------------------------------------------------------------------
# 2. Se não encontrou checkpoint do Trainer, tentar carregar um modelo LoRA final salvo (para avaliação/decisão)
# Esta opção deve ter prioridade sobre o LoRA final salvo, se o usuário explicitamente pediu para continuar com um LoRA.
if not resume_from_trainer_checkpoint and os.path.exists(lora_adapter_path_final) and any(os.scandir(lora_adapter_path_final)):
try:
# Carrega o LoRA final para avaliação/decisão, mas não para retomar o treino do Trainer
model = PeftModel.from_pretrained(model, lora_adapter_path_final, device_map=model_device_map)
for name, param in model.named_parameters():
if "lora" in name:
param.requires_grad = True
lora_adapter_to_load_for_training = lora_adapter_to_load # Usar este LoRA para o treino
log_info(f"Opção 2: Carregado adaptador LoRA de '{lora_adapter_to_load}' para continuar o treino.")
# Se estamos continuando, precisamos saber as épocas já completadas.
# Se lora_adapter_to_load é um checkpoint, podemos tentar ler o trainer_state.json
# Caso contrário, assumimos que é um LoRA final e o treino começa "do zero" em termos de Trainer state.
if "checkpoint-" in lora_adapter_to_load: # Se o LoRA passado é um checkpoint
trainer_state_path = os.path.join(lora_adapter_to_load, "trainer_state.json")
if os.path.exists(trainer_state_path):
with open(trainer_state_path, 'r') as f:
trainer_state = json.load(f)
initial_epochs_completed = int(trainer_state.get('epoch', 0))
log_info(f"LoRA de checkpoint encontrado. Épocas já completadas: {initial_epochs_completed}")
else:
log_warning(f"trainer_state.json não encontrado em {lora_adapter_to_load}. Assumindo 0 épocas completadas para continuação.")
else:
# Se é um LoRA final, não há estado de Trainer para retomar, então initial_epochs_completed permanece 0
log_info("LoRA final carregado para continuação. Treino do Trainer começará do zero.")
except Exception as e:
log_error(f"Falha ao carregar LoRA de '{lora_adapter_to_load}' para continuar o treino: {e}. Iniciando um novo LoRA.")
lora_adapter_to_load = None
lora_adapter_to_load_for_training = None # Resetar para criar um novo LoRA
# 3. Se não há checkpoint do Trainer nem LoRA final, mas um LoRA foi passado para continuar
# Esta opção só deve ser considerada se NENHUMA das opções anteriores foi ativada.
if not resume_from_trainer_checkpoint and not lora_adapter_to_load_for_training and os.path.exists(lora_adapter_path_final) and any(os.scandir(lora_adapter_path_final)):
try:
# Carrega o LoRA final para avaliação/decisão, mas não para retomar o treino do Trainer
model = PeftModel.from_pretrained(model, lora_adapter_path_final, device_map=model_device_map)
for name, param in model.named_parameters():
if "lora" in name:
param.requires_grad = True
lora_model_loaded_for_initial_eval = True
lora_adapter_to_load_for_training = lora_adapter_path_final # Usar este LoRA para o treino
log_info(f"Opção 3: Carregado adaptador LoRA final de: {lora_adapter_path_final}. Será avaliado e aguardará decisão do usuário.")
# Se carregamos um LoRA final, não há estado de Trainer para retomar, então initial_epochs_completed permanece 0
initial_epochs_completed = 0 # Garante que é 0 se não veio de um checkpoint completo
except Exception as e:
log_warning(f"Falha ao carregar LoRA final de {lora_adapter_path_final}: {e}. Prosseguindo sem ele.")
lora_adapter_path_final = None
lora_model_loaded_for_initial_eval = False
lora_adapter_to_load_for_training = None # Resetar para criar um novo LoRA
# --- Carregar o LoRA no modelo base ANTES de inicializar o Trainer ---
# Este bloco garante que o modelo base tenha o LoRA aplicado antes do Trainer ser instanciado.
# Ele só deve ser executado se um LoRA foi identificado para ser carregado
# E se o modelo ainda não é um PeftModel (ou seja, não foi carregado nas opções 1, 2 ou 3 acima)
if lora_adapter_to_load_for_training and not isinstance(model, PeftModel):
try:
model = PeftModel.from_pretrained(model, lora_adapter_to_load_for_training, device_map=model_device_map)
for name, param in model.named_parameters():
if "lora" in name:
param.requires_grad = True
log_info(f"Aplicado adaptador LoRA '{lora_adapter_to_load_for_training}' ao modelo base.")
except Exception as e:
log_error(f"Falha crítica ao aplicar LoRA '{lora_adapter_to_load_for_training}' ao modelo base: {e}. Iniciando um novo LoRA.")
lora_adapter_to_load_for_training = None
resume_from_trainer_checkpoint = None # Não podemos retomar o Trainer se o LoRA falhou
initial_epochs_completed = 0 # Resetar se falhou
# 4. Se nenhuma das opções acima, criar um novo adaptador LoRA
if not lora_adapter_to_load_for_training: # Se nenhum LoRA foi carregado/selecionado
targets = guess_lora_targets(model)
lora_config = LoraConfig(
r=16, lora_alpha=32, target_modules=targets, lora_dropout=0.05, bias="none", task_type=TaskType.CAUSAL_LM
)
model = get_peft_model(model, lora_config)
log_info(f"Opção 1: Nenhum checkpoint ou LoRA anterior encontrado. Novo adaptador LoRA criado (targets={targets}).")
resume_from_trainer_checkpoint = None # Se é um novo LoRA, não há estado de Trainer para retomar
if hasattr(model, "print_trainable_parameters"):
model.print_trainable_parameters()
else:
log_warning("print_trainable_parameters() não disponível neste modelo base.")
