File size: 245,254 Bytes
07b428c 73e9225 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 28f242e 07b428c 5c1c396 07b428c 5c1c396 07b428c 28f242e 07b428c a034f4d 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 57f4b1d 07b428c 7803d72 57f4b1d 7803d72 28f242e 7803d72 28f242e 7803d72 28f242e 7803d72 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 7803d72 07b428c 7803d72 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 57f4b1d 28f242e 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 7803d72 57f4b1d 7803d72 07b428c 7803d72 07b428c 28f242e 07b428c 7803d72 07b428c 7803d72 07b428c 7803d72 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 57f4b1d 28f242e 57f4b1d 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 28f242e 57f4b1d 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 28f242e 07b428c a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 63a70a0 a034f4d 57f4b1d 07b428c 28f242e 57f4b1d 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 57f4b1d 07b428c 28f242e 07b428c 28f242e 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 28f242e 57f4b1d 07b428c 57f4b1d 28f242e 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 28f242e 07b428c 28f242e 07b428c e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 07b428c e0ba36a 07b428c e0ba36a 07b428c e0ba36a 07b428c e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e 57f4b1d e0ba36a 28f242e e0ba36a 28f242e 57f4b1d 28f242e e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e 07b428c 7803d72 07b428c e0ba36a 28f242e e0ba36a 28f242e e0ba36a 28f242e 07b428c 57f4b1d 28f242e 07b428c 28f242e 07b428c 7803d72 07b428c 28f242e 07b428c 7803d72 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 57f4b1d 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 28f242e 57f4b1d 28f242e 57f4b1d 28f242e 57f4b1d 28f242e 07b428c 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 7803d72 28f242e 07b428c 28f242e 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 28f242e 07b428c 28f242e 07b428c a034f4d 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 5c1c396 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 57f4b1d 07b428c 7803d72 07b428c 7803d72 07b428c 7803d72 07b428c 5c1c396 07b428c 57f4b1d 07b428c 28f242e | 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 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 | /* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* hexstate_quantize.c β HexState GGUF Quantizer
*
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* β HPC-Optimized GGUF Quantization Engine β
* β β
* β Architecture: HPCGraph Sensitivity Propagation β
* β Optimization: Complex Amplitude BP + MCMC Scale Search β
* β Enhancements: MSE Grid Search, Importance Matrix Weighting β
* β Output: GGUF v3 (Q2_K) β
* β β
* β "The weight and the quantized are opposite faces." β
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
*
* This tool adapts the HExState HPC Ouroboros factoring engine for
* LLM weight quantization. The core mathematical machinery is reused:
*
* Factoring Domain β Quantization Domain
* βββββββββββββββββββββββββββββββββββββββββββββββββ
* HPCGraph + CZ edges β Block sensitivity graph
* Complex Amplitude BP β Importance propagation
* MCMC period sampler β Optimal scale search
* try_period() validation β Error bound checking
* LLL lattice reduction β (future) Adaptive bit allocation
*
* Additional techniques ported from llm-compressor:
* MSE grid search β Optimal min/max range shrinking
* Importance matrix (imatrix) β Per-channel error weighting
*
* Build:
* make -f Makefile.quantize
*
* Usage:
* ./hexstate_quantize <input> <output.gguf> [options]
*
* Input can be:
* - A single .safetensors file
* - A model directory containing sharded .safetensors files
*
* Options:
* --optimizer hpc|mse|hybrid Scale optimization strategy (default: hybrid)
* --imatrix <file> Importance matrix for weighted quantization
* --verbose Per-block diagnostics
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#include <stdio.h>
#ifdef _OPENMP
#include <omp.h>
#endif
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <time.h>
#include <sys/stat.h>
#include <mpfr.h>
/* HExState headers β reused from the factoring engine */
#include "quhit_triality.h"
#include "hpc_graph.h"
#include "hpc_mobius.h"
#include "s6_exotic.h"
/* Quantization-specific headers */
#include "gguf_format.h"
#include "safetensors_reader.h"
#include "tokenizer_reader.h"
#include "imatrix_reader.h"
#define D 6 /* Preserved from HExState β the triality dimension */
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* OPTIMIZER MODE
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
typedef enum {
OPT_HPC, /* HExState BP only */
OPT_MSE, /* MSE grid search only */
OPT_HYBRID /* HPC sensitivity + MSE */
} OptimizerMode;
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* MODEL ARCHITECTURE AUTO-DETECTION
*
* Infers model architecture metadata from tensor names and shapes.
* Supports: LLaMA, Mistral, Qwen2, Phi-3, Gemma, GPT-NeoX, Falcon, DeepSeek
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
typedef struct {
char architecture[64]; /* "llama", "phi3", "gemma", etc. */
char name[256]; /* Human-readable model name */
uint32_t block_count; /* Number of transformer layers */
uint32_t embedding_length; /* Hidden dimension */
uint32_t head_count; /* Number of attention heads */
uint32_t head_count_kv; /* Number of KV heads (GQA) */
uint32_t vocab_size; /* Vocabulary size */
uint32_t context_length; /* Max context length (default) */
float rope_freq_base; /* RoPE frequency base */
uint32_t feed_forward_length; /* FFN intermediate size */
float rms_norm_eps; /* RMS norm epsilon */
int has_bias; /* Whether attention has biases */
int tie_word_embeddings; /* Whether output = embed_tokens */
} ModelArchitecture;
/* Count tensor names matching a pattern prefix */
static int count_tensors_with_prefix(const STMultiFile *mf, const char *prefix)
{
int count = 0;
int prefix_len = strlen(prefix);
for (int i = 0; i < mf->n_tensors; i++) {
if (strncmp(mf->tensor_map[i].name, prefix, prefix_len) == 0)
count++;
}
return count;
}
/* Find max layer index from tensor names like "model.layers.N.xxx" */
static int find_max_layer_index(const STMultiFile *mf, const char *layer_prefix)
{
int max_idx = -1;
int prefix_len = strlen(layer_prefix);
for (int i = 0; i < mf->n_tensors; i++) {
if (strncmp(mf->tensor_map[i].name, layer_prefix, prefix_len) == 0) {
int idx = atoi(mf->tensor_map[i].name + prefix_len);
if (idx > max_idx) max_idx = idx;
}
}
return max_idx;
}
/* ββ Config.json reader for definitive architecture parameters ββ */
typedef struct {
int valid;
uint32_t hidden_size;
uint32_t intermediate_size;
uint32_t num_attention_heads;
uint32_t num_key_value_heads;
uint32_t num_hidden_layers;
uint32_t vocab_size;
uint32_t max_position_embeddings;
float rope_theta;
float rms_norm_eps;
char model_type[64];
int tie_word_embeddings;
} ConfigJson;
static ConfigJson parse_config_json(const char *path)
{
ConfigJson cfg;
memset(&cfg, 0, sizeof(cfg));
FILE *f = fopen(path, "rb");
if (!f) return cfg;
fseek(f, 0, SEEK_END);
long size = ftell(f);
fseek(f, 0, SEEK_SET);
if (size <= 0) { fclose(f); return cfg; }
char *json = (char *)malloc((size_t)size + 1);
if (!json) { fclose(f); return cfg; }
size_t nread = fread(json, 1, (size_t)size, f);
json[nread] = '\0';
fclose(f);
if (nread == 0) { free(json); return cfg; }
cfg.valid = 1;
/* Simple key-value extraction */
const char *p;
p = tok_find_key(json, "hidden_size");
if (p) cfg.hidden_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "intermediate_size");
if (p) cfg.intermediate_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "num_attention_heads");
if (p) cfg.num_attention_heads = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "num_key_value_heads");
if (p) cfg.num_key_value_heads = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "num_hidden_layers");
if (p) cfg.num_hidden_layers = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "vocab_size");
if (p) cfg.vocab_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "max_position_embeddings");
if (p) cfg.max_position_embeddings = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(json, "rope_theta");
if (p) cfg.rope_theta = (float)strtod(p, NULL);
p = tok_find_key(json, "rms_norm_eps");
if (p) cfg.rms_norm_eps = (float)strtod(p, NULL);
p = tok_find_key(json, "model_type");
if (p && *p == '"') {
char buf[64];
tok_extract_string(p, buf, sizeof(buf));
strncpy(cfg.model_type, buf, sizeof(cfg.model_type) - 1);
}
p = tok_find_key(json, "tie_word_embeddings");
if (p) cfg.tie_word_embeddings = (strncmp(p, "true", 4) == 0);
/* ββ Qwen 3.5/3.6: parameters are nested inside "text_config" ββ */
if (cfg.hidden_size == 0) {
const char *tc = strstr(json, "\"text_config\"");
if (tc) {
const char *tc_brace = strchr(tc, '{');
if (tc_brace) {
p = tok_find_key(tc_brace, "hidden_size");
if (p) cfg.hidden_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "intermediate_size");
if (p) cfg.intermediate_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "num_attention_heads");
if (p) cfg.num_attention_heads = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "num_key_value_heads");
if (p) cfg.num_key_value_heads = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "num_hidden_layers");
if (p) cfg.num_hidden_layers = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "vocab_size");
if (p) cfg.vocab_size = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "max_position_embeddings");
if (p) cfg.max_position_embeddings = (uint32_t)strtol(p, NULL, 10);
p = tok_find_key(tc_brace, "rms_norm_eps");
if (p) cfg.rms_norm_eps = (float)strtod(p, NULL);
p = tok_find_key(tc_brace, "model_type");
if (p && *p == '"') {
char buf2[64];
tok_extract_string(p, buf2, sizeof(buf2));
strncpy(cfg.model_type, buf2, sizeof(cfg.model_type) - 1);
}
p = tok_find_key(tc_brace, "tie_word_embeddings");
if (p) cfg.tie_word_embeddings = (strncmp(p, "true", 4) == 0);
/* Qwen3.6 rope_theta is nested in rope_parameters */
const char *rp = strstr(tc_brace, "\"rope_parameters\"");
if (rp) {
p = tok_find_key(rp, "rope_theta");
if (p) cfg.rope_theta = (float)strtod(p, NULL);
}
}
}
}
free(json);
return cfg;
}
static void detect_architecture(const STMultiFile *mf, ModelArchitecture *arch,
const char *config_json_path)
{
memset(arch, 0, sizeof(*arch));
/* Default values */
strcpy(arch->architecture, "llama");
strcpy(arch->name, "HExState-quantized");
arch->context_length = 4096;
arch->rope_freq_base = 10000.0f;
arch->rms_norm_eps = 1e-5f;
/* ββ Try config.json for definitive parameters ββ */
ConfigJson cfg = {0};
if (config_json_path) {
cfg = parse_config_json(config_json_path);
}
if (cfg.valid) {
/* Map model_type to GGUF architecture name */
if (strcmp(cfg.model_type, "llama") == 0 ||
strcmp(cfg.model_type, "mistral") == 0) {
strcpy(arch->architecture, "llama");
} else if (strcmp(cfg.model_type, "qwen2") == 0) {
strcpy(arch->architecture, "qwen2");
} else if (strcmp(cfg.model_type, "qwen2_moe") == 0) {
strcpy(arch->architecture, "qwen2moe");
} else if (strcmp(cfg.model_type, "qwen3_5") == 0 ||
strcmp(cfg.model_type, "qwen3_5_text") == 0 ||
strcmp(cfg.model_type, "qwen3_5_moe") == 0) {
strcpy(arch->architecture, "qwen2"); /* GGUF arch: qwen2 compat */
} else if (strcmp(cfg.model_type, "phi3") == 0 ||
strcmp(cfg.model_type, "phi") == 0) {
strcpy(arch->architecture, "phi3");
} else if (strcmp(cfg.model_type, "gemma") == 0 ||
strcmp(cfg.model_type, "gemma2") == 0) {
strcpy(arch->architecture, "gemma");
} else if (strcmp(cfg.model_type, "deepseek_v2") == 0) {
strcpy(arch->architecture, "llama");
} else if (strcmp(cfg.model_type, "gpt_neox") == 0) {
strcpy(arch->architecture, "gpt_neox");
} else if (strcmp(cfg.model_type, "falcon") == 0) {
strcpy(arch->architecture, "falcon");
} else if (cfg.model_type[0]) {
/* Unknown β try llama as fallback */
strcpy(arch->architecture, "llama");
}
if (cfg.hidden_size) arch->embedding_length = cfg.hidden_size;
if (cfg.intermediate_size) arch->feed_forward_length = cfg.intermediate_size;
if (cfg.num_attention_heads) arch->head_count = cfg.num_attention_heads;
if (cfg.num_key_value_heads) arch->head_count_kv = cfg.num_key_value_heads;
if (cfg.num_hidden_layers) arch->block_count = cfg.num_hidden_layers;
if (cfg.vocab_size) arch->vocab_size = cfg.vocab_size;
if (cfg.max_position_embeddings) arch->context_length = cfg.max_position_embeddings;
if (cfg.rope_theta > 0) arch->rope_freq_base = cfg.rope_theta;
if (cfg.rms_norm_eps > 0) arch->rms_norm_eps = cfg.rms_norm_eps;
arch->tie_word_embeddings = cfg.tie_word_embeddings;
printf(" Architecture determined from config.json: %s\n", cfg.model_type);
}
/* ββ Fall back to tensor name pattern detection ββ */
int has_model_layers = count_tensors_with_prefix(mf, "model.layers.");
int has_gpt_neox = count_tensors_with_prefix(mf, "gpt_neox.");
int has_transformer = count_tensors_with_prefix(mf, "transformer.");
/* Architecture-specific detection */
int has_qkv_proj = count_tensors_with_prefix(mf, "model.layers.0.self_attn.qkv_proj");
int has_kv_a_proj = count_tensors_with_prefix(mf, "model.layers.0.self_attn.kv_a_proj_with_mqa");
int has_final_norm = (st_multi_find_tensor(mf, "model.final_norm.weight") >= 0);
if (has_qkv_proj > 0 && !cfg.valid) {
strcpy(arch->architecture, "phi3");
} else if (has_kv_a_proj > 0 && !cfg.valid) {
strcpy(arch->architecture, "llama"); /* DeepSeek uses llama arch */
} else if (has_final_norm && !cfg.valid) {
strcpy(arch->architecture, "gemma");
}
if (has_model_layers > 0 && arch->block_count == 0) {
arch->block_count = find_max_layer_index(mf, "model.layers.") + 1;
}
/* Infer dimensions from tensor shapes if not from config.json */
if (arch->embedding_length == 0 || arch->head_count == 0) {
int qproj_idx = st_multi_find_tensor(mf, "model.layers.0.self_attn.q_proj.weight");
int kproj_idx = st_multi_find_tensor(mf, "model.layers.0.self_attn.k_proj.weight");
if (qproj_idx >= 0) {
const STTensorInfo *ti = st_multi_tensor_info(mf, qproj_idx);
int64_t q_out = ti->shape[0];
int64_t hidden = ti->shape[1];
if (arch->embedding_length == 0) arch->embedding_length = hidden;
/* Try common head dimensions: 128, 64, 96 */
int head_dim = 128;
if (q_out % 128 == 0) head_dim = 128;
else if (q_out % 96 == 0) head_dim = 96;
else if (q_out % 64 == 0) head_dim = 64;
if (arch->head_count == 0) arch->head_count = q_out / head_dim;
if (kproj_idx >= 0 && arch->head_count_kv == 0) {
const STTensorInfo *kt = st_multi_tensor_info(mf, kproj_idx);
arch->head_count_kv = kt->shape[0] / head_dim;
}
}
}
if (arch->vocab_size == 0) {
int embed_idx = st_multi_find_tensor(mf, "model.embed_tokens.weight");
if (embed_idx >= 0) {
const STTensorInfo *ti = st_multi_tensor_info(mf, embed_idx);
arch->vocab_size = ti->shape[0];
}
}
if (arch->feed_forward_length == 0) {
int gate_idx = st_multi_find_tensor(mf, "model.layers.0.mlp.gate_proj.weight");
if (gate_idx >= 0) {
const STTensorInfo *ti = st_multi_tensor_info(mf, gate_idx);
arch->feed_forward_length = ti->shape[0];
} else {
int up_idx = st_multi_find_tensor(mf, "model.layers.0.mlp.up_proj.weight");
if (up_idx >= 0) {
const STTensorInfo *ti = st_multi_tensor_info(mf, up_idx);
arch->feed_forward_length = ti->shape[0];
}
}
}
/* Check for attention bias */
arch->has_bias = (st_multi_find_tensor(mf, "model.layers.0.self_attn.q_proj.bias") >= 0);
if (has_gpt_neox > 0 && arch->block_count == 0) {
strcpy(arch->architecture, "gpt_neox");
arch->block_count = find_max_layer_index(mf, "gpt_neox.layers.") + 1;
}
if (has_transformer > 0 && arch->block_count == 0) {
strcpy(arch->architecture, "falcon");
arch->block_count = find_max_layer_index(mf, "transformer.h.") + 1;
}
/* Fill in defaults for anything we couldn't detect */
if (arch->head_count == 0) arch->head_count = 32;
if (arch->head_count_kv == 0) arch->head_count_kv = arch->head_count;
if (arch->embedding_length == 0) arch->embedding_length = 4096;
if (arch->vocab_size == 0) arch->vocab_size = 32000;
if (arch->feed_forward_length == 0)
arch->feed_forward_length = (arch->embedding_length * 8) / 3; /* SwiGLU default */
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* TENSOR NAME MAPPING: HuggingFace β GGUF Standard
*
* Maps SafeTensors tensor names to the standardized GGUF naming
* convention used by llama.cpp for model loading.
*
* Enhanced with mappings for Phi-3, Gemma, DeepSeek, MoE, and bias tensors.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* Returns 1 if this tensor should be skipped (not written to GGUF) */
static int should_skip_tensor(const char *hf_name)
{
/* Rotary embeddings are computed at runtime, not stored */
if (strstr(hf_name, "rotary_emb.inv_freq") != NULL) return 1;
if (strstr(hf_name, "rotary_emb.cos_cached") != NULL) return 1;
if (strstr(hf_name, "rotary_emb.sin_cached") != NULL) return 1;
/* Qwen 3.6 vision encoder β skip all visual.* tensors */
if (strncmp(hf_name, "model.visual.", 13) == 0) return 1;
if (strncmp(hf_name, "visual.", 7) == 0) return 1;
/* MTP (multi-token prediction) layers β not needed for inference */
if (strstr(hf_name, "model.language_model.mtp_") != NULL) return 1;
return 0;
}
static void map_tensor_name(const char *hf_name, char *gguf_name, int buflen)
{
/* Start with identity mapping */
strncpy(gguf_name, hf_name, buflen - 1);
gguf_name[buflen - 1] = '\0';
/* Top-level mappings (common to all architectures) */
struct { const char *from; const char *to; } mappings[] = {
{"model.embed_tokens.weight", "token_embd.weight"},
{"model.language_model.embed_tokens.weight","token_embd.weight"}, /* Qwen 3.6 */
{"model.norm.weight", "output_norm.weight"},
{"model.language_model.norm.weight", "output_norm.weight"}, /* Qwen 3.6 */
{"model.final_norm.weight", "output_norm.weight"}, /* Gemma */
{"lm_head.weight", "output.weight"},
{"model.embed_tokens.bias", "token_embd.bias"},
{"model.norm.bias", "output_norm.bias"},
{NULL, NULL}
};
for (int m = 0; mappings[m].from; m++) {
if (strcmp(hf_name, mappings[m].from) == 0) {
strncpy(gguf_name, mappings[m].to, buflen - 1);
return;
}
}
/* Layer mappings: "model.layers.N.xxx" or "model.language_model.layers.N.xxx" β "blk.N.xxx" */
const char *layer_prefix = NULL;
if (strncmp(hf_name, "model.layers.", 13) == 0)
layer_prefix = hf_name + 13;
else if (strncmp(hf_name, "model.language_model.layers.", 27) == 0)
layer_prefix = hf_name + 27;
if (layer_prefix) {
int layer_idx;
char rest[ST_MAX_NAME_LEN];
if (sscanf(layer_prefix, "%d.%255s", &layer_idx, rest) == 2) {
/* Map sublayer names */
struct { const char *from; const char *to; } layer_maps[] = {
/* Standard attention projections */
{"self_attn.q_proj.weight", "attn_q.weight"},
{"self_attn.k_proj.weight", "attn_k.weight"},
{"self_attn.v_proj.weight", "attn_v.weight"},
{"self_attn.o_proj.weight", "attn_output.weight"},
/* Attention biases */
{"self_attn.q_proj.bias", "attn_q.bias"},
{"self_attn.k_proj.bias", "attn_k.bias"},
{"self_attn.v_proj.bias", "attn_v.bias"},
{"self_attn.o_proj.bias", "attn_output.bias"},
/* Phi-3 fused QKV */
{"self_attn.qkv_proj.weight", "attn_qkv.weight"},
{"self_attn.qkv_proj.bias", "attn_qkv.bias"},
/* DeepSeek MLA */
{"self_attn.kv_a_proj_with_mqa.weight", "attn_kv_a_mqa.weight"},
{"self_attn.kv_b_proj.weight", "attn_kv_b.weight"},
/* Standard FFN (SwiGLU) */
{"mlp.gate_proj.weight", "ffn_gate.weight"},
{"mlp.up_proj.weight", "ffn_up.weight"},
{"mlp.down_proj.weight", "ffn_down.weight"},
/* FFN biases */
{"mlp.gate_proj.bias", "ffn_gate.bias"},
{"mlp.up_proj.bias", "ffn_up.bias"},
{"mlp.down_proj.bias", "ffn_down.bias"},
/* MoE gate */
{"mlp.gate.weight", "ffn_gate_inp.weight"},
/* MoE expert weights */
{"mlp.experts.gate_proj.weight", "ffn_gate_exps.weight"},
{"mlp.experts.up_proj.weight", "ffn_up_exps.weight"},
{"mlp.experts.down_proj.weight", "ffn_down_exps.weight"},
/* Norm layers */
{"input_layernorm.weight", "attn_norm.weight"},
{"post_attention_layernorm.weight", "ffn_norm.weight"},
{"input_layernorm.bias", "attn_norm.bias"},
{"post_attention_layernorm.bias", "ffn_norm.bias"},
/* Gemma pre/post feedforward norm */
{"pre_feedforward_layernorm.weight", "ffn_norm.weight"},
{"post_feedforward_layernorm.weight", "ffn_post_norm.weight"},
/* Qwen 3.6 full attention QK norms */
{"self_attn.q_norm.weight", "attn_q_norm.weight"},
{"self_attn.k_norm.weight", "attn_k_norm.weight"},
/* Qwen 3.6 DeltaNet (Gated Linear Attention) */
{"linear_attn.in_proj_qkv.weight", "ssm_in_qkv.weight"},
{"linear_attn.in_proj_z.weight", "ssm_in_z.weight"},
{"linear_attn.in_proj_a.weight", "ssm_in_a.weight"},
{"linear_attn.in_proj_b.weight", "ssm_in_b.weight"},
{"linear_attn.out_proj.weight", "ssm_out.weight"},
{"linear_attn.conv1d.weight", "ssm_conv1d.weight"},
{"linear_attn.norm.weight", "ssm_norm.weight"},
{"linear_attn.A_log", "ssm_a"},
{"linear_attn.dt_bias", "ssm_dt.bias"},
{NULL, NULL}
};
for (int m = 0; layer_maps[m].from; m++) {
if (strcmp(rest, layer_maps[m].from) == 0) {
snprintf(gguf_name, buflen, "blk.%d.%s",
layer_idx, layer_maps[m].to);
return;
}
}
/* MoE expert layer mapping: model.layers.N.mlp.experts.E.xxx */
int expert_idx;
char expert_rest[ST_MAX_NAME_LEN];
if (sscanf(rest, "mlp.experts.%d.%255s", &expert_idx, expert_rest) == 2) {
struct { const char *from; const char *to; } expert_maps[] = {
{"gate_proj.weight", "ffn_gate_exp.weight"},
{"up_proj.weight", "ffn_up_exp.weight"},
{"down_proj.weight", "ffn_down_exp.weight"},
{NULL, NULL}
};
for (int m = 0; expert_maps[m].from; m++) {
if (strcmp(expert_rest, expert_maps[m].from) == 0) {
snprintf(gguf_name, buflen, "blk.%d.%s.%d",
layer_idx, expert_maps[m].to, expert_idx);
return;
}
}
}
/* Fallback: keep original sub-path */
snprintf(gguf_name, buflen, "blk.%d.%s", layer_idx, rest);
}
}
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* SHOULD THIS TENSOR BE QUANTIZED?
*
* Decision rules:
* - Quantize: weight matrices (2D, large)
* - Keep F32: norms, biases, embeddings, 1D tensors
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static int should_quantize(const STTensorInfo *ti, const char *gguf_name)
{
/* Never quantize 1D tensors (norms, biases) */
if (ti->n_dims < 2) return 0;
/* Never quantize embedding tables (row dimension = vocab) */
if (strstr(gguf_name, "token_embd") != NULL) return 0;
/* Never quantize LM head output β use exact match, not substring,
* to avoid matching "attn_output.weight" */
if (strcmp(gguf_name, "output.weight") == 0) return 0;
/* Never quantize norm weights */
if (strstr(gguf_name, "norm") != NULL) return 0;
/* Never quantize bias tensors */
if (strstr(gguf_name, ".bias") != NULL) return 0;
/* Never quantize MoE gate routing weights */
if (strstr(gguf_name, "ffn_gate_inp") != NULL) return 0;
/* Never quantize DeltaNet state-space parameters (1D or small) */
if (strstr(gguf_name, "ssm_a") != NULL) return 0; /* A_log */
if (strstr(gguf_name, "ssm_dt") != NULL) return 0; /* dt_bias */
if (strstr(gguf_name, "ssm_conv1d") != NULL) return 0; /* conv kernel */
/* Quantize everything else (attention projections, FFN weights, SSM projections) */
return 1;
}
/* Detect attention Q/K/V/O projection tensors.
* These are the most sensitive to quantization β errors in attention scores
* cascade through the entire sequence, causing self-correction loops.
