Spaces:
Running on Zero
Running on Zero
File size: 170,287 Bytes
90ce156 dc79d25 90ce156 35616fd 90ce156 db10d55 c3f6ca2 90ce156 62c372b 90ce156 afd380b 90ce156 96c5ff4 1585384 90ce156 c3f6ca2 90ce156 46a9d20 2a9046b 90ce156 2f79947 90ce156 52300bf b7a0fa0 90ce156 b7a0fa0 90ce156 cf90d70 11e219a 90ce156 afd380b 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 8a696da 90ce156 c3f6ca2 90ce156 62f553a 8dfb377 62f553a 8dfb377 62f553a 2f79947 afd380b 2f79947 afd380b 62f553a 8dfb377 afd380b 2f79947 afd380b 2f79947 62f553a 2f79947 b7a0fa0 2f79947 8dfb377 2f79947 8dfb377 afd380b 2f79947 8dfb377 2f79947 8dfb377 62f553a 2f79947 b7a0fa0 8dfb377 afd380b 2f79947 afd380b 8dfb377 62f553a 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 dc79d25 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 62f553a 2f79947 62f553a 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 afd380b 2f79947 db10d55 2f79947 52300bf 8dfb377 2f79947 afd380b 8dfb377 afd380b 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 8dfb377 2f79947 62f553a 2f79947 62f553a 2f79947 afd380b 2f79947 afd380b 8dfb377 db10d55 29f533f 6ada711 f0ee9e7 cf90d70 90ce156 afd380b 90ce156 2f79947 90ce156 2f79947 90ce156 96c5ff4 b7a0fa0 90ce156 96c5ff4 90ce156 96c5ff4 90ce156 afd380b 90ce156 b7a0fa0 90ce156 624aeae 90ce156 624aeae 90ce156 c3f6ca2 e79d110 c3f6ca2 90ce156 e79d110 90ce156 e79d110 90ce156 624aeae 90ce156 624aeae c3f6ca2 90ce156 e79d110 90ce156 b7a0fa0 90ce156 b7a0fa0 dc79d25 b7a0fa0 90ce156 b7a0fa0 dc79d25 b7a0fa0 90ce156 dc79d25 b7a0fa0 dc79d25 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 dc79d25 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 842f4a7 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 db10d55 b7a0fa0 e18b210 b7a0fa0 db10d55 b7a0fa0 e18b210 b7a0fa0 90ce156 e18b210 dc79d25 e18b210 90ce156 2f79947 90ce156 2f79947 90ce156 2f79947 afd380b 2f79947 afd380b 2f79947 afd380b 2f79947 afd380b 2f79947 afd380b f1aeada db10d55 f1aeada db10d55 f1aeada db10d55 f1aeada db10d55 dc79d25 3586de6 dc79d25 3586de6 dc79d25 3586de6 dc79d25 3586de6 dc79d25 db10d55 dc79d25 db10d55 e18b210 db10d55 dc79d25 db10d55 dc79d25 e18b210 dc79d25 db10d55 e18b210 dc79d25 e18b210 dc79d25 e18b210 dc79d25 e18b210 db10d55 dc79d25 db10d55 dc79d25 3586de6 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 3586de6 dc79d25 db10d55 dc79d25 db10d55 3586de6 db10d55 dc79d25 db10d55 3586de6 db10d55 dc79d25 db10d55 3586de6 db10d55 3586de6 db10d55 dc79d25 db10d55 dc79d25 52300bf db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 dc79d25 db10d55 90ce156 db10d55 90ce156 db10d55 e18b210 db10d55 90ce156 db10d55 90ce156 db10d55 90ce156 db10d55 e18b210 db10d55 afd380b db10d55 afd380b db10d55 90ce156 db10d55 90ce156 db10d55 90ce156 db10d55 e18b210 db10d55 afd380b db10d55 afd380b db10d55 90ce156 5b25029 90ce156 5b25029 90ce156 c65177f 90ce156 c65177f 90ce156 11e219a 90ce156 8a696da 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 afd380b 90ce156 afd380b 90ce156 afd380b 90ce156 afd380b 90ce156 afd380b 90ce156 afd380b 90ce156 96c5ff4 90ce156 cf90d70 90ce156 dc79d25 90ce156 afd380b dc79d25 90ce156 afd380b dc79d25 90ce156 afd380b 90ce156 96c5ff4 90ce156 96c5ff4 90ce156 dc79d25 90ce156 fddaf5e 62c372b 46a9d20 62c372b 46a9d20 62c372b 46a9d20 62c372b 46a9d20 62c372b 46a9d20 62c372b 46a9d20 62c372b 79b62dc fddaf5e 79b62dc b7a0fa0 79b62dc fddaf5e 79b62dc fddaf5e b7a0fa0 96c5ff4 8a696da 96c5ff4 fddaf5e 96c5ff4 8a696da 96c5ff4 8a696da 96c5ff4 8a696da 96c5ff4 8a696da 96c5ff4 8a696da b7a0fa0 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd 96c5ff4 35616fd b7a0fa0 2847bf5 cf90d70 35616fd 2847bf5 90ce156 e18b210 96c5ff4 afd380b 96c5ff4 afd380b 96c5ff4 afd380b 96c5ff4 afd380b 96c5ff4 afd380b 96c5ff4 db10d55 96c5ff4 db10d55 dc79d25 db10d55 dc79d25 db10d55 96c5ff4 2f79947 96c5ff4 db10d55 dc79d25 db10d55 96c5ff4 90ce156 dc79d25 90ce156 96c5ff4 8a696da 90ce156 dc79d25 8a696da 90ce156 1585384 90ce156 62f553a 90ce156 62f553a 90ce156 62f553a 90ce156 62f553a 90ce156 b7a0fa0 90ce156 afd380b b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 842f4a7 cf90d70 b7a0fa0 90ce156 b7a0fa0 90ce156 b7a0fa0 842f4a7 cf90d70 842f4a7 eff87e2 90ce156 b7a0fa0 842f4a7 90ce156 cf90d70 eff87e2 90ce156 b7a0fa0 90ce156 db10d55 90ce156 2f79947 afd380b 90ce156 afd380b 96c5ff4 afd380b 90ce156 afd380b 90ce156 afd380b 96c5ff4 afd380b 96c5ff4 afd380b 90ce156 29f533f eff87e2 29f533f afd380b 90ce156 afd380b cf90d70 afd380b e18b210 afd380b b7a0fa0 90ce156 afd380b 90ce156 6ada711 90ce156 db10d55 90ce156 52300bf f0ee9e7 52300bf 90ce156 b7a0fa0 2f79947 b7a0fa0 2f79947 db10d55 afd380b 2f79947 afd380b 2f79947 afd380b 2f79947 afd380b 2f79947 b7a0fa0 2f79947 afd380b 2f79947 afd380b b7a0fa0 2f79947 db10d55 b7a0fa0 afd380b 90ce156 afd380b 2f79947 afd380b 2f79947 afd380b 90ce156 b7a0fa0 90ce156 db10d55 90ce156 b7a0fa0 90ce156 29f533f b7a0fa0 db10d55 b7a0fa0 db10d55 90ce156 e18b210 90ce156 db10d55 b7a0fa0 db10d55 b7a0fa0 e18b210 b7a0fa0 90ce156 afd380b b7a0fa0 90ce156 dc79d25 90ce156 b7a0fa0 fddaf5e b7a0fa0 90ce156 dc79d25 b7a0fa0 2f79947 b7a0fa0 dc79d25 90ce156 dc79d25 90ce156 afd380b e18b210 90ce156 | 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 | from __future__ import annotations
import argparse
import base64
import concurrent.futures
import gc
import hashlib
import html
import math
import json
import os
import random
import re
import shutil
import subprocess
import sys
import threading
import time
import traceback
from collections import deque
from copy import deepcopy
from datetime import datetime
from pathlib import Path
from typing import Optional
from urllib.parse import quote
os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True,max_split_size_mb:128")
try:
import spaces
except ImportError: # pragma: no cover - keeps local CPU runs working
class _SpacesShim:
@staticmethod
def GPU(*args, **kwargs):
if args and callable(args[0]) and not kwargs:
return args[0]
def decorator(fn):
return fn
return decorator
spaces = _SpacesShim()
import gradio as gr
import torch
from huggingface_hub import snapshot_download
from safetensors import safe_open
from safetensors.torch import load_file, save_file
from transformers import set_seed
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLVisionConfig
from common.utils.logging import get_logger
from common.utils.misc import AutoEncoderParams, tuple_mul
from config.config_factory import DataArguments, InferenceArguments, ModelArguments
from data.data_utils import add_special_tokens
from data.dataset_base import DataConfig, simple_custom_collate
from data.datasets_custom import ValidationDataset
from inference_lance import (
PROMPT_JSON_FILENAME,
apply_inference_defaults,
clean_memory,
init_from_model_path_if_needed,
save_prompt_results,
validate_on_fixed_batch,
)
from modeling.lance import Lance, LanceConfig, Qwen2ForCausalLM
from modeling.qwen2 import Qwen2Tokenizer
from modeling.qwen2.modeling_qwen2 import Qwen2Config
from modeling.vae.wan.model import WanVideoVAE
from modeling.vit.qwen2_5_vl_vit import Qwen2_5_VisionTransformerPretrainedModel
REPO_ROOT = Path(__file__).resolve().parent
GRADIO_TMP_ROOT = Path(os.getenv("LANCE_GRADIO_TMP_ROOT", "/tmp/lance_gradio")).expanduser()
TMP_INPUT_DIR = GRADIO_TMP_ROOT / "inputs"
RESULTS_ROOT = GRADIO_TMP_ROOT / "results"
GLOBAL_RECORDS_FILE = GRADIO_TMP_ROOT / "generation_records.jsonl"
RUN_RECORD_FILENAME = "generation_record.json"
LOCAL_MODEL_BASE_DIR = Path("downloads")
SPACE_MODEL_BASE_DIR = Path("/data/lance_models")
DEFAULT_MODEL_REPO_ID = "bytedance-research/Lance"
DEFAULT_FLASH_ATTN_VERSION = "2.8.3"
DEFAULT_FLASH_ATTN_WHEEL_URL = "https://huggingface.co/strangertoolshf/flash_attention_2_wheelhouse/resolve/main/wheelhouse-flash_attn-2.8.3/linux_x86_64/torch2.8/cu12/abiTRUE/cp310/flash_attn-2.8.3+cu12torch2.8cxx11abiTRUE-cp310-cp310-linux_x86_64.whl"
DEFAULT_MODEL_VARIANT = "video"
MODEL_VARIANT_VIDEO = "video"
MODEL_VARIANT_IMAGE = "image"
MODEL_VARIANT_TO_DIR = {
MODEL_VARIANT_VIDEO: "Lance_3B_Video",
MODEL_VARIANT_IMAGE: "Lance_3B",
}
DEFAULT_MODEL_PATH = LOCAL_MODEL_BASE_DIR / MODEL_VARIANT_TO_DIR[MODEL_VARIANT_VIDEO]
DEFAULT_VIT_TYPE = "qwen_2_5_vl_original"
DEFAULT_TASK = "t2v"
DEFAULT_TIMESTEPS = 30
DEFAULT_TIMESTEP_SHIFT = 3.5
DEFAULT_CFG_TEXT_SCALE = 4.0
DEFAULT_RESOLUTION = "video_480p"
DEFAULT_VIDEO_EDIT_RESOLUTION = "video_480p"
DEFAULT_IMAGE_RESOLUTION = "image_768x768"
DEFAULT_BASIC_SEED = 42
DEFAULT_HEIGHT = 352
DEFAULT_WIDTH = 640
DEFAULT_IMAGE_SIZE = 768
DEFAULT_VIDEO_DURATION_SECONDS = 5
MAX_VIDEO_DURATION_SECONDS = 10
MAX_VIDEO_NUM_FRAMES = 12 * MAX_VIDEO_DURATION_SECONDS + 1
DEFAULT_NUM_FRAMES = 12 * DEFAULT_VIDEO_DURATION_SECONDS + 1
DEFAULT_VIDEO_ASPECT_RATIO = "16:9"
DEFAULT_IMAGE_ASPECT_RATIO = "1:1"
ASPECT_RATIO_CHOICES = ["21:9", "16:9", "3:2", "4:3", "1:1", "3:4", "2:3", "9:16"]
VIDEO_360P_ASPECT_RATIO_TO_SIZE = {
"21:9": (672, 288),
"16:9": (640, 352),
"3:2": (528, 352),
"4:3": (560, 416),
"1:1": (480, 480),
"3:4": (416, 560),
"2:3": (352, 528),
"9:16": (352, 640),
}
VIDEO_480P_ASPECT_RATIO_TO_SIZE = {
"21:9": (976, 416),
"16:9": (848, 480),
"3:2": (784, 528),
"4:3": (736, 560),
"1:1": (640, 640),
"3:4": (560, 736),
"2:3": (528, 784),
"9:16": (480, 848),
}
VIDEO_RESOLUTION_TO_SIZE_MAP = {
"video_360p": VIDEO_360P_ASPECT_RATIO_TO_SIZE,
"video_480p": VIDEO_480P_ASPECT_RATIO_TO_SIZE,
}
IMAGE_ASPECT_RATIO_TO_SIZE = {
"21:9": (1168, 496),
"16:9": (1024, 576),
"3:2": (944, 624),
"4:3": (880, 672),
"1:1": (768, 768),
"3:4": (672, 880),
"2:3": (624, 944),
"9:16": (576, 1024),
}
DEFAULT_GPUS = "0"
DEFAULT_QUEUE_SIZE = 32
DEFAULT_CONCURRENCY_LIMIT = 1
USE_KVCACHE = True
TEXT_TEMPLATE = True
RECORD_WRITE_LOCK = threading.Lock()
LANCE_HOMEPAGE_URL = "https://lance-project.github.io/"
LANCE_PAPER_URL = "http://arxiv.org/abs/2605.18678"
LANCE_HUGGING_FACE_URL = "https://huggingface.co/bytedance-research/Lance"
LANCE_GITHUB_URL = "https://github.com/bytedance/Lance"
LANCE_LOGO_PATH = REPO_ROOT / "assets" / "logo" / "lance-logo.png"
APP_CSS = """
:root {
color-scheme: light;
--lance-accent: #fb923c;
--lance-accent-hover: #f97316;
--lance-surface: #ffffff;
--lance-surface-muted: #f8fafc;
--lance-border: rgba(148, 163, 184, .36);
--lance-text: #111827;
--lance-text-muted: #475569;
--lance-shadow: 0 8px 24px rgba(15, 23, 42, .08);
--body-background-fill: var(--lance-surface);
--background-fill-primary: var(--lance-surface);
--block-background-fill: var(--lance-surface);
--input-background-fill: var(--lance-surface);
--button-primary-background-fill: var(--lance-accent);
--button-primary-background-fill-hover: var(--lance-accent-hover);
--button-primary-text-color: #0f172a;
}
body, .gradio-container, .contain { background: var(--lance-surface) !important; color: var(--lance-text) !important; }
.gradio-container, .contain { max-width: 1180px !important; margin: 0 auto !important; }
.lance-hero { text-align: center; padding: 8px 12px 4px; }
.lance-logo { width: min(150px, 34vw); height: auto; display: block; margin: 0 auto 4px; }
.lance-title { margin: 0 auto 5px; font-size: clamp(22px, 2.4vw, 32px); line-height: 1.08; font-weight: 800; }
.lance-badges { display: flex; flex-wrap: wrap; justify-content: center; gap: 6px; margin: 4px auto 0; }
.lance-badges a { line-height: 0; }
.lance-badges img { height: 20px; width: auto; display: block; }
.lance-status, .lance-run-status { max-width: 1120px; margin: 8px auto !important; }
.lance-run-status p { margin: 0 !important; }
.lance-run-status-pill { display: inline-flex; align-items: center; gap: 8px; padding: 8px 12px; border-radius: 999px; border: 1px solid var(--lance-border); background: var(--lance-surface); color: var(--lance-text-muted); font-size: 14px; font-weight: 700; box-shadow: var(--lance-shadow); }
.lance-run-status-chip { width: 8px; height: 8px; border-radius: 999px; background: var(--lance-accent); box-shadow: 0 0 0 4px rgba(251,146,60,.18); }
.lance-run-status-dots i { display: inline-block; width: 4px; height: 4px; margin-left: 3px; border-radius: 999px; background: currentColor; opacity: .45; animation: lance-dot-pulse 1.1s infinite ease-in-out; }
.lance-run-status-dots i:nth-child(2) { animation-delay: .15s; }
.lance-run-status-dots i:nth-child(3) { animation-delay: .3s; }
@keyframes lance-dot-pulse { 40% { transform: translateY(-1px); opacity: 1; } }
.lance-main-row { display: grid !important; grid-template-columns: minmax(0, 1.16fr) minmax(0, 0.84fr) !important; gap: 18px !important; align-items: start !important; }
.lance-main-column { min-width: 0 !important; width: 100% !important; }
.lance-panel, .lance-control-field, .example-panel { border: 0 !important; box-shadow: none !important; background: transparent !important; padding: 0 !important; }
.lance-panel > .form, .lance-control-field > .form, .lance-label-html, .lance-label-html > div, .lance-label-html .wrap { border: 0 !important; background: transparent !important; box-shadow: none !important; padding: 0 !important; margin: 0 !important; min-height: 0 !important; }
.lance-section-label, .lance-generation-label { margin: 0 0 10px !important; font-weight: 800 !important; color: var(--body-text-color) !important; }
.lance-section-label { font-size: 18px !important; }
.lance-generation-label { font-size: 14px !important; }
.lance-label-icon { display: none !important; }
.lance-output-label { display: inline-flex !important; align-items: center !important; gap: 8px !important; }
.lance-output-label .lance-label-icon { display: inline-flex !important; align-items: center !important; justify-content: center !important; width: 20px !important; height: 20px !important; color: var(--lance-accent) !important; }
.lance-output-label .lance-label-icon svg { width: 18px !important; height: 18px !important; display: block !important; }
.lance-taskbar-wrap { max-width: 1120px; margin: 0 auto 12px !important; }
.task-selector {
overflow-x: auto !important;
padding: 4px 0 12px !important;
scrollbar-width: thin;
display: flex !important;
justify-content: center !important;
}
.task-selector > .wrap, .task-selector .wrap {
width: max-content !important;
max-width: min(100%, 1080px) !important;
margin: 0 auto !important;
padding: 4px !important;
display: flex !important;
justify-content: center !important;
flex-wrap: nowrap !important;
gap: 10px !important;
border-radius: 999px !important;
background: transparent !important;
border: 0 !important;
box-shadow: none !important;
}
.task-selector label {
min-width: max-content !important;
min-height: 38px !important;
padding: 9px 18px !important;
border: 0 !important;
border-radius: 999px !important;
background: #f1f5f9 !important;
color: var(--lance-text-muted) !important;
justify-content: center !important;
white-space: nowrap !important;
}
.task-selector label:has(input:checked) { background: var(--lance-accent) !important; color: #0f172a !important; box-shadow: 0 6px 16px rgba(251,146,60,.22) !important; }
.task-selector input:checked + span { color: #0f172a !important; font-weight: 800 !important; }
.lance-taskbar-wrap,
.lance-taskbar-wrap > div,
.lance-taskbar-wrap > .form,
.lance-taskbar-wrap .block,
.task-selector,
.task-selector > div,
.task-selector > .form,
.task-selector .form,
.task-selector .wrap {
background: transparent !important;
border: 0 !important;
box-shadow: none !important;
}
.task-selector > .wrap,
.task-selector .wrap {
padding: 0 !important;
}
.task-selector label {
background: #f8fafc !important;
border: 1px solid rgba(148,163,184,.25) !important;
box-shadow: 0 3px 10px rgba(15,23,42,.04) !important;
}
.task-selector label:has(input:checked) {
background: var(--lance-accent) !important;
border-color: transparent !important;
color: #0f172a !important;
box-shadow: 0 8px 18px rgba(249,115,22,.24) !important;
}
.task-selector input:checked + span { color: #0f172a !important; }
.lance-task-prompt-panel { max-width: 1040px; margin: 0 auto 10px !important; }
.main-prompt-control, .main-prompt-control > div, .main-prompt-control .wrap { border: 0 !important; background: transparent !important; box-shadow: none !important; }
.main-prompt-control textarea { min-height: 160px !important; padding: 18px !important; border: 1px solid var(--lance-border) !important; border-radius: 16px !important; background: var(--lance-surface) !important; color: var(--lance-text) !important; font-size: 15px !important; line-height: 1.45 !important; box-shadow: var(--lance-shadow) !important; }
.main-prompt-control textarea::placeholder { color: #94a3b8 !important; }
.prompt-options {
position: relative !important;
z-index: 2 !important;
margin: 8px 0 16px !important;
padding: 0 !important;
}
.prompt-options > .form {
display: grid !important;
grid-template-columns: repeat(4, max-content) !important;
align-items: center !important;
justify-content: start !important;
justify-items: start !important;
gap: 6px !important;
width: max-content !important;
max-width: 100% !important;
}
.prompt-chip,
.prompt-chip > .form,
.prompt-chip > div,
.prompt-chip .block,
.prompt-chip .form,
.prompt-chip .container,
.prompt-chip .wrap {
width: 100% !important;
min-width: 0 !important;
background: transparent !important;
border: 0 !important;
box-shadow: none !important;
padding: 0 !important;
margin: 0 !important;
}
.prompt-chip {
display: block !important;
min-width: 0 !important;
width: auto !important;
flex: 0 0 auto !important;
}
.prompt-chip .wrap,
.prompt-chip .container,
.prompt-chip > .form,
.prompt-chip .form {
display: inline-flex !important;
align-items: center !important;
width: auto !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
width: auto !important;
min-width: 58px !important;
min-height: 32px !important;
height: 32px !important;
border-radius: 999px !important;
border: 1px solid var(--lance-border) !important;
outline: 0 !important;
background: var(--lance-surface-muted) !important;
color: var(--lance-text) !important;
font-size: 10px !important;
font-weight: 800 !important;
