File size: 286,697 Bytes
00d4f8c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 | diff --git a/common/arg.cpp b/common/arg.cpp
index 8b302d5d7..49354d98b 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -3096,11 +3096,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("HF_TOKEN"));
add_opt(common_arg(
{"--mtp"},
- "also download the multi-token prediction (MTP) head, if available (default: unused)",
+ "enable multi-token prediction (MTP) self-speculative decoding, and download the MTP head if available (default: unused)",
[](common_params & params) {
params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_MTP);
}
- ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
+ ).set_examples({LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}));
add_opt(common_arg(
{"--dflash"},
"also download the DFlash sidecar, if available (default: unused)",
diff --git a/common/chat.cpp b/common/chat.cpp
index 485671308..ddf9e3ac9 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -906,9 +906,10 @@ common_chat_prompt common_chat_template_direct_apply_impl(
inp[k] = v;
}
}
- if (inputs.add_generation_prompt) {
- inp["add_generation_prompt"] = true;
- }
+ // always set explicitly: templates that default an undefined add_generation_prompt
+ // to true (e.g. openPangu) would otherwise render identical output for both the
+ // with- and without-generation-prompt passes, yielding an empty generation prompt
+ inp["add_generation_prompt"] = inputs.add_generation_prompt;
if (inp.contains("preserve_reasoning") && inp["preserve_reasoning"].is_boolean()) {
bool enabled = inp["preserve_reasoning"].get<bool>();
jinja::caps_apply_preserve_reasoning(ctx, enabled);
@@ -1163,6 +1164,12 @@ static common_chat_params_init_fn common_chat_template_detect_params_init(const
return common_chat_params_init_cohere2moe;
}
+ // openPangu - <|message_start|> turns, <|tool_call_start|> JSON tool calls and thinking reasoning
+ if (is_openpangu_template(src)) {
+ LOG_DBG("Using specialized template: openPangu\n");
+ return common_chat_params_init_openpangu;
+ }
+
if (is_lfm2_template(src)) {
LOG_DBG("Using specialized template: LFM2\n");
return common_chat_params_init_lfm2_tokens;
diff --git a/common/parsers/openpangu.cpp b/common/parsers/openpangu.cpp
new file mode 100644
index 000000000..adc5f50ce
--- /dev/null
+++ b/common/parsers/openpangu.cpp
@@ -0,0 +1,96 @@
+#include "parsers.h"
+
+// openPangu format detection: <|message_start|>{role} turns with
+// <|tool_call_start|>[{"name":..,"arguments":..}]<|tool_call_end|> tool calls and thinking reasoning
+bool is_openpangu_template(const std::string & src) {
+ return src.find("<|tool_call_start|>") != std::string::npos &&
+ src.find("<|message_start|>") != std::string::npos;
+}
+
+common_chat_params common_chat_params_init_openpangu(const common_chat_template & tmpl,
+ const autoparser::generation_params & inputs) {
+ common_chat_params data;
+
+ const std::string TOOL_CALL_START = "<|tool_call_start|>";
+ const std::string TOOL_CALL_END = "<|tool_call_end|>";
+ const std::string THINK_START = " thinking";
+ const std::string THINK_END = "</think>";
+ const std::string GEN_PROMPT = "<|message_start|>assistant\n";
+
+ // the template reads a 'thinking' flag (default true): thinking on force-opens the
+ // think block in the generation prompt ('...assistant\n thinking'), thinking off closes
+ // it immediately ('...assistant\n</think>')
+ const json template_context = { {"thinking", inputs.enable_thinking} };
+
+ data.prompt = common_chat_template_direct_apply_impl(tmpl, inputs, std::nullopt, std::nullopt, template_context);
+ data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, std::nullopt, std::nullopt, template_context);
+ data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
+ data.supports_thinking = true;
+ data.thinking_start_tag = THINK_START;
+ data.thinking_end_tags = { THINK_END };
+ data.preserved_tokens = { TOOL_CALL_START, TOOL_CALL_END, THINK_START, THINK_END };
+
+ auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+ auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+ auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE &&
+ tmpl.source().find(THINK_START) != std::string::npos;
+ auto include_grammar = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+ if (inputs.has_continuation()) {
+ const auto & msg = inputs.continue_msg;
+
+ data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+ if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+ data.generation_prompt += THINK_END + msg.render_content();
+ }
+
+ data.prompt.append(data.generation_prompt);
+ }
+
+ data.parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+ // the parser input is generation_prompt + completion when the caller plumbs the
+ // generation prompt through (server), and just the completion otherwise (cli) --
+ // so the assistant header is optional
+ auto generation_prompt = p.optional(p.literal(GEN_PROMPT));
+ auto end = p.end();
+
+ auto reasoning = p.eps();
+ if (extract_reasoning) {
+ // the template force-opens the think block: ' thinking' is only seen when the
+ // generation prompt is prepended, so the opening tag is optional; with thinking
+ // disabled the template emits a bare '</think>', parsed as empty reasoning
+ reasoning = p.optional(p.optional(p.literal(THINK_START)) + p.reasoning(p.until(THINK_END)) + p.literal(THINK_END));
+ }
+
+ if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+ if (has_response_format) {
+ auto response_format = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
+ return generation_prompt + reasoning + response_format + end;
+ }
+ return generation_prompt + reasoning + p.content(p.rest()) + end;
+ }
+
+ auto tool_calls = p.standard_json_tools(TOOL_CALL_START, TOOL_CALL_END, inputs.tools, inputs.parallel_tool_calls,
+ /* force_tool_calls = */ false,
+ /* name_key = */ "name",
+ /* args_key = */ "arguments",
+ /* array_wrapped = */ true);
+
+ auto content = p.content(p.until(TOOL_CALL_START));
+
+ return generation_prompt + reasoning + content + p.optional(tool_calls) + end;
+ });
+
+ if (include_grammar) {
+ data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+ data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+ data.parser.build_grammar(builder, data.grammar_lazy);
+ });
+
+ data.grammar_triggers = {
+ { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOL_CALL_START }
+ };
+ }
+
+ return data;
+}
diff --git a/common/parsers/parsers.h b/common/parsers/parsers.h
index 8f0a97ce3..5eaf8e8e9 100644
--- a/common/parsers/parsers.h
+++ b/common/parsers/parsers.h
@@ -41,6 +41,8 @@ std::string common_chat_template_generation_prompt_impl(
bool is_lfm2_template(const std::string & src);
+bool is_openpangu_template(const std::string & src);
+
namespace workaround {
void convert_tool_responses_gemma4(json & messages);
@@ -80,6 +82,8 @@ common_chat_params common_chat_params_init_ministral_3(const common_chat_templat
common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+common_chat_params common_chat_params_init_openpangu(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
common_chat_params common_chat_params_init_translate_gemma(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
diff --git a/common/parsers/sources.cmake b/common/parsers/sources.cmake
index 152397074..ca4b11654 100644
--- a/common/parsers/sources.cmake
+++ b/common/parsers/sources.cmake
@@ -19,6 +19,7 @@ set(LLAMA_CHAT_PARSERS_SOURCES
${CMAKE_CURRENT_LIST_DIR}/minimax-m3.cpp
${CMAKE_CURRENT_LIST_DIR}/ministral3.cpp
${CMAKE_CURRENT_LIST_DIR}/muse-glimmer.cpp
+ ${CMAKE_CURRENT_LIST_DIR}/openpangu.cpp
${CMAKE_CURRENT_LIST_DIR}/qwen3-coder.cpp
${CMAKE_CURRENT_LIST_DIR}/translate-gemma.cpp
)
diff --git a/conversion/__init__.py b/conversion/__init__.py
index 8ce4cb733..4302cc1a4 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -228,6 +228,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
"PLaMo3ForCausalLM": "plamo",
"PaddleOCRVLForConditionalGeneration": "ernie",
"PanguEmbeddedForCausalLM": "pangu",
+ "OpenPanguV2ForCausalLM": "pangu",
"Phi3ForCausalLM": "phi",
"Phi4ForCausalLMV": "phi",
"PhiForCausalLM": "phi",
diff --git a/conversion/base.py b/conversion/base.py
index a5a80f27d..c5b389fb8 100644
--- a/conversion/base.py
+++ b/conversion/base.py
@@ -2010,6 +2010,9 @@ class TextModel(ModelBase):
if chkhsh == "877081d19cf6996e2c4ff0e1236341e9b7bde288f5311a56a937f0afbbb3aeb5":
# ref: https://huggingface.co/deepseek-ai/DeepSeek-V3
res = "deepseek-v3"
+ if chkhsh == "65df2fe396b537a53433301848c0a739f56d56f67ad3d35eba27961ac33c12bb":
+ # ref: https://huggingface.co/openpangu/openPangu-2.0-Flash
+ res = "openpangu-v2"
if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5":
# ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
res = "deepseek-r1-qwen"
diff --git a/conversion/pangu.py b/conversion/pangu.py
index 74c76532b..ec2301459 100644
--- a/conversion/pangu.py
+++ b/conversion/pangu.py
@@ -4,6 +4,8 @@ import json
from typing import Iterable, TYPE_CHECKING
+import torch
+
if TYPE_CHECKING:
from torch import Tensor
@@ -45,3 +47,228 @@ class PanguEmbeddedModel(TextModel):
logger.info("Skipping tied output layer 'lm_head.weight'")
return
yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("OpenPanguV2ForCausalLM")
+class OpenPanguV2Model(TextModel):
+ model_arch = gguf.MODEL_ARCH.OPENPANGU_V2
+
+ # simple root tensor map (HF name -> gguf enum)
+ _root_map = {
+ "model.embed_tokens.weight": (gguf.MODEL_TENSOR.TOKEN_EMBD, ".weight"),
+ "model.norm.weight": (gguf.MODEL_TENSOR.OUTPUT_NORM, ".weight"),
+ "lm_head.weight": (gguf.MODEL_TENSOR.OUTPUT, ".weight"),
+ "model.merge_mhc_module.branch_alpha_pre": (gguf.MODEL_TENSOR.HC_HEAD_SCALE, ".weight"),
+ "model.merge_mhc_module.branch_beta_pre": (gguf.MODEL_TENSOR.HC_HEAD_BASE, ".weight"),
+ }
+
+ # per-layer tensor map (suffix after "model.layers.N." -> gguf enum)
+ _layer_map = {
+ "input_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_NORM, ".weight"),
+ "post_attention_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_POST_NORM, ".weight"),
+ "pre_mlp_layernorm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
+ "post_mlp_layernorm.weight": (gguf.MODEL_TENSOR.FFN_POST_NORM, ".weight"),
+ "block_post_layernorm.weight": (gguf.MODEL_TENSOR.BLOCK_POST_NORM, ".weight"),
+ "attn_mhc_module.branch_alpha": (gguf.MODEL_TENSOR.HC_ATTN_SCALE, ".weight"),
+ "attn_mhc_module.branch_beta": (gguf.MODEL_TENSOR.HC_ATTN_BASE, ".weight"),
+ "mlp_mhc_module.branch_alpha": (gguf.MODEL_TENSOR.HC_FFN_SCALE, ".weight"),
+ "mlp_mhc_module.branch_beta": (gguf.MODEL_TENSOR.HC_FFN_BASE, ".weight"),
+ "self_attn.q_a_proj.weight": (gguf.MODEL_TENSOR.ATTN_Q_A, ".weight"),
+ "self_attn.q_a_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_Q_A_NORM, ".weight"),
+ "self_attn.q_b_proj.weight": (gguf.MODEL_TENSOR.ATTN_Q_B, ".weight"),
+ "self_attn.qa_conv.weight": (gguf.MODEL_TENSOR.ATTN_Q_A_CONV, ".weight"),
+ "self_attn.kv_a_proj_with_mqa.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_MQA, ".weight"),
+ "self_attn.kv_a_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_NORM, ".weight"),
+ "self_attn.compresskv_conv.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_CONV, ".weight"),
+ "self_attn.o_conv.weight": (gguf.MODEL_TENSOR.ATTN_O_CONV, ".weight"),
+ "self_attn.o_proj.weight": (gguf.MODEL_TENSOR.ATTN_OUT, ".weight"),
+ "self_attn.param_sink_compressed_kv": (gguf.MODEL_TENSOR.ATTN_SINK_KV, ".weight"),
+ "self_attn.param_sink_k_pe": (gguf.MODEL_TENSOR.ATTN_SINK_K_PE, ".weight"),
+ "self_attn.indexer.wq_b.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_Q_B, ".weight"),
+ "self_attn.indexer.wk.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_K, ".weight"),
+ "self_attn.indexer.k_norm.weight": (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".weight"),
+ "self_attn.indexer.weights_proj.weight": (gguf.MODEL_TENSOR.INDEXER_PROJ, ".weight"),
+ "mlp.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
+ "mlp.e_score_correction.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
+ "mlp.shared_experts.gate_proj.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
+ "mlp.shared_experts.up_proj.weight": (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"),
+ "mlp.shared_experts.down_proj.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
+ "mlp.gate_proj.weight": (gguf.MODEL_TENSOR.FFN_GATE, ".weight"),
+ "mlp.up_proj.weight": (gguf.MODEL_TENSOR.FFN_UP, ".weight"),
+ "mlp.down_proj.weight": (gguf.MODEL_TENSOR.FFN_DOWN, ".weight"),
+ # MTP / NextN heads (layers >= num_hidden_layers)
+ "enorm.weight": (gguf.MODEL_TENSOR.NEXTN_ENORM, ".weight"),
+ "hnorm.weight": (gguf.MODEL_TENSOR.NEXTN_HNORM, ".weight"),
+ "eh_proj.weight": (gguf.MODEL_TENSOR.NEXTN_EH_PROJ, ".weight"),
+ "embed_tokens.weight": (gguf.MODEL_TENSOR.NEXTN_EMBED_TOKENS, ".weight"),
+ "shared_head.norm.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM, ".weight"),
+ "shared_head.head.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, ".weight"),
+ }
+
+ _expert_map = {
+ "gate_proj": gguf.MODEL_TENSOR.FFN_GATE_EXP,
+ "up_proj": gguf.MODEL_TENSOR.FFN_UP_EXP,
+ "down_proj": gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+ }
+
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+ # buffers for mHC gamma-fold and expert stacking
+ self._mhc_buf: dict = {}
+ self._exp_buf: dict = {}
+ # include the MTP / NextN prediction layers in the block count
+ self._n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0) or 0)
+ if self._n_nextn > 0:
+ self.block_count = self.hparams["num_hidden_layers"] + self._n_nextn
+ self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+ def tensor_force_quant(self, name, new_name, bid, n_dims):
+ # tiny depthwise conv kernels, consumed as F32 by ggml_ssm_conv
+ if new_name.endswith("_conv.weight") and ".attn_" in new_name:
+ return gguf.GGMLQuantizationType.F32
+ return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+ def set_vocab(self):
+ # openPangu ships a custom tokenizer_class in tokenizer_config.json, but
+ # tokenizer.json is a standard ByteLevel BPE. Allow AutoTokenizer to load
+ # it without prompting for the custom-code confirmation.
+ from transformers import AutoTokenizer
+ orig = AutoTokenizer.from_pretrained
+
+ def patched(*args, **kwargs):
+ kwargs.setdefault("trust_remote_code", True)
+ return orig(*args, **kwargs)
+
+ AutoTokenizer.from_pretrained = staticmethod(patched)
+ try:
+ self._set_vocab_gpt2()
+ finally:
+ AutoTokenizer.from_pretrained = orig
+
+ def set_gguf_parameters(self):
+ hparams = self.hparams
+ # MLA uses a single compressed KV head
+ hparams.setdefault("num_key_value_heads", 1)
+
+ super().set_gguf_parameters()
+
+ self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+ self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
+ self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+ self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+ self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
+
+ self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+ self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+ self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
+ self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
+ self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+ self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+ self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
+ self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
+ self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
+
+ self.gguf_writer.add_hyper_connection_count(hparams["mhc_num_stream"])
+ self.gguf_writer.add_hyper_connection_sinkhorn_iterations(hparams["mhc_recur_norm"])
+ self.gguf_writer.add_hyper_connection_epsilon(1e-6)
+
+ self.gguf_writer.add_sink_count(hparams["param_sink_number"])
+ self.gguf_writer.add_conv_kernel_size(hparams["router_sliding_window"])
+
+ if self._n_nextn > 0:
+ self.gguf_writer.add_nextn_predict_layers(self._n_nextn)
+
+ # per-layer sliding window: DSA layers get 0, SWA layers get their window.
+ # NOTE: the reference runs the MTP / NextN layers with SWA (window 2048,
+ # sliding_window_list[-1]), but llama.cpp has a single global window size and
+ # the trunk SWA layers use 512, so the MTP layers are converted as global (0).
+ # This diverges from the reference at contexts beyond the MTP window and can
+ # only lower draft acceptance (drafts are verified by the trunk).
+ n_layer = hparams["num_hidden_layers"]
+ swa_layers = hparams["swa_layers"]
+ window_list = hparams["sliding_window_list"]
+ layer_window = []
+ for il in range(self.block_count):
+ if il < n_layer and il in swa_layers:
+ layer_window.append(int(window_list[swa_layers.index(il)]))
+ else:
+ layer_window.append(0)
+ self.gguf_writer.add_sliding_window(int(hparams["sliding_window"]))
+ self.gguf_writer.add_layer_swa_window(layer_window)
+
+ def _fold_mhc(self, key, tensor_enum, bid, data_torch, is_phi):
+ buf = self._mhc_buf.setdefault(key, {})
+ buf["phi" if is_phi else "gamma"] = data_torch
+ if "phi" not in buf or "gamma" not in buf:
+ return []
+ # fold RMSNorm gamma into phi: linear(x*gamma, W) == linear(x, W*gamma)
+ folded = (buf["phi"].float() * buf["gamma"].float())
+ self._mhc_buf.pop(key)
+ return [(self.format_tensor_name(tensor_enum, bid), folded)]
+
+ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+ # MLA absorption: split kv_b_proj into k_b (transposed) and v_b
+ if bid is not None and name.endswith("self_attn.kv_b_proj.weight"):
+ n_head = self.hparams["num_attention_heads"]
+ v_head_dim = self.hparams["v_head_dim"]
+ qk_nope = self.hparams["qk_nope_head_dim"]
+ kv_b = data_torch.view(n_head, v_head_dim + qk_nope, data_torch.shape[-1])
+ k_b, v_b = torch.split(kv_b, [qk_nope, v_head_dim], dim=1)
+ k_b = k_b.transpose(1, 2).contiguous()
+ v_b = v_b.contiguous()
+ return [
+ (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K_B, bid), k_b),
+ (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V_B, bid), v_b),
+ ]
+
+ # mHC phi + norm_gamma get folded together
+ if name == "model.merge_mhc_module.phi.weight":
+ return self._fold_mhc("merge", gguf.MODEL_TENSOR.HC_HEAD_FN, None, data_torch, True)
+ if name == "model.merge_mhc_module.norm_gamma":
+ return self._fold_mhc("merge", gguf.MODEL_TENSOR.HC_HEAD_FN, None, data_torch, False)
+ if name in self._root_map:
+ enum, suffix = self._root_map[name]
+ return [(self.format_tensor_name(enum, None, suffix), data_torch)]
+
+ if bid is not None:
+ suffix = name.split(f"model.layers.{bid}.", 1)[-1]
+
+ for mod, fn_enum in (("attn_mhc_module", gguf.MODEL_TENSOR.HC_ATTN_FN),
+ ("mlp_mhc_module", gguf.MODEL_TENSOR.HC_FFN_FN)):
+ if suffix == f"{mod}.phi.weight":
+ return self._fold_mhc((bid, mod), fn_enum, bid, data_torch, True)
+ if suffix == f"{mod}.norm_gamma":
+ return self._fold_mhc((bid, mod), fn_enum, bid, data_torch, False)
+
+ # routed experts: stack across the expert dimension
+ if ".mlp.experts." in name:
+ parts = suffix.split(".") # mlp experts E proj weight
+ eid = int(parts[2])
+ proj = parts[3]
+ ebuf = self._exp_buf.setdefault((bid, proj), {})
+ ebuf[eid] = data_torch
+ n_experts = self.hparams["n_routed_experts"]
+ if len(ebuf) < n_experts:
+ return []
+ stacked = torch.stack([ebuf[i] for i in range(n_experts)], dim=0)
+ self._exp_buf.pop((bid, proj))
+ return [(self.format_tensor_name(self._expert_map[proj], bid), stacked)]
+
+ if suffix in self._layer_map:
+ enum, sfx = self._layer_map[suffix]
+ out = data_torch
+ # depthwise conv weights are [C, 1, k]; drop the singleton group dim.
+ # store as F32 so ggml_ssm_conv uses them without a per-eval cast
+ # (like Mamba's ssm_conv1d; they are tiny)
+ if enum in (gguf.MODEL_TENSOR.ATTN_Q_A_CONV,
+ gguf.MODEL_TENSOR.ATTN_KV_A_CONV,
+ gguf.MODEL_TENSOR.ATTN_O_CONV):
+ if out.ndim == 3:
+ out = out.squeeze(1)
+ out = out.float()
+ return [(self.format_tensor_name(enum, bid, sfx), out)]
+
+ logger.warning(f"openpangu-v2: unmapped tensor '{name}'")
+ return []
diff --git a/conversion/split_pangu_mtp.py b/conversion/split_pangu_mtp.py
new file mode 100644
index 000000000..bd96af26d
--- /dev/null
+++ b/conversion/split_pangu_mtp.py
@@ -0,0 +1,166 @@
+#!/usr/bin/env python3
+"""Split a combined openPangu-2.0 (openpangu-v2) GGUF into:
+
+ - a base GGUF without the MTP / NextN layers (smaller; use standalone), and
+ - an MTP-only GGUF holding just the NextN head layers plus the shared
+ embedding / output tensors, usable for self-speculative decoding via
+ `--model-draft <mtp.gguf> --mtp`.
+
+The MTP GGUF renumbers the NextN layers to blk.0.. and carries
+block_count == nextn_predict_layers, so hparams.n_layer() == 0 and every
+layer is an MTP layer. Its per-layer SWA window is set to the reference
+value (sliding_window_list[-1], 2048 for openPangu-2.0-Flash) -- the split
+model runs the MTP head with its trained sliding window, which the combined
+GGUF cannot express next to the trunk's window.