#-----------------------
# Antes de calcular total_target_epochs, ajuste initial_epochs_completed
# Determine o número de épocas que o Trainer deve executar
# Este é o valor que será passado para num_train_epochs do TrainingArguments
if train_mode == "new_train":
# Para um novo treino, o Trainer deve executar 'epochs' vezes.
# initial_epochs_completed é 0, o que é handled pelo Trainer.
num_epochs_for_trainer = float(epochs)
log_info(f"Modo 'new_train' detectado. O treino será executado por {num_epochs_for_trainer} épocas.")
elif train_mode == "continue_train":
# Para continuar, o Trainer deve executar até o 'initial_epochs_completed + epochs' total.
# Se você quer fazer mais 'X' épocas, o total final será initial_epochs_completed + X.
num_epochs_for_trainer = float(initial_epochs_completed + epochs)
log_info(f"Modo 'continue_train' detectado. Treino retomará da época {initial_epochs_completed} e visa um total de {num_epochs_for_trainer} épocas.")
else: # Default para new_train
log_warning(f"Modo de treino desconhecido: {train_mode}. Assumindo 'new_train'.")
num_epochs_for_trainer = float(epochs)
#--------------------------------------------------------------------------------------------
model.train()
total_target_epochs = initial_epochs_completed + epochs
# ── Limites de contexto do modelo e tokenizer ──────────────────────
max_len_model = getattr(model.config, "max_position_embeddings", None)
if not (isinstance(max_len_model, int) and max_len_model > 0):
max_len_model = DEFAULT_FALLBACK_CAP
tok_max_len_raw = getattr(tokenizer, "model_max_length", None)
if (tok_max_len_raw is None
or not isinstance(tok_max_len_raw, int)
or tok_max_len_raw <= 0
or tok_max_len_raw > TOKENIZER_SENTINEL_CAP):
tok_max_len_raw = DEFAULT_FALLBACK_CAP
effective_model_max_len = min(max_len_model, tok_max_len_raw)
DYNAMIC_MAX_LEN = ajustar_max_len(
tokenizer_model_max_length=effective_model_max_len,
target_utilization=TARGET_MAX_LEN_UTILIZATION,
base_max_len=BASE_MAX_LEN,
increment=MAX_LEN_INCREMENT,
max_cap=MAX_LEN_CAP,
estimated_base_dataset_ram_gb=ESTIMATED_BASE_DATASET_RAM_GB,
cost_per_increment_gb=COST_PER_INCREMENT_GB,
)
logging.info(f"MAX_LEN dinâmico definido para: {DYNAMIC_MAX_LEN}")
# ── Dataset: carregar do disco ou tokenizar ────────────────────────
tokenized_path = os.path.join(output_dir, "tokenized_dataset")
#----------------------------------------------------------------------------
# Depois de tokenizar o dataset
# enviar para bucket
#----------------------------------------------------------------------------
if os.path.exists(tokenized_path):
log_info("Dataset tokenizado encontrado no disco. A carregar...")
tokenized_datasets = load_from_disk(tokenized_path)
else:
training_data = []
for item in file_data:
if "prompt" in item and "completion" in item:
prompt = item.get("prompt", "")
completion = item.get("completion", "")
if isinstance(prompt, str) and isinstance(completion, str) and prompt and completion:
training_data.append({"text": f"<s>[INST] {prompt.strip()} [/INST] {completion.strip()}</s>"})
elif "text" in item and isinstance(item["text"], str) and item["text"].strip():
training_data.append({"text": f"<s>{item['text'].strip()}</s>"})
if not training_data:
log_error("Nenhum dado válido encontrado para treino após formatação.")
train_progress.update({"status": "error", "message": "Nenhum dado válido para treino."})
return
def tokenize_function(examples):
return tokenizer(examples["text"], truncation=True, max_length=DYNAMIC_MAX_LEN, padding=False)
tokenized_datasets = Dataset.from_list(training_data).map(
tokenize_function,
batched=True,
num_proc=DATALOADER_WORKERS,
remove_columns=["text"],
)
tokenized_datasets.save_to_disk(tokenized_path)
log_success(f"Dataset tokenizado salvo em {tokenized_path} com MAX_LEN={DYNAMIC_MAX_LEN}")
split_dataset = tokenized_datasets.train_test_split(test_size=0.1, seed=42)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
# ── TrainingArguments ──────────────────────────────────────────────
# ATENÇÃO: num_epochs_for_trainer deve ser definido antes desta função.
# Se não existir, usa `epochs` do argumento como fallback seguro.
_epochs_for_trainer = num_epochs_for_trainer if 'num_epochs_for_trainer' in dir() else epochs
training_args = TrainingArguments(
output_dir=output_dir,
num_train_epochs=_epochs_for_trainer,
per_device_train_batch_size=BASE_BATCH_SIZE,
per_device_eval_batch_size=BASE_EVAL_SIZE,
gradient_accumulation_steps=ACCUMULATION_STEPS,
learning_rate=BASE_LEARNING_RATE,
warmup_ratio=WARMUP_RATIO,
lr_scheduler_type=LR_SCHEDULER_TYPE,
weight_decay=WEIGHT_DECAY,
logging_steps=LOGGIN_STEPS,
save_strategy=SAVE_STRATEGY,
save_steps=SAVE_STEPS,
eval_steps=EVAL_STEPS,
optim=OPTIM,
dataloader_num_workers=DATALOADER_WORKERS,
bf16=False,
max_grad_norm=1.0,
report_to="none",
save_total_limit=0,
seed=42,
)