* Promoting these to Q4_0 (~4.5bpw) doubles their precision. */
static int is_attention_tensor(const char *gguf_name)
{
/* Gemma / LLaMA style GGUF names: blk.N.attn_q/k/v/output.weight */
if (strstr(gguf_name, "attn_q.weight") != NULL) return 1;
if (strstr(gguf_name, "attn_k.weight") != NULL) return 1;
if (strstr(gguf_name, "attn_v.weight") != NULL) return 1;
if (strstr(gguf_name, "attn_output.weight") != NULL) return 1;
if (strstr(gguf_name, "attn_qkv.weight") != NULL) return 1;
/* Qwen 3.6 DeltaNet SSM projections β treat as attention-class (Q4_0) */
if (strstr(gguf_name, "ssm_in_qkv.weight") != NULL) return 1;
if (strstr(gguf_name, "ssm_in_z.weight") != NULL) return 1;
if (strstr(gguf_name, "ssm_out.weight") != NULL) return 1;
/* HuggingFace style (fallthrough names) */
if (strstr(gguf_name, "self_attn.q_proj.weight") != NULL) return 1;
if (strstr(gguf_name, "self_attn.k_proj.weight") != NULL) return 1;
if (strstr(gguf_name, "self_attn.v_proj.weight") != NULL) return 1;
if (strstr(gguf_name, "self_attn.o_proj.weight") != NULL) return 1;
return 0;
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* HPC SENSITIVITY GRAPH BUILDER
*
* Creates an HPCGraph where each node represents a weight block.
* For Q2_K: 256-weight superblocks.
*
* The 6 values per site correspond to 6 candidate scale factors:
* v=0: scale * 0.85 (aggressive, high compression)
* v=1: scale * 0.90
* v=2: scale * 0.95
* v=3: scale * 1.00 (standard)
* v=4: scale * 1.05
* v=5: scale * 1.10 (conservative, less compression error)
*
* BP propagates: "if your neighbor block is sensitive, you should be
* conservative too" β creating coherent precision allocation.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* ββ Multi-quhit expanded scale table ββ
* Search grid: 24Γ24 = 576 (d, dmin) candidates
* Quhit encoding: bin 24 β 6 for D=6 quhits (BP operates on 6-state marginals)
* Beam search: operates on all 576 candidates directly */
#define QUHITS_PER_BLOCK 2
#define N_CAND_D 24 /* d multiplier candidates (expanded) */
#define N_CAND_M 24 /* dmin multiplier candidates (expanded) */
#define TOTAL_SCALE_CANDIDATES (N_CAND_D * N_CAND_M)
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* EXPERIMENTAL / CURRENTLY-UNUSED CODE PATHS
*
* Nothing in the live pipeline calls the legacy BP sensitivity graph
* (build_sensitivity_graph + compute_block_error_q2k + SCALE_TABLE) or the
* llm-compressor MSE grid search (mse_grid_search_q2k_subblock); the Shor /
* Viterbi path superseded them. They are preserved behind this flag instead
* of silently shipping as dead code that still costs an init pass.
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#ifdef HEXSTATE_ENABLE_EXPERIMENTAL
#define SCALE_FACTOR_COUNT 6
static const float SCALE_MULTIPLIERS[SCALE_FACTOR_COUNT] = {
0.60f, 0.75f, 0.90f, 1.00f, 1.15f, 1.40f
};
static float SCALE_TABLE[TOTAL_SCALE_CANDIDATES];
static int scale_table_initialized = 0;
static void init_scale_table(void) {
if (scale_table_initialized) return;
/* candidates: uniform spacing centered on 1.0 */
for (int i = 0; i < TOTAL_SCALE_CANDIDATES; i++) {
SCALE_TABLE[i] = 0.50f + (float)i * (1.00f / (float)(TOTAL_SCALE_CANDIDATES - 1));
}
scale_table_initialized = 1;
}
#endif /* HEXSTATE_ENABLE_EXPERIMENTAL */
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* THREAD-LOCAL HPCGRAPH REUSE β Eliminates 776K malloc/free cycles
*
* The sub-block Shor measurement uses a 16-node linear-chain graph that
* is identical in topology every time. Instead of hpc_create()/hpc_destroy()
* inside the OMP hot loop, we reset the same graph to a clean state.
*
* This function resets an existing HPCGraph with n_sites nodes to its
* initial state: clears all edges, resets adjacency lists, reinitializes
* locals. Zero allocations.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static void hpc_reset_for_subblock(HPCGraph *g, uint64_t n_sites)
{
/* Reset edge state */
g->n_edges = 0;
g->cz_edges = 0;
g->phase_edges = 0;
g->syntheme_edges = 0;
g->n_log = 0;
g->min_fidelity = 1.0;
g->avg_fidelity = 1.0;
g->amp_evals = 0;
g->prob_evals = 0;
g->measurements = 0;
/* Reset adjacency lists (just zero the counts, keep allocated buffers) */
for (uint64_t i = 0; i < n_sites; i++) {
g->adj[i].count = 0;
}
/* Reinitialize local quhit states */
for (uint64_t i = 0; i < n_sites; i++)
triality_init(&g->locals[i]);
}
#ifdef HEXSTATE_ENABLE_EXPERIMENTAL
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* FAST POWER APPROXIMATION β Replaces powf(x, 2.4f) in MSE grid search
*
* powf() costs ~50-100 cycles. Use log2f+exp2f (~25 cycles) for the
* exact x^2.4 = x^2 Γ 2^(0.4Β·log2(x)) computation instead.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static inline float fast_pow_2_4(float x)
{
/* x^2.4 = x^2 Γ 2^(0.4 Γ log2(x)). log2f+exp2f β 25 cycles total vs
* 50-100 for powf, and produces the exact ^2.4 norm the grid search needs. */
float x2 = x * x;
return x2 * exp2f(0.4f * log2f(x)); /* x^2 Γ x^0.4 = x^2.4 */
}
/* Compute the Q2_K sub-block reconstruction error for a block at a given
* scale multiplier, optionally weighted by importance vector */
static float compute_block_error_q2k(const float *weights, int block_size,
float scale_mult,
const float *importance, int imp_offset)
{
float min_val = weights[0];
float max_val = weights[0];
for (int j = 1; j < block_size; j++) {
if (weights[j] < min_val) min_val = weights[j];
if (weights[j] > max_val) max_val = weights[j];
}
if (min_val > 0) min_val = 0;
float range = (max_val - min_val) * scale_mult;
if (range < 1e-15f) return 0.0f;
float inv_range = 3.0f / range;
float err = 0.0f;
for (int j = 0; j < block_size; j++) {
float x = weights[j];
int q = (int)((x - min_val * scale_mult) * inv_range + 0.5f);
if (q < 0) q = 0; if (q > 3) q = 3;
float deq = min_val * scale_mult + (float)q * range / 3.0f;
float diff = x - deq;
float w = (importance) ? importance[imp_offset + j] : 1.0f;
err += diff * diff * w;
}
return err;
}
/* Build multi-quhit HPC sensitivity graph.
* 2 quhits per block β 576 scale candidates per block.
*
* Graph layout: sites [0..2*n-1] where:
* site 2*i = coarse quhit for block i
* site 2*i + 1 = fine quhit for block i
*
* Edges:
* Intra-block: CZ(2i, 2i+1) β coarseβfine coupling
* Inter-block: CZ(2i, 2(i+1)) β coarseβcoarse neighbor
* CZ(2i+1, 2(i+1)+1) β fineβfine neighbor */
static HPCGraph *build_sensitivity_graph(const float *weights,
int64_t n_elements,
int block_size,
float temperature,
const float *importance)
{
int64_t n_blocks = n_elements / block_size;
if (n_blocks < 2) return NULL;
init_scale_table();
int64_t graph_blocks = (n_blocks > 8192) ? 8192 : n_blocks;
int64_t stride = n_blocks / graph_blocks;
int64_t n_sites = graph_blocks * QUHITS_PER_BLOCK;
HPCGraph *graph = hpc_create(n_sites);
if (!graph) return NULL;
for (int64_t i = 0; i < n_sites; i++)
triality_dft(&graph->locals[i]);
/* Compute errors for all candidates per block,
* then project onto coarse (quhit 0) and fine (quhit 1) marginals */
for (int64_t i = 0; i < graph_blocks; i++) {
int64_t block_idx = i * stride;
const float *block_weights = weights + block_idx * block_size;
/* Evaluate all candidates */
float errors[TOTAL_SCALE_CANDIDATES];
float min_err = 1e30f;
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++) {
errors[c] = compute_block_error_q2k(block_weights, block_size,
SCALE_TABLE[c],
importance,
(int)(block_idx * block_size));
if (errors[c] < min_err) min_err = errors[c];
}
/* Project onto quhit 0 (coarse): marginalize over fine dimension
* amp_coarse[v0] = Ξ£_{v1} exp(-error(v0*6+v1) / 2T) */
double coarse_re[6], coarse_im[6];
double coarse_norm = 0.0;
for (int v0 = 0; v0 < 6; v0++) {
coarse_re[v0] = 0.0;
coarse_im[v0] = 0.0;
for (int v1 = 0; v1 < 6; v1++) {
int idx = v0 * 6 + v1;
coarse_re[v0] += exp(-(double)(errors[idx] - min_err) /
(2.0 * (double)temperature));
}
coarse_norm += coarse_re[v0] * coarse_re[v0];
}
if (coarse_norm > 1e-30) {
double inv = 1.0 / sqrt(coarse_norm);
for (int v = 0; v < 6; v++) coarse_re[v] *= inv;
}
/* Project onto quhit 1 (fine): marginalize over coarse dimension
* amp_fine[v1] = Ξ£_{v0} exp(-error(v0*6+v1) / 2T) */
double fine_re[6], fine_im[6];
double fine_norm = 0.0;
for (int v1 = 0; v1 < 6; v1++) {
fine_re[v1] = 0.0;
fine_im[v1] = 0.0;
for (int v0 = 0; v0 < 6; v0++) {
int idx = v0 * 6 + v1;
fine_re[v1] += exp(-(double)(errors[idx] - min_err) /
(2.0 * (double)temperature));
}
fine_norm += fine_re[v1] * fine_re[v1];
}
if (fine_norm > 1e-30) {
double inv = 1.0 / sqrt(fine_norm);
for (int v = 0; v < 6; v++) fine_re[v] *= inv;
}
/* Write coarse quhit (site 2*i) */
int64_t s_coarse = 2 * i;
for (int v = 0; v < 6; v++) {
graph->locals[s_coarse].edge_re[v] = coarse_re[v];
graph->locals[s_coarse].edge_im[v] = 0.0;
}
graph->locals[s_coarse].primary = VIEW_EDGE;
graph->locals[s_coarse].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[s_coarse].delta_valid = 0;
triality_update_mask(&graph->locals[s_coarse]);
/* Write fine quhit (site 2*i + 1) */
int64_t s_fine = 2 * i + 1;
for (int v = 0; v < 6; v++) {
graph->locals[s_fine].edge_re[v] = fine_re[v];
graph->locals[s_fine].edge_im[v] = 0.0;
}
graph->locals[s_fine].primary = VIEW_EDGE;
graph->locals[s_fine].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[s_fine].delta_valid = 0;
triality_update_mask(&graph->locals[s_fine]);
}
/* ββ Build edges ββ */
for (int64_t i = 0; i < graph_blocks; i++) {
/* Intra-block: coarse β fine coupling */
hpc_cz(graph, 2 * i, 2 * i + 1);
/* Inter-block: neighbor coupling */
if (i + 1 < graph_blocks) {
hpc_cz(graph, 2 * i, 2 * (i + 1)); /* coarse β coarse */
hpc_cz(graph, 2 * i + 1, 2 * (i + 1) + 1); /* fine β fine */
}
}
return graph;
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* MSE GRID SEARCH (ported from llm-compressor observers/mse.py)
*
* For a Q2_K sub-block, progressively shrink the min/max range to find
* the candidate that minimizes weighted reconstruction error.
*
* for p in [1.0, 1.0 - 1/grid, 1.0 - 2/grid, ...] down to (1 - maxshrink):
* candidate_min = p * min
* candidate_max = p * max
* error = ||x - quantize(x, candidate_min, candidate_max)||^norm
* if error < best: update best
* else: patience--; if patience == 0: break
*
* This is a direct C port of llm-compressor's _grid_search_mse.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
typedef struct {
float maxshrink; /* Maximum shrink factor (0.0 to 1.0) */
int grid; /* Number of grid divisions */
int patience; /* Early stopping patience */
float norm; /* Error norm exponent (2.0 = MSE, 2.4 = ...)*/
} MSEGridConfig;
static const MSEGridConfig MSE_DEFAULT_CONFIG = {
.maxshrink = 0.20f,
.grid = 200,
.patience = 8,
.norm = 2.4f
};
/* Grid search for optimal scale/min for a Q2_K sub-block of n weights
* with nmax = 3 quantization levels.
* Returns optimized scale; stores absolute min in *out_min.
* importance: per-element weights (can be NULL for uniform). */
static float mse_grid_search_q2k_subblock(const float *x, int n, int nmax,
uint8_t *L, float *out_min,
const float *importance,
const MSEGridConfig *cfg)
{
float min_val = x[0], max_val = x[0];
for (int i = 1; i < n; i++) {
if (x[i] < min_val) min_val = x[i];
if (x[i] > max_val) max_val = x[i];
}
if (max_val == min_val) {
for (int i = 0; i < n; i++) L[i] = 0;
*out_min = -min_val;
return 0.0f;
}
if (min_val > 0) min_val = 0;
float best_scale = 0.0f;
float best_min = -min_val;
float best_error = 1e30f;
int no_improve = 0;
int shrink_steps = (int)(cfg->maxshrink * cfg->grid);
if (shrink_steps < 1) shrink_steps = 1;
for (int step = 0; step <= shrink_steps; step++) {
float p = 1.0f - (float)step / (float)cfg->grid;
float cand_min = p * min_val;
float cand_max = p * max_val;
if (cand_max <= cand_min) continue;
float iscale = (float)nmax / (cand_max - cand_min);
float scale = 1.0f / iscale;
/* Quantize and measure error */
float err = 0.0f;
uint8_t tmp_L[256];
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale * (x[i] - cand_min));
if (l < 0) l = 0;
if (l > nmax) l = nmax;
tmp_L[i] = (uint8_t)l;
float deq = cand_min + scale * (float)l;
float diff = fabsf(x[i] - deq);
/* Apply error norm β fast path for default norm=2.4 */
float e = diff;
if (cfg->norm == 2.4f) {
e = fast_pow_2_4(diff);
} else if (cfg->norm != 1.0f) {
e = powf(diff, cfg->norm);
}
/* Apply importance weighting */
if (importance) e *= importance[i];
err += e;
}
if (err < best_error) {
best_error = err;
best_scale = scale;
best_min = -cand_min;
memcpy(L, tmp_L, n);
no_improve = 0;
} else {
no_improve++;
if (no_improve >= cfg->patience) break;
}
}
/* Iterative refinement on the best candidate (from ggml) */
float cur_min = -best_min;
float cur_scale = best_scale;
if (cur_scale > 1e-15f) {
float iscale = 1.0f / cur_scale;
for (int itry = 0; itry < 5; itry++) {
float sumlx = 0;
int suml2 = 0;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale * (x[i] - cur_min));
if (l < 0) l = 0;
if (l > nmax) l = nmax;
L[i] = (uint8_t)l;
sumlx += (x[i] - cur_min) * l;
suml2 += l * l;
}
if (suml2 > 0) cur_scale = sumlx / suml2;
float sum = 0;
for (int i = 0; i < n; i++)
sum += x[i] - cur_scale * L[i];
/* True coordinate-descent optimal: min* = sum/n (no momentum).
* Clamp to β€ 0 since min must be non-positive by convention. */
cur_min = fminf(0.0f, sum / n);
if (cur_scale > 1e-15f) iscale = 1.0f / cur_scale;
}
}
*out_min = -cur_min;
return cur_scale;
}
#endif /* HEXSTATE_ENABLE_EXPERIMENTAL */
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* HPC Q2_K QUANTIZATION β GGML-QUALITY + HPC REFINEMENT
*
* Two-phase approach:
* Phase A: Per-sub-block weighted least-squares (ggml make_qkx2_quants)
* This produces per-sub-block (scale, min) with 16-step search.
* Phase B: HPC BP refines the superblock-level d/dmin rounding.
* 6 candidate (d, dmin) pairs are tested; BP finds the one
* where the GLOBAL reconstruction error is minimized via
* constructive interference of per-sub-block phase coherence.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* Weighted least-squares quantization for a sub-block (ggml make_qkx2_quants).
* Finds optimal (scale, min) by searching 16 candidate iscale values
* and solving weighted least-squares for each.
* Returns scale; *the_min is set to the negative of the optimal min. */
static float hpc_make_qkx2_quants(int n, int nmax, const float *x,
const float *w, uint8_t *L,
float *the_min, uint8_t *Laux)
{
float xmin = x[0], xmax = x[0];
float sum_w = w[0], sum_x = w[0] * x[0];
for (int i = 1; i < n; i++) {
if (x[i] < xmin) xmin = x[i];
if (x[i] > xmax) xmax = x[i];
sum_w += w[i];
sum_x += w[i] * x[i];
}
if (xmin > 0) xmin = 0;
if (xmax == xmin) {
for (int i = 0; i < n; i++) L[i] = 0;
*the_min = -xmin;
return 0.0f;
}
float iscale = (float)nmax / (xmax - xmin);
float scale = 1.0f / iscale;
float best_mad = 0;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale * (x[i] - xmin));
if (l < 0) l = 0;
if (l > nmax) l = nmax;
L[i] = (uint8_t)l;
float diff = scale * (float)l + xmin - x[i];
best_mad += w[i] * fabsf(diff);
}
/* 16 candidate iscale values: search [-0.5, -0.5 + 0.1*15] + nmax */
for (int is = 0; is <= 15; is++) {
float try_iscale = (-0.5f + 0.1f * (float)is + (float)nmax) / (xmax - xmin);
float sl = 0, sl2 = 0, sxl = 0;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(try_iscale * (x[i] - xmin));
if (l < 0) l = 0;
if (l > nmax) l = nmax;
Laux[i] = (uint8_t)l;
sl += w[i] * (float)l;
sl2 += w[i] * (float)(l * l);
sxl += w[i] * (float)l * x[i];
}
float det = sum_w * sl2 - sl * sl;
if (det > 0) {
float this_scale = (sum_w * sxl - sum_x * sl) / det;
float this_min = (sl2 * sum_x - sl * sxl) / det;
if (this_min > 0) {
this_min = 0;
this_scale = sxl / sl2;
}
float mad = 0;
for (int i = 0; i < n; i++) {
float diff = this_scale * (float)Laux[i] + this_min - x[i];
mad += w[i] * fabsf(diff);
}
if (mad < best_mad) {
for (int i = 0; i < n; i++) L[i] = Laux[i];
best_mad = mad;
scale = this_scale;
xmin = this_min;
}
}
}
*the_min = -xmin;
return scale;
}
/* Quantize the scale/min arrays into 4-bit values: make_qp_quants equivalent.
* Returns the optimal d such that scales[j] β d Γ Ls[j]. */
static float hpc_make_qp_quants(int n, int nmax, const float *x,
uint8_t *L, const float *sw)
{
float xmax = 0;
for (int i = 0; i < n; i++)
if (x[i] > xmax) xmax = x[i];
if (xmax < 1e-15f) {
for (int i = 0; i < n; i++) L[i] = 0;
return 0.0f;
}
float iscale = (float)nmax / xmax;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale * x[i]);
if (l < 0) l = 0;
if (l > nmax) l = nmax;
L[i] = (uint8_t)l;
}
float scale = 1.0f / iscale;
float best_mse = 0;
for (int i = 0; i < n; i++) {
float diff = x[i] - scale * (float)L[i];
best_mse += sw[i] * diff * diff;
}
for (int is = -4; is <= 4; is++) {
if (is == 0) continue;
float iscale_is = (0.1f * (float)is + (float)nmax) / xmax;
float scale_is = 1.0f / iscale_is;
float mse = 0;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale_is * x[i]);
if (l < 0) l = 0;
if (l > nmax) l = nmax;
float diff = x[i] - scale_is * (float)l;
mse += sw[i] * diff * diff;
}
if (mse < best_mse) {
best_mse = mse;
iscale = iscale_is;
}
}
/* Recompute with best iscale + iterative refinement */
float sumlx = 0, suml2 = 0;
for (int i = 0; i < n; i++) {
int l = gguf_nearest_int(iscale * x[i]);
if (l < 0) l = 0;
if (l > nmax) l = nmax;
L[i] = (uint8_t)l;
sumlx += sw[i] * x[i] * (float)l;
suml2 += sw[i] * (float)(l * l);
}
/* Iterative greedy refinement */
for (int itry = 0; itry < 5; itry++) {
int n_changed = 0;
for (int i = 0; i < n; i++) {
float wi = sw[i];
float slx = sumlx - wi * x[i] * (float)L[i];
float sl2 = suml2 - wi * (float)(L[i] * L[i]);
if (slx > 0 && sl2 > 0) {
int new_l = gguf_nearest_int(x[i] * sl2 / slx);
if (new_l < 0) new_l = 0;
if (new_l > nmax) new_l = nmax;
if (new_l != L[i]) {
slx += wi * x[i] * (float)new_l;
sl2 += wi * (float)(new_l * new_l);
if (slx * slx * suml2 > sumlx * sumlx * sl2) {
L[i] = (uint8_t)new_l;
sumlx = slx;
suml2 = sl2;
n_changed++;
}
}
}
}
if (!n_changed) break;
}
return suml2 > 0 ? sumlx / suml2 : 0.0f;
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* SHOR'S GRIFFITHS-NIU SEQUENTIAL MEASUREMENT FOR RMSE OPTIMIZATION
* (Ported 1:1 from tesseract_factor.c β replaces BP)
*
* Instead of iterative message-passing (BP), this uses the EXACT sequential
* measurement protocol from Shor's algorithm:
*
* For each block k (MSB β LSB):
* 1. Compute feed-forward phase correction from previously measured blocks
* 2. Compute work factor: C_k(d) = Ξ _j Ξ£_w local_j(w) Γ edge(d,w)
* 3. Bake C_k into locals: Ξ±(d) *= C_k(d)
* 4. Apply phase correction: Ξ±(d) *= e^{-2Οi d ΞΈ_k}
* 5. Apply IDFT6 in-place: interference creates peaks at optimal scales
* 6. Born rule measurement β select optimal scale candidate
* 7. Collapse site + absorb edge weights into neighbors (back-action)
*
* This IS the quantum Fourier transform that creates constructive
* interference at the optimal RMSE configuration, exactly as Shor's
* algorithm creates interference at the correct period.
*
* Domain mapping:
* Factoring: oracle phase 2ΟΓdΓc_k/N β period r
* Quantize: error Boltzmann amplitudes β optimal RMSE block
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* Οβ roots of unity for CZ phase lookup come from hpc_graph.h
* (HPC_W6_RE / HPC_W6_IM) β the file-local duplicates were unused. */
static const double INV_SQRT6 = 0.40824829046386301637; /* 1/β6 */
/* ββ Collapse + Back-Action core (ported from tesseract_factor.c) ββ
* After sampling an outcome, collapse the target site to |outcomeβ©,
* absorb all edge weights into neighbor local states (Magic Pointer
* disentanglement), and remove dead edges from the graph.
*
* This is the EXACT same back-action protocol used in Shor's algorithm
* for the semi-classical QFT: measurement of one site conditions all
* remaining sites through the CZ phase correlations. */
static void shor_collapse_site(HPCGraph *graph, int target_site, int outcome)
{
/* Step 1: Collapse local state to |outcomeβ© */
for (int v = 0; v < 6; v++) {
graph->locals[target_site].edge_re[v] = (v == outcome) ? 1.0 : 0.0;
graph->locals[target_site].edge_im[v] = 0.0;
}
graph->locals[target_site].primary = VIEW_EDGE;
graph->locals[target_site].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[target_site].delta_valid = 0;
/* Step 2: Absorb edge weights into neighbor states (back-action).
* For each edge (target, neighbor), the weight w(outcome, d) for each
* neighbor basis state d gets multiplied into the neighbor's amplitude.
* This is the Magic Pointer disentanglement from tesseract_factor.c. */
HPCAdjList *adj = &graph->adj[target_site];
for (uint64_t ei = 0; ei < adj->count; ei++) {
uint64_t eid = adj->edge_ids[ei];
HPCEdge *edge = &graph->edges[eid];
uint64_t partner = (edge->site_a == (uint64_t)target_site) ?
edge->site_b : edge->site_a;
TrialityQuhit *pq = &graph->locals[partner];
for (int d = 0; d < 6; d++) {
double w_re, w_im;
if (edge->type == HPC_EDGE_CZ) {
int pidx = (outcome * d) % 6;
w_re = HPC_W6_RE[pidx];
w_im = HPC_W6_IM[pidx];
} else {
/* Weighted phase edge */
if (edge->site_a == (uint64_t)target_site) {
w_re = edge->w_re[outcome][d];
w_im = edge->w_im[outcome][d];
} else {
w_re = edge->w_re[d][outcome];
w_im = edge->w_im[d][outcome];
}
}
double old_re = pq->edge_re[d], old_im = pq->edge_im[d];
pq->edge_re[d] = old_re * w_re - old_im * w_im;
pq->edge_im[d] = old_re * w_im + old_im * w_re;
}
pq->dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
pq->delta_valid = 0;
}
/* Step 3: Remove edges touching this site from the graph.
* Mark by setting fidelity to -1 and remove from adj lists. */
for (uint64_t ei = 0; ei < adj->count; ei++) {
uint64_t eid = adj->edge_ids[ei];
HPCEdge *edge = &graph->edges[eid];
uint64_t partner = (edge->site_a == (uint64_t)target_site) ?
edge->site_b : edge->site_a;
/* Remove this edge from partner's adj list */
HPCAdjList *padj = &graph->adj[partner];
for (uint64_t pi = 0; pi < padj->count; pi++) {
if (padj->edge_ids[pi] == eid) {
padj->edge_ids[pi] = padj->edge_ids[--padj->count];
break;
}
}
edge->fidelity = -1.0; /* Mark as dead */
}
adj->count = 0; /* Clear target's adj list */
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* SHOR SEQUENTIAL MEASUREMENT β Griffiths-Niu Protocol for Quantization
*
* Ported 1:1 from tesseract_factor.c lines 2343-2500.
*
* Measures sites MSBβLSB. For each site k:
* 1. Compute feed-forward phase correction ΞΈ_k from previously measured sites
* 2. Compute neighbor contribution C_k(d) analytically
* 3. Bake C_k into locals
* 4. Apply phase correction: Ξ±(d) *= e^{-2Οi d ΞΈ_k}
* 5. Apply IDFT6: Ξ²(v) = (1/β6) Ξ£_d Ξ±'(d) Γ e^{2Οi dv/6}
* 6. Compute |Ξ²(v)|Β² as measurement probabilities
* 7. Sample/argmax β outcome
* 8. Collapse + back-action via shor_collapse_site()
*
* Returns: marginals are written into marg_out[n_sites][6].
