File size: 155,768 Bytes
dfefe0b | 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 | diff --git a/src/transformers/models/blip_2/modeling_blip_2.py b/src/transformers/models/blip_2/modeling_blip_2.py
index 84f0356cecb2..916631da7e8f 100644
--- a/src/transformers/models/blip_2/modeling_blip_2.py
+++ b/src/transformers/models/blip_2/modeling_blip_2.py
@@ -2016,6 +2016,9 @@ def forward(
class Blip2ForConditionalGeneration(Blip2PreTrainedModel, GenerationMixin):
config_class = Blip2Config
main_input_name = "pixel_values"
+ _supports_cache_class = True
+ _supports_static_cache = True
+ _supports_quantized_cache = False # not all LM bacbones support (e.g. T5)
def __init__(self, config: Blip2Config):
super().__init__(config)
diff --git a/src/transformers/models/chameleon/modeling_chameleon.py b/src/transformers/models/chameleon/modeling_chameleon.py
index 0a9421409e25..65322e236ca0 100644
--- a/src/transformers/models/chameleon/modeling_chameleon.py
+++ b/src/transformers/models/chameleon/modeling_chameleon.py
@@ -1284,13 +1284,13 @@ def forward(
if pixel_values is not None:
image_tokens = self.get_image_tokens(pixel_values)
- n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum().item()
- n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
- if n_image_tokens_in_text != n_image_features:
+ special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_tokens.numel():
+ n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum()
+ n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens_in_text}, features {n_image_features}"
)
- special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
image_tokens = image_tokens.to(input_ids.device, input_ids.dtype)
input_ids = input_ids.masked_scatter(special_image_mask, image_tokens)
diff --git a/src/transformers/models/cohere2/modeling_cohere2.py b/src/transformers/models/cohere2/modeling_cohere2.py
index 11353a0a990c..75144c65ecff 100644
--- a/src/transformers/models/cohere2/modeling_cohere2.py
+++ b/src/transformers/models/cohere2/modeling_cohere2.py
@@ -25,7 +25,7 @@
import torch.nn as nn
from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
@@ -701,7 +701,7 @@ def _update_causal_mask(
dtype, device = input_tensor.dtype, input_tensor.device
sequence_length = input_tensor.shape[1]
- if isinstance(past_key_values, HybridCache):
+ if isinstance(past_key_values, (HybridCache, StaticCache)):
target_length = past_key_values.get_max_cache_shape()
else:
target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
diff --git a/src/transformers/models/gemma2/modeling_gemma2.py b/src/transformers/models/gemma2/modeling_gemma2.py
index e9fd43c49000..c977f873dc8c 100644
--- a/src/transformers/models/gemma2/modeling_gemma2.py
+++ b/src/transformers/models/gemma2/modeling_gemma2.py
@@ -25,7 +25,7 @@
import torch.nn as nn
from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_outputs import (
@@ -713,7 +713,7 @@ def _update_causal_mask(
dtype, device = input_tensor.dtype, input_tensor.device
sequence_length = input_tensor.shape[1]
- if isinstance(past_key_values, HybridCache):
+ if isinstance(past_key_values, (HybridCache, StaticCache)):
target_length = past_key_values.get_max_cache_shape()
else:
target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
diff --git a/src/transformers/models/gemma2/modular_gemma2.py b/src/transformers/models/gemma2/modular_gemma2.py
index 4e3c8487c4d8..805e6ba0d2a3 100644
--- a/src/transformers/models/gemma2/modular_gemma2.py
+++ b/src/transformers/models/gemma2/modular_gemma2.py
@@ -20,7 +20,7 @@
import torch.utils.checkpoint
from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
from ...configuration_utils import PretrainedConfig
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_outputs import (
@@ -550,7 +550,7 @@ def _update_causal_mask(
dtype, device = input_tensor.dtype, input_tensor.device
sequence_length = input_tensor.shape[1]
- if isinstance(past_key_values, HybridCache):
+ if isinstance(past_key_values, (HybridCache, StaticCache)):
target_length = past_key_values.get_max_cache_shape()
else:
target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]
diff --git a/src/transformers/models/got_ocr2/configuration_got_ocr2.py b/src/transformers/models/got_ocr2/configuration_got_ocr2.py
index 480252ab1471..fb9a1fb68889 100644
--- a/src/transformers/models/got_ocr2/configuration_got_ocr2.py
+++ b/src/transformers/models/got_ocr2/configuration_got_ocr2.py
@@ -132,8 +132,6 @@ class GotOcr2Config(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 151859):
The image token index to encode the image prompt.
image_seq_length (`int`, *optional*, defaults to 576):
@@ -161,13 +159,11 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=151859,
image_seq_length=576,
pad_token_id=-1,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.image_seq_length = image_seq_length
self.pad_token_id = pad_token_id
diff --git a/src/transformers/models/got_ocr2/modeling_got_ocr2.py b/src/transformers/models/got_ocr2/modeling_got_ocr2.py
index 957e05bea75a..86598ac08965 100644
--- a/src/transformers/models/got_ocr2/modeling_got_ocr2.py
+++ b/src/transformers/models/got_ocr2/modeling_got_ocr2.py
@@ -594,6 +594,8 @@ class GotOcr2PreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
# important: this ported version of GotOcr2 isn't meant for training from scratch - only
@@ -748,89 +750,6 @@ def get_image_features(
image_outputs = self.vision_tower(pixel_values).last_hidden_state
return self.multi_modal_projector(image_outputs)
- def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
- num_images, num_image_patches, embed_dim = image_features.shape
- batch_size, sequence_length = input_ids.shape
- left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == self.config.image_token_index
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # Compute the maximum embed dimension
- max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
- batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
-
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
- nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
- if left_padding:
- new_token_positions += nb_image_pad[:, None] # offset for left padding
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- final_embedding = torch.zeros(
- batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- if labels is not None:
- final_labels = torch.full(
- (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- if labels is not None:
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
- # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
- image_to_overwrite = torch.full(
- (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
- )
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- if left_padding:
- image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
- else:
- mask = torch.ones_like(image_to_overwrite, dtype=torch.bool).cumsum(-1) - 1
- padding_mask = mask <= new_token_positions[:, -1:].to(target_device)
- image_to_overwrite &= padding_mask
-
- if image_to_overwrite.sum() != image_features.shape[:-1].numel():
- raise ValueError(
- f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
- )
-
- final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
- batch_indices, pad_indices = torch.where(input_ids == self.pad_token_id)
- indices_to_mask = new_token_positions[batch_indices, pad_indices]
-
- final_embedding[batch_indices, indices_to_mask] = 0
-
- if labels is None:
- final_labels = None
-
- return final_embedding, final_attention_mask, final_labels, position_ids
-
@add_start_docstrings_to_model_forward(GOT_OCR2_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=GotOcr2CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
def forward(
diff --git a/src/transformers/models/got_ocr2/modular_got_ocr2.py b/src/transformers/models/got_ocr2/modular_got_ocr2.py
index 899075683eb4..fff434ead2e9 100644
--- a/src/transformers/models/got_ocr2/modular_got_ocr2.py
+++ b/src/transformers/models/got_ocr2/modular_got_ocr2.py
@@ -170,8 +170,6 @@ class GotOcr2Config(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 151859):
The image token index to encode the image prompt.
image_seq_length (`int`, *optional*, defaults to 576):
@@ -199,13 +197,11 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=151859,
image_seq_length=576,
pad_token_id=-1,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.image_seq_length = image_seq_length
self.pad_token_id = pad_token_id
diff --git a/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py b/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
index d5153fb3f828..10b6efbc5943 100755
--- a/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
+++ b/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
@@ -51,7 +51,7 @@ class GPTNeoXJapanesePreTrainedModel(PreTrainedModel):
_skip_keys_device_placement = "past_key_values"
_supports_cache_class = True
_supports_quantized_cache = True
- _supports_static_cache = False # TODO (fix me): compilation fails due to a stide error?
+ _supports_static_cache = True
def _init_weights(self, module):
"""Initialize the weights"""
@@ -129,8 +129,8 @@ def forward(
cos, sin = position_embeddings
query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
- query = torch.cat((query, query_pass), dim=-1)
- key = torch.cat((key, key_pass), dim=-1)
+ query = torch.cat((query, query_pass), dim=-1).contiguous()
+ key = torch.cat((key, key_pass), dim=-1).contiguous()
# Cache QKV values
if layer_past is not None:
diff --git a/src/transformers/models/granitemoe/modeling_granitemoe.py b/src/transformers/models/granitemoe/modeling_granitemoe.py
index d877b8323b3b..546e78eac148 100644
--- a/src/transformers/models/granitemoe/modeling_granitemoe.py
+++ b/src/transformers/models/granitemoe/modeling_granitemoe.py
@@ -1108,6 +1108,7 @@ def forward(
router_logits=all_router_logits,
)
+ # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
def _update_causal_mask(
self,
attention_mask: torch.Tensor,
@@ -1116,13 +1117,8 @@ def _update_causal_mask(
past_key_values: Cache,
output_attentions: bool,
):
- # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
- # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
- # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
- # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
-
if self.config._attn_implementation == "flash_attention_2":
- if attention_mask is not None and 0.0 in attention_mask:
+ if attention_mask is not None and (attention_mask == 0.0).any():
return attention_mask
return None
@@ -1143,7 +1139,6 @@ def _update_causal_mask(
return None
dtype, device = input_tensor.dtype, input_tensor.device
- min_dtype = torch.finfo(dtype).min
sequence_length = input_tensor.shape[1]
if using_static_cache:
target_length = past_key_values.get_max_cache_shape()
@@ -1154,25 +1149,17 @@ def _update_causal_mask(
else past_seen_tokens + sequence_length + 1
)
- if attention_mask is not None and attention_mask.dim() == 4:
- # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
- causal_mask = attention_mask
- else:
- causal_mask = torch.full(
- (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
- )
- if sequence_length != 1:
- causal_mask = torch.triu(causal_mask, diagonal=1)
- causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
- causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
- if attention_mask is not None:
- causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
- mask_length = attention_mask.shape[-1]
- padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
- padding_mask = padding_mask == 0
- causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
- padding_mask, min_dtype
- )
+ # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+ causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask,
+ sequence_length=sequence_length,
+ target_length=target_length,
+ dtype=dtype,
+ device=device,
+ cache_position=cache_position,
+ batch_size=input_tensor.shape[0],
+ )
+
if (
self.config._attn_implementation == "sdpa"
and attention_mask is not None
@@ -1182,6 +1169,7 @@ def _update_causal_mask(
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
# Details: https://github.com/pytorch/pytorch/issues/110213
+ min_dtype = torch.finfo(dtype).min
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
return causal_mask
diff --git a/src/transformers/models/instructblip/modeling_instructblip.py b/src/transformers/models/instructblip/modeling_instructblip.py
index b705da44eba4..ea42d65b845c 100644
--- a/src/transformers/models/instructblip/modeling_instructblip.py
+++ b/src/transformers/models/instructblip/modeling_instructblip.py
@@ -1290,6 +1290,9 @@ def forward(
class InstructBlipForConditionalGeneration(InstructBlipPreTrainedModel, GenerationMixin):
config_class = InstructBlipConfig
main_input_name = "pixel_values"
+ _supports_cache_class = True
+ _supports_static_cache = True
+ _supports_quantized_cache = False # not all LM bacbones support (e.g. T5)
def __init__(self, config: InstructBlipConfig):
super().__init__(config)
diff --git a/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py b/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
index dcf77863a149..5183a3c22faf 100644
--- a/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
+++ b/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
@@ -1284,6 +1284,9 @@ def forward(
class InstructBlipVideoForConditionalGeneration(InstructBlipVideoPreTrainedModel, GenerationMixin):
config_class = InstructBlipVideoConfig
main_input_name = "pixel_values"
+ _supports_cache_class = True
+ _supports_static_cache = True
+ _supports_quantized_cache = False # not all LM bacbones support (e.g. T5)
def __init__(self, config: InstructBlipVideoConfig):
super().__init__(config)
diff --git a/src/transformers/models/llava/configuration_llava.py b/src/transformers/models/llava/configuration_llava.py
index d2a3e9747b66..f476591b2eb6 100644
--- a/src/transformers/models/llava/configuration_llava.py
+++ b/src/transformers/models/llava/configuration_llava.py
@@ -37,8 +37,6 @@ class LlavaConfig(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32000):
The image token index to encode the image prompt.
projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -83,7 +81,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32000,
projector_hidden_act="gelu",
vision_feature_select_strategy="default",
@@ -92,7 +89,6 @@ def __init__(
multimodal_projector_bias=True,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.projector_hidden_act = projector_hidden_act
self.image_seq_length = image_seq_length
diff --git a/src/transformers/models/llava/modeling_llava.py b/src/transformers/models/llava/modeling_llava.py
index 36f212e76844..610ab417d92b 100644
--- a/src/transformers/models/llava/modeling_llava.py
+++ b/src/transformers/models/llava/modeling_llava.py
@@ -28,6 +28,7 @@
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
+ is_torchdynamo_compiling,
logging,
replace_return_docstrings,
)
@@ -136,6 +137,8 @@ class LlavaPreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
# important: this ported version of Llava isn't meant for training from scratch - only
@@ -321,89 +324,6 @@ def get_image_features(
image_features = self.multi_modal_projector(selected_image_feature)
return image_features
- def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
- num_images, num_image_patches, embed_dim = image_features.shape
- batch_size, sequence_length = input_ids.shape
- left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == self.config.image_token_index
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # Compute the maximum embed dimension
- max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
- batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
-
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
- nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
- if left_padding:
- new_token_positions += nb_image_pad[:, None] # offset for left padding
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- final_embedding = torch.zeros(
- batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- if labels is not None:
- final_labels = torch.full(
- (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- if labels is not None:
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
- # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
- image_to_overwrite = torch.full(
- (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
- )
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- if left_padding:
- image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
- else:
- mask = torch.ones_like(image_to_overwrite, dtype=torch.bool).cumsum(-1) - 1
- padding_mask = mask <= new_token_positions[:, -1:].to(target_device)
- image_to_overwrite &= padding_mask
-
- if image_to_overwrite.sum() != image_features.shape[:-1].numel():
- raise ValueError(
- f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
- )
-
- final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
- batch_indices, pad_indices = torch.where(input_ids == self.pad_token_id)
- indices_to_mask = new_token_positions[batch_indices, pad_indices]
-
- final_embedding[batch_indices, indices_to_mask] = 0
-
- if labels is None:
- final_labels = None
-
- return final_embedding, final_attention_mask, final_labels, position_ids
-
@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
@@ -499,14 +419,14 @@ def forward(
image_sizes=image_sizes,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum()
- n_image_features = image_features.shape[0] * image_features.shape[1]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0] * image_features.shape[1]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
diff --git a/src/transformers/models/llava_next/configuration_llava_next.py b/src/transformers/models/llava_next/configuration_llava_next.py
index 2610275cedfd..3836dbf71cd2 100644
--- a/src/transformers/models/llava_next/configuration_llava_next.py
+++ b/src/transformers/models/llava_next/configuration_llava_next.py
@@ -36,8 +36,6 @@ class LlavaNextConfig(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32000):
The image token index to encode the image prompt.
projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -88,7 +86,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32000,
projector_hidden_act="gelu",
vision_feature_select_strategy="default",
@@ -99,7 +96,6 @@ def __init__(
multimodal_projector_bias=True,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.projector_hidden_act = projector_hidden_act
self.image_seq_length = image_seq_length
diff --git a/src/transformers/models/llava_next/modeling_llava_next.py b/src/transformers/models/llava_next/modeling_llava_next.py
index 06e1cc63940f..3cdf1b348404 100644
--- a/src/transformers/models/llava_next/modeling_llava_next.py
+++ b/src/transformers/models/llava_next/modeling_llava_next.py
@@ -31,6 +31,7 @@
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
+ is_torchdynamo_compiling,
logging,
replace_return_docstrings,
)
@@ -245,6 +246,8 @@ class LlavaNextPreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
# important: this ported version of LlavaNext isn't meant for training from scratch - only
@@ -405,245 +408,6 @@ def set_decoder(self, decoder):
def get_decoder(self):
return self.language_model.get_decoder()
- def _merge_input_ids_with_image_features(
- self,
- image_features,
- feature_lens,
- inputs_embeds,
- input_ids,
- attention_mask,
- position_ids=None,
- labels=None,
- image_token_index=None,
- ignore_index=-100,
- ):
- """
- Merge input_ids with with image features into final embeddings
-
- Args:
- image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
- All vision vectors of all images in the batch
- feature_lens (`torch.LongTensor` of shape `(num_images)`):
- The length of visual embeddings of each image as stacked in `image_features`
- inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
- Token embeddings before merging with visual embeddings
- input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Input_ids of tokens, possibly filled with image token
- attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Mask to avoid performing attention on padding token indices.
- position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
- config.n_positions - 1]`.
- labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
- :abels need to be recalculated to support training (if provided)
- image_token_index (`int`, *optional*)
- Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
- ignore_index (`int`, *optional*)
- Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
- Returns:
- final_embedding, final_attention_mask, position_ids, final_labels
-
- Explanation:
- each image has variable length embeddings, with length specified by feature_lens
- image_features is concatenation of all visual embed vectors
- task: fill each <image> with the correct number of visual embeddings
- Example:
- X (5 patches), Y (3 patches), Z (8)
- X, Y are in the same sequence (in-context learning)
- if right padding
- input_ids: [
- a b c d e f X g h i j k Y l m
- o p q r Z s t u v _ _ _ _ _ _
- ]
- input_ids should be: [
- a b c d e f X X X X X g h i j k Y Y Y l m
- o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
- ]
- labels should be: [
- a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
- o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
- ]
- elif left padding
- input_ids: [
- a b c d e f X g h i j k Y l m
- _ _ _ _ _ _ o p q r Z s t u v
- ]
- input_ids should be: [
- a b c d e f X X X X X g h i j k Y Y Y l m
- _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
- ]
- labels should be: [
- a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
- _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
- ]
- Edge cases:
- * If tokens are same but image token sizes are different, then cannot infer left or right padding
- ```python
- cat_img = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
- chart_img = Image.open(requests.get("https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true", stream=True).raw)
- prompts = [
- "[INST] <image>\nWhat is shown in this image? [/INST]",
- "[INST] <image>\nWhat is shown in this image? [/INST]",
- ]
- inputs = processor(prompts, [chart_img, cat_img], return_tensors='pt', padding=True).to("cuda")
- chart_img has 2634 tokens, while cat_img has 2340 tokens
- ```
-
- input_ids: [
- a b c d X g h
- i j Y k l m n
- ]
- where X is 3 tokens while Y is 5, this mean after merge
- if left-padding (batched generation)
- input_ids should be: [
- _ _ a b c d X X X g h
- i j Y Y Y Y Y k l m n
- ]
- elif (right padding) (training)
- input_ids should be: [
- a b c d X X X g h _ _
- i j Y Y Y Y Y k l m n
- ]
- """
- image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
- ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index
-
- if self.training and self.padding_side == "left":
- logger.warning_once(
- "Padding side is set to 'left' but the model is in training mode. For training "
- "it is recommended to set `model.padding_side='right' and `processor.tokenizer.padding_side='right'`. "
- "If that's intended, ignore this warning"
- )
- if not self.training and self.padding_side == "right":
- logger.warning_once(
- "Padding side is set to 'right' but the model is in inference mode. For correct "
- "generation results, please set `model.padding_side='left'` and `processor.tokenizer.padding_side='left'`. "
- "If that's intended, ignore this warning"
- )
-
- with torch.no_grad():
- # ! in llava 1.6, number of patches is variable
- num_images = feature_lens.size(0)
- num_image_features, embed_dim = image_features.shape
- if feature_lens.sum() != num_image_features:
- raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
- batch_size = input_ids.shape[0]
- _left_padding = torch.any(attention_mask[:, 0] == 0)
- _right_padding = torch.any(attention_mask[:, -1] == 0)
-
- left_padding = self.padding_side == "left"
- if batch_size > 1:
- if _left_padding and _right_padding:
- raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
- elif _right_padding and left_padding:
- left_padding = False
- elif _left_padding and not left_padding:
- left_padding = True
-
- # Whether to turn off right padding
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == image_token_index
- # special_image_token_mask: [bsz, seqlen]
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # num_special_image_tokens: [bsz]
- # Reserve for padding of num_images
- total_num_special_image_tokens = torch.sum(special_image_token_mask)
- if total_num_special_image_tokens != num_images:
- raise ValueError(
- f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
- )
- # Compute the maximum embed dimension
- # max_image_feature_lens is max_feature_lens per batch
- feature_lens = feature_lens.to(input_ids.device)
- feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
- feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=input_ids.device)
- embed_sequence_lengths = (
- (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
- )
- max_embed_dim = embed_sequence_lengths.max()
-
- batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- # ! instead of special_image_token_mask * (num_image_patches - 1)
- # special_image_token_mask * (num_feature_len - 1)
- special_image_token_mask = special_image_token_mask.long()
- special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
- new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
- if left_padding:
- # shift right token positions so that they are ending at the same number
- # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
- new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]
-
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- final_embedding = torch.zeros(
- batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- final_input_ids = torch.full(
- (batch_size, max_embed_dim), self.pad_token_id, dtype=input_ids.dtype, device=inputs_embeds.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
- input_ids = input_ids.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- final_input_ids[batch_indices, text_to_overwrite] = input_ids[batch_indices, non_image_indices]
- final_labels = None
- if labels is not None:
- labels = labels.to(target_device)
- final_labels = torch.full_like(final_attention_mask, ignore_index).to(torch.long)
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
- # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
- with torch.no_grad():
- image_to_overwrite = torch.full(
- (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
- )
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
- embed_indices = embed_indices.expand(batch_size, max_embed_dim)
- embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)
-
- if left_padding:
- # exclude padding on the left
- max_embed_dim = max_embed_dim.to(target_device)
- val = (max_embed_dim - embed_indices) <= embed_seq_lens
- else:
- # exclude padding on the right
- val = embed_indices < embed_seq_lens
- image_to_overwrite &= val
-
- if image_to_overwrite.sum() != num_image_features:
- raise ValueError(
- f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
- f"The number of image tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of image given to the model is {num_images}. "
- f"This prevents correct indexing and breaks batch generation."