# ── Trainer ────────────────────────────────────────────────────────
# NOTA: DynamicAccumulationCallback(**dynacc_kwargs) — instância com ()
# O bloco inspect.signature foi removido: os parâmetros do callback são
# conhecidos e fixos; passar kwargs directamente é mais claro e seguro.
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=data_collator,
callbacks=[
LogCallback(),
DynamicAccumulationCallback(
base_batch=BASE_BATCH_SIZE,
max_steps=DYNAMIC_ACCUMULATION_MAX_STEPS,
),
SimpleStepTimerCallback(),
OptimizerRefreshCallback(refresh_interval_sec=OPTIMIZER_REFRESH_INTERVAL),
],
)
# --- ADD: resumo final com steps e accumulation/EB
try:
global TOTAL_TRAIN_STEPS
# tentar ler steps efetivos do Trainer ao final
steps_final = None
if hasattr(trainer, "state") and getattr(trainer.state, "global_step", None) is not None:
steps_final = trainer.state.global_step
elif hasattr(trainer, "state") and getattr(trainer.state, "max_steps", None):
steps_final = trainer.state.max_steps
if steps_final is not None and isinstance(steps_final, int) and steps_final > 0:
TOTAL_TRAIN_STEPS = steps_final # ajusta se necessário
origem = LAST_ACCUM_ORIGIN or "desconhecido"
base_bsz = BASE_BATCH_SIZE_EFFECTIVE if 'BASE_BATCH_SIZE_EFFECTIVE' in globals() and BASE_BATCH_SIZE_EFFECTIVE else BASE_BATCH_SIZE
eb = (base_bsz * CURRENT_ACCUM_STEPS) if CURRENT_ACCUM_STEPS else None
training_logs.append("[SUCESSO] Treino concluído (resumo):")
training_logs.append(f" - total_train_steps = {TOTAL_TRAIN_STEPS}")
if CURRENT_ACCUM_STEPS is not None:
training_logs.append(f" - gradient_accumulation_steps final = {CURRENT_ACCUM_STEPS} ({origem})")
if eb is not None:
training_logs.append(f" - Effective Batch final (base={base_bsz}) = {eb}")
except Exception as _e:
training_logs.append(f"[AVISO] não foi possível gerar resumo final detalhado ({_e})")
# --- END ADD
# Estimar total de steps antes do treino
try:
steps_estimados = trainer.args.max_steps
if steps_estimados <= 0:
steps_estimados = len(split_dataset["train"]) // (BASE_BATCH_SIZE * ACCUMULATION_STEPS) * int(epochs)
training_logs.append(f"[INFO] [{datetime.now().strftime('%H:%M:%S')}] A iniciar iteração do treino (0/{steps_estimados} passos estimados)...")
except Exception:
training_logs.append(f"[INFO] [{datetime.now().strftime('%H:%M:%S')}] A iniciar iteração do treino...")
# ── Log de acumulação pré-treino (se override manual activo) ───────
if CURRENT_ACCUM_STEPS is not None:
origem = LAST_ACCUM_ORIGIN or "desconhecido"
training_logs.append(
f"[INFO] gradient_accumulation_steps atual (pré-train): {CURRENT_ACCUM_STEPS} ({origem})"
)
eb = BASE_BATCH_SIZE_EFFECTIVE * CURRENT_ACCUM_STEPS
training_logs.append(f"[INFO] Effective Batch inicial (base={BASE_BATCH_SIZE_EFFECTIVE}) = {eb}")
#------------------------------------------------------------------------------------------------------------------
#INICIO DO TREINO E SALVAMENTO
#-----------------------------------------------------------------------------
try:
log_success("Configuração do Trainer concluída. A iniciar o treino...")
train_progress.update({
"status": "training",
"message": "Iniciando o treinamento do modelo..."
})
trainer_output = trainer.train(resume_from_checkpoint=resume_from_trainer_checkpoint)
# -------------------------------
# Salvar adapter LoRA localmente
# -------------------------------
lora_model_path = os.path.join(output_dir, "lora_model")
os.makedirs(lora_model_path, exist_ok=True)
model.save_pretrained(lora_model_path)
tokenizer.save_pretrained(lora_model_path)
training_logs.append("Adapter LoRA salvo com sucesso!")
# ================================================
# ENVIAR LoRA PARA BUCKET
# ================================================
# ---------------------------------
# Avaliação final do modelo
# ---------------------------------
log_info("A avaliar o modelo final no dataset de validação...")
final_metrics = trainer.evaluate()
eval_loss = final_metrics.get("eval_loss")
if eval_loss is not None:
perplexity = math.exp(eval_loss)
log_success("Avaliação Final Concluída:")
log_info(f" -> Eval Loss: {eval_loss:.4f}")
log_info(f" -> Perplexity: {perplexity:.4f}")
with open(os.path.join(output_dir, "final_metrics.txt"), "w") as f:
f.write(f"Eval Loss: {eval_loss}\nPerplexity: {perplexity}\n")
else:
log_warning("Não foi possível obter 'eval_loss' das métricas finais.")
train_progress.update({
"status": "awaiting_merge",
"percent": 100,
"message": "Treino concluído. LoRA salvo. Decida a próxima ação."
})
training_logs.append("Treino concluído. Adapter LoRA salvo em " + lora_model_path)
log_success("Treino concluído. Adapter LoRA salvo. Aguardando decisão (merge ou continuar).")
except RuntimeError as e:
msg = str(e)
training_logs.append(f"Erro durante treino: {msg}")
train_progress["status"] = "error"
log_error(f"Erro de Runtime durante o treino: {e}")
traceback.print_exc()
except Exception as e:
log_error(f"Erro crítico durante o treino: {e}")
train_progress.update({"status": "error", "message": f"Erro crítico no treino: {e}"})
traceback.print_exc()
return
#----------------------------------------------------
TENSORBOARD_LOGDIR = "trained_model_output\logs"
TENSORBOARD_PORT = 6006
def run_tensorboard():
# Lança o TensorBoard em thread separada para não bloquear
subprocess.Popen([
"tensorboard",
f"--logdir={TENSORBOARD_LOGDIR}",
f"--port={TENSORBOARD_PORT}"
], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
#-----------------------------------------------------------
#----------------------
# Rotas FLASK
#----------------------
@app.route('/')
def home():
""" Rota principal que serve o ficheiro index.html. """
return render_template('index-6.html')
#---------------------------------------
def guess_lora_targets(model):
"""
Função auxiliar para identificar automaticamente os módulos alvo para LoRA.