box-shadow: none !important;
padding: 0 8px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input { min-width: 82px !important; }
.video-resolution-row button,
.video-resolution-row [role="button"],
.video-resolution-row select,
.video-resolution-row input { min-width: 58px !important; }
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input { min-width: 48px !important; }
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input { min-width: 44px !important; }
.output-resolution-row button,
.output-resolution-row [role="button"],
.output-resolution-row select,
.output-resolution-row input { min-width: 70px !important; }
.prompt-chip button,
.prompt-chip [role="button"] { white-space: nowrap !important; }
.prompt-chip .icon-wrap,
.prompt-chip .select-arrow,
.prompt-chip .label-wrap,
.prompt-chip .block-title,
.prompt-chip .block-info,
.prompt-chip label {
background: transparent !important;
border: 0 !important;
box-shadow: none !important;
}
@media (max-width: 1200px) {
.lance-main-row { grid-template-columns: minmax(0, 1.24fr) minmax(0, 0.76fr) !important; }
.prompt-options > .form {
grid-template-columns: repeat(4, max-content) !important;
justify-content: start !important;
gap: 4px !important;
}
.prompt-chip button, .prompt-chip [role="button"], .prompt-chip select, .prompt-chip input {
font-size: 9.5px !important;
min-width: 50px !important;
padding: 0 6px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input { min-width: 76px !important; }
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input { min-width: 42px !important; }
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input { min-width: 40px !important; }
}
.prompt-options {
margin: 8px 0 16px !important;
padding: 0 !important;
}
.prompt-options > .form {
display: inline-flex !important;
flex-wrap: nowrap !important;
justify-content: flex-start !important;
justify-items: start !important;
align-items: center !important;
gap: 6px !important;
width: auto !important;
max-width: 100% !important;
}
.prompt-chip,
.prompt-chip > .form,
.prompt-chip > div,
.prompt-chip .block,
.prompt-chip .form,
.prompt-chip .container,
.prompt-chip .wrap {
width: auto !important;
min-width: 0 !important;
max-width: none !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
width: auto !important;
min-width: 0 !important;
height: 30px !important;
min-height: 30px !important;
font-size: 9.5px !important;
padding: 0 8px !important;
border-radius: 999px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input { min-width: 74px !important; max-width: 82px !important; }
.video-resolution-row button,
.video-resolution-row [role="button"],
.video-resolution-row select,
.video-resolution-row input { min-width: 50px !important; max-width: 58px !important; }
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input { min-width: 44px !important; max-width: 52px !important; }
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input { min-width: 38px !important; max-width: 46px !important; }
.output-resolution-row button,
.output-resolution-row [role="button"],
.output-resolution-row select,
.output-resolution-row input { min-width: 64px !important; max-width: 80px !important; }
@media (max-width: 1200px) {
.prompt-options > .form {
display: inline-flex !important;
flex-wrap: nowrap !important;
justify-content: flex-start !important;
gap: 4px !important;
width: auto !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
font-size: 9px !important;
padding: 0 6px !important;
height: 29px !important;
min-height: 29px !important;
}
}
.lance-display-frame, .lance-display-frame > div, .lance-display-frame textarea, .output-media-control { width: 100% !important; }
.lance-output-panel { background: transparent !important; }
.lance-output-panel .lance-display-frame > div,
.lance-output-panel .lance-display-frame .wrap,
.lance-output-panel .output-media-control,
.lance-output-panel .output-media-control > div {
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
padding: 0 !important;
}
.lance-output-panel .output-media-control video,
.lance-output-panel .output-media-control img,
.lance-output-panel .lance-display-frame textarea {
border-radius: 18px !important;
border: 1px solid rgba(116, 126, 140, .34) !important;
background: linear-gradient(180deg, rgba(250,251,253,.94), rgba(244,246,249,.9)) !important;
box-shadow: 0 10px 28px rgba(15,23,42,.10), inset 0 0 0 1px rgba(255,255,255,.75) !important;
}
.lance-output-panel .lance-display-frame textarea { color: #101828 !important; }
.output-media-control video, .output-media-control img { border-radius: 18px !important; }
.lance-run-button { max-width: 1040px !important; margin: 10px auto 16px !important; border-radius: 12px !important; font-size: 18px !important; font-weight: 800 !important; }
.lance-quota-note {
max-width: 1040px !important;
margin: -8px auto 16px !important;
text-align: center !important;
color: var(--lance-text-muted) !important;
font-size: 13px !important;
line-height: 1.45 !important;
}
.lance-quota-note p {
margin: 0 !important;
}
button.lance-run-button, .lance-run-button button { width: 100% !important; border: 0 !important; border-radius: 12px !important; background: var(--lance-accent) !important; color: #0f172a !important; font-size: 18px !important; font-weight: 800 !important; box-shadow: 0 10px 24px rgba(249,115,22,.22) !important; }
button.lance-run-button:hover, .lance-run-button button:hover { background: var(--lance-accent-hover) !important; color: #0f172a !important; }
button.lance-run-button, .lance-run-button button {
background: var(--lance-accent) !important;
color: #0f172a !important;
box-shadow: 0 10px 24px rgba(249,115,22,.22) !important;
}
button.lance-run-button:hover, .lance-run-button button:hover {
background: var(--lance-accent-hover) !important;
color: #0f172a !important;
}
.lance-advanced-accordion { max-width: 1040px; margin: 8px auto 0 !important; }
.lance-advanced-accordion .label-wrap, .lance-advanced-accordion summary { font-weight: 800 !important; }
.lance-recommended-section { max-width: 1040px; margin: 20px auto 0 !important; }
.lance-recommended-section .lance-section-label { text-align: left !important; font-size: 20px !important; margin-bottom: 12px !important; }
.prompt-example-full-table {
max-height: 420px !important;
overflow: auto !important;
border: 1px solid rgba(148,163,184,.24) !important;
border-radius: 18px !important;
background: linear-gradient(180deg, #ffffff, #f8fafc) !important;
box-shadow: 0 12px 28px rgba(15,23,42,.07) !important;
padding: 12px !important;
}
.prompt-example-full-table > .form { gap: 10px !important; }
.prompt-examples .prompt-example-row-button,
.prompt-examples .prompt-example-row-button button {
width: 100% !important;
height: auto !important;
min-height: 52px !important;
max-height: 150px !important;
padding: 12px 14px !important;
border: 1px solid rgba(148,163,184,.22) !important;
border-radius: 14px !important;
background: #fff !important;
color: var(--lance-text) !important;
text-align: left !important;
justify-content: flex-start !important;
align-items: flex-start !important;
white-space: normal !important;
overflow-y: auto !important;
box-shadow: 0 6px 16px rgba(15,23,42,.045) !important;
transition: transform .12s ease, box-shadow .12s ease, border-color .12s ease !important;
}
.prompt-examples .prompt-example-row-button:hover,
.prompt-examples .prompt-example-row-button button:hover {
transform: translateY(-1px) !important;
border-color: rgba(251,146,60,.48) !important;
box-shadow: 0 10px 22px rgba(15,23,42,.075) !important;
}
.prompt-examples .prompt-example-row-button span,
.prompt-examples .prompt-example-row-button p,
.prompt-examples .prompt-example-row-button div {
white-space: pre-wrap !important;
overflow-wrap: anywhere !important;
word-break: break-word !important;
line-height: 1.38 !important;
color: var(--lance-text) !important;
}
.prompt-example-multimodal-row,
.prompt-example-multimodal-row > .form {
width: 100% !important;
min-width: 0 !important;
margin: 0 !important;
gap: 12px !important;
align-items: stretch !important;
}
.prompt-example-multimodal-row > .form {
display: grid !important;
grid-template-columns: minmax(0, 1fr) 230px !important;
padding: 8px !important;
border: 1px solid rgba(148,163,184,.20) !important;
border-radius: 16px !important;
background: #fff !important;
box-shadow: 0 6px 16px rgba(15,23,42,.045) !important;
}
.prompt-example-prompt-cell,
.prompt-example-prompt-cell > .form,
.prompt-example-media-cell,
.prompt-example-media-cell > .form {
min-width: 0 !important;
width: 100% !important;
margin: 0 !important;
padding: 0 !important;
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
}
.prompt-example-multimodal-row .prompt-example-row-button,
.prompt-example-multimodal-row .prompt-example-row-button button {
height: 100% !important;
min-height: 132px !important;
max-height: 132px !important;
border: 0 !important;
box-shadow: none !important;
background: #f8fafc !important;
}
.prompt-example-media-html,
.prompt-example-media-html > div,
.prompt-example-media-html .wrap {
width: 100% !important;
height: 132px !important;
min-height: 132px !important;
max-height: 132px !important;
margin: 0 !important;
padding: 0 !important;
border: 1px solid rgba(148,163,184,.22) !important;
border-radius: 14px !important;
background: #fff !important;
box-shadow: none !important;
overflow: hidden !important;
}
.prompt-example-media-html video,
.prompt-example-media-html img,
.example-preview-video,
.example-preview-image {
width: 100% !important;
height: 132px !important;
border-radius: 12px !important;
display: block !important;
background: var(--lance-surface-muted) !important;
object-fit: contain !important;
object-position: center center !important;
}
.reference-media-fallback {
width: 100% !important;
height: 132px !important;
border-radius: 12px !important;
display: flex !important;
align-items: center !important;
justify-content: center !important;
background: var(--lance-surface-muted) !important;
color: var(--lance-text-muted) !important;
font-size: 12px !important;
font-weight: 700 !important;
text-align: center !important;
}
@media (max-width: 760px) {
.prompt-example-multimodal-row > .form { grid-template-columns: minmax(0, 1fr) 140px !important; }
.prompt-example-multimodal-row .prompt-example-row-button,
.prompt-example-multimodal-row .prompt-example-row-button button,
.prompt-example-media-html,
.prompt-example-media-html > div,
.prompt-example-media-html .wrap,
.prompt-example-media-html video,
.prompt-example-media-html img,
.example-preview-video,
.example-preview-image {
height: 108px !important;
min-height: 108px !important;
max-height: 108px !important;
}
}
@media (max-width: 900px) { .lance-main-row { grid-template-columns: minmax(0, 1fr) !important; } .prompt-options { margin-top: 8px !important; } }
.prompt-example-full-table {
max-height: none !important;
overflow: visible !important;
padding: 18px !important;
}
.prompt-example-full-table > .form {
gap: 18px !important;
}
.prompt-examples .prompt-example-row-button,
.prompt-examples .prompt-example-row-button button {
min-height: 168px !important;
height: auto !important;
max-height: none !important;
padding: 22px 24px !important;
line-height: 1.62 !important;
overflow: hidden !important;
display: flex !important;
align-items: flex-start !important;
}
.prompt-examples .prompt-example-row-button span,
.prompt-examples .prompt-example-row-button p,
.prompt-examples .prompt-example-row-button div {
line-height: 1.62 !important;
overflow: hidden !important;
}
.prompt-example-multimodal-row .prompt-example-row-button,
.prompt-example-multimodal-row .prompt-example-row-button button,
.prompt-example-media-html,
.prompt-example-media-html > div,
.prompt-example-media-html .wrap,
.prompt-example-media-html video,
.prompt-example-media-html img,
.example-preview-video,
.example-preview-image,
.reference-media-fallback {
min-height: 160px !important;
height: 160px !important;
max-height: 160px !important;
}
.prompt-example-full-table {
max-height: 560px !important;
}
.prompt-examples .prompt-example-row-button,
.prompt-examples .prompt-example-row-button button {
min-height: 96px !important;
max-height: none !important;
padding: 18px 20px !important;
overflow-y: visible !important;
}
.prompt-examples .prompt-example-row-button span,
.prompt-examples .prompt-example-row-button p,
.prompt-examples .prompt-example-row-button div {
line-height: 1.55 !important;
}
.task-selector label:has(input:checked) {
box-shadow: 0 4px 10px rgba(249,115,22,.12) !important;
}
.prompt-options {
margin: 5px 0 14px !important;
}
.prompt-options > .form {
gap: 7px !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
height: 31px !important;
min-height: 31px !important;
font-size: 10.5px !important;
padding: 0 9px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input { min-width: 78px !important; max-width: 88px !important; }
.video-resolution-row button,
.video-resolution-row [role="button"],
.video-resolution-row select,
.video-resolution-row input { min-width: 54px !important; max-width: 62px !important; }
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input { min-width: 48px !important; max-width: 56px !important; }
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input { min-width: 42px !important; max-width: 50px !important; }
.output-resolution-row button,
.output-resolution-row [role="button"],
.output-resolution-row select,
.output-resolution-row input { min-width: 68px !important; max-width: 86px !important; }
.lance-recommended-section { margin-top: 24px !important; }
.prompt-example-full-table {
max-height: 480px !important;
padding: 16px !important;
}
.prompt-example-full-table > .form {
gap: 12px !important;
}
.prompt-examples .prompt-example-row-button,
.prompt-examples .prompt-example-row-button button {
min-height: 66px !important;
padding: 16px 18px !important;
line-height: 1.48 !important;
}
.prompt-examples .prompt-example-row-button span,
.prompt-examples .prompt-example-row-button p,
.prompt-examples .prompt-example-row-button div {
line-height: 1.48 !important;
}
.prompt-example-multimodal-row,
.prompt-example-multimodal-row > .form {
gap: 14px !important;
}
.prompt-example-multimodal-row > .form {
padding: 12px !important;
}
.prompt-example-multimodal-row .prompt-example-row-button,
.prompt-example-multimodal-row .prompt-example-row-button button,
.prompt-example-media-html,
.prompt-example-media-html > div,
.prompt-example-media-html .wrap,
.prompt-example-media-html video,
.prompt-example-media-html img,
.example-preview-video,
.example-preview-image,
.reference-media-fallback {
min-height: 148px !important;
height: 148px !important;
max-height: 148px !important;
}
@media (max-width: 1200px) {
.prompt-options { margin-top: 5px !important; }
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
font-size: 10px !important;
height: 30px !important;
min-height: 30px !important;
padding: 0 7px !important;
}
}
.prompt-example-full-table,
.prompt-example-full-table > .form,
.prompt-examples,
.prompt-examples > .form {
max-height: none !important;
height: auto !important;
overflow: visible !important;
}
.prompt-example-full-table {
padding: 16px !important;
}
.prompt-example-full-table > .form {
gap: 14px !important;
}
.prompt-examples .prompt-example-row-button,
.prompt-examples .prompt-example-row-button button {
min-height: 96px !important;
height: auto !important;
max-height: none !important;
padding: 18px 22px !important;
overflow: visible !important;
white-space: normal !important;
display: block !important;
text-align: left !important;
}
.prompt-examples .prompt-example-row-button span,
.prompt-examples .prompt-example-row-button p,
.prompt-examples .prompt-example-row-button div {
max-height: none !important;
height: auto !important;
overflow: visible !important;
white-space: normal !important;
overflow-wrap: anywhere !important;
word-break: normal !important;
line-height: 1.5 !important;
text-overflow: unset !important;
-webkit-line-clamp: unset !important;
line-clamp: unset !important;
}
.prompt-example-multimodal-row,
.prompt-example-multimodal-row > .form {
max-height: none !important;
overflow: visible !important;
gap: 12px !important;
}
.prompt-example-multimodal-row > .form {
padding: 12px !important;
}
.prompt-example-multimodal-row .prompt-example-row-button,
.prompt-example-multimodal-row .prompt-example-row-button button,
.prompt-example-media-html,
.prompt-example-media-html > div,
.prompt-example-media-html .wrap,
.prompt-example-media-html video,
.prompt-example-media-html img,
.example-preview-video,
.example-preview-image,
.reference-media-fallback {
min-height: 148px !important;
height: 148px !important;
max-height: 148px !important;
}
.lance-output-panel .output-media-control {
min-height: 220px !important;
border: 1px solid rgba(116,126,140,.34) !important;
border-radius: 18px !important;
background: linear-gradient(180deg, rgba(250,251,253,.94), rgba(244,246,249,.9)) !important;
box-shadow: 0 10px 28px rgba(15,23,42,.10), inset 0 0 0 1px rgba(255,255,255,.75) !important;
overflow: hidden !important;
}
.lance-output-panel .output-media-control > div,
.lance-output-panel .output-media-control .wrap {
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
}
.lance-output-panel .output-media-control video,
.lance-output-panel .output-media-control img {
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
border-radius: 18px !important;
width: 100% !important;
height: 100% !important;
object-fit: contain !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input {
min-width: 138px !important;
max-width: 158px !important;
width: auto !important;
font-size: 10.5px !important;
padding-left: 12px !important;
padding-right: 12px !important;
}
@media (max-width: 1200px) {
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input {