- )
- final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- return final_embedding, final_attention_mask, position_ids, final_labels, final_input_ids
-
def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
"""
Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.
@@ -875,14 +639,14 @@ def forward(
image_newline=self.image_newline,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
diff --git a/src/transformers/models/llava_next_video/configuration_llava_next_video.py b/src/transformers/models/llava_next_video/configuration_llava_next_video.py
index 6b85ebb4455e..01450f6b587c 100644
--- a/src/transformers/models/llava_next_video/configuration_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/configuration_llava_next_video.py
@@ -38,8 +38,6 @@ class LlavaNextVideoConfig(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32001):
The image token index to encode the image prompt.
projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -96,7 +94,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32001,
projector_hidden_act="gelu",
multimodal_projector_bias=True,
@@ -116,7 +113,6 @@ def __init__(
self.spatial_pool_stride = spatial_pool_stride
self.image_seq_length = image_seq_length
self.video_seq_length = video_seq_length
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.projector_hidden_act = projector_hidden_act
self.multimodal_projector_bias = multimodal_projector_bias
diff --git a/src/transformers/models/llava_next_video/modeling_llava_next_video.py b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
index f62824947ddf..9ce88c541231 100644
--- a/src/transformers/models/llava_next_video/modeling_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
@@ -32,7 +32,13 @@
from ...image_processing_utils import select_best_resolution
from ...modeling_outputs import ModelOutput
from ...modeling_utils import PreTrainedModel
-from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from ...utils import (
+ add_start_docstrings,
+ add_start_docstrings_to_model_forward,
+ is_torchdynamo_compiling,
+ logging,
+ replace_return_docstrings,
+)
from ...utils.deprecation import deprecate_kwarg
from ..auto import AutoModel, AutoModelForCausalLM
from .configuration_llava_next_video import LlavaNextVideoConfig
@@ -153,6 +159,8 @@ class LlavaNextVideoPreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
# important: this ported version of LlavaNextVideo isn't meant for training from scratch - only
@@ -440,245 +448,6 @@ def set_decoder(self, decoder):
def get_decoder(self):
return self.language_model.get_decoder()
- def _merge_input_ids_with_image_features(
- self,
- image_features,
- feature_lens,
- inputs_embeds,
- input_ids,
- attention_mask,
- position_ids=None,
- labels=None,
- image_token_index=None,
- ignore_index=-100,
- ):
- """
- Merge input_ids with with image features into final embeddings
-
- Args:
- image_features (`torch.Tensor` of shape `(all_feature_lens, embed_dim)`):
- All vision vectors of all images in the batch
- feature_lens (`torch.LongTensor` of shape `(num_images)`):
- The length of visual embeddings of each image as stacked in `image_features`
- inputs_embeds (`torch.Tensor` of shape `(batch_size, sequence_length, embed_dim)`):
- Token embeddings before merging with visual embeddings
- input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Input_ids of tokens, possibly filled with image token
- attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Mask to avoid performing attention on padding token indices.
- position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
- Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
- config.n_positions - 1]`.
- labels (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*)
- :abels need to be recalculated to support training (if provided)
- image_token_index (`int`, *optional*)
- Token id used to indicate the special "image" token. Defaults to `config.image_token_index`
- ignore_index (`int`, *optional*)
- Value that is used to pad `labels` and will be ignored when calculated loss. Default: -100.
- Returns:
- final_embedding, final_attention_mask, position_ids, final_labels
-
- Explanation:
- each image has variable length embeddings, with length specified by feature_lens
- image_features is concatenation of all visual embed vectors
- task: fill each <image> with the correct number of visual embeddings
- Example:
- X (5 patches), Y (3 patches), Z (8)
- X, Y are in the same sequence (in-context learning)
- if right padding
- input_ids: [
- a b c d e f X g h i j k Y l m
- o p q r Z s t u v _ _ _ _ _ _
- ]
- input_ids should be: [
- a b c d e f X X X X X g h i j k Y Y Y l m
- o p q r Z Z Z Z Z Z Z Z s t u v _ _ _ _ _
- ]
- labels should be: [
- a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
- o p q r _ _ _ _ _ _ _ _ s t u v _ _ _ _ _
- ]
- elif left padding
- input_ids: [
- a b c d e f X g h i j k Y l m
- _ _ _ _ _ _ o p q r Z s t u v
- ]
- input_ids should be: [
- a b c d e f X X X X X g h i j k Y Y Y l m
- _ _ _ _ _ o p q r Z Z Z Z Z Z Z Z s t u v
- ]
- labels should be: [
- a b c d e f _ _ _ _ _ g h i j k _ _ _ l m
- _ _ _ _ _ o p q r _ _ _ _ _ _ _ _ s t u v
- ]
- Edge cases:
- * If tokens are same but image token sizes are different, then cannot infer left or right padding
- ```python
- cat_img = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
- chart_img = Image.open(requests.get("https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true", stream=True).raw)
- prompts = [
- "[INST] <image>\nWhat is shown in this image? [/INST]",
- "[INST] <image>\nWhat is shown in this image? [/INST]",
- ]
- inputs = processor(prompts, [chart_img, cat_img], return_tensors='pt', padding=True).to("cuda")
- chart_img has 2634 tokens, while cat_img has 2340 tokens
- ```
-
- input_ids: [
- a b c d X g h
- i j Y k l m n
- ]
- where X is 3 tokens while Y is 5, this mean after merge
- if left-padding (batched generation)
- input_ids should be: [
- _ _ a b c d X X X g h
- i j Y Y Y Y Y k l m n
- ]
- elif (right padding) (training)
- input_ids should be: [
- a b c d X X X g h _ _
- i j Y Y Y Y Y k l m n
- ]
- """
- image_token_index = image_token_index if image_token_index is not None else self.config.image_token_index
- ignore_index = ignore_index if ignore_index is not None else self.config.ignore_index
-
- if self.training and self.padding_side == "left":
- logger.warning_once(
- "Padding side is set to 'left' but the model is in training mode. For training "
- "it is recommended to set `model.padding_side='right' and `processor.tokenizer.padding_side='right'`. "
- "If that's intended, ignore this warning"
- )
- if not self.training and self.padding_side == "right":
- logger.warning_once(
- "Padding side is set to 'right' but the model is in inference mode. For correct "
- "generation results, please set `model.padding_side='left'` and `processor.tokenizer.padding_side='left'`. "
- "If that's intended, ignore this warning"
- )
-
- with torch.no_grad():
- # ! in llava 1.6, number of patches is variable
- num_images = feature_lens.size(0)
- num_image_features, embed_dim = image_features.shape
- if feature_lens.sum() != num_image_features:
- raise ValueError(f"{feature_lens=} / {feature_lens.sum()} != {image_features.shape=}")
- batch_size = input_ids.shape[0]
- _left_padding = torch.any(attention_mask[:, 0] == 0)
- _right_padding = torch.any(attention_mask[:, -1] == 0)
-
- left_padding = self.padding_side == "left"
- if batch_size > 1:
- if _left_padding and _right_padding:
- raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
- elif _right_padding and left_padding:
- left_padding = False
- elif _left_padding and not left_padding:
- left_padding = True
-
- # Whether to turn off right padding
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == image_token_index
- # special_image_token_mask: [bsz, seqlen]
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # num_special_image_tokens: [bsz]
- # Reserve for padding of num_images
- total_num_special_image_tokens = torch.sum(special_image_token_mask)
- if total_num_special_image_tokens != num_images:
- raise ValueError(
- f"Number of image tokens in input_ids ({total_num_special_image_tokens}) different from num_images ({num_images})."
- )
- # Compute the maximum embed dimension
- # max_image_feature_lens is max_feature_lens per batch
- feature_lens = feature_lens.to(input_ids.device)
- feature_lens_batch = feature_lens.split(num_special_image_tokens.tolist(), dim=0)
- feature_lens_batch_sum = torch.tensor([x.sum() for x in feature_lens_batch], device=input_ids.device)
- embed_sequence_lengths = (
- (attention_mask == 1).long().sum(-1) - num_special_image_tokens + feature_lens_batch_sum
- )
- max_embed_dim = embed_sequence_lengths.max()
-
- batch_indices, non_image_indices = torch.where((input_ids != image_token_index) & (attention_mask == 1))
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- # ! instead of special_image_token_mask * (num_image_patches - 1)
- # special_image_token_mask * (num_feature_len - 1)
- special_image_token_mask = special_image_token_mask.long()
- special_image_token_mask[special_image_token_mask == 1] = feature_lens - 1
- new_token_positions = torch.cumsum((special_image_token_mask + 1), -1) - 1
- if left_padding:
- # shift right token positions so that they are ending at the same number
- # the below here was incorrect? new_token_positions += new_token_positions[:, -1].max() - new_token_positions[:, -1:]
- new_token_positions += max_embed_dim - 1 - new_token_positions[:, -1:]
-
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- final_embedding = torch.zeros(
- batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- final_input_ids = torch.full(
- (batch_size, max_embed_dim), self.pad_token_id, dtype=input_ids.dtype, device=inputs_embeds.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
- input_ids = input_ids.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- final_input_ids[batch_indices, text_to_overwrite] = input_ids[batch_indices, non_image_indices]
- final_labels = None
- if labels is not None:
- labels = labels.to(target_device)
- final_labels = torch.full_like(final_attention_mask, ignore_index).to(torch.long)
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
- # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
- with torch.no_grad():
- image_to_overwrite = torch.full(
- (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
- )
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- embed_indices = torch.arange(max_embed_dim).unsqueeze(0).to(target_device)
- embed_indices = embed_indices.expand(batch_size, max_embed_dim)
- embed_seq_lens = embed_sequence_lengths[:, None].to(target_device)
-
- if left_padding:
- # exclude padding on the left
- max_embed_dim = max_embed_dim.to(target_device)
- val = (max_embed_dim - embed_indices) <= embed_seq_lens
- else:
- # exclude padding on the right
- val = embed_indices < embed_seq_lens
- image_to_overwrite &= val
-
- if image_to_overwrite.sum() != num_image_features:
- raise ValueError(
- f"{image_to_overwrite.sum()=} != {num_image_features=} The input provided to the model are wrong. "
- f"The number of image tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of image given to the model is {num_images}. "
- f"This prevents correct indexing and breaks batch generation."