Analisa camadas lineares no modelo e retorna uma lista de nomes de camadas candidatas.
"""
targets = []
for name, module in model.named_modules():
if isinstance(module, torch.nn.Linear) and module.weight.requires_grad:
targets.append(name)
if not targets:
targets = ["q_proj", "v_proj"] # fallback padrão usado por PEFT
return targets
#-------------------------------------
# ROTA DE TREINAMENTO
#-------------------------------------
@app.route('/api/train', methods=['POST'])
def handle_train_request():
global all_data
try:
model_path = request.form.get('model_path')
epochs = int(request.form.get('epochs'))
uploaded_files = request.files.getlist('files')
logging.info(f"🚀 Novo treino iniciado")
logging.info(f"Modelo: {model_path}")
logging.info(f"Épocas: {epochs}")
if not all([model_path, epochs, uploaded_files]):
return jsonify({"status": "error", "message": "Faltam parâmetros: modelo, épocas ou ficheiros."}), 400
# Limpa os dados de treinos anteriores
all_data = []
for file in uploaded_files:
try:
raw_content = file.read()
encoding = chardet.detect(raw_content)['encoding'] or 'utf-8'
content = raw_content.decode(encoding)
for line in content.splitlines():
if not line.strip(): continue
if file.filename.endswith('.jsonl'):
all_data.append(json.loads(line))
else: # Assume .txt ou outros formatos de texto
all_data.append({"text": line.strip()})
except Exception as e:
return jsonify({"status": "error", "message": f"Erro ao ler ficheiro {file.filename}: {e}"}), 400
if not all_data:
return jsonify({"status": "error", "message": "Nenhum dado válido encontrado nos ficheiros."}), 400
output_dir = os.path.join(os.getcwd(), "trained_model_output")
os.makedirs(output_dir, exist_ok=True)
training_logs.append(f"[DEBUG] CWD: {os.getcwd()}")
training_logs.append(f"[DEBUG] OUTPUT_DIR: {output_dir}")
training_logs.append(f"[DEBUG] FILES: {os.listdir(os.getcwd())}")
# Passar um flag indicando que é um "novo" treino (ou re-treino)
thread = Thread(target=train_model_lora, args=(all_data, epochs, model_path, output_dir, ACCUMULATION_STEPS, DATALOADER_WORKERS, "new_train"))
thread.start()
return jsonify({"status": "started", "message": "Requisição de treino recebida. O processo foi iniciado."})
except Exception as e:
return jsonify({"status": "error", "message": f"Erro no servidor: {e}"}), 500
#-------------------------------------------
@app.route("/api/dataset/process", methods=["POST"])
def dataset_process():
global dataset_logs, dataset_status
file = request.files.get("file")
if not file or file.filename == "":
return jsonify({"error": "Nenhum ficheiro enviado."}), 400
# Parâmetros do formulário
schema = request.form.get("schema", "lite")
fmt = request.form.get("format", "jsonl")
chunk_size = int(request.form.get("chunk_size", 1200))
overlap = int(request.form.get("overlap", 150))
chunk_mode = request.form.get("chunk_mode", "chars")
clean_level = request.form.get("clean_level", "basic")
dedup = request.form.get("dedup", "0") == "1"
auto_pc = request.form.get("auto_pc", "0") == "1"
min_chars = int(request.form.get("min_chars", 0))
# Guardar ficheiro recebido em temp
import tempfile, threading
suffix = os.path.splitext(file.filename)[1].lower()
tmp = tempfile.NamedTemporaryFile(delete=False, suffix=suffix)
file.save(tmp.name)
tmp.close()
dataset_logs.clear()
dataset_status.update({"status": "processing", "records": 0, "message": "A iniciar..."})
def _run():
global dataset_logs, dataset_status
try:
from pdf_txt_to_dataset import (
process_pdf_to_records, process_txt_to_records,
process_docx_to_records, process_md_to_records,
process_jsonl_passthrough, process_mediawiki_xml_to_records,
write_jsonl, write_csv,
ensure_outdir, is_pdf, is_txt, is_docx, is_md, is_jsonl
)
import json
ensure_outdir(DATASET_OUT_DIR)
dataset_logs.append(f"[INFO] Ficheiro recebido: {file.filename}")
dataset_logs.append(f"[INFO] Schema: {schema} | Formato: {fmt} | Chunk: {chunk_size} | Overlap: {overlap}")
common = dict(
chunk_size=chunk_size, overlap=overlap, chunk_mode=chunk_mode,
schema=schema, clean_level=clean_level, dedup=dedup, min_chars=min_chars
)
fp = tmp.name
if is_pdf(fp):
recs, report, _ = process_pdf_to_records(fp, **common)
elif is_txt(fp):
recs, report, _ = process_txt_to_records(fp, **common)
elif is_docx(fp):
recs, report, _ = process_docx_to_records(fp, **common)
elif is_md(fp):
recs, report, _ = process_md_to_records(fp, **common)
elif is_jsonl(fp):
recs, report, _ = process_jsonl_passthrough(
fp, schema=schema, clean_level=clean_level,
dedup=dedup, min_chars=min_chars, auto_prompt_completion=auto_pc
)
elif fp.endswith(".xml") or file.filename.lower().endswith(".xml"):
max_wiki = int(request.form.get("max_wiki_articles", 0))
dataset_logs.append(f"[INFO] Wiki XML detectado — max_articles={max_wiki or 'sem limite'}")
recs, report, _ = process_mediawiki_xml_to_records(
fp, max_articles=max_wiki, log_fn=dataset_logs.append, **common
)
else:
dataset_logs.append(f"[ERRO] Formato não suportado: {suffix}")
dataset_status.update({"status": "error", "message": "Formato não suportado."})
return
if not recs:
dataset_logs.append("[WARN] Nenhum registo extraído. Verifica o ficheiro ou parâmetros.")
