File size: 374,909 Bytes
236083b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225 6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645 6646 6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740 6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753 6754 6755 6756 6757 6758 6759 6760 6761 6762 6763 6764 6765 6766 6767 6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899 6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079 7080 7081 7082 7083 7084 7085 7086 7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111 7112 7113 7114 7115 7116 7117 7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134 7135 7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199 7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251 7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281 7282 7283 7284 7285 7286 7287 7288 7289 7290 7291 7292 7293 7294 7295 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306 7307 7308 7309 7310 7311 7312 7313 7314 7315 7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330 7331 7332 7333 7334 7335 7336 7337 7338 7339 7340 7341 7342 | # Copyright Lightning AI. Licensed under the Apache License 2.0, see LICENSE file.
"""Full definition of a decoder-only transformer-based language model, all of it in this single file.
Based on the nanoGPT implementation: https://github.com/karpathy/nanoGPT and
https://github.com/EleutherAI/gpt-neox/tree/main/megatron/model.
"""
import math
from copy import copy
from functools import partial
from typing import Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
from torch.nn.attention.flex_attention import BlockMask, create_block_mask, flex_attention, flex_attention
from typing_extensions import Self
from litgpt.config import Config
from litgpt.scripts.convert_hf_checkpoint import qkv_reassemble
def build_attention(config: Config, attention_idx: int) -> nn.Module:
"""Construct the configured mixer for one attention-layer ordinal."""
if config.kda_enabled:
period = config.kda_mla_ratio + 1
if (attention_idx + 1) % period:
return KimiDeltaAttentionAdapter(config, attention_idx)
return MultiheadLatentAttention(config, attention_idx)
if config.latent_attention:
return MultiheadLatentAttention(config, attention_idx)
if config.attention_variant == "shared_diff":
return SharedDiffCausalSelfAttention(config, attention_idx)
return CausalSelfAttention(config, attention_idx)
class GPT(nn.Module):
def __init__(self, config: Config) -> None:
super().__init__()
assert config.padded_vocab_size is not None
self.config = config
self.lm_head = (
nn.Identity()
if config.mamba3_hierarchical_vocab
else nn.Linear(config.n_embd, config.padded_vocab_size, bias=config.lm_head_bias)
)
if config.mamba3_hierarchical_vocab:
mixtures = config.mamba3_hierarchical_mixtures
self.mamba3_cluster_head = nn.Linear(
config.n_embd, mixtures * 64, bias=False
)
self.mamba3_slot_head = nn.Linear(
config.n_embd, mixtures * 64, bias=False
)
self.mamba3_mixture_head = (
nn.Linear(config.n_embd, mixtures, bias=False)
if mixtures > 1
else None
)
multipliers = torch.tensor(
[1, 2053, 1597, 3135, 817, 3327, 1231, 2871],
dtype=torch.long,
)[:mixtures]
offsets = torch.tensor(
[0, 911, 1877, 283, 3491, 1453, 2579, 3677],
dtype=torch.long,
)[:mixtures]
vocabulary = torch.arange(4096, dtype=torch.long)
self.register_buffer(
"mamba3_vocab_multipliers", multipliers, persistent=False
)
self.register_buffer(
"mamba3_vocab_offsets", offsets, persistent=False
)
self.register_buffer(
"mamba3_vocab_permutations",
(multipliers[:, None] * vocabulary[None] + offsets[:, None])
.bitwise_and(4095),
persistent=False,
)
self.mamba3_position_embedding = (
nn.Embedding(config.mamba3_position_table_size, config.n_embd)
if config.mamba3_position_table_size
else None
)
self.mamba3_phase_embedding = (
nn.Embedding(config.mamba3_phase_table_size, config.n_embd)
if config.mamba3_phase_table_size
else None
)
self.mamba3_token_scalar = (
nn.Embedding(config.mamba3_token_scalar_params, 1)
if config.mamba3_token_scalar_params
else None
)
block_class = RMSBudgetedBlock if config.rms_budgeted_block else Block
if config.mamba3_enabled:
from litgpt.mamba3_hybrid import Mamba3MultiscreenBlock
blocks = nn.ModuleList(Mamba3MultiscreenBlock(config, block_idx) for block_idx in range(config.n_layer))
elif config.rwkv7_enabled:
from litgpt.rwkv7 import RWKV7HybridBlock
if not config.rwkv7_hybrid_multiscreen:
raise NotImplementedError("The prototype currently exposes the RWKV-7 + Multiscreen hybrid.")
blocks = nn.ModuleList(RWKV7HybridBlock(config, block_idx) for block_idx in range(config.n_layer))
elif config.multiscreen_enabled:
blocks = nn.ModuleList(MultiscreenBlock(config, block_idx) for block_idx in range(config.n_layer))
elif config.multiscreen_layer_interval:
blocks = nn.ModuleList(
SelectiveRecallBlock(config, block_idx)
for block_idx in range(config.n_layer)
)
elif config.sandwich_coefficient:
k = config.sandwich_coefficient
architecture = "s" * k + "sf" * (config.n_layer - k) + "f" * k
blocks = nn.ModuleList(
SandwichSublayer(config, sublayer_idx, kind) for sublayer_idx, kind in enumerate(architecture)
)
else:
blocks = nn.ModuleList(block_class(config, block_idx) for block_idx in range(config.n_layer))
depth_memory = nn.ModuleDict()
if config.depth_memory_mode != "none":
for sublayer_idx in range(len(blocks)):
depth = sublayer_idx + 1
if depth >= config.depth_memory_start and depth % config.depth_memory_interval == 0:
depth_memory[str(sublayer_idx)] = AdaptiveDepthMemory(config)
self.transformer = nn.ModuleDict(
dict(
wte=nn.Embedding(
config.padded_vocab_size * config.mamba3_embedding_banks,
config.n_embd,
),
h=blocks,
depth_memory=depth_memory,
ln_f=(
nn.Identity()
if config.multiscreen_enabled
else nn.LayerNorm(
config.n_embd,
eps=config.norm_eps,
bias=not config.rwkv7_exact_10m_norms,
)
if config.rwkv7_enabled and config.rwkv7_exact_10m_norms
else config.norm_class(config.n_embd, eps=config.norm_eps)
),
)
)
mtp_config = config
if config.mtp_num_layers and config.sliding_window_indices is not None:
mtp_config = copy(config)
period = config.sliding_window_indices
mtp_config.sliding_window_indices = [
*period,
*(period[index % len(period)] for index in range(config.mtp_num_layers)),
]
mtp_block_class = MultiscreenBlock if config.multiscreen_enabled else Block
self.mtp_blocks = nn.ModuleList(
mtp_block_class(mtp_config, config.n_layer + index)
for index in range(config.mtp_num_layers)
)
self.mtp_enorm = (
config.norm_class(config.n_embd, eps=config.norm_eps)
if config.mtp_num_layers
else None
)
self.mtp_hnorm = (
config.norm_class(config.n_embd, eps=config.norm_eps)
if config.mtp_num_layers
else None
)
self.mtp_eh_proj = (
nn.Linear(2 * config.n_embd, config.n_embd, bias=False)
if config.mtp_num_layers
else None
)
self.mtp_eagle3_feature_proj = (
nn.Linear(3 * config.n_embd, config.n_embd, bias=False)
if config.mtp_eagle3
else None
)
if config.attn_residual_num_blocks or config.attn_residual_block_size:
num_sublayers = len(blocks)
num_reads = num_sublayers + 1
self.attn_residual_norms = nn.ModuleList(
config.norm_class(config.n_embd, eps=config.norm_eps)
for _ in range(num_reads)
)
self.attn_residual_queries = nn.ModuleList(
nn.Linear(config.n_embd, 1, bias=False)
for _ in range(num_reads)
)
self.attn_residual_block_size = (
config.attn_residual_block_size
or num_sublayers // config.attn_residual_num_blocks
)
else:
self.attn_residual_norms = nn.ModuleList()
self.attn_residual_queries = nn.ModuleList()
self.attn_residual_block_size = 0
if config.multiscreen_enabled:
self.multiscreen_embedding_scale = nn.Parameter(torch.zeros(()))
# The Multiscreen residual initially remains strongly aligned
# with the current normalized token embedding. Because the readout
# is tied to that same normalized embedding table, initializing
# its norm to sqrt(n_embd) gives the current token a logit near
# sqrt(n_embd), even though the target is the next token.
#
# A fixed initial norm of 4 keeps initial CE near log(vocab_size)
# at every residual width. This is still a learnable log-scale.
self.multiscreen_readout_scale = nn.Parameter(
torch.tensor(math.log(4.0))
)
self.register_buffer("depth_memory_scale", torch.tensor(1.0), persistent=False)
self.model_match_scale = (
nn.Parameter(torch.zeros(config.model_match_adapter_params))
if config.model_match_adapter_params
else None
)
self.mask_cache: torch.Tensor | None = None
self.max_seq_length = self.config.block_size
def _apply_model_match_scale(self, x: torch.Tensor) -> torch.Tensor:
if self.model_match_scale is not None:
scale = self.model_match_scale
if scale.numel() < self.config.n_embd:
scale = F.pad(scale, (0, self.config.n_embd - scale.numel()))
elif scale.numel() > self.config.n_embd:
rows = math.ceil(scale.numel() / self.config.n_embd)
scale = F.pad(
scale,
(0, rows * self.config.n_embd - scale.numel()),
).view(rows, self.config.n_embd).mean(dim=0)
x = x * (1.0 + 0.01 * torch.tanh(scale))
return x
def _matched_final_norm(self, x: torch.Tensor) -> torch.Tensor:
return self._apply_model_match_scale(self.transformer.ln_f(x))
def _attention_residual_read(
self,
blocks: list[torch.Tensor],
partial_block: torch.Tensor,
norm: nn.Module,
query: nn.Module,
) -> torch.Tensor:
"""Kimi Block AttnRes inter-block softmax aggregation."""
values = torch.stack([*blocks, partial_block], dim=0)
keys = norm(values)
logits = query(keys).squeeze(-1)
weights = F.softmax(logits, dim=0, dtype=torch.float).to(values.dtype)
return torch.einsum("nbt,nbtd->btd", weights, values)
def prepare_sliding_window_masks(self, sequence_length: int, device: torch.device) -> None:
"""Build block-sparse causal masks before ``torch.compile`` traces the model."""
for module in self.modules():
if isinstance(module, CausalSelfAttention):
module.prepare_sliding_window_mask(sequence_length, device)
def set_sliding_window_enabled(self, enabled: bool) -> int:
"""Toggle configured local-attention layers without changing parameters."""
changed = 0
for module in self.modules():
if not isinstance(module, CausalSelfAttention):
continue
next_value = module.configured_sliding_window_attention and enabled
if module.apply_sliding_window_attention != next_value:
module.apply_sliding_window_attention = next_value
changed += 1
return changed
@property
def max_seq_length(self) -> int:
return self._max_seq_length
@max_seq_length.setter
def max_seq_length(self, value: int) -> None:
"""
When doing inference, the sequences used might be shorter than the model's context length.
This allows setting a smaller number to avoid allocating unused memory
"""
if value > self.config.block_size:
raise ValueError(
f"Cannot attend to {value}, block size is only {self.config.block_size}."
" This is likely because the input text exceeds the supported context length of this model."
)
self._max_seq_length = value
if not hasattr(self, "cos"):
# first call
cos, sin = self.rope_cache()
self.register_buffer("cos", cos, persistent=False)
self.register_buffer("sin", sin, persistent=False)
# override
elif value != self.cos.size(0):
self.cos, self.sin = self.rope_cache(device=self.cos.device)
# the mask and kv cache size will get updated on `set_kv_cache`. we cannot update it here because we don't know
# if the kv cache is expected
if self.mask_cache is not None and self.mask_cache.shape[-1] < value:
print(
f"Warning: KV cache has length {self.mask_cache.shape[-1]} < {value} = max_seq_length. Call 'set_kv_cache' before doing any forwards!"
)
def reset_parameters(self) -> None:
# Trigger resetting the rope-cache
self.cos, self.sin = self.rope_cache(device=self.cos.device)
def _init_weights(self, module: nn.Module) -> None:
"""Meant to be used with `gpt.apply(gpt._init_weights)`."""
if module is self.mtp_eagle3_feature_proj:
# Kimi K3 initializes W_E3=[0 0 I], so enabling EAGLE-3 starts
# exactly from the high-level feature used for MTP pretraining.
with torch.no_grad():
module.weight.zero_()
module.weight[:, -self.config.n_embd :].copy_(
torch.eye(
self.config.n_embd,
device=module.weight.device,
dtype=module.weight.dtype,
)
)
elif isinstance(module, GroupedTopkRouter):
torch.nn.init.normal_(module.weight.data, mean=0.0, std=0.02)
elif isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(
self,
idx: torch.Tensor,
input_pos: torch.Tensor | None = None,
return_hidden: bool = False,
return_mtp_features: bool = False,
input_pos_maxp1: int | None = None,
lm_head_chunk_size: int = 0,
targets: torch.Tensor | None = None,
) -> torch.Tensor | list[torch.Tensor]:
"""
If `input_pos` is provided, the KV cache uses K and V vectors for
positions smaller than entries in `input_pos`. For efficiency, pass
`input_pos_maxp1` as `max(input_pos) + 1` if already available from
your forward algorithm. This slices the KV cache buffers and speeds
up multi-head attention.
Without `input_pos_maxp1`, the computation uses the full KV cache
(`max_seq_length`) with masking applied. Note that inferring
`input_pos_maxp1` from `input_pos` causes graph breaks and prevents
compilation.
Args:
idx: Token indices of input sequences, shape `(B, T)`, where `B`
is batch size.
input_pos: Optional. Positions of input tokens. The default is
`arange(T)`. Can have shape `(T,)` or `(B, T)` (batched index).
input_pos_maxp1: Optional. See above.
lm_head_chunk_size: Optional. If `lm_head_chunk_size > 0`, the final
`lm_head` computation is done in chunks of this size.
Returns:
Logit outputs, shape `(B, T, config.padded_vocab_size)`. If
`lm_head_chunk_size > 0`, this is a list of chunks of shape
`(B, lm_head_chunk_size, config.padded_vocab_size)`, the final
entry can be shorter.
"""
if return_hidden and return_mtp_features:
raise ValueError("Request either the final hidden state or EAGLE-3 features, not both.")
if return_mtp_features and not self.config.mtp_eagle3:
raise ValueError("return_mtp_features requires mtp_eagle3=True.")
T = idx.size(1)
if self.max_seq_length < T:
raise ValueError(f"Cannot forward sequence of length {T}, max seq length is only {self.max_seq_length}.")
if input_pos is not None: # use the kv cache
if input_pos.dim() > 2:
# otherwise, things go wrong in `apply_rope`
raise ValueError(f"input_pos must have 1 or 2 dimensions, input_pos.shape = {input_pos.shape}")
if input_pos.shape[-1] != T:
raise ValueError(f"input_pos.shape[-1] = {input_pos.shape[-1]} != {T} = idx.shape[1], must be the same")
cos = batched_index_select(self.cos, 0, input_pos)
sin = batched_index_select(self.sin, 0, input_pos)
if input_pos.dim() == 1:
cos = cos.unsqueeze(0)
sin = sin.unsqueeze(0)
if self.mask_cache is None:
raise TypeError("You need to call `gpt.set_kv_cache()`")
mask = batched_index_select(self.mask_cache, 2, input_pos)
if mask.dim() > 4:
# the mask cache has a batch dim of 1 in addition to the one
# we get if input_pos has a batch dimension
mask = mask.view(*(mask.shape[0:1] + mask.shape[2:]))
if input_pos_maxp1 is not None:
# Shorten final dimension so it just covers all `input_pos` entries
if input_pos_maxp1 > self.max_seq_length:
raise ValueError(f"Positions in 'input_pos' must be in [0,{self.max_seq_length})")
mask = mask[..., :input_pos_maxp1]
else:
# unsqueeze to have a batch dimension
cos = self.cos[:T].unsqueeze(0)
sin = self.sin[:T].unsqueeze(0)
# `cos`, `sin` have shape (1, T, config.rope_n_elem)
mask = None # defaults to causal mask
input_pos_maxp1 = None
if self.config.multiscreen_enabled and input_pos is not None:
raise NotImplementedError("Multiscreen KV-cache decode is not implemented; use full-sequence evaluation.")
if self.config.multiscreen_enabled:
embedding = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
x = F.embedding(idx, embedding) * self.multiscreen_embedding_scale.exp()
elif self.config.mamba3_hierarchical_vocab:
positions = torch.arange(T, device=idx.device)
bank = positions.remainder(self.config.mamba3_embedding_banks)
striped_idx = idx + bank[None] * self.config.padded_vocab_size
x = self.transformer.wte(striped_idx)
if self.mamba3_position_embedding is not None:
x = x + self.mamba3_position_embedding(
positions.remainder(self.config.mamba3_position_table_size)
)[None]
if self.mamba3_phase_embedding is not None:
x = x + self.mamba3_phase_embedding(
positions.remainder(self.config.mamba3_phase_table_size)
)[None]
if self.mamba3_token_scalar is not None:
x = x + self.mamba3_token_scalar(
idx.remainder(self.config.mamba3_token_scalar_params)
)
else:
x = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
if self.config.scale_embeddings:
x = x * torch.tensor(self.config.n_embd**0.5, dtype=x.dtype)
first_layer_values: torch.Tensor | None = None
depth_memory_states = [x]
if (
self.config.multiscreen_enabled
and self.config.multiscreen_activation_checkpointing
and self.training
and torch.is_grad_enabled()
):
if self.config.rope_indices is not None or self.config.value_residual_mix > 0.0:
raise NotImplementedError(
"Segmented Multiscreen checkpointing does not support per-layer RoPE "
"indices or value residual mixing."
)
group_size = max(1, self.config.multiscreen_checkpoint_group_size)
blocks = tuple(self.transformer.h)
for start in range(0, len(blocks), group_size):
segment_blocks = blocks[start : start + group_size]
def run_segment(
segment_input: torch.Tensor,
segment: tuple[nn.Module, ...] = segment_blocks,
) -> torch.Tensor:
for segment_block in segment:
segment_input = segment_block(
segment_input, cos, sin, None, None, None
)
return segment_input
x = checkpoint(
run_segment,
x,
# Reentrant checkpointing runs the original segment
# forward under no_grad. For Multiscreen this avoids
# retaining the compiled inner autograd graph, which is
# much larger than the single segment input at B=32.
use_reentrant=True,
preserve_rng_state=False,
)
x = F.normalize(x, dim=-1, eps=1e-6) * self.multiscreen_readout_scale.exp()
x = self._apply_model_match_scale(x)
weight = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
if targets is not None and self.config.multiscreen_chunked_lm_loss:
from litgpt.multiscreen_loss import chunked_linear_cross_entropy
loss = chunked_linear_cross_entropy(
x,
weight,
targets,
self.config.multiscreen_lm_loss_chunk_tokens,
)
return (loss, x) if return_hidden else loss
if lm_head_chunk_size > 0:
return [F.linear(x_i, weight) for x_i in x.split(lm_head_chunk_size, dim=1)]
logits = F.linear(x, weight)
return (logits, x) if return_hidden else logits
mamba3_checkpointed = (
self.config.mamba3_enabled
and self.config.mamba3_activation_checkpointing
and self.training
and torch.is_grad_enabled()
)
if mamba3_checkpointed:
group_size = self.config.mamba3_checkpoint_group_size
blocks = tuple(self.transformer.h)
for start in range(0, len(blocks), group_size):
segment_blocks = blocks[start : start + group_size]
def run_mamba3_segment(
segment_input: torch.Tensor,
segment: tuple[nn.Module, ...] = segment_blocks,
) -> torch.Tensor:
for segment_block in segment:
segment_input = segment_block(
segment_input, cos, sin, None, None, None
)
return segment_input
x = checkpoint(
run_mamba3_segment,
x,
# Mamba-3's custom Triton autograd reads saved_tensors more
# than once, which is incompatible with non-reentrant
# checkpoint hooks.
use_reentrant=True,
preserve_rng_state=False,
)
rwkv_first_value: torch.Tensor | None = None
attn_residual_blocks: list[torch.Tensor] | None = None
attn_residual_partial: torch.Tensor | None = None
if self.config.attn_residual_num_blocks or self.config.attn_residual_block_size:
# The paper treats the token embedding as the first completed
# source and the current hidden state as the initial partial sum.
attn_residual_blocks = [x]
attn_residual_partial = x
active_blocks = self.transformer.h if not mamba3_checkpointed else ()
attn_residual_norms = (
self.attn_residual_norms[:-1]
if attn_residual_blocks is not None
else (None,) * len(active_blocks)
)
attn_residual_queries = (
self.attn_residual_queries[:-1]
if attn_residual_blocks is not None
else (None,) * len(active_blocks)
)
for block_idx, (block, attn_residual_norm, attn_residual_query) in enumerate(
zip(active_blocks, attn_residual_norms, attn_residual_queries)
):
if self.config.rope_indices is not None:
block_cos = cos[..., self.config.rope_indices[block_idx]]
block_sin = sin[..., self.config.rope_indices[block_idx]]
else:
block_cos = cos
block_sin = sin
if attn_residual_blocks is not None:
assert attn_residual_partial is not None
# Keep the complete depth-attention calculation in the parent
# forward graph. Calling a shared helper here makes Dynamo
# specialize that helper on every different ``len(blocks)``
# and fall back after eight AttnRes reads.
attn_residual_values = torch.stack(
[*attn_residual_blocks, attn_residual_partial],
dim=0,
)
attn_residual_keys = attn_residual_norm(attn_residual_values)
attn_residual_logits = attn_residual_query(
attn_residual_keys
).squeeze(-1)
attn_residual_weights = F.softmax(
attn_residual_logits,
dim=0,
dtype=torch.float,
).to(attn_residual_values.dtype)
h = torch.einsum(
"nbt,nbtd->btd",
attn_residual_weights,
attn_residual_values,
)
# Block AttnRes commits the completed partial block after its
# final read and starts the next block from the next update.
if block_idx and block_idx % self.attn_residual_block_size == 0:
attn_residual_blocks.append(attn_residual_partial)
attn_residual_partial = None
update = block(
h,
block_cos,
block_sin,
mask,
input_pos,
input_pos_maxp1,
) - h
attn_residual_partial = (
update
if attn_residual_partial is None
else attn_residual_partial + update
)
x = attn_residual_partial
elif self.config.rwkv7_enabled:
x, rwkv_first_value, _ = block(
x,
block_cos,
block_sin,
mask,
input_pos,
input_pos_maxp1,
first_value=rwkv_first_value,
)
elif self.config.value_residual_mix > 0.0:
x, current_values = block(
x,
block_cos,
block_sin,
mask,
input_pos,
input_pos_maxp1,
value_residual=first_layer_values,
)
if first_layer_values is None:
first_layer_values = current_values
else:
x = block(x, block_cos, block_sin, mask, input_pos, input_pos_maxp1)
depth_memory_key = str(block_idx)
if depth_memory_key in self.transformer.depth_memory:
if self.config.depth_memory_mode == "adaptive":
x = self.transformer.depth_memory[depth_memory_key](
x, depth_memory_states, self.depth_memory_scale
)
if self.config.depth_memory_score_scale == "recurrent":
# ADeM-R keeps only the previous committed depth state.
# This avoids retaining and stacking four full residuals.
depth_memory_states = [x]
else:
depth_memory_states.append(x)
mtp_features: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None
if attn_residual_blocks is not None:
assert attn_residual_partial is not None
completed_attnres_blocks = [
*attn_residual_blocks[1:],
attn_residual_partial,
]
if return_mtp_features:
if len(completed_attnres_blocks) < 4:
raise RuntimeError("EAGLE-3 requires at least four completed AttnRes blocks.")
# Kimi K3 uses outputs of the 1st, 4th, and final blocks.
mtp_features = (
completed_attnres_blocks[0],
completed_attnres_blocks[3],
completed_attnres_blocks[-1],
)
attn_residual_values = torch.stack(
[*attn_residual_blocks, attn_residual_partial],
dim=0,
)
attn_residual_keys = self.attn_residual_norms[-1](
attn_residual_values
)
attn_residual_logits = self.attn_residual_queries[-1](
attn_residual_keys
).squeeze(-1)
attn_residual_weights = F.softmax(
attn_residual_logits,
dim=0,
dtype=torch.float,
).to(attn_residual_values.dtype)
x = torch.einsum(
"nbt,nbtd->btd",
attn_residual_weights,
attn_residual_values,
)
if self.config.multiscreen_enabled:
x = F.normalize(x, dim=-1, eps=1e-6) * self.multiscreen_readout_scale.exp()
x = self._apply_model_match_scale(x)
weight = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
if targets is not None and self.config.multiscreen_chunked_lm_loss:
from litgpt.multiscreen_loss import chunked_linear_cross_entropy
loss = chunked_linear_cross_entropy(
x,
weight,
targets,
self.config.multiscreen_lm_loss_chunk_tokens,
)
return (loss, x) if return_hidden else loss
if lm_head_chunk_size > 0:
return [F.linear(x_i, weight) for x_i in x.split(lm_head_chunk_size, dim=1)]
logits = F.linear(x, weight)
return (logits, x) if return_hidden else logits
mtp_hidden = x
x = self._matched_final_norm(x)
if self.config.mamba3_hierarchical_vocab:
logits_or_loss = self._hierarchical_vocab_output(x, targets)
if targets is not None:
return logits_or_loss
logits = logits_or_loss
return (logits, mtp_hidden) if return_hidden else logits
if (
targets is not None
and self.config.multiscreen_chunked_lm_loss
):
if self.config.final_logit_softcapping is not None:
raise NotImplementedError(
"Chunked linear CE does not support logit softcapping."
)
from litgpt.multiscreen_loss import chunked_linear_cross_entropy
loss = chunked_linear_cross_entropy(
x,
self.lm_head.weight,
targets,
self.config.multiscreen_lm_loss_chunk_tokens,
)
return (loss, mtp_hidden) if return_hidden else loss
clamp_head = (
partial(do_softcapping, thresh=self.config.final_logit_softcapping)
if self.config.final_logit_softcapping is not None
else nn.Identity()
)
if lm_head_chunk_size > 0:
return [clamp_head(self.lm_head(x_i)) for x_i in x.split(lm_head_chunk_size, dim=1)]
else:
logits = clamp_head(self.lm_head(x))
if return_mtp_features:
assert mtp_features is not None
return logits, mtp_features
return (logits, mtp_hidden) if return_hidden else logits
def forward_rwkv7_step(
self,
idx: torch.Tensor,
states: list[Any] | None = None,
) -> tuple[torch.Tensor, list[Any]]:
"""Decode one token with fixed-size RWKV and local-screening state."""
if not self.config.rwkv7_enabled:
raise RuntimeError("forward_rwkv7_step requires rwkv7_enabled.")
if idx.ndim != 2 or idx.size(1) != 1:
raise ValueError("forward_rwkv7_step expects token indices shaped (batch, 1).")
if states is not None and len(states) != len(self.transformer.h):
raise ValueError("One recurrent state is required per hybrid block.")
x = self.transformer.wte(idx)
if self.config.scale_embeddings:
x = x * torch.tensor(self.config.n_embd**0.5, dtype=x.dtype, device=x.device)
next_states = []
first_value = None
for index, block in enumerate(self.transformer.h):
layer_state = None if states is None else states[index]
x, first_value, layer_state = block.forward_step(x, first_value, layer_state)
next_states.append(layer_state)
logits = self.lm_head(self._matched_final_norm(x))
return logits, next_states
def forward_multiscreen_step(
self,
idx: torch.Tensor,
states: list[Any] | None = None,
) -> tuple[torch.Tensor, list[Any]]:
"""Decode one token with bounded per-layer Multiscreen key/value caches."""
if not self.config.multiscreen_enabled or self.config.mamba3_enabled:
raise RuntimeError("forward_multiscreen_step requires pure Multiscreen.")
if idx.ndim != 2 or idx.size(1) != 1:
raise ValueError("forward_multiscreen_step expects token indices shaped (batch, 1).")
if states is not None and len(states) != len(self.transformer.h):
raise ValueError("One screening state is required per Multiscreen block.")
embedding = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
x = F.embedding(idx, embedding) * self.multiscreen_embedding_scale.exp()
if self.config.scale_embeddings:
x = x * torch.tensor(self.config.n_embd**0.5, dtype=x.dtype, device=x.device)
next_states = []
for index, block in enumerate(self.transformer.h):
layer_state = None if states is None else states[index]
x, layer_state = block.forward_step(x, layer_state)
next_states.append(layer_state)
x = F.normalize(x, dim=-1, eps=1e-6) * self.multiscreen_readout_scale.exp()
x = self._apply_model_match_scale(x)
weight = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
return F.linear(x, weight), next_states
def forward_mamba3_step(
self,
idx: torch.Tensor,
states: list[Any] | None = None,
) -> tuple[torch.Tensor, list[Any]]:
"""Decode one token with fixed-size Mamba-3 and local-screening state."""
if not self.config.mamba3_enabled:
raise RuntimeError("forward_mamba3_step requires mamba3_enabled.")
if idx.ndim != 2 or idx.size(1) != 1:
raise ValueError("forward_mamba3_step expects token indices shaped (batch, 1).")
if states is not None and len(states) != len(self.transformer.h):
raise ValueError("One recurrent state is required per hybrid block.")
if self.config.mamba3_hierarchical_vocab:
position = 0 if states is None else states[0].next_position
bank = position % self.config.mamba3_embedding_banks
x = self.transformer.wte(idx + bank * self.config.padded_vocab_size)
if self.mamba3_position_embedding is not None:
position_idx = torch.tensor(
[position % self.config.mamba3_position_table_size],
device=idx.device,
)
x = x + self.mamba3_position_embedding(position_idx)[None]
if self.mamba3_phase_embedding is not None:
phase_idx = torch.tensor(
[position % self.config.mamba3_phase_table_size],
device=idx.device,
)
x = x + self.mamba3_phase_embedding(phase_idx)[None]
if self.mamba3_token_scalar is not None:
x = x + self.mamba3_token_scalar(
idx.remainder(self.config.mamba3_token_scalar_params)
)
else:
x = self.transformer.wte(idx)
next_states = []
for index, block in enumerate(self.transformer.h):
layer_state = None if states is None else states[index]
x, layer_state = block.forward_step(x, layer_state)
next_states.append(layer_state)
x = self._matched_final_norm(x)
if self.config.mamba3_hierarchical_vocab:
logits = self._hierarchical_vocab_output(x, None)
return logits, next_states
return self.lm_head(x), next_states
def _hierarchical_vocab_output(
self,
hidden: torch.Tensor,
targets: torch.Tensor | None,
) -> torch.Tensor:
mixtures = self.config.mamba3_hierarchical_mixtures
cluster = self.mamba3_cluster_head(hidden).view(
*hidden.shape[:-1], mixtures, 64
)
slot = self.mamba3_slot_head(hidden).view(
*hidden.shape[:-1], mixtures, 64
)
if mixtures == 1:
if targets is not None:
cluster_targets = torch.div(targets, 64, rounding_mode="floor")
slot_targets = targets.remainder(64)
return F.cross_entropy(
cluster.flatten(0, -2).float(),
cluster_targets.flatten(),
) + F.cross_entropy(
slot.flatten(0, -2).float(),
slot_targets.flatten(),
)
cluster = cluster.squeeze(-2)
slot = slot.squeeze(-2)
return (cluster.unsqueeze(-1) + slot.unsqueeze(-2)).flatten(-2)
cluster_log_probability = F.log_softmax(cluster.float(), dim=-1)
slot_log_probability = F.log_softmax(slot.float(), dim=-1)
mixture_log_probability = F.log_softmax(
self.mamba3_mixture_head(hidden).float(), dim=-1
)
if targets is not None:
mapped = (
targets[..., None] * self.mamba3_vocab_multipliers
+ self.mamba3_vocab_offsets
).bitwise_and(4095)
cluster_target = torch.div(mapped, 64, rounding_mode="floor")
slot_target = mapped.remainder(64)
cluster_target_logp = cluster_log_probability.gather(
-1, cluster_target.unsqueeze(-1)
).squeeze(-1)
slot_target_logp = slot_log_probability.gather(
-1, slot_target.unsqueeze(-1)
).squeeze(-1)
token_log_probability = torch.logsumexp(
mixture_log_probability
+ cluster_target_logp
+ slot_target_logp,
dim=-1,
)
return -token_log_probability.mean()
component_log_probability = (
cluster_log_probability.unsqueeze(-1)
+ slot_log_probability.unsqueeze(-2)
).flatten(-2)
gather_index = self.mamba3_vocab_permutations.view(
*((1,) * (component_log_probability.ndim - 2)),
mixtures,
4096,
).expand_as(component_log_probability)
component_log_probability = component_log_probability.gather(
-1, gather_index
)
return torch.logsumexp(
mixture_log_probability.unsqueeze(-1) + component_log_probability,
dim=-2,
)
def mtp_forward(
self,
hidden,
next_tokens,
return_hidden: bool = False,
targets: torch.Tensor | None = None,
):
T = hidden.size(1)
if self.config.multiscreen_enabled:
embedding_weight = F.normalize(
self.transformer.wte.weight,
dim=-1,
eps=1e-6,
)
embedding = (
F.embedding(next_tokens, embedding_weight)
* self.multiscreen_embedding_scale.exp()
)
else:
embedding = self.transformer.wte(next_tokens)
# DeepSeek-V3 Eq. 21: M_k[RMSNorm(h); RMSNorm(Emb(t_{i+k}))].
x = self.mtp_eh_proj(
torch.cat((self.mtp_hnorm(hidden), self.mtp_enorm(embedding)), dim=-1)
)
cos = self.cos[:T].unsqueeze(0)
sin = self.sin[:T].unsqueeze(0)
for block in self.mtp_blocks:
x = block(x, cos, sin)
if self.config.multiscreen_enabled:
x = F.normalize(x, dim=-1, eps=1e-6) * self.multiscreen_readout_scale.exp()
x = self._apply_model_match_scale(x)
weight = F.normalize(self.transformer.wte.weight, dim=-1, eps=1e-6)
if targets is not None and self.config.multiscreen_chunked_lm_loss:
from litgpt.multiscreen_loss import chunked_linear_cross_entropy
return chunked_linear_cross_entropy(
x,
weight,
targets,
self.config.multiscreen_lm_loss_chunk_tokens,
)
logits = F.linear(x, weight)
else:
x = self._matched_final_norm(x)
if targets is not None and self.config.multiscreen_chunked_lm_loss:
from litgpt.multiscreen_loss import chunked_linear_cross_entropy
return chunked_linear_cross_entropy(
x,
self.lm_head.weight,
targets,
self.config.multiscreen_lm_loss_chunk_tokens,
)
logits = self.lm_head(x)
return (logits, x) if return_hidden else logits
def mtp_forward_multiscreen_step(
self,
hidden: torch.Tensor,
next_tokens: torch.Tensor,
states: list[Any] | None = None,
return_hidden: bool = False,
) -> tuple[torch.Tensor, list[Any]] | tuple[
torch.Tensor,
torch.Tensor,
list[Any],
]:
"""Advance the MTP drafter with bounded Multiscreen recurrent state."""
if not self.config.multiscreen_enabled or self.config.mamba3_enabled:
raise RuntimeError(
"mtp_forward_multiscreen_step requires pure Multiscreen."