min-width: 126px !important;
max-width: 146px !important;
font-size: 10px !important;
padding-left: 10px !important;
padding-right: 10px !important;
}
}
.lance-output-panel .output-text-control {
min-height: 220px !important;
border: 1px solid rgba(116,126,140,.34) !important;
border-radius: 18px !important;
background: linear-gradient(180deg, rgba(250,251,253,.94), rgba(244,246,249,.9)) !important;
box-shadow: 0 10px 28px rgba(15,23,42,.10), inset 0 0 0 1px rgba(255,255,255,.75) !important;
overflow: hidden !important;
padding: 0 !important;
}
.lance-output-panel .output-text-control > div,
.lance-output-panel .output-text-control .wrap,
.lance-output-panel .output-text-control .container {
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
padding: 0 !important;
}
.lance-output-panel .output-text-control textarea {
min-height: 220px !important;
border: 0 !important;
border-radius: 18px !important;
background: transparent !important;
box-shadow: none !important;
color: #101828 !important;
padding: 18px !important;
resize: none !important;
}
.prompt-options > .form {
display: inline-flex !important;
flex-wrap: nowrap !important;
justify-content: flex-start !important;
align-items: center !important;
gap: 8px !important;
width: auto !important;
max-width: 100% !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
height: 36px !important;
min-height: 36px !important;
font-size: 12px !important;
font-weight: 800 !important;
padding-left: 12px !important;
padding-right: 12px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input {
min-width: 166px !important;
max-width: 184px !important;
}
.video-resolution-row button,
.video-resolution-row [role="button"],
.video-resolution-row select,
.video-resolution-row input {
min-width: 74px !important;
max-width: 84px !important;
}
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input {
min-width: 72px !important;
max-width: 82px !important;
}
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input {
min-width: 62px !important;
max-width: 72px !important;
}
.output-resolution-row button,
.output-resolution-row [role="button"],
.output-resolution-row select,
.output-resolution-row input {
min-width: 92px !important;
max-width: 114px !important;
}
@media (max-width: 1200px) {
.prompt-options > .form {
gap: 6px !important;
}
.prompt-chip button,
.prompt-chip [role="button"],
.prompt-chip select,
.prompt-chip input {
height: 34px !important;
min-height: 34px !important;
font-size: 11px !important;
padding-left: 9px !important;
padding-right: 9px !important;
}
.frame-interpolation-row button,
.frame-interpolation-row [role="button"],
.frame-interpolation-row select,
.frame-interpolation-row input {
min-width: 148px !important;
max-width: 166px !important;
}
.video-resolution-row button,
.video-resolution-row [role="button"],
.video-resolution-row select,
.video-resolution-row input {
min-width: 66px !important;
max-width: 76px !important;
}
.aspect-ratio-row button,
.aspect-ratio-row [role="button"],
.aspect-ratio-row select,
.aspect-ratio-row input {
min-width: 64px !important;
max-width: 74px !important;
}
.video-duration-row button,
.video-duration-row [role="button"],
.video-duration-row select,
.video-duration-row input {
min-width: 56px !important;
max-width: 66px !important;
}
}
.lance-run-button {
margin-bottom: 6px !important;
}
.lance-quota-note,
.lance-quota-note > div,
.lance-quota-note .wrap,
.lance-quota-note .prose {
min-height: 0 !important;
padding-top: 0 !important;
padding-bottom: 0 !important;
}
.lance-quota-note {
max-width: 1040px !important;
margin: 0 auto 8px !important;
text-align: center !important;
color: var(--lance-text-muted) !important;
font-size: 12px !important;
line-height: 1.1 !important;
}
.lance-quota-note p {
margin: 0 !important;
padding: 0 !important;
line-height: 1.1 !important;
}
.frame-interpolation-row,
.frame-interpolation-disabled {
display: none !important;
visibility: hidden !important;
width: 0 !important;
max-width: 0 !important;
height: 0 !important;
max-height: 0 !important;
min-height: 0 !important;
margin: 0 !important;
padding: 0 !important;
overflow: hidden !important;
}
"""
APP_JS = None
TASK_T2V = "t2v"
TASK_T2I = "t2i"
TASK_V2T = "v2t"
TASK_X2T = "x2t"
TASK_X2T_VIDEO = "x2t_video"
TASK_X2T_IMAGE = "x2t_image"
TASK_IMAGE_EDIT = "image_edit"
TASK_VIDEO_EDIT = "video_edit"
TASK_LABEL_VIDEO_GENERATION = "Video Generation"
TASK_LABEL_VIDEO_EDIT = "Video Edit"
TASK_LABEL_VIDEO_UNDERSTANDING = "Video Understanding"
TASK_LABEL_IMAGE_GENERATION = "Image Generation"
TASK_LABEL_IMAGE_EDIT = "Image Edit"
TASK_LABEL_IMAGE_UNDERSTANDING = "Image Understanding"
TASK_CHOICES = [
TASK_LABEL_VIDEO_GENERATION,
TASK_LABEL_VIDEO_EDIT,
TASK_LABEL_VIDEO_UNDERSTANDING,
TASK_LABEL_IMAGE_GENERATION,
TASK_LABEL_IMAGE_EDIT,
TASK_LABEL_IMAGE_UNDERSTANDING,
]
TASK_LABEL_TO_INTERNAL = {
TASK_LABEL_VIDEO_GENERATION: TASK_T2V,
TASK_LABEL_VIDEO_EDIT: TASK_VIDEO_EDIT,
TASK_LABEL_VIDEO_UNDERSTANDING: TASK_X2T_VIDEO,
TASK_LABEL_IMAGE_GENERATION: TASK_T2I,
TASK_LABEL_IMAGE_EDIT: TASK_IMAGE_EDIT,
TASK_LABEL_IMAGE_UNDERSTANDING: TASK_X2T_IMAGE,
TASK_T2V: TASK_T2V,
TASK_VIDEO_EDIT: TASK_VIDEO_EDIT,
TASK_V2T: TASK_X2T_VIDEO,
TASK_X2T: TASK_X2T_VIDEO,
TASK_X2T_VIDEO: TASK_X2T_VIDEO,
TASK_T2I: TASK_T2I,
TASK_IMAGE_EDIT: TASK_IMAGE_EDIT,
TASK_X2T_IMAGE: TASK_X2T_IMAGE,
}
GENERATION_TASKS = {TASK_T2V, TASK_T2I, TASK_IMAGE_EDIT, TASK_VIDEO_EDIT}
UNDERSTANDING_TASKS = {TASK_X2T_VIDEO, TASK_X2T_IMAGE}
IMAGE_TASKS = {TASK_T2I, TASK_IMAGE_EDIT, TASK_X2T_IMAGE}
VIDEO_TASKS = {TASK_T2V, TASK_VIDEO_EDIT, TASK_X2T_VIDEO}
EDIT_TASKS = {TASK_IMAGE_EDIT, TASK_VIDEO_EDIT}
VIDEO_RESOLUTION_CHOICES = [DEFAULT_RESOLUTION]
VIDEO_EDIT_RESOLUTION_CHOICES = [DEFAULT_VIDEO_EDIT_RESOLUTION]
IMAGE_RESOLUTION_CHOICES = [DEFAULT_IMAGE_RESOLUTION]
RESOLUTION_CHOICES = VIDEO_RESOLUTION_CHOICES + IMAGE_RESOLUTION_CHOICES
VIDEO_RESOLUTION_DISPLAY_CHOICES = [("360p", "video_360p"), ("480p", "video_480p")]
V2T_QA_SYSTEM_PROMPT = "View the video attentively and provide a suitable answer to the posed question."
I2T_QA_SYSTEM_PROMPT = "View the image attentively and provide a suitable answer to the posed question."
def get_aspect_ratio_choices_for_task(task: str) -> list[tuple[str, str]]:
"""Get Aspect Ratio choices with default/recommended marker for the given task."""
internal_task = normalize_task(task)
default_ratio = DEFAULT_IMAGE_ASPECT_RATIO if internal_task in IMAGE_TASKS else DEFAULT_VIDEO_ASPECT_RATIO
return [
(f"{ratio}" if ratio == default_ratio else ratio, ratio)
for ratio in ASPECT_RATIO_CHOICES
]
def get_video_duration_choices() -> list[tuple[str, int]]:
return [(f"{seconds}s", seconds) for seconds in range(1, 11)]
def env_flag(name: str, default: bool) -> bool:
value = os.getenv(name)
if value is None:
return default
return value.strip().lower() in {"1", "true", "yes", "on"}
def running_on_space() -> bool:
return bool(os.getenv("SPACE_ID") or os.getenv("SPACE_HOST"))
def display_path(path: Path) -> str:
path_text = path.as_posix()
if path.is_absolute():
try:
path_text = path.relative_to(Path.cwd()).as_posix()
except ValueError:
return path_text
if path_text == "." or path_text.startswith("./"):
return path_text
return f"./{path_text}"
def get_model_base_dir() -> Path:
configured = os.getenv("LANCE_MODEL_BASE_DIR")
if configured:
configured_path = Path(configured).expanduser()
if _path_can_be_created_or_written(configured_path):
return configured_path
if LOCAL_MODEL_BASE_DIR.exists():
return LOCAL_MODEL_BASE_DIR
if running_on_space() and SPACE_MODEL_BASE_DIR.exists() and os.access(SPACE_MODEL_BASE_DIR, os.W_OK):
return SPACE_MODEL_BASE_DIR
return LOCAL_MODEL_BASE_DIR
def _path_can_be_created_or_written(path: Path) -> bool:
if path.exists():
return path.is_dir() and os.access(path, os.W_OK)
probe = path.parent
while not probe.exists() and probe != probe.parent:
probe = probe.parent
return probe.exists() and os.access(probe, os.W_OK)
def normalize_model_variant(model_variant: Optional[str] = None) -> str:
variant = (model_variant or os.getenv("LANCE_MODEL_VARIANT", DEFAULT_MODEL_VARIANT)).strip().lower()
if variant in {"image", "t2i", "i2t"}:
return MODEL_VARIANT_IMAGE
return MODEL_VARIANT_VIDEO
def get_model_path(model_variant: Optional[str] = None) -> Path:
variant = normalize_model_variant(model_variant)
variant_env_name = "LANCE_IMAGE_MODEL_PATH" if variant == MODEL_VARIANT_IMAGE else "LANCE_VIDEO_MODEL_PATH"
variant_configured = os.getenv(variant_env_name)
if variant_configured:
return Path(variant_configured).expanduser()
configured = os.getenv("LANCE_MODEL_PATH")
if configured:
return Path(configured).expanduser()
model_dir_name = MODEL_VARIANT_TO_DIR[variant]
return get_model_base_dir() / model_dir_name
def get_required_model_asset_paths(model_base_dir: Path, model_path: Path) -> list[Path]:
return [
model_path / "llm_config.json",
model_path / "model.safetensors",
model_base_dir / "Qwen2.5-VL-ViT" / "vit.safetensors",
model_base_dir / "Wan2.2_VAE.pth",
]
def get_model_download_allow_patterns(model_variant: Optional[str] = None) -> list[str]:
variant = normalize_model_variant(model_variant)
model_dir_name = MODEL_VARIANT_TO_DIR[variant]
return [
f"{model_dir_name}/**",
"Qwen2.5-VL-ViT/**",
"Wan2.2_VAE.pth",
"generation_config.json",
"llm_config.json",
"tokenizer.json",
"tokenizer_config.json",
"vocab.json",
"merges.txt",
"config.json",
]
def _get_safetensors_first_tensor_dtype(path: Path) -> Optional[torch.dtype]:
if not path.exists():
return None
with safe_open(str(path), framework="pt", device="cpu") as f:
keys = list(f.keys())
if not keys:
return None
return f.get_tensor(keys[0]).dtype
def convert_model_weights_to_bf16_inplace(model_path: Path) -> bool:
weight_path = model_path / "model.safetensors"
if not weight_path.exists():
return False
first_dtype = _get_safetensors_first_tensor_dtype(weight_path)
if first_dtype is None or first_dtype == torch.bfloat16:
return False
if first_dtype != torch.float32:
print(
f"[startup] Skipping bf16 conversion for {weight_path} because the first tensor dtype is {first_dtype}.",
flush=True,
)
return False
temp_path = weight_path.with_suffix(".bf16.safetensors.tmp")
print(f"[startup] Converting {weight_path} to bf16 to reduce disk usage.", flush=True)
with safe_open(str(weight_path), framework="pt", device="cpu") as f:
metadata = f.metadata()
tensor_names = list(f.keys())
tensors = {}
for name in tensor_names:
tensor = f.get_tensor(name)
tensors[name] = tensor.to(torch.bfloat16) if tensor.dtype == torch.float32 else tensor
save_file(tensors, str(temp_path), metadata=metadata)
os.replace(temp_path, weight_path)
print(f"[startup] Replaced original fp32 weights with bf16 weights at {weight_path}.", flush=True)
return True
def compact_downloaded_model_weights(model_base_dir: Path, variants: Optional[list[str]] = None) -> None:
model_dir_names = variants or [MODEL_VARIANT_TO_DIR[MODEL_VARIANT_IMAGE], MODEL_VARIANT_TO_DIR[MODEL_VARIANT_VIDEO]]
for model_dir_name in model_dir_names:
model_path = model_base_dir / model_dir_name
try:
convert_model_weights_to_bf16_inplace(model_path)
except Exception as exc:
print(f"[startup] bf16 compaction skipped for {display_path(model_path)}: {exc}", flush=True)
def ensure_model_assets(model_variant: Optional[str] = None) -> Path:
model_base_dir = get_model_base_dir()
os.environ["LANCE_MODEL_BASE_DIR"] = display_path(model_base_dir)
model_path = get_model_path(model_variant)
required_paths = get_required_model_asset_paths(model_base_dir, model_path)
if all(path.exists() for path in required_paths):
compact_downloaded_model_weights(model_base_dir, [MODEL_VARIANT_TO_DIR[normalize_model_variant(model_variant)]])
return model_path
downloads_model_base_dir = Path("downloads")
if model_base_dir == Path(".") and downloads_model_base_dir.exists():
downloads_model_path = downloads_model_base_dir / MODEL_VARIANT_TO_DIR[normalize_model_variant(model_variant)]
downloads_required_paths = get_required_model_asset_paths(downloads_model_base_dir, downloads_model_path)
if all(path.exists() for path in downloads_required_paths):
model_base_dir = downloads_model_base_dir
model_path = downloads_model_path
required_paths = downloads_required_paths
os.environ["LANCE_MODEL_BASE_DIR"] = display_path(model_base_dir)
compact_downloaded_model_weights(model_base_dir, [MODEL_VARIANT_TO_DIR[normalize_model_variant(model_variant)]])
return model_path
auto_download = env_flag("LANCE_AUTO_DOWNLOAD", running_on_space())
if not auto_download:
missing = "\n".join(f"- {display_path(path)}" for path in required_paths if not path.exists())
raise FileNotFoundError(
"Lance model assets are missing. Set LANCE_MODEL_BASE_DIR or enable "
f"LANCE_AUTO_DOWNLOAD=1.\nMissing files:\n{missing}"
)
model_base_dir.mkdir(parents=True, exist_ok=True)
repo_id = os.getenv("LANCE_MODEL_REPO_ID", DEFAULT_MODEL_REPO_ID)
print(f"[startup] Downloading Lance model assets from {repo_id} to {display_path(model_base_dir)}", flush=True)
hub_token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
snapshot_path = Path(
snapshot_download(
repo_id=repo_id,
local_dir=str(model_base_dir),
local_dir_use_symlinks=False,
resume_download=True,
token=hub_token,
allow_patterns=get_model_download_allow_patterns(model_variant),
)
)
if snapshot_path != model_base_dir and not model_path.exists():
os.environ["LANCE_MODEL_BASE_DIR"] = display_path(snapshot_path)
model_path = get_model_path(model_variant)
compact_downloaded_model_weights(model_base_dir, [MODEL_VARIANT_TO_DIR[normalize_model_variant(model_variant)]])
return model_path
def ensure_dirs() -> None:
TMP_INPUT_DIR.mkdir(parents=True, exist_ok=True)
RESULTS_ROOT.mkdir(parents=True, exist_ok=True)
def save_generation_record(record: dict, save_dir: Path) -> None:
ensure_dirs()
run_record_path = save_dir / RUN_RECORD_FILENAME
with run_record_path.open("w", encoding="utf-8") as f:
json.dump(record, f, ensure_ascii=False, indent=2)
with RECORD_WRITE_LOCK:
with GLOBAL_RECORDS_FILE.open("a", encoding="utf-8") as f:
f.write(json.dumps(record, ensure_ascii=False) + "\n")
def normalize_seed(seed: int) -> int:
return random.randint(0, 2**31 - 1) if seed == -1 else seed
def video_seconds_to_num_frames(seconds: int) -> int:
seconds = max(1, min(10, int(seconds)))
return 12 * seconds + 1
def normalize_task(task: str) -> str:
task_key = (task or TASK_LABEL_VIDEO_GENERATION).strip()
task = TASK_LABEL_TO_INTERNAL.get(task_key, TASK_LABEL_TO_INTERNAL.get(task_key.lower(), ""))
if task not in GENERATION_TASKS | UNDERSTANDING_TASKS:
raise ValueError(f"Unsupported task type: {task}")
return task
def normalize_resolution_choice_value(resolution: str, task: str) -> str:
resolution_text = str(resolution or "").strip()
for choice in get_resolution_choices_for_task(task):
if isinstance(choice, tuple):
label, value = choice
if resolution_text in {str(label), str(value)}:
return str(value)
elif resolution_text == str(choice):
return str(choice)
return resolution_text
def get_resolution_choice_values_for_task(task: str) -> list[str]:
return [choice[1] if isinstance(choice, tuple) else choice for choice in get_resolution_choices_for_task(task)]
def get_resolution_choices_for_task(task: str) -> list[str | tuple[str, str]]:
internal_task = normalize_task(task)
if internal_task in IMAGE_TASKS:
return IMAGE_RESOLUTION_CHOICES
if internal_task == TASK_T2V:
return VIDEO_RESOLUTION_DISPLAY_CHOICES
return VIDEO_EDIT_RESOLUTION_CHOICES if internal_task in VIDEO_TASKS else VIDEO_RESOLUTION_CHOICES
def get_default_resolution_for_task(task: str) -> str:
internal_task = normalize_task(task)
if internal_task in IMAGE_TASKS:
return DEFAULT_IMAGE_RESOLUTION
if internal_task == TASK_T2V:
return DEFAULT_RESOLUTION
return DEFAULT_VIDEO_EDIT_RESOLUTION if internal_task in VIDEO_TASKS else DEFAULT_RESOLUTION
def normalize_resolution_for_backend(resolution: str, task: str) -> str:
internal_task = normalize_task(task)
normalized_resolution = normalize_resolution_choice_value(resolution, internal_task)
return normalized_resolution if normalized_resolution in get_resolution_choice_values_for_task(internal_task) else get_default_resolution_for_task(internal_task)
def get_default_aspect_ratio(task: str) -> str:
internal_task = normalize_task(task)
return DEFAULT_IMAGE_ASPECT_RATIO if internal_task in IMAGE_TASKS else DEFAULT_VIDEO_ASPECT_RATIO
def normalize_video_resolution(resolution: Optional[str], task: Optional[str] = None) -> str:
if task is None:
return resolution if resolution in VIDEO_RESOLUTION_CHOICES else DEFAULT_RESOLUTION
normalized_resolution = normalize_resolution_choice_value(resolution, task)
choices = get_resolution_choice_values_for_task(task)
return normalized_resolution if normalized_resolution in choices else get_default_resolution_for_task(task)
def get_size_for_aspect_ratio(task: str, aspect_ratio: str, video_resolution: Optional[str] = None) -> tuple[int, int]:
internal_task = normalize_task(task)
aspect_ratio = aspect_ratio if aspect_ratio in ASPECT_RATIO_CHOICES else get_default_aspect_ratio(internal_task)
if internal_task in IMAGE_TASKS:
size_map = IMAGE_ASPECT_RATIO_TO_SIZE
else:
size_map = VIDEO_RESOLUTION_TO_SIZE_MAP[normalize_video_resolution(video_resolution, internal_task)]
return size_map[aspect_ratio]
def format_size_markdown(task: str, width: int, height: int) -> str:
return "" if normalize_task(task) in UNDERSTANDING_TASKS else f"{width} x {height}"
def get_size_map_for_task(task: str, video_resolution: Optional[str] = None) -> dict[str, tuple[int, int]]:
internal_task = normalize_task(task)
if internal_task in IMAGE_TASKS:
return IMAGE_ASPECT_RATIO_TO_SIZE
return VIDEO_RESOLUTION_TO_SIZE_MAP[normalize_video_resolution(video_resolution, internal_task)]
def get_output_resolution_choices_for_task(task: str, video_resolution: Optional[str] = None) -> list[tuple[str, str]]:
"""Get Output Resolution choices with a one-to-one mapping to aspect ratios."""