dataset_status.update({"status": "error", "message": "Sem registos."})
return
# Guardar
base_name = os.path.splitext(file.filename)[0]
out_path = os.path.join(DATASET_OUT_DIR, f"{base_name}.{fmt}")
if fmt == "jsonl":
write_jsonl(recs, out_path)
else:
write_csv(recs, out_path)
# Relatório
report_path = os.path.join(DATASET_OUT_DIR, f"{base_name}_report.json")
with open(report_path, "w", encoding="utf-8") as f_rep:
json.dump(report, f_rep, ensure_ascii=False, indent=2)
dataset_logs.append(f"[OK] {len(recs)} registos extraídos.")
dataset_logs.append(f"[OK] Dataset gravado em: {out_path}")
dataset_logs.append(f"[OK] Relatório gravado em: {report_path}")
dataset_status.update({"status": "done", "records": len(recs), "message": "Concluído."})
except Exception as e:
import traceback
dataset_logs.append(f"[ERRO] {e}")
dataset_logs.append(traceback.format_exc())
dataset_status.update({"status": "error", "message": str(e)})
finally:
try:
os.remove(tmp.name)
except Exception:
pass
threading.Thread(target=_run, daemon=True).start()
return jsonify({"ok": True, "message": "A processar em background..."})
#-------------------------------------------
@app.route("/api/dataset/status")
def dataset_status_route():
logs_tail = dataset_logs[-200:] if len(dataset_logs) > 200 else dataset_logs[:]
return jsonify({
"status": dataset_status.get("status", "idle"),
"records": dataset_status.get("records", 0),
"message": dataset_status.get("message", ""),
"logs": logs_tail
})
#-------------------------------------------
@app.route("/api/dataset_dir")
def dataset_dir():
listing = []
if os.path.isdir(DATASET_OUT_DIR):
for fname in sorted(os.listdir(DATASET_OUT_DIR)):
fpath = os.path.join(DATASET_OUT_DIR, fname)
size_kb = os.path.getsize(fpath) // 1024
listing.append(f"{fname} ({size_kb} KB)")
else:
listing.append("(pasta ainda não criada)")
return jsonify({"listing": listing})
#-------------------------------------------
@app.route("/api/dataset/download/<path:filename>")
def dataset_download(filename):
"""Serve ficheiro do directório de output do dataset para download."""
import re
# Segurança: só permite nomes de ficheiro simples sem path traversal
safe_name = os.path.basename(filename)
file_path = os.path.join(DATASET_OUT_DIR, safe_name)
if not os.path.isfile(file_path):
return jsonify({"error": "File not found."}), 404
return send_file(file_path, as_attachment=True, download_name=safe_name)
#--------------------------------------------
@app.route("/api/train_status")
def train_status():
progress = train_progress.get("percent", 0)
status = train_progress.get("status", "")
epoch_list = epoch_losses
logs_tail = training_logs[-200:] if len(training_logs) > 200 else training_logs[:]
# Acrescentar as linhas pedidas (apenas texto; simples)
if CURRENT_ACCUM_STEPS is not None:
origem = LAST_ACCUM_ORIGIN or "desconhecido"
logs_tail.append(f"[INFO] gradient_accumulation_steps atual: {CURRENT_ACCUM_STEPS} ({origem})")
# opcional: mostrar Effective Batch
if BASE_BATCH_SIZE_EFFECTIVE:
eb = BASE_BATCH_SIZE_EFFECTIVE * CURRENT_ACCUM_STEPS
logs_tail.append(f"[INFO] Effective Batch (base={BASE_BATCH_SIZE_EFFECTIVE}) = {eb}")
if TOTAL_TRAIN_STEPS is not None:
logs_tail.append(f"[INFO] total_train_steps desta execução: {TOTAL_TRAIN_STEPS}")
return jsonify({
"progress": progress,
"status": status,
"epoch_losses": epoch_list,
"logs": logs_tail
})
#---------------------------------------
@app.route('/api/load_lora_chat', methods=['POST'])
def load_lora_chat():
try:
data = request.get_json()
model_path = data.get("model_path")
if not model_path or "C:/" in model_path:
return jsonify({"status": "error", "message": "Caminho do modelo inválido para Linux (HF Spaces). Use o ID do Hugging Face."}), 400
output_dir = os.path.join(os.getcwd(), "trained_model_output")
lora_model_path = os.path.join(output_dir, "lora_model")
if not os.path.exists(lora_model_path):
return jsonify({"status": "error", "message": f"LoRA não encontrado em: {lora_model_path}"}), 404
# Determinar dtype e device_map
if torch.cuda.is_available():
model_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
model_device_map = "auto"
else:
model_dtype = torch.bfloat16
model_device_map = "cpu"
# Carregar o tokenizer do LoRA primeiro para obter o tamanho de vocabulário correto
tokenizer = AutoTokenizer.from_pretrained(lora_model_path, use_fast=True)
base_model = AutoModelForCausalLM.from_pretrained(
model_path,
dtype=model_dtype,
device_map=model_device_map,
low_cpu_mem_usage=True
)
# CORREÇÃO: Redimensionar o modelo base para o tamanho do vocabulário do tokenizer do LoRA
base_model.resize_token_embeddings(len(tokenizer))
lora_model = PeftModel.from_pretrained(base_model, lora_model_path, device_map=model_device_map)
chat_model = lora_model
chat_tokenizer = tokenizer
log_success("Modelo LoRA carregado com sucesso para chat (sem merge).")