+
+Example:
+ python conversion/split_pangu_mtp.py openPangu-2.0-Flash-BF16.gguf \
+ --base openPangu-2.0-Flash-base-BF16.gguf \
+ --mtp openPangu-2.0-Flash-mtp-BF16.gguf
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+import re
+import sys
+from pathlib import Path
+
+if 'NO_LOCAL_GGUF' not in __import__('os').environ:
+ sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py'))
+
+import gguf
+from gguf import GGUFReader, GGUFWriter, GGUFValueType # noqa: E402
+
+logger = logging.getLogger("split-pangu-mtp")
+
+# non-blk tensors the MTP head graph reads (shared head/embedding fallbacks)
+MTP_SHARED_TENSORS = {"token_embd.weight", "output.weight", "output_norm.weight"}
+
+BLK_RE = re.compile(r"^blk\.(\d+)\.(.+)$")
+
+
+def copy_kv(reader: GGUFReader, writer: GGUFWriter, overrides: dict, remove: set) -> None:
+ for field in reader.fields.values():
+ if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'):
+ continue
+ if field.name in remove:
+ logger.debug('removing %s', field.name)
+ continue
+
+ val_type = field.types[0]
+ sub_type = field.types[-1] if val_type == GGUFValueType.ARRAY else None
+
+ value = field.contents()
+ if field.name in overrides:
+ value = overrides.pop(field.name)
+ logger.info('overriding %s -> %s', field.name, str(value)[:80])
+
+ writer.add_key_value(field.name, value, val_type, sub_type=sub_type)
+
+ for key, (value, val_type) in overrides.items():
+ logger.info('adding %s = %s', key, str(value)[:80])
+ writer.add_key_value(key, value, val_type)
+
+
+def write_split(reader: GGUFReader, path: Path, arch: str, overrides: dict, remove: set,
+ keep_tensor) -> None:
+ writer = GGUFWriter(path, arch)
+ copy_kv(reader, writer, dict(overrides), remove)
+
+ tensors = []
+ for tensor in reader.tensors:
+ new_name = keep_tensor(tensor.name)
+ if new_name is None:
+ continue
+ tensors.append((new_name, tensor))
+ writer.add_tensor_info(new_name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes,
+ tensor.tensor_type)
+
+ writer.write_header_to_file()
+ writer.write_kv_data_to_file()
+ writer.write_ti_data_to_file()
+
+ for i, (new_name, tensor) in enumerate(tensors):
+ logger.info('[%d/%d] writing %s (%s)', i + 1, len(tensors), new_name, path.name)
+ writer.write_tensor_data(tensor.data, tensor_endianess=reader.endianess)
+
+ writer.close()
+
+
+def main() -> None:
+ parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+ parser.add_argument("input", type=Path, help="combined openpangu-v2 GGUF")
+ parser.add_argument("--base", type=Path, help="output path for the base (no-MTP) GGUF")
+ parser.add_argument("--mtp", type=Path, help="output path for the MTP-only GGUF")
+ parser.add_argument("--mtp-swa-window", type=int, default=2048,
+ help="SWA window for the MTP layers (reference: sliding_window_list[-1], default 2048)")
+ parser.add_argument("--verbose", action="store_true")
+ args = parser.parse_args()
+
+ logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+
+ if not args.base and not args.mtp:
+ parser.error("nothing to do: pass --base and/or --mtp")
+
+ reader = GGUFReader(args.input)
+
+ arch = reader.get_field(gguf.Keys.General.ARCHITECTURE).contents()
+ if arch != "openpangu-v2":
+ parser.error(f"expected an openpangu-v2 GGUF, got architecture '{arch}'")
+
+ n_layer_all = reader.get_field(f"{arch}.block_count").contents()
+ field_nextn = reader.get_field(f"{arch}.nextn_predict_layers")
+ n_nextn = field_nextn.contents() if field_nextn else 0
+ if n_nextn <= 0:
+ parser.error("this GGUF has no MTP / NextN layers (nextn_predict_layers missing or 0)")
+ n_base = n_layer_all - n_nextn
+
+ swa_field = reader.get_field(f"{arch}.attention.layer_swa_window")
+ swa_all = list(swa_field.contents()) if swa_field else [0] * n_layer_all
+
+ logger.info("input: %d layers = %d base + %d MTP", n_layer_all, n_base, n_nextn)
+
+ if args.base:
+ def keep_base(name: str):
+ m = BLK_RE.match(name)
+ if m and int(m.group(1)) >= n_base:
+ return None
+ return name
+
+ write_split(reader, args.base, arch,
+ overrides={
+ f"{arch}.block_count": n_base,
+ f"{arch}.attention.layer_swa_window": swa_all[:n_base],
+ },
+ remove={f"{arch}.nextn_predict_layers"},
+ keep_tensor=keep_base)
+ logger.info("base GGUF written to %s", args.base)
+
+ if args.mtp:
+ def keep_mtp(name: str):
+ m = BLK_RE.match(name)
+ if m:
+ il = int(m.group(1))
+ if il < n_base:
+ return None
+ return f"blk.{il - n_base}.{m.group(2)}"
+ if name in MTP_SHARED_TENSORS:
+ return name
+ return None
+
+ write_split(reader, args.mtp, arch,
+ overrides={
+ f"{arch}.block_count": n_nextn,
+ f"{arch}.nextn_predict_layers": n_nextn,
+ f"{arch}.leading_dense_block_count": 0,
+ f"{arch}.attention.layer_swa_window": [args.mtp_swa_window] * n_nextn,
+ },
+ remove=set(),
+ keep_tensor=keep_mtp)
+ logger.info("MTP GGUF written to %s", args.mtp)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index c37ba9471..efff904ab 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -567,6 +567,9 @@ extern "C" {
GGML_OP_LEAKY_RELU,
GGML_OP_TRI,
GGML_OP_FILL,
+ GGML_OP_SINKHORN,
+ GGML_OP_HC_MIX,
+ GGML_OP_DSA_SCORE,
GGML_OP_FLASH_ATTN_EXT,
GGML_OP_FLASH_ATTN_BACK,
@@ -2439,6 +2442,47 @@ extern "C" {
struct ggml_tensor * a,
float c);
+ // Sinkhorn row/column normalization over the leading [ne0, ne1] matrix of each
+ // (i2, i3) batch (used by hyper-connection stream mixing):
+ // 1. softmax along dim0 (per i1)
+ // 2. add eps to every element
+ // 3. normalize along dim1: M[i0,:] /= (eps + sum_i1 M[i0,i1])
+ // 4. repeat (n_iter-1) times:
+ // normalize along dim0: M[:,i1] /= (eps + sum_i0 M[i0,i1]), then step 3
+ GGML_API struct ggml_tensor * ggml_sinkhorn(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ int n_iter,
+ float eps);
+
+ // Hyper-connection stream mixing: for each token (rows of a [2*hc+hc*hc, nt]):
+ // out[0 .. hc) = sigmoid(a*scale[0] + base[i]) + eps (pre)
+ // out[hc .. 2*hc) = 2*sigmoid(a*scale[1] + base[i]) (post)
+ // out[2*hc .. 2*hc+hc*hc) = sinkhorn(a*scale[2] + base[i], n_iter, eps) (comb,
+ // the [hc, hc] matrix laid out dim0-major, same semantics as ggml_sinkhorn)
+ GGML_API struct ggml_tensor * ggml_hc_mix(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ struct ggml_tensor * scale, // [3]
+ struct ggml_tensor * base, // [2*hc+hc*hc]
+ int hc,
+ int n_iter,
+ float eps);
+
+ // DSA lightning-indexer scores (one pass over the cached indexer keys):
+ // out[s,t] = mask[s,t] + sum_h w[h,t] * relu(sum_d ik[d,s] * q[d,h,t])
+ // ik: [d, n_kv] indexer keys (may be a strided view; F16 or F32)
+ // q: [d, h, nt] per-token indexer queries (F32)
+ // w: [h, nt] per-token head weights (F32)
+ // mask: [n_kv, nt] additive mask (F16 or F32)
+ // returns [n_kv, nt] F32
+ GGML_API struct ggml_tensor * ggml_dsa_score(
+ struct ggml_context * ctx,
+ struct ggml_tensor * ik,
+ struct ggml_tensor * q,
+ struct ggml_tensor * w,
+ struct ggml_tensor * mask);
+
// Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
// timesteps: [N,]
// return: [N, dim]
diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp
index e90c2797f..37780d618 100644
--- a/ggml/src/ggml-cpu/amx/amx.cpp
+++ b/ggml/src/ggml-cpu/amx/amx.cpp
@@ -146,6 +146,9 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
if (op->op != GGML_OP_MUL_MAT) {
return false;
}
+ if (op->src[0]->ne[2] > 1 || op->src[0]->ne[3] > 1) {
+ return false; // packing is not implemented for batched tensors
+ }
auto * src0 = op->src[0];
auto * src1 = op->src[1];
diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp
index 9f3a744b5..2d79f714e 100644
--- a/ggml/src/ggml-cpu/amx/mmq.cpp
+++ b/ggml/src/ggml-cpu/amx/mmq.cpp
@@ -418,7 +418,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
// Quantize these floats
const float iscale = 127.f / amax;
- y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
+ y[i].d = 1 / iscale;
const float id = ( amax != 0.0f ) ? iscale : 0.f;
const __m512 vscale = _mm512_set1_ps(id);
@@ -470,12 +470,7 @@ inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
template <>
inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
-#if 1
- // TODO: this is reference impl!
- quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
-#else
quantize_row_q8_K_vnni(x, vy, k);
-#endif
}
// load A from memory to array when nrows can not fill in whole tile
diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c
index 21cd6312a..9d05aac0e 100644
--- a/ggml/src/ggml-cpu/arch/x86/quants.c
+++ b/ggml/src/ggml-cpu/arch/x86/quants.c
@@ -2302,11 +2302,107 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const __m128i mzero = _mm_setzero_si128();
const __m256i mone = _mm256_set1_epi8(1);
- __m256 acc = _mm256_setzero_ps();
-
float summs = 0.f;
- for (int i = 0; i < nb; ++i) {
+ int i = 0;
+
+#if defined(__AVX512F__) && defined(__AVX512BW__) && defined(__AVX512DQ__)
+ __m512 acc512 = _mm512_setzero_ps();
+ const __m256i scale_shuffle = _mm256_broadcastsi128_si256(_mm_setr_epi8(
+ 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11
+ ));
+ const __m256i scale_hi_shuffle = _mm256_broadcastsi128_si256(_mm_setr_epi8(
+ -1, -1, -1, -1, 0, 1, 2, 3, -1, -1, -1, -1, 4, 5, 6, 7
+ ));
+ const __m256i scale_mask = _mm256_setr_epi32(kmask1, kmask2, kmask1, kmask2, kmask1, kmask2, kmask1, kmask2);
+ const __m256i scale_hi_mask = _mm256_setr_epi32(0, kmask3, 0, kmask3, 0, kmask3, 0, kmask3);
+
+ for (; i + 1 < nb; i += 2) {
+ const __m256i scale_bytes = MM256_SET_M128I(
+ _mm_loadu_si128((const __m128i *) x[i + 1].scales),
+ _mm_loadu_si128((const __m128i *) x[i + 0].scales)
+ );
+ const __m256i scale_ordered = _mm256_shuffle_epi8(scale_bytes, scale_shuffle);
+ const __m256i scale_low = _mm256_and_si256(_mm256_blend_epi32(scale_ordered, _mm256_srli_epi16(scale_ordered, 4), 0x88), scale_mask);
+ const __m256i scale_high = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(_mm256_shuffle_epi8(scale_bytes, scale_hi_shuffle), 6), scale_hi_mask), 4);
+ const __m256i mins_and_scales8 = _mm256_or_si256(scale_low, scale_high);
+ const __m256i mins_and_scales0 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(mins_and_scales8, 0));
+ const __m256i mins_and_scales1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(mins_and_scales8, 1));
+
+ const __m256i q8sums0 = _mm256_loadu_si256((const __m256i *) y[i + 0].bsums);
+ const __m128i q8s0 = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums0, 0), _mm256_extracti128_si256(q8sums0, 1));
+ const __m128i prod0 = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales0, 1), q8s0);
+ const __m128i hsum0 = _mm_hadd_epi32(_mm_hadd_epi32(prod0, mzero), mzero);
+ summs -= y[i + 0].d * GGML_CPU_FP16_TO_FP32(x[i + 0].dmin) * _mm_extract_epi32(hsum0, 0);
+
+ const __m256i q8sums1 = _mm256_loadu_si256((const __m256i *) y[i + 1].bsums);
+ const __m128i q8s1 = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums1, 0), _mm256_extracti128_si256(q8sums1, 1));
+ const __m128i prod1 = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales1, 1), q8s1);
+ const __m128i hsum1 = _mm_hadd_epi32(_mm_hadd_epi32(prod1, mzero), mzero);
+ summs -= y[i + 1].d * GGML_CPU_FP16_TO_FP32(x[i + 1].dmin) * _mm_extract_epi32(hsum1, 0);
+
+ const __m128i sc1280 = _mm256_extracti128_si256(mins_and_scales0, 0);
+ const __m128i sc1281 = _mm256_extracti128_si256(mins_and_scales1, 0);
+ const __m256i scales0 = MM256_SET_M128I(sc1280, sc1280);
+ const __m256i scales1 = MM256_SET_M128I(sc1281, sc1281);
+ const __m512i scales = _mm512_inserti64x4(_mm512_castsi256_si512(scales0), scales1, 1);
+
+ const __m256i hbits0 = _mm256_loadu_si256((const __m256i *) x[i + 0].qh);
+ const __m256i hbits1 = _mm256_loadu_si256((const __m256i *) x[i + 1].qh);
+ const __m512i hbits = _mm512_inserti64x4(_mm512_castsi256_si512(hbits0), hbits1, 1);
+ __m512i hmask = _mm512_set1_epi8(1);
+ __m512i sumi = _mm512_setzero_si512();
+
+ int bit = 0;
+
+ for (int j = 0; j < QK_K/64; ++j) {
+ const __m512i scale_0 = _mm512_shuffle_epi8(scales, _mm512_broadcast_i64x4(get_scale_shuffle_k4(2*j + 0)));
+ const __m512i scale_1 = _mm512_shuffle_epi8(scales, _mm512_broadcast_i64x4(get_scale_shuffle_k4(2*j + 1)));
+
+ const __m256i q5bits0 = _mm256_loadu_si256((const __m256i *) (x[i + 0].qs + 32*j));
+ const __m256i q5bits1 = _mm256_loadu_si256((const __m256i *) (x[i + 1].qs + 32*j));
+ const __m512i q5bits = _mm512_inserti64x4(_mm512_castsi256_si512(q5bits0), q5bits1, 1);
+
+ const __m512i q5l_0 = _mm512_and_si512(q5bits, _mm512_set1_epi8(0xF));
+ const __m512i q5h_0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask), bit++), 4);
+ const __m512i q5_0 = _mm512_add_epi8(q5l_0, q5h_0);
+ hmask = _mm512_slli_epi16(hmask, 1);
+
+ const __m512i q5l_1 = _mm512_and_si512(_mm512_srli_epi16(q5bits, 4), _mm512_set1_epi8(0xF));
+ const __m512i q5h_1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask), bit++), 4);
+ const __m512i q5_1 = _mm512_add_epi8(q5l_1, q5h_1);
+ hmask = _mm512_slli_epi16(hmask, 1);
+
+ const __m256i q8_00 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 64*j + 0));
+ const __m256i q8_01 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 64*j + 0));
+ const __m512i q8_0 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_00), q8_01, 1);
+ const __m256i q8_10 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 64*j + 32));
+ const __m256i q8_11 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 64*j + 32));
+ const __m512i q8_1 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_10), q8_11, 1);
+
+ const __m512i p16_0 = _mm512_maddubs_epi16(q5_0, q8_0);
+ const __m512i p16_1 = _mm512_maddubs_epi16(q5_1, q8_1);
+#if defined(__AVX512VNNI__)
+ sumi = _mm512_dpwssd_epi32(sumi, scale_0, p16_0);
+ sumi = _mm512_dpwssd_epi32(sumi, scale_1, p16_1);
+#else
+ const __m512i p32_0 = _mm512_madd_epi16(scale_0, p16_0);
+ const __m512i p32_1 = _mm512_madd_epi16(scale_1, p16_1);
+ sumi = _mm512_add_epi32(sumi, _mm512_add_epi32(p32_0, p32_1));
+#endif
+ }
+
+ __m512 vd = _mm512_castps256_ps512(_mm256_set1_ps(y[i + 0].d * GGML_CPU_FP16_TO_FP32(x[i + 0].d)));
+ vd = _mm512_insertf32x8(vd, _mm256_set1_ps(y[i + 1].d * GGML_CPU_FP16_TO_FP32(x[i + 1].d)), 1);
+ acc512 = _mm512_fmadd_ps(vd, _mm512_cvtepi32_ps(sumi), acc512);
+ }
+
+ __m256 acc = _mm256_add_ps(_mm512_castps512_ps256(acc512), _mm512_extractf32x8_ps(acc512, 1));
+#else
+ __m256 acc = _mm256_setzero_ps();
+#endif
+
+ for (; i < nb; ++i) {
const uint8_t * GGML_RESTRICT q5 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;
@@ -2358,13 +2454,20 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
- __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
- __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+ const __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+ const __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
- p16_0 = _mm256_madd_epi16(scale_0, p16_0);
- p16_1 = _mm256_madd_epi16(scale_1, p16_1);
-
- sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
+ sumi = _mm256_dpwssd_epi32(sumi, scale_0, p16_0);
+ sumi = _mm256_dpwssd_epi32(sumi, scale_1, p16_1);
+#elif defined(__AVXVNNI__)
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale_0, p16_0);
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale_1, p16_1);
+#else
+ const __m256i p32_0 = _mm256_madd_epi16(scale_0, p16_0);
+ const __m256i p32_1 = _mm256_madd_epi16(scale_1, p16_1);
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p32_0, p32_1));
+#endif
}
@@ -2505,9 +2608,105 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const __m256i m3 = _mm256_set1_epi8(3);
const __m256i m15 = _mm256_set1_epi8(15);
+ int i = 0;
+
+#if defined(__AVX512F__) && defined(__AVX512BW__) && defined(__AVX512DQ__)
+ const __m512i m3_512 = _mm512_set1_epi8(3);
+ const __m512i m12_512 = _mm512_set1_epi8(12);
+ const __m512i m15_512 = _mm512_set1_epi8(15);
+ const __m512i m48_512 = _mm512_set1_epi8(48);
+ const __m512i mc0_512 = _mm512_set1_epi8(-64);
+
+ __m512 acc512 = _mm512_setzero_ps();
+
+ for (; i + 1 < nb; i += 2) {
+ const __m256i q8sums0 = _mm256_loadu_si256((const __m256i *) y[i + 0].bsums);
+ const __m256i q8sums1 = _mm256_loadu_si256((const __m256i *) y[i + 1].bsums);
+ const __m512i q8sums = _mm512_inserti64x4(_mm512_castsi256_si512(q8sums0), q8sums1, 1);
+
+ const __m128i scales0 = _mm_loadu_si128((const __m128i *) x[i + 0].scales);
+ const __m128i scales1 = _mm_loadu_si128((const __m128i *) x[i + 1].scales);
+ const __m256i scales = MM256_SET_M128I(scales1, scales0);
+ const __m512i scales_16 = _mm512_cvtepi8_epi16(scales);
+ const __m512i q8sclsub = _mm512_slli_epi32(_mm512_madd_epi16(q8sums, scales_16), 5);
+
+ __m512i sumi = _mm512_setzero_si512();
+
+ for (int j = 0; j < QK_K/128; ++j) {
+ const __m256i q4bits10 = _mm256_loadu_si256((const __m256i *) (x[i + 0].ql + 64*j + 0));
+ const __m256i q4bits11 = _mm256_loadu_si256((const __m256i *) (x[i + 1].ql + 64*j + 0));
+ const __m512i q4bits1 = _mm512_inserti64x4(_mm512_castsi256_si512(q4bits10), q4bits11, 1);
+ const __m256i q4bits20 = _mm256_loadu_si256((const __m256i *) (x[i + 0].ql + 64*j + 32));
+ const __m256i q4bits21 = _mm256_loadu_si256((const __m256i *) (x[i + 1].ql + 64*j + 32));
+ const __m512i q4bits2 = _mm512_inserti64x4(_mm512_castsi256_si512(q4bits20), q4bits21, 1);
+ const __m256i q4bitsH0 = _mm256_loadu_si256((const __m256i *) (x[i + 0].qh + 32*j));
+ const __m256i q4bitsH1 = _mm256_loadu_si256((const __m256i *) (x[i + 1].qh + 32*j));
+ const __m512i q4bitsH = _mm512_inserti64x4(_mm512_castsi256_si512(q4bitsH0), q4bitsH1, 1);
+
+ const __m512i q4h_0 = _mm512_slli_epi16(_mm512_and_si512(q4bitsH, m3_512), 4);
+ const __m512i q4h_1 = _mm512_slli_epi16(_mm512_and_si512(q4bitsH, m12_512), 2);
+ const __m512i q4h_2 = _mm512_and_si512(q4bitsH, m48_512);
+ const __m512i q4h_3 = _mm512_srli_epi16(_mm512_and_si512(q4bitsH, mc0_512), 2);
+
+ const __m512i q4_0 = _mm512_or_si512(_mm512_and_si512(q4bits1, m15_512), q4h_0);
+ const __m512i q4_1 = _mm512_or_si512(_mm512_and_si512(q4bits2, m15_512), q4h_1);
+ const __m512i q4_2 = _mm512_or_si512(_mm512_and_si512(_mm512_srli_epi16(q4bits1, 4), m15_512), q4h_2);
+ const __m512i q4_3 = _mm512_or_si512(_mm512_and_si512(_mm512_srli_epi16(q4bits2, 4), m15_512), q4h_3);
+
+ const __m256i q8_00 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 128*j + 0));
+ const __m256i q8_01 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 128*j + 0));
+ const __m512i q8_0 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_00), q8_01, 1);
+ const __m256i q8_10 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 128*j + 32));
+ const __m256i q8_11 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 128*j + 32));
+ const __m512i q8_1 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_10), q8_11, 1);
+ const __m256i q8_20 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 128*j + 64));
+ const __m256i q8_21 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 128*j + 64));
+ const __m512i q8_2 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_20), q8_21, 1);
+ const __m256i q8_30 = _mm256_loadu_si256((const __m256i *) (y[i + 0].qs + 128*j + 96));
+ const __m256i q8_31 = _mm256_loadu_si256((const __m256i *) (y[i + 1].qs + 128*j + 96));
+ const __m512i q8_3 = _mm512_inserti64x4(_mm512_castsi256_si512(q8_30), q8_31, 1);
+
+ const __m256i scale_shuffle0 = MM256_SET_M128I(get_scale_shuffle(4*j + 0), get_scale_shuffle(4*j + 0));
+ const __m256i scale_shuffle1 = MM256_SET_M128I(get_scale_shuffle(4*j + 1), get_scale_shuffle(4*j + 1));
+ const __m256i scale_shuffle2 = MM256_SET_M128I(get_scale_shuffle(4*j + 2), get_scale_shuffle(4*j + 2));
+ const __m256i scale_shuffle3 = MM256_SET_M128I(get_scale_shuffle(4*j + 3), get_scale_shuffle(4*j + 3));
+
+ const __m512i scale_0 = _mm512_cvtepi8_epi16(_mm256_shuffle_epi8(scales, scale_shuffle0));
+ const __m512i scale_1 = _mm512_cvtepi8_epi16(_mm256_shuffle_epi8(scales, scale_shuffle1));
+ const __m512i scale_2 = _mm512_cvtepi8_epi16(_mm256_shuffle_epi8(scales, scale_shuffle2));
+ const __m512i scale_3 = _mm512_cvtepi8_epi16(_mm256_shuffle_epi8(scales, scale_shuffle3));
+ const __m512i p16_0 = _mm512_maddubs_epi16(q4_0, q8_0);
+ const __m512i p16_1 = _mm512_maddubs_epi16(q4_1, q8_1);
+ const __m512i p16_2 = _mm512_maddubs_epi16(q4_2, q8_2);
+ const __m512i p16_3 = _mm512_maddubs_epi16(q4_3, q8_3);
+
+#if defined(__AVX512VNNI__)
+ sumi = _mm512_dpwssd_epi32(sumi, scale_0, p16_0);
+ sumi = _mm512_dpwssd_epi32(sumi, scale_1, p16_1);
+ sumi = _mm512_dpwssd_epi32(sumi, scale_2, p16_2);
+ sumi = _mm512_dpwssd_epi32(sumi, scale_3, p16_3);
+#else
+ const __m512i p32_0 = _mm512_madd_epi16(scale_0, p16_0);
+ const __m512i p32_1 = _mm512_madd_epi16(scale_1, p16_1);
+ const __m512i p32_2 = _mm512_madd_epi16(scale_2, p16_2);
+ const __m512i p32_3 = _mm512_madd_epi16(scale_3, p16_3);
+ sumi = _mm512_add_epi32(sumi, _mm512_add_epi32(p32_0, p32_1));
+ sumi = _mm512_add_epi32(sumi, _mm512_add_epi32(p32_2, p32_3));
+#endif
+ }
+
+ sumi = _mm512_sub_epi32(sumi, q8sclsub);
+ __m512 vd = _mm512_castps256_ps512(_mm256_set1_ps(y[i + 0].d * GGML_CPU_FP16_TO_FP32(x[i + 0].d)));
+ vd = _mm512_insertf32x8(vd, _mm256_set1_ps(y[i + 1].d * GGML_CPU_FP16_TO_FP32(x[i + 1].d)), 1);
+ acc512 = _mm512_fmadd_ps(vd, _mm512_cvtepi32_ps(sumi), acc512);
+ }
+
+ __m256 acc = _mm256_add_ps(_mm512_castps512_ps256(acc512), _mm512_extractf32x8_ps(acc512, 1));
+#else
__m256 acc = _mm256_setzero_ps();
+#endif
- for (int i = 0; i < nb; ++i) {
+ for (; i < nb; ++i) {
const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
@@ -2544,10 +2743,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
- __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
- __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
- __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
- __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+ const __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+ const __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+ const __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+ const __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
@@ -2555,13 +2754,29 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
is += 4;
- p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
- p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
- p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
- p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+ const __m256i scale16_0 = _mm256_cvtepi8_epi16(scale_0);
+ const __m256i scale16_1 = _mm256_cvtepi8_epi16(scale_1);
+ const __m256i scale16_2 = _mm256_cvtepi8_epi16(scale_2);
+ const __m256i scale16_3 = _mm256_cvtepi8_epi16(scale_3);
- sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
- sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
+ sumi = _mm256_dpwssd_epi32(sumi, scale16_0, p16_0);
+ sumi = _mm256_dpwssd_epi32(sumi, scale16_1, p16_1);
+ sumi = _mm256_dpwssd_epi32(sumi, scale16_2, p16_2);
+ sumi = _mm256_dpwssd_epi32(sumi, scale16_3, p16_3);
+#elif defined(__AVXVNNI__)
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale16_0, p16_0);
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale16_1, p16_1);
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale16_2, p16_2);
+ sumi = _mm256_dpwssd_avx_epi32(sumi, scale16_3, p16_3);
+#else
+ const __m256i p32_0 = _mm256_madd_epi16(scale16_0, p16_0);
+ const __m256i p32_1 = _mm256_madd_epi16(scale16_1, p16_1);
+ const __m256i p32_2 = _mm256_madd_epi16(scale16_2, p16_2);
+ const __m256i p32_3 = _mm256_madd_epi16(scale16_3, p16_3);
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p32_0, p32_1));
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p32_2, p32_3));
+#endif
}
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index af1cebad1..f30d822c5 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -2062,7 +2062,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
UNUSED(ncols_interleaved);
UNUSED(blocklen);
-#if defined(__AVX2__) || defined(__AVX512F__)
+#if defined(__AVX2__)
const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx;
const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
int64_t b_nb = n / QK_K;
@@ -2074,8 +2074,14 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
int64_t xstart = 0;
int anr = nr - nr % 16;; // Used to align nr with boundary of 16
-#if defined(__AVX512BW__) && defined(__AVX512DQ__)
+#if defined(__AVX512F__) && defined(__AVX512BW__) && defined(__AVX512DQ__)
int anc = nc - nc % 16; // Used to align nc with boundary of 16
+
+ const __m512i rhsLeftOrderSp1 = _mm512_setr_epi32(0x00, 0x02, 0x00, 0x02, 0x08, 0x0A, 0x08, 0x0A, 0x10, 0x12, 0x10, 0x12, 0x18, 0x1A, 0x18, 0x1A);
+ const __m512i rhsLeftOrderSp2 = _mm512_setr_epi32(0x01, 0x03, 0x01, 0x03, 0x09, 0x0B, 0x09, 0x0B, 0x11, 0x13, 0x11, 0x13, 0x19, 0x1B, 0x19, 0x1B);
+ const __m512i rhsRightOrderSp1 = _mm512_setr_epi32(0x04, 0x06, 0x04, 0x06, 0x0C, 0x0E, 0x0C, 0x0E, 0x14, 0x16, 0x14, 0x16, 0x1C, 0x1E, 0x1C, 0x1E);
+ const __m512i rhsRightOrderSp2 = _mm512_setr_epi32(0x05, 0x07, 0x05, 0x07, 0x0D, 0x0F, 0x0D, 0x0F, 0x15, 0x17, 0x15, 0x17, 0x1D, 0x1F, 0x1D, 0x1F);
+
// Mask to mask out nibbles from packed bytes expanded to 512 bit length
const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
//Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
@@ -2115,111 +2121,85 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
// Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
for (int sb = 0; sb < QK_K / 64; sb++) {
+ const __m512i rhs_raw_mat_01234567_0 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + sb * 256));
+ const __m512i rhs_raw_mat_01234567_1 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 64 + sb * 256));
+ const __m512i rhs_raw_mat_01234567_2 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 128 + sb * 256));
+ const __m512i rhs_raw_mat_01234567_3 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
- const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
- const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
- const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_0 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_1 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 64 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_2 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 128 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_3 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
- const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
- const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
- const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
- const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+ const __m512i rhs_raw_mat_014589CD_0_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_014589CD_0_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_2367ABEF_0_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_2367ABEF_0_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_0);
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
- const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
- const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+ const __m512i rhs_raw_mat_014589CD_1_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_014589CD_1_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_2367ABEF_1_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_2367ABEF_1_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_1);
- const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
- const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
- const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
- const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+ const __m512i rhs_raw_mat_014589CD_2_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_014589CD_2_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_2367ABEF_2_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_2367ABEF_2_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_2);
- const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
- const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
- const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
- const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+ const __m512i rhs_raw_mat_014589CD_3_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_014589CD_3_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_2367ABEF_3_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_2367ABEF_3_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_3);
- const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
- const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
- const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
- const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+ // 4bit-8bit pattern one - right side input
+ const __m512i rhs_mat_014589CD_00_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_0_sp1, m4bexpanded); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+ const __m512i rhs_mat_014589CD_10_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0_sp1, 4), m4bexpanded); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
- //4-bit -> 8-bit
- const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
- const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
- const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
- const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+ const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0_sp1, m4bexpanded); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+ const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0_sp1, 4), m4bexpanded); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
- const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
- const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
- const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
- const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+ const __m512i rhs_mat_014589CD_01_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_1_sp1, m4bexpanded); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+ const __m512i rhs_mat_014589CD_11_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1_sp1, 4), m4bexpanded); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
- const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
- const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
- const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
- const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+ const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1_sp1, m4bexpanded); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+ const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1_sp1, 4), m4bexpanded); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-1１) B１E(8-１１) B１F(8-１１) B１E(8-１１) B１F(8-１１)
- const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
- const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
- const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
- const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+ const __m512i rhs_mat_014589CD_02_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_2_sp1, m4bexpanded); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+ const __m512i rhs_mat_014589CD_12_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2_sp1, 4), m4bexpanded); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
- // Shuffle pattern one - right side input
- const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
- const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
- const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
- const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
- const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
- const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
- const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
- const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+ const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2_sp1, m4bexpanded); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+ const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2_sp1, 4), m4bexpanded); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
- const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
- const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
- const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
- const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
- const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
- const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
- const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
- const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+ const __m512i rhs_mat_014589CD_03_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_3_sp1, m4bexpanded); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+ const __m512i rhs_mat_014589CD_13_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3_sp1, 4), m4bexpanded); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
- // Shuffle pattern two - right side input
- const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
- const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
- const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
- const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
- const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
- const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
- const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
- const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+ const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3_sp1, m4bexpanded); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+ const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3_sp1, 4), m4bexpanded); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
- const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
- const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
- const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
- const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
- const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
- const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
- const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
- const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+ // 4bit-8bit pattern two - right side input
+ const __m512i rhs_mat_014589CD_00_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_0_sp2, m4bexpanded); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+ const __m512i rhs_mat_014589CD_10_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0_sp2, 4), m4bexpanded); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+
+ const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0_sp2, m4bexpanded); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7] B0F(4-7)
+ const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0_sp2, 4), m4bexpanded); //B12(4-7] B13(4-7] B12(4-7] B13(4-7] B16(4-7] B17(4-七年] B16(四年] B1七七年] B１A（四年]B１B（四年]B１A（四年]B１B（四年]B１E（四年]B１F（四年]B１E（四年]B１F（四年]
+
+ const __m512i rhs_mat_014589CD_01_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_1_sp2, m4bexpanded); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+ const __m512i rhs_mat_014589CD_11_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1_sp2, 4), m4bexpanded); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+
+ const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1_sp2, m4bexpanded); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+ const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1_sp2, 4), m4bexpanded); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+
+ const __m512i rhs_mat_014589CD_02_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_2_sp2, m4bexpanded); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+ const __m512i rhs_mat_014589CD_12_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2_sp2, 4), m4bexpanded); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+
+ const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2_sp2, m4bexpanded); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+ const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2_sp2, 4), m4bexpanded); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+
+ const __m512i rhs_mat_014589CD_03_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_3_sp2, m4bexpanded); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+ const __m512i rhs_mat_014589CD_13_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3_sp2, 4), m4bexpanded); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+
+ const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3_sp2, m4bexpanded); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+ const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3_sp2, 4), m4bexpanded); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
@@ -2273,51 +2253,32 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
for (int rp = 0; rp < 4; rp++) {
-
// Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
- __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb)));