internal_task = normalize_task(task)
default_ratio = get_default_aspect_ratio(internal_task)
size_map = get_size_map_for_task(internal_task, video_resolution)
choices = []
for ratio in ASPECT_RATIO_CHOICES:
width, height = size_map[ratio]
resolution_text = format_size_markdown(internal_task, width, height)
label = f"{resolution_text}" if ratio == default_ratio else resolution_text
choices.append((label, resolution_text))
return choices
def build_lance_label_html(text: str, *extra_classes: str) -> str:
class_names = " ".join(["lance-section-label", *extra_classes]).strip()
return f'<div class="{class_names}">{html.escape(text)}</div>'
def build_lance_icon_label_html(text: str, icon: str, *extra_classes: str) -> str:
icon_map = {
"video": """
<span class="lance-label-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round">
<rect x="3.5" y="6" width="11" height="12" rx="2.2"></rect>
<path d="M15 10.2 20.5 7v10L15 13.8z" fill="currentColor" stroke="none"></path>
</svg>
</span>
""",
"image": """
<span class="lance-label-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round">
<rect x="3.5" y="5.5" width="17" height="13" rx="2.2"></rect>
<circle cx="9" cy="10" r="1.5" fill="currentColor" stroke="none"></circle>
<path d="M5.5 16.5 10 12l2.7 2.7 2.1-2.1 3.7 3.9"></path>
</svg>
</span>
""",
"text": """
<span class="lance-label-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round">
<rect x="3.5" y="5.5" width="17" height="13" rx="2.2"></rect>
<path d="M7 9h10"></path>
<path d="M7 12h7.5"></path>
<path d="M7 15h5.5"></path>
</svg>
</span>
""",
"logs": """
<span class="lance-label-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round">
<rect x="3.5" y="5.5" width="17" height="13" rx="2.2"></rect>
<path d="M7 10.2 10 12l-3 1.8"></path>
<path d="M12.5 15h4"></path>
</svg>
</span>
""",
}
icon_html = icon_map.get(icon, "")
class_names = " ".join(["lance-section-label", "lance-icon-label", *extra_classes]).strip()
return f'<div class="{class_names}">{icon_html}<span>{html.escape(text)}</span></div>'
def update_size_from_aspect_ratio(task: str, aspect_ratio: str, video_resolution: Optional[str] = None):
width, height = get_size_for_aspect_ratio(task, aspect_ratio, video_resolution)
return height, width, gr.update(
choices=get_output_resolution_choices_for_task(task, video_resolution),
value=format_size_markdown(task, width, height),
)
def update_output_resolution_from_video_profile(task: str, aspect_ratio: str, video_resolution: str):
width, height = get_size_for_aspect_ratio(task, aspect_ratio, video_resolution)
return (
gr.update(
choices=get_output_resolution_choices_for_task(task, video_resolution),
value=format_size_markdown(task, width, height),
),
height,
width,
)
def reset_generation_defaults_for_task(task: str):
internal_task = normalize_task(task)
aspect_ratio = get_default_aspect_ratio(internal_task)
resolution = get_default_resolution_for_task(internal_task)
width, height = get_size_for_aspect_ratio(internal_task, aspect_ratio, resolution)
num_frames = DEFAULT_VIDEO_DURATION_SECONDS
return aspect_ratio, height, width, num_frames, resolution, gr.update(
choices=get_output_resolution_choices_for_task(internal_task, resolution),
value=format_size_markdown(internal_task, width, height),
)
def make_prompt_example_click_handler(prompt_text: str, cache_key: str = ""):
"""Create a click handler for custom text-to-visual prompt-example rows.
gr.Dataset and gr.Examples render long text through compact preview cells, so
long prompts/instructions/questions can be truncated before CSS gets a chance
to wrap them. The custom rows below use normal buttons for display and keep
the full prompt string in this closure for click-to-fill behavior.
"""
def _handler(task: str):
defaults = reset_generation_defaults_for_task(task)
return (prompt_text, pack_recommended_cache_carrier(cache_key, task), *defaults)
return _handler
def make_media_prompt_example_click_handler(
prompt_text: str,
input_video_path: Optional[str] = None,
input_image_path: Optional[str] = None,
cache_key: str = "",
):
"""Create a click handler for edit/understanding example rows.
The row button renders the complete prompt/instruction/question, while the
closure also carries the matching media path so one click still fills every
required input component.
"""
def _handler(task: str):
defaults = reset_generation_defaults_for_task(task)
return (prompt_text, input_video_path, input_image_path, pack_recommended_cache_carrier(cache_key, task), *defaults)
return _handler
def get_understanding_system_prompt_choices(task: str) -> list[str]:
internal_task = normalize_task(task)
if internal_task == TASK_X2T_IMAGE:
return [I2T_QA_SYSTEM_PROMPT]
return [V2T_QA_SYSTEM_PROMPT]
def normalize_understanding_system_prompt(task: str, system_prompt: Optional[str]) -> str:
return get_understanding_system_prompt_choices(task)[0]
RECOMMENDED_CACHE_CARRIER_PREFIX = "__LANCE_RECOMMENDED_CASE_KEY__="
def pack_recommended_cache_carrier(cache_key: str, task: str) -> str:
"""Carry a recommended case key through the existing hidden system_prompt input.
This keeps Generate at the original Gradio inputs while carrying only the
example identity. Actual cache hits are validated later with a full request
signature so user-edited parameters never reuse the wrong output.
"""
internal_task = normalize_task(task)
base_prompt = normalize_understanding_system_prompt(internal_task, None) if internal_task in UNDERSTANDING_TASKS else ""
if not cache_key:
return base_prompt
return f"{RECOMMENDED_CACHE_CARRIER_PREFIX}{cache_key}\n{base_prompt}"
def unpack_recommended_cache_carrier(system_prompt: Optional[str]) -> tuple[str, Optional[str]]:
text = str(system_prompt or "")
if not text.startswith(RECOMMENDED_CACHE_CARRIER_PREFIX):
return "", system_prompt
payload = text[len(RECOMMENDED_CACHE_CARRIER_PREFIX):]
cache_key, _, base_prompt = payload.partition("\n")
return cache_key.strip(), (base_prompt if base_prompt else None)
def create_request_json(
task: str,
prompt: str,
input_video: Optional[str],
input_image: Optional[str],
system_prompt: Optional[str] = None,
) -> Path:
ensure_dirs()
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
prompt_file = TMP_INPUT_DIR / f"{task}_{timestamp}.json"
if task == TASK_T2V:
payload = {"000000.mp4": prompt}
elif task == TASK_T2I:
payload = {"000000.png": prompt}
elif task == TASK_VIDEO_EDIT:
if not input_video:
raise ValueError("The video edit task requires an input video.")
payload = {
"000000": {
"interleave_array": [prompt, input_video, input_video],
"element_dtype_array": ["text", "video", "video"],
"istarget_in_interleave": [0, 0, 1],
}
}
elif task == TASK_IMAGE_EDIT:
if not input_image:
raise ValueError("The image edit task requires an input image.")
payload = {
"000000": {
"interleave_array": [prompt, input_image, input_image],
"element_dtype_array": ["text", "image", "image"],
"istarget_in_interleave": [0, 0, 1],
}
}
elif task == TASK_X2T_VIDEO:
if not input_video:
raise ValueError("The video understanding task requires an input video.")
system_prompt = normalize_understanding_system_prompt(task, system_prompt)
payload = {
"000000": {
"interleave_array": [input_video, [system_prompt, prompt, ""]],
"element_dtype_array": ["video", "text"],
"istarget_in_interleave": [0, 1],
}
}
elif task == TASK_X2T_IMAGE:
if not input_image:
raise ValueError("The image understanding task requires an input image.")
system_prompt = normalize_understanding_system_prompt(task, system_prompt)
payload = {
"000000": {
"interleave_array": [input_image, [system_prompt, prompt, ""]],
"element_dtype_array": ["image", "text"],
"istarget_in_interleave": [0, 1],
}
}
else:
raise ValueError(f"Unsupported task type: {task}")
with prompt_file.open("w", encoding="utf-8") as f:
json.dump(payload, f, ensure_ascii=False, indent=2)
return prompt_file
def resolve_example_path(path: str) -> str:
candidate = Path(path)
if candidate.is_absolute():
return str(candidate)
repo_candidate = (REPO_ROOT / candidate)
if repo_candidate.exists():
return str(repo_candidate.resolve())
if candidate.exists():
return str(candidate.resolve())
return path
def resolve_browser_video_example_path(path: str) -> str:
candidate = Path(path)
compatible_candidate = candidate.with_name(f"{candidate.stem}_h264{candidate.suffix}")
repo_compatible_candidate = REPO_ROOT / compatible_candidate
if not compatible_candidate.is_absolute() and repo_compatible_candidate.exists():
return str(repo_compatible_candidate.resolve())
if compatible_candidate.is_absolute() and compatible_candidate.exists():
return str(compatible_candidate.resolve())
repo_candidate = REPO_ROOT / candidate
if not candidate.is_absolute() and repo_candidate.exists():
return str(repo_candidate.resolve())
if candidate.is_absolute() and candidate.exists():
return str(candidate.resolve())
return resolve_example_path(path)
def resolve_video_example_paths(path: str) -> tuple[str, str]:
"""Return (browser_preview_path, model_input_path) for a reference video."""
return resolve_browser_video_example_path(path), resolve_example_path(path)
def _resolve_existing_media_path(media_path: Optional[str]) -> Optional[Path]:
if not media_path:
return None
candidate = Path(str(media_path))
candidates = [candidate] if candidate.is_absolute() else [REPO_ROOT / candidate, candidate]
for item in candidates:
try:
resolved = item.expanduser().resolve()
except Exception:
continue
if resolved.exists():
return resolved
return None
def build_gradio_media_url(media_path: Optional[str]) -> str:
"""Build a Gradio file-serving URL for local recommended-case media."""
existing = _resolve_existing_media_path(media_path)
source = str(existing if existing else media_path or "")
if not source:
return ""
try:
from gradio.route_utils import API_PREFIX
except Exception:
API_PREFIX = ""
return f"{API_PREFIX or ''}/file={quote(source, safe='/:')}"
def build_example_media_html(media_path: Optional[str], media_type: str, fallback_media_path: Optional[str] = None) -> str:
"""Build a lightweight complete-fit media preview for recommended cases."""
if media_type == "video":
sources = []
for candidate in (media_path, fallback_media_path):
url = build_gradio_media_url(candidate)
if url and url not in sources:
sources.append(url)
if not sources:
return '<div class="reference-media-fallback">Video file not found</div>'
source_tags = "".join(
f'<source src="{html.escape(url, quote=True)}" type="video/mp4">'
for url in sources
)
return (
'<video class="example-preview-video" controls muted preload="metadata" playsinline>'
+ source_tags
+ 'Your browser cannot play this reference video.</video>'
)
url = build_gradio_media_url(media_path)
if not url:
return '<div class="reference-media-fallback">Image file not found</div>'
alt_text = html.escape(Path(str(media_path)).name or "example image", quote=True)
return f'<img class="example-preview-image" src="{html.escape(url, quote=True)}" alt="{alt_text}" loading="lazy" />'
# Recommended-case cache under the app.py directory. Runtime generated caches are
# written here by default, so each case can be committed with the repository.
LOCAL_RECOMMENDED_OUTPUT_CACHE_DIR = Path(
os.getenv("LANCE_LOCAL_RECOMMENDED_OUTPUT_CACHE_DIR", str(REPO_ROOT / "lance_gradio" / "recommended_outputs"))
).expanduser()
# Space/runtime cache root. This is kept as a read/query fallback so the app can
# still hit caches that were previously saved on the running Space instance.
SPACE_RECOMMENDED_OUTPUT_CACHE_DIR = Path(
os.getenv("LANCE_SPACE_RECOMMENDED_OUTPUT_CACHE_DIR", str(GRADIO_TMP_ROOT / "recommended_outputs"))
).expanduser()
# Writable cache target used by store_recommended_cached_result(). By default this
# is app.py's directory / lance_gradio / recommended_outputs. Set
# LANCE_RECOMMENDED_OUTPUT_CACHE_DIR to override it explicitly.
RECOMMENDED_OUTPUT_CACHE_DIR = Path(
os.getenv("LANCE_RECOMMENDED_OUTPUT_CACHE_DIR", str(LOCAL_RECOMMENDED_OUTPUT_CACHE_DIR))
).expanduser()
ASSET_RECOMMENDED_OUTPUT_CACHE_DIR = LOCAL_RECOMMENDED_OUTPUT_CACHE_DIR
RECOMMENDED_CASE_CACHE: dict[str, dict] = {}
def _sanitize_cache_token(value: object) -> str:
text = str(value or "").strip()
text = re.sub(r"[^A-Za-z0-9._-]+", "-", text)
return text.strip("-") or "default"
def _recommended_output_type(task: str) -> str:
internal_task = normalize_task(task)
if internal_task in {TASK_T2V, TASK_VIDEO_EDIT}:
return "video"
if internal_task in {TASK_T2I, TASK_IMAGE_EDIT}:
return "image"
return "text"
def _recommended_output_suffixes(output_type: str) -> tuple[str, ...]:
if output_type == "video":
return (".mp4", ".webm", ".mov")
if output_type == "image":
return (".png", ".jpg", ".jpeg", ".webp")
return (".txt", ".json")
def _default_recommended_output_name(task: str, example_id: str) -> str:
output_type = _recommended_output_type(task)
candidate = Path(str(example_id)).name or _sanitize_cache_token(example_id)
suffix = Path(candidate).suffix.lower()
if suffix in _recommended_output_suffixes(output_type):
return candidate
return f"{Path(candidate).stem or _sanitize_cache_token(example_id)}{_recommended_output_suffixes(output_type)[0]}"
def _cache_roots() -> list[Path]:
"""Query the new local cache first, then the Space/runtime saved cache."""
roots = [RECOMMENDED_OUTPUT_CACHE_DIR, SPACE_RECOMMENDED_OUTPUT_CACHE_DIR]
unique_roots: list[Path] = []
seen = set()
for root in roots:
try:
key = str(root.expanduser().resolve())
except Exception:
key = str(root)
if key not in seen:
seen.add(key)
unique_roots.append(root)
return unique_roots
def _infer_aspect_ratio_from_size(task: str, width: int, height: int, resolution: Optional[str]) -> str:
internal_task = normalize_task(task)
try:
size_map = get_size_map_for_task(internal_task, resolution)
requested = (int(width), int(height))
for ratio, size in size_map.items():
if tuple(size) == requested:
return ratio
except Exception:
pass
return get_default_aspect_ratio(internal_task)
def _canonical_float_for_cache(value: object) -> str:
try:
number = float(value)
except Exception:
return str(value or "")
# Keep numeric values stable across Gradio/Python representations while still
# being parameter-sensitive (for example, 3.5 and 3.500 resolve together).
return f"{number:.10g}"
def _cache_media_content_hash_enabled() -> bool:
# On Spaces, Gradio may copy example videos to a temporary file before the
# backend receives them. Path/mtime based identities then differ from local
# runs even when the media bytes are the same. A content hash makes example
# media identities stable across repo paths and Gradio temp paths.
return env_flag("LANCE_CACHE_MEDIA_CONTENT_HASH", True)
def _cache_media_hash_max_bytes() -> int:
try:
return int(os.getenv("LANCE_CACHE_MEDIA_HASH_MAX_BYTES", str(512 * 1024 * 1024)))
except Exception:
return 512 * 1024 * 1024
def _media_content_identity_for_cache(path: Path) -> str:
if not _cache_media_content_hash_enabled():
return ""
try:
stat = path.stat()
max_bytes = _cache_media_hash_max_bytes()
if max_bytes > 0 and stat.st_size > max_bytes:
return ""
digest = hashlib.sha256()
with path.open("rb") as f:
for chunk in iter(lambda: f.read(1024 * 1024), b""):
digest.update(chunk)
return f"sha256:{digest.hexdigest()}:{stat.st_size}"
except Exception:
return ""
def _canonical_media_identity_for_cache(media_path: Optional[str]) -> str:
"""Return a stable identity for media inputs used by recommended-case cache.
Example files may be passed either as repo-relative paths from JSON, resolved
absolute paths, or Space/Gradio temp-file paths. Content hashing is attempted
first so the same example video can match across local and Space even if
Gradio rewrites the path. If hashing is disabled or too expensive, this
falls back to repo-relative identity and then path/stat identity.