- )
- final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- return final_embedding, final_attention_mask, position_ids, final_labels, final_input_ids
-
def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
"""
Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.
@@ -948,14 +717,14 @@ def forward(
image_newline=self.image_newline,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
@@ -970,14 +739,14 @@ def forward(
video_features = torch.cat(video_features, dim=0)
video_feature_lens = torch.tensor(video_feature_lens, dtype=torch.long, device=video_features.device)
- n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
- n_video_features = video_features.shape[0]
- if n_video_tokens != n_video_features:
+ special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != video_features.numel():
+ n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
+ n_video_features = video_features.shape[0]
raise ValueError(
f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
)
- special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, video_features)
diff --git a/src/transformers/models/llava_next_video/modular_llava_next_video.py b/src/transformers/models/llava_next_video/modular_llava_next_video.py
index b2e06c337c1b..8769f8db4131 100644
--- a/src/transformers/models/llava_next_video/modular_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/modular_llava_next_video.py
@@ -30,6 +30,7 @@
from ...configuration_utils import PretrainedConfig
from ...utils import (
+ is_torchdynamo_compiling,
logging,
)
from ..auto import CONFIG_MAPPING, AutoConfig
@@ -52,8 +53,6 @@ class LlavaNextVideoConfig(PretrainedConfig):
The config object or dictionary of the vision backbone.
text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
The config object or dictionary of the text backbone.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32001):
The image token index to encode the image prompt.
projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -110,7 +109,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32001,
projector_hidden_act="gelu",
multimodal_projector_bias=True,
@@ -130,7 +128,6 @@ def __init__(
self.spatial_pool_stride = spatial_pool_stride
self.image_seq_length = image_seq_length
self.video_seq_length = video_seq_length
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.projector_hidden_act = projector_hidden_act
self.multimodal_projector_bias = multimodal_projector_bias
@@ -479,14 +476,14 @@ def forward(
image_newline=self.image_newline,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
@@ -501,14 +498,14 @@ def forward(
video_features = torch.cat(video_features, dim=0)
video_feature_lens = torch.tensor(video_feature_lens, dtype=torch.long, device=video_features.device)
- n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
- n_video_features = video_features.shape[0]
- if n_video_tokens != n_video_features:
+ special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != video_features.numel():
+ n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
+ n_video_features = video_features.shape[0]
raise ValueError(
f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
)
- special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, video_features)
diff --git a/src/transformers/models/llava_onevision/modeling_llava_onevision.py b/src/transformers/models/llava_onevision/modeling_llava_onevision.py
index ed584bda7f5d..e86ce394e13d 100644
--- a/src/transformers/models/llava_onevision/modeling_llava_onevision.py
+++ b/src/transformers/models/llava_onevision/modeling_llava_onevision.py
@@ -30,6 +30,7 @@
from ...modeling_utils import PreTrainedModel
from ...utils import (
add_start_docstrings,
+ is_torchdynamo_compiling,
logging,
)
from ...utils.deprecation import deprecate_kwarg
@@ -250,7 +251,7 @@ class LlavaOnevisionPreTrainedModel(PreTrainedModel):
_skip_keys_device_placement = "past_key_values"
_supports_flash_attn_2 = True
_supports_cache_class = True
- _supports_static_cache = False # Qwen2 doesn't but llava has no reasons to not support
+ _supports_static_cache = True
_supports_quantized_cache = True
_supports_sdpa = True
@@ -712,19 +713,15 @@ def forward(
image_newline=self.image_newline,
vision_aspect_ratio=vision_aspect_ratio,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (
- (input_ids == self.config.image_token_index)
- .unsqueeze(-1)
- .expand_as(inputs_embeds)
- .to(inputs_embeds.device)
- )
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
@@ -741,18 +738,14 @@ def forward(
video_features = torch.cat((video_features, image_newline), dim=1)
video_features = video_features.flatten(0, 1)
- n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
- n_video_features = video_features.shape[0]
- if n_video_tokens != n_video_features:
+ special_video_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
+ special_video_mask = special_video_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != video_features.numel():
+ n_video_tokens = (input_ids == self.config.video_token_index).sum()
+ n_video_features = video_features.shape[0]
raise ValueError(
f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
)
- special_video_mask = (
- (input_ids == self.config.video_token_index)
- .unsqueeze(-1)
- .expand_as(inputs_embeds)
- .to(inputs_embeds.device)
- )
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_video_mask, video_features)
diff --git a/src/transformers/models/opt/modeling_opt.py b/src/transformers/models/opt/modeling_opt.py
index 1969acf2f5b1..f1f1ef1821c7 100644
--- a/src/transformers/models/opt/modeling_opt.py
+++ b/src/transformers/models/opt/modeling_opt.py
@@ -22,10 +22,10 @@
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
+from ...cache_utils import Cache, DynamicCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import (
- _prepare_4d_causal_attention_mask,
- _prepare_4d_causal_attention_mask_for_sdpa,
+ AttentionMaskConverter,
)
from ...modeling_outputs import (
BaseModelOutputWithPast,
@@ -98,6 +98,7 @@ class OPTAttention(nn.Module):
def __init__(
self,
config: OPTConfig,
+ layer_idx: int = None,
**kwargs,
):
super().__init__()
@@ -106,6 +107,13 @@ def __init__(
self.num_heads = config.num_attention_heads
self.dropout = config.attention_dropout
self.enable_bias = config.enable_bias
+ self.layer_idx = layer_idx
+ if layer_idx is None:
+ logger.warning_once(
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+ "when creating this class."
+ )
self.head_dim = self.embed_dim // self.num_heads
self.is_causal = True
@@ -122,9 +130,6 @@ def __init__(
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=self.enable_bias)
- def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int) -> torch.Tensor:
- return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
-
def forward(
self,
hidden_states: torch.Tensor,
@@ -134,52 +139,33 @@ def forward(
output_attentions: bool = False,
# isn't needed in normal attention, but needed in flash attention so to keep the signature same
position_ids: Optional[torch.Tensor] = None,
- ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+ cache_position: Optional[torch.Tensor] = None,
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, _ = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
- # get key, value proj
- key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
- value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
- if past_key_value is not None:
- # reuse k, v, self_attention
- key_states = torch.cat([past_key_value[0], key_states], dim=2)
- value_states = torch.cat([past_key_value[1], value_states], dim=2)
-
- past_key_value = (key_states, value_states)
+ query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
- proj_shape = (bsz * self.num_heads, -1, self.head_dim)
- query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
- key_states = key_states.view(*proj_shape)
- value_states = value_states.view(*proj_shape)
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+ key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
- src_len = key_states.size(1)
- attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
-
- if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
- raise ValueError(
- f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
- f" {attn_weights.size()}"
+ if past_key_value is not None:
+ # save all key/value_states to cache to be re-used for fast auto-regressive generation
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, {"cache_position": cache_position}
)
+ attn_weights = torch.matmul(query_states, key_states.transpose(3, 2))
if attention_mask is not None:
- if attention_mask.size() != (bsz, 1, tgt_len, src_len):
- raise ValueError(
- f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
- )
- attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
- attn_weights = torch.max(
- attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min, device=attn_weights.device)
- )
- attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+ attn_weights = attn_weights + causal_mask
# upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437
- if attn_weights.dtype == torch.float16:
- attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(torch.float16)
- else:
- attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
if layer_head_mask is not None:
if layer_head_mask.size() != (self.num_heads,):
@@ -187,39 +173,19 @@ def forward(
f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
f" {layer_head_mask.size()}"
)
- attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
- attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
-
- if output_attentions:
- # this operation is a bit awkward, but it's required to
- # make sure that attn_weights keeps its gradient.
- # In order to do so, attn_weights have to be reshaped
- # twice and have to be reused in the following
- attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
- attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
- else:
- attn_weights_reshaped = None
+ attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+ attn_output = torch.matmul(attn_probs, value_states)
- attn_output = torch.bmm(attn_probs, value_states)
-
- if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
- raise ValueError(
- f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
- f" {attn_output.size()}"
- )
-
- attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
- attn_output = attn_output.transpose(1, 2)
+ attn_output = attn_output.transpose(1, 2).contiguous()
# Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
# partitioned aross GPUs when using tensor-parallelism.
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
-
attn_output = self.out_proj(attn_output)
- return attn_output, attn_weights_reshaped, past_key_value
+ return attn_output, attn_probs, past_key_value
class OptFlashAttention2(OPTAttention):
@@ -245,33 +211,33 @@ def forward(
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
position_ids: Optional[torch.Tensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
- bsz, _, _ = hidden_states.size()
- # get query proj
- query_states = self.q_proj(hidden_states)
- # get key, value proj
- key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
- value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
- if past_key_value is not None:
- # reuse k, v, self_attention
- key_states = torch.cat([past_key_value[0], key_states], dim=2)
- value_states = torch.cat([past_key_value[1], value_states], dim=2)
+ bsz, query_length, _ = hidden_states.size()
- past_key_value = (key_states, value_states)
+ query_states = self.q_proj(hidden_states)
+ query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim)
- query_length = query_states.shape[1]
- tgt_len = key_states.shape[-2]
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+ key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
- # Flash attention requires the input to have the shape
- # batch_size x seq_length x head_dim x hidden_dim
- query_states = query_states.view(bsz, query_length, self.num_heads, self.head_dim)
- key_states = key_states.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
- value_states = value_states.transpose(1, 2).view(bsz, tgt_len, self.num_heads, self.head_dim)
+ if past_key_value is not None:
+ # save all key/value_states to cache to be re-used for fast auto-regressive generation
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, {"cache_position": cache_position}
+ )
attn_dropout = self.dropout if self.training else 0.0
+ # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+ # to be able to avoid many of these transpose/reshape/view.
+ key_states = key_states.transpose(1, 2)
+ value_states = value_states.transpose(1, 2)
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in float16 just to be sure everything works as expected.