- __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
- __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
- __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb)));
- __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
- __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
- __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb)));
- __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
- __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
- __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb)));
- __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
- __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
- __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb)));
- __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
- __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
- __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb)));
- __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
- __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
- __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb)));
- __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
- __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
- __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb)));
- __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
- __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
-
- __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
- __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
- __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
- __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
- __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
- __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
- __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
- __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
-
- __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
- __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
- __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
- __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
- __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
- __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
- __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
- __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+ __m512i lhs_mat_zmm_0123_00_01 = _mm512_loadu_si512((const __m512i * )((a_ptrs[rp][b].qs + sb * 256)));
+ __m512i lhs_mat_zmm_0123_02_03 = _mm512_loadu_si512((const __m512i * )((a_ptrs[rp][b].qs + 64 + sb * 256)));
+ __m512i lhs_mat_zmm_0123_10_11 = _mm512_loadu_si512((const __m512i * )((a_ptrs[rp][b].qs + 128 + sb * 256)));
+ __m512i lhs_mat_zmm_0123_12_13 = _mm512_loadu_si512((const __m512i * )((a_ptrs[rp][b].qs + 192 + sb * 256)));
+
+ // Loaded as set of 128 bit vectors and repeated and stored into a 512 bit vector
+ __m512i lhs_mat_01_00 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b00000000);
+ __m512i lhs_mat_23_00 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b01010101);
+ __m512i lhs_mat_01_01 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b10101010);
+ __m512i lhs_mat_23_01 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b11111111);
+
+ __m512i lhs_mat_01_02 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b00000000);
+ __m512i lhs_mat_23_02 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b01010101);
+ __m512i lhs_mat_01_03 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b10101010);
+ __m512i lhs_mat_23_03 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b11111111);
+
+ __m512i lhs_mat_01_10 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b00000000);
+ __m512i lhs_mat_23_10 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b01010101);
+ __m512i lhs_mat_01_11 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b10101010);
+ __m512i lhs_mat_23_11 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b11111111);
+
+ __m512i lhs_mat_01_12 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b00000000);
+ __m512i lhs_mat_23_12 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b01010101);
+ __m512i lhs_mat_01_13 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b10101010);
+ __m512i lhs_mat_23_13 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b11111111);
// Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
__m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb)));
@@ -2478,111 +2439,85 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
// Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
for (int sb = 0; sb < QK_K / 64; sb++) {
+ const __m512i rhs_raw_mat_01234567_0 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + sb * 256));
+ const __m512i rhs_raw_mat_01234567_1 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 64 + sb * 256));
+ const __m512i rhs_raw_mat_01234567_2 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 128 + sb * 256));
+ const __m512i rhs_raw_mat_01234567_3 = _mm512_loadu_si512((const __m512i * )(b_ptr_0[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
- const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
- const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
- const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_0 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_1 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 64 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_2 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 128 + sb * 256));
+ const __m512i rhs_raw_mat_89ABCDEF_3 = _mm512_loadu_si512((const __m512i * )(b_ptr_1[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
- const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
- const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
- const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
- const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
+ const __m512i rhs_raw_mat_014589CD_0_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_014589CD_0_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_2367ABEF_0_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_0);
+ const __m512i rhs_raw_mat_2367ABEF_0_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_0, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_0);
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
- const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
- const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+ const __m512i rhs_raw_mat_014589CD_1_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_014589CD_1_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_2367ABEF_1_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_1);
+ const __m512i rhs_raw_mat_2367ABEF_1_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_1, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_1);
- const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
- const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
- const __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240);
- const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 240);
+ const __m512i rhs_raw_mat_014589CD_2_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_014589CD_2_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_2367ABEF_2_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_2);
+ const __m512i rhs_raw_mat_2367ABEF_2_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_2, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_2);
- const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
- const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
- const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
- const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+ const __m512i rhs_raw_mat_014589CD_3_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsLeftOrderSp1, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_014589CD_3_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsLeftOrderSp2, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_2367ABEF_3_sp1 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsRightOrderSp1, rhs_raw_mat_89ABCDEF_3);
+ const __m512i rhs_raw_mat_2367ABEF_3_sp2 = _mm512_permutex2var_epi32(rhs_raw_mat_01234567_3, rhsRightOrderSp2, rhs_raw_mat_89ABCDEF_3);
- const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
- const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
- const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
- const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
+ // 4bit-8bit pattern one - right side input
+ const __m512i rhs_mat_014589CD_00_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_0_sp1, m4bexpanded); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
+ const __m512i rhs_mat_014589CD_10_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0_sp1, 4), m4bexpanded); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
- //4-bit -> 8-bit
- const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
- const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
- const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
- const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
+ const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0_sp1, m4bexpanded); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
+ const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0_sp1, 4), m4bexpanded); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
- const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
- const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
- const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
- const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
+ const __m512i rhs_mat_014589CD_01_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_1_sp1, m4bexpanded); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
+ const __m512i rhs_mat_014589CD_11_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1_sp1, 4), m4bexpanded); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
- const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
- const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
- const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
- const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
+ const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1_sp1, m4bexpanded); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
+ const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1_sp1, 4), m4bexpanded); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-1１) B１E(8-１１) B１F(8-１１) B１E(8-１１) B１F(8-１１)
- const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
- const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
- const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
- const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
+ const __m512i rhs_mat_014589CD_02_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_2_sp1, m4bexpanded); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
+ const __m512i rhs_mat_014589CD_12_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2_sp1, 4), m4bexpanded); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
- // Shuffle pattern one - right side input
- const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3)
- const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3)
- const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11)
- const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11)
- const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
- const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
- const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
- const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+ const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2_sp1, m4bexpanded); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
+ const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2_sp1, 4), m4bexpanded); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
- const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3)
- const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3)
- const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11)
- const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11)
- const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
- const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
- const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
- const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
+ const __m512i rhs_mat_014589CD_03_sp1 = _mm512_and_si512(rhs_raw_mat_014589CD_3_sp1, m4bexpanded); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
+ const __m512i rhs_mat_014589CD_13_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3_sp1, 4), m4bexpanded); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
- // Shuffle pattern two - right side input
- const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
- const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7)
- const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
- const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
- const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
- const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
- const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
- const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+ const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3_sp1, m4bexpanded); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
+ const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3_sp1, 4), m4bexpanded); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
- const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
- const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7)
- const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
- const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
- const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
- const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
- const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
- const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
+ // 4bit-8bit pattern two - right side input
+ const __m512i rhs_mat_014589CD_00_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_0_sp2, m4bexpanded); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7)
+ const __m512i rhs_mat_014589CD_10_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0_sp2, 4), m4bexpanded); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7)
+
+ const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0_sp2, m4bexpanded); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7] B0F(4-7)
+ const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0_sp2, 4), m4bexpanded); //B12(4-7] B13(4-7] B12(4-7] B13(4-7] B16(4-7] B17(4-七年] B16(四年] B1七七年] B１A（四年]B１B（四年]B１A（四年]B１B（四年]B１E（四年]B１F（四年]B１E（四年]B１F（四年]
+
+ const __m512i rhs_mat_014589CD_01_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_1_sp2, m4bexpanded); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15)
+ const __m512i rhs_mat_014589CD_11_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1_sp2, 4), m4bexpanded); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15)
+
+ const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1_sp2, m4bexpanded); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15)
+ const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1_sp2, 4), m4bexpanded); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15)
+
+ const __m512i rhs_mat_014589CD_02_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_2_sp2, m4bexpanded); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
+ const __m512i rhs_mat_014589CD_12_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2_sp2, 4), m4bexpanded); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
+
+ const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2_sp2, m4bexpanded); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
+ const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2_sp2, 4), m4bexpanded); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
+
+ const __m512i rhs_mat_014589CD_03_sp2 = _mm512_and_si512(rhs_raw_mat_014589CD_3_sp2, m4bexpanded); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
+ const __m512i rhs_mat_014589CD_13_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3_sp2, 4), m4bexpanded); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
+
+ const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3_sp2, m4bexpanded); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
+ const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3_sp2, 4), m4bexpanded); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
@@ -2637,50 +2572,31 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
// Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
- __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb)));
- __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
- __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
- __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb)));
- __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
- __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
- __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb)));
- __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
- __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
- __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb)));
- __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
- __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
- __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb)));
- __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
- __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
- __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb)));
- __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
- __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
- __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb)));
- __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
- __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
- __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb)));
- __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
- __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
-
- //Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into a 512 bit vector
- __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
- __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
- __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
- __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
- __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
- __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
- __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
- __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
-
- __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
- __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
- __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
- __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
- __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
- __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
- __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
- __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
+ __m512i lhs_mat_zmm_0123_00_01 = _mm512_loadu_si512((const __m512i * )((a_ptr[b].qs + sb * 256)));
+ __m512i lhs_mat_zmm_0123_02_03 = _mm512_loadu_si512((const __m512i * )((a_ptr[b].qs + 64 + sb * 256)));
+ __m512i lhs_mat_zmm_0123_10_11 = _mm512_loadu_si512((const __m512i * )((a_ptr[b].qs + 128 + sb * 256)));
+ __m512i lhs_mat_zmm_0123_12_13 = _mm512_loadu_si512((const __m512i * )((a_ptr[b].qs + 192 + sb * 256)));
+
+ // Loaded as set of 128 bit vectors and repeated and stored into a 512 bit vector
+ __m512i lhs_mat_01_00 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b00000000);
+ __m512i lhs_mat_23_00 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b01010101);
+ __m512i lhs_mat_01_01 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b10101010);
+ __m512i lhs_mat_23_01 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_00_01, lhs_mat_zmm_0123_00_01, 0b11111111);
+
+ __m512i lhs_mat_01_02 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b00000000);
+ __m512i lhs_mat_23_02 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b01010101);
+ __m512i lhs_mat_01_03 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b10101010);
+ __m512i lhs_mat_23_03 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_02_03, lhs_mat_zmm_0123_02_03, 0b11111111);
+
+ __m512i lhs_mat_01_10 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b00000000);
+ __m512i lhs_mat_23_10 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b01010101);
+ __m512i lhs_mat_01_11 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b10101010);
+ __m512i lhs_mat_23_11 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_10_11, lhs_mat_zmm_0123_10_11, 0b11111111);
+
+ __m512i lhs_mat_01_12 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b00000000);
+ __m512i lhs_mat_23_12 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b01010101);
+ __m512i lhs_mat_01_13 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b10101010);
+ __m512i lhs_mat_23_13 = _mm512_shuffle_i32x4(lhs_mat_zmm_0123_12_13, lhs_mat_zmm_0123_12_13, 0b11111111);
// Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
__m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb)));
@@ -2812,7 +2728,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
xstart = anc/8;
y = 0;
}
-#endif // __AVX512BW__ && __AVX512DQ__
+#endif // __AVX512F__ && __AVX512BW__ && __AVX512DQ__
// Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
for (; y < anr / 4; y += 4) {
@@ -2863,14 +2779,14 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
- const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
- const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+ const __m256i rhs_raw_mat_0145_0 = _mm256_permute2x128_si256(rhs_raw_mat_0123_0, rhs_raw_mat_4567_0, 0x20); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+ const __m256i rhs_raw_mat_2367_0 = _mm256_permute2x128_si256(rhs_raw_mat_0123_0, rhs_raw_mat_4567_0, 0x31); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+ const __m256i rhs_raw_mat_0145_1 = _mm256_permute2x128_si256(rhs_raw_mat_0123_1, rhs_raw_mat_4567_1, 0x20); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+ const __m256i rhs_raw_mat_2367_1 = _mm256_permute2x128_si256(rhs_raw_mat_0123_1, rhs_raw_mat_4567_1, 0x31); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+ const __m256i rhs_raw_mat_0145_2 = _mm256_permute2x128_si256(rhs_raw_mat_0123_2, rhs_raw_mat_4567_2, 0x20); //B00(16-23) B01(16-23) B04(16-23) B05(16-23)
+ const __m256i rhs_raw_mat_2367_2 = _mm256_permute2x128_si256(rhs_raw_mat_0123_2, rhs_raw_mat_4567_2, 0x31); //B02(16-23) B03(16-23) B06(16-23) B07(16-23)
+ const __m256i rhs_raw_mat_0145_3 = _mm256_permute2x128_si256(rhs_raw_mat_0123_3, rhs_raw_mat_4567_3, 0x20); //B00(24-31) B01(24-31) B04(24-31) B05(24-31)
+ const __m256i rhs_raw_mat_2367_3 = _mm256_permute2x128_si256(rhs_raw_mat_0123_3, rhs_raw_mat_4567_3, 0x31); //B02(24-31) B03(24-31) B06(24-31) B07(24-31)
// 4-bit -> 8-bit
// First sub block of the two sub blocks processed in the iteration
@@ -3031,10 +2947,10 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m256i lhs_mat_23_02_sp1 = _mm256_shuffle_epi32(lhs_mat_23_02, 160); //A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19)
const __m256i lhs_mat_01_03_sp1 = _mm256_shuffle_epi32(lhs_mat_01_03, 160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
- const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27)
+ const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27)
- const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
- const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
+ const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3)
+ const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11)
const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11)
@@ -3155,6 +3071,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
}
}
+
for (; y < nr / 4; y++) {
const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
@@ -3196,14 +3113,14 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
// Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
- const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240);
- const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240);
+ const __m256i rhs_raw_mat_0145_0 = _mm256_permute2x128_si256(rhs_raw_mat_0123_0, rhs_raw_mat_4567_0, 0x20); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+ const __m256i rhs_raw_mat_2367_0 = _mm256_permute2x128_si256(rhs_raw_mat_0123_0, rhs_raw_mat_4567_0, 0x31); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+ const __m256i rhs_raw_mat_0145_1 = _mm256_permute2x128_si256(rhs_raw_mat_0123_1, rhs_raw_mat_4567_1, 0x20); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+ const __m256i rhs_raw_mat_2367_1 = _mm256_permute2x128_si256(rhs_raw_mat_0123_1, rhs_raw_mat_4567_1, 0x31); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+ const __m256i rhs_raw_mat_0145_2 = _mm256_permute2x128_si256(rhs_raw_mat_0123_2, rhs_raw_mat_4567_2, 0x20); //B00(16-23) B01(16-23) B04(16-23) B05(16-23)
+ const __m256i rhs_raw_mat_2367_2 = _mm256_permute2x128_si256(rhs_raw_mat_0123_2, rhs_raw_mat_4567_2, 0x31); //B02(16-23) B03(16-23) B06(16-23) B07(16-23)
+ const __m256i rhs_raw_mat_0145_3 = _mm256_permute2x128_si256(rhs_raw_mat_0123_3, rhs_raw_mat_4567_3, 0x20); //B00(24-31) B01(24-31) B04(24-31) B05(24-31)
+ const __m256i rhs_raw_mat_2367_3 = _mm256_permute2x128_si256(rhs_raw_mat_0123_3, rhs_raw_mat_4567_3, 0x31); //B02(24-31) B03(24-31) B06(24-31) B07(24-31)
// 4-bit -> 8-bit
// First sub block of the two sub blocks processed in the iteration
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index ab82e3d6c..5d13dbea1 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -2026,6 +2026,18 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_fill(params, tensor);
} break;
+ case GGML_OP_SINKHORN:
+ {
+ ggml_compute_forward_sinkhorn(params, tensor);
+ } break;
+ case GGML_OP_HC_MIX:
+ {
+ ggml_compute_forward_hc_mix(params, tensor);
+ } break;
+ case GGML_OP_DSA_SCORE:
+ {
+ ggml_compute_forward_dsa_score(params, tensor);
+ } break;
case GGML_OP_FLASH_ATTN_EXT:
{
ggml_compute_forward_flash_attn_ext(params, tensor);
@@ -2266,6 +2278,9 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_CUMSUM:
case GGML_OP_TRI:
case GGML_OP_FILL:
+ case GGML_OP_SINKHORN:
+ case GGML_OP_HC_MIX:
+ case GGML_OP_DSA_SCORE:
{
n_tasks = n_threads;
} break;
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 7a73e138e..56c864e39 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -2334,6 +2334,202 @@ void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor *
}
}
+// ggml_compute_forward_sinkhorn
+
+// shared by GGML_OP_SINKHORN and GGML_OP_HC_MIX: softmax along dim0, add eps,
+// then alternating dim1/dim0 normalization of the [ne0, ne1] matrix m
+static void ggml_sinkhorn_matrix(float * m, const int64_t ne0, const int64_t ne1, const int n_iter, const float eps) {
+ const int64_t nm = ne0*ne1;
+
+ // softmax along dim0, per i1
+ for (int64_t i1 = 0; i1 < ne1; ++i1) {
+ float * c = m + i1*ne0;
+ float vmax = c[0];
+ for (int64_t i0 = 1; i0 < ne0; ++i0) {
+ vmax = MAX(vmax, c[i0]);
+ }
+ float sum = 0.0f;
+ for (int64_t i0 = 0; i0 < ne0; ++i0) {
+ c[i0] = expf(c[i0] - vmax);
+ sum += c[i0];
+ }
+ // reciprocal multiply to match ggml_soft_max rounding
+ const float inv = 1.0f/sum;
+ for (int64_t i0 = 0; i0 < ne0; ++i0) {
+ c[i0] *= inv;
+ }
+ }
+
+ // add eps to every element
+ for (int64_t i = 0; i < nm; ++i) {
+ m[i] += eps;
+ }
+
+ for (int it = 0; it < n_iter; ++it) {
+ // normalize along dim0, per i1 (skipped on the first pass)
+ if (it > 0) {
+ for (int64_t i1 = 0; i1 < ne1; ++i1) {
+ float sum = eps;
+ for (int64_t i0 = 0; i0 < ne0; ++i0) {
+ sum += m[i0 + i1*ne0];
+ }
+ for (int64_t i0 = 0; i0 < ne0; ++i0) {
+ m[i0 + i1*ne0] /= sum;
+ }
+ }
+ }
+ // normalize along dim1, per i0
+ for (int64_t i0 = 0; i0 < ne0; ++i0) {
+ float sum = eps;
+ for (int64_t i1 = 0; i1 < ne1; ++i1) {
+ sum += m[i0 + i1*ne0];
+ }
+ for (int64_t i1 = 0; i1 < ne1; ++i1) {
+ m[i0 + i1*ne0] /= sum;
+ }
+ }
+ }
+}
+
+void ggml_compute_forward_sinkhorn(const ggml_compute_params * params, ggml_tensor * dst) {
+ const ggml_tensor * src0 = dst->src[0];
+
+ GGML_ASSERT(src0->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(src0));
+ GGML_ASSERT(ggml_is_contiguous(dst));
+
+ const int n_iter = ggml_get_op_params_i32(dst, 0);
+ const float eps = ggml_get_op_params_f32(dst, 1);
+
+ const int64_t ne0 = src0->ne[0];
+ const int64_t ne1 = src0->ne[1];
+ const int64_t nm = ne0*ne1; // per-batch matrix size (<= 64, asserted at build)
+ const int64_t nb = src0->ne[2]*src0->ne[3];
+
+ // parallelize over the (i2, i3) batches
+ const int64_t db = (nb + params->nth - 1)/params->nth;
+ const int64_t b0 = db*params->ith;
+ const int64_t b1 = MIN(b0 + db, nb);
+
+ float m[64];
+
+ for (int64_t b = b0; b < b1; ++b) {
+ const float * sp = (const float *) src0->data + b*nm;
+ float * dp = (float *) dst->data + b*nm;
+
+ memcpy(m, sp, nm*sizeof(float));
+
+ ggml_sinkhorn_matrix(m, ne0, ne1, n_iter, eps);
+
+ memcpy(dp, m, nm*sizeof(float));
+ }
+}
+
+// ggml_compute_forward_hc_mix
+
+void ggml_compute_forward_hc_mix(const ggml_compute_params * params, ggml_tensor * dst) {
+ const ggml_tensor * src0 = dst->src[0]; // mixes [2*hc + hc*hc, nt]
+ const ggml_tensor * scale = dst->src[1]; // [3]
+ const ggml_tensor * base = dst->src[2]; // [2*hc + hc*hc]
+
+ GGML_ASSERT(src0->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(src0));
+ GGML_ASSERT(ggml_is_contiguous(dst));
+
+ const int hc = ggml_get_op_params_i32(dst, 0);
+ const int n_iter = ggml_get_op_params_i32(dst, 1);
+ const float eps = ggml_get_op_params_f32(dst, 2);
+
+ const int64_t n0 = src0->ne[0];
+ const int64_t nt = src0->ne[1]*src0->ne[2]*src0->ne[3];
+
+ GGML_ASSERT(n0 == 2*hc + hc*hc);
+
+ const float * sc = (const float *) scale->data;
+ const float * bs = (const float *) base->data;
+
+ // parallelize over the tokens
+ const int64_t dt = (nt + params->nth - 1)/params->nth;
+ const int64_t t0 = dt*params->ith;
+ const int64_t t1 = MIN(t0 + dt, nt);
+
+ float m[64];
+
+ for (int64_t t = t0; t < t1; ++t) {
+ const float * sp = (const float *) src0->data + t*n0;
+ float * dp = (float *) dst->data + t*n0;
+
+ for (int i = 0; i < hc; ++i) {
+ dp[i] = 1.0f/(1.0f + expf(-(sp[i]*sc[0] + bs[i]))) + eps; // pre
+ }
+ for (int i = hc; i < 2*hc; ++i) {
+ dp[i] = 2.0f/(1.0f + expf(-(sp[i]*sc[1] + bs[i]))); // post
+ }
+ for (int i = 2*hc; i < (int) n0; ++i) {
+ m[i - 2*hc] = sp[i]*sc[2] + bs[i]; // comb
+ }
+
+ ggml_sinkhorn_matrix(m, hc, hc, n_iter, eps);
+
+ memcpy(dp + 2*hc, m, hc*hc*sizeof(float));
+ }
+}
+
+// ggml_compute_forward_dsa_score
+
+void ggml_compute_forward_dsa_score(const ggml_compute_params * params, ggml_tensor * dst) {
+ const ggml_tensor * ik = dst->src[0]; // [d, n_kv] (possibly strided rows)
+ const ggml_tensor * q = dst->src[1]; // [d, h, nt]
+ const ggml_tensor * w = dst->src[2]; // [h, nt]
+ const ggml_tensor * mask = dst->src[3]; // [n_kv, nt]
+
+ const int64_t nd = ik->ne[0];
+ const int64_t nkv = ik->ne[1];
+ const int64_t nh = q->ne[1];
+ const int64_t nt = q->ne[2];
+
+ const float * qd = (const float *) q->data;
+ const float * wd = (const float *) w->data;
+
+ // parallelize over the cached positions
+ const int64_t ds = (nkv + params->nth - 1)/params->nth;
+ const int64_t s0 = ds*params->ith;
+ const int64_t s1 = MIN(s0 + ds, nkv);
+
+ for (int64_t s = s0; s < s1; ++s) {
+ const char * ik_row = (const char *) ik->data + s*ik->nb[1];
+
+ for (int64_t t = 0; t < nt; ++t) {
+ float acc = 0.0f;
+ for (int64_t h = 0; h < nh; ++h) {
+ const float * qh = qd + (t*nh + h)*nd;
+ float dot = 0.0f;
+ if (ik->type == GGML_TYPE_F16) {
+ const ggml_fp16_t * ikd = (const ggml_fp16_t *) ik_row;
+ for (int64_t d = 0; d < nd; ++d) {
+ dot += GGML_CPU_FP16_TO_FP32(ikd[d])*qh[d];
+ }
+ } else {
+ const float * ikd = (const float *) ik_row;
+ for (int64_t d = 0; d < nd; ++d) {
+ dot += ikd[d]*qh[d];
+ }
+ }
+ acc += wd[t*nh + h]*(dot > 0.0f ? dot : 0.0f);
+ }
+
+ float m;
+ if (mask->type == GGML_TYPE_F16) {
+ m = GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) ((const char *) mask->data + t*mask->nb[1]))[s]);
+ } else {
+ m = ((const float *) ((const char *) mask->data + t*mask->nb[1]))[s];
+ }
+
+ ((float *) dst->data)[t*nkv + s] = acc + m;
+ }
+ }
+}
+
// ggml_compute_tri
static void ggml_compute_forward_tri_f32(const ggml_compute_params * params, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
index 2728b08b6..58821c0c7 100644
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -86,6 +86,9 @@ void ggml_compute_forward_top_k(const struct ggml_compute_params * params, struc
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_sinkhorn(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_hc_mix(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_dsa_score(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
diff --git a/ggml/src/ggml-cuda/concat.cu b/ggml/src/ggml-cuda/concat.cu
index f8dcd95ab..7286e76eb 100644
--- a/ggml/src/ggml-cuda/concat.cu
+++ b/ggml/src/ggml-cuda/concat.cu
@@ -1,5 +1,6 @@
#include "concat.cuh"
+#include <algorithm>
#include <stdint.h>
// contiguous kernels
@@ -103,39 +104,45 @@ static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
uint64_t nb12,
uint64_t nb13,
int64_t ne0,
- int64_t /*ne1*/,
- int64_t /*ne2*/,
- int64_t /*ne3*/,
+ int64_t ne1,
+ int64_t ne2,
+ int64_t ne3,
uint64_t nb0,
uint64_t nb1,
uint64_t nb2,
uint64_t nb3) {
static_assert(dim >= 0 && dim <= 3, "dim must be in [0, 3]");
- const int64_t i3 = blockIdx.z;
- const int64_t i2 = blockIdx.y;
- const int64_t i1 = blockIdx.x;
+ // rows are flattened into a grid-stride loop over gridDim.x because
+ // ne1/ne2/ne3 can each exceed the CUDA grid.y/grid.z limit of 65535
+ const int64_t n_rows = ne1*ne2*ne3;
const T * x;
- for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
- if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
- x = (const T *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
- } else {
- if constexpr (dim == 0) {
- x = (const T *)(src1 + i3*nb13 + i2*nb12 + i1*nb11 + (i0 - ne00)*nb10);
- } else if constexpr (dim == 1) {
- x = (const T *)(src1 + i3*nb13 + i2*nb12 + (i1 - ne01)*nb11 + i0*nb10);
- } else if constexpr (dim == 2) {
- x = (const T *)(src1 + i3*nb13 + (i2 - ne02)*nb12 + i1*nb11 + i0*nb10);
- } else if constexpr (dim == 3) {
- x = (const T *)(src1 + (i3 - ne03)*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+ for (int64_t ir = blockIdx.x; ir < n_rows; ir += gridDim.x) {
+ const int64_t i3 = ir / (ne1*ne2);
+ const int64_t i2 = (ir - i3*ne1*ne2) / ne1;
+ const int64_t i1 = ir - i3*ne1*ne2 - i2*ne1;
+
+ for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+ if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+ x = (const T *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+ } else {
+ if constexpr (dim == 0) {
+ x = (const T *)(src1 + i3*nb13 + i2*nb12 + i1*nb11 + (i0 - ne00)*nb10);
+ } else if constexpr (dim == 1) {
+ x = (const T *)(src1 + i3*nb13 + i2*nb12 + (i1 - ne01)*nb11 + i0*nb10);
+ } else if constexpr (dim == 2) {
+ x = (const T *)(src1 + i3*nb13 + (i2 - ne02)*nb12 + i1*nb11 + i0*nb10);
+ } else if constexpr (dim == 3) {
+ x = (const T *)(src1 + (i3 - ne03)*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+ }
}
- }
- T * y = (T *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+ T * y = (T *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
- *y = *x;
+ *y = *x;
+ }
}
}
@@ -163,7 +170,8 @@ static void concat_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml
} else {
GGML_ASSERT(!ggml_is_quantized(src0->type));
- dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
+ const int64_t n_rows = dst->ne[1]*dst->ne[2]*dst->ne[3];
+ dim3 grid_dim((unsigned) std::min<int64_t>(n_rows, INT32_MAX));
auto launch_kernel = [&](auto dim) {
concat_non_cont<T, dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
(const char *) src0->data, (const char *) src1->data, (char *) dst->data,
diff --git a/ggml/src/ggml-cuda/dsa-score.cu b/ggml/src/ggml-cuda/dsa-score.cu
new file mode 100644
index 000000000..bf0a057f4
--- /dev/null
+++ b/ggml/src/ggml-cuda/dsa-score.cu
@@ -0,0 +1,217 @@
+#include "dsa-score.cuh"
+
+// DSA lightning-indexer scores, one pass over the cached indexer keys:
+// out[s,t] = mask[s,t] + sum_h w[h,t] * relu(sum_d ik[d,s] * q[d,h,t])
+// One warp per cached position s: the row load is coalesced across the lanes and
+// the per-head dot products are warp-shuffle reductions. q and w for the block's
+// token are staged in shared memory (d*h floats, <= 12 KB for the openPangu
+// 128x24 indexer). Used on the decode path (nt small); prefill keeps the tiled
+// matmul pipeline.