* measured_out[n_sites] receives the measurement outcomes.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static void shor_measure_graph(HPCGraph *graph, int64_t n_sites,
double (*marg_out)[6], int *measured_out,
int deterministic)
{
/* Measure sites from last to first (MSBβLSB, same as Griffiths-Niu) */
for (int64_t k = n_sites - 1; k >= 0; k--) {
int site_k = (int)k;
/* Step 1: Compute feed-forward phase correction from previously
* measured sites. The QFT phase is 2Ο F x / 6^n. For site k,
* the fractional phase from previously measured site j (j > k)
* is measured_out[j] / 6^{j-k+1}.
* Power MUST start at 36.0 (6^2) for the immediately previous site. */
double theta_k = 0.0;
{
double power = 36.0;
for (int64_t j = k + 1; j < n_sites; j++) {
theta_k += (double)measured_out[j] / power;
power *= 6.0;
}
}
/* Step 2: Compute neighbor contribution C_k(d) analytically.
* C_k(d) = Ξ _neighbor Ξ£_{w=0}^{5} local_neighbor(w) Γ edge_weight(d, w)
* Each neighbor is independent (product state). */
double ck_re[6], ck_im[6];
for (int d = 0; d < 6; d++) { ck_re[d] = 1.0; ck_im[d] = 0.0; }
const HPCAdjList *adj = &graph->adj[site_k];
for (uint64_t ei = 0; ei < adj->count; ei++) {
uint64_t eid = adj->edge_ids[ei];
const HPCEdge *edge = &graph->edges[eid];
if (edge->fidelity < 0.0) continue; /* Skip dead edges */
uint64_t partner = (edge->site_a == (uint64_t)site_k) ?
edge->site_b : edge->site_a;
const TrialityQuhit *pq = &graph->locals[partner];
for (int d = 0; d < 6; d++) {
double sr = 0, si = 0;
for (int w = 0; w < 6; w++) {
double lr = pq->edge_re[w], li = pq->edge_im[w];
double wr, wi;
if (edge->type == HPC_EDGE_CZ) {
int pidx = (d * w) % 6;
wr = HPC_W6_RE[pidx]; wi = HPC_W6_IM[pidx];
} else if (edge->site_a == (uint64_t)site_k) {
wr = edge->w_re[d][w]; wi = edge->w_im[d][w];
} else {
wr = edge->w_re[w][d]; wi = edge->w_im[w][d];
}
sr += lr*wr - li*wi;
si += lr*wi + li*wr;
}
double nr = ck_re[d]*sr - ck_im[d]*si;
double ni = ck_re[d]*si + ck_im[d]*sr;
ck_re[d] = nr; ck_im[d] = ni;
}
}
/* Step 3: Bake C_k(d) into locals: Ξ±(d) *= C_k(d) */
for (int d = 0; d < 6; d++) {
double re = graph->locals[site_k].edge_re[d];
double im = graph->locals[site_k].edge_im[d];
graph->locals[site_k].edge_re[d] = re*ck_re[d] - im*ck_im[d];
graph->locals[site_k].edge_im[d] = re*ck_im[d] + im*ck_re[d];
}
/* Step 4: Apply feed-forward phase correction to locals. */
for (int d = 0; d < 6; d++) {
double angle = -2.0 * 3.14159265358979323846 * d * theta_k;
double pr = cos(angle), pi2 = sin(angle);
double re = graph->locals[site_k].edge_re[d];
double im = graph->locals[site_k].edge_im[d];
graph->locals[site_k].edge_re[d] = re*pr - im*pi2;
graph->locals[site_k].edge_im[d] = re*pi2 + im*pr;
}
/* Step 5: Apply IDFT6 in-place: phase basis β computational basis.
* Ξ²(v) = (1/β6) Ξ£_{d=0}^{5} Ξ±'(d) Γ e^{2Οi d v / 6}
* C_k(d) is INSIDE the coherent sum β THIS creates interference
* peaks at the optimal RMSE configuration, exactly as Shor's
* algorithm creates peaks at the correct period. */
{
double alpha_re[6], alpha_im[6];
for (int d = 0; d < 6; d++) {
alpha_re[d] = graph->locals[site_k].edge_re[d];
alpha_im[d] = graph->locals[site_k].edge_im[d];
}
for (int v = 0; v < 6; v++) {
double sum_re = 0.0, sum_im = 0.0;
for (int d = 0; d < 6; d++) {
double angle = 2.0 * 3.14159265358979323846 * d * v / 6.0;
double er = cos(angle), ei = sin(angle);
sum_re += alpha_re[d]*er - alpha_im[d]*ei;
sum_im += alpha_re[d]*ei + alpha_im[d]*er;
}
graph->locals[site_k].edge_re[v] = sum_re * INV_SQRT6;
graph->locals[site_k].edge_im[v] = sum_im * INV_SQRT6;
}
}
/* Step 6: Compute marginals from |local(v)|Β² */
double probs[6];
double total = 0.0;
for (int v = 0; v < 6; v++) {
probs[v] = graph->locals[site_k].edge_re[v] * graph->locals[site_k].edge_re[v] +
graph->locals[site_k].edge_im[v] * graph->locals[site_k].edge_im[v];
total += probs[v];
}
if (total > 1e-30) {
for (int v = 0; v < 6; v++) probs[v] /= total;
} else {
for (int v = 0; v < 6; v++) probs[v] = 1.0 / 6.0;
}
/* Store marginals for downstream beam search */
for (int v = 0; v < 6; v++)
marg_out[k][v] = probs[v];
/* Step 7: Select outcome β deterministic argmax for quantization
* (unlike factoring which uses Born sampling for probabilistic
* period recovery, quantization wants the MAP estimate) */
int outcome;
if (deterministic) {
outcome = 0;
double max_p = probs[0];
for (int v = 1; v < 6; v++) {
if (probs[v] > max_p) { max_p = probs[v]; outcome = v; }
}
} else {
/* Born sampling (for multi-shot refinement) */
static unsigned int shor_rng = 271828;
shor_rng = shor_rng * 1664525u + 1013904223u;
double r01 = (double)(shor_rng >> 8) / 16777216.0;
double cumul = 0.0;
outcome = 5;
for (int v = 0; v < 6; v++) {
cumul += probs[v];
if (r01 <= cumul) { outcome = v; break; }
}
}
measured_out[k] = outcome;
/* Step 8: Collapse + back-action β absorb edge weights into
* neighbor locals (Magic Pointer disentanglement) */
shor_collapse_site(graph, site_k, outcome);
}
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* HPC-OPTIMIZED Q4_0 QUANTIZATION (for attention tensors)
*
* Same architecture as Q2_K HPC pipeline, but simpler:
* - One parameter per block (scale d only, no dmin)
* - Single quhit per block (6 states)
* - 24 candidate scales β bin to 6 for BP
* - 48-beam Hensel search for globally optimal configuration
* - Triality 3-view marginals for robust scoring
*
* Q4_0 block: 32 weights, 16 levels (0β15), dequant: w = (q - 8) * d
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#define Q4_N_CAND 24 /* expanded scale candidates for Q4_0 */
#define Q4_N_BEAMS 48 /* expanded beam width */
/* Tight neighborhood around WLS optimum */
static const float Q4_NEIGHBOR_MULTS[Q4_N_CAND] = {
0.850f, 0.880f, 0.900f, 0.915f, 0.930f, 0.945f, 0.955f, 0.965f,
0.975f, 0.985f, 0.995f, 1.000f, 1.005f, 1.015f, 1.025f, 1.035f,
1.050f, 1.070f, 1.100f, 1.130f, 1.160f, 1.200f, 1.250f, 1.300f
};
static const int Q4_CAND_TO_QUHIT[Q4_N_CAND] = {
0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2,
3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5
};
/* ββ Candidate-selection error metric (shared by Q4_0 and Q2_K) ββ
* Candidates are now scored with the EXACT importance-weighted SSE
* err = Ξ£_i w_i Β· (x_i β deq_i)Β²
* which is the same objective the final assembly/polish phases minimise and
* the same quantity reported as RMSE. The previous 2-point Hadamard form
* (0.5Β·vesica + 0.5Β·wave with pair-AVERAGED weights) is algebraically equal
* to Ξ£ wΜΒ·(e_iΒ² + e_jΒ²), i.e. it silently replaced per-element importance
* weights with the pair mean β a systematic mis-weighting whenever an
* imatrix is supplied. Scoring candidates on a different objective than the
* one being optimised mis-ranks them; aligning the two strictly lowers the
* final weighted RMSE (and is bit-identical when no imatrix is used). */
/* ββ Cross-block prior override ratio ββ
* Q2_K and Q4_0 blocks are decoded INDEPENDENTLY by every GGUF runtime:
* there is no cross-block coupling in the dequantizer, so a smoothness
* prior that keeps a block on a worse candidate can only raise the true
* reconstruction RMSE. With 1.00f the per-block argmin over the candidate
* grid always wins (provably optimal seed for the assembly phase); the HPC
* graph/Viterbi/Born machinery still shapes ties and seeds the search.
* Set to e.g. 0.95f to restore the old 5%-hysteresis smoothness prior. */
#ifndef HEX_GREEDY_OVERRIDE_RATIO
#define HEX_GREEDY_OVERRIDE_RATIO 1.00f
#endif
/* fp16-ULP radius of the monotone (d, dmin) micro-search in the Phase-4.6
* polish (move 3). Larger radii let coordinate descent escape shallower
* local minima at O(radiusΒ²) extra cost per polish iteration. */
#ifndef HEX_POLISH_ULP
#define HEX_POLISH_ULP 4
#endif
/* ββ DC + vesica/wave extended objective (dot-product error cancellation) ββ
*
* The quantity that matters downstream is the layer-output error
* Ξ΅ = Ξ£α΅’ eα΅’Β·aα΅’, E[Ρ²] = eα΅Re, R = activation second-moment matrix.
* Modelling R with three components β per-channel power (diagonal, β
* imatrix), a common mean ΞΌ (rank-1), and correlation c across the
* half-block fold (i β i+n/2) β gives EXACTLY:
*
* E[Ρ²] β Ξ£α΅’ wα΅’eα΅’Β² + ΞΌΒ²Β·(Ξ£α΅’eα΅’)Β² + cΒ·Ξ£_pairs[(eα΅’+eβ±Ό)Β² β (eα΅’βeβ±Ό)Β²]
* βββ = vesicaΒ² β waveΒ² = 4Β·eα΅’eβ±Ό βββ
*
* The vesica/wave decomposition is therefore the natural basis of the
* fold-correlation term: in-phase (vesica) error energy COSTS output
* accuracy, anti-phase (wave) error energy is CREDITED β it cancels in
* the dot product. (The old 0.5/0.5 scorer ADDED the two, which collapses
* to plain SSE; the spectrally meaningful combination SUBTRACTS them.)
* Every selection/acceptance stage scores blocks with
*
* E(block) = Ξ£α΅’ wα΅’eα΅’Β²
* + (HEX_DC_LAMBDA / n) Β· (Ξ£α΅’eα΅’)Β²
* + (HEX_VW_LAMBDA / n) Β· Ξ£_{i<n/2} [(eα΅’+eβ±Ό)Β² β (eα΅’βeβ±Ό)Β²], j = i+n/2
*
* applied CONSISTENTLY to: Q2_K/Q4_0 candidate scoring, the closed-form
* (d, dmin) refit acceptance, the shaping accept guards, every polish
* move, and the Phase-4.7 floor β so no stage optimises a different
* objective than its acceptance test measures. The closed-form solvers
* incorporate the DC term as a rank-1 augmented observation and act as
* proposal generators; acceptance always uses the full extended E.
* Ξ» = 0 on both knobs reduces exactly to the pure weighted-SSE objective.
* Positive-definiteness: the fold coupling adds Β±2Ξ»_vw/n off-diagonal β
* negligible against any sane wα΅’, so E stays a valid quadratic objective.
* NOTE: reported RMSE stays pure reconstruction RMSE; with Ξ» > 0 a small
* RMSE increase is the *intended* price for lower output error. Per-block
* terms are a proxy for row-level structure (the API sees a flat stream);
* the Phase-3.9 rolling-DC pass handles cross-block linkage. */
#ifndef HEX_DC_LAMBDA
#define HEX_DC_LAMBDA 1.0f
#endif
#ifndef HEX_VW_LAMBDA
#define HEX_VW_LAMBDA 1.0f
#endif
/* Default (1, 1): unit-strength spectral prior. Empirically (synthetic
* benchmark, identical inputs): lowers dot-product output error ~0.8-1.4%
* on both mean-only and fold-correlated activation models for ~+0.05%
* weight RMSE. The theoretically optimal Ξ» grows with the deployment
* model's activation mean energy and row length (the per-block term
* under-counts cross-block row coupling); the synthetic sweep kept
* improving monotonically through Ξ» = 4 at ~+0.1% RMSE. Set both to
* 0.0f to recover the exact pure weighted-SSE / minimum-RMSE pipeline. */
/* Spectral penalty of the extended objective for one block: residuals e[n],
* fold at n/2. Negative values are possible (anti-phase credit) β the total
* E remains positive-definite as argued above. */
static inline float hex_spectral_penalty(const float *e, int n)
{
if (HEX_DC_LAMBDA == 0.0f && HEX_VW_LAMBDA == 0.0f) return 0.0f;
float dc = 0.0f, cross = 0.0f;
int half = n / 2;
for (int i = 0; i < half; i++) {
dc += e[i] + e[i + half];
cross += e[i] * e[i + half];
}
return (HEX_DC_LAMBDA / (float)n) * dc * dc
+ (HEX_VW_LAMBDA / (float)n) * 4.0f * cross;
}
static void quantize_tensor_q4_0_hpc(const float *weights, int64_t n_elements,
BlockQ4_0 *output, float *out_total_error,
const float *imat_importance, int verbose)
{
int64_t n_blocks = n_elements / QK4_0;
float total_err = 0.0f;
(void)verbose; /* kept for API symmetry with the Q2_K path */
/* ββ Phase 1: Greedy seed β compute scale per block ββ */
float *greedy_d = (float *)calloc(n_blocks, sizeof(float));
#pragma omp parallel for schedule(dynamic, 64)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bw = weights + blk * QK4_0;
float amax = 0.0f;
for (int j = 0; j < QK4_0; j++) {
float av = fabsf(bw[j]);
if (av > amax) amax = av;
}
greedy_d[blk] = amax / 7.0f;
}
/* ββ Phase 2: WLS-Optimal Candidate Generation for Q4_0 ββ
* First find the true optimal d* via 3-iteration WLS,
* then generate candidates centered on d* with tight spacing. */
float (*cand_errors)[Q4_N_CAND] = (float (*)[Q4_N_CAND])
calloc(n_blocks, sizeof(float[Q4_N_CAND]));
uint16_t (*cand_d16)[Q4_N_CAND] = (uint16_t (*)[Q4_N_CAND])
calloc(n_blocks, sizeof(uint16_t[Q4_N_CAND]));
#pragma omp parallel for schedule(dynamic, 64)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bw = weights + blk * QK4_0;
/* ββ Step 2a: WLS solve to find optimal d* ββ */
float wls_d = greedy_d[blk];
uint16_t prev_wls_d16 = 0;
for (int ls_iter = 0; ls_iter < 5; ls_iter++) {
if (wls_d < 1e-15f) break;
float inv_d = 1.0f / wls_d;
float num = 0.0f, den = 0.0f;
float dcS = 0.0f, dcQ = 0.0f; /* DC rank-1 augmentation sums */
for (int j = 0; j < QK4_0; j++) {
int q = (int)(bw[j] * inv_d + 8.5f);
if (q < 0) q = 0; if (q > 15) q = 15;
float qc = (float)q - 8.0f;
float w = (imat_importance) ?
imat_importance[blk * QK4_0 + j] : 1.0f;
num += w * bw[j] * qc;
den += w * qc * qc;
dcS += bw[j];
dcQ += qc;
}
/* DC term of the extended objective enters the normal equation
* as one extra observation (S ~ dΒ·Q) of weight Ξ»_dc/n. The
* vesica/wave term is handled by extended-E acceptance in the
* ULP search; the solver is a proposal generator. */
num += (HEX_DC_LAMBDA / (float)QK4_0) * dcS * dcQ;
den += (HEX_DC_LAMBDA / (float)QK4_0) * dcQ * dcQ;
if (den > 1e-15f) {
float d_new = num / den;
if (fabsf(d_new) < 4.0f * (greedy_d[blk] + 1e-10f))
wls_d = gguf_fp16_to_fp32(gguf_fp32_to_fp16(d_new));
}
uint16_t cur_wls_d16 = gguf_fp32_to_fp16(wls_d);
if (cur_wls_d16 == prev_wls_d16) break; /* converged in FP16 */
prev_wls_d16 = cur_wls_d16;
}
/* ββ Step 2b: Generate candidates centered on WLS optimum ββ */
for (int ci = 0; ci < Q4_N_CAND; ci++) {
float trial_d = wls_d * Q4_NEIGHBOR_MULTS[ci];
uint16_t d16 = gguf_fp32_to_fp16(trial_d);
float actual_d = gguf_fp16_to_fp32(d16);
cand_d16[blk][ci] = d16;
float id = (actual_d > 1e-15f) ? 1.0f / actual_d : 0.0f;
/* ββ Extended objective over all QK4_0 elements ββ
* Exact importance-weighted SSE + DC + vesica/wave spectral
* penalty β the same objective every acceptance stage uses. */
float err = 0.0f;
float e_arr[QK4_0];
for (int j = 0; j < QK4_0; j++) {
float x = bw[j];
int q = (int)(x * id + 8.5f);
if (q < 0) q = 0; if (q > 15) q = 15;
float deq = ((float)q - 8.0f) * actual_d;
float e = x - deq;
e_arr[j] = e;
float w = (imat_importance) ? imat_importance[blk * QK4_0 + j] : 1.0f;
err += e * e * w;
}
cand_errors[blk][ci] = err + hex_spectral_penalty(e_arr, QK4_0);
}
}
/* ββ Phase 3: HPC graph β single quhit per block ββ */
int *best_candidate = (int *)malloc(n_blocks * sizeof(int));
int hpc_ran_q4 = 0;
for (int64_t i = 0; i < n_blocks; i++)
best_candidate[i] = 11; /* Q4_NEIGHBOR_MULTS[11] = 1.00 */
if (n_blocks >= 2) {
float temperature = 0.5f;
int64_t graph_blocks = (n_blocks > 200) ? 200 : n_blocks;
int64_t stride = n_blocks / graph_blocks;
int64_t n_sites = graph_blocks; /* 1 quhit per block */
HPCGraph *graph = hpc_create(n_sites);
if (graph) {
hpc_ran_q4 = 1;
for (int64_t i = 0; i < n_sites; i++)
triality_dft(&graph->locals[i]);
/* Adaptive temperature from error landscape */
{
double err_accum = 0.0;
int err_count = 0;
for (int64_t gi = 0; gi < graph_blocks && gi < 100; gi++) {
int64_t blk = gi * stride;
float max_e = 0.0f;
for (int c = 0; c < Q4_N_CAND; c++)
if (cand_errors[blk][c] > max_e)
max_e = cand_errors[blk][c];
err_accum += (double)max_e;
err_count++;
}
if (err_count > 0) {
temperature = (float)(err_accum / err_count) * 0.1f;
if (temperature < 1e-10f) temperature = 1e-10f;
}
}
/* Encode stride-group AGGREGATED candidate errors as Boltzmann amplitudes */
for (int64_t i = 0; i < graph_blocks; i++) {
/* Aggregate errors across stride group */
float agg_errors[Q4_N_CAND];
for (int c = 0; c < Q4_N_CAND; c++)
agg_errors[c] = 0.0f;
int64_t blk_start = i * stride;
int64_t blk_end = blk_start + stride;
if (blk_end > n_blocks) blk_end = n_blocks;
int64_t group_size = blk_end - blk_start;
for (int64_t b = blk_start; b < blk_end; b++) {
for (int c = 0; c < Q4_N_CAND; c++)
agg_errors[c] += cand_errors[b][c];
}
if (group_size > 1) {
float inv_gs = 1.0f / (float)group_size;
for (int c = 0; c < Q4_N_CAND; c++)
agg_errors[c] *= inv_gs;
}
float min_err = 1e30f;
for (int c = 0; c < Q4_N_CAND; c++)
if (agg_errors[c] < min_err)
min_err = agg_errors[c];
double amp_re[6];
double amp_norm = 0.0;
for (int qi = 0; qi < 6; qi++) amp_re[qi] = 0.0;
for (int ci = 0; ci < Q4_N_CAND; ci++) {
int qi = Q4_CAND_TO_QUHIT[ci];
amp_re[qi] += exp(-(double)(agg_errors[ci] - min_err) /
(2.0 * (double)temperature));
}
for (int qi = 0; qi < 6; qi++)
amp_norm += amp_re[qi] * amp_re[qi];
if (amp_norm > 1e-30) {
double inv = 1.0 / sqrt(amp_norm);
for (int v = 0; v < 6; v++) amp_re[v] *= inv;
}
for (int v = 0; v < 6; v++) {
graph->locals[i].edge_re[v] = amp_re[v];
graph->locals[i].edge_im[v] = 0.0;
}
graph->locals[i].primary = VIEW_EDGE;
graph->locals[i].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[i].delta_valid = 0;
triality_update_mask(&graph->locals[i]);
}
/* Neighbor edges */
for (int64_t i = 0; i < graph_blocks - 1; i++)
hpc_cz(graph, i, i + 1);
/* ββ Shor's Griffiths-Niu Sequential Measurement ββ
* Replaces BP with exact marginals via IDFT6 + feed-forward +
* collapse/back-action (ported 1:1 from tesseract_factor.c).
* Single pass, no iteration, no message damping. */
double (*marg)[6] = (double (*)[6])calloc(graph_blocks, sizeof(double[6]));
int *shor_measured = (int *)calloc(graph_blocks, sizeof(int));
shor_measure_graph(graph, graph_blocks, marg, shor_measured, 1);
free(shor_measured);
/* Beam search over candidates */
typedef struct { double acc_error; int history_idx; } Q4Beam;
typedef struct { int cand_idx; int parent_idx; } Q4BeamHistory;
Q4Beam beams[Q4_N_BEAMS];
int active_beams = 1;
Q4BeamHistory *history = (Q4BeamHistory *)malloc(n_blocks * Q4_N_BEAMS * sizeof(Q4BeamHistory));
for (int b = 0; b < Q4_N_BEAMS; b++) {
beams[b].acc_error = 0.0;
beams[b].history_idx = -1;
}
for (int64_t i = 0; i < graph_blocks; i++) {
double m_total = 0.0;
for (int v = 0; v < 6; v++) m_total += marg[i][v];
double cand_score[Q4_N_CAND];
int64_t blk = i * stride;
/* Count candidates per quhit bin for normalization */
int q4_bin_count[6] = {0};
for (int ci = 0; ci < Q4_N_CAND; ci++)
q4_bin_count[Q4_CAND_TO_QUHIT[ci]]++;
/* Per-block error normalization: divide by block mean error
* so small-weight blocks don't dominate beam selection */
float blk_mean_err = 0.0f;
for (int ci = 0; ci < Q4_N_CAND; ci++)
blk_mean_err += cand_errors[blk][ci];
blk_mean_err /= (float)Q4_N_CAND;
if (blk_mean_err < 1e-30f) blk_mean_err = 1e-30f;
for (int ci = 0; ci < Q4_N_CAND; ci++) {
int qi = Q4_CAND_TO_QUHIT[ci];
double p = (m_total > 1e-30) ? marg[i][qi] / m_total : 1.0/6.0;
p /= (double)q4_bin_count[qi]; /* normalize by bin occupancy */
cand_score[ci] = p / (cand_errors[blk][ci] / blk_mean_err + 1e-15);
}
typedef struct { double score; int beam_idx; int cand_idx; } Q4Ext;
Q4Ext extensions[Q4_N_BEAMS * Q4_N_CAND];
int n_ext = 0;
for (int b = 0; b < active_beams; b++) {
for (int c = 0; c < Q4_N_CAND; c++) {
double ext_err = beams[b].acc_error + cand_errors[blk][c];
extensions[n_ext].score = cand_score[c] / (ext_err + 1e-15);
extensions[n_ext].beam_idx = b;
extensions[n_ext].cand_idx = c;
n_ext++;
}
}
int top_k = (n_ext < Q4_N_BEAMS) ? n_ext : Q4_N_BEAMS;
int top_indices[Q4_N_BEAMS];
for (int k = 0; k < top_k; k++) {
int best = -1; double best_s = -1e30;
for (int e = 0; e < n_ext; e++) {
if (extensions[e].score > best_s) {
best_s = extensions[e].score; best = e;
}
}
top_indices[k] = best;
extensions[best].score = -2e30;
}
Q4Beam new_beams[Q4_N_BEAMS];
for (int k = 0; k < top_k; k++) {
int ei = top_indices[k];
int sb = extensions[ei].beam_idx;
int cand = extensions[ei].cand_idx;
int hist_idx = i * Q4_N_BEAMS + k;
history[hist_idx].cand_idx = cand;
history[hist_idx].parent_idx = beams[sb].history_idx;
new_beams[k].history_idx = hist_idx;
new_beams[k].acc_error = beams[sb].acc_error + cand_errors[blk][cand];
}
for (int k = 0; k < top_k; k++) beams[k] = new_beams[k];
active_beams = top_k;
}
int curr_hist = beams[0].history_idx;
for (int64_t i = graph_blocks - 1; i >= 0; i--) {
int group_cidx;
if (curr_hist >= 0) {
group_cidx = history[curr_hist].cand_idx;
curr_hist = history[curr_hist].parent_idx;
} else {
group_cidx = 11;
}
if (stride <= 1) {
best_candidate[i] = group_cidx;
} else {
/* Per-block local optimization within stride group.
* Beam picks the quhit bin; each block picks its best
* candidate in that bin from its own error landscape. */
int target_bin = Q4_CAND_TO_QUHIT[group_cidx];
for (int64_t b = i * stride; b < (i+1) * stride && b < n_blocks; b++) {
float best_err = 1e30f;
int best_c = group_cidx;
for (int c = 0; c < Q4_N_CAND; c++) {
if (Q4_CAND_TO_QUHIT[c] != target_bin) continue;
if (cand_errors[b][c] < best_err) {
best_err = cand_errors[b][c];
best_c = c;
}
}
/* Greedy override if global best is >5% better */
float global_best = 1e30f;
int global_best_c = group_cidx;
for (int c = 0; c < Q4_N_CAND; c++) {
if (cand_errors[b][c] < global_best) {
global_best = cand_errors[b][c];
global_best_c = c;
}
}
if (global_best < best_err * HEX_GREEDY_OVERRIDE_RATIO)
best_candidate[b] = global_best_c;
else
best_candidate[b] = best_c;
}
}
}
free(history);
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* Phase 3.5: Born-Rule Multi-Shot Scale Refinement
*
* The beam search found the MAP candidate sequence. But the
* triality marginals encode quantum phase-coherent structure
* that a greedy beam can miss.