)
if len(self.mtp_blocks) != 1:
raise RuntimeError(
"mtp_forward_multiscreen_step requires exactly one MTP layer."
)
if hidden.shape[:2] != next_tokens.shape or next_tokens.size(1) != 1:
raise ValueError(
"MTP hidden states and tokens must both have shape (batch, 1, ...)."
)
if states is not None and len(states) != len(self.mtp_blocks):
raise ValueError("One recurrent state is required per MTP block.")
embedding_weight = F.normalize(
self.transformer.wte.weight,
dim=-1,
eps=1e-6,
)
embedding = (
F.embedding(next_tokens, embedding_weight)
* self.multiscreen_embedding_scale.exp()
)
x = self.mtp_eh_proj(
torch.cat(
(self.mtp_hnorm(hidden), self.mtp_enorm(embedding)),
dim=-1,
)
)
next_states = []
for index, block in enumerate(self.mtp_blocks):
layer_state = None if states is None else states[index]
x, layer_state = block.forward_step(x, layer_state)
next_states.append(layer_state)
draft_hidden = x
x = F.normalize(x, dim=-1, eps=1e-6) * self.multiscreen_readout_scale.exp()
x = self._apply_model_match_scale(x)
logits = F.linear(x, embedding_weight)
if return_hidden:
return logits, draft_hidden, next_states
return logits, next_states
def mtp_eagle3_forward(
self,
features: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
next_tokens: torch.Tensor,
return_hidden: bool = False,
):
"""Run the Kimi K3 EAGLE-3 drafter from three AttnRes block features."""
if self.mtp_eagle3_feature_proj is None:
raise RuntimeError("mtp_eagle3_forward requires mtp_eagle3=True.")
if len(features) != 3:
raise ValueError("EAGLE-3 requires exactly low-, mid-, and high-level features.")
hidden = self.mtp_eagle3_feature_proj(torch.cat(features, dim=-1))
return self.mtp_forward(hidden, next_tokens, return_hidden=return_hidden)
def mtp_eagle3_forward_step(
self,
features: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
next_tokens: torch.Tensor,
input_pos: torch.Tensor,
input_pos_maxp1: int,
return_hidden: bool = False,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
"""Advance EAGLE-3 using the cached pretrained MTP module."""
if self.mtp_eagle3_feature_proj is None:
raise RuntimeError("mtp_eagle3_forward_step requires mtp_eagle3=True.")
if len(features) != 3:
raise ValueError("EAGLE-3 requires exactly low-, mid-, and high-level features.")
hidden = self.mtp_eagle3_feature_proj(torch.cat(features, dim=-1))
return self.mtp_forward_step(
hidden,
next_tokens,
input_pos,
input_pos_maxp1,
return_hidden=return_hidden,
)
def mtp_forward_step(
self,
hidden: torch.Tensor,
next_tokens: torch.Tensor,
input_pos: torch.Tensor,
input_pos_maxp1: int,
return_hidden: bool = False,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
"""Advance the one-depth DeepSeek MTP module with its own KV cache."""
if len(self.mtp_blocks) != 1:
raise RuntimeError("mtp_forward_step requires exactly one MTP module.")
if hidden.shape[:2] != next_tokens.shape or next_tokens.size(1) != input_pos.shape[-1]:
raise ValueError("MTP hidden states, tokens, and positions must have matching sequence lengths.")
if self.mask_cache is None:
raise TypeError("Call set_kv_cache() before cached MTP decoding.")
embedding = self.transformer.wte(next_tokens)
x = self.mtp_eh_proj(
torch.cat((self.mtp_hnorm(hidden), self.mtp_enorm(embedding)), dim=-1)
)
cos = batched_index_select(self.cos, 0, input_pos)
sin = batched_index_select(self.sin, 0, input_pos)
if input_pos.dim() == 1:
cos = cos.unsqueeze(0)
sin = sin.unsqueeze(0)
mask = batched_index_select(self.mask_cache, 2, input_pos)
if mask.dim() > 4:
mask = mask.view(*(mask.shape[0:1] + mask.shape[2:]))
mask = mask[..., :input_pos_maxp1]
for block in self.mtp_blocks:
x = block(
x,
cos,
sin,
mask,
input_pos,
input_pos_maxp1,
)
logits = self.lm_head(self._matched_final_norm(x))
return (logits, x) if return_hidden else logits
@classmethod
def from_name(cls, name: str, **kwargs: Any) -> Self:
return cls(Config.from_name(name, **kwargs))
def rope_cache(self, device: torch.device | None = None) -> tuple[torch.Tensor, torch.Tensor]:
if self.config.rope_adjustments is None:
extra_config = None
else:
# Check for mutually exclusive parameter sets
llama3_params = ["low_freq_factor", "high_freq_factor"]
yarn_params = ["beta_fast", "beta_slow"]
has_llama3 = any(param in self.config.rope_adjustments for param in llama3_params)
has_yarn = any(param in self.config.rope_adjustments for param in yarn_params)
if has_llama3 and has_yarn:
raise ValueError(
"RoPE adjustments cannot contain both Llama3 parameters (low_freq_factor, high_freq_factor) "
"and YaRN parameters (beta_fast, beta_slow). These are mutually exclusive."
)
# Llama3-style RoPE
if has_llama3:
adjusted_params_required = ["factor", "low_freq_factor", "high_freq_factor", "original_max_seq_len"]
params_present = [param in self.config.rope_adjustments for param in adjusted_params_required]
if all(params_present):
extra_config = {name: self.config.rope_adjustments[name] for name in adjusted_params_required}
else:
missing_params = [
param for param, present in zip(adjusted_params_required, params_present) if not present
]
raise ValueError(
f"The following Llama3 RoPE parameters are missing in rope_adjustments: {', '.join(missing_params)}. "
"All Llama3 parameters must be specified together."
)
# YaRN-style RoPE
elif has_yarn:
# Required: factor, beta_fast, beta_slow, original_max_seq_len
# Optional: mscale, mscale_all_dim
yarn_required_params = ["factor", "beta_fast", "beta_slow", "original_max_seq_len"]
params_present = [param in self.config.rope_adjustments for param in yarn_required_params]
if not all(params_present):
missing_params = [
param for param, present in zip(yarn_required_params, params_present) if not present
]
raise ValueError(
f"The following YaRN RoPE parameters are missing in rope_adjustments: {', '.join(missing_params)}. "
"All YaRN required parameters must be specified together."
)
extra_config = {name: self.config.rope_adjustments[name] for name in yarn_required_params}
# Add optional YaRN parameters
for param in ["mscale", "mscale_all_dim"]:
if param in self.config.rope_adjustments:
extra_config[param] = self.config.rope_adjustments[param]
# Linear or standard RoPE
elif "factor" in self.config.rope_adjustments:
# linear RoPE
adjusted_params_required = ["factor"]
extra_config = {name: self.config.rope_adjustments[name] for name in adjusted_params_required}
else:
extra_config = None # uses standard RoPE
return build_rope_cache(
seq_len=self.max_seq_length,
n_elem=self.config.rope_n_elem,
device=device,
condense_ratio=self.config.rope_condense_ratio,
base=self.config.rope_base,
extra_config=extra_config,
rope_local_base_freq=self.config.rope_local_base_freq,
)
def rope_cache_length(self) -> int:
"""
Extract the head dimension (n_elem) from RoPE cache regardless of shape.
The RoPE cache can have different shapes depending on model configuration:
- Standard RoPE: (seq_len, n_elem) - 2D tensor
- Dual RoPE (local/global): (seq_len, n_elem, 2) - 3D tensor
Returns:
int: n_elem (head dimension for RoPE)
"""
return self.cos.size(1)
def set_kv_cache(
self,
batch_size: int,
max_seq_length: int | None = None,
rope_cache_length: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> None:
if rope_cache_length is None:
rope_cache_length = self.rope_cache_length()
if max_seq_length is None:
max_seq_length = self.max_seq_length
# initialize the kv cache for all blocks
for block in self.transformer.h:
if block.attn is None:
continue
block.attn.kv_cache = block.attn.build_kv_cache(
batch_size,
max_seq_length,
rope_cache_length,
device,
dtype,
)
for block in self.mtp_blocks:
block.attn.kv_cache = block.attn.build_kv_cache(
batch_size,
max_seq_length,
rope_cache_length,
device,
dtype,
)
if self.mask_cache is None or self.mask_cache.size(3) != max_seq_length:
# passing `attn_mask` to SDPA disables the flash implementation. since we only need the mask
# for the kv-cache support (only during inference), we only create it in that situation
self.mask_cache = build_mask_cache(max_seq_length, device)
def clear_kv_cache(self) -> None:
self.mask_cache = None
for block in self.transformer.h:
if block.attn is not None:
block.attn.kv_cache = None
for block in self.mtp_blocks:
block.attn.kv_cache = None
class AdaptiveDepthMemory(nn.Module):
"""Bounded, token-adaptive retrieval from earlier residual states.
Retrieval weights are content-addressed independently for every token.
A zero-initialized channel gate makes the adaptive path exactly identical
to the control at initialization while allowing a live first-step gradient.
Earlier states are RMS-matched to the current stream before interpolation,
preventing depth-dependent residual scale from dominating retrieval.
"""
def __init__(self, config: Config) -> None:
super().__init__()
self.max_sources = config.depth_memory_max_sources
self.max_commit = config.depth_memory_max_commit
self.score_scale = config.depth_memory_score_scale
# Direct zeros consume no RNG, preserving every shared parent tensor.
self.channel_gate = nn.Parameter(torch.zeros(config.n_embd))
self.source_bias = nn.Parameter(torch.zeros(self.max_sources))
self.log_temperature = nn.Parameter(torch.zeros(()))
def forward(
self,
x: torch.Tensor,
states: list[torch.Tensor],
scale: torch.Tensor | float = 1.0,
) -> torch.Tensor:
sources = torch.stack(states[-self.max_sources :], dim=-2)
x_float = x.float()
sources_float = sources.float()
source_count = sources.size(-2)
bias = self.source_bias[-source_count:].float()
if self.score_scale in {"factorized", "recurrent"}:
# Factorized ADeM: global depth selection plus a local
# token/channel agreement modulation. This removes the costly
# per-token source softmax while retaining input-dependent gates.
if self.score_scale == "recurrent":
retrieved = sources_float[..., -1, :]
else:
weights = torch.softmax(bias, dim=-1)
retrieved = (weights.unsqueeze(-1) * sources_float).sum(dim=-2)
current_rms = x_float.pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
retrieved_rms = retrieved.pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
retrieved = retrieved * (current_rms / retrieved_rms)
correction = (retrieved - x_float).to(dtype=x.dtype)
agreement = torch.tanh(
x_float * retrieved / (current_rms * current_rms).clamp_min(1e-6)
).to(dtype=x.dtype)
modulation = 1.0 + 0.5 * torch.tanh(self.log_temperature).to(dtype=x.dtype) * agreement
gate = self.max_commit * torch.tanh(self.channel_gate).to(dtype=x.dtype) * modulation
return x + torch.as_tensor(scale, device=x.device, dtype=x.dtype) * gate * correction
query_unit = F.normalize(x_float, dim=-1, eps=1e-6)
source_unit = F.normalize(sources_float, dim=-1, eps=1e-6)
scores = (source_unit * query_unit.unsqueeze(-2)).sum(dim=-1)
if self.score_scale == "inverse_sqrt":
# Retain exact ADeM-v1 checkpoint semantics. ADeM-v2 uses the
# cosine directly because both operands are already unit norm.
scores = scores / math.sqrt(x.size(-1))
temperature = F.softplus(self.log_temperature.float()) + 0.5
weights = torch.softmax(scores * temperature + bias, dim=-1)
retrieved = (weights.unsqueeze(-1) * sources_float).sum(dim=-2)
current_rms = x_float.pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
retrieved_rms = retrieved.pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
retrieved = retrieved * (current_rms / retrieved_rms)
correction = (retrieved - x_float).to(dtype=x.dtype)
gate = self.max_commit * torch.tanh(self.channel_gate).to(dtype=x.dtype)
return x + torch.as_tensor(scale, device=x.device, dtype=x.dtype) * gate * correction
class _PackedMultiscreenProjection(torch.autograd.Function):
"""One GEMM over four original parameters, with gradients returned separately."""
@staticmethod
def forward(ctx, x, packed_weight, *weights):
ctx.save_for_backward(x, packed_weight)
ctx.shapes = tuple(weight.shape for weight in weights)
return F.linear(x, packed_weight)
@staticmethod
def backward(ctx, grad_output):
x, packed_weight = ctx.saved_tensors
x_flat = x.reshape(-1, x.size(-1))
grad_flat = grad_output.reshape(-1, grad_output.size(-1))
grad_x = torch.mm(grad_flat, packed_weight).reshape_as(x)
grad_packed = torch.mm(grad_flat.transpose(0, 1), x_flat)
gradients = []
offset = 0
for heads, hidden, features in ctx.shapes:
width = heads * features
gradient = (
grad_packed[offset : offset + width]
.view(heads, features, hidden)
.permute(0, 2, 1)
.contiguous()
)
gradients.append(gradient)
offset += width
return grad_x, None, *gradients
class ExactMultiscreenResidualAdapter(nn.Module):
"""Compiler-fusible active calibration for a small exact budget gap."""
def __init__(self, width: int, parameter_budget: int) -> None:
super().__init__()
scale_size = min(width, parameter_budget)
bias_size = parameter_budget - scale_size
self.width = width
self.scale = nn.Parameter(torch.zeros(scale_size))
self.bias = nn.Parameter(torch.zeros(bias_size))
if sum(parameter.numel() for parameter in self.parameters()) != parameter_budget:
raise AssertionError("Multiscreen adapter did not consume its exact budget.")
def forward(self, x: torch.Tensor) -> torch.Tensor:
scale = F.pad(self.scale, (0, self.width - self.scale.numel()))
bias = F.pad(self.bias, (0, self.width - self.bias.numel()))
return x * (0.01 * torch.tanh(scale)) + bias
class BudgetedMultiscreenResidualAdapter(nn.Module):
"""Useful low-rank residual that consumes an arbitrary exact budget."""
def __init__(self, width: int, parameter_budget: int) -> None:
super().__init__()
rank, remainder = divmod(parameter_budget, 2 * width)
if rank <= 0:
raise ValueError("Budgeted adapter requires at least 2 * width parameters.")
self.width = width
self.up = nn.Parameter(torch.empty(width, rank))
self.down = nn.Parameter(torch.zeros(rank, width))
scale_size = min(width, remainder)
self.scale = nn.Parameter(torch.zeros(scale_size))
self.bias = nn.Parameter(torch.zeros(remainder - scale_size))
nn.init.orthogonal_(self.up)
if sum(parameter.numel() for parameter in self.parameters()) != parameter_budget:
raise AssertionError("Budgeted adapter did not consume its exact budget.")
def forward(self, x: torch.Tensor) -> torch.Tensor:
low_rank = F.silu(x @ self.up) @ self.down
scale = F.pad(self.scale, (0, self.width - self.scale.numel()))
bias = F.pad(self.bias, (0, self.width - self.bias.numel()))
return low_rank + 0.01 * x * torch.tanh(scale) + bias
class MultiscreenBlock(nn.Module):
"""A paper-faithful gated screening layer from arXiv:2604.01178.
The query loop is only a memory-bounding implementation detail. It computes
the same dense causal screening equation as the unchunked reference.
"""
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
d = config.n_embd
h = config.multiscreen_num_heads
dk = config.multiscreen_key_dim
dv = config.multiscreen_value_dim
self.num_heads = h
self.num_layers = config.n_layer
self.key_dim = dk
self.value_dim = dv
self.window_threshold = float(config.multiscreen_window_threshold)
self.query_chunk_size = config.multiscreen_query_chunk_size
self.landmark_stride = config.multiscreen_landmark_stride
self.triton_inference = config.multiscreen_triton_inference
self.fused_projections = config.multiscreen_fused_projections
self.swe_context_length = config.multiscreen_swe_context_length
self.hard_max_window = config.multiscreen_hard_max_window
self.match_adapter = None
if config.multiscreen_match_adapter_params and block_idx == config.n_layer - 1:
adapter_budget = config.multiscreen_match_adapter_params
self.match_adapter = (
ExactMultiscreenResidualAdapter(d, adapter_budget)
if adapter_budget <= 2 * d
else BudgetedMultiscreenResidualAdapter(d, adapter_budget)
)
self.query_weight = nn.Parameter(torch.empty(h, d, dk))
self.key_weight = nn.Parameter(torch.empty(h, d, dk))
self.value_weight = nn.Parameter(torch.empty(h, d, dv))
self.gate_weight = nn.Parameter(torch.empty(h, d, dv))
self.output_weight = nn.Parameter(torch.empty(h, dv, d))
self.log_window_minus_one = nn.Parameter(torch.empty(h))
self.acceptance_logit = nn.Parameter(torch.zeros(h))
self.log_output_scale = nn.Parameter(torch.empty(h))
self.mlp_norm = (
config.norm_class(d, eps=config.norm_eps) if config.multiscreen_mlp_enabled else None
)
self.mlp = config.mlp_class(config) if config.multiscreen_mlp_enabled else None
self.reset_parameters()
@torch.no_grad()
def reset_parameters(self) -> None:
nn.init.normal_(self.query_weight, std=0.1 / math.sqrt(self.key_dim))
nn.init.normal_(self.key_weight, std=0.1 / math.sqrt(self.key_dim))
nn.init.normal_(self.value_weight, std=0.1 / math.sqrt(self.value_dim))
nn.init.normal_(self.gate_weight, std=0.1)
nn.init.normal_(self.output_weight, std=0.1 / math.sqrt(self.output_weight.shape[-1]))
self.log_window_minus_one.copy_(
torch.linspace(0.0, math.log(self.window_threshold), self.num_heads)
)
self.acceptance_logit.zero_()
self.log_output_scale.fill_(-0.5 * math.log(self.num_heads * self.num_layers))
@staticmethod
def _mipe(x: torch.Tensor, angle: torch.Tensor) -> torch.Tensor:
first, second = x[..., 0], x[..., 1]
cos, sin = angle.cos(), angle.sin()
rotated_pair = torch.stack(
(first * cos - second * sin, first * sin + second * cos),
dim=-1,
)
# Avoid clone + two in-place CopySlices nodes. Those nodes retain the
# full pre-rotation tensor for backward and become expensive at large
# microbatches.
return torch.cat((rotated_pair, x[..., 2:]), dim=-1)
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
) -> torch.Tensor:
del cos, sin, mask, input_pos, input_pos_maxp1
_, sequence_length, _ = x.shape
if self.fused_projections:
from litgpt.multiscreen_projection_triton import direct_multiscreen_projection
h, dk, dv = self.num_heads, self.key_dim, self.value_dim
projected = direct_multiscreen_projection(
x,
self.query_weight,
self.key_weight,
self.value_weight,
self.gate_weight,
)
query, key, value, gate = projected.split(
(h * dk, h * dk, h * dv, h * dv), dim=-1
)
query = query.view(*query.shape[:2], h, dk).transpose(1, 2)
key = key.view(*key.shape[:2], h, dk).transpose(1, 2)
value = value.view(*value.shape[:2], h, dv).transpose(1, 2)
gate = gate.view(*gate.shape[:2], h, dv).transpose(1, 2)
else:
query = torch.einsum("btd,hdk->bhtk", x, self.query_weight)
key = torch.einsum("btd,hdk->bhtk", x, self.key_weight)
value = torch.einsum("btd,hdv->bhtv", x, self.value_weight)
gate = torch.einsum("btd,hdv->bhtv", x, self.gate_weight)
query = F.normalize(query, dim=-1, eps=1e-6)
key = F.normalize(key, dim=-1, eps=1e-6)
value = F.normalize(value, dim=-1, eps=1e-6)
learned_window = self.log_window_minus_one.exp() + 1.0
if self.hard_max_window > 0:
learned_window = learned_window.clamp_max(float(self.hard_max_window))
use_swe = not torch.is_grad_enabled() and self.swe_context_length > 0
if use_swe:
# A negative value is the internal infinity sentinel understood
# by both the reference and Triton screening implementations.
window = torch.where(
learned_window > self.swe_context_length,
torch.full_like(learned_window, -1.0),
learned_window,
)
else:
window = learned_window
finite_window = window > 0.0
safe_window = torch.where(finite_window, window, torch.ones_like(window))
gamma = torch.where(
finite_window & (window < self.window_threshold),
0.5 * (torch.cos(math.pi * safe_window / self.window_threshold) + 1.0),
torch.zeros_like(window),
)
positions = torch.arange(sequence_length, device=x.device, dtype=torch.float32)
query_angle = (math.pi * gamma / safe_window)[:, None] * positions[None, :]
query = self._mipe(query, query_angle[None].to(query.dtype))
key_positions_1d = positions[:: self.landmark_stride]
key = key[:, :, :: self.landmark_stride]
value = value[:, :, :: self.landmark_stride]
key_angle = (
(math.pi * gamma / safe_window)[:, None] * key_positions_1d[None, :]
)
key = self._mipe(key, key_angle[None].to(key.dtype))
acceptance_width = torch.sigmoid(self.acceptance_logit).clamp_min(1e-6)
use_triton = (
self.triton_inference
and query.is_cuda
and query.dtype in {torch.bfloat16, torch.float32}
and self.key_dim == 16
and self.value_dim in {32, 64, 128}
and self.landmark_stride == 1
)
if use_triton:
from litgpt.multiscreen_triton import screening_aggregate_triton
aggregated = screening_aggregate_triton(
query.contiguous(), key.contiguous(), value.contiguous(), acceptance_width, window
)
else:
chunks = []
key_positions = key_positions_1d[None, None, None, :]
for start in range(0, sequence_length, self.query_chunk_size):
end = min(start + self.query_chunk_size, sequence_length)
similarity = torch.einsum("bhqd,bhkd->bhqk", query[:, :, start:end], key)
relevance = F.relu(
1.0 - (1.0 - similarity.float()) / acceptance_width.float()[None, :, None, None]
).square()
query_positions = positions[start:end][None, None, :, None]
distance = key_positions - query_positions
infinite = ~finite_window[None, :, None, None]
valid = (distance <= 0.0) & (
infinite | (distance > -safe_window[None, :, None, None])
)
finite_softmask = 0.5 * (
torch.cos(
math.pi * distance / safe_window[None, :, None, None]
)
+ 1.0
)
softmask = torch.where(infinite, torch.ones_like(finite_softmask), finite_softmask)
relevance = relevance * torch.where(valid, softmask, torch.zeros_like(softmask))
chunks.append(
torch.einsum(
"bhqk,bhkv->bhqv",
relevance.to(value.dtype),
value,
)
)
aggregated = torch.cat(chunks, dim=2)
norm = aggregated.float().norm(dim=-1, keepdim=True)
aggregated = aggregated * (norm.tanh() / norm.clamp_min(1e-6)).to(aggregated.dtype)
gated = aggregated * torch.tanh(F.silu(gate))
gated = gated * self.log_output_scale.exp()[None, :, None, None]
update = torch.einsum("bhtv,hvd->btd", gated, self.output_weight)
x = x + update
if self.mlp is not None:
x = x + self.mlp(self.mlp_norm(x))
if self.match_adapter is not None:
x = x + self.match_adapter(x)
return x
def forward_step(
self,
x: torch.Tensor,
state: tuple[torch.Tensor, torch.Tensor, int] | None = None,
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor, int]]:
"""Decode one token while retaining only the largest learned finite window."""
if x.ndim != 3 or x.size(1) != 1:
raise ValueError("MultiscreenBlock.forward_step expects shape (batch, 1, width).")
if self.swe_context_length > 0:
raise NotImplementedError(
"Bounded cached decode does not support screening-window expansion."
)
if self.fused_projections:
from litgpt.multiscreen_projection_triton import direct_multiscreen_projection
h, dk, dv = self.num_heads, self.key_dim, self.value_dim
projected = direct_multiscreen_projection(
x,
self.query_weight,
self.key_weight,
self.value_weight,
self.gate_weight,
)
query, key, value, gate = projected.split(
(h * dk, h * dk, h * dv, h * dv), dim=-1
)
query = query.view(*query.shape[:2], h, dk).transpose(1, 2)
key = key.view(*key.shape[:2], h, dk).transpose(1, 2)
value = value.view(*value.shape[:2], h, dv).transpose(1, 2)
gate = gate.view(*gate.shape[:2], h, dv).transpose(1, 2)
else:
query = torch.einsum("btd,hdk->bhtk", x, self.query_weight)
key = torch.einsum("btd,hdk->bhtk", x, self.key_weight)
value = torch.einsum("btd,hdv->bhtv", x, self.value_weight)
gate = torch.einsum("btd,hdv->bhtv", x, self.gate_weight)
query = F.normalize(query, dim=-1, eps=1e-6)
key = F.normalize(key, dim=-1, eps=1e-6)
value = F.normalize(value, dim=-1, eps=1e-6)
learned_window = self.log_window_minus_one.exp() + 1.0
if self.hard_max_window > 0:
learned_window = learned_window.clamp_max(float(self.hard_max_window))
safe_window = learned_window.clamp_min(1.0)
gamma = torch.where(
learned_window < self.window_threshold,
0.5 * (torch.cos(math.pi * safe_window / self.window_threshold) + 1.0),
torch.zeros_like(learned_window),
)
position = 0 if state is None else state[2]
angle = (math.pi * gamma / safe_window * position)[None, :, None].to(query.dtype)
query = self._mipe(query, angle)
key = self._mipe(key, angle)
if state is not None:
key = torch.cat((state[0], key), dim=2)
value = torch.cat((state[1], value), dim=2)
cache_tokens = max(1, int(torch.ceil(safe_window.max()).item()))
if key.size(2) > cache_tokens:
key = key[:, :, -cache_tokens:]
value = value[:, :, -cache_tokens:]
history = key.size(2)
distance = torch.arange(
1 - history,
1,
device=x.device,
dtype=torch.float32,
)[None, None, None, :]
similarity = torch.einsum("bhqd,bhkd->bhqk", query, key)
acceptance_width = torch.sigmoid(self.acceptance_logit).clamp_min(1e-6)
relevance = F.relu(
1.0 - (1.0 - similarity.float()) / acceptance_width.float()[None, :, None, None]
).square()
valid = distance > -safe_window[None, :, None, None]
softmask = 0.5 * (
torch.cos(math.pi * distance / safe_window[None, :, None, None]) + 1.0
)
relevance = relevance * torch.where(valid, softmask, torch.zeros_like(softmask))
aggregated = torch.einsum("bhqk,bhkv->bhqv", relevance, value)
norm = aggregated.float().norm(dim=-1, keepdim=True)
aggregated = aggregated * (norm.tanh() / norm.clamp_min(1e-6)).to(aggregated.dtype)
gated = aggregated * torch.tanh(F.silu(gate))
gated = gated * self.log_output_scale.exp()[None, :, None, None]
update = torch.einsum("bhtv,hvd->btd", gated, self.output_weight)
x = x + update
if self.mlp is not None:
x = x + self.mlp(self.mlp_norm(x))
if self.match_adapter is not None:
x = x + self.match_adapter(x)
return x, (key, value, position + 1)
class KimiDeltaAttentionAdapter(nn.Module):
"""LitGPT adapter around Moonshot/FLA's official KDA implementation."""
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
try:
from fla.layers import KimiDeltaAttention
except ImportError as error:
raise ImportError(
"KDA requires the official `fla-core`/`flash-linear-attention` package."
) from error
self.attn = KimiDeltaAttention(
hidden_size=config.n_embd,
expand_v=1,
head_dim=config.kda_head_dim,
num_heads=config.kda_num_heads,
num_v_heads=config.kda_num_heads,
mode="chunk",
use_short_conv=True,
conv_size=config.kda_short_conv_kernel_size,
conv_bias=False,
safe_gate=config.kda_safe_gate,
lower_bound=config.kda_lower_bound if config.kda_safe_gate else None,
layer_idx=block_idx,
norm_eps=config.norm_eps,
)
if config.kda_full_rank_output_gate:
# Kimi K3 Eq. 6: y = Wo[sigmoid(Wg x) * RMSNorm(o)].
# The current FLA layer defaults to Kimi Linear's low-rank Wg;
# replacing it with one bias-free projection preserves the
# official FLA KDA kernel while matching K3's parameterization.
self.attn.g_proj = nn.Linear(
config.n_embd,
self.attn.value_dim,
bias=False,
)
self.kv_cache = None
def build_kv_cache(
self,
batch_size: int,
max_seq_length: int,
rope_cache_length: int,
device: torch.device | None,
dtype: torch.dtype | None,
):
"""Create the official FLA recurrent cache used by KDA decoding."""
del batch_size, max_seq_length, rope_cache_length, device, dtype
from fla.models.utils import Cache
return Cache()
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
) -> torch.Tensor:
use_cache = input_pos is not None
output, _, cache = self.attn(
hidden_states=x,
attention_mask=None,
past_key_values=self.kv_cache if use_cache else None,
use_cache=use_cache,
output_attentions=False,
)
if use_cache:
self.kv_cache = cache
return output
class Block(nn.Module):
def __init__(
self,
config: Config,
block_idx: int,
) -> None:
super().__init__()
if not config.parallel_residual and config.shared_attention_norm:
raise NotImplementedError(
"No checkpoint amongst the ones we support uses this configuration"
" (non-parallel residual and shared attention norm)."
)
self.norm_1 = nn.Identity() if not config.norm_1 else config.norm_class(config.n_embd, eps=config.norm_eps)
self.attn = build_attention(config, block_idx)
self.post_attention_norm = (
config.norm_class(config.n_embd, eps=config.norm_eps) if config.post_attention_norm else nn.Identity()
)
self.norm_2 = (
nn.Identity()
if not config.norm_2
else (None if config.shared_attention_norm else config.norm_class(config.n_embd, eps=config.norm_eps))
)
mlp_config = config
if config.mlp_taper_schedule != "none":
mlp_config = copy(config)
mlp_config.intermediate_size = config.tapered_mlp_intermediate_sizes()[block_idx]
self.mlp = mlp_config.mlp_class(mlp_config)
self.mlp_intermediate_size = mlp_config.intermediate_size
if config.first_k_dense_replace is not None and block_idx < config.first_k_dense_replace:
self.mlp = LLaMAMLP(config)
if hasattr(self.mlp, "set_block_index"):
self.mlp.set_block_index(block_idx, config.n_layer)
self.post_mlp_norm = (
config.norm_class(config.n_embd, eps=config.norm_eps) if config.post_mlp_norm else nn.Identity()
)
self.grouped_mlp_rotation_shift = 0
is_grouped_mlp = (
config.mlp_class_name.startswith("TileRouted") or config.mlp_class_name == "HiddenBlockDSwiGLUMLP"
) and not isinstance(self.mlp, LLaMAMLP)
if is_grouped_mlp and config.grouped_mlp_channel_rotation:
stride = config.grouped_mlp_channel_rotation_stride
if stride is None:
stride = max(1, config.n_embd // max(1, config.sparse_mlp_num_groups))
self.grouped_mlp_rotation_shift = (block_idx * stride) % config.n_embd
self.grouped_mlp_mixer: nn.Module | None = None
if (
is_grouped_mlp
and config.grouped_mlp_mix_every_n_layers > 0
and (block_idx + 1) % config.grouped_mlp_mix_every_n_layers == 0
):
if config.grouped_mlp_mix_rank > 0:
self.grouped_mlp_mixer = nn.Sequential(
nn.Linear(config.n_embd, config.grouped_mlp_mix_rank, bias=False),
nn.Linear(config.grouped_mlp_mix_rank, config.n_embd, bias=False),
)
else:
self.grouped_mlp_mixer = nn.Linear(config.n_embd, config.n_embd, bias=False)
mix_alpha = torch.tensor(float(config.grouped_mlp_mix_alpha))
if config.grouped_mlp_mix_alpha_learnable:
self.grouped_mlp_mix_alpha = nn.Parameter(mix_alpha)
else:
self.register_buffer("grouped_mlp_mix_alpha", mix_alpha, persistent=False)
self.config = config
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
value_residual: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
"""
Non-parallel residual Parallel residual
┌─ x ┌─ x ──────────────────┐ Note: if `shared_attention_norm` is True,
│ ↓ │ ↓ ↓ the output from `norm_1` is reused
│ norm_1 │ norm_1 ───────► norm_2
│ ↓ │ ↓ ↓
│ attn │ attn MLP
│ ↓ │ ↓ ↓
| post_attn_norm | post_attn_norm post_mlp_norm
| ↓ | ↓ ↓
┌─ └► + └► + ◄─────────────────┘
| ↓
│ norm_2
│ ↓
│ MLP
│ ↓
| post_mlp_norm
| ↓
└───► +
"""
x_normed = self.norm_1(x)
if self.config.value_residual_mix > 0.0:
attention_output, current_values = self.attn(
x_normed,
cos,
sin,
mask,
input_pos,
input_pos_maxp1,
value_residual=value_residual,
)
else:
attention_output = self.attn(x_normed, cos, sin, mask, input_pos, input_pos_maxp1)
attention_output = self.post_attention_norm(attention_output)
if self.config.parallel_residual:
if not self.config.shared_attention_norm:
x_normed = self.norm_2(x)
x = attention_output + x
else:
x = attention_output + x
x_normed = self.norm_2(x)
mlp_input = x_normed
if self.grouped_mlp_rotation_shift:
mlp_input = torch.roll(mlp_input, shifts=self.grouped_mlp_rotation_shift, dims=-1)
mlp_output = self.mlp(mlp_input)
if self.grouped_mlp_rotation_shift:
mlp_output = torch.roll(mlp_output, shifts=-self.grouped_mlp_rotation_shift, dims=-1)
if self.grouped_mlp_mixer is not None:
mix_alpha = self.grouped_mlp_mix_alpha.to(device=mlp_output.device, dtype=mlp_output.dtype)
mlp_output = mlp_output + mix_alpha * self.grouped_mlp_mixer(mlp_output)
output = self.post_mlp_norm(mlp_output) + x
if self.config.value_residual_mix > 0.0:
return output, current_values
return output
class SelectiveRecallBlock(nn.Module):
"""Fast Transformer block with a periodic causal landmark screen."""
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
self.base = Block(config, block_idx)
self.screen_enabled = (
(block_idx + 1) % config.multiscreen_layer_interval == 0
)
self.screen_norm = (
config.norm_class(config.n_embd, eps=config.norm_eps)
if self.screen_enabled
else None
)
if self.screen_enabled:
screen_config = copy(config)
screen_config.multiscreen_mlp_enabled = False
self.screen = MultiscreenBlock(screen_config, block_idx)
else:
self.screen = None
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
value_residual: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
base_output = self.base(
x,
cos,
sin,
mask,
input_pos,
input_pos_maxp1,
value_residual,
)
if isinstance(base_output, tuple):
x, current_values = base_output
else:
x, current_values = base_output, None
if self.screen is not None:
normalized = self.screen_norm(x)
x = x + (
self.screen(
normalized,
cos,
sin,
mask,
input_pos,
input_pos_maxp1,
)
- normalized
)
if current_values is not None:
return x, current_values
return x
class SandwichSublayer(nn.Module):
"""One attention or FFN sublayer in the official Sandwich ordering.
The source architecture is ``s^k (sf)^(L-k) f^k`` from
https://github.com/ofirpress/sandwich_transformer. This class only adapts
that ordering to LitGPT's pre-norm residual convention.