@@ -331,6 +297,7 @@ def forward(
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
position_ids: Optional[torch.Tensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if output_attentions or layer_head_mask is not None:
logger.warning_once(
@@ -344,24 +311,24 @@ def forward(
layer_head_mask=layer_head_mask,
past_key_value=past_key_value,
output_attentions=output_attentions,
- ) # TODO after merge add position_ids=position_ids
+ cache_position=cache_position,
+ )
bsz, q_len, _ = hidden_states.size()
- query_states = self.q_proj(hidden_states) * self.scaling
- query_states = self._shape(query_states, -1, bsz)
-
- # get key, value proj
- key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
- value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
- if past_key_value is not None:
- # reuse k, v, self_attention
- key_states = torch.cat([past_key_value[0], key_states], dim=2)
- value_states = torch.cat([past_key_value[1], value_states], dim=2)
+ query_states = self.q_proj(hidden_states)
+ query_states = query_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
- past_key_value = (key_states, value_states)
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+ key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2)
- # shape now is (bsz, num_heads, seq_len, head_dim), all are continuous
+ if past_key_value is not None:
+ # save all key/value_states to cache to be re-used for fast auto-regressive generation
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, {"cache_position": cache_position}
+ )
causal_mask = attention_mask
if attention_mask is not None:
@@ -378,10 +345,6 @@ def forward(
attn_mask=causal_mask,
dropout_p=self.dropout if self.training else 0.0,
is_causal=is_causal,
- # this model uses the scaling factor in the query projection for some reason, but not in Q@K^T
- # so we need to scale to remove scaling in SDPA to have similar results with eager.
- # Maybe needs a change in the model to remove scaling in query projection
- scale=1.0,
)
attn_output = attn_output.transpose(1, 2).contiguous()
@@ -399,11 +362,11 @@ def forward(
class OPTDecoderLayer(nn.Module):
- def __init__(self, config: OPTConfig):
+ def __init__(self, config: OPTConfig, layer_idx: int = None):
super().__init__()
self.embed_dim = config.hidden_size
- self.self_attn = OPT_ATTENTION_CLASSES[config._attn_implementation](config=config)
+ self.self_attn = OPT_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
self.do_layer_norm_before = config.do_layer_norm_before
self.dropout = config.dropout
@@ -425,6 +388,7 @@ def forward(
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = False,
position_ids: Optional[torch.LongTensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
"""
Args:
@@ -440,6 +404,8 @@ def forward(
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+ Indices depicting the position of the input sequence tokens in the sequence..
"""
residual = hidden_states
@@ -456,6 +422,7 @@ def forward(
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
+ cache_position=cache_position,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
@@ -524,6 +491,9 @@ class OPTPreTrainedModel(PreTrainedModel):
_no_split_modules = ["OPTDecoderLayer"]
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_cache_class = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
std = self.config.init_std
@@ -601,6 +571,10 @@ def _init_weights(self, module):
config.n_positions - 1]`. for padding use -1.
[What are position IDs?](../glossary#position-ids)
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+ Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+ this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+ the complete sequence length.
"""
@@ -643,9 +617,7 @@ def __init__(self, config: OPTConfig):
else:
self.final_layer_norm = None
- self.layers = nn.ModuleList([OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)])
- self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
- self._use_sdpa = config._attn_implementation == "sdpa"
+ self.layers = nn.ModuleList([OPTDecoderLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)])
self.gradient_checkpointing = False
# Initialize weights and apply final processing
@@ -657,48 +629,130 @@ def get_input_embeddings(self):
def set_input_embeddings(self, value):
self.embed_tokens = value
+ # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
def _update_causal_mask(
self,
- inputs_embeds: torch.Tensor,
- input_shape: Tuple[int, int],
- past_key_values_length: int,
- attention_mask: Optional[torch.Tensor] = None,
- head_mask: Optional[torch.Tensor] = None,
- output_attentions: Optional[bool] = None,
+ attention_mask: torch.Tensor,
+ input_tensor: torch.Tensor,
+ cache_position: torch.Tensor,
+ past_key_values: Cache,
+ output_attentions: bool,
):
- """
- Updates the causal mask for the decoder.
- """
- batch_size, seq_length = input_shape
- mask_seq_length = past_key_values_length + seq_length
- if self._use_flash_attention_2:
- # 2d mask is passed through the layers
- causal_attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
- attention_mask = (
- torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
- if attention_mask is None
- else attention_mask
+ if self.config._attn_implementation == "flash_attention_2":
+ if attention_mask is not None and (attention_mask == 0.0).any():
+ return attention_mask
+ return None
+
+ # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+ # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+ # to infer the attention mask.
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+ using_static_cache = isinstance(past_key_values, StaticCache)
+
+ # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+ if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+ if AttentionMaskConverter._ignore_causal_mask_sdpa(
+ attention_mask,
+ inputs_embeds=input_tensor,
+ past_key_values_length=past_seen_tokens,
+ is_training=self.training,
+ ):
+ return None
+
+ dtype, device = input_tensor.dtype, input_tensor.device
+ sequence_length = input_tensor.shape[1]
+ if using_static_cache:
+ target_length = past_key_values.get_max_cache_shape()
+ else:
+ target_length = (
+ attention_mask.shape[-1]
+ if isinstance(attention_mask, torch.Tensor)
+ else past_seen_tokens + sequence_length + 1
)
- return causal_attention_mask, attention_mask
+ # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+ causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask,
+ sequence_length=sequence_length,
+ target_length=target_length,
+ dtype=dtype,
+ device=device,
+ cache_position=cache_position,
+ batch_size=input_tensor.shape[0],
+ )
- if attention_mask is None:
- attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
- elif attention_mask.shape[1] != mask_seq_length:
- raise ValueError(
- f"The provided attention mask has length {attention_mask.shape[1]}, but its length should be "
- f"{mask_seq_length} (sum of the lengths of current and past inputs)"
- )
- if self._use_sdpa and not output_attentions and head_mask is None:
- causal_attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
- attention_mask, input_shape, inputs_embeds, past_key_values_length
- )
+ if (
+ self.config._attn_implementation == "sdpa"
+ and attention_mask is not None
+ and attention_mask.device.type in ["cuda", "xpu"]
+ and not output_attentions
+ ):
+ # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+ # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+ # Details: https://github.com/pytorch/pytorch/issues/110213
+ min_dtype = torch.finfo(dtype).min
+ causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+ return causal_mask
+
+ @staticmethod
+ # Copied from transformers.models.llama.modeling_llama.LlamaModel._prepare_4d_causal_attention_mask_with_cache_position
+ def _prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask: torch.Tensor,
+ sequence_length: int,
+ target_length: int,
+ dtype: torch.dtype,
+ device: torch.device,
+ cache_position: torch.Tensor,
+ batch_size: int,
+ **kwargs,
+ ):
+ """
+ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+ `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+ Args:
+ attention_mask (`torch.Tensor`):
+ A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+ `(batch_size, 1, query_length, key_value_length)`.
+ sequence_length (`int`):
+ The sequence length being processed.
+ target_length (`int`):
+ The target length: when generating with static cache, the mask should be as long as the static cache,
+ to account for the 0 padding, the part of the cache that is not filled yet.
+ dtype (`torch.dtype`):
+ The dtype to use for the 4D attention mask.
+ device (`torch.device`):
+ The device to plcae the 4D attention mask on.
+ cache_position (`torch.Tensor`):
+ Indices depicting the position of the input sequence tokens in the sequence.
+ batch_size (`torch.Tensor`):
+ Batch size.
+ """
+ if attention_mask is not None and attention_mask.dim() == 4:
+ # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+ causal_mask = attention_mask
else:
- causal_attention_mask = _prepare_4d_causal_attention_mask(
- attention_mask, input_shape, inputs_embeds, past_key_values_length
+ min_dtype = torch.finfo(dtype).min
+ causal_mask = torch.full(
+ (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
)
+ if sequence_length != 1:
+ causal_mask = torch.triu(causal_mask, diagonal=1)
+ causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+ causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+ if attention_mask is not None:
+ causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
+ mask_length = attention_mask.shape[-1]
+ padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+ causal_mask.device
+ )
+ padding_mask = padding_mask == 0
+ causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+ padding_mask, min_dtype
+ )
- return causal_attention_mask, attention_mask
+ return causal_mask
def forward(
self,
@@ -712,6 +766,7 @@ def forward(
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
position_ids: Optional[torch.LongTensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
r"""
Args:
@@ -764,6 +819,10 @@ def forward(
config.n_positions - 1]`. for padding use -1.
[What are position IDs?](../glossary#position-ids)
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+ Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+ this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+ the complete sequence length.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
@@ -773,51 +832,65 @@ def forward(
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
- # retrieve input_ids and inputs_embeds
- if input_ids is not None and inputs_embeds is not None:
- raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
- elif input_ids is not None:
- input_shape = input_ids.size()
- input_ids = input_ids.view(-1, input_shape[-1])
- elif inputs_embeds is not None:
- input_shape = inputs_embeds.size()[:-1]
- else:
- raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+ if (input_ids is None) ^ (inputs_embeds is not None):
+ raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+ if self.gradient_checkpointing and self.training and use_cache:
+ logger.warning_once(
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+ )
+ use_cache = False
+
+ if input_ids is not None:
+ input_ids = input_ids.view(-1, input_ids.shape[-1])
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
- past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+ return_legacy_cache = False
+ if use_cache and not isinstance(past_key_values, Cache):
+ return_legacy_cache = True
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+ if past_key_values is None:
+ logger.warning_once(
+ "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.53.0. "
+ "You should pass an instance of `DynamicCache` instead, e.g. "
+ "`past_key_values=DynamicCache.from_legacy_cache(past_key_values)`."