"""
if not media_path:
return ""
text = str(media_path)
candidate = Path(text).expanduser()
candidates = [candidate] if candidate.is_absolute() else [REPO_ROOT / candidate, candidate]
for item in candidates:
try:
resolved = item.resolve()
except Exception:
continue
if not resolved.exists():
continue
content_identity = _media_content_identity_for_cache(resolved)
if content_identity:
return content_identity
try:
rel = resolved.relative_to(REPO_ROOT.resolve()).as_posix()
return f"repo:{rel}"
except Exception:
pass
try:
stat = resolved.stat()
return f"file:{resolved.as_posix()}:{stat.st_size}:{int(stat.st_mtime_ns)}"
except Exception:
return f"file:{resolved.as_posix()}"
return f"path:{text}"
def _stable_json_for_cache(payload: dict) -> str:
return json.dumps(payload, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
def _recommended_request_signature_hash(request_signature: Optional[dict]) -> str:
if not request_signature:
return ""
return hashlib.sha256(_stable_json_for_cache(request_signature).encode("utf-8")).hexdigest()[:20]
def _recommended_request_cacheable(request_signature: Optional[dict]) -> bool:
if not request_signature:
return False
# A seed of -1 intentionally means random. The actual seed is sampled inside
# the generation path, so using a pre-existing cache would be misleading.
return int(request_signature.get("seed", 0)) != -1
def _recommended_signatures_equal(left: Optional[dict], right: Optional[dict]) -> bool:
if not left or not right:
return False
return _stable_json_for_cache(left) == _stable_json_for_cache(right)
def _recommended_cache_media_alias_enabled() -> bool:
# Gradio Spaces may copy or transcode example media before the backend sees
# it. For recommended cases, allow legacy cache files to match when all
# non-media parameters are identical and only the media identity differs.
return env_flag("LANCE_RECOMMENDED_CACHE_ALLOW_MEDIA_ALIAS", True)
def _recommended_signatures_equal_ignoring_media(left: Optional[dict], right: Optional[dict]) -> bool:
if not left or not right:
return False
left_copy = dict(left)
right_copy = dict(right)
for key in ("input_video", "input_image"):
left_copy.pop(key, None)
right_copy.pop(key, None)
return _stable_json_for_cache(left_copy) == _stable_json_for_cache(right_copy)
def build_recommended_request_signature(
task: str,
prompt: Optional[str],
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames_ui: int,
seed: int,
resolution: Optional[str],
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
) -> dict:
"""Build a complete cache signature for all user-controllable run params."""
internal_task = normalize_task(task)
normalized_resolution = normalize_resolution_for_backend(str(resolution), internal_task)
normalized_height = int(height)
normalized_width = int(width)
normalized_num_frames_ui = int(num_frames_ui)
aspect_ratio = _infer_aspect_ratio_from_size(
internal_task,
normalized_width,
normalized_height,
normalized_resolution,
)
normalized_system_prompt = (
normalize_understanding_system_prompt(internal_task, system_prompt)
if internal_task in UNDERSTANDING_TASKS
else str(system_prompt or "")
)
return {
"signature_version": 2,
"task": internal_task,
"prompt": str(prompt or "").strip(),
"system_prompt": normalized_system_prompt,
"input_video": _canonical_media_identity_for_cache(input_video),
"input_image": _canonical_media_identity_for_cache(input_image),
"resolution": normalized_resolution,
"aspect_ratio": aspect_ratio,
"height": normalized_height,
"width": normalized_width,
"num_frames_ui": normalized_num_frames_ui,
"num_frames_backend": video_seconds_to_num_frames(normalized_num_frames_ui)
if internal_task == TASK_T2V
else normalized_num_frames_ui,
"seed": int(seed),
"validation_num_timesteps": int(validation_num_timesteps),
"validation_timestep_shift": _canonical_float_for_cache(validation_timestep_shift),
"cfg_text_scale": _canonical_float_for_cache(cfg_text_scale),
"enable_frame_interpolation": bool(enable_frame_interpolation),
}
def _recommended_variant_tokens(
task: str,
resolution: Optional[str],
aspect_ratio: Optional[str],
duration_seconds: Optional[int] = None,
) -> list[str]:
internal_task = normalize_task(task)
normalized_resolution = normalize_resolution_for_backend(
str(resolution or get_default_resolution_for_task(internal_task)),
internal_task,
)
normalized_aspect = aspect_ratio if aspect_ratio in ASPECT_RATIO_CHOICES else get_default_aspect_ratio(internal_task)
tokens = [
_sanitize_cache_token(normalized_resolution),
_sanitize_cache_token(normalized_aspect),
]
# Only Video Generation exposes a user duration selector. Video Editing and
# Understanding use the input media duration, so the UI duration should not
# split their cache.
if internal_task == TASK_T2V:
seconds = int(duration_seconds if duration_seconds is not None else DEFAULT_VIDEO_DURATION_SECONDS)
tokens.append(f"{max(1, min(10, seconds))}s")
return tokens
def _recommended_output_name_for_variant(
task: str,
output_name: str,
resolution: Optional[str],
aspect_ratio: Optional[str],
duration_seconds: Optional[int] = None,
) -> str:
path_obj = Path(str(output_name))
stem = path_obj.stem or _sanitize_cache_token(output_name)
suffix = path_obj.suffix or _recommended_output_suffixes(_recommended_output_type(task))[0]
tokens = "__".join(_recommended_variant_tokens(task, resolution, aspect_ratio, duration_seconds))
return f"{stem}__{tokens}{suffix}" if tokens else f"{stem}{suffix}"
def _recommended_output_name_for_signature(
task: str,
output_name: str,
request_signature: dict,
) -> str:
path_obj = Path(str(output_name))
stem = path_obj.stem or _sanitize_cache_token(output_name)
suffix = path_obj.suffix or _recommended_output_suffixes(_recommended_output_type(task))[0]
signature_hash = _recommended_request_signature_hash(request_signature)
return f"{stem}__sig-{signature_hash}{suffix}"
def register_recommended_case_cache(
task: str,
example_id: str,
output_name: Optional[str] = None,
aspect_ratio: Optional[str] = None,
resolution: Optional[str] = None,
duration_seconds: Optional[int] = None,
prompt_text: Optional[str] = None,
input_video_path: Optional[str] = None,
input_image_path: Optional[str] = None,
) -> str:
internal_task = normalize_task(task)
normalized_resolution = normalize_resolution_for_backend(
str(resolution or get_default_resolution_for_task(internal_task)),
internal_task,
)
normalized_aspect = aspect_ratio if aspect_ratio in ASPECT_RATIO_CHOICES else get_default_aspect_ratio(internal_task)
default_width, default_height = get_size_for_aspect_ratio(internal_task, normalized_aspect, normalized_resolution)
default_duration = int(duration_seconds if duration_seconds is not None else DEFAULT_VIDEO_DURATION_SECONDS)
default_request_signature = build_recommended_request_signature(
task=internal_task,
prompt=prompt_text,
system_prompt=normalize_understanding_system_prompt(internal_task, None) if internal_task in UNDERSTANDING_TASKS else "",
input_video=input_video_path,
input_image=input_image_path,
height=default_height,
width=default_width,
num_frames_ui=default_duration,
seed=DEFAULT_BASIC_SEED,
resolution=normalized_resolution,
validation_num_timesteps=DEFAULT_TIMESTEPS,
validation_timestep_shift=DEFAULT_TIMESTEP_SHIFT,
cfg_text_scale=DEFAULT_CFG_TEXT_SCALE,
enable_frame_interpolation=False,
)
cache_key = f"{internal_task}:{_sanitize_cache_token(example_id)}"
RECOMMENDED_CASE_CACHE[cache_key] = {
"key": cache_key,
"task": internal_task,
"example_id": str(example_id),
"output_name": output_name or _default_recommended_output_name(internal_task, str(example_id)),
"output_type": _recommended_output_type(internal_task),
"resolution": normalized_resolution,
"aspect_ratio": normalized_aspect,
"duration_seconds": default_duration,
"prompt_text": str(prompt_text or ""),
"input_video_path": str(input_video_path or ""),
"input_image_path": str(input_image_path or ""),
"default_request_signature": default_request_signature,
"default_request_signature_hash": _recommended_request_signature_hash(default_request_signature),
}
return cache_key
def infer_recommended_case_key_from_request(
task: str,
prompt: str,
input_video: Optional[str] = None,
input_image: Optional[str] = None,
) -> str:
"""Best-effort fallback for sessions that do not carry the hidden cache key."""
internal_task = normalize_task(task)
prompt_text = str(prompt or "").strip()
input_video_id = _canonical_media_identity_for_cache(input_video)
input_image_id = _canonical_media_identity_for_cache(input_image)
for cache_key, meta in RECOMMENDED_CASE_CACHE.items():
if meta.get("task") != internal_task:
continue
if str(meta.get("prompt_text") or "").strip() != prompt_text:
continue
meta_video = str(meta.get("input_video_path") or "")
meta_image = str(meta.get("input_image_path") or "")
meta_video_id = _canonical_media_identity_for_cache(meta_video)
meta_image_id = _canonical_media_identity_for_cache(meta_image)
if meta_video_id and input_video_id and meta_video_id != input_video_id:
continue
if meta_image_id and input_image_id and meta_image_id != input_image_id:
continue
if meta_video_id and not input_video_id:
continue
if meta_image_id and not input_image_id:
continue
return cache_key
return ""
def _recommended_cache_candidates(
meta: dict,
resolution: Optional[str] = None,
aspect_ratio: Optional[str] = None,
duration_seconds: Optional[int] = None,
request_signature: Optional[dict] = None,
):
task = str(meta["task"])
output_name = str(meta.get("output_name") or _default_recommended_output_name(task, meta.get("example_id", meta["key"])))
output_type = str(meta.get("output_type") or _recommended_output_type(task))
requested_resolution = normalize_resolution_for_backend(str(resolution or meta.get("resolution") or ""), task)
requested_aspect = aspect_ratio if aspect_ratio in ASPECT_RATIO_CHOICES else str(meta.get("aspect_ratio") or get_default_aspect_ratio(task))
requested_duration = int(duration_seconds if duration_seconds is not None else meta.get("duration_seconds", DEFAULT_VIDEO_DURATION_SECONDS))
default_resolution = str(meta.get("resolution") or "")
default_aspect = str(meta.get("aspect_ratio") or get_default_aspect_ratio(task))
default_duration = int(meta.get("duration_seconds") or DEFAULT_VIDEO_DURATION_SECONDS)
default_signature = meta.get("default_request_signature")
is_default_signature = _recommended_signatures_equal(request_signature, default_signature)
is_media_alias_signature = (
_recommended_cache_media_alias_enabled()
and _recommended_signatures_equal_ignoring_media(request_signature, default_signature)
)
stem = Path(output_name).stem or _sanitize_cache_token(meta.get("example_id", meta.get("key", "case")))
names = set()
# New strict cache filenames: every user-controllable parameter is part of
# request_signature, so a changed seed/steps/CFG/media/size/etc. cannot hit
# an output generated under different settings.
if request_signature and _recommended_request_cacheable(request_signature):
signature_hash = _recommended_request_signature_hash(request_signature)
signature_name = _recommended_output_name_for_signature(task, output_name, request_signature)
names.add(signature_name)
for suffix in _recommended_output_suffixes(output_type):
names.add(f"{stem}__sig-{signature_hash}{suffix}")
names.add(f"{_sanitize_cache_token(meta['key'])}__sig-{signature_hash}{suffix}")
# Legacy recommended assets were named only by resolution/aspect/duration, or
# sometimes just by case id. They are safe for the exact default request
# signature registered for that recommended case. On Spaces, Gradio can
# rewrite recommended example videos to temp/transcoded files; in that case
# input_video changes while the user-visible recommended case is still the
# same. Allow legacy candidates when every non-media parameter still matches.
allow_legacy_candidates = request_signature is None or is_default_signature or is_media_alias_signature
if allow_legacy_candidates:
names.add(_recommended_output_name_for_variant(task, output_name, requested_resolution, requested_aspect, requested_duration))
tokens = "__".join(_recommended_variant_tokens(task, requested_resolution, requested_aspect, requested_duration))
for suffix in _recommended_output_suffixes(output_type):
names.add(f"{stem}__{tokens}{suffix}")
names.add(f"{_sanitize_cache_token(meta['key'])}__{tokens}{suffix}")
# Backward compatibility with the older width/height/duration filename format:
# stem__video_360p__640x352__3u.mp4
try:
width, height = get_size_for_aspect_ratio(task, requested_aspect, requested_resolution)
old_tokens = f"{_sanitize_cache_token(requested_resolution)}__{int(width)}x{int(height)}"
if normalize_task(task) == TASK_T2V:
old_tokens = f"{old_tokens}__{requested_duration}u"
for suffix in _recommended_output_suffixes(output_type):
names.add(f"{stem}__{old_tokens}{suffix}")
names.add(f"{_sanitize_cache_token(meta['key'])}__{old_tokens}{suffix}")
except Exception:
pass
# Legacy generic filename is only allowed for the case's default visible spec.
if (
requested_resolution == default_resolution
and requested_aspect == default_aspect
and (normalize_task(task) != TASK_T2V or requested_duration == default_duration)
):
names.add(output_name)
for suffix in _recommended_output_suffixes(output_type):
names.add(f"{stem}{suffix}")
names.add(f"{_sanitize_cache_token(meta['key'])}{suffix}")
for root in _cache_roots():
for folder in (root / str(task), root):
for name in names:
yield folder / name
def _recommended_cache_debug_enabled() -> bool:
return env_flag("LANCE_DEBUG_RECOMMENDED_CACHE", False)
def find_recommended_cached_output(
cache_key: str,
resolution: Optional[str] = None,
aspect_ratio: Optional[str] = None,
duration_seconds: Optional[int] = None,
request_signature: Optional[dict] = None,
) -> Optional[Path]:
meta = RECOMMENDED_CASE_CACHE.get(cache_key or "")
if not meta:
return None
debug = _recommended_cache_debug_enabled()
tried: list[str] = []
for candidate in _recommended_cache_candidates(
meta,
resolution=resolution,
aspect_ratio=aspect_ratio,
duration_seconds=duration_seconds,
request_signature=request_signature,
):
if debug and len(tried) < 24:
tried.append(str(candidate))
try:
if candidate.exists() and candidate.is_file():
return candidate.resolve()
except Exception:
continue
if debug:
default_signature = meta.get("default_request_signature")
print(
"[recommended-cache] Miss "
+ json.dumps(
{
"cache_key": cache_key,
"request_sig": _recommended_request_signature_hash(request_signature),
"default_sig": _recommended_request_signature_hash(default_signature),
"is_default_signature": _recommended_signatures_equal(request_signature, default_signature),
"is_media_alias_signature": _recommended_signatures_equal_ignoring_media(request_signature, default_signature),
"media_alias_enabled": _recommended_cache_media_alias_enabled(),
"roots": [str(root) for root in _cache_roots()],
"sample_candidates": tried,
"request_input_video": (request_signature or {}).get("input_video"),
"default_input_video": (default_signature or {}).get("input_video"),
"request_input_image": (request_signature or {}).get("input_image"),
"default_input_image": (default_signature or {}).get("input_image"),
"request_system_prompt": (request_signature or {}).get("system_prompt"),
"default_system_prompt": (default_signature or {}).get("system_prompt"),
},
ensure_ascii=False,
),
flush=True,
)
return None
def get_recommended_cached_result(
cache_key: str,
task: str,
resolution: Optional[str],
aspect_ratio: Optional[str],
duration_seconds: Optional[int] = None,
request_signature: Optional[dict] = None,
):
meta = RECOMMENDED_CASE_CACHE.get(cache_key or "")
if not meta:
return None
if not _recommended_request_cacheable(request_signature):
return None
cached_path = find_recommended_cached_output(
cache_key,
resolution=resolution,
aspect_ratio=aspect_ratio,
duration_seconds=duration_seconds,
request_signature=request_signature,
)
if cached_path is None:
return None
signature_hash = _recommended_request_signature_hash(request_signature)
print(f"[recommended-cache] Hit {cache_key} sig={signature_hash}: {cached_path}", flush=True)