"""
def __init__(self, config: Config, sublayer_idx: int, kind: str) -> None:
super().__init__()
if kind not in {"s", "f"}:
raise ValueError(f"Unknown Sandwich sublayer kind: {kind!r}")
self.kind = kind
self.config = config
self.sublayer_idx = sublayer_idx
logical_depth = min(sublayer_idx // 2, config.n_layer - 1)
if kind == "s":
# Sandwich starts with ``k`` consecutive attention sublayers, so
# ``sublayer_idx // 2`` maps the first two attention modules to the
# same layer. Use the true attention ordinal for per-layer
# attention schedules (sliding windows, NoPE, and similar flags).
# This is identical to ``logical_depth`` for the ordinary
# interleaved architecture and changes no parameters.
k = config.sandwich_coefficient
attention_idx = sublayer_idx if sublayer_idx < k else k + (sublayer_idx - k) // 2
self.norm = config.norm_class(config.n_embd, eps=config.norm_eps)
self.attn = build_attention(config, attention_idx)
self.mlp = None
self.post_norm = (
config.norm_class(config.n_embd, eps=config.norm_eps)
if config.post_attention_norm
else nn.Identity()
)
else:
self.norm = config.norm_class(config.n_embd, eps=config.norm_eps)
self.attn = None
self.mlp = config.mlp_class(config)
if hasattr(self.mlp, "set_block_index"):
self.mlp.set_block_index(logical_depth, config.n_layer)
self.post_norm = (
config.norm_class(config.n_embd, eps=config.norm_eps) if config.post_mlp_norm else nn.Identity()
)
adaptive_start = 2 * config.n_layer - config.sandwich_adaptive_commit
if kind == "f" and sublayer_idx >= adaptive_start:
# Direct zero initialization consumes no RNG and keeps every shared
# parent parameter and logit bit-identical to the matched control.
self.adaptive_commit_router = nn.Parameter(torch.zeros(config.n_embd))
else:
self.register_parameter("adaptive_commit_router", None)
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
value_residual: torch.Tensor | None = None,
) -> torch.Tensor:
x_normed = self.norm(x)
if self.kind == "s":
update = self.attn(x_normed, cos, sin, mask, input_pos, input_pos_maxp1)
else:
update = self.mlp(x_normed)
update = self.post_norm(update)
if self.adaptive_commit_router is not None:
router_logit = F.linear(
x_normed.float(),
self.adaptive_commit_router.float().unsqueeze(0),
)
commit = (2.0 * torch.sigmoid(router_logit)).to(dtype=update.dtype)
update = commit * update
return x + update
class RMSBudgetedBlock(Block):
"""Transformer block with a protected residual stream and bounded commits.
Strong attention/MLP computation may happen in the branch state, but the
write back to the main residual stream is clamped by an update/input RMS
budget. The block is opt-in through ``Config.rms_budgeted_block`` so the
proven dense and grouped baselines keep their exact execution path.
"""
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__(config, block_idx)
self.block_idx = block_idx
self.rms_budget_dual_stream = bool(config.rms_budget_dual_stream)
self.rms_budget_attn_mlp = bool(config.rms_budget_attn_mlp)
self.rms_budget_direction_aware = bool(config.rms_budget_direction_aware)
self.rms_budget_work_scale = float(config.rms_budget_work_scale)
self.rms_budget_min_scale = float(config.rms_budget_min_scale)
self.rms_budget_exact_target = bool(config.rms_budget_exact_target)
self.rms_budget_target_min = float(config.rms_budget_target_min)
self.rms_budget_target_max = float(config.rms_budget_target_max)
target = float(config.rms_budget_target_init)
target = min(max(target, self.rms_budget_target_min + 1e-6), self.rms_budget_target_max - 1e-6)
if config.rms_budget_target_learnable:
ratio = (target - self.rms_budget_target_min) / (self.rms_budget_target_max - self.rms_budget_target_min)
self.rms_budget_target_raw = nn.Parameter(torch.tensor(math.log(ratio / (1.0 - ratio))))
else:
self.register_buffer("rms_budget_target", torch.tensor(target), persistent=False)
if self.rms_budget_attn_mlp:
self.budget_gate_weight = nn.Parameter(torch.zeros(2, config.n_embd))
self.budget_gate_bias = nn.Parameter(torch.zeros(2))
self.last_residual_ratio: torch.Tensor | None = None
self.last_residual_scale: torch.Tensor | None = None
self.last_attn_budget_mean: torch.Tensor | None = None
self.last_mlp_budget_mean: torch.Tensor | None = None
def _target_ratio(self, x: torch.Tensor) -> torch.Tensor:
if hasattr(self, "rms_budget_target_raw"):
raw = self.rms_budget_target_raw.to(device=x.device, dtype=torch.float32)
target = self.rms_budget_target_min + (self.rms_budget_target_max - self.rms_budget_target_min) * torch.sigmoid(raw)
return target.to(dtype=x.dtype)
return self.rms_budget_target.to(device=x.device, dtype=x.dtype)
def _direction_filter(self, x: torch.Tensor, update: torch.Tensor) -> torch.Tensor:
if not self.rms_budget_direction_aware:
return update
x_float = x.float()
update_float = update.float()
denom = x_float.pow(2).sum(dim=-1, keepdim=True).clamp_min(1e-8)
coeff = (update_float * x_float).sum(dim=-1, keepdim=True) / denom
anti_coeff = coeff.clamp(max=0.0).to(dtype=update.dtype)
return update - anti_coeff * x
def _combine_updates(self, x: torch.Tensor, attention_output: torch.Tensor, mlp_output: torch.Tensor) -> torch.Tensor:
if not self.rms_budget_attn_mlp:
if not torch.is_grad_enabled():
one = x.new_tensor(1.0)
self.last_attn_budget_mean = one
self.last_mlp_budget_mean = one
return attention_output + mlp_output
x_rms = x.float().pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-8)
x_normed = (x.float() / x_rms).to(dtype=x.dtype)
logits = F.linear(
x_normed,
self.budget_gate_weight.to(device=x.device, dtype=x.dtype),
self.budget_gate_bias.to(device=x.device, dtype=x.dtype),
)
weights = torch.softmax(logits.float(), dim=-1).to(dtype=x.dtype) * 2.0
if not torch.is_grad_enabled():
self.last_attn_budget_mean = weights[..., 0].detach().mean()
self.last_mlp_budget_mean = weights[..., 1].detach().mean()
return attention_output * weights[..., 0:1] + mlp_output * weights[..., 1:2]
def _budget_update(self, x: torch.Tensor, update: torch.Tensor) -> torch.Tensor:
update = self._direction_filter(x, update)
input_rms = x.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
update_rms = update.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
target = self._target_ratio(x).float()
desired = target * input_rms
scale = desired / update_rms.clamp_min(1e-8)
if not self.rms_budget_exact_target:
scale = scale.clamp(max=1.0)
if self.rms_budget_min_scale > 0.0:
scale = scale.clamp(min=self.rms_budget_min_scale)
if not torch.is_grad_enabled():
ratio = update_rms / input_rms.clamp_min(1e-8)
self.last_residual_ratio = ratio.detach().mean()
self.last_residual_scale = scale.detach().mean()
return update * scale.to(device=update.device, dtype=update.dtype)
def _mlp_forward_with_hooks(self, mlp_input: torch.Tensor) -> torch.Tensor:
if self.grouped_mlp_rotation_shift:
mlp_input = torch.roll(mlp_input, shifts=self.grouped_mlp_rotation_shift, dims=-1)
mlp_output = self.mlp(mlp_input)
if self.grouped_mlp_rotation_shift:
mlp_output = torch.roll(mlp_output, shifts=-self.grouped_mlp_rotation_shift, dims=-1)
if self.grouped_mlp_mixer is not None:
mix_alpha = self.grouped_mlp_mix_alpha.to(device=mlp_output.device, dtype=mlp_output.dtype)
mlp_output = mlp_output + mix_alpha * self.grouped_mlp_mixer(mlp_output)
return self.post_mlp_norm(mlp_output)
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
) -> torch.Tensor:
x_lm = x
x_normed = self.norm_1(x_lm)
attention_output = self.attn(x_normed, cos, sin, mask, input_pos, input_pos_maxp1)
attention_output = self.post_attention_norm(attention_output)
if self.config.parallel_residual:
if self.rms_budget_dual_stream:
mlp_source = x_lm + self.rms_budget_work_scale * attention_output
else:
mlp_source = x_lm
mlp_input = x_normed if self.config.shared_attention_norm else self.norm_2(mlp_source)
else:
work_scale = self.rms_budget_work_scale if self.rms_budget_dual_stream else 1.0
mlp_input = self.norm_2(x_lm + work_scale * attention_output)
mlp_output = self._mlp_forward_with_hooks(mlp_input)
update = self._combine_updates(x_lm, attention_output, mlp_output)
return x_lm + self._budget_update(x_lm, update)
class _CausalLinearMemory(torch.autograd.Function):
"""Normalized causal linear memory with an explicit reverse-scan backward."""
@staticmethod
def forward(
ctx,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
) -> torch.Tensor:
kv_state = torch.cumsum(
key.unsqueeze(-1) * value.unsqueeze(-2),
dim=1,
)
key_state = torch.cumsum(key, dim=1)
numerator = torch.einsum("btr,btrh->bth", query, kv_state)
denominator_raw = torch.einsum("btr,btr->bt", query, key_state)
denominator = denominator_raw.clamp_min(1e-6)
ctx.save_for_backward(
query,
key,
value,
kv_state,
key_state,
numerator,
denominator,
denominator_raw,
)
return numerator / denominator.unsqueeze(-1)
@staticmethod
def backward(ctx, grad_output: torch.Tensor):
(
query,
key,
value,
kv_state,
key_state,
numerator,
denominator,
denominator_raw,
) = ctx.saved_tensors
grad_numerator = grad_output / denominator.unsqueeze(-1)
grad_denominator = -(
grad_output * numerator
).sum(dim=-1) / denominator.square()
grad_denominator = grad_denominator * (denominator_raw > 1e-6)
grad_query = torch.einsum(
"btrh,bth->btr", kv_state, grad_numerator
)
grad_query = grad_query + key_state * grad_denominator.unsqueeze(-1)
state_contribution = (
query.unsqueeze(-1) * grad_numerator.unsqueeze(-2)
)
grad_state = torch.flip(
torch.cumsum(torch.flip(state_contribution, dims=(1,)), dim=1),
dims=(1,),
)
key_contribution = query * grad_denominator.unsqueeze(-1)
grad_key_state = torch.flip(
torch.cumsum(torch.flip(key_contribution, dims=(1,)), dim=1),
dims=(1,),
)
grad_key = torch.einsum(
"btrh,bth->btr", grad_state, value
) + grad_key_state
grad_value = torch.einsum("btrh,btr->bth", grad_state, key)
return grad_query, grad_key, grad_value
class CausalSelfAttention(nn.Module):
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
self.attention_output_gate = config.attention_output_gate
if self.attention_output_gate == "headwise":
self.gate_size = config.n_head
elif self.attention_output_gate == "elementwise":
self.gate_size = config.n_head * config.head_size
else:
self.gate_size = 0
# key, query and value projections for all heads, but in a batch
self.qkv = nn.Linear(
config.n_embd,
(config.n_head + 2 * config.n_query_groups) * config.head_size + self.gate_size,
bias=config.bias or config.attn_bias,
)
# output projection
self.proj = nn.Linear(config.head_size * config.n_head, config.n_embd, bias=config.bias)
self.dual_path_linear_enabled = (
config.dual_path_linear_enabled
and (block_idx + 1) % config.dual_path_linear_interval == 0
)
self.dual_path_linear_rank = config.dual_path_linear_rank
self.dual_path_linear_explicit_backward = (
config.dual_path_linear_explicit_backward
)
self.dual_path_gate = (
nn.Parameter(
torch.full(
(config.n_head,),
float(config.dual_path_linear_gate_init),
)
)
if self.dual_path_linear_enabled
else None
)
# disabled by default
self.kv_cache: KVCache | MLACompressedKVCache | None = None
self.apply_sliding_window_attention = False
self.configured_sliding_window_attention = False
self.sliding_window_block_mask: BlockMask | None = None
if config.sliding_window_size is not None and config.sliding_window_indices is not None:
self.configured_sliding_window_attention = bool(config.sliding_window_indices[block_idx])
self.apply_sliding_window_attention = self.configured_sliding_window_attention
if config.norm_qk:
norm_q_size = config.n_head * config.head_size if config.norm_qk_type == "olmo2" else config.head_size
norm_k_size = (
config.n_query_groups * config.head_size if config.norm_qk_type == "olmo2" else config.head_size
)
self.norm_q = config.norm_class(norm_q_size, eps=config.norm_eps)
self.norm_k = config.norm_class(norm_k_size, eps=config.norm_eps)
else:
self.norm_q = self.norm_k = None
if config.rope_adjustments is not None:
mscale_all_dim = config.rope_adjustments.get("mscale_all_dim", None)
scaling_factor = config.rope_adjustments.get("factor", None)
if mscale_all_dim and scaling_factor: # YaRN
self.mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
else:
self.mscale = 1.0
else:
self.mscale = 1.0
self.config = config
self.block_idx = block_idx
def prepare_sliding_window_mask(self, sequence_length: int, device: torch.device) -> None:
if not self.configured_sliding_window_attention:
return
window_size = int(self.config.sliding_window_size)
def sliding_causal_mask(
batch_index: torch.Tensor,
head_index: torch.Tensor,
query_index: torch.Tensor,
key_index: torch.Tensor,
) -> torch.Tensor:
del batch_index, head_index
distance = query_index - key_index
return (distance >= 0) & (distance < window_size)
self.sliding_window_block_mask = create_block_mask(
sliding_causal_mask,
B=None,
H=None,
Q_LEN=sequence_length,
KV_LEN=sequence_length,
device=device,
)
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
value_residual: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
# Notation:
# - B | batch size
# - T | time-step (sequence length)
# - C | model's embeddings size (n_embd)
# - C* | attentions's embeddings size
# - hs | head size
# - nh_(q,k,v) | number of heads for query, key and value
# - n_query_groups = nh_k = nh_v | number of query groups sharing key and value heads
# alternative notation: num_kv_groups = n_query_groups
# ┌───┐┌───┐┌───┐┌───┐ ┌───┐ ┌───┐ ┌───┐
# │ v ││ v ││ v ││ v │ │ v │ │ v │ │ v │
# └───┘└───┘└───┘└───┘ └───┘ └───┘ └───┘
# │ │ │ │ │ │ │
# ┌───┐┌───┐┌───┐┌───┐ ┌───┐ ┌───┐ ┌───┐
# │ k ││ k ││ k ││ k │ │ k │ │ k │ │ k │
# └───┘└───┘└───┘└───┘ └───┘ └───┘ └───┘
# │ │ │ │ ┌──┴──┐ ┌──┴──┐ ┌────┬──┴─┬────┐
# ┌───┐┌───┐┌───┐┌───┐ ┌───┐┌───┐┌───┐┌───┐ ┌───┐┌───┐┌───┐┌───┐
# │ q ││ q ││ q ││ q │ │ q ││ q ││ q ││ q │ │ q ││ q ││ q ││ q │
# └───┘└───┘└───┘└───┘ └───┘└───┘└───┘└───┘ └───┘└───┘└───┘└───┘
# ◀──────────────────▶ ◀──────────────────▶ ◀──────────────────▶
# MHA GQA MQA
# n_query_groups=4 n_query_groups=2 n_query_groups=1
#
# credit https://arxiv.org/pdf/2305.13245.pdf
head_size = self.config.head_size
n_head = self.config.n_head
n_query_groups = self.config.n_query_groups
rope_n_elem = self.config.rope_n_elem
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
# Perform a single multiplication operation using a combined QKV matrix to calculate `query`, `key`, and `value`
# instead of individually multiplying the input `x` with the respective weight matrices.
qkv = self.qkv(x) # (B, T, 3xC*)
# Define query, key and value sizes.
# If grouped/multi query is enabled, these sizes are not equal (see the diagram above).
query_size = n_head * head_size
query_projection_size = query_size + self.gate_size
key_size = value_size = n_query_groups * head_size
# Gated Attention appends gate logits to each KV-group's query projection.
q_projection, k, v = qkv.split((query_projection_size, key_size, value_size), dim=-1)
gate_score: torch.Tensor | None = None
if self.attention_output_gate != "none":
queries_per_group = n_head // n_query_groups
q_projection = q_projection.view(B, T, n_query_groups, -1)
gate_features = (
queries_per_group
if self.attention_output_gate == "headwise"
else queries_per_group * head_size
)
q, gate_score = q_projection.split((queries_per_group * head_size, gate_features), dim=-1)
q = q.reshape(B, T, n_head, head_size)
if self.attention_output_gate == "headwise":
gate_score = gate_score.reshape(B, T, n_head, 1)
else:
gate_score = gate_score.reshape(B, T, n_head, head_size)
else:
q = q_projection
current_values = v
if value_residual is not None:
mix = self.config.value_residual_mix
v = mix * value_residual + (1.0 - mix) * v
if self.config.norm_qk and self.config.norm_qk_type == "olmo2":
q = self.norm_q(q)
k = self.norm_k(k)
# To place the num_heads (nh) dimension right after the batch (B) dimension, the first step is to decouple the
# embedding size (C) into num_heads (nh) and head_size (hs).
# The original GQA paper is followed here and the term query groups is used.
# alternative notation: Query groups are also referred to as KV groups.
if self.attention_output_gate == "none":
q = q.view(B, T, n_head, head_size) # (B, T, nh_q, hs)
k = k.view(B, T, n_query_groups, head_size) # (B, T, n_query_groups, hs)
v = v.view(B, T, n_query_groups, head_size) # (B, T, n_query_groups, hs)
# The tensors `query`, `key`, and `value` are now accurately structured: within each batch element (B), there are
# multiple heads (nh), and within each head, there is a sequence of elements (T), each represented by a vector
# of size `hs`.
q = q.transpose(1, 2) # (B, nh_q, T, hs)
k = k.transpose(1, 2) # (B, nh_k, T, hs)
v = v.transpose(1, 2) # (B, nh_v, T, hs)
v_current = v
if self.config.norm_qk and self.config.norm_qk_type == "default":
q = self.norm_q(q)
k = self.norm_k(k)
# SmolLM3 uses a 3:1 RoPE/NoPE pattern by skipping every fourth layer.
use_rope = (
rope_n_elem > 0
and (
self.config.no_rope_layer_interval == 0
or (self.block_idx + 1) % self.config.no_rope_layer_interval != 0
)
)
if use_rope:
if self.config.rope_interleave:
q_roped = apply_rope_interleave(q[..., :rope_n_elem], cos, sin)
k_roped = apply_rope_interleave(k[..., :rope_n_elem], cos, sin)
else:
q_roped = apply_rope(q[..., :rope_n_elem], cos, sin)
k_roped = apply_rope(k[..., :rope_n_elem], cos, sin)
q = torch.cat((q_roped, q[..., rope_n_elem:]), dim=-1) # (B, nh_q, T, hs)
k = torch.cat((k_roped, k[..., rope_n_elem:]), dim=-1) # (B, nh_k, T, hs)
# Apply kv-cache during inference.
if input_pos is not None:
if not isinstance(self.kv_cache, KVCache):
raise TypeError("You need to call `gpt.set_kv_cache()`")
k, v = self.kv_cache(input_pos, k, v)
if input_pos_maxp1 is not None:
# Subselect along sequence dimension
k = k[..., :input_pos_maxp1, :]
v = v[..., :input_pos_maxp1, :]
if self.apply_sliding_window_attention:
# The cache keeps absolute positions so cached prefill and
# decode can use the same semantics as training, including
# prompts longer than one local window. A compact circular
# cache needs chronological reordering and a position-aware
# mask; until that fast path exists, keep correctness exact.
key_position = torch.arange(k.size(2), device=input_pos.device)
window_size = int(self.config.sliding_window_size)
if input_pos.dim() == 1:
distance = input_pos[:, None] - key_position[None, :]
mask = ((distance >= 0) & (distance < window_size))[None, None, :, :]
else:
distance = input_pos[:, :, None] - key_position[None, None, :]
mask = ((distance >= 0) & (distance < window_size))[:, None, :, :]
# k, v: (B, nh_k, input_pos_maxp1, hs)
# If input_pos_maxp1 is None -> max_seq_length
use_flex_window = (
self.apply_sliding_window_attention
and input_pos is None
and self.sliding_window_block_mask is not None
and self.sliding_window_block_mask.shape[-1] == T
and self.config.attention_logit_softcapping is None
)
# Grouped queries: balance the number of heads across all three matrices.
# NOTE: flash attention requires it in training mode.
# Multi-query: this step can be skipped since there is only 1 head, allowing us to use broadcasting.
if n_query_groups != n_head and (input_pos is None or n_query_groups != 1) and not use_flex_window:
q_per_kv = n_head // n_query_groups
k = k.repeat_interleave(q_per_kv, dim=1) # (B, nh_q, T, hs)
v = v.repeat_interleave(q_per_kv, dim=1) # (B, nh_q, T, hs)
if self.apply_sliding_window_attention and not use_flex_window:
"""
Global Window Sliding window Sliding window
attention mask + bias = attention mask
┌────────────────────────┐ ┌───────────────────────┐ ┌─────────────────────────┐
│ True False False False │ │ True True True True │ │ True False False False │
│ True True False False │ │ True True True True │ │ True True False False │
│ True True True False │ │ False True True True │ │ False True True False │
│ True True True True │ │ False False True True │ │ False False True True │
└────────────────────────┘ └───────────────────────┘ └─────────────────────────┘
"""
if input_pos is None:
if mask is None:
mask = torch.ones(T, T, dtype=q.dtype, device=q.device).triu(diagonal=1)
mask.masked_fill_(mask.bool(), float("-inf"))
mask = mask.view(1, 1, *mask.shape)
sliding_window_mask = torch.full((T, T), float("-inf"), dtype=q.dtype, device=q.device)
for i in range(T):
window_start = max(0, i - self.config.sliding_window_size + 1)
sliding_window_mask[i, window_start : i + 1] = 0.0
sliding_window_mask = sliding_window_mask.view(1, 1, T, T)
mask = sliding_window_mask
# Efficient attention using Flash Attention CUDA kernels.
# NOTE: efficient implementation is disabled if `mask` is not None or softcapping is enabled.
# ↓ (B, nh, T, hs) @ (B, nh, T, hs).mT --> (B, nh, T, T) @ (B, nh, T, hs) --> (B, nh, T, hs)
if use_flex_window:
scale = 1.0 / math.sqrt(self.config.attention_scores_scalar or self.config.head_size)
scale = scale * self.mscale * self.mscale
y = flex_attention(
q,
k,
v,
block_mask=self.sliding_window_block_mask,
scale=scale,
enable_gqa=n_query_groups != n_head,
).transpose(1, 2)
else:
y = self.scaled_dot_product_attention(q, k, v, mask)
if self.dual_path_linear_enabled:
# Reuse the already projected MQA features. Positive feature maps
# produce a normalized causal linear-attention recurrence:
# S_t = S_{t-1} + phi(k_t) outer v_t.
rank = self.dual_path_linear_rank
q_feature = F.elu(q[..., :rank].mean(dim=1).float()) + 1.0
k_feature = F.elu(k[:, 0, :, :rank].float()) + 1.0
value_feature = v_current[:, 0].float()
if self.dual_path_linear_explicit_backward:
global_value = _CausalLinearMemory.apply(
q_feature,
k_feature,
value_feature,
)
else:
kv_state = torch.cumsum(
k_feature.unsqueeze(-1) * value_feature.unsqueeze(-2),
dim=1,
)
key_state = torch.cumsum(k_feature, dim=1)
numerator = torch.einsum(
"btr,btrh->bth", q_feature, kv_state
)
denominator = torch.einsum(
"btr,btr->bt", q_feature, key_state
)
global_value = numerator / denominator.clamp_min(
1e-6
).unsqueeze(-1)
global_value = global_value.to(dtype=y.dtype)
gate = torch.tanh(self.dual_path_gate).to(dtype=y.dtype)
y = y + global_value[:, :, None, :] * gate[None, None, :, None]
if gate_score is not None:
y = y * torch.sigmoid(gate_score)
if self.config.xsa_projection:
if v.size(2) == T:
v_projection_base = v.transpose(1, 2)
else:
q_per_kv = n_head // n_query_groups
v_projection_base = v_current.repeat_interleave(q_per_kv, dim=1).transpose(1, 2)
v_projection_base = F.normalize(v_projection_base, dim=-1)
y = y - (y * v_projection_base).sum(dim=-1, keepdim=True) * v_projection_base
# Re-assemble all head outputs side by side.
y = y.reshape(B, T, head_size * n_head)
# Output projection.
output = self.proj(y) # (B, T, C)
if self.config.value_residual_mix > 0.0:
return output, current_values
return output
def scaled_dot_product_attention(
self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, mask: torch.Tensor | None = None
) -> torch.Tensor:
scale = 1.0 / math.sqrt(self.config.attention_scores_scalar or self.config.head_size)
scale = scale * self.mscale * self.mscale
# with softcapping we cannot use SDPA
if self.config.attention_logit_softcapping is not None:
scores = q @ k.mT * scale
scores = do_softcapping(scores, self.config.attention_logit_softcapping)
if mask is None:
mask = torch.ones(q.size(2), q.size(2), dtype=q.dtype, device=q.device).triu(diagonal=1)
mask.masked_fill_(mask.bool(), torch.finfo(q.dtype).min)
scores = scores + mask
scores = F.softmax(scores, dim=-1, dtype=torch.float).to(dtype=q.dtype)
y = scores @ v
else:
y = F.scaled_dot_product_attention(
q, k, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=mask is None
)
return y.transpose(1, 2)
def build_kv_cache(
self,
batch_size: int,
max_seq_length: int,
rope_cache_length: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> "KVCache":
# Keep absolute-position caches for sliding layers. This deliberately
# gives up the potential cache-memory saving until circular-buffer
# decode can preserve chronological order and exact causal masking.
effective_cache_size = max_seq_length
v_shape = (batch_size, self.config.n_query_groups, effective_cache_size, self.config.head_size)
if rope_cache_length is None:
if self.config.rotary_percentage != 1.0:
raise TypeError(
"Please pass the `rope_cache_length` parameter. "
"Use `rope_cache_length=model.rope_cache_length()` to extract it automatically."
)
k_shape = v_shape
else:
k_shape = (
batch_size,
self.config.n_query_groups,
effective_cache_size,
rope_cache_length + self.config.head_size - self.config.rope_n_elem,
)
return KVCache(
k_shape,
v_shape,
device=device,
dtype=dtype,
is_sliding_window=self.apply_sliding_window_attention,
sliding_window_size=self.config.sliding_window_size if self.apply_sliding_window_attention else None,
)
def _load_from_state_dict(self, state_dict: dict, prefix: str, *args: Any, **kwargs: Any) -> None:
"""For compatibility with legacy checkpoints."""
for attr in ("weight", "bias"):
legacy_key = f"{prefix}attn.{attr}"
current_key = f"{prefix}qkv.{attr}"
if legacy_key in state_dict:
state_dict[current_key] = qkv_reassemble(state_dict.pop(legacy_key), self.config)
super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
class SharedDiffCausalSelfAttention(nn.Module):
"""Shared DIFF attention from arXiv:2501.17900.
The differential attention and lambda parameterization follow Microsoft's
reference DIFF implementation. Q and K use the Shared DIFF paper's shared
base projection plus two trainable low-rank updates.
"""
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
self.config = config
self.block_idx = block_idx
self.num_heads = config.n_head // 2
# Config head counts describe the matched baseline. The official DIFF
# geometry halves both query and KV heads while preserving head_dim.
self.num_kv_heads = config.n_query_groups // 2
self.head_dim = config.head_size
self.q_size = self.num_heads * self.head_dim
self.k_size = self.num_kv_heads * self.head_dim
self.v_size = self.num_kv_heads * 2 * self.head_dim
rank = config.shared_diff_rank
self.q_base = nn.Linear(config.n_embd, self.q_size, bias=config.attn_bias)
self.k_base = nn.Linear(config.n_embd, self.k_size, bias=config.attn_bias)
self.v_proj = nn.Linear(config.n_embd, self.v_size, bias=config.attn_bias)
self.proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
self.q_lora_a = nn.Parameter(torch.empty(2, self.num_heads, config.n_embd, rank))
self.q_lora_b = nn.Parameter(torch.empty(2, self.num_heads, rank, self.head_dim))
self.k_lora_a = nn.Parameter(torch.empty(2, self.num_kv_heads, config.n_embd, rank))
self.k_lora_b = nn.Parameter(torch.empty(2, self.num_kv_heads, rank, self.head_dim))
for parameter in (self.q_lora_a, self.q_lora_b, self.k_lora_a, self.k_lora_b):
nn.init.normal_(parameter, mean=0.0, std=0.02)
self.lambda_init = 0.8 - 0.6 * math.exp(-0.3 * block_idx)
self.lambda_q1 = nn.Parameter(torch.empty(self.head_dim).normal_(mean=0.0, std=0.1))
self.lambda_k1 = nn.Parameter(torch.empty(self.head_dim).normal_(mean=0.0, std=0.1))
self.lambda_q2 = nn.Parameter(torch.empty(self.head_dim).normal_(mean=0.0, std=0.1))
self.lambda_k2 = nn.Parameter(torch.empty(self.head_dim).normal_(mean=0.0, std=0.1))
self.subln = RMSNorm(2 * self.head_dim, eps=config.norm_eps)
self.kv_cache: KVCache | None = None
if config.rope_adjustments is not None:
mscale_all_dim = config.rope_adjustments.get("mscale_all_dim")
scaling_factor = config.rope_adjustments.get("factor")
self.mscale = yarn_get_mscale(scaling_factor, mscale_all_dim) if mscale_all_dim and scaling_factor else 1.0
else:
self.mscale = 1.0
@staticmethod
def _low_rank_pair(x: torch.Tensor, a: torch.Tensor, b: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
hidden = torch.einsum("btc,shcr->btshr", x, a)
updates = torch.einsum("btshr,shrd->btshd", hidden, b)
return updates[:, :, 0], updates[:, :, 1]
def _apply_rope(self, tensor: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
rope_n_elem = self.config.rope_n_elem
if self.config.rope_interleave:
roped = apply_rope_interleave(tensor[..., :rope_n_elem], cos, sin)
else:
roped = apply_rope(tensor[..., :rope_n_elem], cos, sin)
return torch.cat((roped, tensor[..., rope_n_elem:]), dim=-1)
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
value_residual: torch.Tensor | None = None,
) -> torch.Tensor:
if value_residual is not None:
raise ValueError("Shared DIFF does not support value residuals.")