+ )
+
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+ if cache_position is None:
+ cache_position = torch.arange(
+ past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+ )
+
+ if attention_mask is None:
+ seq_length = past_seen_tokens + inputs_embeds.shape[1]
+ attention_mask = torch.ones(inputs_embeds.shape[0], seq_length, device=inputs_embeds.device)
- causal_attention_mask, attention_mask = self._update_causal_mask(
- inputs_embeds, input_shape, past_key_values_length, attention_mask, head_mask, output_attentions
+ causal_mask = self._update_causal_mask(
+ attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
)
- # embed positions
+ # embed positions
if position_ids is None:
+ # position_ids = cache_position.unsqueeze(0)
position_ids = torch.cumsum(attention_mask, dim=1)
position_ids = (position_ids * attention_mask - 1).long()
- # cut positions if `past_key_values_length` is > 0
- position_ids = position_ids[:, past_key_values_length:]
+ # cut positions if `past_seen_tokens` is > 0
+ position_ids = position_ids[:, past_seen_tokens:]
- pos_embeds = self.embed_positions(attention_mask, past_key_values_length, position_ids=position_ids)
+ pos_embeds = self.embed_positions(attention_mask, past_seen_tokens, position_ids=position_ids)
if self.project_in is not None:
inputs_embeds = self.project_in(inputs_embeds)
hidden_states = inputs_embeds + pos_embeds.to(inputs_embeds.device)
- if self.gradient_checkpointing and self.training:
- if use_cache:
- logger.warning_once(
- "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
- )
- use_cache = False
-
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
- next_decoder_cache = () if use_cache else None
+ next_decoder_cache = None
# check if head_mask has a correct number of layers specified if desired
for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
@@ -838,34 +911,34 @@ def forward(
if dropout_probability < self.layerdrop:
continue
- past_key_value = past_key_values[idx] if past_key_values is not None else None
-
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
- causal_attention_mask,
+ causal_mask,
head_mask[idx] if head_mask is not None else None,
None,
output_attentions,
use_cache,
position_ids,
+ cache_position,
)
else:
layer_outputs = decoder_layer(
hidden_states,
- attention_mask=causal_attention_mask,
+ attention_mask=causal_mask,
position_ids=position_ids,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
- past_key_value=past_key_value,
+ past_key_value=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
+ cache_position=cache_position,
)
hidden_states = layer_outputs[0]
if use_cache:
- next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
@@ -881,6 +954,9 @@ def forward(
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
+ if return_legacy_cache:
+ next_cache = next_cache.to_legacy_cache()
+
if not return_dict:
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
return BaseModelOutputWithPast(
@@ -930,6 +1006,7 @@ def forward(
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
position_ids: Optional[torch.LongTensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
@@ -950,6 +1027,7 @@ def forward(
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
+ cache_position=cache_position,
)
if not return_dict:
@@ -1008,6 +1086,7 @@ def forward(
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
position_ids: Optional[torch.LongTensor] = None,
+ cache_position: Optional[torch.Tensor] = None,
**kwargs,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""
@@ -1069,6 +1148,10 @@ def forward(
config.n_positions - 1]`. for padding use -1.
[What are position IDs?](../glossary#position-ids)
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+ Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+ this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+ the complete sequence length.
Returns:
@@ -1107,6 +1190,7 @@ def forward(
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
+ cache_position=cache_position,
)
logits = self.lm_head(outputs[0]).contiguous()
diff --git a/src/transformers/models/paligemma/modeling_paligemma.py b/src/transformers/models/paligemma/modeling_paligemma.py
index 9172b98c069e..35ad047a00dd 100644
--- a/src/transformers/models/paligemma/modeling_paligemma.py
+++ b/src/transformers/models/paligemma/modeling_paligemma.py
@@ -29,6 +29,7 @@
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_flash_attn_2_available,
+ is_torchdynamo_compiling,
logging,
replace_return_docstrings,
)
@@ -508,7 +509,7 @@ def forward(
special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
- if inputs_embeds[special_image_mask].numel() != image_features.numel():
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
image_tokens_in_text = torch.sum(input_ids == self.config.image_token_index)
raise ValueError(
f"Number of images does not match number of special image tokens in the input text. "
diff --git a/src/transformers/models/video_llava/configuration_video_llava.py b/src/transformers/models/video_llava/configuration_video_llava.py
index becd20040332..e761481d8259 100644
--- a/src/transformers/models/video_llava/configuration_video_llava.py
+++ b/src/transformers/models/video_llava/configuration_video_llava.py
@@ -38,8 +38,6 @@ class VideoLlavaConfig(PretrainedConfig):
text_config (`Union[AutoConfig, dict]`, *optional*):
The config object of the text backbone. Can be any of `LlamaConfig` or `MistralConfig`.
Defaults to `LlamaConfig` if not indicated.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32000):
The image token index to encode the image prompt.
video_token_index (`int`, *optional*, defaults to 32001):
@@ -88,7 +86,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32000,
video_token_index=32001,
projector_hidden_act="gelu",
@@ -99,7 +96,6 @@ def __init__(
multimodal_projector_bias=True,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.video_token_index = video_token_index
self.projector_hidden_act = projector_hidden_act
diff --git a/src/transformers/models/video_llava/modeling_video_llava.py b/src/transformers/models/video_llava/modeling_video_llava.py
index d8da974b9862..ba4de6537442 100644
--- a/src/transformers/models/video_llava/modeling_video_llava.py
+++ b/src/transformers/models/video_llava/modeling_video_llava.py
@@ -28,6 +28,7 @@
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
+ is_torchdynamo_compiling,
logging,
replace_return_docstrings,
)
@@ -137,6 +138,8 @@ class VideoLlavaPreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
std = (
@@ -276,92 +279,6 @@ def set_decoder(self, decoder):
def get_decoder(self):
return self.language_model.get_decoder()
- def _merge_input_ids_with_visual_features(
- self, visual_features, inputs_embeds, input_ids, attention_mask, labels, num_frames=1
- ):
- num_images, num_image_patches, embed_dim = visual_features.shape
- batch_size, sequence_length = input_ids.shape
- left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
- special_vision_token = self.config.video_token_index if num_frames > 1 else self.config.image_token_index
-
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == special_vision_token
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # Compute the maximum embed dimension
- max_seq_len = (num_special_image_tokens.max() * (num_image_patches * num_frames - 1)) + sequence_length
- batch_indices, non_image_indices = torch.where(input_ids != special_vision_token)
-
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- new_token_positions = (
- torch.cumsum((special_image_token_mask * (num_image_patches * num_frames - 1) + 1), dim=-1) - 1
- )
- nb_image_pad = max_seq_len - 1 - new_token_positions[:, -1]
- if left_padding:
- new_token_positions += nb_image_pad[:, None] # offset for left padding
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- # expand input ids so that the second "merge" with videos does not fail
- final_embedding = torch.zeros(
- batch_size, max_seq_len, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_seq_len, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- final_input_ids = torch.full(
- (batch_size, max_seq_len), self.pad_token_id, dtype=input_ids.dtype, device=inputs_embeds.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- final_input_ids[batch_indices, text_to_overwrite] = input_ids[batch_indices, non_image_indices]
- if labels is not None:
- final_labels = torch.full(
- (batch_size, max_seq_len), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
- )
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
- else:
- final_labels = None
-
- # 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling
- image_to_overwrite = torch.full((batch_size, max_seq_len), True, dtype=torch.bool, device=inputs_embeds.device)
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- if left_padding:
- image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
- else:
- mask = torch.ones_like(image_to_overwrite, dtype=torch.bool).cumsum(-1) - 1
- padding_mask = mask <= new_token_positions[:, -1:].to(target_device)
- image_to_overwrite &= padding_mask
-
- if image_to_overwrite.sum() != visual_features.shape[:-1].numel():
- visual_type = "videos" if num_frames == 8 else "images"
- num_images //= num_frames
- raise ValueError(
- f"The input provided to the model are wrong. The number of {visual_type} tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of {visual_type} given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
- )
-
- final_embedding[image_to_overwrite] = visual_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- return final_embedding, final_attention_mask, final_labels, position_ids, final_input_ids
-
def get_image_features(
self,
pixel_values_images: torch.FloatTensor,
@@ -579,14 +496,14 @@ def forward(
vision_feature_layer=vision_feature_layer,
vision_feature_select_strategy=vision_feature_select_strategy,
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0] * image_features.shape[1]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0] * image_features.shape[1]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
@@ -595,14 +512,14 @@ def forward(
pixel_values_videos=pixel_values_videos, vision_feature_layer=vision_feature_layer
)
- n_video_tokens = (input_ids == self.config.video_token_index).sum().item()
- n_video_features = video_features.shape[0] * video_features.shape[1]
- if n_video_tokens != n_video_features:
+ special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != video_features.numel():
+ n_video_tokens = (input_ids == self.config.video_token_index).sum()
+ n_video_features = video_features.shape[0] * video_features.shape[1]
raise ValueError(
f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
)
- special_image_mask = (input_ids == self.config.video_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, video_features)
diff --git a/src/transformers/models/vipllava/configuration_vipllava.py b/src/transformers/models/vipllava/configuration_vipllava.py
index 94d890c4b84e..ac24cce24129 100644
--- a/src/transformers/models/vipllava/configuration_vipllava.py
+++ b/src/transformers/models/vipllava/configuration_vipllava.py
@@ -37,8 +37,6 @@ class VipLlavaConfig(PretrainedConfig):
Custom vision config or dict
text_config (`Union[AutoConfig, dict]`, *optional*):
The config object of the text backbone. Can be any of `LlamaConfig` or `MistralConfig`.
- ignore_index (`int`, *optional*, defaults to -100):
- The ignore index for the loss function.
image_token_index (`int`, *optional*, defaults to 32000):
The image token index to encode the image prompt.
projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -78,7 +76,6 @@ def __init__(
self,
vision_config=None,
text_config=None,
- ignore_index=-100,
image_token_index=32000,
projector_hidden_act="gelu",
projector_layernorm_eps=1e-5,
@@ -86,7 +83,6 @@ def __init__(
image_seq_length=576,
**kwargs,
):
- self.ignore_index = ignore_index
self.image_token_index = image_token_index
self.projector_hidden_act = projector_hidden_act
self.projector_layernorm_eps = projector_layernorm_eps
diff --git a/src/transformers/models/vipllava/modeling_vipllava.py b/src/transformers/models/vipllava/modeling_vipllava.py
index 71201db2098e..ef4b3bff3958 100644
--- a/src/transformers/models/vipllava/modeling_vipllava.py
+++ b/src/transformers/models/vipllava/modeling_vipllava.py
@@ -28,6 +28,7 @@
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
+ is_torchdynamo_compiling,
logging,
replace_return_docstrings,
)
@@ -137,6 +138,8 @@ class VipLlavaPreTrainedModel(PreTrainedModel):
_supports_cache_class = True
_supports_flash_attn_2 = True
_supports_sdpa = True
+ _supports_quantized_cache = True
+ _supports_static_cache = True
def _init_weights(self, module):
# important: this ported version of VipLlava isn't meant for training from scratch - only
@@ -297,89 +300,6 @@ def get_image_features(self, pixel_values: torch.FloatTensor, vision_feature_lay
image_features = self.multi_modal_projector(image_features)
return image_features
- def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
- num_images, num_image_patches, embed_dim = image_features.shape
- batch_size, sequence_length = input_ids.shape
- left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
- # 1. Create a mask to know where special image tokens are
- special_image_token_mask = input_ids == self.config.image_token_index
- num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
- # Compute the maximum embed dimension
- max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
- batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
-
- # 2. Compute the positions where text should be written
- # Calculate new positions for text tokens in merged image-text sequence.
- # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
- # `torch.cumsum` computes how each image token shifts subsequent text token positions.
- # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
- new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
- nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
- if left_padding:
- new_token_positions += nb_image_pad[:, None] # offset for left padding
- text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
- # 3. Create the full embedding, already padded to the maximum position
- final_embedding = torch.zeros(
- batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
- )
- final_attention_mask = torch.zeros(
- batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
- )
- if labels is not None:
- final_labels = torch.full(
- (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
- )
- # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
- # set the corresponding tensors into their correct target device.