+template <typename T_IK, typename T_MASK>
+static __global__ void dsa_score_f32(
+ const char * ik, const float * q, const float * w, const char * mask, float * dst,
+ const int nd, const int64_t nkv, const int nh, const int64_t nb_ik_row, const int64_t nb_mask_row) {
+ extern __shared__ float smem[]; // [nd*nh + nh]
+
+ const int t = blockIdx.y;
+ const int warp = threadIdx.x/WARP_SIZE;
+ const int lane = threadIdx.x%WARP_SIZE;
+
+ const int warps_per_block = blockDim.x/WARP_SIZE;
+ const int64_t s = (int64_t) blockIdx.x*warps_per_block + warp;
+
+ float * q_s = smem;
+ float * w_s = smem + (size_t) nd*nh;
+
+ for (int i = threadIdx.x; i < nd*nh; i += blockDim.x) {
+ q_s[i] = q[(size_t) t*nd*nh + i];
+ }
+ for (int i = threadIdx.x; i < nh; i += blockDim.x) {
+ w_s[i] = w[(size_t) t*nh + i];
+ }
+ __syncthreads();
+
+ if (s >= nkv) {
+ return;
+ }
+
+ const T_IK * ik_row = (const T_IK *) (ik + s*nb_ik_row);
+
+ // coalesced row load: each lane keeps its d-strided slice in registers
+ float ikl[8]; // nd <= 8*WARP_SIZE (256), asserted host-side
+ const int per_lane = (nd + WARP_SIZE - 1)/WARP_SIZE;
+ for (int i = 0; i < per_lane; ++i) {
+ const int d = lane + i*WARP_SIZE;
+ ikl[i] = d < nd ? (float) ik_row[d] : 0.0f;
+ }
+
+ float acc = 0.0f;
+ for (int h = 0; h < nh; ++h) {
+ const float * qh = q_s + h*nd;
+ float part = 0.0f;
+ for (int i = 0; i < per_lane; ++i) {
+ const int d = lane + i*WARP_SIZE;
+ if (d < nd) {
+ part += ikl[i]*qh[d];
+ }
+ }
+ part = warp_reduce_sum(part);
+ acc += w_s[h]*fmaxf(part, 0.0f);
+ }
+
+ if (lane == 0) {
+ const T_MASK * mask_row = (const T_MASK *) (mask + t*nb_mask_row);
+ dst[(size_t) t*nkv + s] = acc + (float) mask_row[s];
+ }
+}
+
+// Prefill variant: GEMM-like tiled kernel for large token counts. Each block
+// computes a DSA_TILE_S x DSA_TILE_T output tile; the (transposed) ik tile is
+// staged in shared memory once and reused across all heads, the per-head q tile
+// is restaged each head iteration. Each thread accumulates a 4x2 register tile,
+// so the shared-memory loads amortize over 8 FMAs. Never materializes the
+// [n_kv, nh, nt] intermediate the unfused matmul pipeline needs (~5 GB per
+// layer per 512-token ubatch at 100K context).
+#define DSA_TILE_S 64
+#define DSA_TILE_T 32
+#define DSA_TILE_PS (DSA_TILE_S + 4) // padded rows: keeps 8-byte alignment, avoids bank conflicts
+#define DSA_TILE_PT (DSA_TILE_T + 2)
+#define DSA_TILE_THREADS 256
+#define DSA_TILE_ND_MAX 180 // smem = nd*(PS*2 + PT*4) bytes <= 48 KB
+
+template <typename T_MASK>
+static __global__ void dsa_score_tiled_f32(
+ const char * ik, const float * q, const float * w, const char * mask, float * dst,
+ const int nd, const int64_t nkv, const int nh, const int nt,
+ const int64_t nb_ik_row, const int64_t nb_mask_row) {
+ extern __shared__ char smem_raw[];
+ half * sik = (half *) smem_raw; // [nd][DSA_TILE_PS], transposed
+ float * sq = (float *) (smem_raw + (size_t) nd*DSA_TILE_PS*sizeof(half)); // [nd][DSA_TILE_PT], transposed
+
+ const int64_t s0 = (int64_t) blockIdx.x*DSA_TILE_S;
+ const int t0 = blockIdx.y*DSA_TILE_T;
+
+ const int tid = threadIdx.x;
+ const int tt = tid % (DSA_TILE_T/2); // covers t = t0 + tt*2 + {0,1}
+ const int ts = tid / (DSA_TILE_T/2); // covers s = s0 + ts*4 + {0..3}
+
+ // ik tile: coalesced global reads (d contiguous per row), transposed store
+ for (int idx = tid; idx < nd*DSA_TILE_S; idx += DSA_TILE_THREADS) {
+ const int d = idx % nd;
+ const int s = idx / nd;
+ sik[d*DSA_TILE_PS + s] = s0 + s < nkv ? ((const half *) (ik + (s0 + s)*nb_ik_row))[d] : __float2half(0.0f);
+ }
+
+ float acc[4][2] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
+
+ for (int h = 0; h < nh; ++h) {
+ __syncthreads(); // also covers the sik stores on the first iteration
+ for (int idx = tid; idx < nd*DSA_TILE_T; idx += DSA_TILE_THREADS) {
+ const int d = idx % nd;
+ const int t = idx / nd;
+ sq[d*DSA_TILE_PT + t] = t0 + t < nt ? q[((size_t) (t0 + t)*nh + h)*nd + d] : 0.0f;
+ }
+ __syncthreads();
+
+ float dot[4][2] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
+ for (int d = 0; d < nd; ++d) {
+ const half2 * a = (const half2 *) (sik + d*DSA_TILE_PS + ts*4);
+ const float2 a01 = __half22float2(a[0]);
+ const float2 a23 = __half22float2(a[1]);
+ const float2 b = *(const float2 *) (sq + d*DSA_TILE_PT + tt*2);
+ dot[0][0] += a01.x*b.x; dot[0][1] += a01.x*b.y;
+ dot[1][0] += a01.y*b.x; dot[1][1] += a01.y*b.y;
+ dot[2][0] += a23.x*b.x; dot[2][1] += a23.x*b.y;
+ dot[3][0] += a23.y*b.x; dot[3][1] += a23.y*b.y;
+ }
+
+#pragma unroll
+ for (int j = 0; j < 2; ++j) {
+ const int t = t0 + tt*2 + j;
+ const float wv = t < nt ? w[(size_t) t*nh + h] : 0.0f;
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ acc[i][j] += wv*fmaxf(dot[i][j], 0.0f);
+ }
+ }
+ }
+
+#pragma unroll
+ for (int j = 0; j < 2; ++j) {
+ const int t = t0 + tt*2 + j;
+ if (t >= nt) {
+ continue;
+ }
+ const T_MASK * mask_row = (const T_MASK *) (mask + (size_t) t*nb_mask_row);
+#pragma unroll
+ for (int i = 0; i < 4; ++i) {
+ const int64_t s = s0 + ts*4 + i;
+ if (s < nkv) {
+ dst[(size_t) t*nkv + s] = acc[i][j] + (float) mask_row[s];
+ }
+ }
+ }
+}
+
+void ggml_cuda_op_dsa_score(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+ const ggml_tensor * ik = dst->src[0];
+ const ggml_tensor * q = dst->src[1];
+ const ggml_tensor * w = dst->src[2];
+ const ggml_tensor * mask = dst->src[3];
+
+ GGML_ASSERT(dst->type == GGML_TYPE_F32);
+ GGML_ASSERT(q->type == GGML_TYPE_F32 && w->type == GGML_TYPE_F32);
+
+ const int nd = (int) ik->ne[0];
+ const int64_t nkv = ik->ne[1];
+ const int nh = (int) q->ne[1];
+ const int nt = (int) q->ne[2];
+
+ GGML_ASSERT(nd <= 8*WARP_SIZE);
+
+ // large token counts (prefill): tiled kernel, one output tile per block
+ if (nt >= 16 && ik->type == GGML_TYPE_F16 && nd <= DSA_TILE_ND_MAX) {
+ const dim3 grid((nkv + DSA_TILE_S - 1)/DSA_TILE_S, (nt + DSA_TILE_T - 1)/DSA_TILE_T, 1);
+ const size_t smem = (size_t) nd*(DSA_TILE_PS*sizeof(half) + DSA_TILE_PT*sizeof(float));
+
+ if (mask->type == GGML_TYPE_F16) {
+ dsa_score_tiled_f32<half><<<grid, DSA_TILE_THREADS, smem, ctx.stream()>>>(
+ (const char *) ik->data, (const float *) q->data, (const float *) w->data,
+ (const char *) mask->data, (float *) dst->data,
+ nd, nkv, nh, nt, ik->nb[1], mask->nb[1]);
+ } else {
+ dsa_score_tiled_f32<float><<<grid, DSA_TILE_THREADS, smem, ctx.stream()>>>(
+ (const char *) ik->data, (const float *) q->data, (const float *) w->data,
+ (const char *) mask->data, (float *) dst->data,
+ nd, nkv, nh, nt, ik->nb[1], mask->nb[1]);
+ }
+ return;
+ }
+
+ const int warps_per_block = CUDA_DSA_SCORE_BLOCK_SIZE/WARP_SIZE;
+ const dim3 grid((nkv + warps_per_block - 1)/warps_per_block, nt, 1);
+ const size_t smem = ((size_t) nd*nh + nh)*sizeof(float);
+
+ const auto launch = [&](auto ik_dummy, auto mask_dummy) {
+ using T_IK = decltype(ik_dummy);
+ using T_MASK = decltype(mask_dummy);
+ dsa_score_f32<T_IK, T_MASK><<<grid, CUDA_DSA_SCORE_BLOCK_SIZE, smem, ctx.stream()>>>(
+ (const char *) ik->data, (const float *) q->data, (const float *) w->data,
+ (const char *) mask->data, (float *) dst->data,
+ nd, nkv, nh, ik->nb[1], mask->nb[1]);
+ };
+
+ if (ik->type == GGML_TYPE_F16) {
+ if (mask->type == GGML_TYPE_F16) {
+ launch(half{}, half{});
+ } else {
+ launch(half{}, float{});
+ }
+ } else {
+ if (mask->type == GGML_TYPE_F16) {
+ launch(float{}, half{});
+ } else {
+ launch(float{}, float{});
+ }
+ }
+}
diff --git a/ggml/src/ggml-cuda/dsa-score.cuh b/ggml/src/ggml-cuda/dsa-score.cuh
new file mode 100644
index 000000000..564a4ea92
--- /dev/null
+++ b/ggml/src/ggml-cuda/dsa-score.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_DSA_SCORE_BLOCK_SIZE 128
+
+void ggml_cuda_op_dsa_score(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 7c45d0d1f..b835759f7 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -70,6 +70,8 @@
#include "ggml-cuda/cumsum.cuh"
#include "ggml-cuda/fill.cuh"
#include "ggml-cuda/lightning-indexer.cuh"
+#include "ggml-cuda/sinkhorn.cuh"
+#include "ggml-cuda/dsa-score.cuh"
#include "ggml.h"
#include <algorithm>
@@ -2542,6 +2544,15 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_OP_LIGHTNING_INDEXER:
ggml_cuda_lightning_indexer(ctx, dst);
break;
+ case GGML_OP_SINKHORN:
+ ggml_cuda_op_sinkhorn(ctx, dst);
+ break;
+ case GGML_OP_HC_MIX:
+ ggml_cuda_op_hc_mix(ctx, dst);
+ break;
+ case GGML_OP_DSA_SCORE:
+ ggml_cuda_op_dsa_score(ctx, dst);
+ break;
default:
return false;
}
@@ -5874,6 +5885,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_OPT_STEP_ADAMW:
case GGML_OP_OPT_STEP_SGD:
case GGML_OP_FILL:
+ case GGML_OP_SINKHORN:
+ case GGML_OP_HC_MIX:
+ case GGML_OP_DSA_SCORE:
case GGML_OP_CUMSUM:
case GGML_OP_TRI:
case GGML_OP_DIAG:
diff --git a/ggml/src/ggml-cuda/sinkhorn.cu b/ggml/src/ggml-cuda/sinkhorn.cu
new file mode 100644
index 000000000..66add8b91
--- /dev/null
+++ b/ggml/src/ggml-cuda/sinkhorn.cu
@@ -0,0 +1,160 @@
+#include "sinkhorn.cuh"
+
+// Shared by GGML_OP_SINKHORN and GGML_OP_HC_MIX: softmax along dim0, add eps,
+// then alternating dim1/dim0 normalization of the [ne0, ne1] matrix m.
+// The matrices are tiny (hyper-connection stream mixing, ne0*ne1 <= 64), so one
+// thread handles one matrix entirely in local memory. This replaces the
+// ~8 ops x 2 x n_iter graph nodes the unfused form needs per matrix.
+static __device__ void sinkhorn_matrix(float * m, const int ne0, const int ne1, const int n_iter, const float eps) {
+ const int nm = ne0*ne1;
+
+ // softmax along dim0, per i1
+ for (int i1 = 0; i1 < ne1; ++i1) {
+ float * c = m + i1*ne0;
+ float vmax = c[0];
+ for (int i0 = 1; i0 < ne0; ++i0) {
+ vmax = fmaxf(vmax, c[i0]);
+ }
+ float sum = 0.0f;
+ for (int i0 = 0; i0 < ne0; ++i0) {
+ c[i0] = expf(c[i0] - vmax);
+ sum += c[i0];
+ }
+ // reciprocal multiply to match ggml_soft_max rounding
+ const float inv = 1.0f/sum;
+ for (int i0 = 0; i0 < ne0; ++i0) {
+ c[i0] *= inv;
+ }
+ }
+
+ // add eps to every element
+ for (int i = 0; i < nm; ++i) {
+ m[i] += eps;
+ }
+
+ for (int it = 0; it < n_iter; ++it) {
+ // normalize along dim0, per i1 (skipped on the first pass)
+ if (it > 0) {
+ for (int i1 = 0; i1 < ne1; ++i1) {
+ float sum = eps;
+ for (int i0 = 0; i0 < ne0; ++i0) {
+ sum += m[i0 + i1*ne0];
+ }
+ for (int i0 = 0; i0 < ne0; ++i0) {
+ m[i0 + i1*ne0] /= sum;
+ }
+ }
+ }
+ // normalize along dim1, per i0
+ for (int i0 = 0; i0 < ne0; ++i0) {
+ float sum = eps;
+ for (int i1 = 0; i1 < ne1; ++i1) {
+ sum += m[i0 + i1*ne0];
+ }
+ for (int i1 = 0; i1 < ne1; ++i1) {
+ m[i0 + i1*ne0] /= sum;
+ }
+ }
+ }
+}
+
+static __global__ void sinkhorn_f32(const float * x, float * dst,
+ const int ne0, const int ne1, const int64_t nb, const int n_iter, const float eps) {
+ const int64_t b = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+ if (b >= nb) {
+ return;
+ }
+
+ const int nm = ne0*ne1;
+
+ float m[64];
+
+ const float * sp = x + b*nm;
+ float * dp = dst + b*nm;
+
+ for (int i = 0; i < nm; ++i) {
+ m[i] = sp[i];
+ }
+
+ sinkhorn_matrix(m, ne0, ne1, n_iter, eps);
+
+ for (int i = 0; i < nm; ++i) {
+ dp[i] = m[i];
+ }
+}
+
+static __global__ void hc_mix_f32(const float * x, const float * scale, const float * base, float * dst,
+ const int hc, const int64_t nt, const int n_iter, const float eps) {
+ const int64_t t = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+ if (t >= nt) {
+ return;
+ }
+
+ const int n0 = 2*hc + hc*hc;
+
+ float m[64];
+
+ const float * sp = x + t*n0;
+ float * dp = dst + t*n0;
+
+ for (int i = 0; i < hc; ++i) {
+ dp[i] = 1.0f/(1.0f + expf(-(sp[i]*scale[0] + base[i]))) + eps; // pre
+ }
+ for (int i = hc; i < 2*hc; ++i) {
+ dp[i] = 2.0f/(1.0f + expf(-(sp[i]*scale[1] + base[i]))); // post
+ }
+ for (int i = 2*hc; i < n0; ++i) {
+ m[i - 2*hc] = sp[i]*scale[2] + base[i]; // comb
+ }
+
+ sinkhorn_matrix(m, hc, hc, n_iter, eps);
+
+ for (int i = 0; i < hc*hc; ++i) {
+ dp[2*hc + i] = m[i];
+ }
+}
+
+void ggml_cuda_op_sinkhorn(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+ const ggml_tensor * src0 = dst->src[0];
+
+ GGML_ASSERT(src0->type == GGML_TYPE_F32);
+ GGML_ASSERT( dst->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(src0));
+ GGML_ASSERT(src0->ne[0]*src0->ne[1] <= 64);
+
+ const int n_iter = ggml_get_op_params_i32(dst, 0);
+ const float eps = ggml_get_op_params_f32(dst, 1);
+
+ const int64_t nb = src0->ne[2]*src0->ne[3];
+
+ const int64_t num_blocks = (nb + CUDA_SINKHORN_BLOCK_SIZE - 1)/CUDA_SINKHORN_BLOCK_SIZE;
+
+ sinkhorn_f32<<<num_blocks, CUDA_SINKHORN_BLOCK_SIZE, 0, ctx.stream()>>>(
+ (const float *) src0->data, (float *) dst->data,
+ (int) src0->ne[0], (int) src0->ne[1], nb, n_iter, eps);
+}
+
+void ggml_cuda_op_hc_mix(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+ const ggml_tensor * src0 = dst->src[0];
+ const ggml_tensor * scale = dst->src[1];
+ const ggml_tensor * base = dst->src[2];
+
+ GGML_ASSERT(src0->type == GGML_TYPE_F32);
+ GGML_ASSERT( dst->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(src0));
+
+ const int hc = ggml_get_op_params_i32(dst, 0);
+ const int n_iter = ggml_get_op_params_i32(dst, 1);
+ const float eps = ggml_get_op_params_f32(dst, 2);
+
+ GGML_ASSERT(src0->ne[0] == 2*hc + hc*hc);
+ GGML_ASSERT(hc*hc <= 64);
+
+ const int64_t nt = src0->ne[1]*src0->ne[2]*src0->ne[3];
+
+ const int64_t num_blocks = (nt + CUDA_SINKHORN_BLOCK_SIZE - 1)/CUDA_SINKHORN_BLOCK_SIZE;
+
+ hc_mix_f32<<<num_blocks, CUDA_SINKHORN_BLOCK_SIZE, 0, ctx.stream()>>>(
+ (const float *) src0->data, (const float *) scale->data, (const float *) base->data,
+ (float *) dst->data, hc, nt, n_iter, eps);
+}
diff --git a/ggml/src/ggml-cuda/sinkhorn.cuh b/ggml/src/ggml-cuda/sinkhorn.cuh
new file mode 100644
index 000000000..8c1bcc33d
--- /dev/null
+++ b/ggml/src/ggml-cuda/sinkhorn.cuh
@@ -0,0 +1,6 @@
+#include "common.cuh"
+
+#define CUDA_SINKHORN_BLOCK_SIZE 128
+
+void ggml_cuda_op_sinkhorn(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_hc_mix(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/top-k.cu b/ggml/src/ggml-cuda/top-k.cu
index b0f609990..ea30a0204 100644
--- a/ggml/src/ggml-cuda/top-k.cu
+++ b/ggml/src/ggml-cuda/top-k.cu
@@ -46,6 +46,94 @@ using namespace cub;
# endif
#endif // GGML_CUDA_TOP_K_NROWS_THRESHOLD
+// map float bits to a monotonically ordered uint (larger float <=> larger uint)
+static __device__ __forceinline__ uint32_t top_k_float_to_ordered(const float f) {
+ const uint32_t u = __float_as_uint(f);
+ return (u & 0x80000000u) ? ~u : (u | 0x80000000u);
+}
+
+// radix-select top-k, one block per row: 4 rounds of 256-bin byte histograms
+// (MSB first) narrow down the exact bit pattern of the k-th largest value, then
+// a single compaction pass collects the winning indices. O(n) passes instead of
+// a full O(n log n) sort -- at [102400, 512] this is ~5 reads of the row data
+// vs a full segmented radix sort. Output order is unspecified, which matches
+// the GGML_OP_TOP_K contract.