batch_size, sequence_length, _ = x.shape
q_delta1, q_delta2 = self._low_rank_pair(x, self.q_lora_a, self.q_lora_b)
k_delta1, k_delta2 = self._low_rank_pair(x, self.k_lora_a, self.k_lora_b)
q_base = self.q_base(x)
k_base = self.k_base(x)
q_base = q_base.view(batch_size, sequence_length, self.num_heads, self.head_dim)
k_base = k_base.view(batch_size, sequence_length, self.num_kv_heads, self.head_dim)
q1 = q_base + q_delta1
q2 = q_base + q_delta2
k1 = k_base + k_delta1
k2 = k_base + k_delta2
v = self.v_proj(x).view(batch_size, sequence_length, self.num_kv_heads, 2 * self.head_dim)
q1 = self._apply_rope(q1.transpose(1, 2), cos, sin)
q2 = self._apply_rope(q2.transpose(1, 2), cos, sin)
k1 = self._apply_rope(k1.transpose(1, 2), cos, sin)
k2 = self._apply_rope(k2.transpose(1, 2), cos, sin)
v = v.transpose(1, 2)
if input_pos is not None:
if not isinstance(self.kv_cache, KVCache):
raise TypeError("You need to call `gpt.set_kv_cache()`")
cached_k, v = self.kv_cache(input_pos, torch.cat((k1, k2), dim=1), v)
k1, k2 = cached_k.split(self.num_kv_heads, dim=1)
if input_pos_maxp1 is not None:
k1 = k1[..., :input_pos_maxp1, :]
k2 = k2[..., :input_pos_maxp1, :]
v = v[..., :input_pos_maxp1, :]
if self.num_kv_heads != self.num_heads:
repeats = self.num_heads // self.num_kv_heads
k1 = k1.repeat_interleave(repeats, dim=1)
k2 = k2.repeat_interleave(repeats, dim=1)
v = v.repeat_interleave(repeats, dim=1)
scale = self.mscale * self.mscale / math.sqrt(self.head_dim)
is_causal = mask is None
attn1 = F.scaled_dot_product_attention(q1, k1, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=is_causal)
attn2 = F.scaled_dot_product_attention(q2, k2, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=is_causal)
lambda1 = torch.exp(torch.sum(self.lambda_q1 * self.lambda_k1).float()).to(dtype=x.dtype)
lambda2 = torch.exp(torch.sum(self.lambda_q2 * self.lambda_k2).float()).to(dtype=x.dtype)
lambda_full = lambda1 - lambda2 + self.lambda_init
output = self.subln(attn1 - lambda_full * attn2) * (1.0 - self.lambda_init)
output = output.transpose(1, 2).reshape(batch_size, sequence_length, self.config.n_embd)
return self.proj(output)
def build_kv_cache(
self,
batch_size: int,
max_seq_length: int,
rope_cache_length: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> "KVCache":
if rope_cache_length is None:
if self.config.rotary_percentage != 1.0:
raise TypeError("Please pass rope_cache_length when rotary_percentage != 1.0")
key_dim = self.head_dim
else:
key_dim = rope_cache_length + self.head_dim - self.config.rope_n_elem
k_shape = (batch_size, 2 * self.num_kv_heads, max_seq_length, key_dim)
v_shape = (batch_size, self.num_kv_heads, max_seq_length, 2 * self.head_dim)
return KVCache(k_shape, v_shape, device=device, dtype=dtype)
class MultiheadLatentAttention(nn.Module):
def __init__(self, config: Config, block_idx: int) -> None:
super().__init__()
self.q_a_proj = nn.Linear(config.n_embd, config.q_lora_rank, bias=config.attn_bias)
self.q_a_norm = RMSNorm(config.q_lora_rank, eps=config.norm_eps)
self.q_b_proj = nn.Linear(config.q_lora_rank, config.n_head * config.qk_head_dim, bias=config.bias)
self.kv_a_proj_with_mqa = nn.Linear(
config.n_embd, config.kv_lora_rank + config.qk_rope_head_dim, bias=config.attn_bias
)
self.kv_a_norm = RMSNorm(config.kv_lora_rank, eps=config.norm_eps)
self.kv_b_proj = nn.Linear(
config.kv_lora_rank,
config.n_query_groups * (config.qk_nope_head_dim + config.v_head_dim),
bias=config.bias,
)
# output projection
self.proj = nn.Linear(config.n_head * config.v_head_dim, config.n_embd, bias=config.bias)
self.output_gate = (
nn.Linear(
config.n_embd,
config.n_head * config.v_head_dim,
bias=False,
)
if config.mla_use_output_gate
else None
)
# disabled by default
self.kv_cache: KVCache | None = None
if config.rope_adjustments is not None:
mscale_all_dim = config.rope_adjustments.get("mscale_all_dim", None)
scaling_factor = config.rope_adjustments.get("factor", None)
if mscale_all_dim and scaling_factor: # YaRN
self.mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
else:
self.mscale = 1.0
else:
self.mscale = 1.0
self.config = config
self.block_idx = block_idx
def forward(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mask: torch.Tensor | None = None,
input_pos: torch.Tensor | None = None,
input_pos_maxp1: int | None = None,
) -> torch.Tensor:
# Notation:
# - B | batch size
# - T | time-step (sequence length)
# - C | model's embeddings size (n_embd)
# - C* | attentions's embeddings size
# - hs | head size
# - nh_(q,k,v) | number of heads for query, key and value
# - n_query_groups = nh_k = nh_v | number of query groups sharing key and value heads
# alternative notation: num_kv_groups = n_query_groups
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
q = self.q_b_proj(self.q_a_norm(self.q_a_proj(x))) # (B, T, n_head * qk_head_dim)
q = q.view(B, T, -1, self.config.qk_head_dim) # (B, T, n_head, qk_head_dim)
q = q.transpose(1, 2) # (B, n_head, T, qk_head_dim)
q_pass, q_rot = torch.split(q, [self.config.qk_nope_head_dim, self.config.qk_rope_head_dim], dim=-1)
compressed_kv = self.kv_a_proj_with_mqa(x) # (B, T, kv_lora_rank + qk_rope_head_dim)
compressed_latent, k_rot = torch.split(
compressed_kv, [self.config.kv_lora_rank, self.config.qk_rope_head_dim], dim=-1
)
compressed_latent = self.kv_a_norm(compressed_latent)
k_rot = k_rot.view(B, 1, T, self.config.qk_rope_head_dim)
if self.config.mla_use_nope:
q_roped = q_rot
k_roped = k_rot
elif self.config.rope_interleave:
q_roped = apply_rope_interleave(q_rot, cos, sin)
k_roped = apply_rope_interleave(k_rot, cos, sin)
else:
q_roped = apply_rope(q_rot, cos, sin)
k_roped = apply_rope(k_rot, cos, sin)
if input_pos is not None and isinstance(self.kv_cache, MLACompressedKVCache):
# The absorbed compressed-cache path is valid for both one-token
# decode and batched speculative verification. Supporting T > 1
# avoids materializing expanded K/V and preserves prior context.
return self.compressed_cache_attention(
q_pass,
q_roped,
compressed_latent,
k_roped,
input_pos,
input_pos_maxp1,
mask,
x,
)
k_pass = self.kv_b_proj(compressed_latent)
k_pass = k_pass.view(B, T, self.config.n_query_groups, -1)
k_pass = k_pass.transpose(1, 2)
k_pass, v = torch.split(k_pass, [self.config.qk_nope_head_dim, self.config.v_head_dim], dim=-1)
k_roped = k_roped.expand(*k_pass.shape[:-1], -1) # (B, n_head, T, qk_rope_head_dim)
q = torch.cat((q_pass, q_roped), dim=-1)
k = torch.cat((k_pass, k_roped), dim=-1)
# Apply kv-cache during inference.
if input_pos is not None and isinstance(self.kv_cache, KVCache):
if not isinstance(self.kv_cache, KVCache):
raise TypeError("You need to call `gpt.set_kv_cache()`")
k, v = self.kv_cache(input_pos, k, v)
if input_pos_maxp1 is not None:
# Subselect along sequence dimension
k = k[..., :input_pos_maxp1, :]
v = v[..., :input_pos_maxp1, :]
# k, v: (B, nh_k, input_pos_maxp1, hs)
# If input_pos_maxp1 is None -> max_seq_length
# Grouped queries: balance the number of heads across all three matrices.
# NOTE: flash attention requires it in training mode.
# Multi-query: this step can be skipped since there is only 1 head, allowing us to use broadcasting.
if self.config.n_query_groups != self.config.n_head and (
input_pos is None or self.config.n_query_groups != 1 or T > 1
):
q_per_kv = self.config.n_head // self.config.n_query_groups
k = k.repeat_interleave(q_per_kv, dim=1) # (B, nh_q, T, hs)
v = v.repeat_interleave(q_per_kv, dim=1) # (B, nh_q, T, hs)
# Efficient attention using Flash Attention CUDA kernels.
# NOTE: efficient implementation is disabled if `mask` is not None or softcapping is enabled.
# ↓ (B, nh, T, hs) @ (B, nh, T, hs).mT --> (B, nh, T, T) @ (B, nh, T, hs) --> (B, nh, T, hs)
y = self.scaled_dot_product_attention(q, k, v, mask)
# Re-assemble all head outputs side by side.
y = y.reshape(B, T, self.config.n_head * self.config.v_head_dim)
if self.output_gate is not None:
y = torch.sigmoid(self.output_gate(x)) * y
# Output projection.
return self.proj(y) # (B, T, C)
def compressed_cache_attention(
self,
q_pass: torch.Tensor,
q_roped: torch.Tensor,
compressed_latent: torch.Tensor,
k_roped: torch.Tensor,
input_pos: torch.Tensor,
input_pos_maxp1: int | None,
mask: torch.Tensor | None,
gate_input: torch.Tensor,
) -> torch.Tensor:
"""Decode from compressed MLA state without materializing per-head K/V."""
if self.kv_b_proj.bias is not None or self.proj.bias is not None:
raise NotImplementedError("Compressed MLA cache currently requires bias=False.")
latent_cache, rope_cache = self.kv_cache(input_pos, compressed_latent, k_roped)
if input_pos_maxp1 is not None:
latent_cache = latent_cache[:, :input_pos_maxp1, :]
rope_cache = rope_cache[..., :input_pos_maxp1, :]
if mask is not None:
mask = mask[..., :input_pos_maxp1]
h = self.config.n_head
g = self.config.n_query_groups
q_per_group = h // g
r = self.config.kv_lora_rank
k_dim = self.config.qk_nope_head_dim
v_dim = self.config.v_head_dim
kv_weight = self.kv_b_proj.weight.view(g, k_dim + v_dim, r)
k_up = kv_weight[:, :k_dim, :].repeat_interleave(q_per_group, dim=0)
v_up = kv_weight[:, k_dim:, :].repeat_interleave(q_per_group, dim=0)
latent_q = torch.einsum("bhtd,hdr->bhtr", q_pass, k_up)
scores = torch.einsum("bhtr,bsr->bhts", latent_q, latent_cache)
scores = scores + torch.einsum("bhtd,bnsd->bhts", q_roped, rope_cache)
scale = self.mscale * self.mscale / math.sqrt(
self.config.attention_scores_scalar or self.config.qk_head_dim
)
scores = scores * scale
if mask is not None:
scores = scores.masked_fill(~mask, torch.finfo(scores.dtype).min) if mask.dtype == torch.bool else scores + mask
weights = F.softmax(scores, dim=-1, dtype=torch.float).to(dtype=scores.dtype)
latent_context = torch.einsum("bhts,bsr->bhtr", weights, latent_cache)
if self.output_gate is not None:
context = torch.einsum("bhtr,hvr->bhtv", latent_context, v_up)
context = context.transpose(1, 2).reshape(
gate_input.size(0),
gate_input.size(1),
h * v_dim,
)
context = torch.sigmoid(self.output_gate(gate_input)) * context
return self.proj(context)
out_weight = self.proj.weight.view(self.config.n_embd, h, v_dim)
absorbed_v_o = torch.einsum("ohv,hvr->hor", out_weight, v_up)
return torch.einsum("bhtr,hor->bto", latent_context, absorbed_v_o)
def scaled_dot_product_attention(
self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, mask: torch.Tensor | None = None
) -> torch.Tensor:
scale = 1.0 / math.sqrt(self.config.attention_scores_scalar or self.config.qk_head_dim)
scale = scale * self.mscale * self.mscale
# with softcapping we cannot use SDPA
if self.config.attention_logit_softcapping is not None:
scores = q @ k.mT * scale
scores = do_softcapping(scores, self.config.attention_logit_softcapping)
if mask is None:
mask = torch.ones(q.size(2), q.size(2), dtype=q.dtype, device=q.device).triu(diagonal=1)
mask.masked_fill_(mask.bool(), torch.finfo(q.dtype).min)
scores = scores + mask
scores = F.softmax(scores, dim=-1, dtype=torch.float).to(dtype=q.dtype)
y = scores @ v
else:
y = F.scaled_dot_product_attention(
q, k, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=mask is None
)
return y.transpose(1, 2)
def build_kv_cache(
self,
batch_size: int,
max_seq_length: int,
rope_cache_length: int | None = None,
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> "MLACompressedKVCache":
latent_shape = (batch_size, max_seq_length, self.config.kv_lora_rank)
rope_shape = (batch_size, 1, max_seq_length, self.config.qk_rope_head_dim)
if rope_cache_length is not None:
print("Warning: `rope_cache_length` has no effect on MultiheadLatentAttention!")
if self.config.rotary_percentage != 1.0:
print("Warning: `rotary_percentage` has no effect on MultiheadLatentAttention!")
return MLACompressedKVCache(latent_shape, rope_shape, device=device, dtype=dtype)
class GptNeoxMLP(nn.Module):
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.fc = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.proj = nn.Linear(self.intermediate_size, config.n_embd, bias=config.bias)
self.config = config
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.fc(x)
x = F.gelu(x, approximate=self.config.gelu_approximate)
return self.proj(x)
def powlu_gate(x: torch.Tensor, m: float) -> torch.Tensor:
positive_mask = x > 0
positive = torch.where(positive_mask, x, torch.ones_like(x))
exponent = float(m) / (torch.sqrt(positive) + 1.0)
positive_gate = torch.pow(positive, exponent) * torch.sigmoid(x)
return torch.where(positive_mask, positive_gate, F.silu(x))
class LLaMAMLP(nn.Module):
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.fc_1 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.fc_2 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.proj = nn.Linear(self.intermediate_size, config.n_embd, bias=config.bias)
self.config = config
def forward(self, x: torch.Tensor) -> torch.Tensor:
x_fc_1 = self.fc_1(x)
x_fc_2 = self.fc_2(x)
x = F.silu(x_fc_1) * x_fc_2
return self.proj(x)
class KimiSiTUGLUMLP(LLaMAMLP):
"""Kimi K3 Eq. 12 dense SiTU-GLU.
Both multiplicative branches are smoothly capped in FP32 using the
constants released by Moonshot (beta_gate=4, beta_up=25).
"""
def forward(self, x: torch.Tensor) -> torch.Tensor:
gate = self.fc_1(x).float()
up = self.fc_2(x).float()
beta_gate = self.config.kimi_situ_beta
beta_up = self.config.kimi_situ_linear_beta
gate = beta_gate * torch.tanh(gate / beta_gate) * torch.sigmoid(gate)
up = beta_up * torch.tanh(up / beta_up)
return self.proj((gate * up).to(x.dtype))
class _KimiSiTUExpert(nn.Module):
def __init__(self, input_size: int, intermediate_size: int, config: Config) -> None:
super().__init__()
self.gate = nn.Linear(input_size, intermediate_size, bias=False)
self.up = nn.Linear(input_size, intermediate_size, bias=False)
self.down = nn.Linear(intermediate_size, input_size, bias=False)
self.beta_gate = config.kimi_situ_beta
self.beta_up = config.kimi_situ_linear_beta
def forward(self, x: torch.Tensor) -> torch.Tensor:
gate = self.gate(x).float()
up = self.up(x).float()
gate = self.beta_gate * torch.tanh(gate / self.beta_gate) * torch.sigmoid(gate)
up = self.beta_up * torch.tanh(up / self.beta_up)
return self.down((gate * up).to(x.dtype))
class KimiStableLatentMoEMLP(nn.Module):
"""Training-capable Kimi K3 Stable LatentMoE (Eqs. 11-14).
The routed experts operate at latent width, their weighted sum is
RMS-normalized before the full-width up projection, two full-width shared
SiTU experts are always active, and the next batch's dispatch bias is
updated with exact single-device Quantile Balancing.
"""
def __init__(self, config: Config) -> None:
super().__init__()
d = config.n_embd
latent = config.kimi_latent_moe_latent_size
n_experts = config.kimi_latent_moe_num_experts
routed_hidden = config.kimi_latent_moe_expert_intermediate_size
shared_hidden = (
config.kimi_latent_moe_shared_intermediate_size
* config.kimi_latent_moe_num_shared_experts
)
self.config = config
self.router = nn.Linear(d, n_experts, bias=False)
self.routed_down = nn.Linear(d, latent, bias=False)
self.routed_up = nn.Linear(latent, d, bias=False)
self.routed_norm = config.norm_class(latent, eps=config.norm_eps)
# Pack all routed experts so selected experts execute as batched
# matmuls. This is algebraically identical to a ModuleList of
# independent SiTU experts, but avoids one Python/CUDA launch chain
# per expert.
self.routed_gate_weight = nn.Parameter(
torch.empty(n_experts, routed_hidden, latent)
)
self.routed_up_weight = nn.Parameter(
torch.empty(n_experts, routed_hidden, latent)
)
self.routed_down_weight = nn.Parameter(
torch.empty(n_experts, latent, routed_hidden)
)
for parameter in (
self.routed_gate_weight,
self.routed_up_weight,
self.routed_down_weight,
):
nn.init.normal_(parameter, mean=0.0, std=0.02)
# Concatenating the two experts' hidden axes and using one down
# projection is algebraically identical to summing two independent
# full-width GLUs, while issuing three large GEMMs instead of six.
self.shared_experts = _KimiSiTUExpert(d, shared_hidden, config)
self.register_buffer("quantile_bias", torch.zeros(n_experts))
@torch.no_grad()
def _next_quantile_bias(
self,
scores: torch.Tensor,
biased_scores: torch.Tensor,
) -> torch.Tensor:
k = self.config.kimi_latent_moe_top_k
experts = self.config.kimi_latent_moe_num_experts
bins = self.config.kimi_latent_moe_qb_bins
cutoff = biased_scores.topk(k + 1, dim=-1, sorted=True).values[:, k]
# Appendix D: histogram the required bias r_ij = alpha_i - s_ij
# over [b_min - 1, b_max + 1], then recover its k/n quantile.
required_bias = cutoff[:, None] - scores
lower = self.quantile_bias.min().float() - 1.0
upper = self.quantile_bias.max().float() + 1.0
bin_width = (upper - lower) / bins
bin_index = torch.floor(
(required_bias - lower) / bin_width
).to(torch.long).clamp_(0, bins - 1)
expert_offset = (
torch.arange(experts, device=scores.device, dtype=torch.long) * bins
)
flat_index = (bin_index + expert_offset).reshape(-1)
histogram = torch.zeros(
experts * bins,
device=scores.device,
dtype=torch.int32,
)
histogram.scatter_add_(
0,
flat_index,
torch.ones_like(flat_index, dtype=torch.int32),
)
histogram = histogram.view(experts, bins)
cumulative = histogram.cumsum(dim=1)
target_load = (scores.size(0) * k + experts - 1) // experts
selected_bin = (cumulative >= target_load).to(torch.int32).argmax(dim=1)
selected_count = histogram.gather(1, selected_bin[:, None]).squeeze(1)
prior_bin = (selected_bin - 1).clamp_min(0)
prior_count = cumulative.gather(1, prior_bin[:, None]).squeeze(1)
prior_count = torch.where(
selected_bin == 0,
torch.zeros_like(prior_count),
prior_count,
)
fraction = (
(target_load - prior_count).float()
/ selected_count.clamp_min(1).float()
).clamp_(0.0, 1.0)
next_bias = lower + (selected_bin.float() + fraction) * bin_width
return next_bias - next_bias.mean()
def forward(self, x: torch.Tensor) -> torch.Tensor:
shape = x.shape
flat = x.reshape(-1, shape[-1])
scores = torch.sigmoid(self.router(flat).float())
biased_scores = scores + self.quantile_bias.float()
k = self.config.kimi_latent_moe_top_k
indices = biased_scores.topk(k, dim=-1, sorted=False).indices
weights = scores.gather(1, indices)
weights = weights / weights.sum(dim=-1, keepdim=True).clamp_min(1e-20)
latent = self.routed_down(flat)
token_count, route_count = indices.shape
route_input = latent[:, None, :].expand(
token_count,
route_count,
latent.size(-1),
).reshape(-1, latent.size(-1))
flat_indices = indices.reshape(-1)
gate = torch.bmm(
self.routed_gate_weight[flat_indices],
route_input.unsqueeze(-1),
).squeeze(-1).float()
up = torch.bmm(
self.routed_up_weight[flat_indices],
route_input.unsqueeze(-1),
).squeeze(-1).float()
gate = (
self.config.kimi_situ_beta
* torch.tanh(gate / self.config.kimi_situ_beta)
* torch.sigmoid(gate)
)
up = self.config.kimi_situ_linear_beta * torch.tanh(
up / self.config.kimi_situ_linear_beta
)
expert_hidden = (gate * up).to(x.dtype)
expert_output = torch.bmm(
self.routed_down_weight[flat_indices],
expert_hidden.unsqueeze(-1),
).squeeze(-1)
expert_output = expert_output * weights.reshape(-1, 1).to(x.dtype)
routed = torch.zeros_like(latent)
routed.index_add_(
0,
torch.arange(token_count, device=x.device).repeat_interleave(
route_count
),
expert_output,
)
if self.training and self.config.kimi_latent_moe_quantile_balancing:
self.quantile_bias.copy_(
self._next_quantile_bias(scores.detach(), biased_scores.detach())
)
routed = self.routed_up(self.routed_norm(routed))
shared = self.shared_experts(flat)
return (shared + routed).view(shape)
class LLaMAPowLUMLP(LLaMAMLP):
def forward(self, x: torch.Tensor) -> torch.Tensor:
x_fc_1 = self.fc_1(x)
x_fc_2 = self.fc_2(x)
x = powlu_gate(x_fc_1, self.config.powlu_m) * x_fc_2
return self.proj(x)
class DSwiGLUMLP(nn.Module):
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.fc_1 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.fc_2 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.fc_3 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.proj = nn.Linear(self.intermediate_size, config.n_embd, bias=config.bias)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
def _hard_topk_gate(self, logits: torch.Tensor) -> torch.Tensor:
active = max(1, min(self.intermediate_size, int(round(self.intermediate_size * self.config.drelu_target_active))))
threshold = torch.topk(logits.float(), k=active, dim=-1).values[..., -1:].detach()
centered = logits.float() - threshold
soft = torch.sigmoid(self.config.drelu_gate_sharpness * centered).to(dtype=logits.dtype)
hard = (centered >= 0).to(dtype=logits.dtype)
return hard + soft - soft.detach()
def forward(self, x: torch.Tensor) -> torch.Tensor:
representation = F.silu(self.fc_1(x)) * self.fc_2(x)
gate = self._hard_topk_gate(self.fc_3(x))
self.last_gate_sparsity = (gate.detach() == 0).float().mean()
return self.proj(representation * gate)
class AdaptiveDSwiGLUMLP(nn.Module):
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.fc_1 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.fc_2 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.fc_3 = nn.Linear(config.n_embd, self.intermediate_size, bias=config.bias)
self.threshold = nn.Linear(config.n_embd, 1, bias=True)
self.proj = nn.Linear(self.intermediate_size, config.n_embd, bias=config.bias)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
init = min(max(config.drelu_threshold_init, 1e-4), 1 - 1e-4)
torch.nn.init.zeros_(self.threshold.weight)
torch.nn.init.constant_(self.threshold.bias, math.log(init / (1 - init)))
def _adaptive_gate(self, logits: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
probabilities = torch.sigmoid(logits.float())
threshold = torch.sigmoid(self.threshold(x).float())
centered = probabilities - threshold
soft = torch.sigmoid(self.config.drelu_gate_sharpness * centered).to(dtype=logits.dtype)
hard = (centered >= 0).to(dtype=logits.dtype)
self.last_threshold_mean = threshold.detach().mean()
return hard + soft - soft.detach()
def forward(self, x: torch.Tensor) -> torch.Tensor:
representation = F.silu(self.fc_1(x)) * self.fc_2(x)
gate = self._adaptive_gate(self.fc_3(x), x)
self.last_gate_sparsity = (gate.detach() == 0).float().mean()
return self.proj(representation * gate)
class BlockSparseAdaptiveDSwiGLUMLP(nn.Module):
"""Adaptive dSwiGLU with structured conditional compute.
The dense dSwiGLU variants compute every SwiGLU channel and then zero out
activations. This variant splits the intermediate MLP into groups and only
runs the selected groups for each token. The Python implementation is meant
for laptop-scale research and correctness; production speedups need a fused
grouped kernel.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.min_active_groups = config.sparse_mlp_min_active_groups
self.max_active_groups = config.sparse_mlp_max_active_groups
if self.intermediate_size % self.num_groups != 0:
raise ValueError("intermediate_size must be divisible by sparse_mlp_num_groups")
self.group_size = self.intermediate_size // self.num_groups
if config.bias:
raise ValueError("BlockSparseAdaptiveDSwiGLUMLP currently expects bias=False for grouped GEMM.")
self.up_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.group_size))
self.value_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.group_size))
self.down_weight = nn.Parameter(torch.empty(self.num_groups, self.group_size, config.n_embd))
self.router = nn.Linear(config.n_embd, self.num_groups, bias=True)
self.threshold = nn.Linear(config.n_embd, 1, bias=True)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
init = min(max(config.drelu_threshold_init, 1e-4), 1 - 1e-4)
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
self.router.reset_parameters()
torch.nn.init.zeros_(self.threshold.weight)
init = min(max(self.config.drelu_threshold_init, 1e-4), 1 - 1e-4)
torch.nn.init.constant_(self.threshold.bias, math.log(init / (1 - init)))
def _active_group_mask(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
probabilities = torch.sigmoid(self.router(x).float())
threshold = torch.sigmoid(self.threshold(x).float())
_, top_indices = torch.topk(probabilities, k=self.max_active_groups, dim=-1)
top_mask = torch.zeros_like(probabilities, dtype=torch.bool)
top_mask.scatter_(-1, top_indices, True)
threshold_mask = probabilities >= threshold
active_mask = threshold_mask & top_mask
active_count = active_mask.sum(dim=-1)
if self.min_active_groups > 0:
_, min_indices = torch.topk(probabilities, k=self.min_active_groups, dim=-1)
min_mask = torch.zeros_like(active_mask)
min_mask.scatter_(-1, min_indices, True)
active_mask = active_mask | (active_count.unsqueeze(-1) < self.min_active_groups) & min_mask
self.last_gate_sparsity = 1.0 - active_mask.detach().float().mean()
self.last_threshold_mean = threshold.detach().mean()
self.last_active_groups_mean = active_mask.detach().float().sum(dim=-1).mean()
return active_mask, probabilities.to(dtype=x.dtype), threshold.to(dtype=x.dtype)
def _exact_topk_pairs(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
probabilities = torch.sigmoid(self.router(x).float())
top_values, top_indices = torch.topk(probabilities, k=self.max_active_groups, dim=-1)
self.last_gate_sparsity = x.new_tensor(1.0 - self.max_active_groups / self.num_groups)
self.last_threshold_mean = torch.sigmoid(self.threshold(x).float()).detach().mean()
self.last_active_groups_mean = x.new_tensor(float(self.max_active_groups))
token_idx = torch.arange(x.shape[0], device=x.device).repeat_interleave(self.max_active_groups)
return token_idx, top_indices.reshape(-1), top_values.reshape(-1).to(dtype=x.dtype)
@torch._dynamo.disable
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
output = x_flat.new_zeros(x_flat.shape)
if self.min_active_groups == self.max_active_groups:
token_idx, group_idx, scale_values = self._exact_topk_pairs(x_flat)
else:
active_mask, probabilities, _ = self._active_group_mask(x_flat)
active_pairs = torch.nonzero(active_mask, as_tuple=False)
if active_pairs.numel() == 0:
return output.reshape(original_shape)
token_idx = active_pairs[:, 0]
group_idx = active_pairs[:, 1]
scale_values = probabilities[token_idx, group_idx]
order = torch.argsort(group_idx)
token_idx = token_idx.index_select(0, order)
group_idx = group_idx.index_select(0, order)
scale_values = scale_values.index_select(0, order)
selected_x = x_flat.index_select(0, token_idx).contiguous()
counts = torch.bincount(group_idx, minlength=self.num_groups)
offsets = torch.cumsum(counts, dim=0).to(dtype=torch.int32)
if not selected_x.is_cuda:
for group_idx_int in range(self.num_groups):
pair_idx = torch.nonzero(group_idx == group_idx_int, as_tuple=False).flatten()
if pair_idx.numel() == 0:
continue
group_token_idx = token_idx.index_select(0, pair_idx)
x_group = x_flat.index_select(0, group_token_idx)
hidden = F.silu(x_group @ self.up_weight[group_idx_int]) * (x_group @ self.value_weight[group_idx_int])
projected = hidden @ self.down_weight[group_idx_int]
scale = scale_values.index_select(0, pair_idx).unsqueeze(-1)
output.index_add_(0, group_token_idx, projected * scale.to(dtype=projected.dtype))
return output.reshape(original_shape)
grouped_dtype = torch.bfloat16 if selected_x.is_cuda else selected_x.dtype
selected_x_gemm = selected_x.to(dtype=grouped_dtype)
up_weight = self.up_weight.to(dtype=grouped_dtype)
value_weight = self.value_weight.to(dtype=grouped_dtype)
down_weight = self.down_weight.to(dtype=grouped_dtype)
up = torch._grouped_mm(selected_x_gemm, up_weight, offsets)
value = torch._grouped_mm(selected_x_gemm, value_weight, offsets)
hidden = F.silu(up) * value
projected = torch._grouped_mm(hidden, down_weight, offsets).to(dtype=output.dtype)
scale = scale_values.unsqueeze(-1).to(dtype=projected.dtype)
output.index_add_(0, token_idx, projected * scale)
return output.reshape(original_shape)
class HiddenBlockDSwiGLUMLP(nn.Module):
"""Parameter-matched hidden-channel block SwiGLU.
Unlike ``TileRoutedDSwiGLUMLP``, which gives every group the full residual
stream and partitions the intermediate channels, this module partitions the
residual stream itself into blocks. Each block owns a local SwiGLU with the
full configured intermediate width, so total parameter count and matmul work
are close to a dense dSwiGLU at the same ``intermediate_size``.
``grouped_mlp_stack_alpha`` turns the parallel blocks into a small cascade:
each later block can see a scaled residual update from the previous block.
The existing block-level channel rotation hook can rotate hidden channels
across transformer layers, preventing the same channels from being trapped
in the same local block for the whole network.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_blocks = config.sparse_mlp_num_groups
self.active_groups = config.sparse_mlp_max_active_groups
if self.active_groups != self.num_blocks:
raise ValueError("HiddenBlockDSwiGLUMLP is a full-active block variant.")
if config.n_embd % self.num_blocks != 0:
raise ValueError("n_embd must be divisible by sparse_mlp_num_groups")
if config.bias:
raise ValueError("HiddenBlockDSwiGLUMLP currently expects bias=False.")
self.block_hidden_size = config.n_embd // self.num_blocks
self.up_weight = nn.Parameter(torch.empty(self.num_blocks, self.block_hidden_size, self.intermediate_size))
self.value_weight = nn.Parameter(torch.empty(self.num_blocks, self.block_hidden_size, self.intermediate_size))
self.down_weight = nn.Parameter(torch.empty(self.num_blocks, self.intermediate_size, self.block_hidden_size))
stack_alpha = torch.tensor(float(config.grouped_mlp_stack_alpha))
if config.grouped_mlp_stack_alpha_learnable:
self.stack_alpha = nn.Parameter(stack_alpha)
else:
self.register_buffer("stack_alpha", stack_alpha, persistent=False)
output_scale = torch.tensor(float(config.grouped_mlp_stack_output_scale))
if config.grouped_mlp_stack_output_scale_learnable:
self.stack_output_scale = nn.Parameter(output_scale)
else:
self.register_buffer("stack_output_scale", output_scale, persistent=False)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
std *= float(getattr(self.config, "grouped_mlp_stack_init_scale", 1.0))
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
x_blocks = x_flat.reshape(x_flat.shape[0], self.num_blocks, self.block_hidden_size)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
deltas: list[torch.Tensor] = []
carry: torch.Tensor | None = None
for block_idx in range(self.num_blocks):
state = x_blocks[:, block_idx, :]
if carry is not None:
state = state + stack_alpha * carry
hidden = F.silu(state @ self.up_weight[block_idx]) * (state @ self.value_weight[block_idx])
carry = hidden @ self.down_weight[block_idx]
deltas.append(carry)
output = torch.stack(deltas, dim=1).reshape(original_shape)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
output_scale = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
return output * output_scale
class TileRoutedDSwiGLUMLP(nn.Module):
"""Tile-routed grouped dSwiGLU.
This is the hardware-aligned sparse MLP experiment. It keeps one active MLP
group per token, but routes contiguous token tiles instead of individual
tokens. The forward is expressed as batched GEMMs so backward uses efficient
dense kernels instead of per-token atomic accumulation.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
if self.intermediate_size % self.num_groups != 0:
raise ValueError("intermediate_size must be divisible by sparse_mlp_num_groups")
if config.bias:
raise ValueError("TileRoutedDSwiGLUMLP currently expects bias=False.")
self.group_size = self.intermediate_size // self.num_groups
self.active_groups = config.sparse_mlp_max_active_groups
self.up_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.group_size))
self.value_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.group_size))
self.down_weight = nn.Parameter(torch.empty(self.num_groups, self.group_size, config.n_embd))
self.shared_path: LLaMAMLP | None = None
if config.grouped_mlp_shared_intermediate_size > 0:
self.shared_path = LLaMAMLP(config, intermediate_size=config.grouped_mlp_shared_intermediate_size)
shared_alpha = torch.tensor(float(config.grouped_mlp_shared_alpha))
if config.grouped_mlp_shared_alpha_learnable:
self.shared_alpha = nn.Parameter(shared_alpha)
else:
self.register_buffer("shared_alpha", shared_alpha, persistent=False)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
def _apply_shared_path(self, x: torch.Tensor, grouped_output: torch.Tensor) -> torch.Tensor:
if self.shared_path is None:
return grouped_output
shared_alpha = self.shared_alpha.to(device=grouped_output.device, dtype=grouped_output.dtype)
return grouped_output + shared_alpha * self.shared_path(x)
def _gate(self, up: torch.Tensor) -> torch.Tensor:
return F.silu(up)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
# Full-active grouped MLP is mathematically a dense SwiGLU whose
# intermediate channels are partitioned into groups. Collapse grouped
# parameter tensors into dense GEMM views for the 4/4 quality path.
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
output = (self._gate(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
return output.reshape(original_shape)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
if self.active_groups == self.num_groups:
output = self._forward_full_active(x, original_shape, x_flat)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return self._apply_shared_path(x, output)
usable_tokens = x_flat.shape[0] - (x_flat.shape[0] % self.num_groups)
output = x_flat.new_empty(x_flat.shape)
if usable_tokens:
x_tiles = x_flat[:usable_tokens].reshape(self.num_groups, usable_tokens // self.num_groups, x_flat.shape[-1])
out_tiles = torch.zeros_like(x_tiles)
for route_offset in range(self.active_groups):
if route_offset == 0:
up_weight = self.up_weight
value_weight = self.value_weight
down_weight = self.down_weight
else:
up_weight = torch.roll(self.up_weight, shifts=-route_offset, dims=0)
value_weight = torch.roll(self.value_weight, shifts=-route_offset, dims=0)
down_weight = torch.roll(self.down_weight, shifts=-route_offset, dims=0)
up = torch.bmm(x_tiles, up_weight)
value = torch.bmm(x_tiles, value_weight)
out_tiles = out_tiles + torch.bmm(self._gate(up) * value, down_weight)
output[:usable_tokens] = out_tiles.reshape(usable_tokens, x_flat.shape[-1])
if usable_tokens < x_flat.shape[0]:
tail = x_flat[usable_tokens:]
tail_out = torch.zeros_like(tail)
for route_offset in range(self.active_groups):
hidden = self._gate(tail @ self.up_weight[route_offset]) * (tail @ self.value_weight[route_offset])
tail_out = tail_out + hidden @ self.down_weight[route_offset]
output[usable_tokens:] = tail_out
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(1.0 - self.active_groups / self.num_groups)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return self._apply_shared_path(x, output.reshape(original_shape))
class TileRoutedActivationDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped GLU with a configurable gate activation."""
gate_kind = "silu"
def _gate(self, up: torch.Tensor) -> torch.Tensor:
if self.gate_kind == "gelu":
return F.gelu(up, approximate=self.config.gelu_approximate)
if self.gate_kind == "relu":
return F.relu(up)
if self.gate_kind == "squared_relu":
relu = F.relu(up)
return relu * relu
if self.gate_kind == "sigmoid":
return torch.sigmoid(up)
if self.gate_kind == "powlu":
return powlu_gate(up, self.config.powlu_m)
return F.silu(up)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = self._gate(x_flat @ up_weight) * (x_flat @ value_weight)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedGEGLUMLP(TileRoutedActivationDSwiGLUMLP):
gate_kind = "gelu"
class TileRoutedReGLUMLP(TileRoutedActivationDSwiGLUMLP):
gate_kind = "relu"
class TileRoutedSquaredReGLUMLP(TileRoutedActivationDSwiGLUMLP):
gate_kind = "squared_relu"
class TileRoutedSigmoidGLUMLP(TileRoutedActivationDSwiGLUMLP):
gate_kind = "sigmoid"
class TileRoutedPowLUGLUMLP(TileRoutedActivationDSwiGLUMLP):
gate_kind = "powlu"
class TileRoutedEulerGLUMLP(TileRoutedActivationDSwiGLUMLP):
"""Grouped SwiGLU with a normalized Euler maximum envelope.
For ``p = softplus(u)``, ``p ** (1 / p)`` reaches its unique global
maximum at ``p = e``. Normalizing by ``e ** (1 / e)`` bounds the positive
modulation by one while SiLU retains the signed gate behavior.
"""
def _gate(self, up: torch.Tensor) -> torch.Tensor:
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
return F.silu(up) * envelope
class TileRoutedEulerMishBlendGLUMLP(TileRoutedEulerGLUMLP):
"""Euler gate with a zero-start, per-group Mish residual.
The gate begins as the fixed Euler gate exactly. Each contiguous 4/4
group can learn a bounded interpolation toward Mish, preserving the proven
positive Euler envelope while exposing Mish's smooth negative curvature.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Euler-Mish blend requires all groups to be active.")
self.mish_residual = nn.Parameter(torch.zeros(self.num_groups))
def _gate(self, up: torch.Tensor) -> torch.Tensor:
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
euler_gate = F.silu(up) * envelope
mish_gate = up * torch.tanh(p).to(dtype=up.dtype)
mix = torch.tanh(self.mish_residual).repeat_interleave(self.group_size).to(dtype=up.dtype)
return euler_gate + mix * (mish_gate - euler_gate)
class TileRoutedContextualValueRotationEulerGLUMLP(TileRoutedEulerGLUMLP):
"""Euler 4/4 MLP with contextual orthogonal value rotations.
The 1664-channel intermediate is interpreted as 13 aligned tiles, each
containing four 32-channel branches. Four response-conditioned butterfly
rotations mix the branches before the existing dense down projection.
All controller coefficients start at zero, making this exactly the fixed
Euler gate at initialization while retaining nonzero controller gradients.