# Keep cache hits silent in the UI. The output is returned directly without
# exposing cache paths or cache-matching details to end users. Matching is
# sensitive to the full request signature: prompt, media, size, seed, steps,
# shift, CFG scale, duration, resolution, and interpolation flag.
status = ""
output_type = str(meta.get("output_type") or _recommended_output_type(task))
if output_type == "video":
return str(cached_path), None, "", status
if output_type == "image":
return None, str(cached_path), "", status
try:
return None, None, cached_path.read_text(encoding="utf-8"), status
except Exception:
return None, None, str(cached_path), status
def store_recommended_cached_result(
cache_key: str,
result,
resolution: Optional[str],
aspect_ratio: Optional[str],
duration_seconds: Optional[int] = None,
request_signature: Optional[dict] = None,
) -> None:
meta = RECOMMENDED_CASE_CACHE.get(cache_key or "")
if not meta:
return
if not _recommended_request_cacheable(request_signature):
return
if find_recommended_cached_output(
cache_key,
resolution=resolution,
aspect_ratio=aspect_ratio,
duration_seconds=duration_seconds,
request_signature=request_signature,
) is not None:
return
try:
output_video, output_image, output_text, _status = result
target_name = _recommended_output_name_for_signature(
meta["task"],
str(meta["output_name"]),
request_signature,
)
target = RECOMMENDED_OUTPUT_CACHE_DIR / str(meta["task"]) / target_name
target.parent.mkdir(parents=True, exist_ok=True)
if meta["output_type"] == "video" and output_video and Path(str(output_video)).exists():
shutil.copy2(str(output_video), str(target))
elif meta["output_type"] == "image" and output_image and Path(str(output_image)).exists():
shutil.copy2(str(output_image), str(target))
elif meta["output_type"] == "text" and output_text:
target.write_text(str(output_text), encoding="utf-8")
else:
return
print(
f"[recommended-cache] Stored {cache_key} sig={_recommended_request_signature_hash(request_signature)} "
f"at {target} (resolution={resolution}, aspect_ratio={aspect_ratio}, duration={duration_seconds})",
flush=True,
)
except Exception as exc:
print(f"[recommended-cache] Could not store {cache_key}: {exc}", flush=True)
def load_json_examples(relative_path: str) -> dict:
path = REPO_ROOT / relative_path
with path.open("r", encoding="utf-8") as f:
return json.load(f)
T2V_EXAMPLE_SUMMARIES = {
"000000.mp4": "Red panda surfing on a bright seaside wave.",
"000002.mp4": "Panda cub skateboarding in a creative loft.",
"000004.mp4": "Young woman shaping clay in a sunlit pottery workshop.",
"000005.mp4": "Panda boxing a robot in a luxurious palace ring.",
"000008.mp4": "Fantasy pastel horse stepping through a glowing cloud valley.",
}
def make_generation_examples(
task_label: str,
relative_path: str,
limit: int,
image_task: bool,
selected_keys: Optional[list[str]] = None,
summaries: Optional[dict[str, str]] = None,
) -> list[list]:
internal_task = normalize_task(task_label)
data = load_json_examples(relative_path)
items = [(key, data[key]) for key in selected_keys if key in data] if selected_keys else list(data.items())[:limit]
examples = []
for output_name, prompt in items:
cache_key = register_recommended_case_cache(
task=internal_task,
example_id=output_name,
output_name=output_name,
aspect_ratio=get_default_aspect_ratio(internal_task),
resolution=get_default_resolution_for_task(internal_task),
duration_seconds=DEFAULT_VIDEO_DURATION_SECONDS,
prompt_text=prompt,
)
examples.append([prompt, cache_key])
return examples
def make_edit_examples(task_label: str, relative_path: str, limit: int, media_type: str) -> list[list]:
internal_task = normalize_task(task_label)
data = load_json_examples(relative_path)
examples = []
for idx, sample in enumerate(list(data.values())[:limit]):
interleave = sample["interleave_array"]
prompt = interleave[0]
example_id = f"{Path(relative_path).stem}_{idx:06d}"
cache_key = register_recommended_case_cache(
task=internal_task,
example_id=example_id,
output_name=_default_recommended_output_name(internal_task, example_id),
aspect_ratio=get_default_aspect_ratio(internal_task),
resolution=get_default_resolution_for_task(internal_task),
duration_seconds=DEFAULT_VIDEO_DURATION_SECONDS,
prompt_text=prompt,
input_video_path=interleave[1] if media_type == "video" else None,
input_image_path=interleave[1] if media_type == "image" else None,
)
if media_type == "video":
preview_video_path, input_video_path = resolve_video_example_paths(interleave[1])
examples.append([prompt, preview_video_path, input_video_path, None, None, cache_key])
else:
image_path = resolve_example_path(interleave[1])
examples.append([prompt, None, None, image_path, image_path, cache_key])
return examples
def make_understanding_examples(task_label: str, relative_path: str, limit: int, media_type: str) -> list[list]:
internal_task = normalize_task(task_label)
data = load_json_examples(relative_path)
examples = []
for idx, sample in enumerate(list(data.values())[:limit]):
interleave = sample["interleave_array"]
text_payload = interleave[1]
question = text_payload[1] if isinstance(text_payload, list) and len(text_payload) > 1 else ""
example_id = f"{Path(relative_path).stem}_{idx:06d}"
cache_key = register_recommended_case_cache(
task=internal_task,
example_id=example_id,
output_name=_default_recommended_output_name(internal_task, example_id),
aspect_ratio=get_default_aspect_ratio(internal_task),
resolution=get_default_resolution_for_task(internal_task),
duration_seconds=DEFAULT_VIDEO_DURATION_SECONDS,
prompt_text=question,
input_video_path=interleave[0] if media_type == "video" else None,
input_image_path=interleave[0] if media_type == "image" else None,
)
if media_type == "video":
preview_video_path, input_video_path = resolve_video_example_paths(interleave[0])
examples.append([question, preview_video_path, input_video_path, None, None, cache_key])
else:
image_path = resolve_example_path(interleave[0])
examples.append([question, None, None, image_path, image_path, cache_key])
return examples
def make_understanding_system_prompt_map(relative_path: str, task: str) -> dict[str, str]:
data = load_json_examples(relative_path)
system_prompts = {}
for sample in data.values():
interleave = sample["interleave_array"]
text_payload = interleave[1]
if not isinstance(text_payload, list) or len(text_payload) < 2:
continue
system_prompts[text_payload[1]] = normalize_understanding_system_prompt(task, text_payload[0])
return system_prompts
VIDEO_GENERATION_EXAMPLES = make_generation_examples(
TASK_LABEL_VIDEO_GENERATION,
"config/examples/t2v_example.json",
limit=7,
image_task=False,
#selected_keys=["000000.mp4", "000002.mp4", "000005.mp4", "000004.mp4", "000008.mp4"],
selected_keys=["000004.mp4", "000002.mp4", "000000.mp4", "000005.mp4", "000008.mp4", "000007.mp4", "000001.mp4"],
summaries=T2V_EXAMPLE_SUMMARIES,
)
VIDEO_EDIT_EXAMPLES = make_edit_examples(
TASK_LABEL_VIDEO_EDIT,
"config/examples/video_edit_example.json",
limit=3,
media_type="video",
)
VIDEO_UNDERSTANDING_EXAMPLES = make_understanding_examples(
TASK_LABEL_VIDEO_UNDERSTANDING,
"config/examples/x2t_video_example.json",
limit=3,
media_type="video",
)
VIDEO_UNDERSTANDING_SYSTEM_PROMPTS = make_understanding_system_prompt_map(
"config/examples/x2t_video_example.json",
TASK_X2T_VIDEO,
)
IMAGE_GENERATION_EXAMPLES = make_generation_examples(
TASK_LABEL_IMAGE_GENERATION,
"config/examples/t2i_example.json",
limit=9,
image_task=True,
selected_keys=["000000.png", "000003.png", "000002.png", "000005.png", "000006.png", "000007.png", "000008.png", "000009.png", "000010.png"],
)
IMAGE_EDIT_EXAMPLES = make_edit_examples(
TASK_LABEL_IMAGE_EDIT,
"config/examples/image_edit_example.json",
limit=5,
media_type="image",
)
IMAGE_UNDERSTANDING_EXAMPLES = make_understanding_examples(
TASK_LABEL_IMAGE_UNDERSTANDING,
"config/examples/x2t_image_example.json",
limit=3,
media_type="image",
)
IMAGE_UNDERSTANDING_SYSTEM_PROMPTS = make_understanding_system_prompt_map(
"config/examples/x2t_image_example.json",
TASK_X2T_IMAGE,
)
def build_save_dir(task: str) -> Path:
ensure_dirs()
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
return RESULTS_ROOT / f"{task}_{timestamp}_{int(time.time() * 1000) % 1000:03d}"
def find_generated_video(save_dir: Path) -> Optional[Path]:
videos = sorted(save_dir.glob("*.mp4"), key=lambda p: p.stat().st_mtime, reverse=True)
return videos[0] if videos else None
def find_generated_image(save_dir: Path) -> Optional[Path]:
images = sorted(save_dir.glob("*.png"), key=lambda p: p.stat().st_mtime, reverse=True)
return images[0] if images else None
def extract_text_result(save_dir: Path) -> str:
prompt_result_path = save_dir / PROMPT_JSON_FILENAME
if not prompt_result_path.exists():
return ""
with prompt_result_path.open("r", encoding="utf-8") as f:
data = json.load(f)
if not data:
return ""
first_value = next(iter(data.values()))
return first_value if isinstance(first_value, str) else json.dumps(first_value, ensure_ascii=False)
class LanceT2VV2TPipeline:
def __init__(self, device_id: int, model_variant: str = MODEL_VARIANT_VIDEO) -> None:
self._init_lock = threading.Lock()
self._generate_lock = threading.Lock()
self.initialized = False
self.device = device_id
self.model_variant = normalize_model_variant(model_variant)
self.logger = get_logger(f"lance_{self.model_variant}_gpu{device_id}")
self.model: Optional[Lance] = None
self.vae_model: Optional[WanVideoVAE] = None
self.vae_config: Optional[AutoEncoderParams] = None
self.tokenizer: Optional[Qwen2Tokenizer] = None
self.new_token_ids: Optional[dict] = None
self.image_token_id: Optional[int] = None
self.base_model_args: Optional[ModelArguments] = None
self.base_data_args: Optional[DataArguments] = None
self.base_inference_args: Optional[InferenceArguments] = None
def _log_stage(self, stage_name: str, start_time: float, extra: str = "") -> None:
elapsed = time.perf_counter() - start_time
suffix = f" | {extra}" if extra else ""
print(f"[startup][gpu:{self.device}] {stage_name} done in {elapsed:.2f}s{suffix}", flush=True)
def _build_base_model_args(self) -> ModelArguments:
model_path = str(get_model_path(self.model_variant))
return ModelArguments(
model_path=model_path,
vit_type=DEFAULT_VIT_TYPE,
llm_qk_norm=True,
llm_qk_norm_und=True,
llm_qk_norm_gen=True,
tie_word_embeddings=False,
max_num_frames=MAX_VIDEO_NUM_FRAMES,
max_latent_size=64,
latent_patch_size=[1, 1, 1],
)
def _build_base_inference_args(self) -> InferenceArguments:
return InferenceArguments(
validation_num_timesteps=DEFAULT_TIMESTEPS,
validation_timestep_shift=DEFAULT_TIMESTEP_SHIFT,
copy_init_moe=True,
visual_und=True,
visual_gen=True,
vae_model_type="wan",
apply_qwen_2_5_vl_pos_emb=True,
apply_chat_template=False,
cfg_type=0,
validation_data_seed=42,
video_height=DEFAULT_HEIGHT,
video_width=DEFAULT_WIDTH,
num_frames=DEFAULT_NUM_FRAMES,
task=DEFAULT_TASK,
save_path_gen=str(RESULTS_ROOT),
resolution=DEFAULT_RESOLUTION,
text_template=TEXT_TEMPLATE,
use_KVcache=USE_KVCACHE,
)
def initialize(self) -> None:
with self._init_lock:
if self.initialized:
return
ensure_dirs()
resolved_model_path = ensure_model_assets(self.model_variant)
print(
f"[startup][gpu:{self.device}][{self.model_variant}] Using Lance model path: {resolved_model_path}",
flush=True,
)
if not torch.cuda.is_available():
raise RuntimeError("CUDA is unavailable. Lance T2V/V2T Gradio requires a GPU environment.")
if self.device >= torch.cuda.device_count():
raise RuntimeError(
f"GPU {self.device} is unavailable. Detected {torch.cuda.device_count()} GPU(s)."
)
torch.cuda.set_device(self.device)
model_args = self._build_base_model_args()
data_args = DataArguments()
inference_args = self._build_base_inference_args()
apply_inference_defaults(model_args, data_args, inference_args)
inference_args.validation_noise_seed = inference_args.validation_data_seed
self.base_model_args = model_args
self.base_data_args = data_args
self.base_inference_args = inference_args
set_seed(inference_args.global_seed)
stage_start = time.perf_counter()
print(
f"[startup][gpu:{self.device}] Loading LLM config: {Path(model_args.model_path) / 'llm_config.json'}",
flush=True,
)
llm_config: Qwen2Config = Qwen2Config.from_json_file(str(Path(model_args.model_path) / "llm_config.json"))
self._log_stage("LLM config load", stage_start)
llm_config.layer_module = model_args.layer_module
llm_config.qk_norm = model_args.llm_qk_norm
llm_config.qk_norm_und = model_args.llm_qk_norm_und
llm_config.qk_norm_gen = model_args.llm_qk_norm_gen
llm_config.tie_word_embeddings = model_args.tie_word_embeddings
llm_config.freeze_und = inference_args.freeze_und
llm_config.apply_qwen_2_5_vl_pos_emb = inference_args.apply_qwen_2_5_vl_pos_emb
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Initializing LLM weights: {model_args.model_path}", flush=True)
language_model: Qwen2ForCausalLM = Qwen2ForCausalLM(llm_config)
self._log_stage("LLM weight init", stage_start)
vit_model = None
vit_config = None
if inference_args.visual_und:
if model_args.vit_type not in ("qwen2_5_vl", "qwen_2_5_vl_original"):
raise ValueError(f"Unsupported vit_type: {model_args.vit_type}")
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Loading VIT config: {model_args.vit_path}", flush=True)
vit_config = Qwen2_5_VLVisionConfig.from_pretrained(model_args.vit_path)
self._log_stage("VIT config load", stage_start)
stage_start = time.perf_counter()
print(
f"[startup][gpu:{self.device}] Loading VIT weights: {Path(model_args.vit_path) / 'vit.safetensors'}",
flush=True,
)
vit_model = Qwen2_5_VisionTransformerPretrainedModel(vit_config)
vit_weights = load_file(str(Path(model_args.vit_path) / "vit.safetensors"))
vit_model.load_state_dict(vit_weights, strict=True)
self._log_stage("VIT weight load", stage_start)
clean_memory(vit_weights)
if inference_args.visual_gen:
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Initializing VAE", flush=True)
vae_model = WanVideoVAE(device=torch.device("cuda", self.device))
vae_config = deepcopy(vae_model.vae_config)
self._log_stage("VAE init", stage_start)
else:
vae_model = None
vae_config = None
config = LanceConfig(
visual_gen=inference_args.visual_gen,
visual_und=inference_args.visual_und,
llm_config=llm_config,
vit_config=vit_config if inference_args.visual_und else None,
vae_config=vae_config if inference_args.visual_gen else None,
latent_patch_size=model_args.latent_patch_size,
max_num_frames=model_args.max_num_frames,
max_latent_size=model_args.max_latent_size,
vit_max_num_patch_per_side=model_args.vit_max_num_patch_per_side,
connector_act=model_args.connector_act,
interpolate_pos=model_args.interpolate_pos,
timestep_shift=inference_args.timestep_shift,
)
model: Lance = Lance(
language_model=language_model,
vit_model=vit_model if inference_args.visual_und else None,
vit_type=model_args.vit_type,
config=config,
training_args=inference_args,
)
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Casting Lance model to bf16 on CPU", flush=True)
model = model.to(dtype=torch.bfloat16)
self._log_stage("Lance model bf16 cast", stage_start)
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Loading tokenizer: {model_args.model_path}", flush=True)
tokenizer: Qwen2Tokenizer = Qwen2Tokenizer.from_pretrained(model_args.model_path)
tokenizer, new_token_ids, num_new_tokens = add_special_tokens(tokenizer)
self._log_stage("tokenizer load and special token init", stage_start, extra=f"num_new_tokens={num_new_tokens}")
if inference_args.copy_init_moe:
language_model.init_moe()
init_from_model_path_if_needed(model, model_args)
if num_new_tokens > 0:
model.language_model.resize_token_embeddings(len(tokenizer))
model.config.llm_config.vocab_size = len(tokenizer)
model.language_model.config.vocab_size = len(tokenizer)
if model_args.vit_type.lower() == "qwen2_5_vl":
from common.model.hacks import hack_qwen2_5_vl_config
language_model = hack_qwen2_5_vl_config(language_model)
image_token_id = language_model.config.video_token_id
new_token_ids.update({"image_token_id": image_token_id})
model.update_tokenizer(tokenizer=tokenizer)
if model_args.tie_word_embeddings:
model.language_model.untie_lm_head()
model.language_model.copy_new_token_rows_to_lm_head(num_new_tokens)
model_args.tie_word_embeddings = False
llm_config.tie_word_embeddings = False
else:
assert (
model.language_model.get_input_embeddings().weight.data.data_ptr()
!= model.language_model.get_output_embeddings().weight.data.data_ptr()
), "tie_word_embeddings conflict"
stage_start = time.perf_counter()
print(f"[startup][gpu:{self.device}] Moving Lance model to GPU {self.device}", flush=True)
model = model.to(device=self.device)
self._log_stage("Lance model move to GPU", stage_start)
model.eval()
if vae_model is not None and hasattr(vae_model, "eval"):
vae_model.eval()
self.model = model
self.vae_model = vae_model
self.vae_config = vae_config
self.tokenizer = tokenizer
self.new_token_ids = new_token_ids
self.image_token_id = image_token_id
self.initialized = True
print(
f"[startup][gpu:{self.device}][{self.model_variant}] Lance multimodal Gradio model loaded and ready for reuse.",
flush=True,
)
def unload(self) -> None:
with self._init_lock:
if self.model is not None:
self.model.cpu()
if self.vae_model is not None and hasattr(self.vae_model, "vae"):
vae_inner = self.vae_model.vae
if hasattr(vae_inner, "model"):
vae_inner.model.cpu()
self.model = None
self.vae_model = None
self.vae_config = None
self.tokenizer = None
self.new_token_ids = None
self.image_token_id = None
self.base_model_args = None
self.base_data_args = None
self.base_inference_args = None
self.initialized = False
gc.collect()
if torch.cuda.is_available():
with torch.cuda.device(self.device):
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
def _build_request_batch(
self,
prompt_file: Path,
model_args: ModelArguments,
data_args: DataArguments,
inference_args: InferenceArguments,
):
assert self.tokenizer is not None
assert self.new_token_ids is not None
assert self.vae_config is not None
dataset_config = DataConfig.from_yaml(str(prompt_file))
if inference_args.visual_und:
dataset_config.vit_patch_size = model_args.vit_patch_size
dataset_config.vit_patch_size_temporal = model_args.vit_patch_size_temporal
dataset_config.vit_max_num_patch_per_side = model_args.vit_max_num_patch_per_side
if inference_args.visual_gen:
vae_downsample = tuple_mul(
tuple(model_args.latent_patch_size),
(
self.vae_config.downsample_temporal,
self.vae_config.downsample_spatial,
self.vae_config.downsample_spatial,
),
)
dataset_config.latent_patch_size = model_args.latent_patch_size
dataset_config.vae_downsample = vae_downsample
dataset_config.max_latent_size = model_args.max_latent_size
dataset_config.max_num_frames = model_args.max_num_frames
dataset_config.text_cond_dropout_prob = model_args.text_cond_dropout_prob
dataset_config.vae_cond_dropout_prob = model_args.vae_cond_dropout_prob
dataset_config.vit_cond_dropout_prob = model_args.vit_cond_dropout_prob
dataset_config.num_frames = inference_args.num_frames
dataset_config.H = inference_args.video_height
dataset_config.W = inference_args.video_width
dataset_config.task = inference_args.task
dataset_config.resolution = inference_args.resolution
dataset_config.text_template = inference_args.text_template
val_dataset = ValidationDataset(
jsonl_path=str(prompt_file),
tokenizer=self.tokenizer,
data_args=data_args,
model_args=model_args,
training_args=inference_args,
new_token_ids=self.new_token_ids,
dataset_config=dataset_config,
local_rank=0,
world_size=1,
)
return simple_custom_collate([val_dataset[0]])
def generate(
self,
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
):
self.initialize()
internal_task = normalize_task(task)
prompt = (prompt or "").strip()
input_video = str(input_video).strip() if input_video else ""
input_image = str(input_image).strip() if input_image else ""
if internal_task in GENERATION_TASKS and not prompt:
return None, None, "", "Please enter a prompt."
if internal_task in UNDERSTANDING_TASKS and not prompt:
return None, None, "", "Please enter a question."
if internal_task in {TASK_VIDEO_EDIT, TASK_X2T_VIDEO} and not input_video:
return None, None, "", "Please upload an input video."
if internal_task in {TASK_IMAGE_EDIT, TASK_X2T_IMAGE} and not input_image:
return None, None, "", "Please upload an input image."
if height <= 0 or width <= 0:
return None, None, "", "Height and width must be greater than 0."
if num_frames <= 0:
return None, None, "", "The number of frames must be greater than 0."
assert self.model is not None
assert self.tokenizer is not None
assert self.new_token_ids is not None
assert self.image_token_id is not None
assert self.base_model_args is not None
assert self.base_data_args is not None
assert self.base_inference_args is not None
active_model_path = self.base_model_args.model_path
with self._generate_lock:
torch.cuda.set_device(self.device)
actual_seed = normalize_seed(int(seed))
prompt_file = create_request_json(
task=internal_task,
prompt=prompt,
input_video=input_video,
input_image=input_image,
system_prompt=system_prompt,
)
save_dir = build_save_dir(internal_task)
save_dir.mkdir(parents=True, exist_ok=True)
request_started_at = datetime.now().isoformat(timespec="seconds")
request_model_args = deepcopy(self.base_model_args)
request_model_args.cfg_text_scale = float(cfg_text_scale)
request_data_args = deepcopy(self.base_data_args)
request_data_args.val_dataset_config_file = str(prompt_file)
request_inference_args = deepcopy(self.base_inference_args)
request_inference_args.validation_num_timesteps = int(validation_num_timesteps)
request_inference_args.validation_timestep_shift = float(validation_timestep_shift)
request_inference_args.validation_data_seed = actual_seed
request_inference_args.validation_noise_seed = actual_seed
request_inference_args.video_height = int(height)
request_inference_args.video_width = int(width)
request_inference_args.num_frames = int(num_frames)
display_resolution = str(resolution)
backend_resolution = normalize_resolution_for_backend(display_resolution, internal_task)
request_inference_args.resolution = backend_resolution
request_inference_args.save_path_gen = str(save_dir)
request_inference_args.task = internal_task
request_inference_args.text_template = TEXT_TEMPLATE
request_inference_args.prompt_data_dict = {}
try:
print(
"[lance_gradio_t2v_v2t] Start generation "
f"| task={internal_task} | gpu={self.device} | seed={actual_seed} | "
f"size={height}x{width} | frames={num_frames} | resolution={display_resolution}",
flush=True,
)
val_data_cpu = self._build_request_batch(
prompt_file=prompt_file,
model_args=request_model_args,
data_args=request_data_args,
inference_args=request_inference_args,
)