- target_device = inputs_embeds.device
- batch_indices, non_image_indices, text_to_overwrite = (
- batch_indices.to(target_device),
- non_image_indices.to(target_device),
- text_to_overwrite.to(target_device),
- )
- attention_mask = attention_mask.to(target_device)
-
- # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
- # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
- final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
- final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
- if labels is not None:
- final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
- # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
- image_to_overwrite = torch.full(
- (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
- )
- image_to_overwrite[batch_indices, text_to_overwrite] = False
- if left_padding:
- image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
- else:
- mask = torch.ones_like(image_to_overwrite, dtype=torch.bool).cumsum(-1) - 1
- padding_mask = mask <= new_token_positions[:, -1:].to(target_device)
- image_to_overwrite &= padding_mask
-
- if image_to_overwrite.sum() != image_features.shape[:-1].numel():
- raise ValueError(
- f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
- f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
- )
-
- final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
- final_attention_mask |= image_to_overwrite
- position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
- # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
- batch_indices, pad_indices = torch.where(input_ids == self.pad_token_id)
- indices_to_mask = new_token_positions[batch_indices, pad_indices]
-
- final_embedding[batch_indices, indices_to_mask] = 0
-
- if labels is None:
- final_labels = None
-
- return final_embedding, final_attention_mask, final_labels, position_ids
-
@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
@add_start_docstrings_to_model_forward(VIPLLAVA_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=VipLlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
@@ -469,14 +389,14 @@ def forward(
pixel_values=pixel_values, vision_feature_layers=vision_feature_layers
)
- n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
- n_image_features = image_features.shape[0] * image_features.shape[1]
- if n_image_tokens != n_image_features:
+ special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+ special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+ if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+ n_image_tokens = (input_ids == self.config.image_token_index).sum()
+ n_image_features = image_features.shape[0] * image_features.shape[1]
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
- special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
- special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
diff --git a/tests/generation/test_utils.py b/tests/generation/test_utils.py
index ce31cc844f19..3b9700dc20c9 100644
--- a/tests/generation/test_utils.py
+++ b/tests/generation/test_utils.py
@@ -1783,12 +1783,12 @@ def test_generate_from_inputs_embeds_with_static_cache(self):
model.config.use_cache = True
model.config.is_decoder = True
batch_size = input_ids.shape[0]
- max_length = 30
+ max_new_tokens = 10
# here we force to not stop at eos and go until max-length
model.generation_config.eos_token_id = model.config.get_text_config().eos_token_id = -1
generation_kwargs = {
- "max_length": max_length,
+ "max_new_tokens": max_new_tokens,
"cache_implementation": "static",
"return_dict_in_generate": True, # Required to return `past_key_values`
}
@@ -1811,10 +1811,11 @@ def test_generate_from_inputs_embeds_with_static_cache(self):
# we should get `max_length - 1` in shape, not `max_length - embeds_length`.
# -1 because the last generated token isn't yet in the cache.
- cache_shape = (batch_size, num_key_value_heads, max_length - 1, head_dim)
- self.assertTrue(isinstance(outputs.past_key_values, StaticCache))
- self.assertTrue(len(outputs.past_key_values.key_cache) == num_hidden_layers)
- self.assertTrue(outputs.past_key_values.key_cache[0].shape == cache_shape)
+ max_length = max_new_tokens + inputs_embeds.shape[1] - 1
+ cache_shape = [batch_size, num_key_value_heads, max_length, head_dim]
+ self.assertIsInstance(outputs.past_key_values, StaticCache)
+ self.assertEqual(len(outputs.past_key_values.key_cache), num_hidden_layers)
+ self.assertListEqual(list(outputs.past_key_values.key_cache[0].shape), cache_shape)
@pytest.mark.generate
def test_generate_continue_from_past_key_values(self):
@@ -2022,7 +2023,7 @@ def test_generate_with_static_cache(self):
config.is_decoder = True
batch_size = main_input.shape[0]
- seq_length = main_input.shape[-1]
+ seq_length = self.model_tester.seq_length
max_new_tokens = 20
for dtype in (torch.float32, torch.float16):
@@ -2134,7 +2135,15 @@ def test_generate_compile_model_forward(self):
# compilation-specific setup
torch.compiler.reset() # prevent cached compilation from being used in the test
has_defined_cache_implementation = model.generation_config.cache_implementation is not None
- model.generation_config.compile_config._compile_all_devices = True # force compilation (e.g. fast CI, CPU)
+
+ # BLIP is the only exception with custom generate which call `self.lm.generate()`
+ # We should avoid such calls in all subsequent multimodal models and try to make `generate()`
+ # compatible with multimodality
+ if "blip" in model.__class__.__name__.lower():
+ model.language_model.generation_config.compile_config._compile_all_devices = True
+ else:
+ # force compilation (e.g. fast CI, CPU
+ model.generation_config.compile_config._compile_all_devices = True
generation_kwargs = {
"do_sample": False,
@@ -2175,7 +2184,14 @@ def test_generate_compile_model_forward(self):
)
self.assertFalse(isinstance(decoder_cache, DynamicCache))
self.assertTrue(decoder_cache.is_compileable)
- self.assertTrue(hasattr(model, "_compiled_call")) # our auto compile should have been called
+
+ # BLIP is the only exception with custom generate which call `self.lm.generate()`
+ # We should avoid such calls in all subsequent multimodal models and try to make `generate()`
+ # compatible with multimodality
+ if "blip" in model.__class__.__name__.lower():
+ self.assertTrue(hasattr(model.language_model, "_compiled_call"))
+ else:
+ self.assertTrue(hasattr(model, "_compiled_call")) # our auto compile should have been called
for dynamic_result, compiled_result in zip(dynamic_outputs, compiled_outputs):
self._check_similar_generate_outputs(dynamic_result, compiled_result)
@@ -2198,9 +2214,19 @@ def test_generate_compilation_all_outputs(self):
# compilation-specific setup
torch.compiler.reset() # prevent cached compilation from being used in the test
has_defined_cache_implementation = model.generation_config.cache_implementation is not None
- model.generation_config.compile_config._compile_all_devices = True # force compilation (e.g. fast CI, CPU)
- if not has_defined_cache_implementation:
- model.generation_config.cache_implementation = "static"
+
+ # BLIP is the only exception with custom generate which call `self.lm.generate()`
+ # We should avoid such calls in all subsequent multimodal models and try to make `generate()`
+ # compatible with multimodality
+ if "blip" in model.__class__.__name__.lower():
+ model.language_model.generation_config.compile_config._compile_all_devices = True
+ if not has_defined_cache_implementation:
+ model.language_model.generation_config.cache_implementation = "static"
+ else:
+ # force compilation (e.g. fast CI, CPU)
+ model.generation_config.compile_config._compile_all_devices = True
+ if not has_defined_cache_implementation:
+ model.generation_config.cache_implementation = "static"
logits_processor_kwargs = self._get_logits_processor_kwargs(do_sample=False, config=model.config)
output_generate = model.generate(
@@ -2218,8 +2244,10 @@ def test_generate_compilation_all_outputs(self):
**inputs_dict,
)
- # Sanity check: compilation has happened
- self.assertTrue(hasattr(model, "_compiled_call"))
+ if "blip" in model.__class__.__name__.lower():
+ self.assertTrue(hasattr(model.language_model, "_compiled_call"))
+ else:
+ self.assertTrue(hasattr(model, "_compiled_call")) # our auto compile should have been called
if model.config.is_encoder_decoder:
self.assertTrue(output_generate.sequences.shape[-1] == self.max_new_tokens + 1)
diff --git a/tests/models/aria/test_modeling_aria.py b/tests/models/aria/test_modeling_aria.py
index f12ff24b17f1..8b5e62de14c7 100644
--- a/tests/models/aria/test_modeling_aria.py
+++ b/tests/models/aria/test_modeling_aria.py
@@ -286,10 +286,18 @@ def test_generate_from_inputs_embeds_0_greedy(self):
def test_generate_from_inputs_embeds_1_beam_search(self):
pass
- @unittest.skip(reason="Unsupported")
+ @unittest.skip(reason="Dynamic control flow due to MoE")
def test_generate_with_static_cache(self):
pass
+ @unittest.skip(reason="Dynamic control flow due to MoE")
+ def test_generate_from_inputs_embeds_with_static_cache(self):
+ pass
+
+ @unittest.skip(reason="Dynamic control flow due to MoE")
+ def test_generate_compile_model_forward(self):
+ pass
+
@require_torch
class AriaForConditionalGenerationIntegrationTest(unittest.TestCase):
diff --git a/tests/models/blip_2/test_modeling_blip_2.py b/tests/models/blip_2/test_modeling_blip_2.py
index e26232e3eb43..a405a1f97fb3 100644
--- a/tests/models/blip_2/test_modeling_blip_2.py
+++ b/tests/models/blip_2/test_modeling_blip_2.py
@@ -816,6 +816,10 @@ def _prepare_model_kwargs(input_ids, attention_mask, signature):
def test_generate_from_inputs_embeds(self, _, num_beams):
pass
+ @unittest.skip("BLIP2 cannot generate only from input ids, and requires pixel values in all cases to be present")
+ def test_generate_from_inputs_embeds_with_static_cache(self):
+ pass
+
# this class is based on `T5ModelTester` found in tests/models/t5/test_modeling_t5.py
class Blip2TextModelTester:
diff --git a/tests/models/emu3/test_modeling_emu3.py b/tests/models/emu3/test_modeling_emu3.py
index 4563cc17dfce..491fd9f9ec4f 100644
--- a/tests/models/emu3/test_modeling_emu3.py
+++ b/tests/models/emu3/test_modeling_emu3.py
@@ -386,10 +386,6 @@ def test_disk_offload_bin(self):
def test_cpu_offload(self):
pass
- @unittest.skip("Doesn't work, tensors are not almost same") # TODO raushan fixme
- def test_custom_4d_attention_mask(self):
- pass
-
@unittest.skip("VQ-VAE module doesn't initialize weights properly")
def test_initialization(self):
pass
diff --git a/tests/models/got_ocr2/test_modeling_got_ocr2.py b/tests/models/got_ocr2/test_modeling_got_ocr2.py
index ac044de5ca96..178bec98ac62 100644
--- a/tests/models/got_ocr2/test_modeling_got_ocr2.py
+++ b/tests/models/got_ocr2/test_modeling_got_ocr2.py