"""
_tile_size = 32
_rotation_pairs = ((0, 1), (2, 3), (0, 2), (1, 3))
_max_half_angle = math.tan(math.pi / 24.0)
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups or self.num_groups != 4:
raise ValueError("Contextual value rotation requires full-active grouped 4/4.")
if self.group_size % self._tile_size:
raise ValueError("Contextual value rotation requires a group size divisible by 32.")
self.num_value_tiles = self.group_size // self._tile_size
self.cvr_energy_coeff = nn.Parameter(torch.zeros(len(self._rotation_pairs)))
self.cvr_agreement_coeff = nn.Parameter(torch.zeros(len(self._rotation_pairs)))
def _rotate_pair(
self,
left: torch.Tensor,
right: torch.Tensor,
pair_index: int,
) -> tuple[torch.Tensor, torch.Tensor]:
left_float = left.float()
right_float = right.float()
epsilon = torch.finfo(torch.float32).eps
left_energy = left_float.square().mean(dim=-1).clamp_min(epsilon)
right_energy = right_float.square().mean(dim=-1).clamp_min(epsilon)
energy_contrast = torch.tanh(0.5 * (torch.log(left_energy) - torch.log(right_energy)))
agreement = (left_float * right_float).mean(dim=-1) * torch.rsqrt(left_energy * right_energy)
agreement = agreement.clamp(-1.0, 1.0)
controller = (
self.cvr_energy_coeff[pair_index].float() * energy_contrast
+ self.cvr_agreement_coeff[pair_index].float() * agreement
)
half_angle = self._max_half_angle * torch.tanh(controller)
half_angle_squared = half_angle.square()
denominator = 1.0 + half_angle_squared
cosine = ((1.0 - half_angle_squared) / denominator).to(dtype=left.dtype).unsqueeze(-1)
sine = ((2.0 * half_angle) / denominator).to(dtype=left.dtype).unsqueeze(-1)
return cosine * left - sine * right, sine * left + cosine * right
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(
x_flat.shape[0], self.num_groups, self.num_value_tiles, self._tile_size
)
groups = list(hidden.unbind(dim=1))
for pair_index, (left_index, right_index) in enumerate(self._rotation_pairs):
groups[left_index], groups[right_index] = self._rotate_pair(
groups[left_index], groups[right_index], pair_index
)
hidden = torch.stack(groups, dim=1).reshape(x_flat.shape[0], self.intermediate_size)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedContextAdaptiveEulerTemperatureGLUMLP(TileRoutedEulerGLUMLP):
"""Euler 4/4 gate with a bounded, response-conditioned temperature.
The normalized Euler log-envelope is non-positive and has its unique zero
at ``softplus(up) = e``. Each group learns how strongly to sharpen one
side of that maximum while relaxing the other, using the bounded signed
displacement from ``e`` as local context. Zero parameters reproduce the
fixed Euler gate exactly, and the temperature always stays in [0.5, 1.5],
so the envelope remains bounded by one with the same global maximum.
"""
_max_temperature_delta = 0.5
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups or self.num_groups != 4:
raise ValueError("Context-adaptive Euler temperature requires full-active grouped 4/4.")
self.caet_temperature_raw = nn.Parameter(torch.zeros(self.num_groups))
def _gate(self, up: torch.Tensor) -> torch.Tensor:
if up.shape[-1] != self.intermediate_size:
raise RuntimeError("Context-adaptive Euler expects the full active intermediate dimension.")
up_float = up.float()
p = F.softplus(up_float).clamp_min(torch.finfo(torch.float32).tiny)
log_envelope = torch.log(p) / p - (1.0 / math.e)
response = (p - math.e) / (p + math.e)
group_delta = self._max_temperature_delta * torch.tanh(self.caet_temperature_raw.float())
channel_delta = group_delta.repeat_interleave(self.group_size)
temperature = 1.0 + channel_delta * response
envelope = torch.exp(log_envelope * temperature).to(dtype=up.dtype)
return F.silu(up) * envelope
class TileRoutedAdaptiveEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped SwiGLU with a learned SiLU/Euler choice per group.
The 4/4 path stores channels in four contiguous groups. Each group receives
an independent differentiable selector between ordinary SiLU and the
normalized Euler envelope. This adds four scalars per layer, not another
MLP projection, so it can express activation specialization without
changing the model's MLP width or matrix-multiply count.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("TileRoutedAdaptiveEulerGLUMLP requires all groups to be active.")
initial_euler_probability = 0.75
initial_logit = math.log(initial_euler_probability / (1.0 - initial_euler_probability))
self.euler_mix_logits = nn.Parameter(torch.full((self.num_groups,), initial_logit))
def _gate(self, up: torch.Tensor) -> torch.Tensor:
if up.shape[-1] != self.intermediate_size:
raise RuntimeError("Adaptive Euler gating expects the full active intermediate dimension.")
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
euler_weight = torch.sigmoid(self.euler_mix_logits).repeat_interleave(self.group_size)
return F.silu(up) * (1.0 + euler_weight.to(dtype=up.dtype) * (envelope - 1.0))
class TileRoutedDiverseEulerGLUMLP(TileRoutedAdaptiveEulerGLUMLP):
"""Adaptive Euler gating with deliberately diverse group initializations."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
probabilities = torch.linspace(0.125, 0.875, self.num_groups)
with torch.no_grad():
self.euler_mix_logits.copy_(torch.logit(probabilities))
class TileRoutedGroupTokenAttentionEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler-gated 4/4 MLP with token-conditioned attention between groups."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Group-token attention requires all groups to be active.")
self.group_query_scale = nn.Parameter(torch.ones(self.num_groups))
self.group_key_scale = nn.Parameter(torch.ones(self.num_groups))
self.group_value_scale = nn.Parameter(torch.ones(self.num_groups))
self.group_attention_bias = nn.Parameter(torch.zeros(self.num_groups, self.num_groups))
self.group_attention_gain = nn.Parameter(torch.full((self.num_groups,), 0.1))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(-1, self.num_groups, self.group_size)
summaries = hidden.float().mean(dim=-1)
query = summaries * self.group_query_scale.float()
key = summaries * self.group_key_scale.float()
scores = query.unsqueeze(-1) * key.unsqueeze(-2) + self.group_attention_bias.float()
attention = F.softmax(scores, dim=-1)
context = (attention @ (summaries * self.group_value_scale.float()).unsqueeze(-1)).squeeze(-1)
gains = 1.0 + torch.tanh(context * self.group_attention_gain.float())
hidden = hidden * gains.to(dtype=hidden.dtype).unsqueeze(-1)
return (hidden.reshape(-1, self.intermediate_size) @ down_weight).reshape(original_shape)
class TileRoutedHierarchicalEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler-gated 4/4 MLP with a learned 2x2 macro-group stage."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.num_groups != 4 or self.active_groups != self.num_groups:
raise ValueError("Hierarchical Euler MLP is specialized for full-active grouped 4/4.")
self.macro_mix_delta = nn.Parameter(torch.zeros(2, 2))
self.macro_gain = nn.Parameter(torch.full((2, 2), 0.1))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(-1, 2, 2, self.group_size)
macro = hidden.float().mean(dim=2)
mix = torch.eye(2, device=macro.device, dtype=macro.dtype) + self.macro_mix_delta.float()
macro = torch.einsum('tpg,pq->tqg', macro, mix)
macro_summary = macro.mean(dim=-1)
macro_gains = 1.0 + torch.tanh(
macro_summary.unsqueeze(-1) * self.macro_gain.float().unsqueeze(0)
)
hidden = hidden * macro_gains.to(dtype=hidden.dtype).unsqueeze(-1)
return (hidden.reshape(-1, self.intermediate_size) @ down_weight).reshape(original_shape)
class TileRoutedTokenExpertEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Token- and group-routed mixture of SiLU, Euler, and GELU gates."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Token expert routing requires all groups to be active.")
self.expert_router_slope = nn.Parameter(torch.zeros(self.num_groups, 3))
initial_bias = torch.tensor([-2.2, 2.9, -2.2])
self.expert_router_bias = nn.Parameter(initial_bias.repeat(self.num_groups, 1))
def _gate(self, up: torch.Tensor) -> torch.Tensor:
if up.shape[-1] != self.intermediate_size:
raise RuntimeError("Token expert routing expects the full active intermediate dimension.")
grouped_up = up.reshape(-1, self.num_groups, self.group_size)
summaries = grouped_up.float().mean(dim=-1)
logits = summaries.unsqueeze(-1) * self.expert_router_slope.float() + self.expert_router_bias.float()
weights = F.softmax(logits, dim=-1)
p = F.softplus(grouped_up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
silu_gate = F.silu(grouped_up)
gates = torch.stack((silu_gate, silu_gate * envelope, F.gelu(grouped_up)), dim=-1)
return (gates * weights.to(dtype=up.dtype).unsqueeze(-2)).sum(dim=-1).reshape_as(up)
class TileRoutedDynamicGroupMixerEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler-gated 4/4 MLP with a token-conditioned 4x4 group mixer."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.num_groups != 4 or self.active_groups != self.num_groups:
raise ValueError("Dynamic group mixing is specialized for full-active grouped 4/4.")
self.mixer_slope = nn.Parameter(torch.zeros(4, 4))
self.mixer_bias = nn.Parameter(torch.full((4, 4), -3.0))
with torch.no_grad():
self.mixer_bias.diagonal().fill_(3.0)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(-1, 4, self.group_size)
summaries = hidden.float().mean(dim=-1)
logits = self.mixer_bias.float().unsqueeze(0) + summaries.unsqueeze(-1) * self.mixer_slope.float()
mixer = F.softmax(logits, dim=-1)
hidden = torch.einsum('tij,tjg->tig', mixer.to(dtype=hidden.dtype), hidden)
return (hidden.reshape(-1, self.intermediate_size) @ down_weight).reshape(original_shape)
class TileRoutedGroupCompetitionEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler-gated 4/4 MLP with token-wise conserved group energy."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Group competition requires all groups to be active.")
self.competition_temperature_raw = nn.Parameter(torch.tensor(0.0))
self.competition_strength_raw = nn.Parameter(torch.tensor(-2.0))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(-1, self.num_groups, self.group_size)
energy = hidden.float().square().mean(dim=-1).add(1e-6).log()
temperature = F.softplus(self.competition_temperature_raw.float())
allocation = F.softmax(energy * temperature, dim=-1) * self.num_groups
strength = torch.sigmoid(self.competition_strength_raw.float())
gains = 1.0 + strength * (allocation - 1.0)
hidden = hidden * gains.to(dtype=hidden.dtype).unsqueeze(-1)
return (hidden.reshape(-1, self.intermediate_size) @ down_weight).reshape(original_shape)
class TileRoutedEulerGroupStateMemoryGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler-gated full-active grouped MLP with token-conditioned group state."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.group_state_memory_gain_max = float(config.group_state_memory_gain_max)
self.group_state_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, 1))
def _apply_group_state_memory(self, hidden: torch.Tensor) -> torch.Tensor:
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
group_mean = hidden_groups.float().mean(dim=-1, keepdim=True)
group_rms = hidden_groups.float().square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
gain = self.group_state_memory_gain_max * torch.tanh(self.group_state_memory_gain_raw).view(1, self.num_groups, 1)
return (hidden_groups * (1.0 + gain.to(dtype=hidden_groups.dtype) * state)).reshape_as(hidden)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = F.silu(up) * envelope * (x_flat @ value_weight)
hidden = self._apply_group_state_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedEulerFusedGroupMixGLUMLP(TileRoutedDSwiGLUMLP):
"""Fixed Euler 4/4 MLP with group mixing folded into down weights.
Mixing hidden groups before the down projection is algebraically identical
to mixing the four down-weight tiles. Folding it avoids a token-side
einsum and retains the dense full-active GEMM path.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
logits = torch.full((self.num_groups, self.num_groups), -6.0)
logits.fill_diagonal_(6.0)
self.group_mix_logits = nn.Parameter(logits)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = F.silu(up) * envelope * value
mix = F.softmax(self.group_mix_logits.float(), dim=-1).to(dtype=hidden.dtype)
effective_down = torch.einsum('og,osh->gsh', mix, self.down_weight)
return (hidden @ effective_down.reshape(self.intermediate_size, hidden_size)).reshape(original_shape)
class TileRoutedAdaptiveOperatorFieldEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler 4/4 MLP with a token-conditioned group operator field.
A small controller reads the full residual token and selects a convex
combination of 4x4 operator bases. Because the operator changes per token,
it cannot be folded into the down projection like a static group mixer.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.num_groups != 4 or self.active_groups != self.num_groups:
raise ValueError("Adaptive operator field is specialized for full-active grouped 4/4.")
controller_dim = 16
num_bases = 4
self.operator_controller = nn.Linear(config.n_embd, controller_dim, bias=False)
self.operator_router = nn.Linear(controller_dim, num_bases, bias=False)
self.operator_bases = nn.Parameter(torch.empty(num_bases, self.num_groups, self.num_groups))
self.operator_strength_raw = nn.Parameter(torch.tensor(-2.2))
nn.init.normal_(self.operator_controller.weight, mean=0.0, std=0.02)
nn.init.normal_(self.operator_router.weight, mean=0.0, std=0.02)
nn.init.normal_(self.operator_bases, mean=0.0, std=0.02)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
hidden = (F.silu(up) * envelope * value).reshape(-1, self.num_groups, self.group_size)
controller = F.rms_norm(x_flat.float(), (x_flat.shape[-1],)).to(dtype=x_flat.dtype)
router = F.softmax(self.operator_router(F.silu(self.operator_controller(controller))).float(), dim=-1)
delta = torch.einsum('tk,kij->tij', router, self.operator_bases.float())
strength = torch.sigmoid(self.operator_strength_raw.float())
identity = torch.eye(self.num_groups, device=hidden.device, dtype=torch.float32)
operator = identity.unsqueeze(0) + strength * torch.tanh(delta)
hidden = torch.bmm(operator.to(dtype=hidden.dtype), hidden)
return (hidden.reshape(-1, self.intermediate_size) @ down_weight).reshape(original_shape)
class TileRoutedAdaptiveBasisEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler 4/4 MLP with token-dependent low-rank feature construction.
A controller selects one of two low-rank bases that update the up and value
features before the Euler gate. This changes the token's feature basis and
is not equivalent to a post-activation group mixer.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Adaptive basis Euler MLP requires all groups to be active.")
rank = 4
num_bases = 2
self.basis_input = nn.Parameter(torch.empty(config.n_embd, rank))
self.basis_up = nn.Parameter(torch.empty(num_bases, rank, self.intermediate_size))
self.basis_value = nn.Parameter(torch.empty(num_bases, rank, self.intermediate_size))
self.basis_router = nn.Linear(config.n_embd, num_bases, bias=False)
self.basis_strength_raw = nn.Parameter(torch.tensor(-1.4))
nn.init.normal_(self.basis_input, mean=0.0, std=0.02)
nn.init.normal_(self.basis_up, mean=0.0, std=0.02)
nn.init.normal_(self.basis_value, mean=0.0, std=0.02)
nn.init.zeros_(self.basis_router.weight)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
normalized_x = F.rms_norm(x_flat.float(), (x_flat.shape[-1],)).to(dtype=x_flat.dtype)
router = F.softmax(self.basis_router(normalized_x).float(), dim=-1)
latent = x_flat @ self.basis_input.to(dtype=x_flat.dtype)
basis_up = torch.einsum('tk,kri->tri', router, self.basis_up.to(dtype=x_flat.dtype))
basis_value = torch.einsum('tk,kri->tri', router, self.basis_value.to(dtype=x_flat.dtype))
strength = torch.sigmoid(self.basis_strength_raw.float()).to(dtype=x_flat.dtype)
up = x_flat @ up_weight + strength * torch.bmm(latent.unsqueeze(1), basis_up).squeeze(1)
value = x_flat @ value_weight + strength * torch.bmm(latent.unsqueeze(1), basis_value).squeeze(1)
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
return ((F.silu(up) * envelope * value) @ down_weight).reshape(original_shape)
class TileRoutedFusedAdaptiveBasisEulerGLUMLP(TileRoutedAdaptiveBasisEulerGLUMLP):
"""Adaptive basis Euler MLP without materializing token-specific bases."""
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
normalized_x = F.rms_norm(x_flat.float(), (x_flat.shape[-1],)).to(dtype=x_flat.dtype)
router = F.softmax(self.basis_router(normalized_x).float(), dim=-1).to(dtype=x_flat.dtype)
latent = x_flat @ self.basis_input.to(dtype=x_flat.dtype)
delta_up = torch.zeros(x_flat.shape[0], self.intermediate_size, device=x_flat.device, dtype=x_flat.dtype)
delta_value = torch.zeros_like(delta_up)
for basis_idx in range(self.basis_up.shape[0]):
weighted_latent = latent * router[:, basis_idx : basis_idx + 1]
delta_up = delta_up + weighted_latent @ self.basis_up[basis_idx].to(dtype=x_flat.dtype)
delta_value = delta_value + weighted_latent @ self.basis_value[basis_idx].to(dtype=x_flat.dtype)
strength = torch.sigmoid(self.basis_strength_raw.float()).to(dtype=x_flat.dtype)
up = x_flat @ up_weight + strength * delta_up
value = x_flat @ value_weight + strength * delta_value
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
return ((F.silu(up) * envelope * value) @ down_weight).reshape(original_shape)
class TileRoutedAdaptiveDirectionEulerGLUMLP(TileRoutedDSwiGLUMLP):
"""Euler 4/4 MLP with a token-selected rank-one feature direction.
Two learned directions are mixed by a controller, then multiplied by a
token-specific scalar projection. This is a conditional low-rank update to
the up/value features without materializing token-specific matrices.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("Adaptive direction Euler MLP requires all groups to be active.")
num_directions = 2
self.direction_input = nn.Parameter(torch.empty(config.n_embd))
self.direction_up = nn.Parameter(torch.empty(num_directions, self.intermediate_size))
self.direction_value = nn.Parameter(torch.empty(num_directions, self.intermediate_size))
self.direction_router = nn.Linear(config.n_embd, num_directions, bias=False)
self.direction_strength_raw = nn.Parameter(torch.tensor(-2.0))
nn.init.normal_(self.direction_input, mean=0.0, std=0.02)
nn.init.normal_(self.direction_up, mean=0.0, std=0.02)
nn.init.normal_(self.direction_value, mean=0.0, std=0.02)
nn.init.zeros_(self.direction_router.weight)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
normalized_x = F.rms_norm(x_flat.float(), (x_flat.shape[-1],)).to(dtype=x_flat.dtype)
router = F.softmax(self.direction_router(normalized_x).float(), dim=-1).to(dtype=x_flat.dtype)
token_scalar = (x_flat @ self.direction_input.to(dtype=x_flat.dtype)).unsqueeze(-1)
up_direction = router @ self.direction_up.to(dtype=x_flat.dtype)
value_direction = router @ self.direction_value.to(dtype=x_flat.dtype)
strength = torch.sigmoid(self.direction_strength_raw.float()).to(dtype=x_flat.dtype)
up = x_flat @ up_weight + strength * token_scalar * up_direction
value = x_flat @ value_weight + strength * token_scalar * value_direction
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
return ((F.silu(up) * envelope * value) @ down_weight).reshape(original_shape)
class TileRoutedAdaptiveDirectionEulerFusedGLUMLP(TileRoutedAdaptiveDirectionEulerGLUMLP):
"""Checkpoint-compatible two-route fused form of Adaptive Direction."""
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
normalized_x = F.rms_norm(x_flat.float(), (x_flat.shape[-1],)).to(dtype=x_flat.dtype)
router_weight = self.direction_router.weight.to(dtype=x_flat.dtype)
route_first = torch.sigmoid(normalized_x @ (router_weight[0] - router_weight[1]))
token_scalar = (x_flat @ self.direction_input.to(dtype=x_flat.dtype)).unsqueeze(-1)
up_base = self.direction_up[1].to(dtype=x_flat.dtype)
value_base = self.direction_value[1].to(dtype=x_flat.dtype)
up_direction = up_base + route_first.unsqueeze(-1) * (self.direction_up[0].to(dtype=x_flat.dtype) - up_base)
value_direction = value_base + route_first.unsqueeze(-1) * (self.direction_value[0].to(dtype=x_flat.dtype) - value_base)
strength = torch.sigmoid(self.direction_strength_raw.float()).to(dtype=x_flat.dtype)
up = x_flat @ up_weight + strength * token_scalar * up_direction
value = x_flat @ value_weight + strength * token_scalar * value_direction
p = F.softplus(up.float()).clamp_min(torch.finfo(torch.float32).tiny)
envelope = torch.exp(torch.log(p) / p - (1.0 / math.e)).to(dtype=up.dtype)
return ((F.silu(up) * envelope * value) @ down_weight).reshape(original_shape)
class TileRoutedGateLawDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Vectorized grouped4 GLU with a compile-friendly gate law.
The full-active grouped 4/4 path stays as the same three dense GEMMs as
``TileRoutedDSwiGLUMLP``. Subclasses only add cheap elementwise transforms
between the up/value GEMMs and the down GEMM.
"""
use_powrat_gate = False
use_sin_gate = False
use_spon_shift = False
default_spon_value = False
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.gate_alpha_max = float(config.grouped_gate_alpha_max)
self.gate_sin_eps = float(config.grouped_gate_sin_eps)
self.gate_sin_freq = float(config.grouped_gate_sin_freq)
self.gate_spon_max = float(config.grouped_gate_spon_max)
self.gate_layer_schedule = str(config.grouped_gate_layer_schedule)
self.spon_value_enabled = bool(config.grouped_gate_spon_value or self.default_spon_value)
self._gate_schedule_initialized = False
if self.use_powrat_gate:
self.gate_alpha_raw = nn.Parameter(torch.empty(self.num_groups, 1, 1))
self.gate_beta = nn.Parameter(torch.full((self.num_groups, 1, 1), float(config.grouped_gate_beta_init)))
self._fill_alpha(float(config.grouped_gate_alpha_init))
if self.use_spon_shift:
self.spon_up_shift = nn.Parameter(torch.zeros(self.intermediate_size))
self.spon_value_shift = nn.Parameter(torch.zeros(self.intermediate_size))
self.last_gate_alpha_mean: torch.Tensor | None = None
self.last_gate_beta_mean: torch.Tensor | None = None
self.last_spon_up_absmean: torch.Tensor | None = None
self.last_spon_value_absmean: torch.Tensor | None = None
self.last_gate_outlier_rate: torch.Tensor | None = None
self.last_hidden_outlier_rate: torch.Tensor | None = None
def _fill_alpha(self, alpha: float) -> None:
alpha = max(-0.999 * self.gate_alpha_max, min(0.999 * self.gate_alpha_max, alpha))
raw = math.atanh(alpha / self.gate_alpha_max)
with torch.no_grad():
self.gate_alpha_raw.fill_(raw)
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
if self.use_powrat_gate and not self._gate_schedule_initialized:
self._fill_alpha(self._scheduled_alpha(block_idx, n_layer))
self._gate_schedule_initialized = True
def _scheduled_alpha(self, block_idx: int, n_layer: int) -> float:
if self.gate_layer_schedule == "constant":
return float(self.config.grouped_gate_alpha_init)
if self.gate_layer_schedule == "late_positive":
if n_layer <= 1:
return self.gate_alpha_max
pos = block_idx / max(1, n_layer - 1)
return 0.0 if pos < 0.67 else self.gate_alpha_max * (pos - 0.67) / 0.33
observed = [-0.074, -0.090, -0.089, -0.090, -0.084, -0.084, -0.060, -0.043, -0.020, 0.107, 0.235, 0.351]
if n_layer <= 1:
value = observed[-1]
else:
pos = block_idx * (len(observed) - 1) / max(1, n_layer - 1)
lo = int(math.floor(pos))
hi = min(len(observed) - 1, lo + 1)
frac = pos - lo
value = observed[lo] * (1.0 - frac) + observed[hi] * frac
scale = self.gate_alpha_max / max(abs(v) for v in observed)
return max(-self.gate_alpha_max, min(self.gate_alpha_max, value * scale))
def _dense_weights(self, hidden_size: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return up_weight, value_weight, down_weight
def _spon_biases(self, dtype: torch.dtype) -> tuple[torch.Tensor | None, torch.Tensor | None]:
if not self.use_spon_shift or self.gate_spon_max == 0.0:
return None, None
up_bias = self.spon_up_shift.to(dtype=dtype)
value_bias = self.spon_value_shift.to(dtype=dtype) if self.spon_value_enabled else None
return up_bias, value_bias
def _gate_law(self, up: torch.Tensor) -> torch.Tensor:
if not self.use_powrat_gate:
return F.silu(up)
up_groups = up.reshape(up.shape[0], self.num_groups, self.group_size)
alpha = self.gate_alpha_max * torch.tanh(
self.gate_alpha_raw.to(device=up.device, dtype=up.dtype)
).view(1, self.num_groups, 1)
beta = self.gate_beta.to(device=up.device, dtype=up.dtype).abs().clamp_min(1e-4).view(1, self.num_groups, 1)
modifier = 1.0 + alpha * up_groups * torch.rsqrt(1.0 + beta * up_groups.square())
gate = F.silu(up_groups) * modifier
if self.use_sin_gate and self.gate_sin_eps > 0.0:
gate = gate * (1.0 + self.gate_sin_eps * torch.sin(self.gate_sin_freq * up_groups))
return gate.reshape_as(up)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight, value_weight, down_weight = self._dense_weights(hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
up_bias, value_bias = self._spon_biases(x_flat.dtype)
if up_bias is not None:
up.add_(up_bias)
if value_bias is not None:
value.add_(value_bias)
gate = self._gate_law(up)
hidden = gate * value
output = hidden @ down_weight
if not torch.is_grad_enabled():
if self.use_powrat_gate:
alpha = self.gate_alpha_max * torch.tanh(self.gate_alpha_raw.detach().float())
self.last_gate_alpha_mean = alpha.mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_gate_beta_mean = self.gate_beta.detach().float().abs().mean().to(device=x_flat.device, dtype=x_flat.dtype)
if self.use_spon_shift:
self.last_spon_up_absmean = self.spon_up_shift.detach().float().abs().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_spon_value_absmean = self.spon_value_shift.detach().float().abs().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_gate_outlier_rate = (gate.detach().float().abs() > 8.0).float().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_hidden_outlier_rate = (hidden.detach().float().abs() > 8.0).float().mean().to(device=x_flat.device, dtype=x_flat.dtype)
return output.reshape(original_shape)
class TileRoutedPowRatGLUMLP(TileRoutedGateLawDSwiGLUMLP):
use_powrat_gate = True
class TileRoutedSinPowRatGLUMLP(TileRoutedGateLawDSwiGLUMLP):
use_powrat_gate = True
use_sin_gate = True
class TileRoutedSPONGLUMLP(TileRoutedGateLawDSwiGLUMLP):
use_spon_shift = True
class TileRoutedTrainOnlySPONGLUMLP(TileRoutedSPONGLUMLP):
"""Use SPON as a training-time activation perturbation with free inference."""
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
if not torch.is_grad_enabled():
return TileRoutedDSwiGLUMLP._forward_full_active(self, x, original_shape, x_flat)
return super()._forward_full_active(x, original_shape, x_flat)
class TileRoutedSPONPowRatGLUMLP(TileRoutedGateLawDSwiGLUMLP):
use_powrat_gate = True
use_spon_shift = True
class TileRoutedAttentionMemoryGateDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Grouped4 GLU whose gate is steered by a small learned memory attention.
The main compute path keeps grouped4's full-active dense GEMMs:
up/value/down are still single matmuls. The fixed SwiGLU nonlinearity is
replaced by ``up * sigmoid(up + memory_delta)`` where ``memory_delta`` is a
per-group token-conditioned lookup over learned hidden memory slots.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.attn_gate_rank = int(config.attn_gate_rank)
self.attn_gate_slots = int(config.attn_gate_slots)
self.attn_gate_alpha_max = float(config.attn_gate_alpha_max)
self.attn_gate_temperature = float(config.attn_gate_temperature)
self.attn_q_weight = nn.Parameter(torch.empty(self.num_groups, self.group_size, self.attn_gate_rank))
self.attn_k = nn.Parameter(torch.empty(self.num_groups, self.attn_gate_slots, self.attn_gate_rank))
self.attn_v = nn.Parameter(torch.empty(self.num_groups, self.attn_gate_slots, self.group_size))
self.attn_alpha_raw = nn.Parameter(torch.empty(self.num_groups, 1, 1))
self._reset_attention_memory(float(config.attn_gate_alpha_init))
self.last_attn_gate_alpha_mean: torch.Tensor | None = None
self.last_attn_gate_entropy: torch.Tensor | None = None
self.last_attn_gate_delta_rms: torch.Tensor | None = None
self.last_attn_gate_up_rms: torch.Tensor | None = None
def _reset_attention_memory(self, alpha_init: float) -> None:
nn.init.normal_(self.attn_q_weight, mean=0.0, std=self.group_size**-0.5)
nn.init.normal_(self.attn_k, mean=0.0, std=self.attn_gate_rank**-0.5)
nn.init.normal_(self.attn_v, mean=0.0, std=0.01)
alpha = max(-0.999 * self.attn_gate_alpha_max, min(0.999 * self.attn_gate_alpha_max, alpha_init))
raw = math.atanh(alpha / self.attn_gate_alpha_max) if self.attn_gate_alpha_max > 0.0 else 0.0
with torch.no_grad():
self.attn_alpha_raw.fill_(raw)
def _dense_weights(self, hidden_size: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return up_weight, value_weight, down_weight
def _memory_delta(self, up_groups: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
dtype = up_groups.dtype
up_norm = up_groups * torch.rsqrt(up_groups.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + 1e-6)
q = torch.einsum("ngc,gcr->ngr", up_norm, self.attn_q_weight.to(device=up_groups.device, dtype=dtype))
k = self.attn_k.to(device=up_groups.device, dtype=dtype)
logits = torch.einsum("ngr,gsr->ngs", q, k)
logits = logits * (self.attn_gate_rank**-0.5) / max(self.attn_gate_temperature, 1e-4)
attn = torch.softmax(logits.float(), dim=-1).to(dtype=dtype)
delta = torch.einsum("ngs,gsc->ngc", attn, self.attn_v.to(device=up_groups.device, dtype=dtype))
return delta, attn
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight, value_weight, down_weight = self._dense_weights(hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
up_groups = up.reshape(up.shape[0], self.num_groups, self.group_size)
delta, attn = self._memory_delta(up_groups)
alpha = self.attn_gate_alpha_max * torch.tanh(
self.attn_alpha_raw.to(device=up.device, dtype=up.dtype)
).view(1, self.num_groups, 1)
gate_groups = up_groups * torch.sigmoid(up_groups + alpha * delta)
hidden = gate_groups.reshape_as(up) * value
output = hidden @ down_weight
if not torch.is_grad_enabled():
entropy = -(attn.float().clamp_min(1e-8) * attn.float().clamp_min(1e-8).log()).sum(dim=-1).mean()
self.last_attn_gate_alpha_mean = alpha.detach().float().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_attn_gate_entropy = entropy.to(device=x_flat.device, dtype=x_flat.dtype)
self.last_attn_gate_delta_rms = delta.detach().float().square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_attn_gate_up_rms = up_groups.detach().float().square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
return output.reshape(original_shape)
class TileRoutedInnerAttentionFFNMLP(TileRoutedDSwiGLUMLP):
"""Activation-free grouped FFN with an attention block inside the FFN.
This removes pointwise gate activations from the full-active path. The
hidden groups become four tiny tokens: ``up`` builds q/k, ``value`` builds
values, group attention mixes values, and the result is down-projected.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.inner_attn_rank = int(config.inner_attn_rank)
self.inner_attn_alpha_max = float(config.inner_attn_alpha_max)
self.inner_attn_temperature = float(config.inner_attn_temperature)
self.inner_q_weight = nn.Parameter(torch.empty(self.num_groups, self.group_size, self.inner_attn_rank))
self.inner_k_weight = nn.Parameter(torch.empty(self.num_groups, self.group_size, self.inner_attn_rank))
self.inner_attn_bias = nn.Parameter(torch.empty(self.num_groups, self.num_groups))
self.inner_alpha_raw = nn.Parameter(torch.empty(self.num_groups, 1, 1))
self._reset_inner_attention(float(config.inner_attn_alpha_init))
self.last_inner_attn_alpha_mean: torch.Tensor | None = None
self.last_inner_attn_entropy: torch.Tensor | None = None
self.last_inner_attn_offdiag: torch.Tensor | None = None
self.last_inner_attn_update_rms: torch.Tensor | None = None
self.last_inner_attn_value_rms: torch.Tensor | None = None
def _reset_inner_attention(self, alpha_init: float) -> None:
nn.init.normal_(self.inner_q_weight, mean=0.0, std=self.group_size**-0.5)
nn.init.normal_(self.inner_k_weight, mean=0.0, std=self.group_size**-0.5)
with torch.no_grad():
self.inner_attn_bias.fill_(-2.0)
self.inner_attn_bias.diagonal().fill_(2.0)
alpha = max(-0.999 * self.inner_attn_alpha_max, min(0.999 * self.inner_attn_alpha_max, alpha_init))
raw = math.atanh(alpha / self.inner_attn_alpha_max) if self.inner_attn_alpha_max > 0.0 else 0.0
with torch.no_grad():
self.inner_alpha_raw.fill_(raw)
def _dense_weights(self, hidden_size: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return up_weight, value_weight, down_weight
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight, value_weight, down_weight = self._dense_weights(hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
up_groups = up.reshape(up.shape[0], self.num_groups, self.group_size)
value_groups = value.reshape(value.shape[0], self.num_groups, self.group_size)
dtype = up_groups.dtype
up_norm = up_groups * torch.rsqrt(up_groups.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + 1e-6)
q = torch.einsum("ngc,gcr->ngr", up_norm, self.inner_q_weight.to(device=up.device, dtype=dtype))
k = torch.einsum("ngc,gcr->ngr", up_norm, self.inner_k_weight.to(device=up.device, dtype=dtype))
logits = torch.einsum("ngr,nhr->ngh", q, k)
logits = logits * (self.inner_attn_rank**-0.5) / max(self.inner_attn_temperature, 1e-4)
logits = logits + self.inner_attn_bias.to(device=up.device, dtype=dtype).view(1, self.num_groups, self.num_groups)
attn = torch.softmax(logits.float(), dim=-1).to(dtype=dtype)
mixed = torch.einsum("ngh,nhc->ngc", attn, value_groups)
alpha = self.inner_attn_alpha_max * torch.tanh(
self.inner_alpha_raw.to(device=up.device, dtype=dtype)
).view(1, self.num_groups, 1)
hidden_groups = value_groups + alpha * (mixed - value_groups)
output = hidden_groups.reshape_as(value) @ down_weight
if not torch.is_grad_enabled():
probs = attn.detach().float().clamp_min(1e-8)
entropy = -(probs * probs.log()).sum(dim=-1).mean()
diag = torch.diagonal(attn.detach().float(), dim1=-2, dim2=-1).mean()
update = (mixed - value_groups).detach().float()
value_float = value_groups.detach().float()
self.last_inner_attn_alpha_mean = alpha.detach().float().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_inner_attn_entropy = entropy.to(device=x_flat.device, dtype=x_flat.dtype)
self.last_inner_attn_offdiag = (1.0 - diag).to(device=x_flat.device, dtype=x_flat.dtype)
self.last_inner_attn_update_rms = update.square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_inner_attn_value_rms = value_float.square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
return output.reshape(original_shape)
class TileRoutedBilinearMemoryStepFFNMLP(TileRoutedDSwiGLUMLP):
"""Grouped FFN with a small bounded inner update loop instead of a gate.
