File size: 104,715 Bytes
e4ab0d4 | 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 | import os
import sys
import logging
import torch
import importlib
import math
import datetime
from tqdm import tqdm
import folder_paths
import comfy.model_management as mm
from comfy.cli_args import args
from comfy.ldm.modules.attention import wrap_attn, optimized_attention, attention_pytorch
import comfy.utils
import comfy.sd
try:
from comfy_api.latest import io
v3_available = True
except ImportError:
v3_available = False
logging.warning("ComfyUI v3 node API not available, please update ComfyUI to access latest v3 nodes.")
sageattn_modes = ["disabled", "auto", "sageattn_qk_int8_pv_fp16_cuda", "sageattn_qk_int8_pv_fp16_triton", "sageattn_qk_int8_pv_fp8_cuda", "sageattn_qk_int8_pv_fp8_cuda++", "sageattn3", "sageattn3_per_block_mean"]
def get_sage_func(sage_attention, allow_compile=False):
logging.info(f"Using sage attention mode: {sage_attention}")
from sageattention import sageattn
if sage_attention == "auto":
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD"):
return sageattn(q, k, v, is_causal=is_causal, attn_mask=attn_mask, tensor_layout=tensor_layout)
elif sage_attention == "sageattn_qk_int8_pv_fp16_cuda":
from sageattention import sageattn_qk_int8_pv_fp16_cuda
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD"):
return sageattn_qk_int8_pv_fp16_cuda(q, k, v, is_causal=is_causal, attn_mask=attn_mask, pv_accum_dtype="fp32", tensor_layout=tensor_layout)
elif sage_attention == "sageattn_qk_int8_pv_fp16_triton":
from sageattention import sageattn_qk_int8_pv_fp16_triton
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD"):
return sageattn_qk_int8_pv_fp16_triton(q, k, v, is_causal=is_causal, attn_mask=attn_mask, tensor_layout=tensor_layout)
elif sage_attention == "sageattn_qk_int8_pv_fp8_cuda":
from sageattention import sageattn_qk_int8_pv_fp8_cuda
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD"):
return sageattn_qk_int8_pv_fp8_cuda(q, k, v, is_causal=is_causal, attn_mask=attn_mask, pv_accum_dtype="fp32+fp32", tensor_layout=tensor_layout)
elif sage_attention == "sageattn_qk_int8_pv_fp8_cuda++":
from sageattention import sageattn_qk_int8_pv_fp8_cuda
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD"):
return sageattn_qk_int8_pv_fp8_cuda(q, k, v, is_causal=is_causal, attn_mask=attn_mask, pv_accum_dtype="fp32+fp16", tensor_layout=tensor_layout)
elif "sageattn3" in sage_attention:
from sageattn3 import sageattn3_blackwell
def sage_func(q, k, v, is_causal=False, attn_mask=None, tensor_layout="NHD", **kwargs):
q, k, v = [x.transpose(1, 2) if tensor_layout == "NHD" else x for x in (q, k, v)]
out = sageattn3_blackwell(q, k, v, is_causal=is_causal, attn_mask=attn_mask, per_block_mean=(sage_attention == "sageattn3_per_block_mean"))
return out.transpose(1, 2) if tensor_layout == "NHD" else out
if not allow_compile:
sage_func = torch.compiler.disable()(sage_func)
@wrap_attn
def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
if kwargs.get("low_precision_attention", True) is False:
return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
in_dtype = v.dtype
if q.dtype == torch.float32 or k.dtype == torch.float32 or v.dtype == torch.float32:
q, k, v = q.to(torch.float16), k.to(torch.float16), v.to(torch.float16)
if skip_reshape:
b, _, _, dim_head = q.shape
tensor_layout="HND"
else:
b, _, dim_head = q.shape
dim_head //= heads
q, k, v = map(
lambda t: t.view(b, -1, heads, dim_head),
(q, k, v),
)
tensor_layout="NHD"
if mask is not None:
# add a batch dimension if there isn't already one
if mask.ndim == 2:
mask = mask.unsqueeze(0)
# add a heads dimension if there isn't already one
if mask.ndim == 3:
mask = mask.unsqueeze(1)
out = sage_func(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout).to(in_dtype)
if tensor_layout == "HND":
if not skip_output_reshape:
out = (
out.transpose(1, 2).reshape(b, -1, heads * dim_head)
)
else:
if skip_output_reshape:
out = out.transpose(1, 2)
else:
out = out.reshape(b, -1, heads * dim_head)
return out
return attention_sage
from comfy.patcher_extension import CallbacksMP
class PathchSageAttentionKJ():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"sage_attention": (sageattn_modes, {"default": False, "tooltip": "Global patch comfy attention to use sageattn, once patched to revert back to normal you would need to run this node again with disabled option."}),
},
"optional": {
"allow_compile": ("BOOLEAN", {"default": False, "tooltip": "Allow the use of torch.compile for the sage attention function, requires latest sageattn 2.2.0 or higher."})
}
}
RETURN_TYPES = ("MODEL", )
FUNCTION = "patch"
DESCRIPTION = "Experimental node for patching attention mode. This doesn't use the model patching system and thus can't be disabled without running the node again with 'disabled' option."