# Keep the allocator from fragmenting before the heavy forward pass.
clean_memory()
generate_start = time.perf_counter()
validate_on_fixed_batch(
fsdp_model=self.model,
vae_model=self.vae_model,
tokenizer=self.tokenizer,
val_data_cpu=val_data_cpu,
training_args=request_inference_args,
model_args=request_model_args,
inference_args=request_inference_args,
new_token_ids=self.new_token_ids,
image_token_id=self.image_token_id,
device=self.device,
save_source_video=False,
save_path_gen=request_inference_args.save_path_gen,
save_path_gt="",
)
elapsed = time.perf_counter() - generate_start
save_prompt_results(request_inference_args.prompt_data_dict, request_inference_args.save_path_gen, self.logger)
clean_memory()
video_path = find_generated_video(save_dir) if internal_task in {TASK_T2V, TASK_VIDEO_EDIT} else None
original_video_path = video_path
frame_interpolation_enabled = False
image_path = find_generated_image(save_dir) if internal_task in {TASK_T2I, TASK_IMAGE_EDIT} else None
text_result = extract_text_result(save_dir) if internal_task in UNDERSTANDING_TASKS else ""
record = {
"request_started_at": request_started_at,
"request_finished_at": datetime.now().isoformat(timespec="seconds"),
"status": "success",
"task": internal_task,
"model_variant": self.model_variant,
"model_path": active_model_path,
"gpu": self.device,
"prompt": prompt,
"system_prompt": normalize_understanding_system_prompt(internal_task, system_prompt)
if internal_task in UNDERSTANDING_TASKS
else "",
"input_video": input_video,
"input_image": input_image,
"seed": actual_seed,
"height": int(height),
"width": int(width),
"num_frames": int(num_frames),
"resolution": display_resolution,
"backend_resolution": backend_resolution,
"validation_num_timesteps": int(validation_num_timesteps),
"validation_timestep_shift": float(validation_timestep_shift),
"cfg_text_scale": float(cfg_text_scale),
"frame_interpolation": frame_interpolation_enabled,
"elapsed_seconds": round(elapsed, 3),
"prompt_file": str(prompt_file),
"output_dir": str(save_dir),
"original_video_path": str(original_video_path) if original_video_path is not None else "",
"video_path": str(video_path) if video_path is not None else "",
"image_path": str(image_path) if image_path is not None else "",
"text_result": text_result,
"rife_error": "",
}
if internal_task in {TASK_T2V, TASK_VIDEO_EDIT} and video_path is None:
record["status"] = "completed_without_video"
if internal_task in {TASK_T2I, TASK_IMAGE_EDIT} and image_path is None:
record["status"] = "completed_without_image"
if internal_task in UNDERSTANDING_TASKS and not text_result:
record["status"] = "completed_without_text"
save_generation_record(record, save_dir)
if internal_task in {TASK_T2V, TASK_VIDEO_EDIT}:
if video_path is None:
status = (
"Inference completed, but no output video was found.\n\n"
f"- Task: `{internal_task}`\n"
f"- Model: `{self.model_variant}`\n"
f"- Model path: `{active_model_path}`\n"
f"- GPU: `{self.device}`\n"
f"- Actual seed: `{actual_seed}`\n"
f"- Output directory: `{save_dir}`"
)
return None, None, "", status
return str(video_path), None, "", ""
if internal_task in {TASK_T2I, TASK_IMAGE_EDIT}:
if image_path is None:
status = (
"Inference completed, but no output image was found.\n\n"
f"- Task: `{internal_task}`\n"
f"- Model: `{self.model_variant}`\n"
f"- Model path: `{active_model_path}`\n"
f"- GPU: `{self.device}`\n"
f"- Actual seed: `{actual_seed}`\n"
f"- Output directory: `{save_dir}`"
)
return None, None, "", status
return None, str(image_path), "", ""
return None, None, text_result, ""
except Exception:
error_trace = traceback.format_exc()
print(error_trace, flush=True)
record = {
"request_started_at": request_started_at,
"request_finished_at": datetime.now().isoformat(timespec="seconds"),
"status": "failed",
"task": internal_task,
"model_variant": self.model_variant,
"model_path": active_model_path,
"gpu": self.device,
"prompt": prompt,
"input_video": input_video,
"input_image": input_image,
"seed": actual_seed,
"height": int(height),
"width": int(width),
"num_frames": int(num_frames),
"resolution": display_resolution,
"backend_resolution": backend_resolution,
"validation_num_timesteps": int(validation_num_timesteps),
"validation_timestep_shift": float(validation_timestep_shift),
"cfg_text_scale": float(cfg_text_scale),
"prompt_file": str(prompt_file),
"output_dir": str(save_dir),
"video_path": "",
"image_path": "",
"text_result": "",
"error": error_trace,
}
save_generation_record(record, save_dir)
status = (
"Inference failed.\n\n"
f"- Task: `{internal_task}`\n"
f"- Model: `{self.model_variant}`\n"
f"- Model path: `{active_model_path}`\n"
f"- GPU: `{self.device}`\n"
f"- Actual seed: `{actual_seed}`\n"
f"- Resolution: `{display_resolution}`\n"
f"- Output directory: `{save_dir}`"
)
return None, None, "", status
class PipelinePool:
def __init__(self, gpu_ids: list[int], model_variant: str = MODEL_VARIANT_VIDEO) -> None:
if not gpu_ids:
raise ValueError("At least one GPU must be configured.")
self.gpu_ids = gpu_ids
self.model_variant = normalize_model_variant(model_variant)
self.pipelines = [
LanceT2VV2TPipeline(device_id=gpu_id, model_variant=self.model_variant)
for gpu_id in gpu_ids
]
self._available = deque(self.pipelines)
self._condition = threading.Condition()
@property
def size(self) -> int:
return len(self.pipelines)
@property
def gpu_summary(self) -> str:
return ",".join(str(gpu_id) for gpu_id in self.gpu_ids)
@property
def is_initialized(self) -> bool:
return all(pipeline.initialized for pipeline in self.pipelines)
def initialize_all(self) -> None:
if self.is_initialized:
return
print(f"[startup][{self.model_variant}] Preparing parallel GPU preload: {self.gpu_ids}", flush=True)
exceptions: list[Exception] = []
with concurrent.futures.ThreadPoolExecutor(max_workers=self.size) as executor:
futures = {
executor.submit(pipeline.initialize): pipeline.device for pipeline in self.pipelines
}
for future in concurrent.futures.as_completed(futures):
gpu_id = futures[future]
try:
future.result()
except Exception as exc:
print(f"[startup][gpu:{gpu_id}][{self.model_variant}] Preload failed: {exc}", flush=True)
exceptions.append(exc)
if exceptions:
raise RuntimeError(
f"{self.model_variant} preload failed on {len(exceptions)} GPU(s). Please check the terminal logs."
) from exceptions[0]
print(
f"[startup][{self.model_variant}] GPU preload finished. Ready to handle {self.size} concurrent request(s).",
flush=True,
)
def acquire(self) -> LanceT2VV2TPipeline:
with self._condition:
while not self._available:
self._condition.wait()
return self._available.popleft()
def release(self, pipeline: LanceT2VV2TPipeline) -> None:
with self._condition:
self._available.append(pipeline)
self._condition.notify()
def unload_all(self) -> None:
print(f"[runtime][{self.model_variant}] Unloading model pool from GPU(s): {self.gpu_ids}", flush=True)
with self._condition:
while len(self._available) != len(self.pipelines):
self._condition.wait()
for pipeline in self.pipelines:
pipeline.unload()
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
print(f"[runtime][{self.model_variant}] Model pool unloaded.", flush=True)
def generate(
self,
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
):
pipeline = self.acquire()
try:
return pipeline.generate(
task=task,
prompt=prompt,
system_prompt=system_prompt,
input_video=input_video,
input_image=input_image,
height=height,
width=width,
num_frames=num_frames,
seed=seed,
resolution=resolution,
validation_num_timesteps=validation_num_timesteps,
validation_timestep_shift=validation_timestep_shift,
cfg_text_scale=cfg_text_scale,
enable_frame_interpolation=enable_frame_interpolation,
)
finally:
self.release(pipeline)
ACTIVE_PIPELINE_POOL: Optional[PipelinePool] = None
ACTIVE_POOL_LOCK = threading.Lock()
QUEUE_MAX_SIZE = DEFAULT_QUEUE_SIZE
QUEUE_CONCURRENCY_LIMIT = DEFAULT_CONCURRENCY_LIMIT
def get_task_model_variant(task: str) -> str:
internal_task = normalize_task(task)
return MODEL_VARIANT_IMAGE if internal_task in IMAGE_TASKS else MODEL_VARIANT_VIDEO
def get_env_int(name: str, default: int) -> int:
"""Read an integer environment variable, falling back safely on invalid values."""
try:
return int(os.getenv(name, str(default)))
except (TypeError, ValueError):
return default
def ensure_flash_attn_installed() -> None:
try:
from importlib.metadata import PackageNotFoundError, version as package_version
current_version = package_version("flash_attn")
if current_version == DEFAULT_FLASH_ATTN_VERSION:
print(f"[startup] flash-attn {current_version} already installed.", flush=True)
return
print(
f"[startup] flash-attn {current_version} detected; reinstalling {DEFAULT_FLASH_ATTN_VERSION} from wheel.",
flush=True,
)
except Exception:
print(
f"[startup] flash-attn not available; installing {DEFAULT_FLASH_ATTN_VERSION} from wheel.",
flush=True,
)
command = [
sys.executable,
"-m",
"pip",
"install",
"--no-cache-dir",
"--no-deps",
"--force-reinstall",
DEFAULT_FLASH_ATTN_WHEEL_URL,
]
subprocess.check_call(command)
print(f"[startup] flash-attn {DEFAULT_FLASH_ATTN_VERSION} installed from wheel.", flush=True)
def get_zerogpu_duration_cap() -> int:
"""Fixed duration requested from ZeroGPU for each run.
The duration value is a ZeroGPU reservation/timeout hint. Shorter values can
improve queue priority and reduce wasted quota, but the value must still cover
model warm-up plus inference. Override per deployment when needed:
LANCE_ZEROGPU_MAX_DURATION_SECONDS=300
"""
return max(1, get_env_int("LANCE_ZEROGPU_MAX_DURATION_SECONDS", 300))
def clamp_zerogpu_duration(seconds: int) -> int:
return max(1, min(int(seconds), get_zerogpu_duration_cap()))
ZERO_GPU_RUN_TASK_DURATION_SECONDS = get_zerogpu_duration_cap()
def is_pipeline_pool_ready_for_variant(model_variant: str) -> bool:
normalized_variant = normalize_model_variant(model_variant)
with ACTIVE_POOL_LOCK:
return bool(
ACTIVE_PIPELINE_POOL is not None
and ACTIVE_PIPELINE_POOL.model_variant == normalized_variant
and ACTIVE_PIPELINE_POOL.is_initialized
)
def is_pipeline_pool_ready_for_task(task: str) -> bool:
return is_pipeline_pool_ready_for_variant(get_task_model_variant(task))
def get_pipeline_pool(task: str) -> PipelinePool:
global ACTIVE_PIPELINE_POOL
if not torch.cuda.is_available():
raise RuntimeError(
"Lance inference requires a GPU. The Gradio UI can start on CPU, but generation is disabled "
"until GPU hardware is attached."
)
model_variant = get_task_model_variant(task)
gpu_ids = parse_gpu_ids(os.getenv("LANCE_GPUS", DEFAULT_GPUS))
with ACTIVE_POOL_LOCK:
if ACTIVE_PIPELINE_POOL is not None and ACTIVE_PIPELINE_POOL.model_variant == model_variant:
if not ACTIVE_PIPELINE_POOL.is_initialized:
ACTIVE_PIPELINE_POOL.initialize_all()
return ACTIVE_PIPELINE_POOL
if ACTIVE_PIPELINE_POOL is not None:
previous_variant = ACTIVE_PIPELINE_POOL.model_variant
print(
f"[runtime] Switching Lance model from {previous_variant} to {model_variant}.",
flush=True,
)
ACTIVE_PIPELINE_POOL.unload_all()
ACTIVE_PIPELINE_POOL = None
ACTIVE_PIPELINE_POOL = PipelinePool(gpu_ids, model_variant=model_variant)
ACTIVE_PIPELINE_POOL.initialize_all()
return ACTIVE_PIPELINE_POOL
def finalize_zerogpu_duration(estimated_seconds: float, task: str) -> int:
"""Clamp a heuristic duration to the deployment cap with a small safety margin."""
task_key = normalize_task(task)
raw_seconds = float(estimated_seconds)
if raw_seconds <= 0:
raw_seconds = _estimate_zerogpu_duration_seconds(
task_key,
prompt="",
system_prompt=None,
input_video=None,
input_image=None,
height=0,
width=0,
num_frames=0,
seed=0,
resolution="",
validation_num_timesteps=0,
validation_timestep_shift=0.0,
cfg_text_scale=0.0,
enable_frame_interpolation=False,
)
return clamp_zerogpu_duration(math.ceil(raw_seconds * 1.15) + 5)
def _estimate_zerogpu_duration_seconds(
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
) -> int:
internal_task = normalize_task(task)
prompt_length = len((prompt or "").strip())
has_video_input = bool((input_video or "").strip())
has_image_input = bool((input_image or "").strip())
pool_ready = is_pipeline_pool_ready_for_task(internal_task)
is_video_task = internal_task in {TASK_T2V, TASK_VIDEO_EDIT, TASK_X2T_VIDEO}
is_image_task = internal_task in {TASK_T2I, TASK_IMAGE_EDIT, TASK_X2T_IMAGE}
if internal_task == TASK_T2I:
return 90 if pool_ready else 150
if internal_task == TASK_IMAGE_EDIT:
return 100 if pool_ready else 150
if internal_task == TASK_X2T_IMAGE:
return 90 if pool_ready else 150
if internal_task == TASK_X2T_VIDEO:
return 120 if pool_ready else 200
if internal_task == TASK_VIDEO_EDIT:
base = 170 if pool_ready else 300
base += min(30 if pool_ready else 48, max(0, num_frames - 37) // 3)
base += 24 if enable_frame_interpolation else 0
base += 16 if has_video_input else 0
base += 10 if resolution == "video_480p" else 0
return base
if internal_task == TASK_T2V:
if pool_ready:
base = 130 if resolution == "video_360p" else 150
base += min(36, max(0, num_frames - 37) // 3)
base += 18 if enable_frame_interpolation else 0
base += min(12, prompt_length // 320)
return base
base = 224 if resolution == "video_360p" else 264
base += min(56, max(0, num_frames - 37) // 2)
base += 28 if enable_frame_interpolation else 0
base += min(20, prompt_length // 260)
return base
if is_video_task:
base = 150 if pool_ready else 240
base += min(28 if pool_ready else 40, max(0, num_frames - 37) // 3)
base += 18 if enable_frame_interpolation else 0
return base
if is_image_task:
return 100 if pool_ready else 120
return 160
def get_run_task_gpu_duration(
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
) -> int:
enable_frame_interpolation = False
estimated_seconds = _estimate_zerogpu_duration_seconds(
task=task,
prompt=prompt,
system_prompt=system_prompt,
input_video=input_video,
input_image=input_image,
height=height,
width=width,
num_frames=num_frames,
seed=seed,
resolution=resolution,
validation_num_timesteps=validation_num_timesteps,
validation_timestep_shift=validation_timestep_shift,
cfg_text_scale=cfg_text_scale,
enable_frame_interpolation=enable_frame_interpolation,
)
return finalize_zerogpu_duration(estimated_seconds, task)
def run_task(
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
):
internal_task = normalize_task(task)
recommended_case_key, clean_system_prompt = unpack_recommended_cache_carrier(system_prompt)
system_prompt = clean_system_prompt
if not recommended_case_key:
recommended_case_key = infer_recommended_case_key_from_request(internal_task, prompt, input_video, input_image)
if internal_task in UNDERSTANDING_TASKS and not prompt:
return None, None, "", "Please enter a question."
if internal_task in {TASK_VIDEO_EDIT, TASK_X2T_VIDEO} and not input_video:
return None, None, "", "Please upload an input video."
if internal_task in {TASK_IMAGE_EDIT, TASK_X2T_IMAGE} and not input_image:
return None, None, "", "Please upload an input image."
if height <= 0 or width <= 0:
return None, None, "", "Height and width must be greater than 0."
if num_frames <= 0:
return None, None, "", "The number of frames must be greater than 0."
num_frames_ui = int(num_frames)
normalized_resolution = normalize_resolution_for_backend(str(resolution), internal_task)
aspect_ratio = _infer_aspect_ratio_from_size(internal_task, int(width), int(height), normalized_resolution)
# Ignore any stale interpolation value from old browser sessions before
# building the cache signature, because interpolation is disabled in this UI.
enable_frame_interpolation = False
request_signature = build_recommended_request_signature(
task=internal_task,
prompt=prompt,
system_prompt=system_prompt,
input_video=input_video,
input_image=input_image,
height=int(height),
width=int(width),
num_frames_ui=num_frames_ui,
seed=int(seed),
resolution=normalized_resolution,
validation_num_timesteps=int(validation_num_timesteps),
validation_timestep_shift=float(validation_timestep_shift),
cfg_text_scale=float(cfg_text_scale),
enable_frame_interpolation=enable_frame_interpolation,
)
cached_result = get_recommended_cached_result(
recommended_case_key,
internal_task,
resolution=normalized_resolution,
aspect_ratio=aspect_ratio,
duration_seconds=num_frames_ui,
request_signature=request_signature,
)
if cached_result is not None:
return cached_result
if internal_task == TASK_T2V:
num_frames = video_seconds_to_num_frames(num_frames_ui)
result = run_task_gpu(
task=task,
prompt=prompt,
system_prompt=system_prompt,
input_video=input_video,
input_image=input_image,
height=height,
width=width,
num_frames=num_frames,
seed=seed,
resolution=normalized_resolution,
validation_num_timesteps=validation_num_timesteps,
validation_timestep_shift=validation_timestep_shift,
cfg_text_scale=cfg_text_scale,
enable_frame_interpolation=enable_frame_interpolation,
)
store_recommended_cached_result(
recommended_case_key,
result,
resolution=normalized_resolution,
aspect_ratio=aspect_ratio,
duration_seconds=num_frames_ui,
request_signature=request_signature,
)
return result
@spaces.GPU(size="large", duration=get_run_task_gpu_duration)
def run_task_gpu(
task: str,
prompt: str,
system_prompt: Optional[str],
input_video: Optional[str],
input_image: Optional[str],
height: int,
width: int,
num_frames: int,
seed: int,
resolution: str,
validation_num_timesteps: int,
validation_timestep_shift: float,
cfg_text_scale: float,
enable_frame_interpolation: bool,
):
pipeline_pool = get_pipeline_pool(task)
return pipeline_pool.generate(
task=task,
prompt=prompt,
system_prompt=system_prompt,
input_video=input_video,
input_image=input_image,
height=height,
width=width,
num_frames=num_frames,
seed=seed,
resolution=resolution,
validation_num_timesteps=validation_num_timesteps,
validation_timestep_shift=validation_timestep_shift,
cfg_text_scale=cfg_text_scale,
enable_frame_interpolation=enable_frame_interpolation,
)
def build_status_markdown() -> str:
gpu_text = "unknown"
pipeline_slots = 0
active_variant = "none"
with ACTIVE_POOL_LOCK:
if ACTIVE_PIPELINE_POOL is not None:
active_variant = ACTIVE_PIPELINE_POOL.model_variant
gpu_text = ACTIVE_PIPELINE_POOL.gpu_summary
pipeline_slots = ACTIVE_PIPELINE_POOL.size
return (
f"**Status** GPU: `{gpu_text}` | Queue concurrency: `{QUEUE_CONCURRENCY_LIMIT}` | "
f"Pipeline slots: `{pipeline_slots}` | Queue limit: `{QUEUE_MAX_SIZE}` | "
f"Active model: `{active_variant}`"
)
def build_running_status_markdown() -> str:
return "Running..."
def get_logo_data_uri() -> str:
if not LANCE_LOGO_PATH.exists():
return ""
encoded_logo = base64.b64encode(LANCE_LOGO_PATH.read_bytes()).decode("ascii")
return f"data:image/webp;base64,{encoded_logo}"
def build_header_html() -> str:
logo_data_uri = get_logo_data_uri()
logo_html = (
f'<img class="lance-logo" src="{logo_data_uri}" alt="Lance logo">'
if logo_data_uri
else ""
)
return f"""
<div class="lance-hero">
{logo_html}
<h1 class="lance-title">Lance: Unified Multimodal Modeling by Multi-Task Synergy</h1>
<div class="lance-badges">
<a href="{LANCE_HOMEPAGE_URL}" target="_blank" rel="noopener noreferrer">
<img alt="Homepage" src="https://img.shields.io/badge/Homepage-Lance-2563eb?style=flat&labelColor=475569">
</a>
<a href="{LANCE_PAPER_URL}" target="_blank" rel="noopener noreferrer">
<img alt="Paper" src="https://img.shields.io/badge/Paper-arXiv-2563eb?style=flat&labelColor=475569&logo=arxiv">
</a>
<a href="{LANCE_HUGGING_FACE_URL}" target="_blank" rel="noopener noreferrer">
<img alt="Hugging Face" src="https://img.shields.io/badge/Model-HuggingFace-2563eb?style=flat&labelColor=475569&logo=huggingface">
</a>
<a href="{LANCE_GITHUB_URL}" target="_blank" rel="noopener noreferrer">
<img alt="GitHub" src="https://img.shields.io/badge/Code-GitHub-2563eb?style=flat&labelColor=475569&logo=github">
</a>
</div>
</div>
"""
def update_task_ui(task: str):
internal_task = normalize_task(task)
is_image_task = internal_task in IMAGE_TASKS
is_video_task = internal_task in VIDEO_TASKS
is_edit_task = internal_task in EDIT_TASKS
is_understanding_task = internal_task in UNDERSTANDING_TASKS
is_generation_task = internal_task in GENERATION_TASKS
is_text_to_visual_task = internal_task in {TASK_T2V, TASK_T2I}
show_media_input = is_edit_task or is_understanding_task
resolution_choices = get_resolution_choices_for_task(internal_task)
resolution_value = get_default_resolution_for_task(internal_task)
aspect_ratio_value = DEFAULT_IMAGE_ASPECT_RATIO if is_image_task else DEFAULT_VIDEO_ASPECT_RATIO
width_value, height_value = get_size_for_aspect_ratio(internal_task, aspect_ratio_value, resolution_value)
size_markdown = format_size_markdown(internal_task, width_value, height_value)
system_prompt_choices = get_understanding_system_prompt_choices(internal_task)
if is_text_to_visual_task:
text_label = "Prompt"
text_placeholder = "Describe what you want to generate..."