This is not a pointwise activation replacement. The FFN makes group hidden
states, repeatedly applies a bilinear product correction, RMS-bounds that
correction, shares a small mean-bus between groups, and only then writes
through the down projection.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.bilinear_memory_depth = int(config.bilinear_memory_depth)
self.bilinear_memory_alpha_max = float(config.bilinear_memory_alpha_max)
self.bilinear_memory_target_ratio = float(config.bilinear_memory_target_ratio)
self.bilinear_memory_bus_mix = float(config.bilinear_memory_bus_mix)
self.bilinear_memory_state_update = float(config.bilinear_memory_state_update)
self.memory_scale = nn.Parameter(torch.empty(1, self.num_groups, self.group_size))
self.inner_alpha_raw = nn.Parameter(torch.empty(self.num_groups, 1, 1))
self._reset_bilinear_memory(float(config.bilinear_memory_alpha_init))
self.last_bilinear_alpha_mean: torch.Tensor | None = None
self.last_bilinear_update_ratio: torch.Tensor | None = None
self.last_bilinear_hidden_rms: torch.Tensor | None = None
self.last_bilinear_state_rms: torch.Tensor | None = None
def _reset_bilinear_memory(self, alpha_init: float) -> None:
nn.init.normal_(self.memory_scale, mean=1.0, std=0.02)
alpha = max(-0.999 * self.bilinear_memory_alpha_max, min(0.999 * self.bilinear_memory_alpha_max, alpha_init))
raw = math.atanh(alpha / self.bilinear_memory_alpha_max) if self.bilinear_memory_alpha_max > 0.0 else 0.0
with torch.no_grad():
self.inner_alpha_raw.fill_(raw)
def _dense_weights(self, hidden_size: int) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return up_weight, value_weight, down_weight
@staticmethod
def _rms_norm(x: torch.Tensor) -> torch.Tensor:
return x * torch.rsqrt(x.float().square().mean(dim=-1, keepdim=True).to(dtype=x.dtype) + 1e-6)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight, value_weight, down_weight = self._dense_weights(hidden_size)
up = x_flat @ up_weight
value = x_flat @ value_weight
state = self._rms_norm(up.reshape(up.shape[0], self.num_groups, self.group_size))
hidden = value.reshape(value.shape[0], self.num_groups, self.group_size)
alpha = self.bilinear_memory_alpha_max * torch.tanh(
self.inner_alpha_raw.to(device=up.device, dtype=up.dtype)
).view(1, self.num_groups, 1)
memory = self.memory_scale.to(device=up.device, dtype=up.dtype)
last_ratio = hidden.new_tensor(0.0)
for _ in range(self.bilinear_memory_depth):
correction = state * hidden * memory
if self.bilinear_memory_bus_mix > 0.0:
bus = correction.mean(dim=1, keepdim=True)
correction = correction + self.bilinear_memory_bus_mix * (bus - correction)
hidden_rms = hidden.float().square().mean(dim=-1, keepdim=True).sqrt().to(dtype=hidden.dtype).clamp_min(1e-6)
correction_rms = correction.float().square().mean(dim=-1, keepdim=True).sqrt().to(dtype=hidden.dtype).clamp_min(1e-6)
scale = (self.bilinear_memory_target_ratio * hidden_rms / correction_rms).detach()
correction = correction * scale
hidden = hidden + alpha * correction
if self.bilinear_memory_state_update > 0.0:
state = self._rms_norm(state + self.bilinear_memory_state_update * correction)
if not torch.is_grad_enabled():
last_ratio = (correction.detach().float().square().mean().sqrt() / hidden.detach().float().square().mean().sqrt().clamp_min(1e-6)).to(dtype=hidden.dtype)
output = hidden.reshape_as(value) @ down_weight
if not torch.is_grad_enabled():
self.last_bilinear_alpha_mean = alpha.detach().float().mean().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_bilinear_update_ratio = last_ratio.to(device=x_flat.device, dtype=x_flat.dtype)
self.last_bilinear_hidden_rms = hidden.detach().float().square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
self.last_bilinear_state_rms = state.detach().float().square().mean().sqrt().to(device=x_flat.device, dtype=x_flat.dtype)
return output.reshape(original_shape)
class TileRoutedRationalGateDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""SwiGLU plus a tiny learned rational gate correction.
Inspired by KAT/KAN-style learnable activations, but initialized as exactly
SwiGLU so the model starts from the known-good grouped path.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.rational_num = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
self.rational_den = nn.Parameter(torch.ones(self.num_groups, 1, 1))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
output = torch.zeros_like(x_flat)
for group_idx in range(self.num_groups):
up = x_flat @ self.up_weight[group_idx]
value = x_flat @ self.value_weight[group_idx]
correction = self.rational_num[group_idx] * up.square() / (1.0 + self.rational_den[group_idx].abs() * up.abs())
hidden = (F.silu(up) + correction) * value
output = output + hidden @ self.down_weight[group_idx]
return output.reshape(original_shape)
class TileRoutedExpGateDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""SwiGLU with a learnable stabilized exponential gate multiplier."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.exp_gate_alpha = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
output = torch.zeros_like(x_flat)
for group_idx in range(self.num_groups):
up = x_flat @ self.up_weight[group_idx]
value = x_flat @ self.value_weight[group_idx]
multiplier = torch.exp(self.exp_gate_alpha[group_idx].to(dtype=up.dtype) * torch.tanh(up))
hidden = F.silu(up) * multiplier * value
output = output + hidden @ self.down_weight[group_idx]
return output.reshape(original_shape)
class TileRoutedHiddenRMSNormDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Zero-init hidden RMS correction before the down projection."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.hidden_norm_alpha = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
output = torch.zeros_like(x_flat)
for group_idx in range(self.num_groups):
hidden = F.silu(x_flat @ self.up_weight[group_idx]) * (x_flat @ self.value_weight[group_idx])
rms = hidden.float().pow(2).mean(dim=-1, keepdim=True).add(self.config.norm_eps).rsqrt().to(dtype=hidden.dtype)
hidden_normed = hidden * rms
alpha = self.hidden_norm_alpha[group_idx].to(device=hidden.device, dtype=hidden.dtype).clamp(-1.0, 1.0)
hidden = hidden + alpha * (hidden_normed - hidden)
output = output + hidden @ self.down_weight[group_idx]
return output.reshape(original_shape)
class TileRoutedLateTokenHiddenRMSPreserveDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Late, token-gated hidden RMS correction with the grouped fastpath preserved."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.late_layer_start = int(config.grouped_mlp_hidden_rms_late_layer_start)
self.alpha_max = float(config.grouped_mlp_hidden_rms_alpha_max)
alpha_init = max(
-0.999 * self.alpha_max,
min(0.999 * self.alpha_max, float(config.grouped_mlp_hidden_rms_alpha_init)),
)
raw_alpha = math.atanh(alpha_init / self.alpha_max)
self.hidden_norm_alpha_raw = nn.Parameter(torch.full((self.num_groups, 1, 1), raw_alpha))
self.hidden_norm_gate = nn.Linear(config.n_embd, self.num_groups, bias=True)
nn.init.zeros_(self.hidden_norm_gate.weight)
gate_init = max(1e-4, min(1.0 - 1e-4, float(config.grouped_mlp_hidden_rms_gate_init)))
nn.init.constant_(self.hidden_norm_gate.bias, math.log(gate_init / (1.0 - gate_init)))
self.hidden_norm_output_scale = float(config.grouped_mlp_hidden_rms_output_scale)
self.hidden_norm_token_gate = bool(config.grouped_mlp_hidden_rms_token_gate)
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
if self.block_idx < self.late_layer_start or self.hidden_norm_output_scale == 0.0:
return (hidden @ down_weight).reshape(original_shape)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
rms = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True).add(self.config.norm_eps).rsqrt()
hidden_normed = hidden_groups * rms.to(dtype=hidden_groups.dtype)
alpha = self.alpha_max * torch.tanh(
self.hidden_norm_alpha_raw.to(device=hidden.device, dtype=hidden.dtype)
).view(1, self.num_groups, 1)
if self.hidden_norm_token_gate:
gate = torch.sigmoid(self.hidden_norm_gate(x_flat).float()).to(dtype=hidden.dtype).unsqueeze(-1)
else:
gate = 1.0
correction = (hidden_normed - hidden_groups) * (alpha * gate * self.hidden_norm_output_scale)
corrected_hidden = (hidden_groups + correction).reshape(hidden.shape[0], self.intermediate_size)
return (corrected_hidden @ down_weight).reshape(original_shape)
class TileRoutedRMSMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped SwiGLU with normalized hidden group memory exchange.
The base 4/4 grouped SwiGLU is preserved. Hidden group activations are RMS
normalized, softly mixed across groups, then blended back before the down
projection. This moves communication into hidden scratch space instead of
adding an unbounded late residual correction.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
logits = torch.full((self.num_groups, self.num_groups), float(config.rms_memory_mix_offdiag_logit))
logits.fill_diagonal_(float(config.rms_memory_mix_identity_logit))
self.group_memory_mix_logits = nn.Parameter(logits)
self.rms_memory_alpha_max = float(config.rms_memory_alpha_max)
self.hidden_memory_alpha_raw = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
target = float(config.rms_memory_target_init)
target_min = float(config.rms_memory_target_min)
target_max = float(config.rms_memory_target_max)
target = min(max(target, target_min + 1e-6), target_max - 1e-6)
ratio = (target - target_min) / (target_max - target_min)
self.hidden_rms_target_raw = nn.Parameter(torch.full((self.num_groups, 1, 1), math.log(ratio / (1.0 - ratio))))
self.rms_memory_target_min = target_min
self.rms_memory_target_max = target_max
self.rms_memory_token_gate = bool(config.rms_memory_token_gate)
if self.rms_memory_token_gate:
self.memory_gate_weight = nn.Parameter(torch.zeros(self.num_groups, config.n_embd))
gate_init = min(max(float(config.rms_memory_gate_init), 1e-4), 1.0 - 1e-4)
self.memory_gate_bias = nn.Parameter(torch.full((self.num_groups,), math.log(gate_init / (1.0 - gate_init))))
self.rms_weight_reparam = bool(config.rms_weight_reparam)
if self.rms_weight_reparam:
self.up_weight_gain = nn.Parameter(torch.ones(self.num_groups, 1, 1))
self.value_weight_gain = nn.Parameter(torch.ones(self.num_groups, 1, 1))
self.down_weight_gain = nn.Parameter(torch.ones(self.num_groups, 1, 1))
self.block_idx = 0
self.n_layer = config.n_layer
self._scheduled_alpha_initialized = False
self.last_hidden_rms_before: torch.Tensor | None = None
self.last_hidden_rms_after: torch.Tensor | None = None
self.last_memory_alpha_mean: torch.Tensor | None = None
self.last_memory_mix_offdiag: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
if not self._scheduled_alpha_initialized:
alpha = self._scheduled_alpha(block_idx, n_layer)
alpha = max(-0.999 * self.rms_memory_alpha_max, min(0.999 * self.rms_memory_alpha_max, alpha))
with torch.no_grad():
self.hidden_memory_alpha_raw.fill_(math.atanh(alpha / self.rms_memory_alpha_max))
self._scheduled_alpha_initialized = True
def _scheduled_alpha(self, block_idx: int, n_layer: int) -> float:
observed = [-0.074, -0.090, -0.089, -0.090, -0.084, -0.084, -0.060, -0.043, -0.020, 0.107, 0.235, 0.351]
if n_layer <= 1:
return observed[-1]
pos = block_idx * (len(observed) - 1) / max(1, n_layer - 1)
lo = int(math.floor(pos))
hi = min(len(observed) - 1, lo + 1)
frac = pos - lo
return observed[lo] * (1.0 - frac) + observed[hi] * frac
def _target_rms(self, hidden: torch.Tensor) -> torch.Tensor:
raw = self.hidden_rms_target_raw.to(device=hidden.device, dtype=torch.float32)
target = self.rms_memory_target_min + (self.rms_memory_target_max - self.rms_memory_target_min) * torch.sigmoid(raw)
return target.to(dtype=hidden.dtype).view(1, self.num_groups, 1)
def _maybe_reparam_weight(self, weight: torch.Tensor, gain: torch.Tensor | None, target_std: float) -> torch.Tensor:
if not self.rms_weight_reparam or gain is None:
return weight
rms = weight.float().pow(2).mean(dim=(-2, -1), keepdim=True).sqrt().clamp_min(1e-8)
scale = (target_std / rms).to(device=weight.device, dtype=weight.dtype)
return weight * scale * gain.to(device=weight.device, dtype=weight.dtype)
def _weights(self) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
down_std = 1.0 / math.sqrt(self.config.n_embd) / self.config.n_layer
up = self._maybe_reparam_weight(self.up_weight, getattr(self, "up_weight_gain", None), std)
value = self._maybe_reparam_weight(self.value_weight, getattr(self, "value_weight_gain", None), std)
down = self._maybe_reparam_weight(self.down_weight, getattr(self, "down_weight_gain", None), down_std)
return up, value, down
def _token_gate(self, x_flat: torch.Tensor) -> torch.Tensor | float:
if not self.rms_memory_token_gate:
return 1.0
x_rms = x_flat.float().pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-8)
x_normed = (x_flat.float() / x_rms).to(dtype=x_flat.dtype)
logits = F.linear(
x_normed,
self.memory_gate_weight.to(device=x_flat.device, dtype=x_flat.dtype),
self.memory_gate_bias.to(device=x_flat.device, dtype=x_flat.dtype),
)
return torch.sigmoid(logits.float()).to(dtype=x_flat.dtype).unsqueeze(-1)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight, value_weight, down_weight = self._weights()
up_dense = up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_dense = value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
hidden = F.silu(x_flat @ up_dense) * (x_flat @ value_dense)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_rms = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
hidden_normed = hidden_groups * torch.rsqrt(
hidden_groups.float().pow(2).mean(dim=-1, keepdim=True) + self.config.norm_eps
).to(dtype=hidden_groups.dtype)
hidden_normed = hidden_normed * self._target_rms(hidden_groups)
mix = torch.softmax(self.group_memory_mix_logits.float(), dim=-1).to(device=hidden.device, dtype=hidden.dtype)
mixed = torch.einsum("og,ngs->nos", mix, hidden_normed)
alpha = self.rms_memory_alpha_max * torch.tanh(
self.hidden_memory_alpha_raw.to(device=hidden.device, dtype=hidden.dtype)
).view(1, self.num_groups, 1)
gate = self._token_gate(x_flat)
hidden_groups = hidden_groups + alpha * gate * (mixed - hidden_groups)
down_dense = down_weight.reshape(self.intermediate_size, hidden_size)
output = hidden_groups.reshape(hidden.shape[0], self.intermediate_size) @ down_dense
if not torch.is_grad_enabled():
self.last_hidden_rms_before = hidden_rms.detach().mean()
self.last_hidden_rms_after = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True).sqrt().detach().mean()
self.last_memory_alpha_mean = alpha.detach().mean()
eye = torch.eye(self.num_groups, device=mix.device, dtype=torch.bool)
self.last_memory_mix_offdiag = mix.masked_select(~eye).detach().mean()
return output.reshape(original_shape)
class TileRoutedFastRMSMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Speed-first hidden RMS memory for full-active grouped SwiGLU.
The early layers run the exact grouped4 dense-GEMM fastpath. Enabled late
layers keep the same up/value/down GEMMs, but insert a cheap hidden-space
mean-bus exchange before the down projection. The exchange preserves each
group's hidden RMS, so it changes direction more than magnitude.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.fast_rms_memory_schedule = str(config.fast_rms_memory_schedule)
self.fast_rms_memory_alpha_max = float(config.fast_rms_memory_alpha_max)
self.fast_rms_memory_use_rms_norm = bool(config.fast_rms_memory_use_rms_norm)
self.hidden_memory_alpha_raw = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
target = float(config.rms_memory_target_init)
target_min = float(config.rms_memory_target_min)
target_max = float(config.rms_memory_target_max)
target = min(max(target, target_min + 1e-6), target_max - 1e-6)
ratio = (target - target_min) / (target_max - target_min)
self.hidden_rms_target_raw = nn.Parameter(torch.full((self.num_groups, 1, 1), math.log(ratio / (1.0 - ratio))))
self.rms_memory_target_min = target_min
self.rms_memory_target_max = target_max
self.block_idx = 0
self.n_layer = config.n_layer
self.memory_enabled = False
self._scheduled_alpha_initialized = False
self.last_hidden_rms_before: torch.Tensor | None = None
self.last_hidden_rms_after: torch.Tensor | None = None
self.last_memory_alpha_mean: torch.Tensor | None = None
self.last_memory_bus_rms: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
if not self._scheduled_alpha_initialized:
alpha = self._scheduled_alpha(block_idx, n_layer)
self.memory_enabled = abs(alpha) > 0.0
alpha = max(-0.999 * self.fast_rms_memory_alpha_max, min(0.999 * self.fast_rms_memory_alpha_max, alpha))
with torch.no_grad():
self.hidden_memory_alpha_raw.fill_(math.atanh(alpha / self.fast_rms_memory_alpha_max))
self._scheduled_alpha_initialized = True
def _scheduled_alpha(self, block_idx: int, n_layer: int) -> float:
if self.fast_rms_memory_schedule == "late8_meanbus":
if n_layer == 12:
schedule = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.05, 0.10, 0.18, 0.25]
return schedule[block_idx]
start = max(0, n_layer - 4)
late = [0.05, 0.10, 0.18, 0.25]
return late[block_idx - start] if block_idx >= start else 0.0
if self.fast_rms_memory_schedule == "late9_meanbus":
if n_layer == 12:
schedule = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.08, 0.16, 0.25]
return schedule[block_idx]
start = max(0, n_layer - 3)
late = [0.08, 0.16, 0.25]
return late[block_idx - start] if block_idx >= start else 0.0
if self.fast_rms_memory_schedule == "all_tiny_meanbus":
observed = [-0.074, -0.090, -0.089, -0.090, -0.084, -0.084, -0.060, -0.043, -0.020, 0.107, 0.235, 0.351]
if n_layer <= 1:
value = observed[-1]
else:
pos = block_idx * (len(observed) - 1) / max(1, n_layer - 1)
lo = int(math.floor(pos))
hi = min(len(observed) - 1, lo + 1)
frac = pos - lo
value = observed[lo] * (1.0 - frac) + observed[hi] * frac
return max(-0.03, min(0.08, value))
raise ValueError(f"unknown fast_rms_memory_schedule={self.fast_rms_memory_schedule!r}")
def _target_scale(self, hidden: torch.Tensor) -> torch.Tensor:
raw = self.hidden_rms_target_raw.to(device=hidden.device, dtype=torch.float32)
target = self.rms_memory_target_min + (self.rms_memory_target_max - self.rms_memory_target_min) * torch.sigmoid(raw)
return target.to(dtype=hidden.dtype).view(1, self.num_groups, 1)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
if not self.memory_enabled:
return super()._forward_full_active(x, original_shape, x_flat)
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
alpha = self.fast_rms_memory_alpha_max * torch.tanh(
self.hidden_memory_alpha_raw.to(device=hidden.device, dtype=hidden.dtype)
)
if not self.fast_rms_memory_use_rms_norm:
# Linear mean-bus exchange can be folded into the down projection:
# mixed_o = (1-a_o) h_o + a_o * mean_g(h_g).
alpha_weight = alpha.to(device=self.down_weight.device, dtype=self.down_weight.dtype)
shared_down = (alpha_weight * self.down_weight).mean(dim=0, keepdim=True)
effective_down = (1.0 - alpha_weight) * self.down_weight + shared_down
output = hidden @ effective_down.reshape(self.intermediate_size, hidden_size)
if not torch.is_grad_enabled():
self.last_hidden_rms_before = hidden.float().pow(2).mean(dim=-1).sqrt().detach().mean()
self.last_hidden_rms_after = self.last_hidden_rms_before
self.last_memory_alpha_mean = alpha.detach().mean()
self.last_memory_bus_rms = hidden.new_tensor(0.0)
return output.reshape(original_shape)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_var = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True)
hidden_rms = torch.sqrt(hidden_var + self.config.norm_eps)
hidden_normed = hidden_groups * torch.rsqrt(hidden_var + self.config.norm_eps).to(dtype=hidden_groups.dtype)
bus = hidden_normed.mean(dim=1, keepdim=True)
bus_var = bus.float().pow(2).mean(dim=-1, keepdim=True)
bus = bus * torch.rsqrt(bus_var + self.config.norm_eps).to(dtype=bus.dtype)
memory = bus * hidden_rms.to(dtype=hidden_groups.dtype) * self._target_scale(hidden_groups)
alpha = alpha.view(1, self.num_groups, 1)
hidden_groups = hidden_groups + alpha * (memory - hidden_groups)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
output = hidden_groups.reshape(hidden.shape[0], self.intermediate_size) @ down_weight
if not torch.is_grad_enabled():
self.last_hidden_rms_before = hidden_rms.detach().mean()
self.last_hidden_rms_after = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True).sqrt().detach().mean()
self.last_memory_alpha_mean = alpha.detach().mean()
self.last_memory_bus_rms = bus.float().pow(2).mean(dim=-1, keepdim=True).sqrt().detach().mean()
return output.reshape(original_shape)
class TileRoutedCoAdaptHiddenControllerDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Late-only train-time hidden controller for grouped4.
This keeps grouped4's full-active up/value/down GEMM path. In late layers,
a zero-init low-rank controller can add a tiny RMS-bounded hidden-space
correction before the down projection. The design is the train-time,
co-adapted version of the oracle-positive hidden write probe; at
initialization it is exactly the grouped4 fastpath.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.controller_rank = int(config.hidden_controller_rank)
self.controller_rho_frac = float(config.hidden_controller_rho_frac)
self.controller_late_layer_start = int(config.hidden_controller_late_layer_start)
self.controller_down = nn.Parameter(torch.empty(self.num_groups, self.group_size, self.controller_rank))
self.controller_up = nn.Parameter(torch.zeros(self.num_groups, self.controller_rank, self.group_size))
nn.init.normal_(self.controller_down, mean=0.0, std=self.group_size**-0.5)
self.last_hidden_controller_ratio: torch.Tensor | None = None
self.last_hidden_controller_raw_rms: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
enabled = self._controller_enabled()
self.controller_down.requires_grad_(enabled)
self.controller_up.requires_grad_(enabled)
def _controller_enabled(self) -> bool:
return self.controller_rho_frac > 0.0 and self.block_idx >= self.controller_late_layer_start
def _hidden_controller(self, hidden_groups: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
dtype = hidden_groups.dtype
hidden_rms = hidden_groups.float().square().mean(dim=-1, keepdim=True).sqrt().to(dtype=dtype)
h_norm = hidden_groups * torch.rsqrt(
hidden_groups.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + self.config.norm_eps
)
low = torch.einsum("ngs,gsr->ngr", h_norm, self.controller_down.to(dtype=dtype))
raw = torch.einsum("ngr,grs->ngs", torch.tanh(low), self.controller_up.to(dtype=dtype))
raw_rms = torch.sqrt(raw.float().square().mean(dim=-1, keepdim=True) + self.config.norm_eps)
raw_rms = raw_rms.to(dtype=dtype)
target = self.controller_rho_frac * hidden_rms
correction = raw * (target / raw_rms).clamp(max=1.0)
ratio = correction.float().square().mean().sqrt() / hidden_groups.float().square().mean().sqrt().clamp_min(1e-8)
return hidden_groups + correction, ratio, raw_rms.detach().mean()
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
if not self._controller_enabled():
return super()._forward_full_active(x, original_shape, x_flat)
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_groups, ratio, raw_rms = self._hidden_controller(hidden_groups)
if not torch.is_grad_enabled():
self.last_hidden_controller_ratio = ratio.detach()
self.last_hidden_controller_raw_rms = raw_rms.detach()
output = hidden_groups.reshape(hidden.shape[0], self.intermediate_size) @ down_weight
return output.reshape(original_shape)
class TileRoutedHiddenDirectionMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Late-only hidden key/value direction memory for grouped4.
Each group owns a tiny bank of correction directions. Hidden activations
query the bank, select a value direction with softmax, and commit only an
RMS-bounded correction before the down projection. Values are zero-init, so
the model starts exactly as grouped4 and must learn useful directions during
pretraining instead of receiving an unbounded residual update.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.memory_slots = int(config.hidden_direction_memory_slots)
self.memory_rho_frac = float(config.hidden_direction_memory_rho_frac)
self.memory_late_layer_start = int(config.hidden_direction_memory_late_layer_start)
self.memory_temperature = float(config.hidden_direction_memory_temperature)
self.memory_key = nn.Parameter(torch.empty(self.num_groups, self.group_size, self.memory_slots))
self.memory_value = nn.Parameter(torch.zeros(self.num_groups, self.memory_slots, self.group_size))
nn.init.normal_(self.memory_key, mean=0.0, std=self.group_size**-0.5)
self.last_direction_memory_entropy: torch.Tensor | None = None
self.last_direction_memory_ratio: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
enabled = self._memory_enabled()
self.memory_key.requires_grad_(enabled)
self.memory_value.requires_grad_(enabled)
def _memory_enabled(self) -> bool:
return self.memory_rho_frac > 0.0 and self.block_idx >= self.memory_late_layer_start
def _direction_memory(self, hidden_groups: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
dtype = hidden_groups.dtype
hidden_rms = hidden_groups.float().square().mean(dim=-1, keepdim=True).sqrt().to(dtype=dtype)
h_norm = hidden_groups * torch.rsqrt(
hidden_groups.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + self.config.norm_eps
)
key = self.memory_key.to(dtype=dtype)
value = self.memory_value.to(dtype=dtype)
logits = torch.einsum("ngs,gsm->ngm", h_norm, key)
logits = logits / (math.sqrt(self.group_size) * self.memory_temperature)
weights = torch.softmax(logits.float(), dim=-1).to(dtype=dtype)
raw = torch.einsum("ngm,gms->ngs", weights, value)
raw_rms = torch.sqrt(raw.float().square().mean(dim=-1, keepdim=True) + self.config.norm_eps).to(dtype=dtype)
target = self.memory_rho_frac * hidden_rms
correction = raw * (target / raw_rms).clamp(max=1.0)
ratio = correction.float().square().mean().sqrt() / hidden_groups.float().square().mean().sqrt().clamp_min(1e-8)
entropy = -(weights.float() * weights.float().clamp_min(1e-8).log()).sum(dim=-1).mean()
return hidden_groups + correction, ratio, entropy
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return super()._forward_full_active(x, original_shape, x_flat)
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_groups, ratio, entropy = self._direction_memory(hidden_groups)
if not torch.is_grad_enabled():
self.last_direction_memory_ratio = ratio.detach()
self.last_direction_memory_entropy = entropy.detach()
output = hidden_groups.reshape(hidden.shape[0], self.intermediate_size) @ down_weight
return output.reshape(original_shape)
class TileRoutedScalarDirectionMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Speed-first hidden direction memory with no extra hidden matmul.
Each group stores one learned direction. A token reads that memory by
cosine-like alignment with its normalized hidden state, then writes a tiny
RMS-capped signed correction before the down projection. This is less
expressive than a key/value bank but keeps the extra work to reductions and
elementwise ops in late layers.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.scalar_memory_rho_frac = float(config.scalar_direction_memory_rho_frac)
self.scalar_memory_late_layer_start = int(config.scalar_direction_memory_late_layer_start)
self.scalar_memory_direction = nn.Parameter(torch.empty(self.num_groups, self.group_size))
self.scalar_memory_alpha_raw = nn.Parameter(torch.zeros(self.num_groups, 1, 1))
nn.init.normal_(self.scalar_memory_direction, mean=0.0, std=1.0)
self.last_scalar_memory_alpha: torch.Tensor | None = None
self.last_scalar_memory_gate_abs: torch.Tensor | None = None
self.last_scalar_memory_ratio: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
enabled = self._memory_enabled()
self.scalar_memory_direction.requires_grad_(enabled)
self.scalar_memory_alpha_raw.requires_grad_(enabled)
def _memory_enabled(self) -> bool:
return self.scalar_memory_rho_frac > 0.0 and self.block_idx >= self.scalar_memory_late_layer_start
def _scalar_memory(self, hidden_groups: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
dtype = hidden_groups.dtype
hidden_rms = hidden_groups.float().square().mean(dim=-1, keepdim=True).sqrt().to(dtype=dtype)
h_norm = hidden_groups * torch.rsqrt(
hidden_groups.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + self.config.norm_eps
)
direction = self.scalar_memory_direction.to(dtype=dtype)
direction = direction * torch.rsqrt(direction.float().square().mean(dim=-1, keepdim=True).to(dtype=dtype) + self.config.norm_eps)
direction = direction.unsqueeze(0)
gate = torch.tanh((h_norm * direction).mean(dim=-1, keepdim=True))
alpha = torch.tanh(self.scalar_memory_alpha_raw.to(dtype=dtype)).view(1, self.num_groups, 1)
correction = alpha * gate * direction * (self.scalar_memory_rho_frac * hidden_rms)
ratio = correction.float().square().mean().sqrt() / hidden_groups.float().square().mean().sqrt().clamp_min(1e-8)
return hidden_groups + correction, ratio, alpha.detach().abs().mean(), gate.detach().abs().mean()
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return super()._forward_full_active(x, original_shape, x_flat)
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_groups, ratio, alpha_abs, gate_abs = self._scalar_memory(hidden_groups)
if not torch.is_grad_enabled():
self.last_scalar_memory_ratio = ratio.detach()
self.last_scalar_memory_alpha = alpha_abs.detach()
self.last_scalar_memory_gate_abs = gate_abs.detach()
output = hidden_groups.reshape(hidden.shape[0], self.intermediate_size) @ down_weight
return output.reshape(original_shape)
class TileRoutedChannelMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Foldable FFN memory-value gain for grouped4.
FFN hidden channels act like keys; down-projection rows are the memory
values. This class learns a bounded per-channel value strength. It starts
exactly as grouped4 and can be folded into ``down_weight`` for inference,
so it changes what the FFN remembers without adding a new token-time
reasoning path.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.channel_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, self.group_size, 1))
self.last_channel_memory_gain_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.channel_memory_gain_raw.requires_grad_(self._memory_enabled())
def _memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _effective_down_weight(self) -> torch.Tensor:
if not self._memory_enabled():
return self.down_weight
gain = self.channel_memory_gain_max * torch.tanh(
self.channel_memory_gain_raw.to(device=self.down_weight.device, dtype=self.down_weight.dtype)
)
if not torch.is_grad_enabled():
self.last_channel_memory_gain_abs = gain.detach().abs().mean()
return self.down_weight * (1.0 + gain)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self._effective_down_weight().reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedChannelStateMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Per-channel token-conditioned memory for grouped4.
The previous scalar group-state memory could move argmaxes but only had one
learned gain per group. This variant keeps the same cheap group-state signal
but gives each hidden channel its own bounded value-control coefficient.
Zero gain is exactly grouped4.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.channel_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, self.group_size))
self.last_channel_state_gain_abs: torch.Tensor | None = None
self.last_channel_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.channel_memory_gain_raw.requires_grad_(self._memory_enabled())
def _memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _apply_channel_state_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
group_rms = hidden_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
group_state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
gain = self.channel_memory_gain_max * torch.tanh(
self.channel_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).unsqueeze(0)
if not torch.is_grad_enabled():
self.last_channel_state_gain_abs = gain.detach().abs().mean()
self.last_channel_state_abs = group_state.detach().abs().mean()
hidden_groups = hidden_groups * (1.0 + gain * group_state)
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_channel_state_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedCenteredChannelStateMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Mean-preserving per-channel token-conditioned memory for grouped4."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.channel_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, self.group_size))
self.last_channel_state_gain_abs: torch.Tensor | None = None
self.last_channel_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.channel_memory_gain_raw.requires_grad_(self._memory_enabled())
def _memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _apply_centered_channel_state_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
centered = hidden_groups - group_mean.to(dtype=hidden_groups.dtype)
group_rms = hidden_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
group_state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
gain = self.channel_memory_gain_max * torch.tanh(
self.channel_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).unsqueeze(0)
if not torch.is_grad_enabled():
self.last_channel_state_gain_abs = gain.detach().abs().mean()
self.last_channel_state_abs = group_state.detach().abs().mean()
hidden_groups = hidden_groups + centered * (gain * group_state)
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_centered_channel_state_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedCenteredChannelGroupStateMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Mean-preserving channel memory plus scalar group-state memory."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.group_state_memory_gain_max = float(config.group_state_memory_gain_max)
self.group_state_memory_late_layer_start = int(config.group_state_memory_late_layer_start)
self.channel_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, self.group_size))
self.group_state_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, 1))
self.last_channel_state_gain_abs: torch.Tensor | None = None
self.last_group_state_gain_abs: torch.Tensor | None = None
self.last_group_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.channel_memory_gain_raw.requires_grad_(self._channel_memory_enabled())
self.group_state_memory_gain_raw.requires_grad_(self._group_state_memory_enabled())
def _channel_memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _group_state_memory_enabled(self) -> bool:
return self.group_state_memory_gain_max > 0.0 and self.block_idx >= self.group_state_memory_late_layer_start
def _apply_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._channel_memory_enabled() and not self._group_state_memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
group_rms = hidden_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
group_state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
if self._channel_memory_enabled():
centered = hidden_groups - group_mean.to(dtype=hidden_groups.dtype)
channel_gain = self.channel_memory_gain_max * torch.tanh(
self.channel_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).unsqueeze(0)
if not torch.is_grad_enabled():
self.last_channel_state_gain_abs = channel_gain.detach().abs().mean()
hidden_groups = hidden_groups + centered * (channel_gain * group_state)
if self._group_state_memory_enabled():
group_gain = self.group_state_memory_gain_max * torch.tanh(
self.group_state_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).view(1, self.num_groups, 1)
if not torch.is_grad_enabled():
self.last_group_state_gain_abs = group_gain.detach().abs().mean()
self.last_group_state_abs = group_state.detach().abs().mean()
hidden_groups = hidden_groups * (1.0 + group_gain * group_state)
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedKeyedChannelMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Prototype-keyed token-conditioned memory for grouped4.