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
{
#define Q4_BORN_SHOTS 128
/* Build per-block CDFs from triality marginals */
unsigned int born_rng = 314159;
/* Compute tail error once (blocks beyond graph coverage) */
float tail_err_q4 = 0.0f;
for (int64_t bi = graph_blocks * stride; bi < n_blocks; bi++)
tail_err_q4 += cand_errors[bi][best_candidate[bi]];
/* Beam-search baseline over the SAME set of blocks a Born
* shot covers: stride representatives + tail. The previous
* code summed the baseline over ALL blocks (including
* mid-stride blocks the shots never touch), making shot_err
* systematically smaller than the baseline and letting
* strictly worse configurations be adopted whenever
* stride > 1. */
float beam_total_err = tail_err_q4;
for (int64_t gi = 0; gi < graph_blocks; gi++) {
int64_t rep = gi * stride;
beam_total_err += cand_errors[rep][best_candidate[rep]];
}
/* Sparse shot buffer: only track stride-sampled blocks */
int *shot_sparse_q4 = (int *)malloc(graph_blocks * sizeof(int));
for (int shot = 0; shot < Q4_BORN_SHOTS; shot++) {
float shot_err = tail_err_q4;
for (int64_t gi = 0; gi < graph_blocks; gi++) {
/* Normalize marginals to CDF */
double m_total = 0.0;
for (int v = 0; v < 6; v++) m_total += marg[gi][v];
/* Born sample: CDF inversion (same as born_sample) */
born_rng = born_rng * 1664525u + 1013904223u;
double rnd = (double)(born_rng >> 8) / 16777216.0;
double target = rnd * m_total;
double cum = 0.0;
int sampled_qi = 5;
for (int v = 0; v < 6; v++) {
cum += marg[gi][v];
if (cum > target) { sampled_qi = v; break; }
}
/* Find the best candidate WITHIN this quhit bin */
int64_t blk = gi * stride;
float best_bin_err = 1e30f;
int best_bin_cand = 11; /* default */
for (int ci = 0; ci < Q4_N_CAND; ci++) {
if (Q4_CAND_TO_QUHIT[ci] == sampled_qi) {
if (cand_errors[blk][ci] < best_bin_err) {
best_bin_err = cand_errors[blk][ci];
best_bin_cand = ci;
}
}
}
shot_sparse_q4[gi] = best_bin_cand;
shot_err += cand_errors[blk][best_bin_cand];
}
/* Metropolis acceptance: adopt if better than current best */
if (shot_err < beam_total_err) {
for (int64_t gi = 0; gi < graph_blocks; gi++)
best_candidate[gi * stride] = shot_sparse_q4[gi];
beam_total_err = shot_err;
}
}
free(shot_sparse_q4);
}
/* Born refinement pass: non-stride blocks were set during beam
* traceback and never revisited by Born shots. For each such block
* pick the lowest-error candidate within the same quhit bin that
* the winning Born shot chose for its stride-representative. */
if (stride > 1) {
for (int64_t b = 0; b < n_blocks; b++) {
if (b % stride == 0) continue;
int64_t rep = (b / stride) * stride;
int target_bin = Q4_CAND_TO_QUHIT[best_candidate[rep]];
float best_b_err = 1e30f;
int best_b_cand = best_candidate[rep];
for (int ci = 0; ci < Q4_N_CAND; ci++) {
if (Q4_CAND_TO_QUHIT[ci] != target_bin) continue;
if (cand_errors[b][ci] < best_b_err) {
best_b_err = cand_errors[b][ci];
best_b_cand = ci;
}
}
best_candidate[b] = best_b_cand;
}
}
free(marg);
hpc_destroy(graph);
}
}
/* Fallback when the HPC graph never ran (single block, or hpc_create
* failure): pick the per-block argmin over the candidate grid instead
* of silently leaving every block on the neutral Γ1.00 candidate. */
if (!hpc_ran_q4) {
#pragma omp parallel for schedule(static)
for (int64_t blk = 0; blk < n_blocks; blk++) {
float best_e = cand_errors[blk][0];
int best_c = 0;
for (int c = 1; c < Q4_N_CAND; c++) {
if (cand_errors[blk][c] < best_e) {
best_e = cand_errors[blk][c];
best_c = c;
}
}
best_candidate[blk] = best_c;
}
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 4: Assemble blocks via least-squares scale extraction
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#pragma omp parallel for schedule(dynamic, 64) reduction(+:total_err)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bw = weights + blk * QK4_0;
int cidx = best_candidate[blk];
/* Start from the grid-selected scale (the "assembled frequency") */
float d_current = gguf_fp16_to_fp32(cand_d16[blk][cidx]);
/* Analog assembly: iterate to full convergence. */
for (int ls_iter = 0; ls_iter < 5; ls_iter++) {
if (d_current < 1e-15f) break;
float id = 1.0f / d_current;
int qs_tmp[QK4_0];
for (int j = 0; j < QK4_0; j++) {
int q = (int)(bw[j] * id + 8.5f);
if (q < 0) q = 0; if (q > 15) q = 15;
qs_tmp[j] = q;
}
float num = 0.0f, den = 0.0f;
float dc4S = 0.0f, dc4Q = 0.0f;
for (int j = 0; j < QK4_0; j++) {
float q_centered = (float)qs_tmp[j] - 8.0f;
float w = (imat_importance) ?
imat_importance[blk * QK4_0 + j] : 1.0f;
num += w * bw[j] * q_centered;
den += w * q_centered * q_centered;
dc4S += bw[j];
dc4Q += q_centered;
}
num += (HEX_DC_LAMBDA / (float)QK4_0) * dc4S * dc4Q;
den += (HEX_DC_LAMBDA / (float)QK4_0) * dc4Q * dc4Q;
if (den > 1e-15f) {
float d_new = num / den;
float d_seed = gguf_fp16_to_fp32(cand_d16[blk][cidx]);
if (fabsf(d_new) < 4.0f * (fabsf(d_seed) + 1e-10f)) {
uint16_t d16 = gguf_fp32_to_fp16(d_new);
d_current = gguf_fp16_to_fp32(d16);
}
}
}
/* ββ FP16 ULP neighborhood search + sign-flip exploration ββ */
{
uint16_t base_d16 = gguf_fp32_to_fp16(d_current);
uint16_t best_d16 = base_d16;
float best_ulp_err = 1e30f;
/* Try Β±8 ULP neighborhood + sign flip = up to 34 candidates */
uint16_t ulp_candidates[35];
int n_ulp = 0;
for (int delta = -8; delta <= 8; delta++) {
int cand16 = (int)base_d16 + delta;
if (cand16 >= 0 && cand16 <= 0x7BFF)
ulp_candidates[n_ulp++] = (uint16_t)cand16;
}
{
float neg_d = -d_current;
uint16_t neg_d16 = gguf_fp32_to_fp16(neg_d);
for (int delta = -8; delta <= 8; delta++) {
int cand16 = (int)neg_d16 + delta;
if (cand16 >= 0 && cand16 <= 0x7BFF)
ulp_candidates[n_ulp++] = (uint16_t)cand16;
}
}
for (int ui = 0; ui < n_ulp; ui++) {
float trial_d = gguf_fp16_to_fp32(ulp_candidates[ui]);
float trial_id = (fabsf(trial_d) > 1e-15f) ? 1.0f / trial_d : 0.0f;
float err = 0.0f;
float e_ulp[QK4_0];
for (int j = 0; j < QK4_0; j++) {
int q = (int)(bw[j] * trial_id + 8.5f);
if (q < 0) q = 0; if (q > 15) q = 15;
float deq = ((float)q - 8.0f) * trial_d;
float w = (imat_importance) ? imat_importance[blk * QK4_0 + j] : 1.0f;
e_ulp[j] = bw[j] - deq;
err += e_ulp[j] * e_ulp[j] * w;
}
err += hex_spectral_penalty(e_ulp, QK4_0);
if (err < best_ulp_err) {
best_ulp_err = err;
best_d16 = ulp_candidates[ui];
}
}
d_current = gguf_fp16_to_fp32(best_d16);
}
output[blk].d = gguf_fp32_to_fp16(d_current);
float actual_d = d_current;
float id = (fabsf(actual_d) > 1e-15f) ? 1.0f / actual_d : 0.0f;
/* ββ Dβ Hadamard Error Shaping with Simulated Annealing ββ */
int q_base[QK4_0], q_shaped[QK4_0];
float q_cont[QK4_0];
for (int j = 0; j < QK4_0; j++) {
q_cont[j] = bw[j] * id + 8.0f;
q_base[j] = (int)(q_cont[j] + 0.5f);
if (q_base[j] < 0) q_base[j] = 0;
if (q_base[j] > 15) q_base[j] = 15;
}
memcpy(q_shaped, q_base, QK4_0 * sizeof(int));
{
float e_live[QK4_0];
for (int j = 0; j < QK4_0; j++) {
float deq = ((float)q_shaped[j] - 8.0f) * actual_d;
e_live[j] = bw[j] - deq;
}
float v_live[QK4_0 / 2];
float vesica_cur = 0.0f, dc_cur = 0.0f;
for (int j = 0; j < QK4_0 / 2; j++) {
v_live[j] = e_live[j] + e_live[j + QK4_0 / 2];
vesica_cur += v_live[j] * v_live[j];
}
for (int j = 0; j < QK4_0; j++) dc_cur += e_live[j];
float metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
/* Deterministic greedy descent: only strict improvements.
* The previous SA acceptance called rand() inside an OpenMP
* parallel region (data race in the shared PRNG state, and
* non-reproducible output). Uphill moves were pointless anyway:
* the base-vs-shaped MSE guard below discards any shaped result
* that ends up worse, so accepted uphill excursions could only
* waste the pass budget or strand the descent. */
for (int pass = 0; pass < QK4_0; pass++) {
int best_k = -1;
int best_q_alt = 0;
float best_delta = 0.0f; /* strictly positive threshold */
for (int k = 0; k < QK4_0; k++) {
int q_cur = q_shaped[k];
int q_try = (q_cont[k] - (float)q_cur >= 0.0f)
? q_cur + 1 : q_cur - 1;
if (q_try < 0 || q_try > 15) continue;
float deq_try = ((float)q_try - 8.0f) * actual_d;
float e_new = bw[k] - deq_try;
float de = e_new - e_live[k];
int pi = (k < QK4_0 / 2) ? k : k - QK4_0 / 2;
float v_old = v_live[pi];
float v_new = v_old + de;
float vesica_alt = vesica_cur - v_old * v_old + v_new * v_new;
float dc_alt = dc_cur + de;
float metric_alt = 4.0f * vesica_alt + dc_alt * dc_alt;
float delta = metric_cur - metric_alt;
if (delta > best_delta) {
best_delta = delta;
best_k = k;
best_q_alt = q_try;
}
}
if (best_k < 0) break; /* converged β no improving flip */
q_shaped[best_k] = best_q_alt;
{
float deq_commit = ((float)best_q_alt - 8.0f) * actual_d;
float e_new_commit = bw[best_k] - deq_commit;
float de_commit = e_new_commit - e_live[best_k];
int pi_commit = (best_k < QK4_0 / 2) ? best_k : best_k - QK4_0 / 2;
float v_old_commit = v_live[pi_commit];
float v_new_commit = v_old_commit + de_commit;
vesica_cur += v_new_commit * v_new_commit - v_old_commit * v_old_commit;
dc_cur += de_commit;
metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
v_live[pi_commit] = v_new_commit;
e_live[best_k] = e_new_commit;
}
}
}
float err_base = 0.0f, err_shaped = 0.0f;
float e_gb[QK4_0], e_gs[QK4_0];
for (int j = 0; j < QK4_0; j++) {
float w = (imat_importance) ? imat_importance[blk * QK4_0 + j] : 1.0f;
float deq_b = ((float)q_base[j] - 8.0f) * actual_d;
float deq_s = ((float)q_shaped[j] - 8.0f) * actual_d;
e_gb[j] = bw[j] - deq_b;
e_gs[j] = bw[j] - deq_s;
err_base += e_gb[j] * e_gb[j] * w;
err_shaped += e_gs[j] * e_gs[j] * w;
}
err_base += hex_spectral_penalty(e_gb, QK4_0);
err_shaped += hex_spectral_penalty(e_gs, QK4_0);
int *q_final = (err_shaped <= err_base) ? q_shaped : q_base;
for (int j = 0; j < QK4_0 / 2; j++) {
int q0 = q_final[j];
int q1 = q_final[j + QK4_0/2];
output[blk].qs[j] = (uint8_t)(q0 | (q1 << 4));
float deq0 = ((float)q0 - 8.0f) * actual_d;
float deq1 = ((float)q1 - 8.0f) * actual_d;
total_err += (bw[j] - deq0) * (bw[j] - deq0) + (bw[j + QK4_0/2] - deq1) * (bw[j + QK4_0/2] - deq1);
}
}
*out_total_error = total_err;
free(greedy_d);
free(cand_errors);
free(cand_d16);
free(best_candidate);
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* Q8_0 HPC QUANTIZER β Shor pipeline at 8 bits
*
* Same pipeline as Q4_0: WLS scale + tight candidate grid scored on the
* extended objective (weighted SSE + DC + vesica/wave), triality-quhit
* graph with Boltzmann-encoded candidate errors, CZ chain entanglement,
* Shor Griffiths-Niu sequential measurement for bin consensus, greedy
* override (HEX_GREEDY_OVERRIDE_RATIO), then per-block ULP polish, the
* vesica/DC error-shaping descent with an extended-objective guard, and
* the candidate floor. Intended for embedding / LM-head tensors (tied
* embeddings especially), where 2-4 bit codes destroy logit precision.
* At 8 bits the candidate grid is tight (Β±1.5%) β the win over naive
* amax/127 rounding comes from WLS + ULP + spectral selection, not from
* coarse scale exploration.
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#ifndef QK8_0
#define QK8_0 32
#endif
typedef struct { uint16_t d; int8_t qs[QK8_0]; } hex_block_q8_0;
#define Q8_N_CAND 24
static const float Q8_NEIGHBOR_MULTS[Q8_N_CAND] = {
0.9850f, 0.9865f, 0.9880f, 0.9895f, 0.9910f, 0.9925f,
0.9940f, 0.9952f, 0.9964f, 0.9976f, 0.9988f, 1.0000f,
1.0010f, 1.0020f, 1.0030f, 1.0040f, 1.0052f, 1.0064f,
1.0076f, 1.0088f, 1.0100f, 1.0115f, 1.0130f, 1.0150f,
};
/* 24 candidates β 6 quhit states (4 per bin), same folding as Q4_0 */
static const int Q8_CAND_TO_QUHIT[Q8_N_CAND] = {
0,0,0,0, 1,1,1,1, 2,2,2,2, 3,3,3,3, 4,4,4,4, 5,5,5,5
};
static inline float q8_block_ext_err(const float *bw, const float *iw,
float d, int8_t *qs_out)
{
float e_arr[QK8_0];
float id = (fabsf(d) > 1e-20f) ? 1.0f / d : 0.0f;
float err = 0.0f;
for (int j = 0; j < QK8_0; j++) {
int q = gguf_nearest_int(bw[j] * id);
if (q < -127) q = -127; if (q > 127) q = 127;
if (qs_out) qs_out[j] = (int8_t)q;
float e = bw[j] - (float)q * d;
e_arr[j] = e;
float w = iw ? iw[j] : 1.0f;
err += e * e * w;
}
return err + hex_spectral_penalty(e_arr, QK8_0);
}
static void quantize_tensor_q8_0_hpc(const float *weights, int64_t n_elements,
hex_block_q8_0 *output,
float *out_total_error,
const float *imat_importance, int verbose)
{
int64_t n_blocks = n_elements / QK8_0;
float total_err = 0.0f;
(void)verbose;
float (*cand_errors)[Q8_N_CAND] = (float (*)[Q8_N_CAND])
calloc(n_blocks, sizeof(float[Q8_N_CAND]));
uint16_t (*cand_d16)[Q8_N_CAND] = (uint16_t (*)[Q8_N_CAND])
calloc(n_blocks, sizeof(uint16_t[Q8_N_CAND]));
int *best_candidate = (int *)malloc(n_blocks * sizeof(int));
if (!cand_errors || !cand_d16 || !best_candidate) {
free(cand_errors); free(cand_d16); free(best_candidate);
if (out_total_error) *out_total_error = -1.0f;
return;
}
/* ββ Phase 1+2: WLS-refined scale + tight candidate grid ββ */
#pragma omp parallel for schedule(dynamic, 256)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bw = weights + blk * QK8_0;
const float *iw = imat_importance ? imat_importance + blk * QK8_0 : NULL;
float amax = 0.0f;
for (int j = 0; j < QK8_0; j++) {
float av = fabsf(bw[j]);
if (av > amax) amax = av;
}
float wls_d = amax / 127.0f;
/* ggml-style fixed-point WLS with DC rank-1 augmentation */
for (int it = 0; it < 3 && wls_d > 1e-20f; it++) {
float inv_d = 1.0f / wls_d;
float num = 0.0f, den = 0.0f, dcS = 0.0f, dcQ = 0.0f;
for (int j = 0; j < QK8_0; j++) {
int q = gguf_nearest_int(bw[j] * inv_d);
if (q < -127) q = -127; if (q > 127) q = 127;
float qf = (float)q;
float w = iw ? iw[j] : 1.0f;
num += w * bw[j] * qf;
den += w * qf * qf;
dcS += bw[j];
dcQ += qf;
}
num += (HEX_DC_LAMBDA / (float)QK8_0) * dcS * dcQ;
den += (HEX_DC_LAMBDA / (float)QK8_0) * dcQ * dcQ;
if (den > 1e-15f) {
float d_new = num / den;
if (d_new > 1e-20f) wls_d = d_new;
}
}
for (int ci = 0; ci < Q8_N_CAND; ci++) {
float trial_d = wls_d * Q8_NEIGHBOR_MULTS[ci];
uint16_t d16 = gguf_fp32_to_fp16(trial_d);
float actual_d = gguf_fp16_to_fp32(d16);
cand_d16 [blk][ci] = d16;
cand_errors[blk][ci] = q8_block_ext_err(bw, iw, actual_d, NULL);
}
best_candidate[blk] = 11; /* Γ1.0000 neutral seed */
}
/* ββ Phase 3: Shor graph β triality quhits, CZ chain, GN measurement ββ */
int shor_ran = 0;
if (n_blocks >= 2) {
int64_t graph_blocks = (n_blocks > 200) ? 200 : n_blocks;
int64_t stride = n_blocks / graph_blocks;
HPCGraph *graph = hpc_create(graph_blocks);
if (graph) {
shor_ran = 1;
/* Adaptive temperature from the candidate-error landscape */
float temperature = 1e-10f;
{
double err_accum = 0.0;
int err_count = 0;
for (int64_t gi = 0; gi < graph_blocks && gi < 100; gi++) {
int64_t blk = gi * stride;
float max_e = 0.0f;
for (int c = 0; c < Q8_N_CAND; c++)
if (cand_errors[blk][c] > max_e)
max_e = cand_errors[blk][c];
err_accum += (double)max_e;
err_count++;
}
if (err_count > 0) {
temperature = (float)(err_accum / err_count) * 0.1f;
if (temperature < 1e-10f) temperature = 1e-10f;
}
}
/* Boltzmann-encode stride-aggregated candidate errors as
* quhit amplitudes (24 candidates folded into 6 states) */
for (int64_t i = 0; i < graph_blocks; i++) {
float agg_errors[Q8_N_CAND];
for (int c = 0; c < Q8_N_CAND; c++) agg_errors[c] = 0.0f;
int64_t blk_start = i * stride;
int64_t blk_end = blk_start + stride;
if (blk_end > n_blocks) blk_end = n_blocks;
for (int64_t b = blk_start; b < blk_end; b++)
for (int c = 0; c < Q8_N_CAND; c++)
agg_errors[c] += cand_errors[b][c];
float min_err = 1e30f;
for (int c = 0; c < Q8_N_CAND; c++)
if (agg_errors[c] < min_err) min_err = agg_errors[c];
double amp_re[6] = {0,0,0,0,0,0};
double amp_norm = 0.0;
for (int ci = 0; ci < Q8_N_CAND; ci++)
amp_re[Q8_CAND_TO_QUHIT[ci]] +=
exp(-(double)(agg_errors[ci] - min_err) /
(2.0 * (double)temperature));
for (int v = 0; v < 6; v++) amp_norm += amp_re[v] * amp_re[v];
if (amp_norm > 1e-30) {
double inv = 1.0 / sqrt(amp_norm);
for (int v = 0; v < 6; v++) amp_re[v] *= inv;
}
for (int v = 0; v < 6; v++) {
graph->locals[i].edge_re[v] = amp_re[v];
graph->locals[i].edge_im[v] = 0.0;
}
graph->locals[i].primary = VIEW_EDGE;
graph->locals[i].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[i].delta_valid = 0;
triality_update_mask(&graph->locals[i]);
}
for (int64_t i = 0; i < graph_blocks - 1; i++)
hpc_cz(graph, i, i + 1);
double (*marg)[6] = (double (*)[6])calloc(graph_blocks, sizeof(double[6]));
int *measured = (int *)calloc(graph_blocks, sizeof(int));
if (marg && measured) {
shor_measure_graph(graph, graph_blocks, marg, measured, 1);
/* Per-block selection: best candidate inside the Shor-
* measured bin, then greedy override against the global
* argmin β identical Step-F semantics to Q2_K/Q4_0. */
for (int64_t i = 0; i < graph_blocks; i++) {
int bin = measured[i];
if (bin < 0 || bin > 5) {
double bm = -1.0; bin = 0;
for (int v = 0; v < 6; v++)
if (marg[i][v] > bm) { bm = marg[i][v]; bin = v; }
}
int64_t blk_start = i * stride;
int64_t blk_end = blk_start + stride;
if (blk_end > n_blocks) blk_end = n_blocks;
for (int64_t b = blk_start; b < blk_end; b++) {
float bin_best = 1e30f; int bin_cand = -1;
float g_best = 1e30f; int g_cand = 0;
for (int c = 0; c < Q8_N_CAND; c++) {
float e = cand_errors[b][c];
if (e < g_best) { g_best = e; g_cand = c; }
if (Q8_CAND_TO_QUHIT[c] == bin && e < bin_best) {
bin_best = e; bin_cand = c;
}
}
int sel = (bin_cand >= 0) ? bin_cand : g_cand;
if (g_best < cand_errors[b][sel] * HEX_GREEDY_OVERRIDE_RATIO)
sel = g_cand;
best_candidate[b] = sel;
}
}
}
free(marg); free(measured);
hpc_destroy(graph);
}
}
if (!shor_ran) {
for (int64_t blk = 0; blk < n_blocks; blk++) {
float g_best = cand_errors[blk][0]; int g_cand = 0;
for (int c = 1; c < Q8_N_CAND; c++)
if (cand_errors[blk][c] < g_best) {
g_best = cand_errors[blk][c]; g_cand = c;
}
best_candidate[blk] = g_cand;
}
}
/* ββ Phase 4: ULP polish + vesica/DC shaping guard + floor ββ */
#pragma omp parallel for schedule(dynamic, 256) reduction(+:total_err)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bw = weights + blk * QK8_0;
const float *iw = imat_importance ? imat_importance + blk * QK8_0 : NULL;
int cidx = best_candidate[blk];
uint16_t best_d16 = cand_d16[blk][cidx];
float best_err = cand_errors[blk][cidx];
/* Β±8 fp16 ULP joint search on the extended objective */
for (int du = -8; du <= 8; du++) {
if (du == 0) continue;
int c16 = (int)cand_d16[blk][cidx] + du;
if (c16 <= 0 || c16 > 0x7BFF) continue;
float td = gguf_fp16_to_fp32((uint16_t)c16);
float err = q8_block_ext_err(bw, iw, td, NULL);
if (err < best_err) { best_err = err; best_d16 = (uint16_t)c16; }
}
/* Candidate floor: final β€ best raw grid candidate (by construction
* the ULP search already starts from it, so this is implicit). */
float d = gguf_fp16_to_fp32(best_d16);
int8_t qs[QK8_0];
(void)q8_block_ext_err(bw, iw, d, qs);
/* Vesica/DC greedy shaping with extended-objective guard */
{
int8_t qs_shaped[QK8_0];
memcpy(qs_shaped, qs, QK8_0);
float e_live[QK8_0], v_live[QK8_0 / 2];
float vesica_cur = 0.0f, dc_cur = 0.0f;
for (int k = 0; k < QK8_0; k++)
e_live[k] = bw[k] - (float)qs_shaped[k] * d;
for (int p = 0; p < QK8_0 / 2; p++) {
v_live[p] = e_live[p] + e_live[p + QK8_0 / 2];
vesica_cur += v_live[p] * v_live[p];
dc_cur += v_live[p];
}
float metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
for (int pass = 0; pass < QK8_0; pass++) {
int best_k = -1, best_q_alt = 0;
float best_delta = 0.0f;
for (int k = 0; k < QK8_0; k++) {
int q_try = (e_live[k] >= 0.0f) ? qs_shaped[k] + 1
: qs_shaped[k] - 1;
if (q_try < -127 || q_try > 127) continue;
float e_new = bw[k] - (float)q_try * d;
float de = e_new - e_live[k];
int pi = (k < QK8_0 / 2) ? k : k - QK8_0 / 2;
float v_new = v_live[pi] + de;
float ves_a = vesica_cur - v_live[pi] * v_live[pi]
+ v_new * v_new;
float dc_a = dc_cur + de;
float delta = metric_cur - (4.0f * ves_a + dc_a * dc_a);
if (delta > best_delta) {
best_delta = delta; best_k = k; best_q_alt = q_try;
}
}
if (best_k < 0) break;
{
float e_new = bw[best_k] - (float)best_q_alt * d;
float de = e_new - e_live[best_k];
int pi = (best_k < QK8_0 / 2) ? best_k
: best_k - QK8_0 / 2;
float v_new = v_live[pi] + de;
vesica_cur += v_new * v_new - v_live[pi] * v_live[pi];
dc_cur += de;
metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
v_live[pi] = v_new;
e_live[best_k] = e_new;
qs_shaped[best_k] = (int8_t)best_q_alt;
}
}
/* Guard on the extended objective vs originals */
float e_b[QK8_0], e_s[QK8_0];
float err_b = 0.0f, err_s = 0.0f;
for (int k = 0; k < QK8_0; k++) {
float w = iw ? iw[k] : 1.0f;
e_b[k] = bw[k] - (float)qs[k] * d;
e_s[k] = bw[k] - (float)qs_shaped[k] * d;
err_b += e_b[k] * e_b[k] * w;
err_s += e_s[k] * e_s[k] * w;
}
err_b += hex_spectral_penalty(e_b, QK8_0);
err_s += hex_spectral_penalty(e_s, QK8_0);
if (err_s < err_b) memcpy(qs, qs_shaped, QK8_0);
}
output[blk].d = best_d16;
for (int k = 0; k < QK8_0; k++) {
output[blk].qs[k] = qs[k];
float e = bw[k] - (float)qs[k] * d;
total_err += e * e; /* pure reconstruction SSE report */
}
}
free(cand_errors);
free(cand_d16);
free(best_candidate);
if (out_total_error) *out_total_error = total_err;
}
/* Re-derive the 4-bit sub-scale codes (Ls, Lm) for a candidate (d, dmin)
* pair from the Phase-1 float scales/mins. Bit-identical to the Phase-2b
* candidate generation, so stored codes are unnecessary. */
static inline void hex_derive_subscales(const float *scales, const float *mins,
float actual_dm, float actual_mm,
uint8_t *Ls, uint8_t *Lm)
{
for (int j = 0; j < 16; j++) {
if (actual_dm > 1e-15f) {
int ls = gguf_nearest_int(scales[j] / actual_dm);
if (ls < 0) ls = 0; if (ls > 15) ls = 15;
Ls[j] = (uint8_t)ls;
} else { Ls[j] = 0; }
if (actual_mm > 1e-15f) {
int lm = gguf_nearest_int(mins[j] / actual_mm);