train_progress.update({"status": "awaiting_merge", "message": "Modelo LoRA carregado para chat. Decida a próxima ação."})
return jsonify({"status": "ok", "message": "Modelo LoRA carregado para chat (sem merge)."})
except Exception as e:
log_error(f"Erro ao carregar LoRA para chat: {e}")
train_progress.update({"status": "error", "message": f"Erro ao carregar LoRA para chat: {e}"})
return jsonify({"status": "error", "message": str(e)}), 500
#--------------------------------------------
@app.route('/api/chat', methods=['POST'])
def handle_chat():
global chat_model, chat_tokenizer
if not chat_model or not chat_tokenizer:
return jsonify({
"response": "ERRO: O modelo ainda não está carregado. Faça o 'merge' ou 'load_lora_chat' primeiro."
}), 400
try:
data = request.get_json()
prompt = data.get('prompt')
if not prompt:
return jsonify({"response": "Erro: 'prompt' não foi enviado."}), 400
formatted_prompt = f"<s>[INST] {prompt} [/INST]"
inputs = chat_tokenizer(formatted_prompt, return_tensors="pt").to(chat_model.device)
outputs = chat_model.generate(
**inputs,
max_new_tokens=256,
num_return_sequences=1,
eos_token_id=chat_tokenizer.eos_token_id,
pad_token_id=chat_tokenizer.eos_token_id,
do_sample=True,
temperature=0.01, # <--- Sugestão: Reduzir temperatura para respostas mais determinísticas
top_p=0.9,
)
response_full = chat_tokenizer.decode(outputs[0], skip_special_tokens=True)
response_only = response_full.split("[/INST]")[-1].strip()
return jsonify({"response": response_only})
except Exception as e:
return jsonify({"response": f"Erro interno no chat: {str(e)}"}), 500
#------------------------------------------------
@app.route('/api/decide_merge', methods=['POST'])
def decide_merge():
global chat_model, chat_tokenizer, train_progress
try:
decision = request.form.get('decision')
model_path = request.form.get('model_path')
output_dir = os.path.join(os.getcwd(), "trained_model_output")
lora_model_path = os.path.join(output_dir, "lora_model")
if torch.cuda.is_available():
model_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
model_device_map = "auto"
else:
model_dtype = torch.bfloat16
model_device_map = "cpu"
if decision == "merge":
log_info("Iniciando o merge do adapter LoRA com o modelo base...")
train_progress.update({"status": "merging", "message": "Iniciando a fusão do adapter LoRA..."})
# Carregar o tokenizer do LoRA primeiro para obter o tamanho de vocabulário correto
tokenizer = AutoTokenizer.from_pretrained(lora_model_path, use_fast=True)
base_model = AutoModelForCausalLM.from_pretrained(
model_path,
dtype=model_dtype,
device_map=model_device_map,
low_cpu_mem_usage=True
)
# CORREÇÃO: Redimensionar o modelo base para o tamanho do vocabulário do tokenizer do LoRA
base_model.resize_token_embeddings(len(tokenizer))
lora_model = PeftModel.from_pretrained(base_model, lora_model_path)
merged_model = lora_model.merge_and_unload()
final_model_path = os.path.join(output_dir, "merged_model")
os.makedirs(final_model_path, exist_ok=True)
merged_model.save_pretrained(final_model_path, safe_serialization=True)
# O tokenizer já foi carregado acima
tokenizer.save_pretrained(final_model_path)
chat_model = merged_model
chat_tokenizer = tokenizer
log_success("Modelo fundido com sucesso e pronto para o chat!")
train_progress.update({"status": "finished", "percent": 100, "message": "Modelo LoRA fundido e salvo."})
return jsonify({"status": "merged", "message": "Modelo fundido com sucesso."})
#-----------------------------------------------------------------------------------------
elif decision == "continue":
log_info("A continuar o treino por mais épocas a partir do adapter salvo...")
epochs_to_add = int(request.form.get('epochs', 1)) # Renomeado para clareza
if not os.path.exists(lora_model_path):
log_error("Nenhum adapter LoRA encontrado para retomar o treino.")
train_progress.update({"status": "continuing_training", "message": f"Continuando treino por mais {epochs_to_add} épocas..."})
return jsonify({"status": "error", "message": "Nenhum adapter LoRA encontrado para retomar o treino."}), 400
# --- Determinar épocas já completadas ---
initial_epochs_completed = 0
output_dir = os.path.join(os.getcwd(), "trained_model_output") # Certifique-se que output_dir está definido
checkpoint_dirs = [d for d in os.listdir(output_dir) if d.startswith("checkpoint-")]
if checkpoint_dirs:
checkpoint_dirs.sort(key=lambda x: int(x.split('-')[1]), reverse=True)
latest_candidate_checkpoint_dir = os.path.join(output_dir, checkpoint_dirs[0])
trainer_state_path = os.path.join(latest_candidate_checkpoint_dir, "trainer_state.json")
if os.path.exists(trainer_state_path):
try:
with open(trainer_state_path, 'r') as f:
trainer_state = json.load(f)
initial_epochs_completed = int(trainer_state.get('epoch', 0))
log_info(f"Detectado que {initial_epochs_completed} épocas foram completadas no último checkpoint.")
except Exception as e:
log_warning(f"Erro ao ler trainer_state.json do checkpoint: {e}. Assumindo 0 épocas completadas.")