This reads a richer token signal than group mean/RMS: each group compares
its normalized centered hidden state to learned prototype keys, then uses
zero-init bounded value vectors to modulate centered channels. With zero
values the function is exactly grouped4.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.keyed_channel_memory_rank = int(config.keyed_channel_memory_rank)
key = torch.randn(self.num_groups, self.keyed_channel_memory_rank, self.group_size)
key = key * (self.group_size ** -0.5)
self.keyed_memory_key = nn.Parameter(key)
self.keyed_memory_value_raw = nn.Parameter(
torch.zeros(self.num_groups, self.keyed_channel_memory_rank, self.group_size)
)
self.last_keyed_memory_gain_abs: torch.Tensor | None = None
self.last_keyed_memory_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
enabled = self._memory_enabled()
self.keyed_memory_key.requires_grad_(enabled)
self.keyed_memory_value_raw.requires_grad_(enabled)
def _memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _apply_keyed_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
centered = hidden_groups - group_mean.to(dtype=hidden_groups.dtype)
centered_float = centered.float()
centered_rms = centered_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
centered_norm = centered_float / centered_rms
key = self.keyed_memory_key.to(device=hidden.device, dtype=torch.float32)
key = key / key.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
state = torch.tanh((centered_norm.unsqueeze(2) * key.unsqueeze(0)).mean(dim=-1))
value = self.channel_memory_gain_max * torch.tanh(
self.keyed_memory_value_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
)
channel_gain = (state.to(dtype=hidden_groups.dtype).unsqueeze(-1) * value.unsqueeze(0)).sum(dim=2)
channel_gain = channel_gain * (self.keyed_channel_memory_rank ** -0.5)
if not torch.is_grad_enabled():
self.last_keyed_memory_gain_abs = channel_gain.detach().abs().mean()
self.last_keyed_memory_state_abs = state.detach().abs().mean()
hidden_groups = hidden_groups + centered * channel_gain
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_keyed_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedGroupStateMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Token-conditioned bounded group memory for grouped4.
This keeps grouped4's full-active GEMM path, then lets each hidden group
scale its own channels from a signed normalized group state. Zero gain is
exactly grouped4, while nonzero gain gives a cheap token-dependent memory
write before the down projection.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.group_state_memory_gain_max = float(config.group_state_memory_gain_max)
self.group_state_memory_late_layer_start = int(config.group_state_memory_late_layer_start)
self.group_state_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, 1))
self.last_group_state_gain_abs: torch.Tensor | None = None
self.last_group_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.group_state_memory_gain_raw.requires_grad_(self._memory_enabled())
def _memory_enabled(self) -> bool:
return self.group_state_memory_gain_max > 0.0 and self.block_idx >= self.group_state_memory_late_layer_start
def _apply_group_state_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
group_rms = hidden_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
group_state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
gain = self.group_state_memory_gain_max * torch.tanh(
self.group_state_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).view(1, self.num_groups, 1)
if not torch.is_grad_enabled():
self.last_group_state_gain_abs = gain.detach().abs().mean()
self.last_group_state_abs = group_state.detach().abs().mean()
hidden_groups = hidden_groups * (1.0 + gain * group_state)
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_group_state_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedChannelGroupStateMemoryDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Combined foldable channel values plus token-conditioned group memory."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.block_idx = 0
self.n_layer = config.n_layer
self.channel_memory_gain_max = float(config.channel_memory_gain_max)
self.channel_memory_late_layer_start = int(config.channel_memory_late_layer_start)
self.group_state_memory_gain_max = float(config.group_state_memory_gain_max)
self.group_state_memory_late_layer_start = int(config.group_state_memory_late_layer_start)
self.channel_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, self.group_size, 1))
self.group_state_memory_gain_raw = nn.Parameter(torch.zeros(self.num_groups, 1))
self.last_channel_memory_gain_abs: torch.Tensor | None = None
self.last_group_state_gain_abs: torch.Tensor | None = None
self.last_group_state_abs: torch.Tensor | None = None
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
self.channel_memory_gain_raw.requires_grad_(self._channel_memory_enabled())
self.group_state_memory_gain_raw.requires_grad_(self._group_state_memory_enabled())
def _channel_memory_enabled(self) -> bool:
return self.channel_memory_gain_max > 0.0 and self.block_idx >= self.channel_memory_late_layer_start
def _group_state_memory_enabled(self) -> bool:
return self.group_state_memory_gain_max > 0.0 and self.block_idx >= self.group_state_memory_late_layer_start
def _effective_down_weight(self) -> torch.Tensor:
if not self._channel_memory_enabled():
return self.down_weight
gain = self.channel_memory_gain_max * torch.tanh(
self.channel_memory_gain_raw.to(device=self.down_weight.device, dtype=self.down_weight.dtype)
)
if not torch.is_grad_enabled():
self.last_channel_memory_gain_abs = gain.detach().abs().mean()
return self.down_weight * (1.0 + gain)
def _apply_group_state_memory(self, hidden: torch.Tensor) -> torch.Tensor:
if not self._group_state_memory_enabled():
return hidden
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
hidden_float = hidden_groups.float()
group_mean = hidden_float.mean(dim=-1, keepdim=True)
group_rms = hidden_float.square().mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-6)
group_state = torch.tanh(group_mean / group_rms).to(dtype=hidden_groups.dtype)
gain = self.group_state_memory_gain_max * torch.tanh(
self.group_state_memory_gain_raw.to(device=hidden.device, dtype=hidden_groups.dtype)
).view(1, self.num_groups, 1)
if not torch.is_grad_enabled():
self.last_group_state_gain_abs = gain.detach().abs().mean()
self.last_group_state_abs = group_state.detach().abs().mean()
hidden_groups = hidden_groups * (1.0 + gain * group_state)
return hidden_groups.reshape(hidden.shape)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self._effective_down_weight().reshape(self.intermediate_size, hidden_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden = self._apply_group_state_memory(hidden)
return (hidden @ down_weight).reshape(original_shape)
class TileRoutedGRNDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Grouped hidden Global Response Normalization before the down projection."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.grn_gamma = nn.Parameter(torch.zeros(self.num_groups, self.group_size))
self.grn_beta = nn.Parameter(torch.zeros(self.num_groups, self.group_size))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
group_norm = hidden_groups.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
response = group_norm / group_norm.mean(dim=1, keepdim=True).clamp_min(1e-6)
gamma = self.grn_gamma.to(device=hidden.device, dtype=hidden.dtype).unsqueeze(0)
beta = self.grn_beta.to(device=hidden.device, dtype=hidden.dtype).unsqueeze(0)
hidden_groups = hidden_groups + gamma * (hidden_groups * response.to(dtype=hidden.dtype)) + beta
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return (hidden_groups.reshape_as(hidden) @ down_weight).reshape(original_shape)
class TileRoutedGroupMixDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Learn a tiny soft group-to-group hidden mixer before the down weights."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
logits = torch.full((self.num_groups, self.num_groups), -6.0)
logits.fill_diagonal_(6.0)
self.group_mix_logits = nn.Parameter(logits)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
mix = torch.softmax(self.group_mix_logits.to(device=hidden.device, dtype=torch.float32), dim=-1).to(dtype=hidden.dtype)
hidden_groups = torch.einsum("og,ngs->nos", mix, hidden_groups)
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return (hidden_groups.reshape_as(hidden) @ down_weight).reshape(original_shape)
class TileRoutedTokenGroupGateDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Token-dependent competition between full-active groups."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.group_gate_alpha = nn.Parameter(torch.tensor(0.05))
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = original_shape[-1]
up_weight = self.up_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.value_weight.permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
hidden = F.silu(x_flat @ up_weight) * (x_flat @ value_weight)
hidden_groups = hidden.reshape(hidden.shape[0], self.num_groups, self.group_size)
scores = hidden_groups.float().pow(2).mean(dim=-1).sqrt()
gates = torch.softmax(scores, dim=-1).to(dtype=hidden.dtype).unsqueeze(-1) * self.num_groups
alpha = self.group_gate_alpha.to(device=hidden.device, dtype=hidden.dtype).clamp(-0.5, 0.5)
hidden_groups = hidden_groups * (1.0 + alpha * (gates - 1.0))
down_weight = self.down_weight.reshape(self.intermediate_size, hidden_size)
return (hidden_groups.reshape_as(hidden) @ down_weight).reshape(original_shape)
class TileRoutedBlockNormDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped MLP with block-local RMS normalization per group.
Hard hidden blocks damaged LM quality because each FFN block lost access to
most of the residual stream. This softer variant keeps the standard
full-hidden grouped MLP: every intermediate group still sees all hidden
channels. The non-absorbed part is an input-dependent local RMS normalization
over hidden-channel blocks before each group path.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
if config.n_embd % self.num_groups != 0:
raise ValueError("n_embd must be divisible by sparse_mlp_num_groups")
self.block_hidden_size = config.n_embd // self.num_groups
self.group_norm_weight = nn.Parameter(torch.ones(self.num_groups, config.n_embd))
def _block_norm(self, x_flat: torch.Tensor, group_idx: int) -> torch.Tensor:
x_blocks = x_flat.reshape(x_flat.shape[0], self.num_groups, self.block_hidden_size)
scale = torch.rsqrt(x_blocks.pow(2).mean(dim=-1, keepdim=True) + self.config.norm_eps)
x_normed = (x_blocks * scale).reshape_as(x_flat)
weight = self.group_norm_weight[group_idx].to(device=x_flat.device, dtype=x_flat.dtype)
return x_normed * weight
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
output = torch.zeros_like(x_flat)
for group_idx in range(self.num_groups):
group_input = self._block_norm(x_flat, group_idx)
hidden = F.silu(group_input @ self.up_weight[group_idx]) * (group_input @ self.value_weight[group_idx])
output = output + hidden @ self.down_weight[group_idx]
return output.reshape(original_shape)
class TileRoutedStackedDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped dSwiGLU with serial group communication.
At ``grouped_mlp_stack_alpha == 0`` this is equivalent to full-active
grouped 4/4. With positive alpha, each group sees the residual state
updated by earlier group outputs, making the groups a small ordered
sub-FFN instead of independent parallel subspaces.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
stack_alpha = torch.tensor(float(config.grouped_mlp_stack_alpha))
if config.grouped_mlp_stack_alpha_learnable:
self.stack_alpha = nn.Parameter(stack_alpha)
else:
self.register_buffer("stack_alpha", stack_alpha, persistent=False)
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
output = torch.zeros_like(x_flat)
state = x_flat
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for group_idx in range(self.num_groups):
hidden = F.silu(state @ self.up_weight[group_idx]) * (state @ self.value_weight[group_idx])
delta = hidden @ self.down_weight[group_idx]
output = output + delta
if group_idx + 1 < self.num_groups:
state = state + stack_alpha * delta
return output.reshape(original_shape)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
if self.active_groups != self.num_groups:
return super().forward(x)
output = self._forward_full_active(x, original_shape, x_flat)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return self._apply_shared_path(x, output)
class TileRoutedChunkedStackedDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full-active grouped MLP with serial stage chunks inside existing groups.
This keeps the exact ``TileRoutedDSwiGLUMLP`` parameter layout. The
intermediate channels in each group are split into ``grouped_mlp_stack_depth``
chunks. At alpha=0, summing all chunks is exactly the normal grouped MLP.
Non-zero alpha lets later chunks see a small residual correction from earlier
chunks, testing the ARC margin effect without changing initialization layout.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.stack_depth = config.grouped_mlp_stack_depth
if self.stack_depth < 1:
raise ValueError("grouped_mlp_stack_depth must be >= 1")
if self.group_size % self.stack_depth != 0:
raise ValueError("group_size must be divisible by grouped_mlp_stack_depth")
self.stage_group_size = self.group_size // self.stack_depth
self.stage_intermediate_size = self.stage_group_size * self.num_groups
stack_alpha = torch.tensor(float(config.grouped_mlp_stack_alpha))
if config.grouped_mlp_stack_alpha_learnable:
self.stack_alpha = nn.Parameter(stack_alpha)
else:
self.register_buffer("stack_alpha", stack_alpha, persistent=False)
output_scale = torch.tensor(float(config.grouped_mlp_stack_output_scale))
if config.grouped_mlp_stack_output_scale_learnable:
self.stack_output_scale = nn.Parameter(output_scale)
else:
self.register_buffer("stack_output_scale", output_scale, persistent=False)
def _stage_forward(self, stage_idx: int, state: torch.Tensor) -> torch.Tensor:
hidden_size = state.shape[-1]
start = stage_idx * self.stage_group_size
stop = (stage_idx + 1) * self.stage_group_size
up_weight = self.up_weight[:, :, start:stop].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
value_weight = self.value_weight[:, :, start:stop].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
down_weight = self.down_weight[:, start:stop, :].reshape(self.stage_intermediate_size, hidden_size)
return (F.silu(state @ up_weight) * (state @ value_weight)) @ down_weight
def _forward_full_active(self, x: torch.Tensor, original_shape: torch.Size, x_flat: torch.Tensor) -> torch.Tensor:
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
output = output + delta
if stage_idx + 1 < self.stack_depth:
state = state + stack_alpha * delta
output_scale = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
return (output * output_scale).reshape(original_shape)
class TileRoutedGatedStackedDSwiGLUMLP(nn.Module):
"""Grouped MLP main path plus a gated independent split-stage branch.
The total intermediate budget is conserved:
- ``intermediate_size - grouped_mlp_contrastive_intermediate_size`` goes to
a normal full-active grouped MLP main path.
- ``grouped_mlp_contrastive_intermediate_size`` goes to a split-stage
branch, the architecture that produced stronger ARC margins.
This tests whether the split-stage branch is useful as a bounded
contrastive/ranking perturbation instead of replacing the whole MLP.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.active_groups = config.sparse_mlp_max_active_groups
self.stack_depth = config.grouped_mlp_stack_depth
self.branch_intermediate_size = int(config.grouped_mlp_contrastive_intermediate_size)
if self.active_groups != self.num_groups:
raise ValueError("TileRoutedGatedStackedDSwiGLUMLP is a full-active grouped variant.")
if self.stack_depth < 1:
raise ValueError("grouped_mlp_stack_depth must be >= 1")
if config.bias:
raise ValueError("TileRoutedGatedStackedDSwiGLUMLP currently expects bias=False.")
if self.branch_intermediate_size <= 0:
raise ValueError("grouped_mlp_contrastive_intermediate_size must be > 0")
if self.branch_intermediate_size >= self.intermediate_size:
raise ValueError("contrastive branch must be smaller than intermediate_size")
self.main_intermediate_size = self.intermediate_size - self.branch_intermediate_size
if self.main_intermediate_size % self.num_groups != 0:
raise ValueError("main intermediate must divide across sparse_mlp_num_groups")
if self.branch_intermediate_size % (self.num_groups * self.stack_depth) != 0:
raise ValueError("contrastive branch intermediate must divide across groups * stack depth")
self.main_group_size = self.main_intermediate_size // self.num_groups
self.branch_stage_intermediate_size = self.branch_intermediate_size // self.stack_depth
self.branch_group_size = self.branch_stage_intermediate_size // self.num_groups
self.main_up_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.main_group_size))
self.main_value_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.main_group_size))
self.main_down_weight = nn.Parameter(torch.empty(self.num_groups, self.main_group_size, config.n_embd))
self.branch_up_weight = nn.Parameter(
torch.empty(self.stack_depth, self.num_groups, config.n_embd, self.branch_group_size)
)
self.branch_value_weight = nn.Parameter(
torch.empty(self.stack_depth, self.num_groups, config.n_embd, self.branch_group_size)
)
self.branch_down_weight = nn.Parameter(
torch.empty(self.stack_depth, self.num_groups, self.branch_group_size, config.n_embd)
)
stack_alpha = torch.tensor(float(config.grouped_mlp_stack_alpha))
if config.grouped_mlp_stack_alpha_learnable:
self.stack_alpha = nn.Parameter(stack_alpha)
else:
self.register_buffer("stack_alpha", stack_alpha, persistent=False)
gate_max = float(config.grouped_mlp_contrastive_gate_max)
if gate_max <= 0:
raise ValueError("grouped_mlp_contrastive_gate_max must be > 0")
self.contrastive_gate_max = gate_max
gate = float(config.grouped_mlp_contrastive_gate)
if config.grouped_mlp_contrastive_gate_learnable:
ratio = max(-0.999, min(0.999, gate / gate_max))
self.raw_contrastive_gate = nn.Parameter(torch.tensor(math.atanh(ratio)))
else:
self.register_buffer("contrastive_gate", torch.tensor(gate), persistent=False)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.main_up_weight, mean=0.0, std=std)
nn.init.normal_(self.main_value_weight, mean=0.0, std=std)
nn.init.normal_(self.main_down_weight, mean=0.0, std=std)
branch_std = std * float(getattr(self.config, "grouped_mlp_stack_init_scale", 1.0))
nn.init.normal_(self.branch_up_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_value_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_down_weight, mean=0.0, std=branch_std)
def _main_forward(self, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = x_flat.shape[-1]
up_weight = self.main_up_weight.permute(1, 0, 2).reshape(hidden_size, self.main_intermediate_size)
value_weight = self.main_value_weight.permute(1, 0, 2).reshape(hidden_size, self.main_intermediate_size)
down_weight = self.main_down_weight.reshape(self.main_intermediate_size, hidden_size)
return (F.silu(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
def _branch_stage_forward(self, stage_idx: int, state: torch.Tensor) -> torch.Tensor:
hidden_size = state.shape[-1]
up_weight = self.branch_up_weight[stage_idx].permute(1, 0, 2).reshape(
hidden_size, self.branch_stage_intermediate_size
)
value_weight = self.branch_value_weight[stage_idx].permute(1, 0, 2).reshape(
hidden_size, self.branch_stage_intermediate_size
)
down_weight = self.branch_down_weight[stage_idx].reshape(self.branch_stage_intermediate_size, hidden_size)
return (F.silu(state @ up_weight) * (state @ value_weight)) @ down_weight
def _branch_forward(self, x_flat: torch.Tensor) -> torch.Tensor:
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._branch_stage_forward(stage_idx, state)
output = output + delta
if stage_idx + 1 < self.stack_depth:
state = state + stack_alpha * delta
return output
def _gate(self, x_flat: torch.Tensor) -> torch.Tensor:
if hasattr(self, "raw_contrastive_gate"):
gate = self.contrastive_gate_max * torch.tanh(self.raw_contrastive_gate)
else:
gate = self.contrastive_gate
return gate.to(device=x_flat.device, dtype=x_flat.dtype)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
output = self._main_forward(x_flat) + self._gate(x_flat) * self._branch_forward(x_flat)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainGatedStackedDSwiGLUMLP(TileRoutedGatedStackedDSwiGLUMLP):
"""Full grouped MLP plus a ReZero-style split-stack residual branch.
``TileRoutedGatedStackedDSwiGLUMLP`` conserves total intermediate width by
moving channels from the normal grouped MLP into the split-stack branch. That
is useful as a fixed-budget control, but it weakens the LM path whenever the
branch gate starts small. This variant keeps the full grouped MLP width and
adds the split-stack branch on top behind the same bounded learnable gate.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.main_intermediate_size = self.intermediate_size
self.main_group_size = self.intermediate_size // self.num_groups
self.main_up_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.main_group_size))
self.main_value_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.main_group_size))
self.main_down_weight = nn.Parameter(torch.empty(self.num_groups, self.main_group_size, config.n_embd))
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.main_up_weight, mean=0.0, std=std)
nn.init.normal_(self.main_value_weight, mean=0.0, std=std)
nn.init.normal_(self.main_down_weight, mean=0.0, std=std)
branch_std = std * float(getattr(self.config, "grouped_mlp_stack_init_scale", 1.0))
nn.init.normal_(self.branch_up_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_value_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_down_weight, mean=0.0, std=branch_std)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self._branch_forward(x_flat)
output = main_output + self._gate(x_flat) * branch_output
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainRMSBoundedGatedStackedDSwiGLUMLP(TileRoutedFullMainGatedStackedDSwiGLUMLP):
"""Full-main gated split-stack with final RMS residual scaling.
The full-main gated screen showed that a small branch gate alone does not
prevent the whole MLP update from growing to split-stack scale. This variant
interprets ``grouped_mlp_stack_output_scale`` as the target output/input RMS
ratio for the complete MLP output.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
output_scale = torch.tensor(float(config.grouped_mlp_stack_output_scale))
if config.grouped_mlp_stack_output_scale_learnable:
self.stack_output_scale = nn.Parameter(output_scale)
else:
self.register_buffer("stack_output_scale", output_scale, persistent=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self._branch_forward(x_flat)
output = main_output + self._gate(x_flat) * branch_output
target_ratio = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
input_rms = x_flat.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
output_rms = output.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
scale = target_ratio * input_rms / output_rms.clamp_min(torch.finfo(output.dtype).eps)
output = output * scale
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP(
TileRoutedFullMainRMSBoundedGatedStackedDSwiGLUMLP
):
def _rms_bound_output(self, x_flat: torch.Tensor, output: torch.Tensor) -> torch.Tensor:
target_ratio = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
input_rms = x_flat.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
output_rms = output.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
scale = target_ratio * input_rms / output_rms.clamp_min(torch.finfo(output.dtype).eps)
return output * scale
def _match_token_rms(self, source: torch.Tensor, reference: torch.Tensor) -> torch.Tensor:
source_rms = source.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
reference_rms = reference.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
scale = reference_rms / source_rms.clamp_min(1e-8)
return source * scale.to(device=source.device, dtype=source.dtype)
class TileRoutedFullMainOrthogonalRMSGatedStackedDSwiGLUMLP(
TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP
):
"""Add only the branch component that is orthogonal to the main grouped update."""
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self._branch_forward(x_flat)
dot = (branch_output.float() * main_output.float()).sum(dim=-1, keepdim=True)
denom = main_output.float().pow(2).sum(dim=-1, keepdim=True).clamp_min(1e-8)
branch_output = branch_output - (dot / denom).to(dtype=branch_output.dtype) * main_output
branch_output = self._match_token_rms(branch_output, main_output)
output = main_output + self._gate(x_flat) * branch_output
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainAlignedRMSGatedStackedDSwiGLUMLP(
TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP
):
"""Convert the branch into a bounded token-wise modulation of the main update."""
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self._branch_forward(x_flat)
dot = (branch_output.float() * main_output.float()).sum(dim=-1, keepdim=True)
denom = main_output.float().pow(2).sum(dim=-1, keepdim=True).clamp_min(1e-8)
coeff = (dot / denom).clamp(min=-1.0, max=1.0).to(dtype=main_output.dtype)
output = main_output * (1.0 + self._gate(x_flat) * coeff)
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainNormedBranchRMSGatedStackedDSwiGLUMLP(
TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP
):
"""Normalize branch token RMS to the main update before the gated add."""
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self._match_token_rms(self._branch_forward(x_flat), main_output)
output = main_output + self._gate(x_flat) * branch_output
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedFullMainCompressedRMSGatedStackedDSwiGLUMLP(
TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP
):
"""Compress the split branch through a small dense bottleneck before adding it."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
rank = int(config.grouped_mlp_mix_rank) if config.grouped_mlp_mix_rank > 0 else max(1, config.n_embd // 4)
self.branch_compress = nn.Linear(config.n_embd, rank, bias=False)
self.branch_expand = nn.Linear(rank, config.n_embd, bias=False)
nn.init.normal_(self.branch_compress.weight, mean=0.0, std=0.02)
nn.init.normal_(self.branch_expand.weight, mean=0.0, std=0.02)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
branch_output = self.branch_expand(torch.tanh(self.branch_compress(self._branch_forward(x_flat))))
branch_output = self._match_token_rms(branch_output, main_output)
output = main_output + self._gate(x_flat) * branch_output
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedLateOnlyRMSGatedStackedDSwiGLUMLP(TileRoutedFullMainDirectionalRMSGatedStackedDSwiGLUMLP):
"""Use the branch only in later transformer blocks."""
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_output = self._main_forward(x_flat)
start_layer = int(getattr(self.config, "grouped_mlp_branch_start_layer", 0))
if getattr(self, "block_idx", 0) >= start_layer:
branch_output = self._match_token_rms(self._branch_forward(x_flat), main_output)
output = main_output + self._gate(x_flat) * branch_output
else:
output = main_output
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedAdditiveLateTokenRMSGatedDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Full grouped 4/4 MLP plus a tiny late token-conditioned branch.
This keeps the base grouped MLP parameter names and shapes unchanged so a
trained ``TileRoutedDSwiGLUMLP`` checkpoint can initialize the full LM path
exactly. Only the additive branch and its token gate are new.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
branch_intermediate_size = int(config.grouped_mlp_contrastive_intermediate_size)
if branch_intermediate_size <= 0:
raise ValueError("grouped_mlp_contrastive_intermediate_size must be > 0")
if branch_intermediate_size % self.num_groups != 0:
raise ValueError("contrastive branch intermediate must divide across sparse_mlp_num_groups")
self.branch_intermediate_size = branch_intermediate_size
self.branch_group_size = branch_intermediate_size // self.num_groups
self.branch_up_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.branch_group_size))
self.branch_value_weight = nn.Parameter(torch.empty(self.num_groups, config.n_embd, self.branch_group_size))
self.branch_down_weight = nn.Parameter(torch.empty(self.num_groups, self.branch_group_size, config.n_embd))
self.token_gate = nn.Linear(config.n_embd, 1, bias=True)
gate_max = float(config.grouped_mlp_contrastive_gate_max)
if gate_max <= 0:
raise ValueError("grouped_mlp_contrastive_gate_max must be > 0")
self.contrastive_gate_max = gate_max
gate = float(config.grouped_mlp_contrastive_gate)
ratio = max(-0.999, min(0.999, gate / gate_max))
self.initial_gate_raw = math.atanh(ratio)
self.block_idx = 0
self.n_layer = 1
self.reset_branch_parameters()
def reset_branch_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
branch_std = std * float(getattr(self.config, "grouped_mlp_stack_init_scale", 1.0))
nn.init.normal_(self.branch_up_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_value_weight, mean=0.0, std=branch_std)
nn.init.normal_(self.branch_down_weight, mean=0.0, std=branch_std)
nn.init.zeros_(self.token_gate.weight)
nn.init.constant_(self.token_gate.bias, self.initial_gate_raw)
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.n_layer = n_layer
def _branch_forward(self, x_flat: torch.Tensor) -> torch.Tensor:
hidden_size = x_flat.shape[-1]
up_weight = self.branch_up_weight.permute(1, 0, 2).reshape(hidden_size, self.branch_intermediate_size)
value_weight = self.branch_value_weight.permute(1, 0, 2).reshape(hidden_size, self.branch_intermediate_size)
down_weight = self.branch_down_weight.reshape(self.branch_intermediate_size, hidden_size)
return (F.silu(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
def _match_token_rms(self, source: torch.Tensor, reference: torch.Tensor) -> torch.Tensor:
source_rms = source.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
reference_rms = reference.float().pow(2).mean(dim=-1, keepdim=True).sqrt()
scale = reference_rms / source_rms.clamp_min(1e-8)
return source * scale.to(device=source.device, dtype=source.dtype)
def _token_gate(self, x_flat: torch.Tensor) -> torch.Tensor:
x_rms = x_flat.float().pow(2).mean(dim=-1, keepdim=True).sqrt().clamp_min(1e-8)
x_normed = (x_flat.float() / x_rms).to(dtype=x_flat.dtype)
raw_gate = self.token_gate(x_normed)
return self.contrastive_gate_max * torch.tanh(raw_gate).to(dtype=x_flat.dtype)
def forward(self, x: torch.Tensor) -> torch.Tensor:
main_output = super().forward(x)
start_layer = int(getattr(self.config, "grouped_mlp_branch_start_layer", 0))
if self.block_idx < start_layer:
return main_output
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
main_flat = main_output.reshape(-1, original_shape[-1])
branch_output = self._match_token_rms(self._branch_forward(x_flat), main_flat)
output = main_flat + self._token_gate(x_flat) * branch_output
return output.reshape(original_shape)
class TileRoutedStackedLayersDSwiGLUMLP(nn.Module):
"""Compute-matched stack of full-active grouped 4/4 MLP stages.
The total intermediate budget is split across ``grouped_mlp_stack_depth``
stages. Each stage is a full-active grouped 4/4 MLP. Earlier stages update
the state consumed by later stages, giving group layers a serial path for
communication without increasing total MLP parameters or matmul FLOPs.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.stack_depth = config.grouped_mlp_stack_depth
if self.stack_depth < 1:
raise ValueError("grouped_mlp_stack_depth must be >= 1")
if config.sparse_mlp_max_active_groups != self.num_groups:
raise ValueError("TileRoutedStackedLayersDSwiGLUMLP is a full-active 4/4 variant.")
if self.intermediate_size % (self.num_groups * self.stack_depth) != 0:
raise ValueError("intermediate_size must divide across stack_depth * sparse_mlp_num_groups")
if config.bias:
raise ValueError("TileRoutedStackedLayersDSwiGLUMLP currently expects bias=False.")
self.stage_intermediate_size = self.intermediate_size // self.stack_depth
self.group_size = self.stage_intermediate_size // self.num_groups
self.active_groups = config.sparse_mlp_max_active_groups
self.up_weight = nn.Parameter(torch.empty(self.stack_depth, self.num_groups, config.n_embd, self.group_size))
self.value_weight = nn.Parameter(torch.empty(self.stack_depth, self.num_groups, config.n_embd, self.group_size))
self.down_weight = nn.Parameter(torch.empty(self.stack_depth, self.num_groups, self.group_size, config.n_embd))
stack_alpha = torch.tensor(float(config.grouped_mlp_stack_alpha))
if config.grouped_mlp_stack_alpha_learnable:
self.stack_alpha = nn.Parameter(stack_alpha)
else:
self.register_buffer("stack_alpha", stack_alpha, persistent=False)
output_scale = torch.tensor(float(config.grouped_mlp_stack_output_scale))
if config.grouped_mlp_stack_output_scale_learnable:
self.stack_output_scale = nn.Parameter(output_scale)
else:
self.register_buffer("stack_output_scale", output_scale, persistent=False)
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
std *= float(getattr(self.config, "grouped_mlp_stack_init_scale", 1.0))
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
def _stage_forward(self, stage_idx: int, state: torch.Tensor) -> torch.Tensor:
hidden_size = state.shape[-1]
up_weight = self.up_weight[stage_idx].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
value_weight = self.value_weight[stage_idx].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
down_weight = self.down_weight[stage_idx].reshape(self.stage_intermediate_size, hidden_size)
return (F.silu(state @ up_weight) * (state @ value_weight)) @ down_weight
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
output = output + delta
if stage_idx + 1 < self.stack_depth:
state = state + stack_alpha * delta
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
output_scale = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
return (output * output_scale).reshape(original_shape)
class TileRoutedRMSBoundedStackedLayersDSwiGLUMLP(TileRoutedStackedLayersDSwiGLUMLP):
"""Stacked grouped layers with bounded residual update RMS.
The unbounded split-stack d2 alpha -0.03 improves some choice benchmarks
but produces MLP updates about four times stronger than dense/grouped MLPs.
Here ``grouped_mlp_stack_output_scale`` is interpreted as a target
output/input RMS ratio, keeping the serial group computation while limiting
how hard it can perturb the residual stream.
"""
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
output = output + delta
if stage_idx + 1 < self.stack_depth:
state = state + stack_alpha * delta
target_ratio = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
input_rms = x_flat.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
output_rms = output.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
scale = target_ratio * input_rms / output_rms.clamp_min(torch.finfo(output.dtype).eps)
output = output * scale
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedBlockLayeredNormDSwiGLUMLP(nn.Module):
"""Deep MLP with block-local RMS normalization at each layer.
Unlike split-stack which uses serial state updates, this has real layers
where the output of one layer feeds into the next layer. Each layer has
block-normalized grouped MLPs. This tests whether deep blocknorm layers
improve LM quality without the serial state perturbation of split-stack.
The depth is controlled by `grouped_mlp_stack_depth`, but each layer
is a full feedforward pass, not a state update.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.num_layers = config.grouped_mlp_stack_depth
self.config = config
if config.n_embd % self.num_groups != 0:
raise ValueError("n_embd must be divisible by sparse_mlp_num_groups")
self.block_hidden_size = config.n_embd // self.num_groups
self.group_size = self.intermediate_size // self.num_groups
# Each layer has its own grouped weights and block norm weights
self.layer_up_weight = nn.Parameter(torch.empty(self.num_layers, self.num_groups, config.n_embd, self.group_size))
self.layer_value_weight = nn.Parameter(torch.empty(self.num_layers, self.num_groups, config.n_embd, self.group_size))
self.layer_down_weight = nn.Parameter(torch.empty(self.num_layers, self.num_groups, self.group_size, config.n_embd))
self.layer_norm_weight = nn.Parameter(torch.ones(self.num_layers, self.num_groups, config.n_embd))
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.layer_up_weight, mean=0.0, std=std)
nn.init.normal_(self.layer_value_weight, mean=0.0, std=std)
nn.init.normal_(self.layer_down_weight, mean=0.0, std=std)
def _block_norm(self, x_flat: torch.Tensor, layer_idx: int, group_idx: int) -> torch.Tensor:
x_blocks = x_flat.reshape(x_flat.shape[0], self.num_groups, self.block_hidden_size)
scale = torch.rsqrt(x_blocks.pow(2).mean(dim=-1, keepdim=True) + self.config.norm_eps)
x_normed = (x_blocks * scale).reshape_as(x_flat)
weight = self.layer_norm_weight[layer_idx, group_idx].to(device=x_flat.device, dtype=x_flat.dtype)
return x_normed * weight
def _layer_forward(self, x_flat: torch.Tensor, layer_idx: int) -> torch.Tensor:
hidden_size = x_flat.shape[-1]
# Reshape layer weights for grouped computation
up_weight = self.layer_up_weight[layer_idx].permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
value_weight = self.layer_value_weight[layer_idx].permute(1, 0, 2).reshape(hidden_size, self.intermediate_size)
down_weight = self.layer_down_weight[layer_idx].reshape(self.intermediate_size, hidden_size)
# Apply block normalization and compute grouped output
output = torch.zeros_like(x_flat)
for group_idx in range(self.num_groups):
group_input = self._block_norm(x_flat, layer_idx, group_idx)
hidden = F.silu(group_input @ up_weight) * (group_input @ value_weight)
output = output + hidden @ down_weight
return output
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
# Pass through each layer sequentially (real layers, not state updates)
for layer_idx in range(self.num_layers):
x_flat = self._layer_forward(x_flat, layer_idx)
return x_flat.reshape(original_shape)
class TileRoutedRegulatedStackedLayersDSwiGLUMLP(TileRoutedRMSBoundedStackedLayersDSwiGLUMLP):
def _match_global_rms(self, source: torch.Tensor, reference: torch.Tensor) -> torch.Tensor:
source_rms = source.float().pow(2).mean().sqrt()
reference_rms = reference.float().pow(2).mean().sqrt()
scale = reference_rms / source_rms.clamp_min(1e-8)
return source * scale.to(device=source.device, dtype=source.dtype)
def _rms_bound_output(self, x_flat: torch.Tensor, output: torch.Tensor) -> torch.Tensor:
target_ratio = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
input_rms = x_flat.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
output_rms = output.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
scale = target_ratio * input_rms / output_rms.clamp_min(torch.finfo(output.dtype).eps)
return output * scale
class TileRoutedStageNormRMSBoundedStackedLayersDSwiGLUMLP(TileRoutedRegulatedStackedLayersDSwiGLUMLP):
"""Normalize each inter-stage state update before the next stage reads it."""