if (lm < 0) lm = 0; if (lm > 15) lm = 15;
Lm[j] = (uint8_t)lm;
} else { Lm[j] = 0; }
}
}
static void quantize_tensor_q2k_hpc(const float *weights, int64_t n_elements,
BlockQ2K *output, float *out_total_error,
OptimizerMode opt_mode,
const float *imat_importance,
int verbose)
{
int64_t n_blocks = n_elements / QK_K;
float total_err = 0.0f;
const int N_SUB = QK_K / 16;
/* ββ Outlier Clamping for WLS Seeds ββ
* Protects the Phase 1 greedy seed from being violently warped by extreme
* >4.0 sigma outliers, which creates better centering for the grid search. */
double t_sum_sq = 0.0, t_sum_4 = 0.0;
for (int64_t i = 0; i < n_elements; i++) {
double w2 = (double)weights[i] * (double)weights[i];
t_sum_sq += w2;
t_sum_4 += w2 * w2;
}
float w_sigma = sqrtf((float)(t_sum_sq / (double)n_elements));
/* ββ Adaptive outlier clamp (kurtosis-driven) ββ
* The fixed 3.5Ο clamp suppressed the heavy-tail mass that dominates
* reconstruction error, inflating RMSE on near-Gaussian tensors that did
* not need clamping at all. Instead, gate the clamp on the tensor's raw
* kurtosis (Gaussian = 3): leave near-Gaussian tensors untouched and only
* apply a stabilising clamp to genuinely heavy-tailed tensors, where the
* final (d, dmin) refit later recovers fidelity against the UNCLIPPED
* weights anyway. */
double t_var = t_sum_sq / (double)n_elements;
double t_kurt = (t_var > 1e-30) ? (t_sum_4 / (double)n_elements) / (t_var * t_var) : 3.0;
float clamp_sigma;
if (t_kurt <= 6.0) clamp_sigma = 1.0e9f; /* ~Gaussian: effectively no clamp */
else if (t_kurt <= 20.0) clamp_sigma = 6.0f; /* moderately heavy tails */
else clamp_sigma = 4.0f; /* very heavy tails: stabilise seed */
float clamp_val = w_sigma * clamp_sigma;
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 1: Greedy quantization β produce seed (d, dmin) per block
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
typedef struct {
float dm, mm;
uint16_t d_fp16, dmin_fp16;
uint8_t Ls[16], Lm[16];
float scales[16], mins[16], sw[16];
} BlockSeed;
BlockSeed *seeds = (BlockSeed *)calloc(n_blocks, sizeof(BlockSeed));
#pragma omp parallel for schedule(dynamic, 64)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *block_x = weights + blk * QK_K;
uint8_t L[QK_K], Laux[16];
float wt[16];
float sumx2 = 0;
for (int i = 0; i < QK_K; i++) sumx2 += block_x[i] * block_x[i];
float sigma2 = sumx2 / (float)QK_K;
/* Phase 1 WLS uses clamped values to generate stable seeds */
float sx_clipped[16];
for (int j = 0; j < N_SUB; j++) {
const float *sx = block_x + 16 * j;
seeds[blk].sw[j] = 0;
for (int l = 0; l < 16; l++) {
float imp = (imat_importance) ? imat_importance[blk * QK_K + 16 * j + l] : 1.0f;
float v = sx[l];
if (v > clamp_val) v = clamp_val;
if (v < -clamp_val) v = -clamp_val;
sx_clipped[l] = v;
/* Activation-aware weighting: an imatrix entry already encodes
* E[a^2] for that column, which is the correct weight for
* minimising output (dot-product) error. Use it directly rather
* than re-multiplying by the |w| magnitude heuristic, which
* double-counts magnitude. Without an imatrix, fall back to the
* magnitude-relative heuristic. */
wt[l] = (imat_importance)
? imp
: sqrtf(sigma2 + sx_clipped[l] * sx_clipped[l]);
seeds[blk].sw[j] += wt[l];
}
seeds[blk].scales[j] = hpc_make_qkx2_quants(16, 3, sx_clipped, wt,
L + 16 * j, &seeds[blk].mins[j], Laux);
}
seeds[blk].dm = hpc_make_qp_quants(N_SUB, 15, seeds[blk].scales,
seeds[blk].Ls, seeds[blk].sw);
seeds[blk].mm = hpc_make_qp_quants(N_SUB, 15, seeds[blk].mins,
seeds[blk].Lm, seeds[blk].sw);
seeds[blk].d_fp16 = gguf_fp32_to_fp16(seeds[blk].dm);
seeds[blk].dmin_fp16 = gguf_fp32_to_fp16(seeds[blk].mm);
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 2: WLS-Optimal Candidate Generation
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* Expanded neighborhood around WLS optimum: Β±30% with 24 candidates */
/* d is the sensitive axis, so concentrate resolution near 1.0 while
* keeping wide tails for blocks whose WLS seed is off. 1.000 stays at
* index 11 so the neutral-candidate fallback/init remains valid. */
static const float NEIGHBOR_MULTS_D[N_CAND_D] = {
0.780f, 0.835f, 0.880f, 0.915f, 0.943f, 0.963f,
0.978f, 0.988f, 0.994f, 0.997f, 0.999f, 1.000f,
1.002f, 1.005f, 1.011f, 1.021f, 1.035f, 1.054f,
1.080f, 1.115f, 1.160f, 1.215f, 1.275f, 1.340f
};
static const float NEIGHBOR_MULTS_M[N_CAND_M] = {
0.750f, 0.800f, 0.840f, 0.870f, 0.900f, 0.920f,
0.940f, 0.955f, 0.970f, 0.985f, 0.995f, 1.000f,
1.005f, 1.015f, 1.030f, 1.045f, 1.060f, 1.080f,
1.100f, 1.130f, 1.160f, 1.200f, 1.250f, 1.300f
};
/* Map 24 candidates β 6 quhit states for BP encoding */
static const int CAND_TO_QUHIT[24] = {
0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2,
3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5
};
float (*candidate_errors)[TOTAL_SCALE_CANDIDATES] = NULL;
uint16_t (*candidate_d)[TOTAL_SCALE_CANDIDATES] = NULL;
uint16_t (*candidate_dmin)[TOTAL_SCALE_CANDIDATES] = NULL;
candidate_errors = (float (*)[TOTAL_SCALE_CANDIDATES])calloc(n_blocks,
sizeof(float[TOTAL_SCALE_CANDIDATES]));
candidate_d = (uint16_t (*)[TOTAL_SCALE_CANDIDATES])calloc(n_blocks,
sizeof(uint16_t[TOTAL_SCALE_CANDIDATES]));
candidate_dmin = (uint16_t (*)[TOTAL_SCALE_CANDIDATES])calloc(n_blocks,
sizeof(uint16_t[TOTAL_SCALE_CANDIDATES]));
/* NOTE: the per-candidate sub-scale codes (Ls/Lm) are NOT stored.
* They are a pure function of (seeds[blk].scales/mins, candidate fp16
* d/dmin) and are re-derived where needed. Storing them cost
* n_blocks Γ 576 Γ 16 Γ 2 bytes β 18 KB/superblock β multiple GB of
* peak RSS on large FFN tensors β for data used at exactly one index. */
#pragma omp parallel for schedule(dynamic, 16)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *block_x = weights + blk * QK_K;
/* ββ Step 2a: WLS solve to find optimal (d*, dmin*) ββ */
float wls_dm = seeds[blk].dm;
float wls_mm = seeds[blk].mm;
uint8_t wls_Ls[16], wls_Lm[16];
memcpy(wls_Ls, seeds[blk].Ls, 16);
memcpy(wls_Lm, seeds[blk].Lm, 16);
/* Generate soft-clipped buffer for WLS internal stability */
float clipped_block_x[QK_K];
for(int i=0; i<QK_K; i++) {
float v = block_x[i];
if (v > clamp_val) v = clamp_val;
if (v < -clamp_val) v = -clamp_val;
clipped_block_x[i] = v;
}
for (int ls_iter = 0; ls_iter < 5; ls_iter++) {
uint8_t L_wls[QK_K];
for (int j = 0; j < N_SUB; j++) {
float d_sub = wls_dm * (float)wls_Ls[j];
float m_sub = wls_mm * (float)wls_Lm[j];
if (d_sub < 1e-15f) {
for (int k = 0; k < 16; k++) L_wls[16*j+k] = 0;
continue;
}
for (int k = 0; k < 16; k++) {
int q = gguf_nearest_int((clipped_block_x[16*j+k] + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
L_wls[16*j+k] = (uint8_t)q;
}
}
double Saa = 0, Sab = 0, Sbb = 0, Sxa = 0, Sxb = 0;
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)wls_Ls[j];
float lm_f = (float)wls_Lm[j];
for (int k = 0; k < 16; k++) {
float x = clipped_block_x[16*j+k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j+k] : 1.0f;
float a = ls_f * (float)L_wls[16*j+k];
float b = lm_f;
Saa += w * a * a;
Sab += w * a * b;
Sbb += w * b * b;
Sxa += w * x * a;
Sxb += w * x * b;
}
}
double det = Saa * Sbb - Sab * Sab;
if (fabs(det) > 1e-30) {
double d_new = (Sbb * Sxa - Sab * Sxb) / det;
double dm_new = (Sab * Sxa - Saa * Sxb) / det;
if (d_new > 0.0 && d_new < 4.0 * (seeds[blk].dm + 1e-10))
wls_dm = gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)d_new));
if (dm_new > 0.0 && dm_new < 4.0 * (seeds[blk].mm + 1e-10))
wls_mm = gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)dm_new));
}
for (int j = 0; j < N_SUB; j++) {
if (wls_dm > 1e-15f) {
int ls = gguf_nearest_int(seeds[blk].scales[j] / wls_dm);
if (ls < 0) ls = 0; if (ls > 15) ls = 15;
wls_Ls[j] = (uint8_t)ls;
} else { wls_Ls[j] = 0; }
if (wls_mm > 1e-15f) {
int lm = gguf_nearest_int(seeds[blk].mins[j] / wls_mm);
if (lm < 0) lm = 0; if (lm > 15) lm = 15;
wls_Lm[j] = (uint8_t)lm;
} else { wls_Lm[j] = 0; }
}
}
/* ββ Step 2b: Generate Candidates ββ */
for (int di = 0; di < N_CAND_D; di++) {
float trial_dm = wls_dm * NEIGHBOR_MULTS_D[di];
uint16_t trial_d16 = gguf_fp32_to_fp16(trial_dm);
float actual_dm = gguf_fp16_to_fp32(trial_d16);
for (int mi = 0; mi < N_CAND_M; mi++) {
int cidx = di * N_CAND_M + mi;
float trial_mm = wls_mm * NEIGHBOR_MULTS_M[mi];
uint16_t trial_dmin16 = gguf_fp32_to_fp16(trial_mm);
float actual_mm = gguf_fp16_to_fp32(trial_dmin16);
candidate_d[blk][cidx] = trial_d16;
candidate_dmin[blk][cidx] = trial_dmin16;
uint8_t trial_Ls[16], trial_Lm[16];
for (int j = 0; j < N_SUB; j++) {
if (actual_dm > 1e-15f) {
int ls = gguf_nearest_int(seeds[blk].scales[j] / actual_dm);
if (ls < 0) ls = 0; if (ls > 15) ls = 15;
trial_Ls[j] = (uint8_t)ls;
} else { trial_Ls[j] = 0; }
if (actual_mm > 1e-15f) {
int lm = gguf_nearest_int(seeds[blk].mins[j] / actual_mm);
if (lm < 0) lm = 0; if (lm > 15) lm = 15;
trial_Lm[j] = (uint8_t)lm;
} else { trial_Lm[j] = 0; }
}
/* Error evaluation MUST use the non-clipped original weights.
* Exact importance-weighted SSE β the same objective the
* assembly/polish phases minimise and the reported RMSE. */
float err = 0.0f;
float e_arr[QK_K];
for (int i = 0; i < QK_K; i++) {
int jj = i >> 4;
float d = actual_dm * (float)trial_Ls[jj];
float m = actual_mm * (float)trial_Lm[jj];
float x = block_x[i];
float w = (imat_importance) ? imat_importance[blk * QK_K + i] : 1.0f;
float e;
if (d < 1e-15f) {
/* Decoder semantics: deq = dΒ·lsΒ·q β dminΒ·lm = βm here */
e = x + m;
} else {
int q = gguf_nearest_int((x + m) / d);
if (q < 0) q = 0; if (q > 3) q = 3;
e = x - (d * (float)q - m);
}
e_arr[i] = e;
err += e * e * w;
}
candidate_errors[blk][cidx] =
err + hex_spectral_penalty(e_arr, QK_K);
}
}
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 3: HPC Graph β Shor's Griffiths-Niu Measurement
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
int *best_candidate = (int *)malloc(n_blocks * sizeof(int));
for (int64_t i = 0; i < n_blocks; i++)
best_candidate[i] = 11 * N_CAND_M + 11; /* index 11 = 1.0 multiplier */
if (opt_mode != OPT_MSE && n_blocks >= 2) {
int64_t graph_blocks = (n_blocks > 2000) ? 2000 : n_blocks;
int64_t stride = n_blocks / graph_blocks;
float temperature = 0.5f;
int64_t n_sites = graph_blocks * QUHITS_PER_BLOCK;
HPCGraph *graph = hpc_create(n_sites);
if (graph) {
for (int64_t i = 0; i < n_sites; i++)
triality_dft(&graph->locals[i]);
{
double err_accum = 0.0;
int err_count = 0;
for (int64_t gi = 0; gi < graph_blocks && gi < 100; gi++) {
int64_t blk = gi * stride;
float max_e = 0.0f;
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++)
if (candidate_errors[blk][c] > max_e)
max_e = candidate_errors[blk][c];
err_accum += (double)max_e;
err_count++;
}
if (err_count > 0) {
float median_err = (float)(err_accum / err_count);
temperature = median_err * 0.1f;
if (temperature < 1e-10f) temperature = 1e-10f;
}
}
for (int64_t i = 0; i < graph_blocks; i++) {
float agg_errors[TOTAL_SCALE_CANDIDATES];
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++) agg_errors[c] = 0.0f;
int64_t blk_start = i * stride;
int64_t blk_end = blk_start + stride;
if (blk_end > n_blocks) blk_end = n_blocks;
int64_t group_size = blk_end - blk_start;
for (int64_t b = blk_start; b < blk_end; b++) {
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++)
agg_errors[c] += candidate_errors[b][c];
}
if (group_size > 1) {
float inv_gs = 1.0f / (float)group_size;
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++)
agg_errors[c] *= inv_gs;
}
float min_err = 1e30f;
for (int c = 0; c < TOTAL_SCALE_CANDIDATES; c++)
if (agg_errors[c] < min_err)
min_err = agg_errors[c];
double coarse_re[6];
double coarse_norm = 0.0;
for (int qi = 0; qi < 6; qi++) coarse_re[qi] = 0.0;
for (int di = 0; di < N_CAND_D; di++) {
int qi = CAND_TO_QUHIT[di];
for (int mi = 0; mi < N_CAND_M; mi++) {
int cidx = di * N_CAND_M + mi;
coarse_re[qi] += exp(-(double)(agg_errors[cidx] - min_err) /
(2.0 * (double)temperature));
}
}
for (int qi = 0; qi < 6; qi++) coarse_norm += coarse_re[qi] * coarse_re[qi];
if (coarse_norm > 1e-30) {
double inv = 1.0 / sqrt(coarse_norm);
for (int v = 0; v < 6; v++) coarse_re[v] *= inv;
}
double fine_re[6];
double fine_norm = 0.0;
for (int qi = 0; qi < 6; qi++) fine_re[qi] = 0.0;
for (int mi = 0; mi < N_CAND_M; mi++) {
int qi = CAND_TO_QUHIT[mi];
for (int di = 0; di < N_CAND_D; di++) {
int cidx = di * N_CAND_M + mi;
fine_re[qi] += exp(-(double)(agg_errors[cidx] - min_err) /
(2.0 * (double)temperature));
}
}
for (int qi = 0; qi < 6; qi++) fine_norm += fine_re[qi] * fine_re[qi];
if (fine_norm > 1e-30) {
double inv = 1.0 / sqrt(fine_norm);
for (int v = 0; v < 6; v++) fine_re[v] *= inv;
}
int64_t s0 = 2 * i, s1 = 2 * i + 1;
for (int v = 0; v < 6; v++) {
graph->locals[s0].edge_re[v] = coarse_re[v];
graph->locals[s0].edge_im[v] = 0.0;
graph->locals[s1].edge_re[v] = fine_re[v];
graph->locals[s1].edge_im[v] = 0.0;
}
graph->locals[s0].primary = VIEW_EDGE;
graph->locals[s0].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[s0].delta_valid = 0;
triality_update_mask(&graph->locals[s0]);
graph->locals[s1].primary = VIEW_EDGE;
graph->locals[s1].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
graph->locals[s1].delta_valid = 0;
triality_update_mask(&graph->locals[s1]);
}
for (int64_t i = 0; i < graph_blocks; i++) {
hpc_cz(graph, 2 * i, 2 * i + 1);
if (i + 1 < graph_blocks) {
hpc_cz(graph, 2 * i, 2 * (i + 1));
hpc_cz(graph, 2 * i + 1, 2 * (i + 1) + 1);
}
}
double (*shor_marg)[6] = (double (*)[6])calloc(n_sites, sizeof(double[6]));
int *shor_measured = (int *)calloc(n_sites, sizeof(int));
shor_measure_graph(graph, n_sites, shor_marg, shor_measured, 1);
double (*coarse_marg)[6] = (double (*)[6])calloc(graph_blocks, sizeof(double[6]));
double (*fine_marg)[6] = (double (*)[6])calloc(graph_blocks, sizeof(double[6]));
for (int64_t i = 0; i < graph_blocks; i++) {
for (int v = 0; v < 6; v++) {
coarse_marg[i][v] = shor_marg[2 * i][v];
fine_marg[i][v] = shor_marg[2 * i + 1][v];
}
}
free(shor_marg);
free(shor_measured);
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 3 β DETERMINISTIC VITERBI DP
*
* Replaces the probabilistic beam-search + Born-rule Monte-Carlo
* shots with an exact, fully-deterministic DP over the 36-state
* Shor quhit space (6 coarse bins Γ 6 fine bins).
*
* For each graph block i and combined state s = qi_d*6 + qi_m:
*
* bin_best_err[i][s] = min candidate error in that (d,m)-bin
* aggregated over the stride group
* bin_log_prior[i][s] = log P_coarse(qi_d) + log P_fine(qi_m)
* from Shor marginals β HPC prior bonus
*
* Local Viterbi cost (lower = better):
* vcost[i][s] = bin_best_err[i][s]
* β VITERBI_BETA Γ scale_err Γ bin_log_prior[i][s]
*
* Transition cost (cross-block smoothness prior):
* trans(sβ²βs) = VITERBI_ALPHA Γ scale_err
* Γ (|qi_d β qi_dβ²| + |qi_m β qi_mβ²|)
*
* DP recurrence:
* dp[0][s] = vcost[0][s]
* dp[i][s] = vcost[i][s] + min_{sβ²}(dp[i-1][sβ²] + trans(sβ²βs))
*
* Traceback yields the globally optimal sequence of bin choices,
* which is then mapped to per-block best_candidate[] indices.
* A 5%-threshold greedy override rescues blocks where the local
* MSE-optimal candidate is meaningfully better than the bin winner.
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#define VIT_N_STATES 36 /* 6 coarse Γ 6 fine quhit bins */
#define VITERBI_BETA 0.25f /* log-prior bonus weight */
#define VITERBI_ALPHA 0.08f /* cross-block smoothness penalty weight */
{
int64_t vit_gi, vit_b;
int vit_s, vit_sp;
/* Per-graph-block per-state workspace */
float (*vit_bin_err )[VIT_N_STATES] =
(float (*)[VIT_N_STATES])malloc(graph_blocks * sizeof(float[VIT_N_STATES]));
int (*vit_bin_cand)[VIT_N_STATES] =
(int (*)[VIT_N_STATES])malloc(graph_blocks * sizeof(int [VIT_N_STATES]));
float (*vit_log_pri )[VIT_N_STATES] =
(float (*)[VIT_N_STATES])malloc(graph_blocks * sizeof(float[VIT_N_STATES]));
float (*vit_dp )[VIT_N_STATES] =
(float (*)[VIT_N_STATES])malloc(graph_blocks * sizeof(float[VIT_N_STATES]));
int (*vit_back )[VIT_N_STATES] =
(int (*)[VIT_N_STATES])malloc(graph_blocks * sizeof(int [VIT_N_STATES]));
/* ββ Step A: build per-block per-bin statistics ββ */
for (vit_gi = 0; vit_gi < graph_blocks; vit_gi++) {
double c_tot = 0.0, f_tot = 0.0;
for (vit_s = 0; vit_s < VIT_N_STATES; vit_s++) {
vit_bin_err [vit_gi][vit_s] = 1e30f;
vit_bin_cand[vit_gi][vit_s] = -1;
}
/* Best candidate per (qi_d, qi_m) bin over stride group */
for (vit_b = vit_gi * stride;
vit_b < (vit_gi + 1) * stride && vit_b < n_blocks;
vit_b++) {
int vit_c;
for (vit_c = 0; vit_c < TOTAL_SCALE_CANDIDATES; vit_c++) {
int qi_d = CAND_TO_QUHIT[vit_c / N_CAND_M];
int qi_m = CAND_TO_QUHIT[vit_c % N_CAND_M];
vit_s = qi_d * 6 + qi_m;
float e = candidate_errors[vit_b][vit_c];
if (e < vit_bin_err[vit_gi][vit_s]) {
vit_bin_err[vit_gi][vit_s] = e;
/* Canonical candidate = stride-rep block's best */
if (vit_b == vit_gi * stride)
vit_bin_cand[vit_gi][vit_s] = vit_c;
}
}
}
/* HPC log-prior from Shor marginals */
for (int v = 0; v < 6; v++) {
c_tot += coarse_marg[vit_gi][v];
f_tot += fine_marg [vit_gi][v];
}
for (vit_s = 0; vit_s < VIT_N_STATES; vit_s++) {
int qi_d = vit_s / 6, qi_m = vit_s % 6;
double pc = (c_tot > 1e-30)
? coarse_marg[vit_gi][qi_d] / c_tot : 1.0/6.0;
double pf = (f_tot > 1e-30)
? fine_marg [vit_gi][qi_m] / f_tot : 1.0/6.0;
vit_log_pri[vit_gi][vit_s] =
(float)(log(pc + 1e-30) + log(pf + 1e-30));
}
}
/* ββ Step B: scale_err normaliser for transition cost ββ */
float vit_scale_err = 0.0f;
int vit_scale_cnt = 0;
for (vit_gi = 0; vit_gi < graph_blocks; vit_gi++) {
for (vit_s = 0; vit_s < VIT_N_STATES; vit_s++) {
if (vit_bin_err[vit_gi][vit_s] < 1e29f) {
vit_scale_err += vit_bin_err[vit_gi][vit_s];
vit_scale_cnt++;
}
}
}
vit_scale_err = (vit_scale_cnt > 0)
? vit_scale_err / (float)vit_scale_cnt : 1e-10f;
if (vit_scale_err < 1e-20f) vit_scale_err = 1e-20f;
/* ββ Step C: Forward Viterbi pass ββ */
/* Block 0 β no predecessor */
for (vit_s = 0; vit_s < VIT_N_STATES; vit_s++) {
float local = (vit_bin_err[0][vit_s] < 1e29f)
? vit_bin_err[0][vit_s]
- VITERBI_BETA * vit_scale_err * vit_log_pri[0][vit_s]
: 1e30f;
vit_dp [0][vit_s] = local;
vit_back[0][vit_s] = -1;
}
/* Blocks 1..graph_blocks-1 */
for (vit_gi = 1; vit_gi < graph_blocks; vit_gi++) {
for (vit_s = 0; vit_s < VIT_N_STATES; vit_s++) {
float local;
float best_pred = 1e30f;
int best_sp = 0;
int qi_d = vit_s / 6;
int qi_m = vit_s % 6;
if (vit_bin_err[vit_gi][vit_s] > 1e29f) {
vit_dp [vit_gi][vit_s] = 1e30f;
vit_back[vit_gi][vit_s] = 0;
continue;
}
local = vit_bin_err[vit_gi][vit_s]
- VITERBI_BETA * vit_scale_err * vit_log_pri[vit_gi][vit_s];
/* Min-cost predecessor with Manhattan transition penalty */
for (vit_sp = 0; vit_sp < VIT_N_STATES; vit_sp++) {
float prev = vit_dp[vit_gi - 1][vit_sp];
if (prev > 1e29f) continue;
int td = abs(qi_d - (vit_sp / 6));
int tm = abs(qi_m - (vit_sp % 6));
float trans = VITERBI_ALPHA * vit_scale_err * (float)(td + tm);
float total = prev + trans;
if (total < best_pred) {
best_pred = total;
best_sp = vit_sp;
}
}
vit_dp [vit_gi][vit_s] = (best_pred < 1e29f)
? best_pred + local : 1e30f;
vit_back[vit_gi][vit_s] = best_sp;
}
}
/* ββ Step D: Traceback ββ */
int *vit_path = (int *)malloc(graph_blocks * sizeof(int));
{
int best_s = 0;
float best_f = vit_dp[graph_blocks - 1][0];
for (vit_s = 1; vit_s < VIT_N_STATES; vit_s++) {
if (vit_dp[graph_blocks - 1][vit_s] < best_f) {
best_f = vit_dp[graph_blocks - 1][vit_s];
best_s = vit_s;
}
}
vit_path[graph_blocks - 1] = best_s;
for (vit_gi = graph_blocks - 2; vit_gi >= 0; vit_gi--)
vit_path[vit_gi] = vit_back[vit_gi + 1][vit_path[vit_gi + 1]];
}
/* ββ Step E: Map Viterbi path β best_candidate[] ββ */
for (vit_gi = 0; vit_gi < graph_blocks; vit_gi++) {
vit_s = vit_path[vit_gi];
int qi_d = vit_s / 6;
int qi_m = vit_s % 6;
int64_t blk_rep = vit_gi * stride;
/* Stride-representative block: use precomputed bin winner */
if (vit_bin_cand[vit_gi][vit_s] >= 0)
best_candidate[blk_rep] = vit_bin_cand[vit_gi][vit_s];
/* Non-representative blocks in the stride group */
for (vit_b = blk_rep + 1;
vit_b < (vit_gi + 1) * stride && vit_b < n_blocks;
vit_b++) {
int vit_c;
float best_e = 1e30f;
int best_c = best_candidate[blk_rep];
for (vit_c = 0; vit_c < TOTAL_SCALE_CANDIDATES; vit_c++) {
if (CAND_TO_QUHIT[vit_c / N_CAND_M] != qi_d) continue;
if (CAND_TO_QUHIT[vit_c % N_CAND_M] != qi_m) continue;
if (candidate_errors[vit_b][vit_c] < best_e) {
best_e = candidate_errors[vit_b][vit_c];
best_c = vit_c;
}
}
best_candidate[vit_b] = best_c;
}
}
/* ββ Step F: 5 % greedy override (pure MSE safety net) ββ */
for (vit_b = 0; vit_b < n_blocks; vit_b++) {
int vit_c;
float cur_err = candidate_errors[vit_b][best_candidate[vit_b]];
float g_best = cur_err;
int g_cand = best_candidate[vit_b];
for (vit_c = 0; vit_c < TOTAL_SCALE_CANDIDATES; vit_c++) {
if (candidate_errors[vit_b][vit_c] < g_best) {
g_best = candidate_errors[vit_b][vit_c];
g_cand = vit_c;
}
}
if (g_best < cur_err * HEX_GREEDY_OVERRIDE_RATIO)
best_candidate[vit_b] = g_cand;
}
free(vit_path);
free(vit_dp);
free(vit_back);
free(vit_bin_err);
free(vit_bin_cand);
free(vit_log_pri);
}
free(coarse_marg);
free(fine_marg);
hpc_destroy(graph);
}
} else {
for (int64_t blk = 0; blk < n_blocks; blk++) {
float best_err = candidate_errors[blk][0];
int best_idx = 0;
for (int c = 1; c < TOTAL_SCALE_CANDIDATES; c++) {
if (candidate_errors[blk][c] < best_err) {
best_err = candidate_errors[blk][c];
best_idx = c;
}
}
best_candidate[blk] = best_idx;
}
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 3.9 β ROLLING DC BOUNDARY CONDITION PRE-PASS
*
* Transforms the tensor from a collection of isolated 256-element
* Q2_K superblocks into a single, continuous error-cancelling waveform.