+#define CUDA_TOP_K_RADIX_BLOCK_SIZE 512
+
+static __global__ void top_k_radix_select_f32(const float * src, int * dst, const int ncols, const int k) {
+ const float * x = src + (size_t) blockIdx.x*ncols;
+ int * out = dst + (size_t) blockIdx.x*k;
+
+ __shared__ int hist[256];
+ __shared__ uint32_t s_prefix;
+ __shared__ int s_remaining; // elements of the current prefix group still needed
+ __shared__ int s_pos_gt, s_pos_eq;
+
+ const int tid = threadIdx.x;
+ if (tid == 0) {
+ s_prefix = 0;
+ s_remaining = k;
+ }
+
+ for (int round = 0; round < 4; ++round) {
+ const int shift = (3 - round)*8;
+
+ for (int i = tid; i < 256; i += blockDim.x) {
+ hist[i] = 0;
+ }
+ __syncthreads();
+
+ const uint32_t prefix = s_prefix;
+ const uint32_t prefix_mask = round == 0 ? 0 : 0xFFFFFFFFu << (shift + 8);
+ for (int i = tid; i < ncols; i += blockDim.x) {
+ const uint32_t u = top_k_float_to_ordered(x[i]);
+ if ((u & prefix_mask) == prefix) {
+ const int bin = (u >> shift) & 255;
+ // warp-aggregate equal bins before hitting shared memory: masked-out
+ // rows produce millions of identical -inf entries otherwise
+ const unsigned int peers = __match_any_sync(__activemask(), bin);
+ if ((__ffs(peers) - 1) == (int) (threadIdx.x % WARP_SIZE)) {
+ atomicAdd(&hist[bin], __popc(peers));
+ }
+ }
+ }
+ __syncthreads();
+
+ if (tid == 0) {
+ int rem = s_remaining;
+ int b = 255;
+ while (b > 0 && hist[b] < rem) {
+ rem -= hist[b];
+ --b;
+ }
+ s_prefix = prefix | ((uint32_t) b << shift);
+ s_remaining = rem;
+ }
+ __syncthreads();
+ }
+
+ const uint32_t T = s_prefix; // ordered bits of the k-th largest value
+ const int n_gt = k - s_remaining; // elements strictly greater than T
+
+ if (tid == 0) {
+ s_pos_gt = 0;
+ s_pos_eq = 0;
+ }
+ __syncthreads();
+
+ for (int i = tid; i < ncols; i += blockDim.x) {
+ const uint32_t u = top_k_float_to_ordered(x[i]);
+ if (u > T) {
+ out[atomicAdd(&s_pos_gt, 1)] = i;
+ } else if (u == T) {
+ const int slot = atomicAdd(&s_pos_eq, 1);
+ if (slot < k - n_gt) {
+ out[n_gt + slot] = i;
+ }
+ }
+ }
+}
+
#ifdef CUB_TOP_K_AVAILABLE
static void top_k_cub(ggml_cuda_pool & pool,
@@ -296,6 +384,15 @@ void ggml_cuda_op_top_k(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int device = ggml_cuda_get_device();
+ // large batched rows: radix select, O(n) instead of a full sort (order is
+ // unspecified). single rows stay on the sort paths -- one block per row
+ // leaves the GPU idle at nrows == 1
+ if (ncols >= 4096 && nrows > 1) {
+ top_k_radix_select_f32<<<nrows, CUDA_TOP_K_RADIX_BLOCK_SIZE, 0, stream>>>(
+ src0_d, dst_d, ncols, k);
+ return;
+ }
+
#ifdef CUB_TOP_K_AVAILABLE
// a single row always uses DeviceTopK if available
const bool bitonic_short = nrows > 1 && ncols <= GGML_CUDA_TOP_K_NCOLS_THRESHOLD_BITONIC;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 01dc8d05c..408d39f99 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1077,6 +1077,9 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"LEAKY_RELU",
"TRI",
"FILL",
+ "SINKHORN",
+ "HC_MIX",
+ "DSA_SCORE",
"FLASH_ATTN_EXT",
"FLASH_ATTN_BACK",
@@ -1112,7 +1115,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
-static_assert(GGML_OP_COUNT == 101, "GGML_OP_COUNT != 101");
+static_assert(GGML_OP_COUNT == 104, "GGML_OP_COUNT != 104");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1192,6 +1195,9 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"leaky_relu(x)",
"tri(x)",
"fill(x, c)",
+ "sinkhorn(x)",
+ "hc_mix(x)",
+ "dsa_score(x)",
"flash_attn_ext(x)",
"flash_attn_back(x)",
@@ -1227,7 +1233,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
-static_assert(GGML_OP_COUNT == 101, "GGML_OP_COUNT != 101");
+static_assert(GGML_OP_COUNT == 104, "GGML_OP_COUNT != 104");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -3355,8 +3361,8 @@ static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return (t0->ne[0] == t1->ne[0]) &&
- (t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
- (t1->ne[3]%t0->ne[3] == 0);
+ (t0->ne[2] == 0 || t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
+ (t0->ne[3] == 0 || t1->ne[3]%t0->ne[3] == 0);
}
struct ggml_tensor * ggml_mul_mat(
@@ -3442,8 +3448,8 @@ static inline bool ggml_can_out_prod(const struct ggml_tensor * t0, const struct
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return (t0->ne[1] == t1->ne[1]) &&
- (t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
- (t1->ne[3]%t0->ne[3] == 0);
+ (t0->ne[2] == 0 || t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
+ (t0->ne[3] == 0 || t1->ne[3]%t0->ne[3] == 0);
}
struct ggml_tensor * ggml_out_prod(
@@ -5440,6 +5446,93 @@ struct ggml_tensor * ggml_fill_inplace(
return ggml_fill_impl(ctx, a, c, true);
}
+// ggml_sinkhorn
+
+struct ggml_tensor * ggml_sinkhorn(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ int n_iter,
+ float eps) {
+ GGML_ASSERT(a->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(a));
+ GGML_ASSERT(n_iter >= 1);
+ // the per-batch matrix is kept in registers/stack by the backends
+ GGML_ASSERT(a->ne[0]*a->ne[1] <= 64);
+
+ struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+ ggml_set_op_params_i32(result, 0, n_iter);
+ ggml_set_op_params_f32(result, 1, eps);
+
+ result->op = GGML_OP_SINKHORN;
+ result->src[0] = a;
+
+ return result;
+}
+
+// ggml_hc_mix
+
+struct ggml_tensor * ggml_hc_mix(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ struct ggml_tensor * scale,
+ struct ggml_tensor * base,
+ int hc,
+ int n_iter,
+ float eps) {
+ GGML_ASSERT(a->type == GGML_TYPE_F32 && scale->type == GGML_TYPE_F32 && base->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(a));
+ GGML_ASSERT(a->ne[0] == 2*hc + hc*hc);
+ GGML_ASSERT(ggml_nelements(scale) == 3);
+ GGML_ASSERT(ggml_nelements(base) == a->ne[0]);
+ GGML_ASSERT(n_iter >= 1);
+ GGML_ASSERT(hc*hc <= 64);
+
+ struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+ ggml_set_op_params_i32(result, 0, hc);
+ ggml_set_op_params_i32(result, 1, n_iter);
+ ggml_set_op_params_f32(result, 2, eps);
+
+ result->op = GGML_OP_HC_MIX;
+ result->src[0] = a;
+ result->src[1] = scale;
+ result->src[2] = base;
+
+ return result;
+}
+
+// ggml_dsa_score
+
+struct ggml_tensor * ggml_dsa_score(
+ struct ggml_context * ctx,
+ struct ggml_tensor * ik,
+ struct ggml_tensor * q,
+ struct ggml_tensor * w,
+ struct ggml_tensor * mask) {
+ GGML_ASSERT(ik->type == GGML_TYPE_F16 || ik->type == GGML_TYPE_F32);
+ GGML_ASSERT(q->type == GGML_TYPE_F32 && ggml_is_contiguous(q));
+ GGML_ASSERT(w->type == GGML_TYPE_F32 && ggml_is_contiguous(w));
+ GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
+ GGML_ASSERT(ggml_is_contiguous(mask));
+
+ GGML_ASSERT(ik->ne[0] == q->ne[0]); // d
+ GGML_ASSERT(w->ne[0] == q->ne[1]); // h
+ GGML_ASSERT(w->ne[1] == q->ne[2]); // nt
+ GGML_ASSERT(mask->ne[0] == ik->ne[1]); // n_kv
+ GGML_ASSERT(mask->ne[1] >= q->ne[2]);
+
+ struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ik->ne[1], q->ne[2]);
+
+ result->op = GGML_OP_DSA_SCORE;
+ result->src[0] = ik;
+ result->src[1] = q;
+ result->src[2] = w;
+ result->src[3] = mask;
+
+ return result;
+}
+
// ggml_argsort
struct ggml_tensor * ggml_argsort(
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index a39d046fe..2dd64902f 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -224,6 +224,9 @@ class Keys:
ROPE_PATTERN = "{arch}.attention.rope_pattern"
VALUE_EXPERT_COUNT = "{arch}.attention.value_expert_count"
VALUE_EXPERT_USED_COUNT = "{arch}.attention.value_expert_used_count"
+ SINK_COUNT = "{arch}.attention.sink_count"
+ CONV_KERNEL_SIZE = "{arch}.attention.conv_kernel_size"
+ LAYER_SWA_WINDOW = "{arch}.attention.layer_swa_window"
class Indexer:
HEAD_COUNT = "{arch}.attention.indexer.head_count"
@@ -674,6 +677,7 @@ class MODEL_ARCH(IntEnum):
MINIMAXM3 = auto()
RND1 = auto()
PANGU_EMBED = auto()
+ OPENPANGU_V2 = auto()
MISTRAL3 = auto()
EAGLE3 = auto()
DFLASH = auto()
@@ -887,6 +891,12 @@ class MODEL_TENSOR(IntEnum):
PLE_NORM_QUERY = auto() # qwen4exp
PLE_NORM_CONV = auto() # qwen4exp
PLE_CONV1D = auto() # qwen4exp
+ ATTN_Q_A_CONV = auto() # openpangu-v2 MoME short conv
+ ATTN_KV_A_CONV = auto()
+ ATTN_O_CONV = auto()
+ ATTN_SINK_KV = auto() # openpangu-v2 learned param-sink compressed kv
+ ATTN_SINK_K_PE = auto()
+ BLOCK_POST_NORM = auto() # openpangu-v2 mHC stream post-norm
ATTN_COMPRESSOR_WKV = auto()
ATTN_COMPRESSOR_WGATE = auto()
ATTN_COMPRESSOR_APE = auto()
@@ -1479,6 +1489,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.COGVLM: "cogvlm",
MODEL_ARCH.RND1: "rnd1",
MODEL_ARCH.PANGU_EMBED: "pangu-embedded",
+ MODEL_ARCH.OPENPANGU_V2: "openpangu-v2",
MODEL_ARCH.MISTRAL3: "mistral3",
MODEL_ARCH.EAGLE3: "eagle3",
MODEL_ARCH.DFLASH: "dflash",
@@ -1690,6 +1701,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.PLE_NORM_QUERY: "blk.{bid}.ple_norm_query", # qwen4exp
MODEL_TENSOR.PLE_NORM_CONV: "blk.{bid}.ple_norm_conv", # qwen4exp
MODEL_TENSOR.PLE_CONV1D: "blk.{bid}.ple_conv1d", # qwen4exp
+ MODEL_TENSOR.ATTN_Q_A_CONV: "blk.{bid}.attn_q_a_conv",
+ MODEL_TENSOR.ATTN_KV_A_CONV: "blk.{bid}.attn_kv_a_conv",
+ MODEL_TENSOR.ATTN_O_CONV: "blk.{bid}.attn_o_conv",
+ MODEL_TENSOR.ATTN_SINK_KV: "blk.{bid}.attn_sink_kv",
+ MODEL_TENSOR.ATTN_SINK_K_PE: "blk.{bid}.attn_sink_k_pe",
+ MODEL_TENSOR.BLOCK_POST_NORM: "blk.{bid}.block_post_norm",
MODEL_TENSOR.ATTN_COMPRESSOR_WKV: "blk.{bid}.attn_compressor_kv",
MODEL_TENSOR.ATTN_COMPRESSOR_WGATE: "blk.{bid}.attn_compressor_gate",
MODEL_TENSOR.ATTN_COMPRESSOR_APE: "blk.{bid}.attn_compressor_ape",
@@ -4201,6 +4218,59 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
],
+ MODEL_ARCH.OPENPANGU_V2: [
+ MODEL_TENSOR.TOKEN_EMBD,
+ MODEL_TENSOR.OUTPUT_NORM,
+ MODEL_TENSOR.OUTPUT,
+ MODEL_TENSOR.HC_HEAD_FN,
+ MODEL_TENSOR.HC_HEAD_BASE,
+ MODEL_TENSOR.HC_HEAD_SCALE,
+ MODEL_TENSOR.ATTN_NORM,
+ MODEL_TENSOR.ATTN_POST_NORM,
+ MODEL_TENSOR.BLOCK_POST_NORM,
+ MODEL_TENSOR.HC_ATTN_FN,
+ MODEL_TENSOR.HC_ATTN_BASE,
+ MODEL_TENSOR.HC_ATTN_SCALE,
+ MODEL_TENSOR.HC_FFN_FN,
+ MODEL_TENSOR.HC_FFN_BASE,
+ MODEL_TENSOR.HC_FFN_SCALE,
+ MODEL_TENSOR.ATTN_Q_A,
+ MODEL_TENSOR.ATTN_Q_A_NORM,
+ MODEL_TENSOR.ATTN_Q_B,
+ MODEL_TENSOR.ATTN_Q_A_CONV,
+ MODEL_TENSOR.ATTN_KV_A_MQA,
+ MODEL_TENSOR.ATTN_KV_A_NORM,
+ MODEL_TENSOR.ATTN_KV_A_CONV,
+ MODEL_TENSOR.ATTN_K_B,
+ MODEL_TENSOR.ATTN_V_B,
+ MODEL_TENSOR.ATTN_O_CONV,
+ MODEL_TENSOR.ATTN_OUT,
+ MODEL_TENSOR.ATTN_SINK_KV,
+ MODEL_TENSOR.ATTN_SINK_K_PE,
+ MODEL_TENSOR.INDEXER_K_NORM,
+ MODEL_TENSOR.INDEXER_PROJ,
+ MODEL_TENSOR.INDEXER_ATTN_K,
+ MODEL_TENSOR.INDEXER_ATTN_Q_B,
+ MODEL_TENSOR.FFN_GATE_INP,
+ MODEL_TENSOR.FFN_EXP_PROBS_B,
+ MODEL_TENSOR.FFN_NORM,
+ MODEL_TENSOR.FFN_POST_NORM,
+ MODEL_TENSOR.FFN_GATE,
+ MODEL_TENSOR.FFN_DOWN,
+ MODEL_TENSOR.FFN_UP,
+ MODEL_TENSOR.FFN_GATE_EXP,
+ MODEL_TENSOR.FFN_DOWN_EXP,
+ MODEL_TENSOR.FFN_UP_EXP,
+ MODEL_TENSOR.FFN_GATE_SHEXP,
+ MODEL_TENSOR.FFN_DOWN_SHEXP,
+ MODEL_TENSOR.FFN_UP_SHEXP,
+ MODEL_TENSOR.NEXTN_EH_PROJ,
+ MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+ MODEL_TENSOR.NEXTN_ENORM,
+ MODEL_TENSOR.NEXTN_HNORM,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+ MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+ ],
MODEL_ARCH.ERNIE4_5_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index cc2bd8e7b..d1bc15bba 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -837,6 +837,15 @@ class GGUFWriter:
def add_indexer_kpool_select_tail(self, value: bool) -> None:
self.add_bool(Keys.Attention.Indexer.KPOOL_SELECT_TAIL.format(arch=self.arch), value)
+ def add_sink_count(self, count: int) -> None:
+ self.add_uint32(Keys.Attention.SINK_COUNT.format(arch=self.arch), count)
+
+ def add_conv_kernel_size(self, size: int) -> None:
+ self.add_uint32(Keys.Attention.CONV_KERNEL_SIZE.format(arch=self.arch), size)
+
+ def add_layer_swa_window(self, values: Sequence[int]) -> None:
+ self.add_array(Keys.Attention.LAYER_SWA_WINDOW.format(arch=self.arch), values)
+
def add_max_alibi_bias(self, bias: float) -> None:
self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index d1dd4cbae..e57d7c348 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -84,6 +84,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_DEEPSEEK2OCR, "deepseek2-ocr" },
{ LLM_ARCH_DEEPSEEK32, "deepseek32" },
{ LLM_ARCH_DEEPSEEK4, "deepseek4" },
+ { LLM_ARCH_OPENPANGU_V2, "openpangu-v2" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_GLM4, "glm4" },
{ LLM_ARCH_GLM4_MOE, "glm4moe" },
@@ -304,6 +305,9 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_INDEXER_TYPES, "%s.attention.indexer.types" },
{ LLM_KV_ATTENTION_INDEXER_KPOOL, "%s.attention.indexer.kpool" },
{ LLM_KV_ATTENTION_INDEXER_KPOOL_SELECT_TAIL, "%s.attention.indexer.kpool_select_tail" },
+ { LLM_KV_ATTENTION_SINK_COUNT, "%s.attention.sink_count" },
+ { LLM_KV_ATTENTION_CONV_KERNEL_SIZE, "%s.attention.conv_kernel_size" },
+ { LLM_KV_ATTENTION_LAYER_SWA_WINDOW, "%s.attention.layer_swa_window" },
{ LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT, "%s.attention.output_group_count" },
{ LLM_KV_ATTENTION_OUTPUT_LORA_RANK, "%s.attention.output_lora_rank" },
{ LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE, "%s.attention.compress_rope_freq_base" },
@@ -611,6 +615,12 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_PLE_NORM_QUERY, "blk.%d.ple_norm_query" },
{ LLM_TENSOR_PLE_NORM_CONV, "blk.%d.ple_norm_conv" },
{ LLM_TENSOR_PLE_CONV1D, "blk.%d.ple_conv1d" },
+ { LLM_TENSOR_ATTN_Q_A_CONV, "blk.%d.attn_q_a_conv" },
+ { LLM_TENSOR_ATTN_KV_A_CONV, "blk.%d.attn_kv_a_conv" },
+ { LLM_TENSOR_ATTN_O_CONV, "blk.%d.attn_o_conv" },
+ { LLM_TENSOR_ATTN_SINK_KV, "blk.%d.attn_sink_kv" },
+ { LLM_TENSOR_ATTN_SINK_K_PE, "blk.%d.attn_sink_k_pe" },
+ { LLM_TENSOR_BLOCK_POST_NORM, "blk.%d.block_post_norm" },
{ LLM_TENSOR_ATTN_COMPRESSOR_WKV, "blk.%d.attn_compressor_kv" },
{ LLM_TENSOR_ATTN_COMPRESSOR_WGATE, "blk.%d.attn_compressor_gate" },
{ LLM_TENSOR_ATTN_COMPRESSOR_APE, "blk.%d.attn_compressor_ape" },
@@ -879,6 +889,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_PLE_NORM_QUERY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_PLE_NORM_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_PLE_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+ {LLM_TENSOR_ATTN_Q_A_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+ {LLM_TENSOR_ATTN_KV_A_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+ {LLM_TENSOR_ATTN_O_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
+ {LLM_TENSOR_ATTN_SINK_KV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+ {LLM_TENSOR_ATTN_SINK_K_PE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+ {LLM_TENSOR_BLOCK_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_COMPRESSOR_WKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_COMPRESSOR_WGATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_ATTN_COMPRESSOR_APE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
@@ -1212,6 +1228,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_QWEN4EXP:
case LLM_ARCH_DEEPSEEK4:
case LLM_ARCH_MINIMAX_01:
+ case LLM_ARCH_OPENPANGU_V2:
return true;
default:
return false;
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 4e8e53421..9306fe562 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -89,6 +89,7 @@ enum llm_arch {
LLM_ARCH_DEEPSEEK2OCR,
LLM_ARCH_DEEPSEEK32,
LLM_ARCH_DEEPSEEK4,
+ LLM_ARCH_OPENPANGU_V2,
LLM_ARCH_CHATGLM,
LLM_ARCH_GLM4,
LLM_ARCH_GLM4_MOE,
@@ -309,6 +310,9 @@ enum llm_kv {
LLM_KV_ATTENTION_INDEXER_TYPES,
LLM_KV_ATTENTION_INDEXER_KPOOL,
LLM_KV_ATTENTION_INDEXER_KPOOL_SELECT_TAIL,
+ LLM_KV_ATTENTION_SINK_COUNT,
+ LLM_KV_ATTENTION_CONV_KERNEL_SIZE,
+ LLM_KV_ATTENTION_LAYER_SWA_WINDOW,
LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,
LLM_KV_ATTENTION_OUTPUT_LORA_RANK,
LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE,
@@ -655,6 +659,12 @@ enum llm_tensor {
LLM_TENSOR_PLE_NORM_QUERY, // qwen4exp
LLM_TENSOR_PLE_NORM_CONV, // qwen4exp
LLM_TENSOR_PLE_CONV1D, // qwen4exp
+ LLM_TENSOR_ATTN_Q_A_CONV, // openpangu-v2 MoME short conv
+ LLM_TENSOR_ATTN_KV_A_CONV, // openpangu-v2
+ LLM_TENSOR_ATTN_O_CONV, // openpangu-v2
+ LLM_TENSOR_ATTN_SINK_KV, // openpangu-v2 learned param-sink compressed kv
+ LLM_TENSOR_ATTN_SINK_K_PE, // openpangu-v2
+ LLM_TENSOR_BLOCK_POST_NORM,// openpangu-v2 mHC stream post-norm
LLM_TENSOR_ATTN_COMPRESSOR_WKV,
LLM_TENSOR_ATTN_COMPRESSOR_WGATE,
LLM_TENSOR_ATTN_COMPRESSOR_APE,
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 9ee64b566..5177c40c5 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2506,7 +2506,10 @@ void llama_context::output_reorder() {
uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
uint32_t res;
- if (model.arch == LLM_ARCH_KIMI_K3 || model.arch == LLM_ARCH_GLM5_NEXT) {
+ if (model.arch == LLM_ARCH_OPENPANGU_V2) {
+ // mHC (per-layer sinkhorn over many iterations) creates a very large graph
+ res = std::max<uint32_t>(n_tokens * 64, 160u * model.n_tensors());
+ } else if (model.arch == LLM_ARCH_KIMI_K3 || model.arch == LLM_ARCH_GLM5_NEXT) {
// the n_tokens*40 budget below is exhausted at ubatch 3840
res = std::max<uint32_t>(n_tokens * 160, 64u * model.n_tensors());
} else if (model.arch == LLM_ARCH_HRM_TEXT) {
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 6cadda971..913b6a68a 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -527,7 +527,7 @@ void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
if (self_k_idxs && self_k_idxs->buffer) {
mctx->set_input_k_idxs(self_k_idxs, ubatch);
}
- if (self_v_idxs && self_v_idxs->buffer) {
+ if (self_v_idxs && self_v_idxs->buffer) { // not built for K-only MLA caches
mctx->set_input_v_idxs(self_v_idxs, ubatch);
}
@@ -1148,7 +1148,9 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
mctx->get_attn()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
- mctx->get_attn()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+ if (inp_attn->self_v_idxs) { // not built for K-only MLA caches
+ mctx->get_attn()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+ }
mctx->get_attn()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
@@ -1243,7 +1245,9 @@ void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
// base tensors may not be allocated if there are no non-SWA attention layers
if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
attn_ctx->get_base()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
- attn_ctx->get_base()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+ if (inp_attn->self_v_idxs) { // not built for K-only MLA caches
+ attn_ctx->get_base()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+ }
}
if (inp_attn->self_kq_mask && inp_attn->self_kq_mask->buffer) {
@@ -1253,7 +1257,9 @@ void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
// swa tensors may not be allocated if there are no SWA attention layers
if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
attn_ctx->get_swa()->set_input_k_idxs(inp_attn->self_k_idxs_swa, ubatch);
- attn_ctx->get_swa()->set_input_v_idxs(inp_attn->self_v_idxs_swa, ubatch);
+ if (inp_attn->self_v_idxs_swa) { // not built for K-only MLA caches
+ attn_ctx->get_swa()->set_input_v_idxs(inp_attn->self_v_idxs_swa, ubatch);
+ }
}
if (inp_attn->self_kq_mask_swa && inp_attn->self_kq_mask_swa->buffer) {
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index b83f45ba9..f025555ee 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -216,6 +216,12 @@ uint32_t llama_hparams::n_embd_r() const {
return n_embd * (n_shortconv_l_cache - 1);
}
+ if (openpangu_conv_k != 0) {
+ // openPangu MoME: three causal convs (q_a, kv-compressed, attn-out) per layer
+ const uint32_t o_dim = n_head() * n_embd_head_v_mla();
+ return (openpangu_conv_k - 1) * (n_lora_q + n_lora_kv + o_dim);
+ }
+
if (n_embd_head_kda != 0) {
// for Kimi KDA layers
// Conv state for Q, K, V: 3 * (d_conv - 1) * n_head * head_dim
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 922164a54..a316d2521 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -351,6 +351,14 @@ struct llama_hparams {
// PLE conv history rows: (kernel - 1) * ngram_size; 0 without a PLE module
uint32_t ple_conv_state() const;
+ // openPangu-2.0 (openpangu-v2)
+ uint32_t openpangu_n_sink = 0; // learned param-sink KV entries per layer
+ uint32_t openpangu_conv_k = 0; // MoME causal conv kernel width
+
+ // reserved always-visible rows at the head of each K cache stream slice
+ // (learned attention sinks live there; K-only/MLA caches only)
+ uint32_t n_k_sink_prefix = 0;
+
// qwen3vl deepstack
// When parsed from GGUF, this implies the first N layers consume the first
// N deepstack embeddings. Use deepstack_mapping_arr if you need a more
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index b7dfadef4..baf89e59a 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -230,7 +230,14 @@ llama_kv_cache::llama_kv_cache(
const bool has_k = true;
const bool has_v = !is_mla;
- ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr;
+ // reserved always-visible rows at the head of each K stream slice (attention sinks)
+ const uint32_t n_k_prefix = hparams.n_k_sink_prefix;
+ if (n_k_prefix > 0) {
+ GGML_ASSERT(n_stream == 1 && "K sink prefix requires a unified KV cache");
+ GGML_ASSERT(!has_v && "K sink prefix is only supported for K-only (MLA) caches");
+ }
+
+ ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size + n_k_prefix, n_stream) : nullptr;
ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr;
has_k && ggml_format_name(k, "cache_%sk_l%d", name_tag, il);
@@ -240,7 +247,8 @@ llama_kv_cache::llama_kv_cache(
std::vector<ggml_tensor *> v_stream;
for (uint32_t s = 0; s < n_stream; ++s) {
- k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr);
+ // the per-stream cell region starts after the reserved sink prefix rows
+ k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2] + n_k_prefix*k->nb[1]) : nullptr);
v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
}
@@ -335,6 +343,12 @@ llama_kv_cache::llama_kv_cache(
attn_rot_k = true;
}
+ // openPangu-v2's custom MLA+sink attention does not apply the rotation,
+ // so do not create rotation inputs its graph would leave unused
+ if (model.arch == LLM_ARCH_OPENPANGU_V2) {
+ attn_rot_k = false;
+ }
+
const bool attn_rot_v =
!attn_rot_disable &&
n_embd_head_v_all > 0 &&
@@ -1300,19 +1314,23 @@ ggml_tensor * llama_kv_cache::get_k(ggml_context * ctx, int32_t il, uint32_t n_k
auto * k = layers[ikv].k;
- const uint64_t kv_size = get_size();
const uint64_t n_embd_k_gqa = k->ne[0];
+ // per-stream slice: [n_k_sink_prefix reserved rows | kv_size cell rows]
+ const uint64_t slice_size = k->ne[1];
+ const uint32_t n_k_prefix = hparams.n_k_sink_prefix;
+
assert(n_embd_k_gqa == hparams.n_embd_k_gqa(il));
const uint32_t ns = sinfo.s1 - sinfo.s0 + 1;
+ // the view starts at the sink prefix so attention sees [sinks | cells] contiguously
return ggml_view_4d(ctx, k,
- hparams.n_embd_head_k(il), hparams.n_head_kv(il), n_kv, ns,
+ hparams.n_embd_head_k(il), hparams.n_head_kv(il), n_kv + n_k_prefix, ns,
ggml_row_size(k->type, hparams.n_embd_head_k(il)),
ggml_row_size(k->type, n_embd_k_gqa),
- ggml_row_size(k->type, n_embd_k_gqa*kv_size),
- ggml_row_size(k->type, n_embd_k_gqa*kv_size)*sinfo.s0);
+ ggml_row_size(k->type, n_embd_k_gqa*slice_size),
+ ggml_row_size(k->type, n_embd_k_gqa*slice_size)*sinfo.s0);
}
ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
@@ -1366,9 +1384,13 @@ ggml_tensor * llama_kv_cache::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggm
k_cur = ggml_view_2d(ctx, k_cur, n_embd_gqa, n_tokens, k_cur->nb[2], 0);
- const int64_t n_stream = k->ne[2];
+ const int64_t n_stream = k->ne[2];
+ const uint32_t n_k_prefix = hparams.n_k_sink_prefix;
- if (n_stream > 1) {
+ if (n_k_prefix > 0) {
+ // cell rows start after the reserved sink prefix (n_stream == 1, asserted at init)
+ k = ggml_view_2d(ctx, k, n_embd_gqa, get_size(), k->nb[1], n_k_prefix*k->nb[1]);
+ } else if (n_stream > 1) {
const int64_t kv_size = get_size();
assert(n_embd_gqa == k->ne[0]);
@@ -1450,6 +1472,11 @@ ggml_tensor * llama_kv_cache::build_input_k_idxs(ggml_context * ctx, const llama
}
ggml_tensor * llama_kv_cache::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+ if (hparams.is_mla()) {
+ // K-only MLA cache: there are no V tensors, so no graph consumes v_idxs
+ return nullptr;
+ }
+
const uint32_t n_tokens = ubatch.n_tokens;
ggml_tensor * v_idxs;
@@ -2058,6 +2085,8 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
n_rot, n_head_kv, get_size()*n_stream,
ggml_row_size(layer.k->type, n_embd_head_k),
ggml_row_size(layer.k->type, n_embd_k_gqa),
+ // cell rows start after the reserved sink prefix (n_stream == 1 when prefix > 0)
+ hparams.n_k_sink_prefix*ggml_row_size(layer.k->type, n_embd_k_gqa) +
ggml_row_size(layer.k->type, n_embd_nope));
ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, inp->k_rot, rope_factors, freq_base_l, freq_scale_l, il);
diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp
index 2893d819e..0627ae6b8 100644
--- a/src/llama-memory-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -26,7 +26,10 @@ llama_memory_recurrent::llama_memory_recurrent(
uint32_t n_seq_max,
uint32_t n_rs_seq,
const layer_filter_cb & filter) : hparams(model.hparams), n_seq_max(n_seq_max) {
- const int32_t n_layer = hparams.n_layer();
+ // size to the full layer count (incl. MTP/NextN layers) so a filtered MTP
+ // context can allocate recurrent state for layers >= n_layer(); get_r_l/get_s_l
+ // index by the original layer id, and the filter leaves unused layers null.