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
output = output + delta
if stage_idx + 1 < self.stack_depth:
state_delta = self._match_global_rms(delta, state)
state = state + stack_alpha * state_delta
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedWeightedStageRMSBoundedStackedLayersDSwiGLUMLP(TileRoutedRegulatedStackedLayersDSwiGLUMLP):
"""Learn a soft blend over stage deltas instead of summing stages equally."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.raw_stage_weight = nn.Parameter(torch.zeros(self.stack_depth))
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
state = x_flat
output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
weights = F.softmax(self.raw_stage_weight.float(), dim=0).to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
output = output + weights[stage_idx] * self.stack_depth * delta
if stage_idx + 1 < self.stack_depth:
state_delta = self._match_global_rms(delta, state)
state = state + stack_alpha * state_delta
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedMomentumRMSBoundedStackedLayersDSwiGLUMLP(TileRoutedRegulatedStackedLayersDSwiGLUMLP):
"""Use a momentum stream over stage deltas before updating later stages."""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
beta = min(max(float(config.grouped_mlp_mix_alpha), 1e-4), 1 - 1e-4)
self.raw_momentum_beta = nn.Parameter(torch.tensor(math.log(beta / (1.0 - beta))))
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
state = x_flat
output = torch.zeros_like(x_flat)
momentum: torch.Tensor | None = None
beta = torch.sigmoid(self.raw_momentum_beta).to(device=x_flat.device, dtype=x_flat.dtype)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
momentum = delta if momentum is None else beta * momentum + (1.0 - beta) * delta
output = output + momentum
if stage_idx + 1 < self.stack_depth:
state_delta = self._match_global_rms(momentum, state)
state = state + stack_alpha * state_delta
output = self._rms_bound_output(x_flat, output)
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedLateLayerStackedLayersDSwiGLUMLP(TileRoutedStackedLayersDSwiGLUMLP):
"""Split-stack only active in late layers where ranking matters most.
Early layers use standard grouped 4/4 MLP for clean LM learning.
Late layers (config.splitstack_late_layer_start and above) use split-stack
for stronger answer-choice ranking signal. This follows the momentum merge
result where late8 5% injection preserved LM quality while improving OpenBookQA.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.late_layer_start = getattr(config, "splitstack_late_layer_start", 0)
self.block_idx = 0
self.is_late_layer = False
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
self.is_late_layer = block_idx >= self.late_layer_start
def forward(self, x: torch.Tensor) -> torch.Tensor:
if not self.is_late_layer:
# Early layers: use single-stage grouped MLP (standard grouped 4/4)
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
hidden_size = x_flat.shape[-1]
up_weight = self.up_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
value_weight = self.value_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
down_weight = self.down_weight[0].reshape(self.stage_intermediate_size, hidden_size)
output = (F.silu(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
output_scale = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
return (output * output_scale).reshape(original_shape)
else:
# Late layers: use full split-stack
return super().forward(x)
class TileRoutedGatedStackedLayersDSwiGLUMLP(TileRoutedRMSBoundedStackedLayersDSwiGLUMLP):
"""Split-stack with learned per-layer contribution gate.
A scalar gate controls how much split-stack contributes vs a baseline
grouped MLP. This allows the model to learn when strong residual perturbations
are helpful (choice ranking) vs harmful (general LM modeling).
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
initial_gate = getattr(config, "splitstack_initial_gate", 0.1)
self.raw_gate = nn.Parameter(torch.tensor(math.log(initial_gate / (1.0 - initial_gate))))
self.block_idx = 0
def set_block_index(self, block_idx: int, n_layer: int) -> None:
self.block_idx = block_idx
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
# Baseline grouped MLP (single-stage)
hidden_size = x_flat.shape[-1]
up_weight = self.up_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
value_weight = self.value_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
down_weight = self.down_weight[0].reshape(self.stage_intermediate_size, hidden_size)
baseline_output = (F.silu(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
# Full split-stack output
state = x_flat
stack_output = torch.zeros_like(x_flat)
stack_alpha = self.stack_alpha.to(device=x_flat.device, dtype=x_flat.dtype)
for stage_idx in range(self.stack_depth):
delta = self._stage_forward(stage_idx, state)
stack_output = stack_output + delta
if stage_idx + 1 < self.stack_depth:
state = state + stack_alpha * delta
# RMS-bound the split-stack output
target_ratio = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
input_rms = x_flat.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
stack_rms = stack_output.float().pow(2).mean().sqrt().to(dtype=x_flat.dtype)
scale = target_ratio * input_rms / stack_rms.clamp_min(torch.finfo(stack_output.dtype).eps)
stack_output = stack_output * scale
# Learned gate between baseline and split-stack
gate = torch.sigmoid(self.raw_gate).to(device=x_flat.device, dtype=x_flat.dtype)
output = (1.0 - gate) * baseline_output + gate * stack_output
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(0.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return output.reshape(original_shape)
class TileRoutedAttentionOutputStackedLayersDSwiGLUMLP(TileRoutedStackedLayersDSwiGLUMLP):
"""Apply split-stack to attention output instead of MLP input.
The split-stack processes the attention residual stream, then the standard
grouped MLP processes the combined stream. This moves the strong residual
perturbation to the attention path where it may be less disruptive to
general LM modeling while still providing ranking signal.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
self.attention_stack_alpha = getattr(config, "splitstack_attention_alpha", -0.03)
self.attention_stack_alpha_tensor = torch.tensor(self.attention_stack_alpha)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# This MLP processes the combined stream after attention + split-stack
# The split-stack logic is handled at the Block level
# For now, use standard grouped MLP behavior
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
hidden_size = x_flat.shape[-1]
up_weight = self.up_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
value_weight = self.value_weight[0].permute(1, 0, 2).reshape(hidden_size, self.stage_intermediate_size)
down_weight = self.down_weight[0].reshape(self.stage_intermediate_size, hidden_size)
output = (F.silu(x_flat @ up_weight) * (x_flat @ value_weight)) @ down_weight
output_scale = self.stack_output_scale.to(device=x_flat.device, dtype=x_flat.dtype)
return (output * output_scale).reshape(original_shape)
class TileRoutedStaticDSwiGLUMLP(TileRoutedDSwiGLUMLP):
"""Static tile-routed dSwiGLU without the per-route Python math loop.
The active route pattern is fixed: tile group `g` uses expert groups
`g, g + 1, ...` modulo `num_groups`. This keeps the same structural routing
as `TileRoutedDSwiGLUMLP`, but executes the active routes as one batched
matmul stack instead of rolling weights and launching one MLP path per
route.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__(config, intermediate_size)
route_indices = [
(torch.arange(self.num_groups, dtype=torch.long) + route_offset) % self.num_groups
for route_offset in range(self.active_groups)
]
self.register_buffer("route_indices", torch.stack(route_indices, dim=0), persistent=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, original_shape[-1])
usable_tokens = x_flat.shape[0] - (x_flat.shape[0] % self.num_groups)
output = x_flat.new_empty(x_flat.shape)
if usable_tokens:
tokens_per_group = usable_tokens // self.num_groups
x_tiles = x_flat[:usable_tokens].reshape(self.num_groups, tokens_per_group, x_flat.shape[-1])
route_idx = self.route_indices.to(device=x_flat.device)
route_up = self.up_weight.index_select(0, route_idx.reshape(-1)).reshape(
self.active_groups, self.num_groups, x_flat.shape[-1], self.group_size
)
route_value = self.value_weight.index_select(0, route_idx.reshape(-1)).reshape(
self.active_groups, self.num_groups, x_flat.shape[-1], self.group_size
)
route_down = self.down_weight.index_select(0, route_idx.reshape(-1)).reshape(
self.active_groups, self.num_groups, self.group_size, x_flat.shape[-1]
)
x_routes = x_tiles.unsqueeze(0).expand(self.active_groups, -1, -1, -1)
up = torch.matmul(x_routes, route_up)
value = torch.matmul(x_routes, route_value)
out_tiles = torch.matmul(F.silu(up) * value, route_down).sum(dim=0)
output[:usable_tokens] = out_tiles.reshape(usable_tokens, x_flat.shape[-1])
if usable_tokens < x_flat.shape[0]:
tail = x_flat[usable_tokens:]
tail_out = torch.zeros_like(tail)
for route_offset in range(self.active_groups):
group_id = route_offset % self.num_groups
hidden = F.silu(tail @ self.up_weight[group_id]) * (tail @ self.value_weight[group_id])
tail_out = tail_out + hidden @ self.down_weight[group_id]
output[usable_tokens:] = tail_out
if not torch.is_grad_enabled():
self.last_gate_sparsity = x_flat.new_tensor(1.0 - self.active_groups / self.num_groups)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(float(self.active_groups))
return self._apply_shared_path(x, output.reshape(original_shape))
class TileRoutedDSwiGLUMLPStaticA2(nn.Module):
"""Hard-specialized fixed-route tile dSwiGLU for the 1M model.
This variant intentionally trades flexibility for a simpler execution
shape:
- hidden size 128
- intermediate size 1024
- 8 fixed tile groups
- exactly 2 active fixed routes per tile
There is no router, threshold, argsort, gather/scatter, or torch.roll in
the forward path. Route selection is encoded directly in the parameter
layout: `[active_route, tile_group, hidden, group_size]`.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.active_groups = config.sparse_mlp_max_active_groups
if config.bias:
raise ValueError("TileRoutedDSwiGLUMLPStaticA2 expects bias=False.")
if config.n_embd != 128:
raise ValueError("TileRoutedDSwiGLUMLPStaticA2 is specialized for hidden size 128.")
if self.intermediate_size != 1024:
raise ValueError("TileRoutedDSwiGLUMLPStaticA2 is specialized for intermediate size 1024.")
if self.num_groups != 8:
raise ValueError("TileRoutedDSwiGLUMLPStaticA2 is specialized for 8 groups.")
if self.active_groups != 2 or config.sparse_mlp_min_active_groups != 2:
raise ValueError("TileRoutedDSwiGLUMLPStaticA2 is specialized for active groups 2.")
self.group_size = self.intermediate_size // self.num_groups
self.up_weight = nn.Parameter(torch.empty(2, 8, 128, self.group_size))
self.value_weight = nn.Parameter(torch.empty(2, 8, 128, self.group_size))
self.down_weight = nn.Parameter(torch.empty(2, 8, self.group_size, 128))
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, 128)
usable_tokens = x_flat.shape[0] - (x_flat.shape[0] % 8)
output = x_flat.new_empty(x_flat.shape)
if usable_tokens:
tokens_per_group = usable_tokens // 8
x_tiles = x_flat[:usable_tokens].reshape(8, tokens_per_group, 128)
x_routes = x_tiles.unsqueeze(0)
up = torch.matmul(x_routes, self.up_weight)
value = torch.matmul(x_routes, self.value_weight)
out_tiles = torch.matmul(F.silu(up) * value, self.down_weight).sum(dim=0)
output[:usable_tokens] = out_tiles.reshape(usable_tokens, 128)
if usable_tokens < x_flat.shape[0]:
tail = x_flat[usable_tokens:]
tail_routes = tail.view(1, 1, -1, 128)
up = torch.matmul(tail_routes, self.up_weight[:, :1])
value = torch.matmul(tail_routes, self.value_weight[:, :1])
tail_out = torch.matmul(F.silu(up) * value, self.down_weight[:, :1]).sum(dim=(0, 1))
output[usable_tokens:] = tail_out
self.last_gate_sparsity = x_flat.new_tensor(0.75)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(2.0)
return output.reshape(original_shape)
class TileRoutedDSwiGLUMLPStaticGPTS(nn.Module):
"""Hard-specialized fixed-route tile dSwiGLU for GPT-S-5M shape.
Exact shape:
- hidden size 192
- intermediate size 672
- 6 fixed tile groups
- 2 active fixed routes per tile
This intentionally uses a route-major parameter layout to avoid dynamic
routing overhead during training: no router, threshold, argsort,
gather/scatter, or torch.roll in the forward path.
"""
def __init__(self, config: Config, intermediate_size: int | None = None) -> None:
super().__init__()
self.intermediate_size = intermediate_size or config.intermediate_size
self.num_groups = config.sparse_mlp_num_groups
self.active_groups = config.sparse_mlp_max_active_groups
if config.bias:
raise ValueError("TileRoutedDSwiGLUMLPStaticGPTS expects bias=False.")
if config.n_embd != 192:
raise ValueError("TileRoutedDSwiGLUMLPStaticGPTS is specialized for hidden size 192.")
if self.intermediate_size != 672:
raise ValueError("TileRoutedDSwiGLUMLPStaticGPTS is specialized for intermediate size 672.")
if self.num_groups != 6:
raise ValueError("TileRoutedDSwiGLUMLPStaticGPTS is specialized for 6 groups.")
if self.active_groups != 2 or config.sparse_mlp_min_active_groups != 2:
raise ValueError("TileRoutedDSwiGLUMLPStaticGPTS is specialized for active groups 2.")
self.group_size = self.intermediate_size // self.num_groups
self.up_weight = nn.Parameter(torch.empty(2, 6, 192, self.group_size))
self.value_weight = nn.Parameter(torch.empty(2, 6, 192, self.group_size))
self.down_weight = nn.Parameter(torch.empty(2, 6, self.group_size, 192))
self.config = config
self.last_gate_sparsity: torch.Tensor | None = None
self.last_threshold_mean: torch.Tensor | None = None
self.last_active_groups_mean: torch.Tensor | None = None
self.reset_parameters()
def reset_parameters(self) -> None:
std = math.sqrt(2.0 / 5 / self.config.n_embd)
nn.init.normal_(self.up_weight, mean=0.0, std=std)
nn.init.normal_(self.value_weight, mean=0.0, std=std)
nn.init.normal_(self.down_weight, mean=0.0, std=std)
def forward(self, x: torch.Tensor) -> torch.Tensor:
original_shape = x.shape
x_flat = x.reshape(-1, 192)
usable_tokens = x_flat.shape[0] - (x_flat.shape[0] % 6)
output = x_flat.new_empty(x_flat.shape)
if usable_tokens:
tokens_per_group = usable_tokens // 6
x_tiles = x_flat[:usable_tokens].reshape(6, tokens_per_group, 192)
x_routes = x_tiles.unsqueeze(0)
up = torch.matmul(x_routes, self.up_weight)
value = torch.matmul(x_routes, self.value_weight)
out_tiles = torch.matmul(F.silu(up) * value, self.down_weight).sum(dim=0)
output[:usable_tokens] = out_tiles.reshape(usable_tokens, 192)
if usable_tokens < x_flat.shape[0]:
tail = x_flat[usable_tokens:]
tail_routes = tail.view(1, 1, -1, 192)
up = torch.matmul(tail_routes, self.up_weight[:, :1])
value = torch.matmul(tail_routes, self.value_weight[:, :1])
tail_out = torch.matmul(F.silu(up) * value, self.down_weight[:, :1]).sum(dim=(0, 1))
output[usable_tokens:] = tail_out
self.last_gate_sparsity = x_flat.new_tensor(2.0 / 3.0)
self.last_threshold_mean = x_flat.new_tensor(0.0)
self.last_active_groups_mean = x_flat.new_tensor(2.0)
return output.reshape(original_shape)
class GemmaMLP(LLaMAMLP):
def forward(self, x: torch.Tensor) -> torch.Tensor:
x_fc_1 = self.fc_1(x)
x_fc_2 = self.fc_2(x)
x = F.gelu(x_fc_1, approximate=self.config.gelu_approximate) * x_fc_2
return self.proj(x)
class LLaMAMoE(nn.Module):
def __init__(self, config: Config) -> None:
super().__init__()
self.gate = (
nn.Linear(config.n_embd, config.n_expert, bias=False)
if not config.n_expert_groups
else GroupedTopkRouter(config)
)
self.experts = nn.ModuleList(
LLaMAMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(config.n_expert)
)
if config.n_shared_expert:
self.shared_experts = LLaMAMLP(
config, intermediate_size=config.moe_intermediate_size * config.n_shared_expert
)
self.config = config
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Derived from: https://github.com/mistralai/mistral-src/blob/b46d6/moe_one_file_ref.py#L203-L219
See also figure 1 in https://arxiv.org/abs/2211.15841
"""
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
residual_x = x.clone()
x = x.view(-1, C) # (B*T, C)
if not self.config.n_expert_groups:
router = self.gate(x) # (B*T, n_expert)
probs, indices = torch.topk(router, self.config.n_expert_per_token) # (B*T, n_expert_per_token)
probs = probs.softmax(dim=1, dtype=torch.float).to(dtype=x.dtype)
else:
probs, indices = self.gate(x)
if self.config.routed_scaling_factor != 1.0:
probs = probs * self.config.routed_scaling_factor
masks = indices.unsqueeze(-1) == torch.arange(self.config.n_expert, device=x.device)
masks = masks.permute(2, 0, 1) # (n_expert, B*T, n_expert_per_token)
y = torch.zeros_like(x) # (B*T, C)
for mask, expert in zip(masks, self.experts):
token_idx, expert_idx = torch.where(mask)
y[token_idx] += probs[token_idx, expert_idx, None] * expert(x[token_idx])
y = y.view(B, T, C)
if self.config.n_shared_expert:
y = y + self.shared_experts(residual_x)
return y
class GroupedTopkRouter(nn.Module):
"""
Derived from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/deepseek_v3/modeling_deepseek_v3.py.
DeepseekV3TopkRouter class.
"""
def __init__(self, config: Config) -> None:
super().__init__()
self.config = config
self.weight = nn.Parameter(torch.empty(config.n_expert, config.n_embd))
self.register_buffer("e_score_correction_bias", torch.zeros(config.n_expert))
@torch.no_grad()
def get_topk_indices(self, scores: torch.Tensor) -> torch.Tensor:
scores_for_choice = scores.view(-1, self.config.n_expert) + self.e_score_correction_bias.unsqueeze(0)
group_scores = (
scores_for_choice.view(-1, self.config.n_expert_groups, self.config.n_expert // self.config.n_expert_groups)
.topk(self.config.n_topk_scores_per_group, dim=-1)[0] # Top k scores for each group
.sum(dim=-1)
)
group_idx = torch.topk(group_scores, k=self.config.n_topk_groups, dim=-1, sorted=False)[1]
group_mask = torch.zeros_like(group_scores)
group_mask.scatter_(1, group_idx, 1)
score_mask = (
group_mask.unsqueeze(-1)
.expand(-1, self.config.n_expert_groups, self.config.n_expert // self.config.n_expert_groups)
.reshape(-1, self.config.n_expert)
)
scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
topk_indices = torch.topk(scores_for_choice, k=self.config.n_expert_per_token, dim=-1, sorted=False)[1]
return topk_indices
def forward(self, x: torch.Tensor) -> torch.Tensor:
router_logits = F.linear(x.type(torch.float32), self.weight.type(torch.float32))
scores = router_logits.sigmoid()
topk_indices = self.get_topk_indices(scores)
topk_weights = scores.gather(1, topk_indices)
if self.config.norm_topk_prob:
denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
topk_weights /= denominator
return topk_weights, topk_indices
# ROPE: YaRN (Yet another RoPE extensioN) scaling function for extended context
def yarn_get_mscale(scale=1, mscale=1):
if scale <= 1:
return 1.0
return 0.1 * mscale * math.log(scale) + 1.0
def build_rope_cache(
seq_len: int,
n_elem: int,
device: torch.device | None = None,
base: int = 10000,
condense_ratio: int = 1,
extra_config: dict | None = None,
rope_local_base_freq: float | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Enhanced Transformer with Rotary Position Embedding.
Args:
seq_len (int): Sequence length.
n_elem (int): Number of elements (head dimension).
device (torch.device, optional): Device for tensor allocations.
base (int, optional): Base for computing inverse frequencies.
condense_ratio (int, optional): Ratio to condense the position indices.
extra_config (dict, optional): Configuration parameters for frequency adjustments (used by Llama 3.1 and 3.2)
Returns:
Tuple[torch.Tensor, torch.Tensor]: Cosine and sine caches for RoPE.
Shapes are `(seq_len, n_elem)`.
"""
# Compute the inverse frequencies theta
theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, device=device).float() / n_elem))
# Initialize attention scaling factor (modified for YaRN)
attention_scaling = 1.0
if extra_config is not None:
factor = extra_config["factor"]
# Check YaRN first (has beta_fast/beta_slow)
if "beta_fast" in extra_config or "beta_slow" in extra_config:
# YaRN-style RoPE scaling
beta_fast = extra_config["beta_fast"]
beta_slow = extra_config["beta_slow"]
original_max_seq_len = extra_config["original_max_seq_len"]
# Calculate attention scaling factor based on mscale and mscale_all_dim
mscale = extra_config.get("mscale")
mscale_all_dim = extra_config.get("mscale_all_dim")
if mscale and mscale_all_dim:
attention_scaling = yarn_get_mscale(factor, mscale) / yarn_get_mscale(factor, mscale_all_dim)
elif mscale_all_dim:
attention_scaling = yarn_get_mscale(factor, mscale_all_dim)
elif mscale:
attention_scaling = yarn_get_mscale(factor, mscale)
# else: attention_scaling remains 1.0
# Create two frequency sets: extrapolation (unscaled) and interpolation (scaled)
pos_freqs = base ** (torch.arange(0, n_elem, 2, device=device).float() / n_elem)
theta_extrapolation = 1.0 / pos_freqs
theta_interpolation = 1.0 / (factor * pos_freqs)
# Find correction range based on rotation counts
# Inverse dimension formula to find dimension based on number of rotations
def find_correction_dim(num_rotations, dim, base_val, max_pos):
return (dim * math.log(max_pos / (num_rotations * 2 * math.pi))) / (2 * math.log(base_val))
low_dim = find_correction_dim(beta_fast, n_elem, base, original_max_seq_len)
high_dim = find_correction_dim(beta_slow, n_elem, base, original_max_seq_len)
# Apply truncation if specified
if extra_config.get("truncate", True):
low_dim = math.floor(low_dim)
high_dim = math.ceil(high_dim)
low_dim = max(low_dim, 0)
high_dim = min(high_dim, n_elem // 2 - 1)
# Create linear ramp factor for blending
dim_range = torch.arange(n_elem // 2, device=device, dtype=torch.float32)
if low_dim == high_dim:
high_dim += 0.001 # Prevent singularity
linear_func = (dim_range - low_dim) / (high_dim - low_dim)
ramp_func = torch.clamp(linear_func, 0.0, 1.0)
# Blend extrapolation and interpolation frequencies
# ramp_func = 0 -> use interpolation (scaled), ramp_func = 1 -> use extrapolation (unscaled)
theta_extrapolation_factor = ramp_func
theta = (
theta_interpolation * (1 - theta_extrapolation_factor)
+ theta_extrapolation * theta_extrapolation_factor
)
elif "original_max_seq_len" in extra_config:
# Llama3-style RoPE scaling
orig_context_len = extra_config["original_max_seq_len"]
low_freq_factor = extra_config["low_freq_factor"]
high_freq_factor = extra_config["high_freq_factor"]
wavelen = 2 * torch.pi / theta
ratio = orig_context_len / wavelen
smooth_factor = (ratio - low_freq_factor) / (high_freq_factor - low_freq_factor)
smooth_factor = torch.clamp(smooth_factor, min=0.0, max=1.0)
# Compute adjusted_theta without masked indexing
adjusted_theta = (1 - smooth_factor) * (theta / factor) + smooth_factor * theta
theta = adjusted_theta
else:
# Linear scaling fallback
theta = theta / factor
# Create position indices `[0, 1, ..., seq_len - 1]`
seq_idx = torch.arange(seq_len, device=device).float() / condense_ratio
# Calculate the product of position index and $\theta_i$
idx_theta = torch.outer(seq_idx, theta).repeat(1, 2)
# If `n_elem` is odd, the final dimension of `idx_theta` has size
# `n_elem + 1`, so need to cut something off.
# Due to a current bug in Hugging Face, in the case `n_elem == 1`, we leave
# `idx_theta`, `cos`, `sin` as is. Things work out in `apply_rope` due to
# broadcasting. If we shorten `idx_theta`, unit tests comparing to
# Hugging Face fail.
# https://github.com/huggingface/transformers/issues/35233
if idx_theta.shape[-1] > n_elem > 1:
idx_theta = idx_theta[..., :n_elem]
# if rope_local_base_freq is given, have a separate rope value for local embedding
# For now, we use default RoPE for local embedding
if rope_local_base_freq is not None:
local_theta = 1.0 / (rope_local_base_freq ** (torch.arange(0, n_elem, 2, device=device).float() / n_elem))
local_idx_theta = torch.outer(seq_idx, local_theta)
local_idx_theta = local_idx_theta.repeat(1, 2)
if local_idx_theta.shape[-1] > n_elem > 1:
local_idx_theta = local_idx_theta[..., :n_elem]
idx_theta = torch.stack((idx_theta, local_idx_theta), dim=-1)
cos = torch.cos(idx_theta) * attention_scaling
sin = torch.sin(idx_theta) * attention_scaling
return cos, sin
def batched_index_select(t, dim, idx):
"""index_select for batched index and unbatched t"""
if idx.dim() == 1:
return torch.index_select(t, dim, idx)
*batch_shape, idx_size = idx.shape
res = torch.index_select(t, dim, idx.reshape(-1)) # flat index
# split out single batch idx
res = res.view(*t.shape[:dim], -1, idx_size, *t.shape[dim + 1 :])
if dim > 0:
# move batch dim to front, this is np.rollaxis(res, dim, 0) for tensors
dims = [dim] + list(range(res.dim()))
del dims[dim + 1]
res = res.permute(dims)
# unflatten batch dims
res = res.view(*batch_shape, *res.shape[1:])
return res
def batched_index_copy_(t, dim, idx, val):
"""Index copy for batched t, idx, val"""
if t.device.type == "mps":
# Normalize negative dimensions
if dim < 0:
dim = t.dim() + dim
if idx.dim() == 1:
idx_shape = [1] * val.dim()
idx_shape[dim] = -1
idx_expanded = idx.view(*idx_shape)
idx_expanded = idx_expanded.expand_as(val)
t.scatter_(dim, idx_expanded, val)
return t
elif idx.dim() == 2:
assert dim != 0, "Cannot index the batch dimension"
batch_size = idx.size(0)
idx_size = idx.size(1)
assert batch_size == t.size(0) == val.size(0)
idx_shape = [batch_size] + [1] * (val.dim() - 1)
idx_shape[dim] = idx_size
idx_expanded = idx.view(*idx_shape)
idx_expanded = idx_expanded.expand_as(val)
t.scatter_(dim, idx_expanded, val)
return t
else:
raise NotImplementedError(f"idx.dim() == {idx.dim()} not supported")
else:
if idx.dim() == 1:
return t.index_copy_(dim, idx, val)
assert idx.dim() == 2, f"multiple batch dims not yet {idx.shape=}"
assert dim != 0, f"cannot index batch dim {dim=}"
batch_size, idx_size = idx.shape
assert batch_size == t.size(0)
assert batch_size == val.size(0)
# if we can view the batch and indexed dimensions together, we could
# do index trickery. This is, sadly, not the case for kvcache so we
# fall back to for loop
for i in range(batch_size):
unbatched_dim = dim if dim < 0 else dim - 1
t[i].index_copy_(unbatched_dim, idx[i], val[i])
return t
def apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
"""
Applies RoPE transform to `x`. Note that `cos`, `sin` need to have a batch
dimension.
Args:
x: Input tensor, `(B, ..., T, head_size)`
cos: Cached cosines, `(B, T, head_size)` or `(1, T, head_size)`
sin: Cached sines, `(B, T, head_size)` or `(1, T, head_size)`
Returns:
Encoded tensor, `(B, ..., T, head_size)`
"""
if cos.dim() != 3:
raise ValueError(f"cos must be three-dimensional, but shape is {cos.shape}")
if cos.shape != sin.shape:
raise ValueError(f"cos, sin must have same shape, but cos.shape={cos.shape}, sin.shape={sin.shape}")
head_size_half = x.size(-1) // 2
x1 = x[..., :head_size_half] # (B, ..., T, head_size/2)
x2 = x[..., head_size_half:] # (B, ..., T, head_size/2)
rotated = torch.cat((-x2, x1), dim=-1) # (B, ..., T, head_size)
dims_diff = x.dim() - cos.dim()
if dims_diff > 0:
# Ensure that shapes of `x`, `cos`, `sin` align
new_shape = cos.shape[0:1] + (1,) * dims_diff + cos.shape[1:]
cos = cos.view(*new_shape)
sin = sin.view(*new_shape)
roped = (x * cos) + (rotated * sin)
return roped.to(dtype=x.dtype)
def apply_rope_interleave(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
"""Apply rotary position embeddings with interleaved tensor layout.
This version rearranges the input tensor to group even/odd indices separately
before applying the standard RoPE rotation, matching HuggingFace's
apply_rotary_pos_emb_interleave behavior.
Args:
x: Input tensor of shape (..., seq_len, head_dim)
cos: Cosine component of shape (B, seq_len, head_dim) or (1, seq_len, head_dim)
sin: Sine component of shape (B, seq_len, head_dim) or (1, seq_len, head_dim)
Returns:
Tensor with RoPE applied, same shape as input
"""
if cos.dim() != 3:
raise ValueError(f"cos must be three-dimensional, but shape is {cos.shape}")
if cos.shape != sin.shape:
raise ValueError(f"cos, sin must have same shape, but cos.shape={cos.shape}, sin.shape={sin.shape}")
# Rearrange tensor to group even/odd indices: [x0,x1,x2,x3,...] -> [x0,x2,x4,...,x1,x3,x5,...]
*batch_dims, d = x.shape
x = x.view(*batch_dims, d // 2, 2).transpose(-1, -2).reshape(*batch_dims, d)
# Standard rotation logic (same as apply_rope)
head_size_half = x.size(-1) // 2
x1 = x[..., :head_size_half]
x2 = x[..., head_size_half:]
rotated = torch.cat((-x2, x1), dim=-1)
# Auto-detect dimension mismatch and reshape cos/sin
dims_diff = x.dim() - cos.dim()
if dims_diff > 0:
new_shape = cos.shape[0:1] + (1,) * dims_diff + cos.shape[1:]
cos = cos.view(*new_shape)
sin = sin.view(*new_shape)
roped = (x * cos) + (rotated * sin)
return roped.to(dtype=x.dtype)
def do_softcapping(x: torch.Tensor, thresh: float) -> torch.Tensor:
return torch.tanh(x / thresh) * thresh
class MLACompressedKVCache(nn.Module):
"""Cache normalized c^KV and the shared RoPE key instead of expanded K/V."""
def __init__(
self,
latent_shape: tuple[int, int, int],
rope_shape: tuple[int, int, int, int],
device: torch.device | None = None,
dtype: torch.dtype | None = None,
) -> None:
super().__init__()
self.register_buffer("latent", torch.zeros(latent_shape, device=device, dtype=dtype), persistent=False)
self.register_buffer("rope", torch.zeros(rope_shape, device=device, dtype=dtype), persistent=False)
def forward(
self, input_pos: torch.Tensor, latent: torch.Tensor, rope: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
if self.latent.dtype != latent.dtype:
self.latent = self.latent.to(latent.dtype)
if self.rope.dtype != rope.dtype:
self.rope = self.rope.to(rope.dtype)
bs = latent.size(0)
if input_pos.dim() == 1:
self.latent[:bs].index_copy_(1, input_pos, latent)
self.rope[:bs].index_copy_(2, input_pos, rope)
else:
batch = torch.arange(bs, device=input_pos.device)[:, None]
self.latent[batch, input_pos] = latent
self.rope[:bs, 0][batch, input_pos] = rope[:, 0]
return self.latent[:bs], self.rope[:bs]
@property
def bytes_per_token(self) -> int:
return (self.latent.shape[-1] + self.rope.shape[-1]) * self.latent.element_size()
class KVCache(nn.Module):
"""
Buffers `k`, `v` have shape
`(batch_size, n_query_groups, max_seq_length, head_size)`.
"""
def __init__(
self,
k_shape: tuple[int, int, int, int],
v_shape: tuple[int, int, int, int],
device: torch.device | None = None,
dtype: torch.dtype | None = None,
is_sliding_window: bool = False,
sliding_window_size: int | None = None,
) -> None:
super().__init__()
self.register_buffer("k", torch.zeros(k_shape, device=device, dtype=dtype), persistent=False)
self.register_buffer("v", torch.zeros(v_shape, device=device, dtype=dtype), persistent=False)
self.is_sliding_window = is_sliding_window
self.sliding_window_size = sliding_window_size
self.max_cache_len = k_shape[2]
def forward(self, input_pos: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Writes new values `k` and `v` into the cache at the positions specified
by `input_pos` along the sequence dimension (`max_seq_length`). The batch
size of `k` and `v` (`bs`) must be smaller or equal to `KVCache` batch
size. Returns the full buffers, adjusted to the batch size `bs`.
Args:
input_pos: Position index, `(bs, T)` or `(T,)`
k: New values, `(bs, n_query_groups, T, head_size)`
v: New values, `(bs, n_query_groups, T, head_size)`
Returns:
k_full, v_full, `(bs, n_query_groups, max_seq_length, head_size)`
"""
# move the buffer to the activation dtype for when AMP is used
if self.k.dtype != k.dtype:
self.k = self.k.to(k.dtype)
if self.v.dtype != v.dtype:
self.v = self.v.to(v.dtype)
# update the cache
bs = k.size(0)
if self.is_sliding_window:
# Circular buffer for sliding window
prefill_len = input_pos.shape[-1]
if prefill_len > self.max_cache_len:
raise ValueError(
f"Prefill length ({prefill_len}) exceeds the sliding window size ({self.max_cache_len}). "
f"This causes the ring-buffer KV cache to overwrite entries, but the attention mask is not "
f"rebuilt to reflect the true positions, which silently violates causality. "
f"Please use chunked prefill with chunk size <= {self.max_cache_len} to avoid this issue."
)
cache_positions = input_pos % self.max_cache_len
k = batched_index_copy_(self.k[:bs, ...], -2, cache_positions, k)
v = batched_index_copy_(self.v[:bs, ...], -2, cache_positions, v)
max_pos = input_pos.max().item()
if max_pos < self.max_cache_len:
k = k[:, :, : max_pos + 1, :]
v = v[:, :, : max_pos + 1, :]
else:
# Standard KV cache (global attention)
k = batched_index_copy_(self.k[:bs, ...], -2, input_pos, k)
v = batched_index_copy_(self.v[:bs, ...], -2, input_pos, v)
return k, v
def reset_parameters(self) -> None:
torch.nn.init.zeros_(self.k)
torch.nn.init.zeros_(self.v)
def build_mask_cache(max_seq_length: int, device: torch.device | None = None) -> torch.Tensor:
ones = torch.ones((max_seq_length, max_seq_length), device=device, dtype=torch.bool)
return torch.tril(ones).unsqueeze(0).unsqueeze(0)
class RMSNorm(torch.nn.Module):
"""Root Mean Square Layer Normalization.
Derived from https://github.com/bzhangGo/rmsnorm/blob/master/rmsnorm_torch.py. BSD 3-Clause License:
https://github.com/bzhangGo/rmsnorm/blob/master/LICENSE.
"""
def __init__(self, size: int, dim: int = -1, eps: float = 1e-6, add_unit_offset: bool = False) -> None:
super().__init__()
self.weight = torch.nn.Parameter(torch.ones(size))
self.eps = eps
self.dim = dim
self.add_unit_offset = add_unit_offset
def forward(self, x: torch.Tensor) -> torch.Tensor:
dtype = x.dtype
x = x.float()
# NOTE: the original RMSNorm paper implementation is not equivalent
norm_x = torch.mean(x * x, dim=self.dim, keepdim=True)
x_normed = x * torch.rsqrt(norm_x + self.eps)
weight = (1 + self.weight) if self.add_unit_offset else self.weight
return (x_normed * weight.float()).to(dtype=dtype)
def reset_parameters(self) -> None:
torch.nn.init.ones_(self.weight)
|