*
* After Phase 3 has selected the optimal (d, dmin) candidate for every
* block, this sequential pass computes the net DC residual left by each
* block using a cheap round-nearest forward quantization, then feeds the
* negated, exponentially-decayed residual as a correction bias into the
* WLS solver of the immediately following block.
*
* Mathematically, for block N with final DC residual R_N = Ξ£ Ξ΅α΅’:
*
* dc_bias[N+1] = βDC_DECAY Γ R_N / QK_K (per-element offset)
*
* Block N+1's WLS targets become xβ²α΅’ = xα΅’ β dc_bias[N+1], steering the
* quantizer toward codes whose reconstruction deq β xβ², so that
*
* Ξ£ (xα΅’ β deqα΅’) β dc_bias[N+1] Γ QK_K = βDC_DECAY Γ R_N
*
* The accumulated cross-block DC collapses geometrically:
*
* Rβ, DC_DECAYΒ·Rβ, DC_DECAYΒ²Β·Rβ, β¦ β 0
*
* The result is written into block_dc_bias[n_blocks]. Phase 4 reads
* this array (safe: written sequentially before the parallel loop).
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
#define DC_DECAY 0.85f /* Boundary-condition leak factor (0 = isolated, 1 = full) */
float *block_dc_bias = (float *)calloc(n_blocks, sizeof(float));
if (block_dc_bias) {
float rolling_dc = 0.0f;
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *bx = weights + blk * QK_K;
int cidx = best_candidate[blk];
float dm0 = gguf_fp16_to_fp32(candidate_d [blk][cidx]);
float mm0 = gguf_fp16_to_fp32(candidate_dmin[blk][cidx]);
uint8_t dc_Ls[16], dc_Lm[16];
hex_derive_subscales(seeds[blk].scales, seeds[blk].mins,
dm0, mm0, dc_Ls, dc_Lm);
/* Bias applied to THIS block's WLS targets */
float dc_bias = (DC_DECAY * rolling_dc) / (float)QK_K;
block_dc_bias[blk] = dc_bias;
/* Quick round-nearest quant to estimate DC residual for NEXT block.
* We quantize the adjusted target xβ² = x β dc_bias, then measure
* the residual of the ORIGINAL weight against the chosen code. */
float dc_res = 0.0f;
int j, k;
for (j = 0; j < N_SUB; j++) {
float d_sub = dm0 * (float)dc_Ls[j];
float m_sub = mm0 * (float)dc_Lm[j];
for (k = 0; k < 16; k++) {
float x_adj = bx[16*j + k] - dc_bias;
int q = 0;
if (d_sub >= 1e-15f) {
q = gguf_nearest_int((x_adj + m_sub) / d_sub);
if (q < 0) q = 0;
if (q > 3) q = 3;
}
float deq = d_sub * (float)q - m_sub;
/* Residual against ORIGINAL weight (not adjusted) */
dc_res += bx[16*j + k] - deq;
}
}
rolling_dc = dc_res;
}
}
/* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PHASE 4: Assemble blocks via least-squares (d, dmin) extraction
* ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
int _n_omp_threads = 1;
#ifdef _OPENMP
_n_omp_threads = omp_get_max_threads();
#endif
HPCGraph **_tl_graphs = (HPCGraph **)calloc(_n_omp_threads, sizeof(HPCGraph *));
for (int _ti = 0; _ti < _n_omp_threads; _ti++)
_tl_graphs[_ti] = hpc_create(N_SUB);
#pragma omp parallel for schedule(dynamic, 64) reduction(+:total_err)
for (int64_t blk = 0; blk < n_blocks; blk++) {
const float *block_x = weights + blk * QK_K;
int cidx = best_candidate[blk];
uint8_t Ls_blk[16], Lm_blk[16];
/* ββ Rolling DC boundary condition ββββββββββββββββββββββββββββββ
* dc_adj shifts every WLS target in this block so that the net
* quantisation error steers toward cancelling the previous block's
* DC residual (written by the sequential Phase 3.9 pre-pass). */
float dc_adj = (block_dc_bias) ? block_dc_bias[blk] : 0.0f;
/* Adjusted weight view β WLS and Shor work on this array;
* the final error is always reported against the original block_x. */
float adj_block_x[QK_K];
{
int _i;
for (_i = 0; _i < QK_K; _i++)
adj_block_x[_i] = block_x[_i] - dc_adj;
}
float dm = gguf_fp16_to_fp32(candidate_d[blk][cidx]);
float mm = gguf_fp16_to_fp32(candidate_dmin[blk][cidx]);
hex_derive_subscales(seeds[blk].scales, seeds[blk].mins,
dm, mm, Ls_blk, Lm_blk);
uint16_t prev_dm16 = 0, prev_mm16 = 0;
for (int ls_iter = 0; ls_iter < 5; ls_iter++) {
uint8_t state_ls[N_SUB][6];
uint8_t state_lm[N_SUB][6];
float state_err[N_SUB][6];
for (int j = 0; j < N_SUB; j++) {
const float *sx = adj_block_x + 16 * j;
for (int v = 0; v < 6; v++) state_err[j][v] = 1e30f;
for (int try_ls = 0; try_ls <= 15; try_ls++) {
float d_sub = dm * (float)try_ls;
for (int try_lm = 0; try_lm <= 15; try_lm++) {
float m_sub = mm * (float)try_lm;
float sub_err = 0.0f;
for (int k = 0; k < 16; k++) {
float x = sx[k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j + k] : 1.0f;
int q = 0;
if (d_sub >= 1e-15f) {
q = gguf_nearest_int((x + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
}
float deq = d_sub * (float)q - m_sub;
float diff = x - deq;
sub_err += diff * diff * w;
}
for (int v = 0; v < 6; v++) {
if (sub_err < state_err[j][v]) {
for (int u = 5; u > v; u--) {
state_err[j][u] = state_err[j][u-1];
state_ls[j][u] = state_ls[j][u-1];
state_lm[j][u] = state_lm[j][u-1];
}
state_err[j][v] = sub_err;
state_ls[j][v] = (uint8_t)try_ls;
state_lm[j][v] = (uint8_t)try_lm;
break;
}
}
}
}
}
int _tid = 0;
#ifdef _OPENMP
_tid = omp_get_thread_num();
#endif
HPCGraph *sg = _tl_graphs[_tid];
hpc_reset_for_subblock(sg, N_SUB);
{
float min_sub_err[N_SUB];
for (int j = 0; j < N_SUB; j++) min_sub_err[j] = state_err[j][0];
for (int j = 0; j < N_SUB; j++) {
triality_dft(&sg->locals[j]);
double amp_re[6];
double amp_norm = 0.0;
for (int v = 0; v < 6; v++) {
float err_spread = state_err[j][5] - state_err[j][0];
float sub_temp = (err_spread > 1e-15f) ? err_spread * 0.3f : 0.1f;
if (sub_temp < 1e-12f) sub_temp = 1e-12f;
amp_re[v] = exp(-(double)(state_err[j][v] - min_sub_err[j]) / (double)sub_temp);
amp_norm += amp_re[v] * amp_re[v];
}
if (amp_norm > 1e-30) {
double inv = 1.0 / sqrt(amp_norm);
for (int v = 0; v < 6; v++) amp_re[v] *= inv;
}
for (int v = 0; v < 6; v++) {
sg->locals[j].edge_re[v] = amp_re[v];
sg->locals[j].edge_im[v] = 0.0;
}
sg->locals[j].primary = VIEW_EDGE;
sg->locals[j].dirty = DIRTY_VERTEX | DIRTY_DIAGONAL | DIRTY_FOLDED;
sg->locals[j].delta_valid = 0;
triality_update_mask(&sg->locals[j]);
}
for (int j = 0; j < N_SUB - 1; j++)
hpc_cz(sg, j, j + 1);
double sub_marg[N_SUB][6];
int sub_measured[N_SUB];
memset(sub_marg, 0, sizeof(sub_marg));
memset(sub_measured, 0, sizeof(sub_measured));
shor_measure_graph(sg, N_SUB, sub_marg, sub_measured, 1);
for (int j = 0; j < N_SUB; j++) {
double best_prob = -1.0;
int best_v = 0;
for (int v = 0; v < 6; v++) {
if (sub_marg[j][v] > best_prob) {
best_prob = sub_marg[j][v];
best_v = v;
}
}
Ls_blk[j] = state_ls[j][best_v];
Lm_blk[j] = state_lm[j][best_v];
}
}
uint8_t L[QK_K];
for (int j = 0; j < N_SUB; j++) {
float d_sub = dm * (float)Ls_blk[j];
float m_sub = mm * (float)Lm_blk[j];
if (d_sub < 1e-15f) {
for (int k = 0; k < 16; k++) L[16*j+k] = 0;
continue;
}
for (int k = 0; k < 16; k++) {
int q = gguf_nearest_int((adj_block_x[16*j+k] + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
L[16*j+k] = (uint8_t)q;
}
}
double Saa = 0, Sab = 0, Sbb = 0, Sxa = 0, Sxb = 0;
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)Ls_blk[j];
float lm_f = (float)Lm_blk[j];
for (int k = 0; k < 16; k++) {
float x = adj_block_x[16*j+k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j+k] : 1.0f;
float a = ls_f * (float)L[16*j+k];
float b = lm_f;
Saa += w * a * a;
Sab += w * a * b;
Sbb += w * b * b;
Sxa += w * x * a;
Sxb += w * x * b;
}
}
double det = Saa * Sbb - Sab * Sab;
if (fabs(det) > 1e-30) {
double d_new = (Sbb * Sxa - Sab * Sxb) / det;
double dm_new = (Sab * Sxa - Saa * Sxb) / det;
float d_seed = gguf_fp16_to_fp32(candidate_d[blk][cidx]);
float m_seed = gguf_fp16_to_fp32(candidate_dmin[blk][cidx]);
if (d_new > 0.0 && d_new < 4.0 * (d_seed + 1e-10))
dm = gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)d_new));
if (dm_new > 0.0 && dm_new < 4.0 * (m_seed + 1e-10))
mm = gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)dm_new));
}
uint16_t cur_dm16 = gguf_fp32_to_fp16(dm);
uint16_t cur_mm16 = gguf_fp32_to_fp16(mm);
if (cur_dm16 == prev_dm16 && cur_mm16 == prev_mm16) break;
prev_dm16 = cur_dm16;
prev_mm16 = cur_mm16;
}
/* ββ FP16 ULP neighborhood search for (d, dmin) β Expanded to Β±8 ββ */
{
uint16_t base_d16 = gguf_fp32_to_fp16(dm);
uint16_t base_m16 = gguf_fp32_to_fp16(mm);
uint16_t best_d16 = base_d16, best_m16 = base_m16;
float best_ulp_err = 1e30f;
for (int dd = -8; dd <= 8; dd++) {
int cd16 = (int)base_d16 + dd;
if (cd16 < 0 || cd16 > 0x7BFF) continue;
float trial_dm = gguf_fp16_to_fp32((uint16_t)cd16);
for (int dm_delta = -8; dm_delta <= 8; dm_delta++) {
int cm16 = (int)base_m16 + dm_delta;
if (cm16 < 0 || cm16 > 0x7BFF) continue;
float trial_mm = gguf_fp16_to_fp32((uint16_t)cm16);
float err = 0.0f;
for (int j = 0; j < N_SUB; j++) {
float d_sub = trial_dm * (float)Ls_blk[j];
float m_sub = trial_mm * (float)Lm_blk[j];
for (int k = 0; k < 16; k++) {
float x = adj_block_x[16*j+k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j+k] : 1.0f;
int q;
if (d_sub < 1e-15f) { q = 0; }
else {
q = gguf_nearest_int((x + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
}
float deq = d_sub * (float)q - m_sub;
float diff = x - deq;
err += diff * diff * w;
}
}
if (err < best_ulp_err) {
best_ulp_err = err;
best_d16 = (uint16_t)cd16;
best_m16 = (uint16_t)cm16;
}
}
}
dm = gguf_fp16_to_fp32(best_d16);
mm = gguf_fp16_to_fp32(best_m16);
}
for (int j = 0; j < N_SUB; j++) {
const float *sx = adj_block_x + 16 * j;
float best_sub_err = 1e30f;
uint8_t best_ls = Ls_blk[j], best_lm = Lm_blk[j];
for (int try_ls = 0; try_ls <= 15; try_ls++) {
float d_sub = dm * (float)try_ls;
for (int try_lm = 0; try_lm <= 15; try_lm++) {
float m_sub = mm * (float)try_lm;
float sub_err = 0.0f;
for (int k = 0; k < 16; k++) {
float x = sx[k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j + k] : 1.0f;
int q;
if (d_sub < 1e-15f) { q = 0; }
else {
q = gguf_nearest_int((x + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
}
float deq = d_sub * (float)q - m_sub;
float diff = x - deq;
sub_err += diff * diff * w;
}
if (sub_err < best_sub_err) {
best_sub_err = sub_err;
best_ls = (uint8_t)try_ls;
best_lm = (uint8_t)try_lm;
}
}
}
Ls_blk[j] = best_ls;
Lm_blk[j] = best_lm;
}
output[blk].d = gguf_fp32_to_fp16(dm);
output[blk].dmin = gguf_fp32_to_fp16(mm);
for (int j = 0; j < N_SUB; j++)
output[blk].scales[j] = Ls_blk[j] | (Lm_blk[j] << 4);
/* ββ Final quantization: Dβ Hadamard Greedy Descent (deterministic) ββ
*
* The original Simulated Annealing acceptance rule is replaced by a
* strict greedy descent: only accept a flip if it strictly reduces the
* Dβ Hadamard metric (4Β·βvesicaβΒ² + DCΒ²). This makes error shaping
* fully deterministic and thread-safe (no rand() inside omp parallel),
* consistent with the Viterbi philosophy applied in Phase 3.
*
* The metric measures both:
* - Vesica Piscis term: correlated error between weights i and i+QK_K/2
* (targets the first non-DC harmonic β halfwave symmetry)
* - DC term: total signed error across the 256-weight superblock
* (captured and propagated to the next block by Phase 3.9)
*/
uint8_t L[QK_K];
{
float q_cont_all[QK_K];
int q_base_all[QK_K];
int q_shaped_all[QK_K];
for (int i = 0; i < QK_K; i++) {
int jj = i >> 4;
float d_s = dm * (float)(output[blk].scales[jj] & 0xF);
float m_s = mm * (float)(output[blk].scales[jj] >> 4);
if (d_s < 1e-15f) {
q_cont_all[i] = 0.0f;
q_base_all[i] = 0;
} else {
/* Quantize the DC-adjusted target */
float qc = (adj_block_x[i] + m_s) / d_s;
q_cont_all[i] = qc;
int qr = gguf_nearest_int(qc);
if (qr < 0) qr = 0; if (qr > 3) qr = 3;
q_base_all[i] = qr;
}
}
memcpy(q_shaped_all, q_base_all, QK_K * sizeof(int));
float e_live[QK_K];
for (int i = 0; i < QK_K; i++) {
int jj = i >> 4;
float d_s = dm * (float)(output[blk].scales[jj] & 0xF);
float m_s = mm * (float)(output[blk].scales[jj] >> 4);
/* Decoder semantics: deq = d_sΒ·q β m_s, which is βm_s when
* d_s == 0 (NOT 0 β the βdminΒ·lm term always applies). */
float deq = d_s * (float)q_shaped_all[i] - m_s;
/* Residual against the adjusted target (DC-corrected view) */
e_live[i] = adj_block_x[i] - deq;
}
float v_live[QK_K / 2];
float vesica_cur = 0.0f, dc_cur = 0.0f;
for (int i = 0; i < QK_K / 2; i++) {
v_live[i] = e_live[i] + e_live[i + QK_K / 2];
vesica_cur += v_live[i] * v_live[i];
}
for (int i = 0; i < QK_K; i++) dc_cur += e_live[i];
float metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
/* Deterministic greedy descent: accept only strict improvements */
for (int pass = 0; pass < QK_K; pass++) {
int best_k = -1;
int best_q_alt = 0;
float best_delta = 0.0f; /* strictly positive threshold */
for (int k = 0; k < QK_K; k++) {
int jj = k >> 4;
float d_s = dm * (float)(output[blk].scales[jj] & 0xF);
if (d_s < 1e-15f) continue;
int q_cur = q_shaped_all[k];
int q_try = (q_cont_all[k] - (float)q_cur >= 0.0f)
? q_cur + 1 : q_cur - 1;
if (q_try < 0 || q_try > 3) continue;
float m_s = mm * (float)(output[blk].scales[jj] >> 4);
float e_new = adj_block_x[k] - (d_s * (float)q_try - m_s);
float de = e_new - e_live[k];
int pi = (k < QK_K / 2) ? k : k - QK_K / 2;
float v_new = v_live[pi] + de;
float vesica_alt = vesica_cur - v_live[pi]*v_live[pi] + v_new*v_new;
float dc_alt = dc_cur + de;
float delta = metric_cur - (4.0f * vesica_alt + dc_alt * dc_alt);
if (delta > best_delta) {
best_delta = delta;
best_k = k;
best_q_alt = q_try;
}
}
if (best_k < 0) break; /* converged β no further improvement */
q_shaped_all[best_k] = best_q_alt;
{
int jj_c = best_k >> 4;
float d_c = dm * (float)(output[blk].scales[jj_c] & 0xF);
float m_c = mm * (float)(output[blk].scales[jj_c] >> 4);
float e_new_c = adj_block_x[best_k] - (d_c * (float)best_q_alt - m_c);
float de_c = e_new_c - e_live[best_k];
int pi_c = (best_k < QK_K / 2) ? best_k : best_k - QK_K / 2;
float v_new_c = v_live[pi_c] + de_c;
vesica_cur += v_new_c * v_new_c - v_live[pi_c] * v_live[pi_c];
dc_cur += de_c;
metric_cur = 4.0f * vesica_cur + dc_cur * dc_cur;
v_live[pi_c] = v_new_c;
e_live[best_k]= e_new_c;
}
}
/* Choose base vs shaped on the EXTENDED objective vs originals */
float err_base = 0.0f, err_shaped = 0.0f;
float e_qb[QK_K], e_qs[QK_K];
for (int i = 0; i < QK_K; i++) {
int jj = i >> 4;
float d_s = dm * (float)(output[blk].scales[jj] & 0xF);
float m_s = mm * (float)(output[blk].scales[jj] >> 4);
float w = (imat_importance) ? imat_importance[blk * QK_K + i] : 1.0f;
float deq_b = d_s * (float)q_base_all[i] - m_s; /* βm_s when d_s==0 */
float deq_s = d_s * (float)q_shaped_all[i] - m_s;
float xv = block_x[i]; /* original weight for error report */
e_qb[i] = xv - deq_b;
e_qs[i] = xv - deq_s;
err_base += e_qb[i] * e_qb[i] * w;
err_shaped += e_qs[i] * e_qs[i] * w;
}
err_base += hex_spectral_penalty(e_qb, QK_K);
err_shaped += hex_spectral_penalty(e_qs, QK_K);
{
int use_shaped = (err_shaped <= err_base);
for (int i = 0; i < QK_K; i++)
L[i] = (uint8_t)(use_shaped ? q_shaped_all[i] : q_base_all[i]);
}
}
/* ββ Cross-weight error diffusion β intra-sub-block Floyd-Steinberg ββ
*
* Implements cross-weight error diffusion within each 16-weight sub-block.
* After the greedy descent has committed quantisation codes, the residual
* of each weight is partially propagated forward to the next position in
* the same sub-block (7/16 of the error), re-quantising if the diffused
* target falls in a different bin.
*
* This is the "cross-weight" dimension of the error-diffusion request:
* neighbouring weights share and partially absorb each other's rounding
* error, shaping the within-block spectrum away from the DC component
* that Phase 3.9 already propagates between blocks.
*
* Staying within sub-blocks avoids scale-mismatch artefacts that would
* arise from diffusing across the dm * Ls[j] boundary between sub-blocks.
*
* The diffused codes are accepted only when they reduce the weighted MSE
* against the ORIGINAL weight (not the adjusted target), so the diffusion
* cannot increase the total reconstruction error.
*/
{
int fs_j, fs_k;
for (fs_j = 0; fs_j < N_SUB; fs_j++) {
int base = fs_j * 16;
float d_s = dm * (float)(output[blk].scales[fs_j] & 0xF);
float m_s = mm * (float)(output[blk].scales[fs_j] >> 4);
if (d_s < 1e-15f) continue;
float carry = 0.0f; /* FS carry from position k-1 */
for (fs_k = 0; fs_k < 16; fs_k++) {
int idx = base + fs_k;
float x_orig = block_x[idx];
float x_adj = adj_block_x[idx] + carry; /* adjusted + diffused */
/* Propose new code from diffused target */
int q_fs = gguf_nearest_int((x_adj + m_s) / d_s);
if (q_fs < 0) q_fs = 0; if (q_fs > 3) q_fs = 3;
if (q_fs != (int)L[idx]) {
/* Accept only when MSE against original weight improves */
float w_imp = (imat_importance)
? imat_importance[blk * QK_K + idx] : 1.0f;
float deq_old = d_s * (float)L[idx] - m_s;
float deq_new = d_s * (float)q_fs - m_s;
float e_old = (x_orig - deq_old) * (x_orig - deq_old) * w_imp;
float e_new = (x_orig - deq_new) * (x_orig - deq_new) * w_imp;
if (e_new < e_old)
L[idx] = (uint8_t)q_fs;
}
/* Propagate 7/16 of the residual (adj target vs committed code) */
{
float deq_final = d_s * (float)L[idx] - m_s;
float residual = (adj_block_x[idx] - deq_final);
carry = (fs_k < 15) ? residual * (7.0f / 16.0f) : 0.0f;
}
}
}
}
/* ββ Final closed-form (d, dmin) refit against the UNCLIPPED weights ββ
* (issues #2 / #5)
*
* Every earlier (d, dmin) solve fits the DC-adjusted, soft-clipped
* target and runs BEFORE the greedy descent and Floyd-Steinberg passes
* mutate the committed 2-bit codes. Once L[], and the 4-bit sub-block
* scale codes (Ls = scales & 0xF, Lm = scales >> 4), are final, the two
* fp16 scalars (d, dmin) that minimise the importance-weighted SSE
* against the ORIGINAL weights have a closed form. Solve it and adopt it
* only when it lowers the weighted block error β so it can never raise
* RMSE, and because the integer codes are held fixed, the vesica/wave/DC
* error shaping baked into them is preserved intact. */
{
double rSaa = 0, rSab = 0, rSbb = 0, rSxa = 0, rSxb = 0;
double rA = 0, rB = 0, rS = 0; /* DC rank-1 augmentation */
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)(output[blk].scales[j] & 0xF);
float lm_f = (float)(output[blk].scales[j] >> 4);
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float x = block_x[idx]; /* unclipped original */
float w = (imat_importance) ? imat_importance[blk * QK_K + idx] : 1.0f;
float a = ls_f * (float)L[idx];
float b = lm_f;
rSaa += (double)w * a * a;
rSab += (double)w * a * b;
rSbb += (double)w * b * b;
rSxa += (double)w * x * a;
rSxb += (double)w * x * b;
rA += a; rB += b; rS += x;
}
}
/* DC term as one augmented observation (S ~ AΒ·d β BΒ·m), weight
* Ξ»_dc/n; vesica/wave handled by the extended-E acceptance. */
{
double rw = (double)HEX_DC_LAMBDA / (double)QK_K;
rSaa += rw * rA * rA; rSab += rw * rA * rB;
rSbb += rw * rB * rB; rSxa += rw * rS * rA;
rSxb += rw * rS * rB;
}
double rdet = rSaa * rSbb - rSab * rSab;
if (fabs(rdet) > 1e-30) {
double d_ref = (rSbb * rSxa - rSab * rSxb) / rdet;
double m_ref = (rSab * rSxa - rSaa * rSxb) / rdet;
if (d_ref > 0.0) {
float dm_try = gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)d_ref));
float mm_try = (m_ref > 0.0)
? gguf_fp16_to_fp32(gguf_fp32_to_fp16((float)m_ref))
: mm;
/* Extended-objective acceptance test vs original weights. */
float err_cur = 0.0f, err_try = 0.0f;
float e_rc[QK_K], e_rt[QK_K];
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)(output[blk].scales[j] & 0xF);
float lm_f = (float)(output[blk].scales[j] >> 4);
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float x = block_x[idx];
float w = (imat_importance) ? imat_importance[blk * QK_K + idx] : 1.0f;
float qf = (float)L[idx];
float dc = dm * ls_f * qf - mm * lm_f;
float dt = dm_try * ls_f * qf - mm_try * lm_f;
e_rc[idx] = x - dc;
e_rt[idx] = x - dt;
err_cur += e_rc[idx] * e_rc[idx] * w;
err_try += e_rt[idx] * e_rt[idx] * w;
}
}
err_cur += hex_spectral_penalty(e_rc, QK_K);
err_try += hex_spectral_penalty(e_rt, QK_K);
if (err_try < err_cur) { dm = dm_try; mm = mm_try; }
}
}
output[blk].d = gguf_fp32_to_fp16(dm);
output[blk].dmin = gguf_fp32_to_fp16(mm);
}
/* ββ PHASE 4.6: MONOTONE COORDINATE-DESCENT POLISH (RMSE-guaranteed) ββ
*
* Objective-function mismatch fix: the final passes that commit the
* 2-bit codes β the 16Γ16 (ls, lm) sub-block search, the Β±8 ULP
* (d, dmin) neighborhood search, and the greedy-descent error shaping
* β all minimise error against the DC-ADJUSTED target adj_block_x.