@@ -256,12 +256,6 @@ def test_generate_from_inputs_embeds_with_static_cache(self):
def test_past_key_values_format(self):
pass
- @unittest.skip(
- reason="GotOcr2 needs a dynamic control flow to pass pixel values to the forward function only in the first generation step"
- )
- def test_generate_compile_1_end_to_end(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
diff --git a/tests/models/idefics/test_modeling_idefics.py b/tests/models/idefics/test_modeling_idefics.py
index 5d19f5b02025..32c45d6e71f7 100644
--- a/tests/models/idefics/test_modeling_idefics.py
+++ b/tests/models/idefics/test_modeling_idefics.py
@@ -838,6 +838,14 @@ def test_contrastive_generate_low_memory(self):
def test_custom_4d_attention_mask(self):
pass
+ @unittest.skip(reason="IDEFICS cannot compile due to dynamic control flow when checking inputs")
+ def test_generate_with_static_cache(self):
+ pass
+
+ @unittest.skip(reason="IDEFICS cannot compile due to dynamic control flow when checking inputs")
+ def test_generate_compile_model_forward(self):
+ pass
+
@unittest.skip(reason="We only test the model that takes in multiple images")
def test_model(self):
pass
diff --git a/tests/models/instructblip/test_modeling_instructblip.py b/tests/models/instructblip/test_modeling_instructblip.py
index e072499ad3f1..bbf877289040 100644
--- a/tests/models/instructblip/test_modeling_instructblip.py
+++ b/tests/models/instructblip/test_modeling_instructblip.py
@@ -530,6 +530,12 @@ def test_save_load_fast_init_from_base(self):
def test_save_load_fast_init_to_base(self):
pass
+ @unittest.skip(
+ "InstructBLIP cannot generate only from input ids, and requires pixel values in all cases to be present"
+ )
+ def test_generate_from_inputs_embeds_with_static_cache(self):
+ pass
+
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
diff --git a/tests/models/instructblipvideo/test_modeling_instructblipvideo.py b/tests/models/instructblipvideo/test_modeling_instructblipvideo.py
index 0534b4f5ea73..351dea3d6fae 100644
--- a/tests/models/instructblipvideo/test_modeling_instructblipvideo.py
+++ b/tests/models/instructblipvideo/test_modeling_instructblipvideo.py
@@ -546,6 +546,12 @@ def test_save_load_fast_init_from_base(self):
def test_save_load_fast_init_to_base(self):
pass
+ @unittest.skip(
+ "InstructBLIPVideo cannot generate only from input ids, and requires pixel values in all cases to be present"
+ )
+ def test_generate_from_inputs_embeds_with_static_cache(self):
+ pass
+
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
diff --git a/tests/models/llava/test_modeling_llava.py b/tests/models/llava/test_modeling_llava.py
index 25e1a747ce9f..b47423a02ec7 100644
--- a/tests/models/llava/test_modeling_llava.py
+++ b/tests/models/llava/test_modeling_llava.py
@@ -316,14 +316,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
- @unittest.skip(reason="Compile not yet supported because in LLava models")
- def test_sdpa_can_compile_dynamic(self):
- pass
-
- @unittest.skip(reason="Compile not yet supported because in LLava models")
- def test_sdpa_can_dispatch_on_flash(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
diff --git a/tests/models/llava_next/test_modeling_llava_next.py b/tests/models/llava_next/test_modeling_llava_next.py
index eaeda3cecb7b..0c75df53c1bb 100644
--- a/tests/models/llava_next/test_modeling_llava_next.py
+++ b/tests/models/llava_next/test_modeling_llava_next.py
@@ -365,22 +365,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
- @unittest.skip(reason="Feedforward chunking is not yet supported")
- def test_feed_forward_chunking(self):
- pass
-
- @unittest.skip(reason="CPU offload is not yet supported")
- def test_cpu_offload(self):
- pass
-
- @unittest.skip(reason="Compile not yet supported because in LLava models")
- def test_sdpa_can_compile_dynamic(self):
- pass
-
- @unittest.skip(reason="Compile not yet supported because in LLava models")
- def test_sdpa_can_dispatch_on_flash(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
@@ -391,6 +375,10 @@ def test_flash_attn_2_fp32_ln(self):
def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self):
pass
+ @unittest.skip("LLaVA Next has dynamic control flow in unpadding")
+ def test_generate_compile_model_forward(self):
+ pass
+
@require_torch
class LlavaNextForConditionalGenerationIntegrationTest(unittest.TestCase):
diff --git a/tests/models/llava_next_video/test_modeling_llava_next_video.py b/tests/models/llava_next_video/test_modeling_llava_next_video.py
index 0f4642402644..6d4df92f5c22 100644
--- a/tests/models/llava_next_video/test_modeling_llava_next_video.py
+++ b/tests/models/llava_next_video/test_modeling_llava_next_video.py
@@ -382,26 +382,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
- @unittest.skip(reason="Feedforward chunking is not yet supported")
- def test_feed_forward_chunking(self):
- pass
-
- @unittest.skip(reason="CPU offload is not yet supported")
- def test_cpu_offload(self):
- pass
-
- @unittest.skip(
- reason="Compile not yet supported because in LLava models (https://github.com/huggingface/transformers/issues/29891)"
- )
- def test_sdpa_can_compile_dynamic(self):
- pass
-
- @unittest.skip(
- reason="Compile not yet supported because in LLava models (https://github.com/huggingface/transformers/issues/29891)"
- )
- def test_sdpa_can_dispatch_on_flash(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
@@ -412,6 +392,10 @@ def test_flash_attn_2_fp32_ln(self):
def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self):
pass
+ @unittest.skip("LLaVA Next Video has dynamic control flow in unpadding")
+ def test_generate_compile_model_forward(self):
+ pass
+
@require_torch
class LlavaNextVideoForConditionalGenerationIntegrationTest(unittest.TestCase):
diff --git a/tests/models/llava_onevision/test_modeling_llava_onevision.py b/tests/models/llava_onevision/test_modeling_llava_onevision.py
index 63be10a774db..c9bb448278e7 100644
--- a/tests/models/llava_onevision/test_modeling_llava_onevision.py
+++ b/tests/models/llava_onevision/test_modeling_llava_onevision.py
@@ -346,6 +346,10 @@ def test_flash_attn_2_fp32_ln(self):
def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self):
pass
+ @unittest.skip("LLaVA OneVision has dynamic control flow in unpadding")
+ def test_generate_compile_model_forward(self):
+ pass
+
@require_torch
class LlavaOnevisionForConditionalGenerationIntegrationTest(unittest.TestCase):
diff --git a/tests/models/mt5/test_modeling_mt5.py b/tests/models/mt5/test_modeling_mt5.py
index 3c3256da8b24..994d88444809 100644
--- a/tests/models/mt5/test_modeling_mt5.py
+++ b/tests/models/mt5/test_modeling_mt5.py
@@ -540,7 +540,6 @@ def prepare_config_and_inputs_for_common(self):
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
- "use_cache": False,
}
return config, inputs_dict
diff --git a/tests/models/opt/test_modeling_opt.py b/tests/models/opt/test_modeling_opt.py
index 3e3d2159a022..dad740cde721 100644
--- a/tests/models/opt/test_modeling_opt.py
+++ b/tests/models/opt/test_modeling_opt.py
@@ -81,7 +81,7 @@ def __init__(
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
- max_position_embeddings=20,
+ max_position_embeddings=50,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
@@ -89,7 +89,6 @@ def __init__(
num_labels=3,
word_embed_proj_dim=16,
type_sequence_label_size=2,
- attn_implementation="eager",
):
self.parent = parent
self.batch_size = batch_size
@@ -113,7 +112,6 @@ def __init__(
self.type_sequence_label_size = type_sequence_label_size
self.word_embed_proj_dim = word_embed_proj_dim
self.is_encoder_decoder = False
- self.attn_implementation = attn_implementation
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
@@ -143,7 +141,6 @@ def get_config(self):
embed_dim=self.embed_dim,
is_encoder_decoder=False,
word_embed_proj_dim=self.word_embed_proj_dim,
- attn_implementation=self.attn_implementation,
)
def get_pipeline_config(self):
diff --git a/tests/models/t5/test_modeling_t5.py b/tests/models/t5/test_modeling_t5.py
index 9886684d6088..a0439550f8f0 100644
--- a/tests/models/t5/test_modeling_t5.py
+++ b/tests/models/t5/test_modeling_t5.py
@@ -545,7 +545,6 @@ def prepare_config_and_inputs_for_common(self):
"attention_mask": attention_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
- "use_cache": False,
}
return config, inputs_dict
diff --git a/tests/models/video_llava/test_modeling_video_llava.py b/tests/models/video_llava/test_modeling_video_llava.py
index b8d4d4167e57..528f125693f7 100644
--- a/tests/models/video_llava/test_modeling_video_llava.py
+++ b/tests/models/video_llava/test_modeling_video_llava.py
@@ -226,14 +226,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
- @unittest.skip(reason="Pass because video-LLava requires `attention_mask is not None`")
- def test_sdpa_can_compile_dynamic(self):
- pass
-
- @unittest.skip(reason="Pass because video-LLava requires `attention_mask is not None`")
- def test_sdpa_can_dispatch_on_flash(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
diff --git a/tests/models/vipllava/test_modeling_vipllava.py b/tests/models/vipllava/test_modeling_vipllava.py
index f6a601c8a02d..24f99d4b0b18 100644
--- a/tests/models/vipllava/test_modeling_vipllava.py
+++ b/tests/models/vipllava/test_modeling_vipllava.py
@@ -306,14 +306,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
- @unittest.skip(reason="Compile not yet supported because it is not yet supported in LLava")
- def test_sdpa_can_compile_dynamic(self):
- pass
-
- @unittest.skip(reason="Compile not yet supported because in LLava models")
- def test_sdpa_can_dispatch_on_flash(self):
- pass
-
@unittest.skip("FlashAttention only support fp16 and bf16 data type")
def test_flash_attn_2_fp32_ln(self):
pass
diff --git a/tests/test_modeling_common.py b/tests/test_modeling_common.py
index 9dd5877c8b90..a707b25a3110 100755
--- a/tests/test_modeling_common.py
+++ b/tests/test_modeling_common.py
@@ -4324,10 +4324,6 @@ def test_sdpa_can_dispatch_on_flash(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
inputs_dict = self._prepare_for_class(inputs_dict, model_class)
- if config.model_type in ["llava", "llava_next", "vipllava", "video_llava"]:
- self.skipTest(
- reason="Llava-like models currently (transformers==4.39.1) requires an attention_mask input"
- )
if config.model_type in ["paligemma"]:
self.skipTest(
"PaliGemma-like models currently (transformers==4.41.0) requires an attention_mask input"
@@ -4778,6 +4774,9 @@ def test_custom_4d_attention_mask(self):
model = model_class(config).to(device=torch_device, dtype=torch.float32)
set_model_for_less_flaky_test(model)
+ if "position_ids" not in inspect.signature(model.forward).parameters:
+ continue # this model doesn't accept position ids as input
+
(
input_ids,
position_ids,
|