+ const int32_t n_layer = hparams.n_layer_all;
head = 0;
size = mem_size;
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index ded790b2b..8aa61719e 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -206,6 +206,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
return new llama_model_dots3note(params);
case LLM_ARCH_DEEPSEEK4:
return new llama_model_deepseek4(params);
+ case LLM_ARCH_OPENPANGU_V2:
+ return new llama_model_openpangu_v2(params);
case LLM_ARCH_GLM_DSA:
return new llama_model_glm_dsa(params);
case LLM_ARCH_MISTRAL4:
@@ -3069,10 +3071,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
default:
{
// Dense MTP heads use a plain attention KV cache instead of the hybrid wrapper.
+ // openPangu-v2 also qualifies: its MTP layers skip the recurrent MoME conv.
const bool mtp_on_hybrid_qwen =
params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
(arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE ||
- arch == LLM_ARCH_BAILINGMOE3);
+ arch == LLM_ARCH_BAILINGMOE3 || arch == LLM_ARCH_OPENPANGU_V2);
const bool mtp_on_hybrid_nemotron =
params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && arch == LLM_ARCH_NEMOTRON_H_MOE;
@@ -3096,7 +3099,17 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
// a null filter_idx means the GGUF has no indexer tensors
llama_memory_hybrid::layer_filter_cb filter_idx = nullptr;
const bool needs_mem_idx = (arch == LLM_ARCH_QWEN4EXP);
- if (arch == LLM_ARCH_FALCON_H1) {
+ if (arch == LLM_ARCH_OPENPANGU_V2 && hparams.n_layer_nextn > 0) {
+ // every layer has both attention (MLA) and a recurrent conv state;
+ // split the trunk (default context) from the MTP heads (MTP context)
+ const bool is_mtp = params.ctx_type == LLAMA_CONTEXT_TYPE_MTP;
+ auto filt = [is_mtp, this](uint32_t il) {
+ return is_mtp ? (il >= hparams.n_layer()) : (il < hparams.n_layer());
+ };
+ filter_attn = filt;
+ filter_recr = filt;
+ } else if (arch == LLM_ARCH_FALCON_H1 || arch == LLM_ARCH_OPENPANGU_V2) {
+ // every layer has both attention (MLA) and a recurrent conv state
filter_attn = [&](uint32_t) { return true; };
filter_recr = [&](uint32_t) { return true; };
} else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
@@ -3492,6 +3505,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_DEEPSEEK32:
case LLM_ARCH_DEEPSEEK4:
case LLM_ARCH_MUSE_GLIMMER:
+ case LLM_ARCH_OPENPANGU_V2:
case LLM_ARCH_PLM:
case LLM_ARCH_CHATGLM:
case LLM_ARCH_GRANITE:
diff --git a/src/llama-model.h b/src/llama-model.h
index bcfbf1e2c..ae137b181 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -537,6 +537,15 @@ struct llama_layer {
struct ggml_tensor * hc_ffn_fn = nullptr;
struct ggml_tensor * hc_ffn_base = nullptr;
struct ggml_tensor * hc_ffn_scale = nullptr;
+
+ // openPangu-2.0 (openpangu-v2)
+ struct ggml_tensor * attn_q_a_conv = nullptr;
+ struct ggml_tensor * attn_kv_a_conv = nullptr;
+ struct ggml_tensor * attn_o_conv = nullptr;
+ struct ggml_tensor * attn_sink_kv = nullptr;
+ struct ggml_tensor * attn_sink_k_pe = nullptr;
+ struct ggml_tensor * block_post_norm = nullptr;
+
struct ggml_tensor * attn_comp_wkv = nullptr;
struct ggml_tensor * attn_comp_wgate = nullptr;
struct ggml_tensor * attn_comp_ape = nullptr;
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 9c6066796..33456ef2f 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -351,6 +351,17 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * param
quantize &= name.compare(0, 9, "decision.") != 0;
}
+ // openPangu-2.0: keep MoME convs, mHC phi/scale/base, and param sinks unquantized.
+ // The hc_* names are shared with DeepSeek-V4, whose (larger) hc tensors do get
+ // quantized, so restrict these rules to this arch.
+ if (arch == LLM_ARCH_OPENPANGU_V2) {
+ quantize &= name.find("_conv.weight") == std::string::npos;
+ quantize &= name.find("attn_sink_") == std::string::npos;
+ quantize &= name.find("hc_attn") == std::string::npos;
+ quantize &= name.find("hc_ffn") == std::string::npos;
+ quantize &= name.find("output_hc") == std::string::npos;
+ }
+
// do not quantize RWKV's small yet 2D weights
quantize &= name.find("time_mix_first.weight") == std::string::npos;
quantize &= name.find("time_mix_w0.weight") == std::string::npos;
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index 9b5cc2e1c..b956e541e 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -437,6 +437,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
break;
case LLAMA_VOCAB_PRE_TYPE_GPT4O:
case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
+ case LLAMA_VOCAB_PRE_TYPE_OPENPANGU_V2:
regex_exprs = {
// original regex from tokenizer.json
// "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -2475,6 +2476,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "talkie") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
clean_spaces = false;
+ } else if (
+ tokenizer_pre == "openpangu-v2") {
+ pre_type = LLAMA_VOCAB_PRE_TYPE_OPENPANGU_V2;
+ clean_spaces = false;
} else if (
tokenizer_pre == "granite-embed-multi-97m") {
pre_type = LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI;
@@ -3051,6 +3056,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
|| t.first == "<｜end▁of▁sentence｜>" // deepseek-ocr
|| t.first == "[e~[" // minimax-m2/m3
|| t.first == "<|ifm|im_end|>" // k2-horizon
+ || t.first == "<|message_end|>" // openpangu
+ || t.first == "<|pangu_text_end|>" // openpangu
) {
special_eog_ids.insert(t.second);
if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
diff --git a/src/llama-vocab.h b/src/llama-vocab.h
index 898a23ae6..8222b0068 100644
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -70,6 +70,7 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_UFAKZEKA = 59,
LLAMA_VOCAB_PRE_TYPE_MMBERT = 60,
LLAMA_VOCAB_PRE_TYPE_K2_HORIZON = 61,
+ LLAMA_VOCAB_PRE_TYPE_OPENPANGU_V2 = 62,
};
struct LLM_KV;
diff --git a/src/models/models.h b/src/models/models.h
index 88fd07d97..18d046347 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1407,6 +1407,81 @@ struct llama_model_deepseek4 : public llama_model_base {
};
+struct llama_model_openpangu_v2 : public llama_model_base {
+ llama_model_openpangu_v2(const struct llama_model_params & params) : llama_model_base(params) {}
+ void load_arch_hparams(llama_model_loader & ml) override;
+ void load_arch_tensors(llama_model_loader & ml) override;
+
+ // true when every model device supports the fused GGML_OP_SINKHORN; computed
+ // lazily on the first graph build (backends without the op, e.g. Metal, get
+ // the unfused op sequence instead)
+ bool sinkhorn_fused() const;
+ mutable int sinkhorn_fused_probe = -1;
+
+ // shared helpers used by both the trunk graph and the MTP head graph
+ struct graph_base : public llm_graph_context {
+ graph_base(const llm_graph_params & params, bool sinkhorn_fused)
+ : llm_graph_context(params), use_fused_sinkhorn(sinkhorn_fused) {}
+
+ // whether build_hc_sinkhorn emits the fused op or the unfused sequence
+ const bool use_fused_sinkhorn;
+
+ // hoisted per-graph mask tensors (built once by the concrete graph ctor):
+ // [n_k_sink_prefix | n_kv] combined masks and the standalone sink block
+ ggml_tensor * kq_mask_sinked = nullptr;
+ ggml_tensor * kq_mask_sinked_swa = nullptr;
+ ggml_tensor * sink_mask_blk = nullptr;
+
+ // sink-prefix mask block [n_k_sink_prefix, n_tokens]: real sinks visible, padding masked
+ ggml_tensor * build_sink_mask_blk(ggml_tensor * kq_mask) const;
+
+ // write the (static) param-sink rows into the reserved K-cache prefix
+ void build_sink_write(const llama_layer & layer, ggml_tensor * k_row, int il) const;
+
+ // mHC (manifold hyper-connections), gamma pre-folded into phi at conversion.
+ // xt/residual_t is the transposed stream state [hc, n_embd, nt], computed once
+ // per block and shared by the pre- and post-merge batched matmuls; when null
+ // (large batches) the per-stream loop form is used instead.
+ ggml_tensor * build_hc_pre(ggml_tensor * x, ggml_tensor * xt,
+ ggml_tensor * hc_fn, ggml_tensor * hc_scale,
+ ggml_tensor * hc_base, ggml_tensor ** post, ggml_tensor ** comb, int il) const;
+ ggml_tensor * build_hc_post(ggml_tensor * x, ggml_tensor * residual, ggml_tensor * residual_t,
+ ggml_tensor * post, ggml_tensor * comb, int il) const;
+ ggml_tensor * build_hc_head(ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale,
+ ggml_tensor * hc_base) const;
+ ggml_tensor * build_hc_weighted_sum(ggml_tensor * x, ggml_tensor * xt, ggml_tensor * weights) const;
+ ggml_tensor * build_hc_sinkhorn(ggml_tensor * comb, int il) const;
+
+ // MoME width-k causal depthwise conv + identity residual, with recurrent decode state
+ ggml_tensor * build_mome_conv(ggml_tensor * x, ggml_tensor * conv_w,
+ ggml_tensor * conv_rs, ggml_tensor * conv_state, int64_t rs_head,
+ int64_t state_off, int64_t n_seqs, int64_t n_seq_tokens) const;
+
+ ggml_tensor * build_attention(const llama_model & model, llm_graph_input_mem_hybrid_iswa * inp,
+ ggml_tensor * cur, ggml_tensor * inp_pos, float kq_scale, int il) const;
+
+ // MTP-head attention: MLA + param sinks over a plain (non-hybrid) iSWA KV cache,
+ // without the recurrent MoME conv (drafts are verified, so the conv is skipped).
+ ggml_tensor * build_attention_mtp(const llama_model & model, llm_graph_input_attn_kv_iswa * inp_attn,
+ ggml_tensor * cur, ggml_tensor * inp_pos, float kq_scale, int il) const;
+
+ // MoE + shared-expert block (used by both trunk and MTP layers)
+ ggml_tensor * build_moe_block(const llama_model & model, ggml_tensor * cur, int il) const;
+ };
+
+ struct graph : public graph_base {
+ graph(const llama_model & model, const llm_graph_params & params);
+ };
+
+ // MTP / NextN head graph for self-speculative decoding
+ struct graph_mtp : public graph_base {
+ graph_mtp(const llama_model & model, const llm_graph_params & params);
+ };
+
+ std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
struct llama_model_deepseek2ocr : public llama_model_base {
llama_model_deepseek2ocr(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
diff --git a/src/models/openpangu-v2.cpp b/src/models/openpangu-v2.cpp
new file mode 100644
index 000000000..38f1a8c14
--- /dev/null
+++ b/src/models/openpangu-v2.cpp
@@ -0,0 +1,1048 @@
+#include "models.h"
+
+#include "ggml-cpp.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
+#include "llama-memory-hybrid.h"
+#include "llama-memory-hybrid-iswa.h"
+#include "llama-memory-recurrent.h"
+
+// openPangu-2.0-Flash (openpangu-v2)
+//
+// Composition of existing infra:
+// - 4-stream mHC residual (manifold hyper-connections), same family as DeepSeek-V4.
+// The learned RMSNorm gamma is folded into the phi weight at conversion time.
+// - MLA attention (absorbed), same as DeepSeek-V2, scale 1/sqrt(qk_head_dim).
+// - MoME: width-k causal depthwise conv + identity residual on q_a, kv-compressed, and attn-out.
+// - 128 learned param-sink KV entries per layer (always visible, not indexer-scored).
+// - Sigmoid-gated MoE with expert bias + shared expert, leading dense layers.
+// - DSA lightning indexer on the non-SWA trunk layers: per-token top-k selection over
+// the cached tokens, applied as a mask on the full causal attention. The indexer key
+// rides in the tail of the cached MLA K row (no separate indexer cache).
+//
+// The MoME conv state is carried across ubatches via the recurrent memory; SWA layers use
+// the iSWA split cache.
+
+static ggml_tensor * opv2_view_1d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t i0) {
+ return ggml_view_1d(ctx, t, ne0, i0 * ggml_element_size(t));
+}
+
+static ggml_tensor * opv2_view_2d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t ne1, int64_t i0) {
+ return ggml_view_2d(ctx, t, ne0, ne1, t->nb[1], i0 * ggml_element_size(t));
+}
+
+static ggml_tensor * opv2_hc_affine(ggml_context * ctx, ggml_tensor * x, ggml_tensor * scale, ggml_tensor * base) {
+ x = ggml_mul(ctx, x, scale);
+ x = ggml_add(ctx, x, base);
+ return x;
+}
+
+void llama_model_openpangu_v2::load_arch_hparams(llama_model_loader & ml) {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+ ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false);
+ ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+
+ ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+ ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+ ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false);
+ ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
+
+ // DSA lightning indexer (top-k token selection on the non-SWA trunk layers)
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head, false);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size, false);
+ ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k, false);
+
+ ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
+ ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
+ ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
+ ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);
+
+ // mHC (reuse the DeepSeek-V4 hyper-connection fields)
+ ml.get_key(LLM_KV_HYPER_CONNECTION_COUNT, hparams.dsv4_hc_mult);
+ ml.get_key(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, hparams.dsv4_hc_sinkhorn_iters);
+ ml.get_key(LLM_KV_HYPER_CONNECTION_EPSILON, hparams.dsv4_hc_eps);
+
+ // required: the param-sink and MoME conv tensors are sized from these
+ ml.get_key(LLM_KV_ATTENTION_SINK_COUNT, hparams.openpangu_n_sink);
+ ml.get_key(LLM_KV_ATTENTION_CONV_KERNEL_SIZE, hparams.openpangu_conv_k);
+
+ // the param sinks live in reserved rows at the head of the K cache; pad the
+ // prefix to the flash-attention stride so the total KV length stays aligned
+ hparams.n_k_sink_prefix = GGML_PAD(hparams.openpangu_n_sink, 256);
+
+ // per-layer sliding window: SWA layers use a local window, DSA layers are global.
+ // Both read from the same compressed MLA KV cache; the iSWA split keeps the
+ // windowed layers cheap in memory at long context.
+ std::array<uint32_t, LLAMA_MAX_LAYERS> swa_window;
+ swa_window.fill(0);
+ ml.get_arr(LLM_KV_ATTENTION_LAYER_SWA_WINDOW, swa_window, false);
+ uint32_t swa = 0;
+ for (uint32_t il = 0; il < hparams.n_layer_all; ++il) {
+ hparams.is_swa_impl[il] = swa_window[il] != 0;
+ if (swa_window[il] != 0) {
+ swa = swa_window[il];
+ }
+ }
+ if (swa == 0) {
+ // the graph and memory creation assume the hybrid-iswa layout
+ throw std::runtime_error("openpangu-v2 requires attention.layer_swa_window with at least one non-zero window");
+ }
+ hparams.n_swa = swa;
+ hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+
+ // MLA stores a single compressed KV head: size the KV cache to the
+ // compressed dims (kv_lora + rope for K, kv_lora for V). The real
+ // per-head attention dims come from n_embd_head_{k,v}_mla().
+ // DSA layers append their indexer key to the same cache row; layers
+ // without an indexer (SWA, MTP) zero-pad the tail.
+ hparams.n_embd_head_k_full = hparams.n_lora_kv + hparams.n_rot() + hparams.indexer_head_size;
+ hparams.n_embd_head_v_full = hparams.n_lora_kv;
+ // SWA layers share the same compressed MLA KV geometry (used to size the iSWA SWA sub-cache)
+ hparams.n_embd_head_k_swa = hparams.n_embd_head_k_full;
+ hparams.n_embd_head_v_swa = hparams.n_embd_head_v_full;
+ hparams.n_rot_swa = hparams.n_rot();
+
+ type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_openpangu_v2::load_arch_tensors(llama_model_loader &) {
+ LLAMA_LOAD_LOCALS;
+
+ const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+ const int64_t q_lora_rank = hparams.n_lora_q;
+ const int64_t kv_lora_rank = hparams.n_lora_kv;
+ const int64_t n_ff_exp = hparams.n_ff_exp();
+ const int64_t n_expert_shared = hparams.n_expert_shared;
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t conv_k = hparams.openpangu_conv_k;
+ const int64_t n_sink = hparams.openpangu_n_sink;
+
+ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+ output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+ output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+
+ // model-level mHC stream merge (pre_only); absent in split MTP-only GGUFs
+ const int mhc_flags = n_layer > 0 ? 0 : TENSOR_NOT_REQUIRED;
+ hc_head_fn = create_tensor(tn(LLM_TENSOR_HC_HEAD_FN, "weight"), {hc * n_embd, hc}, mhc_flags);
+ hc_head_base = create_tensor(tn(LLM_TENSOR_HC_HEAD_BASE, "weight"), {hc}, mhc_flags);
+ hc_head_scale = create_tensor(tn(LLM_TENSOR_HC_HEAD_SCALE, "weight"), {1}, mhc_flags);
+
+ const int64_t hc_mix_dim = (2 + hc) * hc;
+
+ for (int i = 0; i < (int) hparams.n_layer_all; ++i) {
+ auto & layer = layers[i];
+ const bool is_mtp = i >= n_layer; // MTP / NextN heads have no mHC or indexer
+
+ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+ layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+ layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+ if (!is_mtp) {
+ layer.block_post_norm = create_tensor(tn(LLM_TENSOR_BLOCK_POST_NORM, "weight", i), {hc * n_embd}, TENSOR_NOT_REQUIRED);
+
+ // mHC per-layer
+ layer.hc_attn_fn = create_tensor(tn(LLM_TENSOR_HC_ATTN_FN, "weight", i), {hc * n_embd, hc_mix_dim}, 0);
+ layer.hc_attn_base = create_tensor(tn(LLM_TENSOR_HC_ATTN_BASE, "weight", i), {hc_mix_dim}, 0);
+ layer.hc_attn_scale = create_tensor(tn(LLM_TENSOR_HC_ATTN_SCALE, "weight", i), {3}, 0);
+ layer.hc_ffn_fn = create_tensor(tn(LLM_TENSOR_HC_FFN_FN, "weight", i), {hc * n_embd, hc_mix_dim}, 0);
+ layer.hc_ffn_base = create_tensor(tn(LLM_TENSOR_HC_FFN_BASE, "weight", i), {hc_mix_dim}, 0);
+ layer.hc_ffn_scale = create_tensor(tn(LLM_TENSOR_HC_FFN_SCALE, "weight", i), {3}, 0);
+ } else {
+ // MTP / NextN head projections (shared_head/embed fall back to model.tok_embd/output)
+ layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, 0);
+ layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, 0);
+ layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, 0);
+ layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+ layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+ }
+
+ // MLA
+ layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+ layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+ layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+ layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+ layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA,"weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+ layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+ layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+
+ // MoME convs (stored [k, dim])
+ layer.attn_q_a_conv = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_CONV, "weight", i), {conv_k, q_lora_rank}, 0);
+ layer.attn_kv_a_conv = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_CONV, "weight", i), {conv_k, kv_lora_rank}, 0);
+ layer.attn_o_conv = create_tensor(tn(LLM_TENSOR_ATTN_O_CONV, "weight", i), {conv_k, n_head * n_embd_head_v_mla}, 0);
+
+ // learned param sinks
+ layer.attn_sink_kv = create_tensor(tn(LLM_TENSOR_ATTN_SINK_KV, "weight", i), {kv_lora_rank, n_sink}, 0);
+ layer.attn_sink_k_pe = create_tensor(tn(LLM_TENSOR_ATTN_SINK_K_PE, "weight", i), {n_embd_head_qk_rope, n_sink}, 0);
+
+ // DSA lightning indexer (present only on the non-SWA trunk layers)
+ const int64_t idx_h = hparams.indexer_n_head;
+ const int64_t idx_d = hparams.indexer_head_size;
+ layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, idx_h * idx_d}, TENSOR_NOT_REQUIRED);
+ layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, idx_d}, TENSOR_NOT_REQUIRED);
+ layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {idx_d}, TENSOR_NOT_REQUIRED);
+ layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, idx_h}, TENSOR_NOT_REQUIRED);
+
+ if (i < (int) hparams.n_layer_dense_lead) {
+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
+ } else {
+ layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+ layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+ layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+ layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+ layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+ layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+ layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_exp * n_expert_shared, n_embd}, 0);
+ layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+ }
+ }
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_sink_mask_blk(ggml_tensor * kq_mask) const {
+ const int64_t n_sink = hparams.openpangu_n_sink;
+ const int64_t n_pfx = hparams.n_k_sink_prefix;
+
+ ggml_tensor * m_vis = ggml_new_tensor_4d(ctx0, kq_mask->type, n_sink,
+ kq_mask->ne[1], kq_mask->ne[2], kq_mask->ne[3]);
+ m_vis = ggml_fill(ctx0, m_vis, 0.0f); // real sinks: always visible
+ ggml_tensor * blk = m_vis;
+ if (n_pfx > n_sink) {
+ ggml_tensor * m_pad = ggml_new_tensor_4d(ctx0, kq_mask->type, n_pfx - n_sink,
+ kq_mask->ne[1], kq_mask->ne[2], kq_mask->ne[3]);
+ m_pad = ggml_fill(ctx0, m_pad, -INFINITY); // padding rows: masked out
+ blk = ggml_concat(ctx0, m_vis, m_pad, 0);
+ }
+ return blk;
+}
+
+void llama_model_openpangu_v2::graph_base::build_sink_write(const llama_layer & layer, ggml_tensor * k_row, int il) const {
+ const int64_t n_sink = hparams.openpangu_n_sink;
+ const int64_t n_pfx = hparams.n_k_sink_prefix;
+ const int64_t idx_d = hparams.indexer_head_size;
+
+ // sink rows [row_width, n_pfx]: RMS-norm'd compressed KV + k_pe (not RoPE'd),
+ // zero indexer tail and zero (masked) padding rows. The values are static;
+ // rewriting the small prefix every graph avoids a load-time hook.