# Atualiza o status de progresso para indicar que o treino vai continuar
train_progress.update({"status": "continuing_training", "message": f"Continuando treino por mais {epochs_to_add} épocas..."})
thread = Thread(
target=train_model_lora,
args=(
all_data,
epochs_to_add,
model_path,
output_dir,
ACCUMULATION_STEPS,
DATALOADER_WORKERS,
"continue_train",
lora_model_path,
initial_epochs_completed,
"continue_train"
)
)
thread.start()
return jsonify({
"status": "continued",
"message": f"Treino adicional de {epochs_to_add} épocas iniciado."
})
else:
log_warning(f"Decisão inválida para /decide_merge: {decision}")
train_progress.update({"status": "error", "message": f"Decisão inválida: {decision}"})
return jsonify({"status": "error", "message": "Decisão inválida."}), 400
except Exception as e:
log_error(f"Erro durante o merge/continuação: {e}")
train_progress.update({"status": "error", "message": f"Erro durante o merge/continuação: {e}"})
return jsonify({"status": "error", "message": str(e)}), 500
#--------------------------------------------------------------------
@app.post("/api/tensorboard/start")
def start_tensorboard():
threading.Thread(target=run_tensorboard, daemon=True).start()
return jsonify({"status": "ok", "url": f"http://localhost:{TENSORBOARD_PORT}"}), 200
# =========================================================================
# ROTAS DE DIRETORIO DO OUTPUT
# =========================================================================
@app.route('/api/model_dir', methods=['GET'])
def list_model_dir():
try:
# Tenta vários caminhos possíveis no Hugging Face
paths_to_check = [
os.path.join(os.getcwd(), "trained_model_output"),
"/app/trained_model_output",
"./trained_model_output"
]
target = paths_to_check[0]
for p in paths_to_check:
if os.path.exists(p):
target = p
break
# Se não existir, cria agora para o botão não falhar
if not os.path.exists(target):
os.makedirs(target, exist_ok=True)
return jsonify({"files": ["Pasta criada agora. Está vazia."]}), 200
files_list = []
for root, dirs, filenames in os.walk(target):
for f in filenames:
rel_path = os.path.relpath(os.path.join(root, f), target)
files_list.append(rel_path)
if not files_list:
return jsonify({"files": ["Pasta encontrada, mas não tem ficheiros lá dentro."]}), 200
return jsonify({"files": sorted(files_list)}), 200
except Exception as e:
return jsonify({"files": [f"Erro ao ler pasta: {str(e)}"]}), 200 # Retornamos 200 para o JS não dar erro
#-------------------------------------------------------------------------
@app.route('/api/download/<path:filename>')
def download_file(filename):
# Caminho absoluto para a pasta de output
directory = os.path.join(os.getcwd(), "trained_model_output")
# send_from_directory é seguro e evita que utilizadores acedam a pastas do sistema
try:
return send_from_directory(directory, filename, as_attachment=True)
except Exception as e:
return str(e), 404
# =========================================================================
# ROTAS DE CONFIGURAÇÃO (GET para Ler, POST para Guardar)
# =========================================================================
@app.route('/api/config', methods=['GET', 'POST'])
def api_config():
from config_manager import set_runtime_globals, _load_constants_from_file, map_backend_to_frontend
set_runtime_globals({
'logical_cpus': LOGICAL_CPUS,
'omp_threads': OMP_THREADS,
'dataloader_workers': DATALOADER_WORKERS,
'total_ram_gb': TOTAL_RAM_GB,
})
constants = _load_constants_from_file()
return jsonify(map_backend_to_frontend(constants))
# =========================================================================
# Endpoints
# =========================================================================
@app.route("/api/gguf/convert", methods=["POST"])
def gguf_convert():
"""
Inicia conversão do modelo merged para GGUF em background.
Body JSON: { "model_path": "...", "format": "f16" | "q8_0" }
"""
global gguf_logs, gguf_status
import threading
data = request.get_json(force=True) or {}
model_path = data.get("model_path", "").strip()
fmt = data.get("format", "f16").lower() # "f16" ou "q8_0"
if not model_path:
return jsonify({"error": "model_path é obrigatório."}), 400
if not os.path.isdir(model_path):
return jsonify({"error": f"Diretório não encontrado: {model_path}"}), 400
if fmt not in ("f16", "q8_0"):
return jsonify({"error": "format deve ser 'f16' ou 'q8_0'."}), 400
gguf_logs.clear()
gguf_status.update({"status": "processing", "message": "A iniciar conversão..."})
def _run():
global gguf_logs, gguf_status
import subprocess, sys
os.makedirs(GGUF_OUT_DIR, exist_ok=True)
model_name = os.path.basename(model_path.rstrip("/\\")) or "model"
out_filename = f"{model_name}_{fmt}.gguf"
out_path = os.path.join(GGUF_OUT_DIR, out_filename)
gguf_logs.append(f"[INFO] Modelo: {model_path}")
gguf_logs.append(f"[INFO] Formato: {fmt.upper()}")
gguf_logs.append(f"[INFO] Output: {out_path}")
try:
# ── Método 1: convert_hf_to_gguf.py do llama.cpp ──────────────
# Procura o script em localizações comuns
convert_script = _find_convert_script()
if convert_script:
gguf_logs.append(f"[INFO] Usando: {convert_script}")
cmd = [
sys.executable, convert_script,
model_path,
"--outfile", out_path,
"--outtype", fmt,
]
gguf_logs.append(f"[CMD] {' '.join(cmd)}")
result = subprocess.run(
cmd,
capture_output=True, text=True, timeout=3600
)
if result.stdout:
for line in result.stdout.splitlines():
gguf_logs.append(line)
if result.stderr:
for line in result.stderr.splitlines():
gguf_logs.append(f"[STDERR] {line}")
if result.returncode != 0:
raise RuntimeError(f"Conversão falhou (exit {result.returncode})")
else:
# ── Método 2: llama-cpp-python (se instalado) ──────────────
gguf_logs.append("[INFO] convert_hf_to_gguf.py não encontrado.")