* The reported RMSE, however, is measured against the ORIGINAL
* weights. The codes are therefore stranded at the optimum of a
* SHIFTED objective, while only the scalar (d, dmin) refit above
* targets the true one (and it holds all codes frozen).
*
* This polish runs alternating coordinate descent on the TRUE
* objective (importance-weighted SSE vs the original weights):
*
* (1) For each 16-weight sub-block, an exact joint re-search of
* (ls, lm) over the full 16Γ16 grid with per-weight optimal
* q β {0..3}, committed only on strict improvement of the
* extended objective E. With Ξ»_dc = Ξ»_vw = 0 sub-blocks are
* independent given (d, dmin); with spectral terms active the
* coupling (DC: all subs; fold: sub j β sub jβ8) is handled
* exactly via live residual bookkeeping.
* (2) Closed-form weighted LS refit of the two fp16 scalars
* (d, dmin) with all codes held fixed, committed only on
* strict improvement (same guard as the refit above).
*
* All moves are accept-only-if-better on E β the extended block
* objective is monotonically non-increasing; at Ξ» = 0 this reduces
* to RMSE-monotone (final RMSE can only go DOWN relative to the
* unpatched pipeline), at Ξ» > 0 small RMSE giveback is permitted
* exactly where it buys dot-product error cancellation. The state space is finite
* (4-bit codes, fp16 scalars), so the loop terminates; in practice
* it converges in 2β3 sweeps. The vesica/DC spectral shaping baked
* into L survives wherever it is SSE-neutral, and is overridden
* only where it was costing true reconstruction error. */
{
uint8_t pl_Ls[16], pl_Lm[16];
for (int j = 0; j < N_SUB; j++) {
pl_Ls[j] = output[blk].scales[j] & 0xF;
pl_Lm[j] = output[blk].scales[j] >> 4;
}
for (int pol_iter = 0; pol_iter < 6; pol_iter++) {
int pol_improved = 0;
/* ββ (1) Exact per-sub-block (ls, lm, q) re-search on the
* EXTENDED objective. Under the spectral terms sub-blocks
* are no longer independent: every sub couples to all others
* through the DC term and to its fold partner (sub j β 8,
* i.e. weights i β i+128) through vesicaΒ² β waveΒ². The
* search therefore keeps live residuals pe[] and scores each
* candidate against the whole-block penalty with the partner
* residuals held fixed β exact coordinate descent on E. */
float pe[QK_K];
float sub_sse[16], sub_dc[16], pair_cross[8];
float dc_tot = 0.0f, cross_tot = 0.0f;
for (int j = 0; j < N_SUB; j++) {
float d_sub = dm * (float)pl_Ls[j];
float m_sub = mm * (float)pl_Lm[j];
sub_sse[j] = 0.0f;
sub_dc[j] = 0.0f;
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
/* deq = dΒ·lsΒ·q β dminΒ·lm; equals βm_sub at ls==0 */
float e = block_x[idx] - (d_sub * (float)L[idx] - m_sub);
pe[idx] = e;
sub_sse[j] += e * e * w;
sub_dc[j] += e;
}
dc_tot += sub_dc[j];
}
for (int p = 0; p < 8; p++) {
pair_cross[p] = 0.0f;
for (int k = 0; k < 16; k++)
pair_cross[p] += pe[16*p + k] * pe[16*(p+8) + k];
cross_tot += pair_cross[p];
}
for (int j = 0; j < N_SUB; j++) {
const float *sx = block_x + 16 * j;
int pi = j & 7; /* fold-pair index */
int pj = j ^ 8; /* partner sub-block */
const float *ppe = pe + 16 * pj; /* partner residuals */
float dc_rest = dc_tot - sub_dc[j];
float cross_rest = cross_tot - pair_cross[pi];
/* Extended score of the CURRENT committed state */
float best_sub = sub_sse[j]
+ (HEX_DC_LAMBDA / (float)QK_K) * dc_tot * dc_tot
+ (HEX_VW_LAMBDA / (float)QK_K) * 4.0f * cross_tot;
int best_ls = -1, best_lm = 0;
uint8_t best_q[16];
float best_e[16];
float best_sse = 0.0f, best_dcc = 0.0f, best_cxc = 0.0f;
for (int try_ls = 0; try_ls <= 15; try_ls++) {
float d_sub = dm * (float)try_ls;
for (int try_lm = 0; try_lm <= 15; try_lm++) {
float m_sub = mm * (float)try_lm;
float sub_err = 0.0f, dcc = 0.0f, cxc = 0.0f;
uint8_t q_loc[16];
float e_loc[16];
int aborted = 0;
for (int k = 0; k < 16; k++) {
float x = sx[k];
float w = (imat_importance) ?
imat_importance[blk * QK_K + 16*j + k] : 1.0f;
int q = 0;
if (d_sub >= 1e-15f) {
q = gguf_nearest_int((x + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
}
q_loc[k] = (uint8_t)q;
/* deq = dΒ·lsΒ·q β dminΒ·lm; βm_sub at ls==0 */
float e = x - (d_sub * (float)q - m_sub);
e_loc[k] = e;
sub_err += e * e * w;
dcc += e;
cxc += e * ppe[k];
/* SSE-partial prune is a valid lower bound
* only while the spectral terms are β₯ 0,
* i.e. when the (signable) vw credit is off */
if (HEX_VW_LAMBDA == 0.0f &&
sub_err >= best_sub) { aborted = 1; break; }
}
if (aborted) continue;
float score = sub_err
+ (HEX_DC_LAMBDA / (float)QK_K)
* (dc_rest + dcc) * (dc_rest + dcc)
+ (HEX_VW_LAMBDA / (float)QK_K) * 4.0f
* (cross_rest + cxc);
if (score < best_sub) {
best_sub = score;
best_ls = try_ls;
best_lm = try_lm;
memcpy(best_q, q_loc, 16);
memcpy(best_e, e_loc, sizeof(e_loc));
best_sse = sub_err;
best_dcc = dcc;
best_cxc = cxc;
}
}
}
if (best_ls >= 0) { /* strict improvement in E found */
pl_Ls[j] = (uint8_t)best_ls;
pl_Lm[j] = (uint8_t)best_lm;
memcpy(L + 16 * j, best_q, 16);
memcpy(pe + 16 * j, best_e, sizeof(best_e));
sub_sse[j] = best_sse;
sub_dc[j] = best_dcc;
pair_cross[pi] = best_cxc;
dc_tot = dc_rest + best_dcc;
cross_tot = cross_rest + best_cxc;
pol_improved = 1;
}
}
/* ββ (2) Closed-form (d, dmin) refit vs ORIGINAL, codes fixed ββ */
{
double pSaa = 0, pSab = 0, pSbb = 0, pSxa = 0, pSxb = 0;
double pA = 0, pB = 0, pS = 0; /* DC rank-1 augmentation */
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)pl_Ls[j];
float lm_f = (float)pl_Lm[j];
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float x = block_x[idx];
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
float a = ls_f * (float)L[idx];
float b = lm_f;
pSaa += (double)w * a * a;
pSab += (double)w * a * b;
pSbb += (double)w * b * b;
pSxa += (double)w * x * a;
pSxb += (double)w * x * b;
pA += a; pB += b; pS += x;
}
}
{
double pw = (double)HEX_DC_LAMBDA / (double)QK_K;
pSaa += pw * pA * pA; pSab += pw * pA * pB;
pSbb += pw * pB * pB; pSxa += pw * pS * pA;
pSxb += pw * pS * pB;
}
double pdet = pSaa * pSbb - pSab * pSab;
if (fabs(pdet) > 1e-30) {
double d_ref = (pSbb * pSxa - pSab * pSxb) / pdet;
double m_ref = (pSab * pSxa - pSaa * pSxb) / pdet;
if (d_ref > 0.0) {
float dm_try = gguf_fp16_to_fp32(
gguf_fp32_to_fp16((float)d_ref));
float mm_try = (m_ref > 0.0)
? gguf_fp16_to_fp32(
gguf_fp32_to_fp16((float)m_ref))
: mm;
float err_cur = 0.0f, err_try = 0.0f;
float e_pc[QK_K], e_pt[QK_K];
for (int j = 0; j < N_SUB; j++) {
float ls_f = (float)pl_Ls[j];
float lm_f = (float)pl_Lm[j];
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float x = block_x[idx];
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
float qf = (float)L[idx];
float dc = dm * ls_f * qf - mm * lm_f;
float dt = dm_try * ls_f * qf - mm_try * lm_f;
e_pc[idx] = x - dc;
e_pt[idx] = x - dt;
err_cur += e_pc[idx] * e_pc[idx] * w;
err_try += e_pt[idx] * e_pt[idx] * w;
}
}
err_cur += hex_spectral_penalty(e_pc, QK_K);
err_try += hex_spectral_penalty(e_pt, QK_K);
if (err_try < err_cur) {
dm = dm_try;
mm = mm_try;
pol_improved = 1;
}
}
}
}
if (!pol_improved) {
/* ββ (3) Β±2 ULP joint (d, dmin) micro-search vs ORIGINAL ββ
* The closed-form refit rounds its real-valued optimum to
* fp16, which can land 1β2 ULP away from the best
* representable pair (and the earlier Β±8 ULP search ran
* against the DC-shifted objective). With codes fixed,
* scan the (2Β·HEX_POLISH_ULP+1)Β² fp16 neighborhood on the
* true objective;
* accept only strict improvement, then loop once more so
* move (1) can re-optimise codes for the new scalars.
* Monotone β final RMSE can only decrease. */
uint16_t base_d16 = gguf_fp32_to_fp16(dm);
uint16_t base_m16 = gguf_fp32_to_fp16(mm);
float cur_err = 0.0f;
float e_u[QK_K];
for (int j = 0; j < N_SUB; j++) {
float d_sub = dm * (float)pl_Ls[j];
float m_sub = mm * (float)pl_Lm[j];
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
e_u[idx] = block_x[idx] -
(d_sub * (float)L[idx] - m_sub);
cur_err += e_u[idx] * e_u[idx] * w;
}
}
cur_err += hex_spectral_penalty(e_u, QK_K);
float best_err = cur_err;
uint16_t best_d16 = base_d16, best_m16 = base_m16;
for (int dd = -HEX_POLISH_ULP; dd <= HEX_POLISH_ULP; dd++) {
int cd16 = (int)base_d16 + dd;
if (cd16 < 0 || cd16 > 0x7BFF) continue;
float t_dm = gguf_fp16_to_fp32((uint16_t)cd16);
for (int dmm = -HEX_POLISH_ULP; dmm <= HEX_POLISH_ULP; dmm++) {
if (dd == 0 && dmm == 0) continue;
int cm16 = (int)base_m16 + dmm;
if (cm16 < 0 || cm16 > 0x7BFF) continue;
float t_mm = gguf_fp16_to_fp32((uint16_t)cm16);
float err = 0.0f;
/* SSE-partial prune valid only without the
* signable vesica/wave credit */
for (int j = 0;
j < N_SUB && (HEX_VW_LAMBDA != 0.0f ||
err < best_err); j++) {
float d_sub = t_dm * (float)pl_Ls[j];
float m_sub = t_mm * (float)pl_Lm[j];
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
e_u[idx] = block_x[idx] -
(d_sub * (float)L[idx] - m_sub);
err += e_u[idx] * e_u[idx] * w;
}
}
if (HEX_DC_LAMBDA != 0.0f || HEX_VW_LAMBDA != 0.0f)
err = (err < best_err || HEX_VW_LAMBDA != 0.0f)
? err + hex_spectral_penalty(e_u, QK_K)
: err;
if (err < best_err) {
best_err = err;
best_d16 = (uint16_t)cd16;
best_m16 = (uint16_t)cm16;
}
}
}
if (best_d16 != base_d16 || best_m16 != base_m16) {
dm = gguf_fp16_to_fp32(best_d16);
mm = gguf_fp16_to_fp32(best_m16);
pol_improved = 1;
}
}
if (!pol_improved) break; /* converged on true objective */
}
/* Write back polished codes and scalars */
for (int j = 0; j < N_SUB; j++)
output[blk].scales[j] = pl_Ls[j] | (pl_Lm[j] << 4);
output[blk].d = gguf_fp32_to_fp16(dm);
output[blk].dmin = gguf_fp32_to_fp16(mm);
}
/* ββ PHASE 4.7: CANDIDATE FLOOR (worst-case bound) ββ
*
* candidate_errors[blk][c] is the EXACT weighted SSE of a directly
* encodable configuration (fp16 d/dmin + derived Ls/Lm + nearest
* rounding vs the original weights). The multi-stage assembly
* (DC-shifted WLS, shaping, diffusion, polish) usually improves on
* its seed, but each stage optimises a slightly different objective
* and coordinate descent can land in a worse basin. Compare the
* finished block against the best raw candidate and fall back when
* the pipeline ended up worse β guaranteeing
* final weighted SSE β€ min_c candidate_errors[blk][c]. */
{
float fin_err = 0.0f;
float e_f[QK_K];
for (int j = 0; j < N_SUB; j++) {
float d_sub = dm * (float)(output[blk].scales[j] & 0xF);
float m_sub = mm * (float)(output[blk].scales[j] >> 4);
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
float w = (imat_importance) ?
imat_importance[blk * QK_K + idx] : 1.0f;
e_f[idx] = block_x[idx] -
(d_sub * (float)L[idx] - m_sub);
fin_err += e_f[idx] * e_f[idx] * w;
}
}
fin_err += hex_spectral_penalty(e_f, QK_K);
float g_best = candidate_errors[blk][0];
int g_cand = 0;
for (int c = 1; c < TOTAL_SCALE_CANDIDATES; c++) {
if (candidate_errors[blk][c] < g_best) {
g_best = candidate_errors[blk][c];
g_cand = c;
}
}
if (g_best < fin_err) {
/* Rebuild the block exactly as the candidate was scored */
float c_dm = gguf_fp16_to_fp32(candidate_d [blk][g_cand]);
float c_mm = gguf_fp16_to_fp32(candidate_dmin[blk][g_cand]);
uint8_t c_Ls[16], c_Lm[16];
hex_derive_subscales(seeds[blk].scales, seeds[blk].mins,
c_dm, c_mm, c_Ls, c_Lm);
for (int j = 0; j < N_SUB; j++) {
float d_sub = c_dm * (float)c_Ls[j];
float m_sub = c_mm * (float)c_Lm[j];
for (int k = 0; k < 16; k++) {
int idx = 16 * j + k;
int q = 0;
if (d_sub >= 1e-15f) {
q = gguf_nearest_int((block_x[idx] + m_sub) / d_sub);
if (q < 0) q = 0; if (q > 3) q = 3;
}
L[idx] = (uint8_t)q;
}
output[blk].scales[j] = c_Ls[j] | (c_Lm[j] << 4);
}
dm = c_dm; mm = c_mm;
output[blk].d = candidate_d [blk][g_cand];
output[blk].dmin = candidate_dmin[blk][g_cand];
}
}
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; l++) {
output[blk].qs[j / 4 + l] = L[j + l]
| (L[j + l + 32] << 2)
| (L[j + l + 64] << 4)
| (L[j + l + 96] << 6);
}
}
float berr = gguf_q2_k_block_error(block_x, &output[blk]);
if (isnan(berr)) {
printf("NaN block error at blk %ld! dm=%f mm=%f\n", (long)blk, dm, mm);
for (int j=0; j<16; j++) printf("Ls[%d]=%d Lm[%d]=%d\n", j, Ls_blk[j], j, Lm_blk[j]);
exit(1);
}
total_err += berr;
}
for (int _ti = 0; _ti < _n_omp_threads; _ti++)
hpc_destroy(_tl_graphs[_ti]);
free(_tl_graphs);
free(block_dc_bias);
free(seeds);
free(candidate_errors);
free(candidate_d);
free(candidate_dmin);
free(best_candidate);
if (out_total_error) *out_total_error = total_err;
if (verbose) {
float rmse = sqrtf(total_err / (float)n_elements);
double w_sum2 = 0.0;
for (int64_t i = 0; i < n_elements; i++)
w_sum2 += (double)weights[i] * (double)weights[i];
w_sigma = (float)sqrt(w_sum2 / (double)n_elements);
float rmse_over_sigma = (w_sigma > 1e-15f) ? rmse / w_sigma : 0.0f;
const char *fidelity_class;
const char *fidelity_icon;
if (rmse <= 1.0e-04f) {
fidelity_class = "ULTRA (β€1e-04)";
fidelity_icon = "β
β
β
β
";
} else if (rmse <= 3.0e-04f) {
fidelity_class = "HIGH (β€3e-04)";
fidelity_icon = "β
β
β
β";
} else if (rmse <= 1.0e-03f) {
fidelity_class = "GOOD (β€1e-03)";
fidelity_icon = "β
β
ββ";
} else {
fidelity_class = "STANDARD";
fidelity_icon = "β
βββ";
}
printf("\n βββββ Shor Measurement Q2_K Report βββββββββββββββββββββββββββββββββ\n");
printf(" β Elements: %-12lld Blocks: %-12lld β\n",
(long long)n_elements, (long long)(n_elements / QK_K));
printf(" β Weight Ο: %-12.4e Range: [%.4e, %.4e] β\n",
w_sigma, w_sigma * -4.0f, w_sigma * 4.0f);
printf(" β Total MSE: %-12.6f β\n", total_err);
printf(" β RMSE: %-12.4e RMSE/Ο: %-8.4f β\n",
rmse, rmse_over_sigma);
printf(" β Fidelity: %s %-14s β\n",
fidelity_icon, fidelity_class);
printf(" β Engine: Shor Griffiths-Niu (IDFT6 + feed-forward) β\n");
printf(" βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n");
}
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* PROGRESS REPORTING
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static void print_progress_bar(int current, int total, const char *label,
time_t start_time)
{
if (total <= 0) return;
float pct = (float)current / (float)total;
int bar_width = 40;
int filled = (int)(pct * bar_width);
/* Wall-clock elapsed: clock() sums CPU time over all OpenMP threads,
* which inflated elapsed/ETA by ~the thread count on multicore. */
double elapsed = difftime(time(NULL), start_time);
double eta = (pct > 0.01f) ? elapsed / pct * (1.0 - pct) : 0.0;
printf("\r [");
for (int i = 0; i < bar_width; i++) {
if (i < filled) printf("β");
else if (i == filled) printf("β");
else printf("β");
}
printf("] %3d%% (%d/%d) %.0fs ETA:%.0fs %s",
(int)(pct * 100), current, total, elapsed, eta, label);
fflush(stdout);
if (current == total) printf("\n");
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* GGUF FILE WRITER β Assembles the complete output file
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
static int write_gguf(const char *output_path, const STMultiFile *mf,
const ModelArchitecture *arch,
const TokenizerData *tokenizer,
OptimizerMode opt_mode,
const IMatrixData *imatrix,
int verbose)
{
FILE *fp = fopen(output_path, "wb");
if (!fp) {
fprintf(stderr, " ERROR: Cannot open '%s' for writing\n", output_path);
return -1;
}
printf("\n ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n");
printf(" β WRITING GGUF FILE β\n");
printf(" ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n\n");
/* ββ Determine which tensors to include ββ */
int *include_list = (int *)calloc(mf->n_tensors, sizeof(int));
int n_include = 0;
for (int i = 0; i < mf->n_tensors; i++) {
if (!should_skip_tensor(mf->tensor_map[i].name)) {
include_list[n_include++] = i;
} else {
if (verbose) printf(" SKIP: %s (not needed in GGUF)\n", mf->tensor_map[i].name);
}
}
/* ββ Count metadata KV pairs ββ */
int n_kv = 0;
n_kv++; /* general.architecture */
n_kv++; /* general.name */
n_kv++; /* general.quantization_version */
n_kv++; /* general.file_type */
n_kv++; /* {arch}.context_length */
n_kv++; /* {arch}.embedding_length */
n_kv++; /* {arch}.block_count */
n_kv++; /* {arch}.feed_forward_length */
n_kv++; /* {arch}.attention.head_count */
n_kv++; /* {arch}.attention.head_count_kv */
n_kv++; /* {arch}.attention.layer_norm_rms_epsilon */
n_kv++; /* {arch}.rope.freq_base */
n_kv++; /* {arch}.vocab_size */
/* Tokenizer metadata KV count */
int has_tokenizer = (tokenizer != NULL && tokenizer->vocab_size > 0);
if (has_tokenizer) {
n_kv++; /* tokenizer.ggml.model */
n_kv++; /* tokenizer.ggml.tokens */
n_kv++; /* tokenizer.ggml.scores */
n_kv++; /* tokenizer.ggml.token_type */
n_kv++; /* tokenizer.ggml.bos_token_id */
n_kv++; /* tokenizer.ggml.eos_token_id */
n_kv++; /* tokenizer.ggml.unknown_token_id */
if (tokenizer->n_merges > 0)
n_kv++; /* tokenizer.ggml.merges */
}
/* ββ Check for weight tying ββ
* If tie_word_embeddings is set and there's no separate lm_head,
* llama.cpp handles this internally β do NOT duplicate the tensor.