+ ggml_tensor * sink_in = ggml_cast(ctx0, layer.attn_sink_kv, GGML_TYPE_F32);
+ ggml_tensor * sink_kv = build_norm(sink_in, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+ ggml_tensor * sink_pe = ggml_cast(ctx0, layer.attn_sink_k_pe, sink_kv->type);
+ ggml_tensor * sink_k = ggml_concat(ctx0, sink_kv, sink_pe, 0);
+ if (idx_d > 0 || n_pfx > n_sink) {
+ sink_k = ggml_pad(ctx0, sink_k, idx_d, n_pfx - n_sink, 0, 0);
+ }
+ sink_k = ggml_cast(ctx0, sink_k, k_row->type);
+
+ ggml_tensor * dst = ggml_view_2d(ctx0, k_row, k_row->ne[0], n_pfx, k_row->nb[2], 0);
+ ggml_build_forward_expand(gf, ggml_cpy(ctx0, sink_k, dst));
+}
+
+// The mHC stream reductions have two shapes: a batched-matmul form (few kernel
+// launches -- wins at decode, where launch overhead dominates) and a per-stream
+// loop form (no stream-transpose copies or tiny batched GEMMs -- wins at prefill).
+// The transposed stream state xt selects the form: non-null => batched.
+static const int64_t OPV2_HC_BATCHED_NT_MAX = 8;
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_hc_weighted_sum(ggml_tensor * x, ggml_tensor * xt, ggml_tensor * weights) const {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t nt = x->ne[2];
+
+ if (xt) {
+ // out[e,t] = sum_h xt[h,e,t] * w[h,t]: one batched matvec over the streams
+ if (!ggml_is_contiguous(weights)) {
+ weights = ggml_cont(ctx0, weights); // hc_mix output views are strided
+ }
+ ggml_tensor * wr = ggml_reshape_3d(ctx0, weights, hc, 1, nt);
+ ggml_tensor * out = ggml_mul_mat(ctx0, xt, wr); // [n_embd, 1, nt]
+ return ggml_reshape_2d(ctx0, out, n_embd, nt);
+ }
+
+ ggml_tensor * acc = nullptr;
+ for (int64_t ih = 0; ih < hc; ++ih) {
+ ggml_tensor * xh = ggml_view_2d(ctx0, x, n_embd, nt, x->nb[2], ih * x->nb[1]);
+ ggml_tensor * wh = ggml_view_2d(ctx0, weights, 1, nt, weights->nb[1], ih * weights->nb[0]);
+ ggml_tensor * cur = ggml_mul(ctx0, xh, wh);
+ acc = acc ? ggml_add(ctx0, acc, cur) : cur;
+ }
+ return acc;
+}
+
+bool llama_model_openpangu_v2::sinkhorn_fused() const {
+ if (sinkhorn_fused_probe < 0) {
+ sinkhorn_fused_probe = 1;
+
+ for (const llama_device & ldev : devices) {
+ if (ldev.is_meta) {
+ continue;
+ }
+ ggml_backend_dev_t dev = ldev.dev;
+
+ ggml_init_params params = {
+ /*.mem_size =*/ ggml_tensor_overhead()*16,
+ /*.mem_buffer =*/ NULL,
+ /*.no_alloc =*/ true,
+ };
+ ggml_context_ptr ctx { ggml_init(params) };
+ GGML_ASSERT(ctx);
+
+ ggml_backend_buffer_type_t buft = ggml_backend_dev_buffer_type(dev);
+ ggml_backend_buffer_ptr buf { ggml_backend_buft_alloc_buffer(buft, 0) };
+
+ ggml_tensor * a = ggml_new_tensor_3d(ctx.get(), GGML_TYPE_F32, 4, 4, 1);
+ ggml_tensor * mixes = ggml_new_tensor_2d(ctx.get(), GGML_TYPE_F32, 24, 1);
+ ggml_tensor * sc = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 3);
+ ggml_tensor * bs = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, 24);
+ ggml_tensor * ik = ggml_new_tensor_2d(ctx.get(), GGML_TYPE_F16, 128, 16);
+ ggml_tensor * dq = ggml_new_tensor_3d(ctx.get(), GGML_TYPE_F32, 128, 24, 1);
+ ggml_tensor * dw = ggml_new_tensor_2d(ctx.get(), GGML_TYPE_F32, 24, 1);
+ ggml_tensor * dm = ggml_new_tensor_2d(ctx.get(), GGML_TYPE_F16, 16, 1);
+
+ ggml_tensor * op_sinkhorn = ggml_sinkhorn(ctx.get(), a, 20, 1e-6f);
+ ggml_tensor * op_hc_mix = ggml_hc_mix(ctx.get(), mixes, sc, bs, 4, 20, 1e-6f);
+ ggml_tensor * op_dsa_score = ggml_dsa_score(ctx.get(), ik, dq, dw, dm);
+
+ a->buffer = buf.get();
+ mixes->buffer = buf.get();
+ sc->buffer = buf.get();
+ bs->buffer = buf.get();
+ ik->buffer = buf.get();
+ dq->buffer = buf.get();
+ dw->buffer = buf.get();
+ dm->buffer = buf.get();
+
+ if (!ggml_backend_dev_supports_op(dev, op_sinkhorn) ||
+ !ggml_backend_dev_supports_op(dev, op_hc_mix) ||
+ !ggml_backend_dev_supports_op(dev, op_dsa_score)) {
+ LLAMA_LOG_WARN("%s: device %s does not support the fused hyper-connection ops - using the unfused op sequence\n",
+ __func__, ggml_backend_dev_name(dev));
+ sinkhorn_fused_probe = 0;
+ break;
+ }
+ }
+ }
+ return sinkhorn_fused_probe != 0;
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_hc_sinkhorn(ggml_tensor * comb, int il) const {
+ GGML_UNUSED(il);
+
+ if (use_fused_sinkhorn) {
+ // fused softmax + eps + alternating row/column normalization: one op instead
+ // of the ~8 nodes x 2 x n_iter the unfused form needs per [hc, hc] matrix
+ return ggml_sinkhorn(ctx0, comb, hparams.dsv4_hc_sinkhorn_iters, hparams.dsv4_hc_eps);
+ }
+
+ // unfused fallback for backends without GGML_OP_SINKHORN (numerically equivalent)
+ comb = ggml_soft_max(ctx0, comb);
+
+ ggml_tensor * eps = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+ eps = ggml_fill(ctx0, eps, hparams.dsv4_hc_eps);
+ comb = ggml_add(ctx0, comb, eps);
+
+ auto norm_cols = [&]() {
+ ggml_tensor * comb_src_dst = ggml_cont(ctx0, ggml_permute(ctx0, comb, 1, 0, 2, 3));
+ ggml_tensor * col_sum = ggml_sum_rows(ctx0, comb_src_dst);
+ col_sum = ggml_add(ctx0, col_sum, eps);
+ col_sum = ggml_permute(ctx0, col_sum, 1, 0, 2, 3);
+ comb = ggml_div(ctx0, comb, col_sum);
+ };
+ auto norm_rows = [&]() {
+ ggml_tensor * row_sum = ggml_sum_rows(ctx0, comb);
+ row_sum = ggml_add(ctx0, row_sum, eps);
+ comb = ggml_div(ctx0, comb, row_sum);
+ };
+
+ norm_cols();
+ for (uint32_t i = 1; i < hparams.dsv4_hc_sinkhorn_iters; ++i) {
+ norm_rows();
+ norm_cols();
+ }
+ return comb;
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_hc_pre(ggml_tensor * x, ggml_tensor * xt,
+ ggml_tensor * hc_fn, ggml_tensor * hc_scale,
+ ggml_tensor * hc_base, ggml_tensor ** post, ggml_tensor ** comb, int il) const {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t hc_dim = hc * n_embd;
+ const int64_t nt = x->ne[2];
+
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc_dim, nt);
+ ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+ ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm);
+
+ if (use_fused_sinkhorn) {
+ // fused affine + sigmoid gates + sinkhorn: one op instead of ~12 per call
+ ggml_tensor * mixed = ggml_hc_mix(ctx0, mixes, hc_scale, hc_base,
+ (int) hc, (int) hparams.dsv4_hc_sinkhorn_iters, hparams.dsv4_hc_eps);
+
+ const size_t es = ggml_element_size(mixed);
+ ggml_tensor * pre = ggml_view_2d(ctx0, mixed, hc, nt, mixed->nb[1], 0);
+ *post = ggml_view_2d(ctx0, mixed, hc, nt, mixed->nb[1], hc*es);
+ *comb = ggml_view_3d(ctx0, mixed, hc, hc, nt, hc*es, mixed->nb[1], 2*hc*es);
+
+ return build_hc_weighted_sum(x, xt, pre);
+ }
+
+ ggml_tensor * scale_pre = opv2_view_1d(ctx0, hc_scale, 1, 0);
+ ggml_tensor * scale_post = opv2_view_1d(ctx0, hc_scale, 1, 1);
+ ggml_tensor * scale_comb = opv2_view_1d(ctx0, hc_scale, 1, 2);
+
+ ggml_tensor * base_pre = opv2_view_1d(ctx0, hc_base, hc, 0);
+ ggml_tensor * base_post = opv2_view_1d(ctx0, hc_base, hc, hc);
+ ggml_tensor * base_comb = opv2_view_1d(ctx0, hc_base, hc * hc, 2 * hc);
+
+ ggml_tensor * pre = opv2_view_2d(ctx0, mixes, hc, nt, 0);
+ pre = opv2_hc_affine(ctx0, pre, scale_pre, base_pre);
+ pre = ggml_sigmoid(ctx0, pre);
+ pre = ggml_scale_bias(ctx0, pre, 1.0f, hparams.dsv4_hc_eps);
+
+ *post = opv2_view_2d(ctx0, mixes, hc, nt, hc);
+ *post = opv2_hc_affine(ctx0, *post, scale_post, base_post);
+ *post = ggml_sigmoid(ctx0, *post);
+ *post = ggml_scale(ctx0, *post, 2.0f);
+
+ *comb = opv2_view_2d(ctx0, mixes, hc * hc, nt, 2 * hc);
+ *comb = opv2_hc_affine(ctx0, *comb, scale_comb, base_comb);
+ *comb = ggml_reshape_3d(ctx0, *comb, hc, hc, nt);
+ *comb = build_hc_sinkhorn(*comb, il);
+
+ return build_hc_weighted_sum(x, xt, pre);
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_hc_post(ggml_tensor * x, ggml_tensor * residual,
+ ggml_tensor * residual_t, ggml_tensor * post, ggml_tensor * comb, int il) const {
+ GGML_UNUSED(il);
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t nt = x->ne[1];
+
+ if (residual_t) {
+ // merged[e,dst,t] = sum_src residual_t[src,e,t] * comb[dst,src,t]: batched matmul.
+ // comb is laid out [dst, src, t]; mul_mat contracts dim0, so transpose the
+ // (tiny) mixing matrix to [src, dst, t] first
+ ggml_tensor * comb_t = ggml_cont(ctx0, ggml_permute(ctx0, comb, 1, 0, 2, 3));
+ ggml_tensor * merged = ggml_mul_mat(ctx0, residual_t, comb_t); // [n_embd, hc, nt]
+
+ // xpost[e,dst,t] = x[e,t] * post[dst,t]: rank-1 outer product per token
+ if (!ggml_is_contiguous(post)) {
+ post = ggml_cont(ctx0, post); // hc_mix output views are strided
+ }
+ ggml_tensor * xr = ggml_reshape_3d(ctx0, x, 1, n_embd, nt);
+ ggml_tensor * pr = ggml_reshape_3d(ctx0, post, 1, hc, nt);
+ ggml_tensor * xpost = ggml_mul_mat(ctx0, xr, pr); // [n_embd, hc, nt]
+
+ return ggml_add(ctx0, merged, xpost);
+ }
+
+ ggml_tensor * out = nullptr;
+ for (int64_t dst = 0; dst < hc; ++dst) {
+ ggml_tensor * post_dst = ggml_view_2d(ctx0, post, 1, nt, post->nb[1], dst * post->nb[0]);
+ ggml_tensor * cur = ggml_mul(ctx0, x, post_dst);
+ for (int64_t src = 0; src < hc; ++src) {
+ ggml_tensor * res_src = ggml_view_2d(ctx0, residual, n_embd, nt, residual->nb[2], src * residual->nb[1]);
+ ggml_tensor * comb_src_dst = ggml_view_2d(ctx0, comb, 1, nt, comb->nb[2], dst * comb->nb[0] + src * comb->nb[1]);
+ cur = ggml_add(ctx0, cur, ggml_mul(ctx0, res_src, comb_src_dst));
+ }
+ cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, nt);
+ out = out ? ggml_concat(ctx0, out, cur, 1) : cur;
+ }
+ return out;
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_hc_head(ggml_tensor * x, ggml_tensor * hc_fn,
+ ggml_tensor * hc_scale, ggml_tensor * hc_base) const {
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const int64_t hc_dim = hc * n_embd;
+ const int64_t nt = x->ne[2];
+
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc_dim, nt);
+ ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+ ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm);
+
+ // pre_only merge: plain sigmoid, no eps (the reference only adds hc_eps on
+ // the full pre/post/comb path, not in the pre_only head)
+ ggml_tensor * pre = opv2_hc_affine(ctx0, mixes, hc_scale, hc_base);
+ pre = ggml_sigmoid(ctx0, pre);
+
+ ggml_tensor * xt = nt <= OPV2_HC_BATCHED_NT_MAX
+ ? ggml_cont(ctx0, ggml_permute(ctx0, x, 1, 0, 2, 3)) // [hc, n_embd, nt]
+ : nullptr;
+ return build_hc_weighted_sum(x, xt, pre);
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_mome_conv(ggml_tensor * x, ggml_tensor * conv_w,
+ ggml_tensor * conv_rs, ggml_tensor * conv_state, int64_t rs_head,
+ int64_t state_off, int64_t n_seqs, int64_t n_seq_tokens) const {
+ // x: [dim, n_tokens]; conv_w: [k, dim] (bf16 in gguf, cast to f32 for ssm_conv).
+ // Causal depthwise conv with an identity residual. The previous (k-1) input
+ // columns are carried across ubatches via the recurrent state.
+ const int64_t dim = x->ne[0];
+ const int64_t nt = x->ne[1];
+ const int64_t k = conv_w->ne[0];
+ const int64_t d_conv = k - 1; // state width per channel
+ const int64_t esz = ggml_element_size(conv_state);
+
+ ggml_tensor * w = conv_w->type == GGML_TYPE_F32 ? conv_w : ggml_cast(ctx0, conv_w, GGML_TYPE_F32);
+
+ // previous-token state for this conv: strided view of the gathered recurrent state
+ ggml_tensor * state = ggml_view_3d(ctx0, conv_rs, d_conv, dim, n_seqs,
+ d_conv * esz, conv_rs->nb[1], state_off * esz); // [k-1, dim, n_seqs]
+
+ // input as [n_seq_tokens, dim, n_seqs]; concat handles the non-contiguous permute
+ ggml_tensor * xt = ggml_reshape_3d(ctx0, x, dim, n_seq_tokens, n_seqs);
+ xt = ggml_permute(ctx0, xt, 1, 0, 2, 3); // [n_seq_tokens, dim, n_seqs]
+
+ ggml_tensor * sx = ggml_concat(ctx0, state, xt, 0); // [k-1+n_seq_tokens, dim, n_seqs]
+
+ // store the last (k-1) input columns as the next state
+ ggml_tensor * new_state = ggml_view_3d(ctx0, sx, d_conv, dim, n_seqs, sx->nb[1], sx->nb[2],
+ n_seq_tokens * sx->nb[0]);
+ ggml_tensor * dst = ggml_view_2d(ctx0, conv_state, d_conv * dim, n_seqs, conv_state->nb[1],
+ rs_head * conv_state->nb[1] + state_off * esz);
+ ggml_build_forward_expand(gf, ggml_cpy(ctx0,
+ ggml_reshape_2d(ctx0, ggml_cont(ctx0, new_state), d_conv * dim, n_seqs), dst));
+
+ ggml_tensor * conv = ggml_ssm_conv(ctx0, sx, w); // [dim, n_seq_tokens, n_seqs]
+ conv = ggml_reshape_2d(ctx0, conv, dim, nt);
+
+ return ggml_add(ctx0, x, conv); // identity residual
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_attention(const llama_model & model, llm_graph_input_mem_hybrid_iswa * inp,
+ ggml_tensor * cur, ggml_tensor * inp_pos, float kq_scale, int il) const {
+ const auto & layer = model.layers[il];
+ auto * inp_attn = inp->get_attn();
+ const bool is_swa = hparams.is_swa(il);
+
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+ const int64_t kv_lora_rank = hparams.n_lora_kv;
+ const int64_t nt = cur->ne[1];
+
+ // recurrent MoME conv state (previous k-1 input columns per conv, per sequence)
+ const int64_t n_seqs = ubatch.n_seqs;
+ const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+ const int64_t d_conv = hparams.openpangu_conv_k - 1;
+ const int64_t off_qa = 0;
+ const int64_t off_kv = d_conv * hparams.n_lora_q;
+ const int64_t off_o = d_conv * (hparams.n_lora_q + hparams.n_lora_kv);
+
+ const auto * rmctx = static_cast<const llama_memory_hybrid_iswa_context *>(mctx)->get_recr();
+ ggml_tensor * conv_state = rmctx->get_r_l(il);
+ const int64_t rs_head = rmctx->get_head();
+ ggml_tensor * conv_rs = build_rs(inp->get_recr(), conv_state, hparams.n_embd_r(), n_seqs);
+
+ // q down -> conv -> norm -> up
+ ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+ q = build_mome_conv(q, layer.attn_q_a_conv, conv_rs, conv_state, rs_head, off_qa, n_seqs, n_seq_tokens);
+ ggml_tensor * q_lora = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+ q = ggml_mul_mat(ctx0, layer.wq_b, q_lora);
+
+ ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, nt,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+ ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, nt,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+ ggml_row_size(q->type, n_embd_head_qk_nope));
+
+ // kv down: split compressed (kv_lora) and rope (k_pe); conv applies to compressed only
+ ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+ ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, nt,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+ ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, nt,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+ kv_cmpr = ggml_cont(ctx0, kv_cmpr);
+ kv_cmpr = build_mome_conv(kv_cmpr, layer.attn_kv_a_conv, conv_rs, conv_state, rs_head, off_kv, n_seqs, n_seq_tokens);
+ kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+
+ q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+ // absorb q_nope into the compressed space via wk_b
+ q_nope = ggml_cont(ctx0, ggml_permute(ctx0, q_nope, 0, 2, 1, 3)); // [nope, nt, n_head]
+ ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+ q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3); // [kv_lora, n_head, nt]
+
+ ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0); // [kv_lora+rope, n_head, nt]
+
+ kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, nt);
+ ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0); // [kv_lora+rope, 1, nt]
+
+ // DSA lightning indexer: the per-token indexer key rides in the tail of the
+ // cached K row; layers without an indexer (SWA) zero-pad the tail instead
+ const int64_t idx_d = hparams.indexer_head_size;
+ const int64_t idx_h = hparams.indexer_n_head;
+ const bool is_dsa = idx_d > 0 && hparams.indexer_top_k > 0 && !is_swa &&
+ layer.indexer_attn_q_b && layer.indexer_attn_k && layer.indexer_proj;
+
+ ggml_tensor * indexer_q = nullptr;
+ if (is_dsa) {
+ // query heads from the post-conv, post-norm q_lora (the wq_b input);
+ // rope on the leading n_rot dims of each head, as in the main attention
+ indexer_q = ggml_mul_mat(ctx0, layer.indexer_attn_q_b, q_lora); // [idx_h*idx_d, nt]
+ ggml_tensor * iq_pe = ggml_view_3d(ctx0, indexer_q, n_embd_head_qk_rope, idx_h, nt,
+ ggml_row_size(indexer_q->type, idx_d), ggml_row_size(indexer_q->type, idx_d) * idx_h, 0);
+ ggml_tensor * iq_nope = ggml_view_3d(ctx0, indexer_q, idx_d - n_embd_head_qk_rope, idx_h, nt,
+ ggml_row_size(indexer_q->type, idx_d), ggml_row_size(indexer_q->type, idx_d) * idx_h,
+ ggml_row_size(indexer_q->type, n_embd_head_qk_rope));
+ iq_pe = ggml_rope_ext(ctx0, iq_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ indexer_q = ggml_concat(ctx0, iq_pe, iq_nope, 0); // [idx_d, idx_h, nt]
+
+ // single shared key from the attention input: RMS norm, same rope split
+ ggml_tensor * ik = ggml_mul_mat(ctx0, layer.indexer_attn_k, cur); // [idx_d, nt]
+ ik = build_norm(ik, layer.indexer_k_norm, nullptr, LLM_NORM_RMS, il);
+ ggml_tensor * ik_pe = ggml_view_3d(ctx0, ik, n_embd_head_qk_rope, 1, nt,
+ ggml_row_size(ik->type, idx_d), ggml_row_size(ik->type, idx_d), 0);
+ ggml_tensor * ik_nope = ggml_view_3d(ctx0, ik, idx_d - n_embd_head_qk_rope, 1, nt,
+ ggml_row_size(ik->type, idx_d), ggml_row_size(ik->type, idx_d),
+ ggml_row_size(ik->type, n_embd_head_qk_rope));
+ ik_pe = ggml_rope_ext(ctx0, ik_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ ik = ggml_concat(ctx0, ik_pe, ik_nope, 0); // [idx_d, 1, nt]
+
+ Kcur = ggml_concat(ctx0, Kcur, ik, 0); // [kv_lora+rope+idx_d, 1, nt]
+ } else if (idx_d > 0) {
+ Kcur = ggml_pad(ctx0, Kcur, idx_d, 0, 0, 0);
+ }
+
+ // store compressed K to the right iSWA sub-cache (full for DSA/global layers,
+ // windowed for SWA layers). MLA: V is a view of K, decompressed by wv_b.
+ ggml_build_forward_expand(gf, Qcur);
+ ggml_build_forward_expand(gf, Kcur);
+ const auto * attn_mctx = is_swa ? inp_attn->mctx->get_swa() : inp_attn->mctx->get_base();
+ ggml_tensor * k_idxs = is_swa ? inp_attn->get_k_idxs_swa() : inp_attn->get_k_idxs();
+ ggml_build_forward_expand(gf, attn_mctx->cpy_k(ctx0, Kcur, k_idxs, il));
+
+ const int64_t n_pfx = hparams.n_k_sink_prefix;
+
+ ggml_tensor * k_row = attn_mctx->get_k(ctx0, il); // [kv_lora+rope(+idx_d), 1, n_pfx+n_kv]
+
+ // cache-resident param sinks: refresh the reserved prefix rows (cheap, static values)
+ build_sink_write(layer, k_row, il);
+
+ ggml_tensor * k = idx_d == 0 ? k_row
+ : ggml_view_4d(ctx0, k_row, kv_lora_rank + n_embd_head_qk_rope, k_row->ne[1], k_row->ne[2], k_row->ne[3],
+ k_row->nb[1], k_row->nb[2], k_row->nb[3], 0); // MLA part: [kv_lora+rope, 1, n_pfx+n_kv]
+
+ ggml_tensor * v = ggml_view_4d(ctx0, k_row, kv_lora_rank, k_row->ne[1], k_row->ne[2], k_row->ne[3],
+ k_row->nb[1], k_row->nb[2], k_row->nb[3], 0);
+
+ // hoisted combined mask [n_pfx | n_kv]; DSA layers refine the cell part below
+ ggml_tensor * kq_mask_all = is_swa ? kq_mask_sinked_swa : kq_mask_sinked;
+ ggml_tensor * kq_mask = is_swa ? inp_attn->get_kq_mask_swa() : inp_attn->get_kq_mask();
+
+ // while the whole cache fits in the top-k budget the selection is a no-op:
+ // skip the scoring (the indexer keys are still written to the cache above)
+ if (is_dsa && k_row->ne[2] - n_pfx > (int64_t) hparams.indexer_top_k) {
+ // score the cached indexer keys (cell rows only; the param sinks are not
+ // scored -- they stay always-visible through the mask prefix):
+ // score[t,s] = sum_h w[t,h] * relu(q[t,h,:] . k[s,:]) (no scaling, per the reference)
+ ggml_tensor * ik_cache = ggml_view_4d(ctx0, k_row, idx_d, k_row->ne[1], k_row->ne[2] - n_pfx, k_row->ne[3],
+ k_row->nb[1], k_row->nb[2], k_row->nb[3],
+ n_pfx*k_row->nb[2] + ggml_row_size(k_row->type, kv_lora_rank + n_embd_head_qk_rope)); // [idx_d, 1, n_kv]
+
+ ggml_tensor * iw = ggml_mul_mat(ctx0, layer.indexer_proj, cur); // [idx_h, nt]
+
+ ggml_tensor * iscore;
+ if (use_fused_sinkhorn && k_row->ne[3] == 1) {
+ // one fused pass over the cached indexer keys (incl. relu, head-weighting
+ // and the additive mask); the CUDA backend has a warp-per-position decode
+ // kernel and a tiled prefill kernel, so no [n_kv, nh, nt] intermediate is
+ // ever materialized. multi-stream caches fall back to the matmul pipeline.