elif is_edit_task:
text_label = "Instruction"
text_placeholder = "Describe the edit you want..."
else:
text_label = "Question"
text_placeholder = "Ask a question about the input..."
if internal_task in {TASK_T2V, TASK_VIDEO_EDIT}:
output_label = "Output Video"
elif internal_task in {TASK_T2I, TASK_IMAGE_EDIT}:
output_label = "Output Image"
else:
output_label = "Output Text"
output_icon = "video" if output_label == "Output Video" else "image" if output_label == "Output Image" else "text"
show_generation_settings = is_generation_task or is_edit_task
show_aspect_ratio = is_text_to_visual_task
show_input_video = internal_task in {TASK_VIDEO_EDIT, TASK_X2T_VIDEO}
show_input_image = internal_task in {TASK_IMAGE_EDIT, TASK_X2T_IMAGE}
show_frame_interpolation_settings = False
show_video_resolution_settings = internal_task == TASK_T2V
return (
gr.update(value=build_lance_label_html(text_label, "lance-prompt-label")),
gr.update(
label=text_label,
placeholder=text_placeholder,
visible=True,
value="",
),
gr.update(
choices=system_prompt_choices,
value=system_prompt_choices[0],
visible=False,
),
# Switching task pages should always start from a clean input state.
# Clear both visual input boxes even if one of them stays visible across tasks.
gr.update(label="Input Video", visible=show_input_video, value=None),
gr.update(label="Input Image", visible=show_input_image, value=None),
gr.update(visible=False),
gr.update(visible=show_aspect_ratio),
gr.update(visible=False),
gr.update(visible=internal_task == TASK_T2V),
gr.update(visible=show_video_resolution_settings),
gr.update(choices=get_aspect_ratio_choices_for_task(internal_task), value=aspect_ratio_value, visible=show_aspect_ratio),
gr.update(value=height_value),
gr.update(value=width_value),
gr.update(visible=False, value=False),
gr.update(choices=get_output_resolution_choices_for_task(internal_task, resolution_value), value=size_markdown, visible=False),
gr.update(visible=internal_task == TASK_T2V, value=DEFAULT_VIDEO_DURATION_SECONDS),
gr.update(choices=resolution_choices, value=resolution_value, visible=show_video_resolution_settings),
gr.update(value=build_lance_icon_label_html(output_label, output_icon, "lance-output-label")),
gr.update(visible=internal_task in {TASK_T2V, TASK_VIDEO_EDIT}),
gr.update(visible=internal_task in {TASK_T2I, TASK_IMAGE_EDIT}),
gr.update(visible=is_understanding_task, value=""),
gr.update(visible=internal_task == TASK_T2V),
gr.update(visible=internal_task == TASK_VIDEO_EDIT),
gr.update(visible=internal_task == TASK_X2T_VIDEO),
gr.update(visible=internal_task == TASK_T2I),
gr.update(visible=internal_task == TASK_IMAGE_EDIT),
gr.update(visible=internal_task == TASK_X2T_IMAGE),
"",
)
def build_demo() -> gr.Blocks:
with gr.Blocks(title="Lance", css=APP_CSS, js=APP_JS) as demo:
gr.HTML(build_header_html())
gr.Markdown(build_status_markdown(), elem_classes=["lance-status"], visible=False)
with gr.Column(elem_classes=["lance-taskbar-wrap"]):
task = gr.Radio(
label="Task",
show_label=False,
choices=TASK_CHOICES,
value=TASK_LABEL_VIDEO_GENERATION,
elem_classes=["task-selector"],
)
with gr.Row(elem_classes=["lance-main-row"]):
with gr.Column(scale=1, elem_classes=["lance-main-column", "lance-input-column"]):
with gr.Column(elem_classes=["lance-panel", "lance-task-prompt-panel"]):
prompt_label = gr.HTML(build_lance_label_html("Prompt", "lance-prompt-label"), elem_classes=["lance-label-html"])
prompt = gr.Textbox(
label="Prompt",
show_label=False,
lines=6,
placeholder="Describe the video you want to generate...",
elem_classes=["main-prompt-control"],
)
with gr.Row(elem_classes=["prompt-options"]):
with gr.Group(elem_classes=["prompt-chip", "video-resolution-row"]) as video_resolution_row:
resolution = gr.Dropdown(
label="Video Resolution",
show_label=False,
choices=VIDEO_RESOLUTION_DISPLAY_CHOICES,
value=DEFAULT_RESOLUTION,
allow_custom_value=True,
elem_classes=["generation-control"],
)
with gr.Group(elem_classes=["prompt-chip", "aspect-ratio-row"]) as aspect_ratio_row:
aspect_ratio = gr.Dropdown(
label="Aspect Ratio",
show_label=False,
choices=get_aspect_ratio_choices_for_task(TASK_T2V),
value=DEFAULT_VIDEO_ASPECT_RATIO,
elem_classes=["generation-control"],
)
with gr.Group(elem_classes=["prompt-chip", "video-duration-row"]) as video_duration_row:
num_frames = gr.Dropdown(
label="Video Duration",
show_label=False,
choices=get_video_duration_choices(),
value=DEFAULT_VIDEO_DURATION_SECONDS,
elem_classes=["generation-control"],
)
with gr.Group(visible=False, elem_classes=["prompt-chip", "output-resolution-row"]) as output_resolution_row:
real_size = gr.Dropdown(
label="Output Resolution",
show_label=False,
choices=get_output_resolution_choices_for_task(TASK_T2V),
value=format_size_markdown(TASK_T2V, DEFAULT_WIDTH, DEFAULT_HEIGHT),
interactive=False,
visible=False,
allow_custom_value=True,
elem_classes=["generation-control"],
)
# Hidden compatibility components for old callbacks; frame interpolation is disabled.
with gr.Group(visible=False, elem_classes=["frame-interpolation-row", "frame-interpolation-disabled"]) as frame_interpolation_row:
enable_frame_interpolation = gr.Checkbox(value=False, visible=False)
system_prompt = gr.Dropdown(
label="System Prompt",
choices=get_understanding_system_prompt_choices(TASK_X2T_VIDEO),
value=V2T_QA_SYSTEM_PROMPT,
visible=False,
allow_custom_value=True,
)
input_video = gr.Video(label="Input Video", visible=False, elem_classes=["lance-display-frame"])
input_image = gr.Image(label="Input Image", type="filepath", visible=False, elem_classes=["lance-display-frame"])
height = gr.Number(value=DEFAULT_HEIGHT, precision=0, visible=False)
width = gr.Number(value=DEFAULT_WIDTH, precision=0, visible=False)
with gr.Accordion("Advanced Parameters", open=False, elem_classes=["lance-advanced-accordion"]):
seed = gr.Number(label="Seed (-1 for random seed)", value=DEFAULT_BASIC_SEED, precision=0)
validation_num_timesteps = gr.Slider(
minimum=1,
maximum=50,
step=1,
value=DEFAULT_TIMESTEPS,
label="Validation Num Timesteps",
)
with gr.Row():
validation_timestep_shift = gr.Number(label="Validation Timestep Shift", value=DEFAULT_TIMESTEP_SHIFT)
cfg_text_scale = gr.Number(label="CFG Text Scale", value=DEFAULT_CFG_TEXT_SCALE)
with gr.Column(scale=1, elem_classes=["lance-main-column", "lance-output-column"]):
with gr.Column(elem_classes=["lance-panel", "lance-output-panel"]):
output_label = gr.HTML(
build_lance_icon_label_html("Output Video", "video", "lance-output-label"),
elem_classes=["lance-label-html"],
)
output_video = gr.Video(label="Output Video", show_label=False, elem_classes=["lance-display-frame", "output-media-control"])
output_image = gr.Image(label="Output Image", show_label=False, type="filepath", visible=False, elem_classes=["lance-display-frame", "output-media-control"])
output_text = gr.Textbox(label="Output Text", show_label=False, lines=3, visible=False, elem_classes=["lance-display-frame", "output-text-control"])
status = gr.Markdown("", elem_classes=["lance-run-status"])
recommended_case_key = gr.State("")
run_button = gr.Button("🚀 Generate", variant="primary", elem_classes=["lance-run-button"])
gr.Markdown(
"**Note**: Video-related features may consume more GPU quota and take longer. Cached recommended cases and image tasks are lighter.",
elem_classes=["lance-quota-note"],
)
def build_prompt_example_table(examples: list[list], media_type: Optional[str] = None):
"""Recommended example list with complete-fit reference media previews."""
example_buttons = []
with gr.Column(elem_classes=["prompt-example-full-table"]):
for row in examples:
example_prompt = str(row[0]) if row else ""
example_cache_key = str(row[-1]) if row and str(row[-1]) in RECOMMENDED_CASE_CACHE else ""
preview_video_path = input_video_path = None
preview_image_path = input_image_path = None
if media_type == "video":
preview_video_path = str(row[1]) if len(row) > 1 and row[1] else None
input_video_path = str(row[2]) if len(row) > 2 and row[2] else preview_video_path
elif media_type == "image":
preview_image_path = str(row[3]) if len(row) > 3 and row[3] else (str(row[2]) if len(row) > 2 and row[2] else None)
input_image_path = str(row[4]) if len(row) > 4 and row[4] else preview_image_path
button_label = example_prompt if len(example_prompt) <= 360 else f"{example_prompt[:357]}..."
if media_type in {"video", "image"}:
with gr.Row(elem_classes=["prompt-example-multimodal-row"]):
with gr.Column(elem_classes=["prompt-example-prompt-cell"]):
example_button = gr.Button(
button_label,
variant="secondary",
elem_classes=["prompt-example-row-button"],
)
with gr.Column(elem_classes=["prompt-example-media-cell"]):
if media_type == "video":
gr.HTML(
build_example_media_html(preview_video_path, "video", fallback_media_path=input_video_path),
elem_classes=["prompt-example-media-html"],
)
else:
gr.HTML(
build_example_media_html(preview_image_path, "image"),
elem_classes=["prompt-example-media-html"],
)
else:
example_button = gr.Button(
button_label,
variant="secondary",
elem_classes=["prompt-example-row-button"],
)
example_buttons.append((example_button, example_prompt, input_video_path, input_image_path, example_cache_key))
return example_buttons
def examples_section(title: str, examples: list[list], media_type: Optional[str] = None, visible: bool = False):
with gr.Column(visible=visible, elem_classes=["lance-recommended-section"]) as group:
gr.HTML(build_lance_label_html(title, "lance-section-label"), elem_classes=["lance-label-html"])
with gr.Group(elem_classes=["example-panel", "prompt-examples"]):
buttons = build_prompt_example_table(examples, media_type=media_type)
return group, buttons
video_generation_examples_group, video_generation_example_buttons = examples_section(
"Video generation recommended cases", VIDEO_GENERATION_EXAMPLES, visible=True
)
video_edit_examples_group, video_edit_example_buttons = examples_section(
"Video edit recommended cases", VIDEO_EDIT_EXAMPLES, media_type="video"
)
video_understanding_examples_group, video_understanding_example_buttons = examples_section(
"Video understanding recommended cases", VIDEO_UNDERSTANDING_EXAMPLES, media_type="video"
)
image_generation_examples_group, image_generation_example_buttons = examples_section(
"Image generation recommended cases", IMAGE_GENERATION_EXAMPLES
)
image_edit_examples_group, image_edit_example_buttons = examples_section(
"Image edit recommended cases", IMAGE_EDIT_EXAMPLES, media_type="image"
)
image_understanding_examples_group, image_understanding_example_buttons = examples_section(
"Image understanding recommended cases", IMAGE_UNDERSTANDING_EXAMPLES, media_type="image"
)
task.change(
fn=update_task_ui,
inputs=[task],
outputs=[
prompt_label,
prompt,
system_prompt,
input_video,
input_image,
frame_interpolation_row,
aspect_ratio_row,
output_resolution_row,
video_duration_row,
video_resolution_row,
aspect_ratio,
height,
width,
enable_frame_interpolation,
real_size,
num_frames,
resolution,
output_label,
output_video,
output_image,
output_text,
video_generation_examples_group,
video_edit_examples_group,
video_understanding_examples_group,
image_generation_examples_group,
image_edit_examples_group,
image_understanding_examples_group,
recommended_case_key,
],
)
aspect_ratio.change(
fn=update_size_from_aspect_ratio,
inputs=[task, aspect_ratio, resolution],
outputs=[height, width, real_size],
queue=False,
show_api=False,
)
# real_size is hidden and derived from task/resolution/aspect_ratio.
# Do not attach a .change handler here: dynamic Dropdown choices can briefly
# contain 360p values while the selected value is 480p (or vice versa),
# which makes Gradio reject the stale value during preprocessing.
resolution.change(
fn=update_output_resolution_from_video_profile,
inputs=[task, aspect_ratio, resolution],
outputs=[real_size, height, width],
queue=False,
show_api=False,
)
for example_button, example_prompt, _, _, example_cache_key in video_generation_example_buttons + image_generation_example_buttons:
example_button.click(
fn=make_prompt_example_click_handler(example_prompt, example_cache_key),
inputs=[task],
outputs=[prompt, system_prompt, aspect_ratio, height, width, num_frames, resolution, real_size],
queue=False,
show_api=False,
)
for example_button, example_prompt, example_video, example_image, example_cache_key in (
video_edit_example_buttons
+ video_understanding_example_buttons
+ image_edit_example_buttons
+ image_understanding_example_buttons
):
example_button.click(
fn=make_media_prompt_example_click_handler(example_prompt, example_video, example_image, example_cache_key),
inputs=[task],
outputs=[prompt, input_video, input_image, system_prompt, aspect_ratio, height, width, num_frames, resolution, real_size],
queue=False,
show_api=False,
)
run_button.click(
fn=build_running_status_markdown,
inputs=[],
outputs=[status],
queue=False,
show_api=False,
).then(
fn=run_task,
inputs=[
task,
prompt,
system_prompt,
input_video,
input_image,
height,
width,
num_frames,
seed,
resolution,
validation_num_timesteps,
validation_timestep_shift,
cfg_text_scale,
enable_frame_interpolation,
],
outputs=[output_video, output_image, output_text, status],
show_progress="minimal",
)
return demo
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Lance multimodal Gradio")
parser.add_argument("--server-name", default=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"))
parser.add_argument("--server-port", type=int, default=int(os.getenv("GRADIO_SERVER_PORT", "7860")))
parser.add_argument("--share", action="store_true", default=env_flag("GRADIO_SHARE", False))
parser.add_argument(
"--gpus",
default=os.getenv("LANCE_GPUS", DEFAULT_GPUS),
help="Comma-separated GPU list, for example: 0,1,2,3,4,5,6",
)
parser.add_argument(
"--queue-size",
type=int,
default=int(os.getenv("LANCE_QUEUE_SIZE", str(DEFAULT_QUEUE_SIZE))),
help="Maximum number of queued Gradio requests.",
)
parser.add_argument(
"--concurrency-limit",
type=int,
default=int(os.getenv("LANCE_CONCURRENCY_LIMIT", str(DEFAULT_CONCURRENCY_LIMIT))),
help="Maximum number of Gradio jobs that may execute concurrently. Use 2 for most GPU Spaces; raise it only when enough GPU memory/pipeline slots are available.",
)
return parser.parse_args()
def parse_gpu_ids(gpu_string: str) -> list[int]:
gpu_ids: list[int] = []
for item in gpu_string.split(","):
item = item.strip()
if not item:
continue
gpu_ids.append(int(item))
if not gpu_ids:
raise ValueError("No valid GPU IDs were parsed.")
return gpu_ids
def prefetch_model_assets_before_launch() -> None:
"""Download and compact model files before the first ZeroGPU request.
On ZeroGPU, time spent downloading model snapshots inside @spaces.GPU burns
the first user's GPU reservation. Prefetching only touches CPU/disk and keeps
the visible UI unchanged. Set LANCE_PREFETCH_MODEL_ASSETS=0 to skip this at
Space startup, or LANCE_PREFETCH_MODEL_VARIANTS=video to prefetch less.
"""
if running_on_space() or env_flag("LANCE_INSTALL_FLASH_ATTN_ON_STARTUP", False):
try:
ensure_flash_attn_installed()
except Exception as exc:
print(f"[startup] flash-attn startup install failed and will be retried lazily during inference: {exc}", flush=True)
if not env_flag("LANCE_PREFETCH_MODEL_ASSETS", running_on_space()):
print("[startup] Model asset prefetch disabled.", flush=True)
return
variants_text = os.getenv("LANCE_PREFETCH_MODEL_VARIANTS", f"{MODEL_VARIANT_VIDEO},{MODEL_VARIANT_IMAGE}")
variants: list[str] = []
for raw_variant in variants_text.split(","):
raw_variant = raw_variant.strip()
if not raw_variant:
continue
variant = normalize_model_variant(raw_variant)
if variant not in variants:
variants.append(variant)
for variant in variants:
try:
start = time.perf_counter()
model_path = ensure_model_assets(variant)
elapsed = time.perf_counter() - start
print(
f"[startup][{variant}] Model assets are ready at {display_path(model_path)} "
f"before ZeroGPU inference. elapsed={elapsed:.2f}s",
flush=True,
)
except Exception as exc:
print(
f"[startup][{variant}] Model asset prefetch failed and will be retried lazily during inference: {exc}",
flush=True,
)
if __name__ == "__main__":
args = parse_args()
os.environ["LANCE_GPUS"] = args.gpus
QUEUE_MAX_SIZE = args.queue_size
QUEUE_CONCURRENCY_LIMIT = max(1, args.concurrency_limit)
prefetch_model_assets_before_launch()
print(
"[startup] Skipping GPU model preload. UI will launch first, and Lance weights will be prefetched on CPU before ZeroGPU inference. If that prefetch fails, inference will fall back to lazy loading.",
flush=True,
)
print(
f"[startup] Gradio queue configured with max_size={QUEUE_MAX_SIZE}, default_concurrency_limit={QUEUE_CONCURRENCY_LIMIT}.",
flush=True,
)
demo = build_demo()
demo.queue(
max_size=QUEUE_MAX_SIZE,
default_concurrency_limit=QUEUE_CONCURRENCY_LIMIT,
).launch(
server_name=args.server_name,
server_port=args.server_port,
share=args.share,
allowed_paths=[str(REPO_ROOT.resolve()), str(GRADIO_TMP_ROOT.resolve())],
ssr_mode=False,
)
|