* Only add output.weight if the model has a separate lm_head.weight. */
int has_lm_head = (st_multi_find_tensor(mf, "lm_head.weight") >= 0);
int total_tensors = n_include;
if (arch->tie_word_embeddings && !has_lm_head) {
printf(" Weight-tied embeddings detected β llama.cpp handles internally\n\n");
}
/* ββ Prepare tensor info ββ */
char (*gguf_names)[ST_MAX_NAME_LEN] = calloc(total_tensors, ST_MAX_NAME_LEN);
GGMLType *tensor_types = calloc(total_tensors, sizeof(GGMLType));
int64_t *tensor_sizes = calloc(total_tensors, sizeof(int64_t));
uint64_t data_offset = 0;
uint64_t *tensor_offsets = calloc(total_tensors, sizeof(uint64_t));
int *tensor_src_idx = calloc(total_tensors, sizeof(int)); /* map to unified ST index */
char (*tensor_hf_names)[ST_MAX_NAME_LEN] = calloc(total_tensors, ST_MAX_NAME_LEN);
GGMLType quant_type = GGML_TYPE_Q2_K;
for (int i = 0; i < n_include; i++) {
int src = include_list[i];
const STTensorInfo *ti = st_multi_tensor_info(mf, src);
map_tensor_name(mf->tensor_map[src].name, gguf_names[i], ST_MAX_NAME_LEN);
strncpy(tensor_hf_names[i], mf->tensor_map[src].name, ST_MAX_NAME_LEN - 1);
tensor_src_idx[i] = src;
if (should_quantize(ti, gguf_names[i])) {
if (is_attention_tensor(gguf_names[i])) {
tensor_types[i] = GGML_TYPE_Q4_0;
int64_t n_blocks_q4 = (ti->n_elements + QK4_0 - 1) / QK4_0;
tensor_sizes[i] = n_blocks_q4 * sizeof(BlockQ4_0);
if (verbose)
printf(" [ATTNβQ4_0] %s (%ld elements)\n",
gguf_names[i], (long)ti->n_elements);
} else {
tensor_types[i] = quant_type;
tensor_sizes[i] = ggml_type_size(quant_type, ti->n_elements);
}
} else if (ti->n_dims >= 2) {
tensor_types[i] = GGML_TYPE_F16;
tensor_sizes[i] = ti->n_elements * sizeof(uint16_t);
} else {
tensor_types[i] = GGML_TYPE_F32;
tensor_sizes[i] = ti->n_elements * sizeof(float);
}
tensor_offsets[i] = data_offset;
data_offset += tensor_sizes[i];
data_offset = (data_offset + GGUF_DEFAULT_ALIGNMENT - 1) &
~(uint64_t)(GGUF_DEFAULT_ALIGNMENT - 1);
}
/* ββ Write header ββ */
gguf_write_header(fp, total_tensors, n_kv);
/* ββ Write metadata KV pairs ββ */
gguf_write_kv_string(fp, "general.architecture", arch->architecture);
gguf_write_kv_string(fp, "general.name", arch->name);
gguf_write_kv_uint32(fp, "general.quantization_version", 2);
gguf_write_kv_uint32(fp, "general.file_type", 10); /* Q2_K = 10 */
char kbuf[128];
snprintf(kbuf, sizeof(kbuf), "%s.context_length", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->context_length);
snprintf(kbuf, sizeof(kbuf), "%s.embedding_length", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->embedding_length);
snprintf(kbuf, sizeof(kbuf), "%s.block_count", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->block_count);
snprintf(kbuf, sizeof(kbuf), "%s.feed_forward_length", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->feed_forward_length);
snprintf(kbuf, sizeof(kbuf), "%s.attention.head_count", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->head_count);
snprintf(kbuf, sizeof(kbuf), "%s.attention.head_count_kv", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->head_count_kv);
snprintf(kbuf, sizeof(kbuf), "%s.attention.layer_norm_rms_epsilon", arch->architecture);
gguf_write_kv_float32(fp, kbuf, arch->rms_norm_eps);
snprintf(kbuf, sizeof(kbuf), "%s.rope.freq_base", arch->architecture);
gguf_write_kv_float32(fp, kbuf, arch->rope_freq_base);
snprintf(kbuf, sizeof(kbuf), "%s.vocab_size", arch->architecture);
gguf_write_kv_uint32(fp, kbuf, arch->vocab_size);
/* ββ Write tokenizer metadata ββ */
if (has_tokenizer) {
gguf_write_kv_string(fp, "tokenizer.ggml.model", tokenizer->model_type);
gguf_write_kv_string_array(fp, "tokenizer.ggml.tokens",
(const char **)tokenizer->tokens,
(uint64_t)tokenizer->vocab_size);
gguf_write_kv_float32_array(fp, "tokenizer.ggml.scores",
tokenizer->scores,
(uint64_t)tokenizer->vocab_size);
gguf_write_kv_int32_array(fp, "tokenizer.ggml.token_type",
tokenizer->token_types,
(uint64_t)tokenizer->vocab_size);
gguf_write_kv_uint32(fp, "tokenizer.ggml.bos_token_id",
(uint32_t)tokenizer->bos_id);
gguf_write_kv_uint32(fp, "tokenizer.ggml.eos_token_id",
(uint32_t)tokenizer->eos_id);
gguf_write_kv_uint32(fp, "tokenizer.ggml.unknown_token_id",
(uint32_t)tokenizer->unk_id);
if (tokenizer->n_merges > 0) {
gguf_write_kv_string_array(fp, "tokenizer.ggml.merges",
(const char **)tokenizer->merges,
(uint64_t)tokenizer->n_merges);
}
printf(" Tokenizer metadata written (%d tokens, %d merges)\n\n",
tokenizer->vocab_size, tokenizer->n_merges);
}
/* ββ Write tensor info descriptors ββ */
for (int i = 0; i < total_tensors; i++) {
int src = tensor_src_idx[i];
const STTensorInfo *ti = st_multi_tensor_info(mf, src);
uint64_t dims[ST_MAX_DIMS];
int nd = ti->n_dims;
for (int d = 0; d < nd; d++) {
dims[d] = (uint64_t)ti->shape[nd - 1 - d];
}
gguf_write_tensor_info(fp, gguf_names[i],
ti->n_dims, dims,
tensor_types[i], tensor_offsets[i]);
}
/* ββ Alignment padding before data section ββ */
gguf_write_padding(fp, GGUF_DEFAULT_ALIGNMENT);
/* ββ Write tensor data ββ */
printf(" Quantizing and writing %d tensors...\n\n", total_tensors);
float total_error_sum = 0.0f;
int quant_count = 0;
int64_t total_elements_quantized = 0;
int64_t total_bytes_quantized = 0;
int64_t total_bytes_unquantized = 0;
time_t quant_start = time(NULL);
for (int i = 0; i < total_tensors; i++) {
int src = tensor_src_idx[i];
const STTensorInfo *ti = st_multi_tensor_info(mf, src);
print_progress_bar(i, total_tensors, gguf_names[i], quant_start);
if (tensor_types[i] == GGML_TYPE_Q2_K) {
float *f32_data = st_multi_tensor_to_f32(mf, src);
if (!f32_data) {
fprintf(stderr, "\n ERROR: Failed to convert tensor '%s' to F32\n",
ti->name);
continue;
}
int64_t n_elements = ti->n_elements;
float tensor_error = 0.0f;
int64_t padded = (n_elements + QK_K - 1) / QK_K * QK_K;
if (padded > n_elements) {
float *grown = realloc(f32_data, padded * sizeof(float));
if (!grown) {
fprintf(stderr, "\n ERROR: Out of memory padding '%s'\n",
ti->name);
free(f32_data);
continue;
}
f32_data = grown;
for (int64_t j = n_elements; j < padded; j++)
f32_data[j] = 0.0f;
n_elements = padded;
}
int64_t n_blocks = n_elements / QK_K;
BlockQ2K *quant_data = calloc(n_blocks, sizeof(BlockQ2K));
const float *imp = NULL;
if (imatrix) {
const IMatrixEntry *ime = imatrix_find_any(imatrix,
gguf_names[i], tensor_hf_names[i]);
if (ime && ime->n_values > 0) {
imp = ime->normalized;
if (verbose)
printf("\n imatrix: using %d importance weights for %s\n",
ime->n_values, gguf_names[i]);
}
}
quantize_tensor_q2k_hpc(f32_data, n_elements,
quant_data, &tensor_error,
opt_mode, imp, verbose);
fwrite(quant_data, sizeof(BlockQ2K), n_blocks, fp);
float rmse = sqrtf(tensor_error / (float)ti->n_elements);
double wss = 0.0;
for (int64_t j = 0; j < ti->n_elements; j++)
wss += (double)f32_data[j] * (double)f32_data[j];
float w_sig = (float)sqrt(wss / (double)ti->n_elements);
const char *fid;
if (rmse <= 1.0e-04f) fid = "β
β
β
β
ULTRA";
else if (rmse <= 3.0e-04f) fid = "β
β
β
β HIGH";
else if (rmse <= 1.0e-03f) fid = "β
β
ββ GOOD";
else fid = "β
βββ STD";
if (verbose) {
printf("\n [Q2_KΒ·Shor] %-47s\n", gguf_names[i]);
printf(" %10ld elements β %ld bytes Ο=%.2e RMSE=%.4e %s\n",
(long)ti->n_elements,
(long)(n_blocks * sizeof(BlockQ2K)),
w_sig, rmse, fid);
}
total_error_sum += tensor_error;
total_elements_quantized += ti->n_elements;
total_bytes_quantized += n_blocks * sizeof(BlockQ2K);
quant_count++;
free(quant_data);
free(f32_data);
} else if (tensor_types[i] == GGML_TYPE_Q4_0) {
float *f32_data = st_multi_tensor_to_f32(mf, src);
if (!f32_data) {
fprintf(stderr, "\n ERROR: Failed to convert tensor '%s' to F32\n",
ti->name);
continue;
}
int64_t n_elements = ti->n_elements;
int64_t padded = (n_elements + QK4_0 - 1) / QK4_0 * QK4_0;
if (padded > n_elements) {
float *grown = realloc(f32_data, padded * sizeof(float));
if (!grown) {
fprintf(stderr, "\n ERROR: Out of memory padding '%s'\n",
ti->name);
free(f32_data);
continue;
}
f32_data = grown;
for (int64_t j = n_elements; j < padded; j++)
f32_data[j] = 0.0f;
n_elements = padded;
}
int64_t n_blocks_q4 = n_elements / QK4_0;
BlockQ4_0 *q4_data = calloc(n_blocks_q4, sizeof(BlockQ4_0));
float tensor_error = 0.0f;
const float *imp = NULL;
if (imatrix) {
const IMatrixEntry *ime = imatrix_find_any(imatrix,
gguf_names[i], tensor_hf_names[i]);
if (ime && ime->n_values > 0) {
imp = ime->normalized;
if (verbose)
printf("\n imatrix: using %d importance weights for %s\n",
ime->n_values, gguf_names[i]);
}
}
quantize_tensor_q4_0_hpc(f32_data, n_elements,
q4_data, &tensor_error,
imp, verbose);
fwrite(q4_data, sizeof(BlockQ4_0), n_blocks_q4, fp);
float rmse = sqrtf(tensor_error / (float)ti->n_elements);
double wss4 = 0.0;
for (int64_t j = 0; j < ti->n_elements; j++)
wss4 += (double)f32_data[j] * (double)f32_data[j];
float w_sig4 = (float)sqrt(wss4 / (double)ti->n_elements);
const char *fid4;
if (rmse <= 1.0e-04f) fid4 = "β
β
β
β
ULTRA";
else if (rmse <= 3.0e-04f) fid4 = "β
β
β
β HIGH";
else if (rmse <= 1.0e-03f) fid4 = "β
β
ββ GOOD";
else fid4 = "β
βββ STD";
if (verbose) {
printf("\n [Q4_0Β·Shor] %-47s\n", gguf_names[i]);
printf(" %10ld elements β %ld bytes Ο=%.2e RMSE=%.4e %s\n",
(long)ti->n_elements,
(long)(n_blocks_q4 * sizeof(BlockQ4_0)),
w_sig4, rmse, fid4);
}
total_error_sum += tensor_error;
total_elements_quantized += ti->n_elements;
total_bytes_quantized += n_blocks_q4 * sizeof(BlockQ4_0);
quant_count++;
free(q4_data);
free(f32_data);
} else if (tensor_types[i] == GGML_TYPE_F16) {
float *f32_data = st_multi_tensor_to_f32(mf, src);
if (!f32_data) {
fprintf(stderr, "\n ERROR: Failed to convert tensor '%s'\n",
ti->name);
continue;
}
uint16_t *f16_data = (uint16_t *)malloc(ti->n_elements * sizeof(uint16_t));
for (int64_t j = 0; j < ti->n_elements; j++)
f16_data[j] = gguf_fp32_to_fp16(f32_data[j]);
fwrite(f16_data, sizeof(uint16_t), ti->n_elements, fp);
total_bytes_unquantized += ti->n_elements * sizeof(uint16_t);
if (verbose) {
printf("\n [F16 ] %-50s %10ld elements β %ld bytes\n",
gguf_names[i], (long)ti->n_elements,
(long)(ti->n_elements * sizeof(uint16_t)));
}
free(f16_data);
free(f32_data);
} else {
float *f32_data = st_multi_tensor_to_f32(mf, src);
if (!f32_data) {
fprintf(stderr, "\n ERROR: Failed to convert tensor '%s'\n",
ti->name);
continue;
}
fwrite(f32_data, sizeof(float), ti->n_elements, fp);
total_bytes_unquantized += ti->n_elements * sizeof(float);
if (verbose) {
printf("\n [F32 ] %-50s %10ld elements β %ld bytes\n",
gguf_names[i], (long)ti->n_elements,
(long)(ti->n_elements * sizeof(float)));
}
free(f32_data);
}
gguf_write_padding(fp, GGUF_DEFAULT_ALIGNMENT);
}
print_progress_bar(total_tensors, total_tensors, "done", quant_start);
long final_size = ftell(fp);
fclose(fp);
int64_t original_f32_size = 0;
for (int i = 0; i < total_tensors; i++) {
const STTensorInfo *ti = st_multi_tensor_info(mf, tensor_src_idx[i]);
original_f32_size += ti->n_elements * sizeof(float);
}
float compression_ratio = (original_f32_size > 0) ?
(float)original_f32_size / (float)final_size : 0.0f;
float effective_bpw = (total_elements_quantized > 0) ?
8.0f * (float)total_bytes_quantized / (float)total_elements_quantized :
0.0f;
float total_rmse = (total_elements_quantized > 0) ?
sqrtf(total_error_sum / (float)total_elements_quantized) : 0.0f;
float mean_mse_per_tensor = (quant_count > 0) ?
total_error_sum / (float)quant_count : 0.0f;
const char *overall_fid, *overall_icon;
if (total_rmse <= 1.0e-04f) { overall_fid = "ULTRA (β€1e-04)"; overall_icon = "β
β
β
β
"; }
else if (total_rmse <= 3.0e-04f) { overall_fid = "HIGH (β€3e-04)"; overall_icon = "β
β
β
β"; }
else if (total_rmse <= 1.0e-03f) { overall_fid = "GOOD (β€1e-03)"; overall_icon = "β
β
ββ"; }
else { overall_fid = "STANDARD"; overall_icon = "β
βββ"; }
printf("\n ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n");
printf(" β SHOR-OPTIMIZED QUANTIZATION SUMMARY β\n");
printf(" β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β β\n");
printf(" β Engine: Griffiths-Niu Sequential Measurement β\n");
printf(" β Protocol: IDFT6 β feed-forward β Born β collapse β\n");
printf(" β Origin: tesseract_factor.c (Shor's algorithm) β\n");
printf(" β β\n");
printf(" β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β Tensors quantized: %-33d β\n", quant_count);
printf(" β Elements quantized: %15ld β\n",
(long)total_elements_quantized);
printf(" β Quantized data: %12ld bytes (%6.1f MB) β\n",
(long)total_bytes_quantized,
(double)total_bytes_quantized / (1024.0 * 1024.0));
printf(" β Unquantized data: %12ld bytes (%6.1f MB) β\n",
(long)total_bytes_unquantized,
(double)total_bytes_unquantized / (1024.0 * 1024.0));
printf(" β Effective bits/weight: %15.2f β\n",
effective_bpw);
printf(" β Compression ratio: %15.1fx β\n",
compression_ratio);
printf(" β β\n");
printf(" β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β FIDELITY METRICS (target: 1e-04) β\n");
printf(" β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β β\n");
printf(" β Total MSE: %15.6e β\n",
total_error_sum);
printf(" β Per-element RMSE: %15.4e β\n",
total_rmse);
printf(" β Mean MSE/tensor: %15.6e β\n",
mean_mse_per_tensor);
printf(" β β\n");
printf(" β Fidelity class: %s %-14s β\n",
overall_icon, overall_fid);
if (total_rmse <= 1.0e-04f)
printf(" β β RMSE β€ 1e-04: TARGET MET β maximum fidelity achieved β\n");
else if (total_rmse <= 3.0e-04f)
printf(" β β RMSE β€ 3e-04: near target β high fidelity achieved β\n");
else
printf(" β β RMSE > 3e-04: below target β weight Ο may be large β\n");
printf(" β β\n");
printf(" β βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β Output file: %ld bytes (%.1f MB)%*sβ\n",
final_size, (double)final_size / (1024.0 * 1024.0),
(int)(27 - snprintf(NULL, 0, "%ld bytes (%.1f MB)",
final_size, (double)final_size / (1024.0 * 1024.0))), "");
printf(" ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n\n");
free(include_list);
free(gguf_names);
free(tensor_types);
free(tensor_sizes);
free(tensor_offsets);
free(tensor_src_idx);
free(tensor_hf_names);
return 0;
}
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* LIBRARY API β Exported functions for Python ctypes integration
*
* When built with -DHEXSTATE_LIBRARY, these are the only public symbols.
* The Python GGUF pipeline handles metadata/IO; C handles HPC quantization.
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
/* Initialize HExState subsystems (must be called once before quantization) */
void hexstate_init(void)
{
static int initialized = 0;
if (!initialized) {
srand(42); /* Deterministic for reproducibility */
triality_exotic_init();
s6_exotic_init();
triality_stats_reset();
initialized = 1;
}
}
/* Quantize a single tensor's F32 data to Q2_K using HPC optimization.
*
* Parameters:
* weights: input F32 data (must be padded to multiple of 256)
* n_elements: number of elements (must be multiple of 256)
* output: output buffer (must be n_elements/256 * 84 bytes)
* out_error: pointer to receive total MSE (can be NULL)
* opt_mode: 0=HPC, 1=MSE, 2=Hybrid (recommended)
* verbose: 1 for per-block diagnostics
*/
void hexstate_quantize_tensor_q2k(const float *weights, int64_t n_elements,
void *output, float *out_error,
int opt_mode, int verbose)
{
hexstate_init();
quantize_tensor_q2k_hpc(weights, n_elements,
(BlockQ2K *)output, out_error,
(OptimizerMode)opt_mode, NULL, verbose);
}
/* Same as above but with importance matrix weights */
void hexstate_quantize_tensor_q2k_imat(const float *weights, int64_t n_elements,
void *output, float *out_error,
int opt_mode,
const float *imat_importance,
int verbose)
{
hexstate_init();
quantize_tensor_q2k_hpc(weights, n_elements,
(BlockQ2K *)output, out_error,
(OptimizerMode)opt_mode, imat_importance, verbose);
}
/* Get the block size for Q2_K (84 bytes per 256 elements) */
int hexstate_q2k_block_bytes(void) { return sizeof(BlockQ2K); }
int hexstate_q2k_block_elements(void) { return QK_K; }
/* HPC-optimized Q4_0 quantization for attention tensors.
* Called from Python requantizer via ctypes.
* weights: input F32 weights
* n_elements: number of elements (must be multiple of 32)
* output: output buffer (must be n_elements/32 * 18 bytes)
* out_error: pointer to receive total MSE (can be NULL)
* imat_importance: optional per-element importance weights
* verbose: 1 for per-block diagnostics
*/
void hexstate_quantize_tensor_q4_0_hpc(const float *weights, int64_t n_elements,
void *output, float *out_error,
const float *imat_importance,
int verbose)
{
hexstate_init();
float err = 0.0f;
quantize_tensor_q4_0_hpc(weights, n_elements,
(BlockQ4_0 *)output, &err,
imat_importance, verbose);
if (out_error) *out_error = err;
}
int hexstate_q8_0_block_bytes(void) { return (int)sizeof(hex_block_q8_0); }
int hexstate_q8_0_block_elements(void) { return QK8_0; }
void hexstate_quantize_tensor_q8_0_hpc(const float *weights, int64_t n_elements,
void *output, float *out_error,
const float *imat_importance, int verbose)
{
quantize_tensor_q8_0_hpc(weights, n_elements,
(hex_block_q8_0 *)output, out_error,
imat_importance, verbose);
}
#ifndef HEXSTATE_LIBRARY
/* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
* MAIN
* βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ */
int main(int argc, char **argv)
{
srand(time(NULL));
/* Initialize HExState subsystems */
triality_exotic_init();
s6_exotic_init();
triality_stats_reset();
printf("\n");
printf(" ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n");
printf(" β β\n");
printf(" β HExState GGUF QUANTIZER v3.0 β Shor-Optimized β\n");
printf(" β β\n");
printf(" β Architecture: HPCGraph Sensitivity Propagation β\n");
printf(" β Optimization: Shor's Griffiths-Niu Measurement + iMatrix β\n");
printf(" β Output: GGUF v3 (Q2_K, 2.625 bpw) β\n");
printf(" β β\n");
printf(" β \"The weight and the quantized are opposite faces.\" β\n");
printf(" β β\n");
printf(" ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n\n");
if (argc < 3) {
printf(" Usage: %s <input> <output.gguf> [options]\n\n", argv[0]);
printf(" Input:\n");
printf(" Single .safetensors file, or\n");
printf(" Model directory with sharded .safetensors files\n\n");
printf(" Options:\n");
printf(" --optimizer hpc|mse|hybrid Scale optimization (default: hybrid)\n");
printf(" --imatrix <file> Importance matrix for Q2_K quality\n");
printf(" --config <file> Explicit config.json for arch detection\n");
printf(" --qwen Force Qwen 3.5/3.6 architecture\n");
printf(" --verbose Per-block diagnostics\n\n");
return 1;
}
const char *input_path = argv[1];
const char *output_path = argv[2];
OptimizerMode opt_mode = OPT_HYBRID;
const char *imatrix_path = NULL;
const char *config_override = NULL;
int verbose = 0;
int force_qwen = 0;
/* Parse options */
for (int i = 3; i < argc; i++) {
if (strcmp(argv[i], "--optimizer") == 0 && i + 1 < argc) {
i++;
if (strcmp(argv[i], "hpc") == 0) opt_mode = OPT_HPC;
else if (strcmp(argv[i], "mse") == 0) opt_mode = OPT_MSE;
else if (strcmp(argv[i], "hybrid") == 0) opt_mode = OPT_HYBRID;
else {
fprintf(stderr, " ERROR: Unknown optimizer '%s'. Use hpc, mse, or hybrid.\n", argv[i]);
return 1;
}
} else if (strcmp(argv[i], "--imatrix") == 0 && i + 1 < argc) {
imatrix_path = argv[++i];
} else if (strcmp(argv[i], "--config") == 0 && i + 1 < argc) {
config_override = argv[++i];
} else if (strcmp(argv[i], "--qwen") == 0) {
force_qwen = 1;
} else if (strcmp(argv[i], "--verbose") == 0) {
verbose = 1;
} else {
fprintf(stderr, " ERROR: Unknown option '%s'\n", argv[i]);
return 1;
}
}
const char *opt_names[] = {"HPC (BP only)", "MSE (grid search)", "Hybrid (HPC+MSE)"};
printf(" Input: %s\n", input_path);
printf(" Output: %s\n", output_path);
printf(" Quant type: Q2_K (2.625 bpw)\n");
printf(" Optimizer: %s\n", opt_names[opt_mode]);
if (imatrix_path) printf(" iMatrix: %s\n", imatrix_path);
if (config_override) printf(" Config: %s\n", config_override);
if (force_qwen) printf(" Model: Qwen 3.5/3.6 (forced via --qwen)\n");
printf("\n");
/* ββ Phase 1: Load model ββ */
printf(" Phase 1: Loading model...\n");
time_t t_start = time(NULL);
/* Determine if input is a file or directory */
struct stat st;
if (stat(input_path, &st) != 0) {
fprintf(stderr, " ERROR: Cannot access '%s'\n", input_path);
return 1;
}
STMultiFile *mf = NULL;
char input_dir[512] = "";
if (S_ISDIR(st.st_mode)) {
/* Input is a directory β open all shards */
mf = st_open_dir(input_path);
strncpy(input_dir, input_path, sizeof(input_dir) - 2);
input_dir[sizeof(input_dir) - 2] = '\0';
int dlen = strlen(input_dir);
if (dlen > 0 && input_dir[dlen - 1] != '/') {
input_dir[dlen] = '/';
input_dir[dlen + 1] = '\0';
}
} else {
/* Input is a single file β wrap in STMultiFile */
STFile *sf = st_open(input_path);
if (!sf) {
fprintf(stderr, " ERROR: Failed to open '%s'\n", input_path);
return 1;
}
mf = (STMultiFile *)calloc(1, sizeof(STMultiFile));
mf->shards[0] = sf;
mf->n_shards = 1;
for (int i = 0; i < sf->n_tensors && mf->n_tensors < ST_MAX_TENSORS; i++) {
strncpy(mf->tensor_map[mf->n_tensors].name,
sf->tensors[i].name, ST_MAX_NAME_LEN - 1);
mf->tensor_map[mf->n_tensors].shard_idx = 0;
mf->tensor_map[mf->n_tensors].tensor_idx = i;
mf->n_tensors++;
}
/* Extract directory from file path */
strncpy(input_dir, input_path, sizeof(input_dir) - 1);
input_dir[sizeof(input_dir) - 1] = '\0';
char *last_slash = strrchr(input_dir, '/');
if (last_slash) {
*(last_slash + 1) = '\0';
} else {
strcpy(input_dir, "./");
}
}
if (!mf) {
fprintf(stderr, " ERROR: Failed to load model from '%s'\n", input_path);
return 1;
}
st_multi_print_summary(mf);
time_t t_load = time(NULL);
printf(" Loaded in %.0f seconds\n\n", difftime(t_load, t_start));
/* ββ Phase 2: Detect architecture ββ */
printf(" Phase 2: Detecting model architecture...\n");
/* Try to read config.json from model directory */
char config_path[1024];
snprintf(config_path, sizeof(config_path), "%sconfig.json", input_dir);
const char *config_ptr = NULL;
{
FILE *check = fopen(config_path, "rb");
if (check) {
fclose(check);
config_ptr = config_path;
printf(" Found config.json: %s\n", config_path);
}
}
ModelArchitecture arch;
detect_architecture(mf, &arch, config_ptr);
/* --qwen override: force Qwen 3.5/3.6 architecture parameters */
if (force_qwen) {
strcpy(arch.architecture, "qwen2");
strcpy(arch.name, "Qwen3.6-HExState-Q2K");
printf(" [--qwen] Forcing qwen2-compatible architecture\n");
}
printf(" βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n");
printf(" β Model Architecture β\n");
printf(" β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ£\n");
printf(" β Architecture: %-40s β\n", arch.architecture);
printf(" β Layers: %-40u β\n", arch.block_count);
printf(" β Hidden size: %-40u β\n", arch.embedding_length);
printf(" β Attention heads: %-40u β\n", arch.head_count);
printf(" β KV heads: %-40u β\n", arch.head_count_kv);
printf(" β Vocab size: %-40u β\n", arch.vocab_size);
printf(" β FFN size: %-40u β\n", arch.feed_forward_length);
printf(" β Context length: %-40u β\n", arch.context_length);
printf(" β Has bias: %-40s β\n", arch.has_bias ? "yes" : "no");
printf(" β Tied embeddings: %-40s β\n", arch.tie_word_embeddings ? "yes" : "no");
printf(" βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ\n\n");
/* ββ Phase 2b: Load tokenizer ββ */
printf(" Phase 2b: Loading tokenizer...\n");
TokenizerData *tokenizer = NULL;
{
char tok_json[512], tok_config[512];
snprintf(tok_json, sizeof(tok_json), "%stokenizer.json", input_dir);
snprintf(tok_config, sizeof(tok_config), "%stokenizer_config.json", input_dir);
tokenizer = tok_load(tok_json, tok_config);
if (tokenizer) {
tok_print_summary(tokenizer);
} else {
printf(" No tokenizer found in '%s'\n", input_dir);
printf(" (Output GGUF will lack tokenizer data β not inference-ready)\n\n");
}
}
/* ββ Phase 2c: Load importance matrix (optional) ββ */
IMatrixData *imatrix = NULL;
if (imatrix_path) {
printf(" Phase 2c: Loading importance matrix...\n");
imatrix = imatrix_load(imatrix_path);
if (imatrix) {
imatrix_print_summary(imatrix);
} else {
printf(" WARNING: Failed to load imatrix from '%s'\n", imatrix_path);
printf(" Proceeding without importance weighting.\n\n");
}
}
/* ββ Phase 3-5: Quantize and write GGUF ββ */
printf(" Phase 3: HPC-Optimized Q2_K Quantization + GGUF Output...\n");
int result = write_gguf(output_path, mf, &arch, tokenizer,
opt_mode, imatrix, verbose);
/* Wall-clock total: clock() sums CPU time over all OpenMP threads */
time_t t_end = time(NULL);
printf(" Total time: %.0f seconds\n\n", difftime(t_end, t_start));
if (imatrix) imatrix_free(imatrix);
if (tokenizer) tok_free(tokenizer);
st_multi_close(mf);
return result;
}
#endif /* HEXSTATE_LIBRARY */ |