+ ggml_tensor * ikv = ggml_view_2d(ctx0, k_row, idx_d, k_row->ne[2] - n_pfx, k_row->nb[2],
+ n_pfx*k_row->nb[2] + ggml_row_size(k_row->type, kv_lora_rank + n_embd_head_qk_rope));
+ iscore = ggml_dsa_score(ctx0, ikv, indexer_q, iw, kq_mask); // [n_kv, nt]
+ } else {
+ // prefill: tiled matmul pipeline
+ // split the batch into streams if needed
+ const int64_t n_stream = ik_cache->ne[3];
+ indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream,
+ indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
+ iw = ggml_view_4d(ctx0, iw, iw->ne[0], iw->ne[1]/n_stream, iw->ne[2], n_stream,
+ iw->nb[1], iw->nb[2]/n_stream, iw->nb[3]/n_stream, 0);
+
+ indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3); // [idx_d, nt, idx_h]
+ ik_cache = ggml_permute(ctx0, ik_cache, 0, 2, 1, 3); // [idx_d, n_kv, 1]
+
+ ggml_tensor * ikq = ggml_mul_mat(ctx0, ik_cache, indexer_q); // [n_kv, nt, idx_h]
+ ikq = ggml_relu(ctx0, ikq); // contiguous, no cont needed
+ ikq = ggml_cont(ctx0, ggml_permute(ctx0, ikq, 1, 2, 0, 3)); // [idx_h, n_kv, nt]
+
+ // weighted sum over the heads as a batched matvec
+ ggml_tensor * iwv = ggml_view_4d(ctx0, iw, iw->ne[0], 1, iw->ne[1], iw->ne[3],
+ iw->nb[2], iw->nb[1], iw->nb[3], 0); // [idx_h, 1, nt]
+ iscore = ggml_mul_mat(ctx0, ikq, iwv); // [n_kv, 1, nt]
+ iscore = ggml_reshape_4d(ctx0, iscore, iscore->ne[0], iscore->ne[2], 1, iscore->ne[3]); // [n_kv, nt, 1]
+
+ // mask before top-k so invalid/future cells cannot claim slots
+ iscore = ggml_add(ctx0, iscore,
+ kq_mask->type == GGML_TYPE_F32 ? kq_mask : ggml_cast(ctx0, kq_mask, GGML_TYPE_F32));
+ }
+
+ const int64_t n_top_k = std::min<int64_t>(iscore->ne[0], hparams.indexer_top_k);
+ // GGML_OP_TOP_K output is already contiguous (the cont dated from the argsort+view form)
+ ggml_tensor * top_k = ggml_top_k(ctx0, iscore, n_top_k);
+
+ if (nt == 1) {
+ // gather-based decode attention: physically gather the selected rows and
+ // attend over [sinks | top-k] only -- attention cost stays constant with
+ // context instead of reading (and masking) the entire cache
+ ggml_tensor * ids = ggml_reshape_1d(ctx0, top_k, n_top_k);
+
+ ggml_tensor * cells = ggml_view_2d(ctx0, k_row, k_row->ne[0], k_row->ne[2] - n_pfx,
+ k_row->nb[2], n_pfx*k_row->nb[2]); // [row_w, n_kv]
+ ggml_tensor * gk = ggml_get_rows(ctx0, cells, ids); // [row_w, n_top_k] F32
+ gk = ggml_cast(ctx0, gk, k_row->type);
+
+ ggml_tensor * prefix = ggml_view_2d(ctx0, k_row, k_row->ne[0], n_pfx, k_row->nb[2], 0);
+ ggml_tensor * kg = ggml_concat(ctx0, prefix, gk, 1); // [row_w, n_pfx+n_top_k]
+
+ const int64_t n_rows = n_pfx + n_top_k;
+ k = ggml_view_4d(ctx0, kg, kv_lora_rank + n_embd_head_qk_rope, 1, n_rows, 1,
+ kg->nb[1], kg->nb[1], kg->nb[1]*n_rows, 0);
+ v = ggml_view_4d(ctx0, kg, kv_lora_rank, 1, n_rows, 1,
+ kg->nb[1], kg->nb[1], kg->nb[1]*n_rows, 0);
+
+ // gather the mask entries of the selected cells (keeps padding/other-seq
+ // cells that sneak into the top-k masked out)
+ ggml_tensor * m_r = ggml_reshape_2d(ctx0, kq_mask, 1, kq_mask->ne[0]);
+ ggml_tensor * gm = ggml_get_rows(ctx0, m_r, ids); // [1, n_top_k] F32
+ gm = ggml_reshape_2d(ctx0, gm, n_top_k, 1);
+ if (gm->type != sink_mask_blk->type) {
+ gm = ggml_cast(ctx0, gm, sink_mask_blk->type);
+ }
+ kq_mask_all = ggml_concat(ctx0, sink_mask_blk, gm, 0); // [n_pfx+n_top_k, 1]
+ } else {
+ // prefill: every query has its own top-k set, so keep the mask-based form --
+ // unmask the selected cells on an all-masked copy, then AND with the causal
+ // mask (same construction as the DSA build_attn overload in llama-graph.cpp)
+ ggml_tensor * mask_inf = ggml_fill(ctx0, kq_mask, -INFINITY);
+ mask_inf = ggml_view_4d(ctx0, mask_inf, 1, mask_inf->ne[0], mask_inf->ne[1], mask_inf->ne[3],
+ mask_inf->nb[0], mask_inf->nb[1], mask_inf->nb[2], 0);
+ ggml_tensor * top_k_3d = ggml_view_4d(ctx0, top_k, top_k->ne[0], top_k->ne[1], top_k->ne[3], 1,
+ top_k->nb[1], top_k->nb[2], top_k->ne[3]*top_k->nb[3], 0);
+ ggml_tensor * zeros = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 1, top_k_3d->ne[0], top_k_3d->ne[1], top_k_3d->ne[2]);
+ zeros = ggml_fill(ctx0, zeros, 0.0f);
+ ggml_tensor * unmasked = ggml_set_rows(ctx0, mask_inf, zeros, top_k_3d);
+ unmasked = ggml_view_4d(ctx0, unmasked, unmasked->ne[1], unmasked->ne[2], 1, unmasked->ne[3],
+ unmasked->nb[2], unmasked->nb[3], unmasked->nb[3], 0);
+ kq_mask = ggml_add(ctx0, unmasked, kq_mask);
+
+ // per-DSA-layer combined mask (the top-k selection differs per layer)
+ kq_mask_all = ggml_concat(ctx0, sink_mask_blk, kq_mask, 0);
+ }
+ }
+
+ ggml_tensor * attn_out = build_attn_mha(Qcur, k, v, nullptr, kq_mask_all, nullptr, layer.wv_b, 0, kq_scale, il);
+
+ attn_out = build_mome_conv(attn_out, layer.attn_o_conv, conv_rs, conv_state, rs_head, off_o, n_seqs, n_seq_tokens);
+ attn_out = ggml_mul_mat(ctx0, layer.wo, attn_out); // o_proj
+ return attn_out;
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_attention_mtp(const llama_model & model, llm_graph_input_attn_kv_iswa * inp_attn,
+ ggml_tensor * cur, ggml_tensor * inp_pos, float kq_scale, int il) const {
+ const auto & layer = model.layers[il];
+ // in the combined GGUF the MTP layers are converted as full attention (the single
+ // global n_swa cannot express the reference's 2048 next to the trunk's 512); a
+ // split MTP-only GGUF carries its own n_swa = 2048 and runs SWA here, matching
+ // the reference -- see conversion/pangu.py and conversion/split_pangu_mtp.py
+ const bool is_swa = hparams.is_swa(il);
+
+ const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+ const int64_t n_embd_head_qk_rope = hparams.n_rot();
+ const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+ const int64_t kv_lora_rank = hparams.n_lora_kv;
+ const int64_t nt = cur->ne[1];
+
+ // q down -> norm -> up (no MoME conv on the MTP head)
+ ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+ q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+ q = ggml_mul_mat(ctx0, layer.wq_b, q);
+
+ ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, nt,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+ ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, nt,
+ ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+ ggml_row_size(q->type, n_embd_head_qk_nope));
+
+ ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+ ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, nt,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+ ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, nt,
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+ ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+ kv_cmpr = ggml_cont(ctx0, kv_cmpr);
+ kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+
+ q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_embd_head_qk_rope, GGML_ROPE_TYPE_NEOX,
+ n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+ q_nope = ggml_cont(ctx0, ggml_permute(ctx0, q_nope, 0, 2, 1, 3));
+ ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+ q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+ ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+
+ kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, nt);
+ ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+
+ // MTP layers have no indexer -- zero-pad the DSA tail of the cache row
+ const int64_t idx_d = hparams.indexer_head_size;
+ if (idx_d > 0) {
+ Kcur = ggml_pad(ctx0, Kcur, idx_d, 0, 0, 0);
+ }
+
+ ggml_build_forward_expand(gf, Qcur);
+ ggml_build_forward_expand(gf, Kcur);
+ const auto * attn_mctx = is_swa ? inp_attn->mctx->get_swa() : inp_attn->mctx->get_base();
+ ggml_tensor * k_idxs = is_swa ? inp_attn->get_k_idxs_swa() : inp_attn->get_k_idxs();
+ ggml_build_forward_expand(gf, attn_mctx->cpy_k(ctx0, Kcur, k_idxs, il));
+
+ ggml_tensor * k_row = attn_mctx->get_k(ctx0, il); // [row_w, 1, n_pfx+n_kv]
+
+ // cache-resident param sinks: refresh the reserved prefix rows
+ build_sink_write(layer, k_row, il);
+
+ ggml_tensor * k = idx_d == 0 ? k_row
+ : ggml_view_4d(ctx0, k_row, kv_lora_rank + n_embd_head_qk_rope, k_row->ne[1], k_row->ne[2], k_row->ne[3],
+ k_row->nb[1], k_row->nb[2], k_row->nb[3], 0);
+
+ ggml_tensor * v = ggml_view_4d(ctx0, k_row, kv_lora_rank, k_row->ne[1], k_row->ne[2], k_row->ne[3],
+ k_row->nb[1], k_row->nb[2], k_row->nb[3], 0);
+
+ // hoisted combined mask
+ ggml_tensor * kq_mask_all = is_swa ? kq_mask_sinked_swa : kq_mask_sinked;
+
+ ggml_tensor * attn_out = build_attn_mha(Qcur, k, v, nullptr, kq_mask_all, nullptr, layer.wv_b, 0, kq_scale, il);
+ attn_out = ggml_mul_mat(ctx0, layer.wo, attn_out); // o_proj (no MoME conv)
+ return attn_out;
+}
+
+ggml_tensor * llama_model_openpangu_v2::graph_base::build_moe_block(const llama_model & model, ggml_tensor * cur, int il) const {
+ if (il < (int) hparams.n_layer_dense_lead) {
+ return build_ffn(cur,
+ model.layers[il].ffn_up, nullptr, nullptr,
+ model.layers[il].ffn_gate, nullptr, nullptr,
+ model.layers[il].ffn_down, nullptr, nullptr,
+ nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ }
+ ggml_tensor * moe_out = build_moe_ffn(cur,
+ model.layers[il].ffn_gate_inp,
+ model.layers[il].ffn_up_exps,
+ model.layers[il].ffn_gate_exps,
+ model.layers[il].ffn_down_exps,
+ model.layers[il].ffn_exp_probs_b,
+ n_expert, n_expert_used,
+ LLM_FFN_SILU, hparams.expert_weights_norm,
+ hparams.expert_weights_scale,
+ (llama_expert_gating_func_type) hparams.expert_gating_func,
+ il);
+ ggml_tensor * ffn_shexp = build_ffn(cur,
+ model.layers[il].ffn_up_shexp, nullptr, nullptr,
+ model.layers[il].ffn_gate_shexp, nullptr, nullptr,
+ model.layers[il].ffn_down_shexp, nullptr, nullptr,
+ nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+ return ggml_add(ctx0, moe_out, ffn_shexp);
+}
+
+std::unique_ptr<llm_graph_context> llama_model_openpangu_v2::build_arch_graph(const llm_graph_params & params) const {
+ if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+ return std::make_unique<graph_mtp>(*this, params);
+ }
+ return std::make_unique<graph>(*this, params);
+}
+
+llama_model_openpangu_v2::graph::graph(const llama_model & model, const llm_graph_params & params) :
+ graph_base(params, static_cast<const llama_model_openpangu_v2 &>(model).sinkhorn_fused()) {
+ GGML_ASSERT(n_layer > 0 && "openpangu-v2: no trunk layers -- a split MTP-only GGUF can only be used as --model-draft with --mtp");
+
+ const int64_t hc = hparams.dsv4_hc_mult;
+ const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k_mla()));
+
+ ggml_tensor * cur;
+ ggml_tensor * inp = build_inp_embd(model.tok_embd);
+ ggml_tensor * inp_pos = build_inp_pos();
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+ auto * inp_hybrid = build_inp_mem_hybrid_iswa();
+
+ // hoisted sink masks: one [n_pfx | n_kv] concat per variant per graph
+ {
+ ggml_tensor * mask_base = inp_hybrid->get_attn()->get_kq_mask();
+ ggml_tensor * mask_swa = inp_hybrid->get_attn()->get_kq_mask_swa();
+ sink_mask_blk = build_sink_mask_blk(mask_base);
+ kq_mask_sinked = ggml_concat(ctx0, sink_mask_blk, mask_base, 0);
+ kq_mask_sinked_swa = ggml_concat(ctx0, build_sink_mask_blk(mask_swa), mask_swa, 0);
+ }
+
+ ggml_tensor * inpL = ggml_reshape_3d(ctx0, inp, n_embd, 1, n_tokens);
+ inpL = ggml_repeat_4d(ctx0, inpL, n_embd, hc, n_tokens, 1);
+
+ // transposed stream state [hc, n_embd, nt] for the batched-matmul mHC form,
+ // shared by the pre- and post-merge of a block; null selects the loop form
+ auto make_xt = [&](ggml_tensor * s) -> ggml_tensor * {
+ return s->ne[2] <= OPV2_HC_BATCHED_NT_MAX
+ ? ggml_cont(ctx0, ggml_permute(ctx0, s, 1, 0, 2, 3))
+ : nullptr;
+ };
+
+ for (int il = 0; il < n_layer; ++il) {
+ ggml_tensor * post = nullptr;
+ ggml_tensor * comb = nullptr;
+
+ // attention block
+ ggml_tensor * residual = inpL;
+ ggml_tensor * residual_t = make_xt(inpL);
+ cur = build_hc_pre(inpL, residual_t, model.layers[il].hc_attn_fn, model.layers[il].hc_attn_scale,
+ model.layers[il].hc_attn_base, &post, &comb, il);
+ cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+ cur = build_attention(model, inp_hybrid, cur, inp_pos, kq_scale, il);
+ cur = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
+ inpL = build_hc_post(cur, residual, residual_t, post, comb, il);
+
+ // the last layer's FFN only needs the requested output rows -- gather early,
+ // unless the unmasked MTP extraction needs the full-length hidden state
+ if (il == n_layer - 1 && inp_out_ids &&
+ (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, inpL, n_embd * hc, n_tokens);
+ flat = ggml_get_rows(ctx0, flat, inp_out_ids);
+ inpL = ggml_reshape_3d(ctx0, flat, n_embd, hc, n_outputs);
+ }
+
+ // feed-forward block
+ residual = inpL;
+ residual_t = make_xt(inpL);
+ cur = build_hc_pre(inpL, residual_t, model.layers[il].hc_ffn_fn, model.layers[il].hc_ffn_scale,
+ model.layers[il].hc_ffn_base, &post, &comb, il);
+ cur = build_norm(cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+
+ cur = build_moe_block(model, cur, il);
+
+ cur = build_norm(cur, model.layers[il].ffn_post_norm, nullptr, LLM_NORM_RMS, il);
+ inpL = build_hc_post(cur, residual, residual_t, post, comb, il);
+
+ // block post norm on the stream-concatenated residual (subset of layers)
+ if (model.layers[il].block_post_norm) {
+ const int64_t nt_l = inpL->ne[2];
+ ggml_tensor * flat = ggml_reshape_2d(ctx0, inpL, n_embd * hc, nt_l);
+ flat = build_norm(flat, model.layers[il].block_post_norm, nullptr, LLM_NORM_RMS, il);
+ inpL = ggml_reshape_3d(ctx0, flat, n_embd, hc, nt_l);
+ }
+
+ inpL = build_cvec(inpL, il);
+ }
+
+ // Collapse the mHC streams. In the unmasked-MTP case inpL still holds all
+ // n_tokens rows so t_h_nextn covers every position; otherwise the last-layer
+ // gather above already reduced it to the output rows.
+ cur = build_hc_head(inpL, model.hc_head_fn, model.hc_head_scale, model.hc_head_base);
+
+ // trunk hidden state (pre-output-norm), consumed by the MTP head for speculative drafting
+ cb(cur, "h_nextn", -1);
+ res->t_h_nextn = cur;
+
+ if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ }
+
+ cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+ res->t_embd = cur;
+
+ cur = ggml_mul_mat(ctx0, model.output, cur);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
+
+llama_model_openpangu_v2::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) :
+ graph_base(params, static_cast<const llama_model_openpangu_v2 &>(model).sinkhorn_fused()) {
+ GGML_ASSERT(hparams.n_layer_nextn > 0 && "openpangu-v2 MTP requires n_layer_nextn > 0");
+
+ const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+ GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+ cparams.nextn_layer_offset < (int) hparams.n_layer_nextn);
+ const auto & layer = model.layers[il];
+ GGML_ASSERT(layer.nextn.eh_proj && layer.nextn.enorm && layer.nextn.hnorm);
+
+ const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k_mla()));
+
+ // inputs: the next token id + the previous hidden state (mtp_h_input)
+ auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);
+ inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+ ggml_set_input(inp->tokens);
+ inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+ ggml_set_input(inp->embd);
+ ggml_set_name(inp->embd, "mtp_h_input");
+
+ ggml_tensor * h_input = inp->embd;
+ ggml_tensor * embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+ ggml_tensor * tok_embd = ggml_get_rows(ctx0, embd_w, inp->tokens);
+ res->add_input(std::move(inp));
+
+ ggml_tensor * inp_pos = build_inp_pos();
+ auto * inp_attn = build_attn_inp_kv_iswa();
+
+ // hoisted sink masks (the swa variant is live for split MTP-only models, where
+ // the MTP layers run their reference SWA-2048 window)
+ {
+ ggml_tensor * mask_base = inp_attn->get_kq_mask();
+ ggml_tensor * mask_swa = inp_attn->get_kq_mask_swa();
+ sink_mask_blk = build_sink_mask_blk(mask_base);
+ kq_mask_sinked = ggml_concat(ctx0, sink_mask_blk, mask_base, 0);
+ kq_mask_sinked_swa = ggml_concat(ctx0, build_sink_mask_blk(mask_swa), mask_swa, 0);
+ }
+
+ // eh_proj( concat[ enorm(embed(next_tok)), hnorm(prev_hidden) ] )
+ ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+ ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+ ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, ggml_concat(ctx0, e_norm, h_norm, 0));
+
+ // plain (non-mHC) decoder layer: sandwich norms + MLA(+sinks, no conv) + MoE
+ ggml_tensor * res_sa = cur;
+ cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+ cur = build_attention_mtp(model, inp_attn, cur, inp_pos, kq_scale, il);
+ cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+ cur = ggml_add(ctx0, cur, res_sa);
+
+ ggml_tensor * res_ffn = cur;
+ cur = build_norm(cur, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);
+ cur = build_moe_block(model, cur, il);
+ cur = build_norm(cur, layer.ffn_post_norm, nullptr, LLM_NORM_RMS, il);
+ cur = ggml_add(ctx0, cur, res_ffn);
+
+ // hidden state to seed the next chained MTP head: keep all n_tokens rows,
+ // the unmasked nextn extraction reads them densely by token position
+ cb(cur, "h_nextn", -1);
+ res->t_h_nextn = cur;
+
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
+ if (inp_out_ids) {
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+ }
+
+ ggml_tensor * head_norm = layer.nextn.shared_head_norm ? layer.nextn.shared_head_norm : model.output_norm;
+ ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+ cur = build_norm(cur, head_norm, nullptr, LLM_NORM_RMS, -1);
+ cur = build_lora_mm(head_w, cur);
+ res->t_logits = cur;
+
+ ggml_build_forward_expand(gf, cur);
+}
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index ba7f7eb4e..05a0c7d27 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -8821,6 +8821,91 @@ struct test_fill : public test_case {
}
};
+// GGML_OP_SINKHORN
+struct test_sinkhorn : public test_case {
+ const std::array<int64_t, 4> ne;
+ const int n_iter;
+ const float eps;
+
+ std::string vars() override { return VARS_TO_STR3(ne, n_iter, eps); }
+
+ test_sinkhorn(std::array<int64_t, 4> ne = { 4, 4, 64, 2 }, int n_iter = 20, float eps = 1e-6f)
+ : ne(ne), n_iter(n_iter), eps(eps) {}
+
+ ggml_tensor * build_graph(ggml_context * ctx) override {
+ ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], ne[3]);
+ ggml_set_name(a, "a");
+
+ ggml_tensor * out = ggml_sinkhorn(ctx, a, n_iter, eps);
+
+ ggml_set_name(out, "out");
+
+ return out;
+ }
+};
+
+// GGML_OP_HC_MIX
+struct test_hc_mix : public test_case {
+ const int hc;
+ const int64_t nt;
+ const int n_iter;
+ const float eps;
+
+ std::string vars() override { return VARS_TO_STR4(hc, nt, n_iter, eps); }
+
+ test_hc_mix(int hc = 4, int64_t nt = 37, int n_iter = 20, float eps = 1e-6f)
+ : hc(hc), nt(nt), n_iter(n_iter), eps(eps) {}
+
+ ggml_tensor * build_graph(ggml_context * ctx) override {
+ const int64_t n0 = 2*hc + hc*hc;
+
+ ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n0, nt);
+ ggml_tensor * scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3);
+ ggml_tensor * base = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n0);
+ ggml_set_name(a, "a");
+ ggml_set_name(scale, "scale");
+ ggml_set_name(base, "base");
+
+ ggml_tensor * out = ggml_hc_mix(ctx, a, scale, base, hc, n_iter, eps);
+
+ ggml_set_name(out, "out");
+
+ return out;
+ }
+};
+
+// GGML_OP_DSA_SCORE
+struct test_dsa_score : public test_case {
+ const int64_t nd;
+ const int64_t nkv;
+ const int64_t nh;
+ const int64_t nt;
+ const ggml_type type_ik;
+
+ std::string vars() override { return VARS_TO_STR5(nd, nkv, nh, nt, type_ik); }
+
+ test_dsa_score(int64_t nd = 128, int64_t nkv = 4096, int64_t nh = 24, int64_t nt = 1,
+ ggml_type type_ik = GGML_TYPE_F16)
+ : nd(nd), nkv(nkv), nh(nh), nt(nt), type_ik(type_ik) {}
+
+ ggml_tensor * build_graph(ggml_context * ctx) override {
+ ggml_tensor * ik = ggml_new_tensor_2d(ctx, type_ik, nd, nkv);
+ ggml_tensor * q = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, nd, nh, nt);
+ ggml_tensor * w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nh, nt);
+ ggml_tensor * mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nkv, nt);
+ ggml_set_name(ik, "ik");
+ ggml_set_name(q, "q");
+ ggml_set_name(w, "w");
+ ggml_set_name(mask, "mask");
+
+ ggml_tensor * out = ggml_dsa_score(ctx, ik, q, w, mask);
+
+ ggml_set_name(out, "out");
+
+ return out;
+ }
+};
+
// GGML_OP_SOLVE_TRI
struct test_solve_tri : public test_case {
const ggml_type type;
@@ -11425,6 +11510,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_top_k_inf({248320, 1, 1, 1}, k));
}
+ test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {102400, 8, 1, 1}, 2048)); // DSA indexer selection (CUDA radix select)
for (int i = 0; i < 20; ++i) {
for (int k : {1, 2, 3, 7, 15, 100, 500, 1023, 9999}) {
if (k <= 1<<i) {
@@ -11594,6 +11680,24 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_fill(0.0f));
test_cases.emplace_back(new test_fill(2.0f, GGML_TYPE_F32, { 303, 207, 11, 3 }));
test_cases.emplace_back(new test_fill(-152.0f, GGML_TYPE_F32, { 800, 600, 4, 4 }));
+
+ test_cases.emplace_back(new test_sinkhorn());
+ test_cases.emplace_back(new test_sinkhorn({ 4, 4, 512, 1 }, 20, 1e-6f));
+ test_cases.emplace_back(new test_sinkhorn({ 3, 5, 33, 2 }, 7, 1e-5f));
+ test_cases.emplace_back(new test_sinkhorn({ 8, 8, 1, 1 }, 1, 1e-6f));
+
+ test_cases.emplace_back(new test_hc_mix());
+ test_cases.emplace_back(new test_hc_mix(4, 512, 20, 1e-6f));
+ test_cases.emplace_back(new test_hc_mix(2, 1, 5, 1e-5f));
+
+ test_cases.emplace_back(new test_dsa_score());
+ test_cases.emplace_back(new test_dsa_score(128, 2560, 24, 4, GGML_TYPE_F16));
+ test_cases.emplace_back(new test_dsa_score(64, 1000, 8, 2, GGML_TYPE_F32));
+ test_cases.emplace_back(new test_dsa_score(128, 102400, 24, 1, GGML_TYPE_F16));
+ test_cases.emplace_back(new test_dsa_score(128, 4096, 24, 512, GGML_TYPE_F16)); // CUDA tiled prefill kernel
+ test_cases.emplace_back(new test_dsa_score(128, 1000, 24, 100, GGML_TYPE_F16)); // tiled, ragged tiles
+ test_cases.emplace_back(new test_dsa_score(64, 999, 8, 37, GGML_TYPE_F16)); // tiled, nd < tile depth
+ test_cases.emplace_back(new test_dsa_score(128, 4096, 24, 512, GGML_TYPE_F32)); // large-nt fallback (warp kernel)
test_cases.emplace_back(new test_fill(3.5f, GGML_TYPE_F32, { 2048, 512, 2, 2 }));
test_cases.emplace_back(new test_fill(2.0f, GGML_TYPE_F16, { 303, 207, 11, 3 }));
test_cases.emplace_back(new test_fill(-152.0f, GGML_TYPE_F16, { 800, 600, 4, 4 }));
|