gguf_logs.append("[INFO] A tentar llama_cpp...")
try:
from llama_cpp import llama_model_quantize_params, llama_model_quantize
gguf_logs.append("[WARN] llama_cpp encontrado mas API de quantização não suportada directamente.")
gguf_logs.append("[WARN] Instala o llama.cpp e coloca convert_hf_to_gguf.py no PATH.")
raise RuntimeError("Sem método de conversão disponível.")
except ImportError:
gguf_logs.append("[ERRO] llama-cpp-python não está instalado.")
gguf_logs.append("[INFO] Instala com: pip install llama-cpp-python")
gguf_logs.append("[INFO] Ou clona: https://github.com/ggerganov/llama.cpp")
raise RuntimeError("llama-cpp-python não instalado.")
if not os.path.isfile(out_path):
raise RuntimeError(f"Ficheiro GGUF não foi criado: {out_path}")
size_mb = os.path.getsize(out_path) / (1024 * 1024)
gguf_logs.append(f"[OK] Conversão concluída: {out_filename} ({size_mb:.1f} MB)")
gguf_status.update({"status": "done", "message": "Concluído."})
except Exception as e:
import traceback
gguf_logs.append(f"[ERRO] {e}")
gguf_logs.append(traceback.format_exc())
gguf_status.update({"status": "error", "message": str(e)})
threading.Thread(target=_run, daemon=True).start()
return jsonify({"ok": True, "message": "Conversão iniciada em background..."})
def _find_convert_script():
"""Procura convert_hf_to_gguf.py em localizações comuns."""
import shutil
candidates = [
"convert_hf_to_gguf.py", # mesmo diretório
os.path.join(os.path.dirname(__file__), "convert_hf_to_gguf.py"),
os.path.expanduser("~/llama.cpp/convert_hf_to_gguf.py"),
"/app/convert_hf_to_gguf.py", # HuggingFace Space
"/llama.cpp/convert_hf_to_gguf.py",
]
for path in candidates:
if os.path.isfile(path):
return path
# Tenta encontrar no PATH
found = shutil.which("convert_hf_to_gguf.py")
if found:
return found
return None
#-------------------------------------------
@app.route("/api/gguf/status")
def gguf_status_route():
logs_tail = gguf_logs[-200:] if len(gguf_logs) > 200 else gguf_logs[:]
return jsonify({
"status": gguf_status.get("status", "idle"),
"message": gguf_status.get("message", ""),
"logs": logs_tail,
})
#-------------------------------------------
@app.route("/api/gguf/dir")
def gguf_dir():
listing = []
if os.path.isdir(GGUF_OUT_DIR):
for fname in sorted(os.listdir(GGUF_OUT_DIR)):
fpath = os.path.join(GGUF_OUT_DIR, fname)
if os.path.isfile(fpath):
size_mb = os.path.getsize(fpath) / (1024 * 1024)
listing.append(f"{fname} ({size_mb:.1f} MB)")
else:
listing.append("(pasta ainda não criada)")
return jsonify({"listing": listing})
#-------------------------------------------
@app.route("/api/gguf/download/<path:filename>")
def gguf_download(filename):
"""Serve ficheiro GGUF para download."""
safe_name = os.path.basename(filename)
file_path = os.path.join(GGUF_OUT_DIR, safe_name)
if not os.path.isfile(file_path):
return jsonify({"error": "File not found."}), 404
return send_file(file_path, as_attachment=True, download_name=safe_name)
#------------------------------------
@app.post("/api/adjust_accumulation")
def adjust_accumulation():
"""
Ajuste manual relativo do gradient_accumulation_steps.
Usa o valor atual real do treino se CURRENT_ACCUM_STEPS não estiver definido.
Envia { "action": "auto" } para voltar ao modo automático.
"""
global CURRENT_ACCUM_STEPS
data = request.get_json(force=True)
delta = int(data.get("delta", 0))
action = data.get("action")
# Modo automático
if action == "auto":
CURRENT_ACCUM_STEPS = None
logging.info("🔄 Modo automático de accumulation reativado.")
return jsonify({"status": "ok", "mode": "auto", "message": "Modo automático reativado."})
# Se CURRENT_ACCUM_STEPS ainda não definido, usa valor atual do treino
if CURRENT_ACCUM_STEPS is None:
CURRENT_ACCUM_STEPS = train_progress.get("current_accum_steps", 16) # valor padrão caso não exista
# Aplica delta e limita
CURRENT_ACCUM_STEPS = max(1, min(DYNAMIC_ACCUMULATION_MAX_STEPS, CURRENT_ACCUM_STEPS + delta))
logging.info(f"🎚 Ajuste manual recebido → CURRENT_ACCUM_STEPS = {CURRENT_ACCUM_STEPS} (delta={delta})")
return jsonify({
"status": "ok",
"current_steps": CURRENT_ACCUM_STEPS,
"mode": "manual"
})
#----------------------------------------------------
@app.route('/api/create_model', methods=['POST'])
def create_model():
global model
data = request.json
model_name = data.get("model_name")
if not model_name:
return jsonify({"status": "error", "message": "model_name não fornecido."}), 400
# O modelo real é carregado no início do treino via train_model_lora
return jsonify({"status": "ok", "message": f"Modelo '{model_name}' será carregado ao iniciar o treino."})
#---------------------------------------------------------------
#if __name__ == "__main__" and os.environ.get("WERKZEUG_RUN_MAIN") == "true":
# código de inicialização
if __name__ == '__main__':
app.run(host='0.0.0.0', port=7860, debug=False) |