EXPERIMENTAL = True
CATEGORY = "KJNodes/experimental"
def patch(self, model, sage_attention, allow_compile=False):
if sage_attention == "disabled":
return model,
model_clone = model.clone()
new_attention = get_sage_func(sage_attention, allow_compile=allow_compile)
def attention_override_sage(func, *args, **kwargs):
return new_attention.__wrapped__(*args, **kwargs)
# attention override
model_clone.model_options["transformer_options"]["optimized_attention_override"] = attention_override_sage
return model_clone,
class CheckpointLoaderKJ():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"ckpt_name": (folder_paths.get_filename_list("checkpoints"), {"tooltip": "The name of the checkpoint (model) to load."}),
"weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2", "fp16", "bf16", "fp32"],),
"compute_dtype": (["default", "fp16", "bf16", "fp32"], {"default": "default", "tooltip": "The compute dtype to use for the model."}),
"patch_cublaslinear": ("BOOLEAN", {"default": False, "tooltip": "Enable or disable the cublas_ops arg"}),
"sage_attention": (sageattn_modes, {"default": False, "tooltip": "Patch comfy attention to use sageattn."}),
"enable_fp16_accumulation": ("BOOLEAN", {"default": False, "tooltip": "Enable torch.backends.cuda.matmul.allow_fp16_accumulation, required minimum pytorch version 2.7.1"}),
}}
RETURN_TYPES = ("MODEL", "CLIP", "VAE")
FUNCTION = "load"
DESCRIPTION = "Experimental node for patching torch.nn.Linear with CublasLinear."
EXPERIMENTAL = True
CATEGORY = "KJNodes/model_loaders"
def load(self, ckpt_name, weight_dtype, compute_dtype, patch_cublaslinear, sage_attention, enable_fp16_accumulation):
DTYPE_MAP = {
"fp8_e4m3fn": torch.float8_e4m3fn,
"fp8_e5m2": torch.float8_e5m2,
"fp16": torch.float16,
"bf16": torch.bfloat16,
"fp32": torch.float32
}
model_options = {}
if dtype := DTYPE_MAP.get(weight_dtype):
model_options["dtype"] = dtype
logging.info(f"Setting {ckpt_name} weight dtype to {dtype}")
if weight_dtype == "fp8_e4m3fn_fast":
model_options["dtype"] = torch.float8_e4m3fn
model_options["fp8_optimizations"] = True
if patch_cublaslinear:
args.fast.add("cublas_ops")
else:
args.fast.discard("cublas_ops")
ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True)
model, clip, vae, _ = comfy.sd.load_state_dict_guess_config(
sd,
output_vae=True,
output_clip=True,
embedding_directory=folder_paths.get_folder_paths("embeddings"),
metadata=metadata,
model_options=model_options)
if dtype := DTYPE_MAP.get(compute_dtype):
model.set_model_compute_dtype(dtype)
model.force_cast_weights = False
logging.info(f"Setting {ckpt_name} compute dtype to {dtype}")
if enable_fp16_accumulation:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = True
else:
raise RuntimeError("Failed to set fp16 accumulation, requires pytorch version 2.7.1 or higher")
else:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = False
if sage_attention != "disabled":
new_attention = get_sage_func(sage_attention)
def attention_override_sage(func, *args, **kwargs):
return new_attention.__wrapped__(*args, **kwargs)
# attention override
model.model_options["transformer_options"]["optimized_attention_override"] = attention_override_sage
return model, clip, vae
class DiffusionModelSelector():
@classmethod
def INPUT_TYPES(s):
ltx2_connector_models = folder_paths.get_filename_list("text_encoders")
ltx2_connector_models = [m for m in ltx2_connector_models if "connector" in m.lower()]
return {"required": {
"model_name": (folder_paths.get_filename_list("diffusion_models") + ltx2_connector_models, {"tooltip": "The name of the checkpoint (model) to load."}),
},
}
RETURN_TYPES = ("STRING",)
RETURN_NAMES = ("model_path",)
FUNCTION = "get_path"
DESCRIPTION = "Returns the path to the model as a string."
EXPERIMENTAL = True
CATEGORY = "KJNodes/model_loaders"
def get_path(self, model_name):
if "connector" in model_name.lower():
model_path = folder_paths.get_full_path_or_raise("text_encoders", model_name)
else:
model_path = folder_paths.get_full_path_or_raise("diffusion_models", model_name)
return (model_path,)
class DiffusionModelLoaderKJ():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model_name": (folder_paths.get_filename_list("diffusion_models"), {"tooltip": "The name of the checkpoint (model) to load."}),
"weight_dtype": (["default", "fp8_e4m3fn", "fp8_e4m3fn_fast", "fp8_e5m2", "fp16", "bf16", "fp32"],),
"compute_dtype": (["default", "fp16", "bf16", "fp32"], {"default": "default", "tooltip": "The compute dtype to use for the model."}),
"patch_cublaslinear": ("BOOLEAN", {"default": False, "tooltip": "Enable or disable the cublas_ops arg"}),
"sage_attention": (sageattn_modes, {"default": False, "tooltip": "Patch comfy attention to use sageattn."}),
"enable_fp16_accumulation": ("BOOLEAN", {"default": False, "tooltip": "Enable torch.backends.cuda.matmul.allow_fp16_accumulation, requires pytorch 2.7.0 nightly."}),
},
"optional": {
"extra_state_dict": ("STRING", {"forceInput": True, "tooltip": "The full path to an additional state dict to load, this will be merged with the main state dict. Useful for example to add VACE module to a WanVideoModel. You can use DiffusionModelSelector to easily get the path."}),
}
}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch_and_load"
DESCRIPTION = "Node for patching torch.nn.Linear with CublasLinear."
EXPERIMENTAL = True
CATEGORY = "KJNodes/model_loaders"
def patch_and_load(self, model_name, weight_dtype, compute_dtype, patch_cublaslinear, sage_attention, enable_fp16_accumulation, extra_state_dict=None):
DTYPE_MAP = {
"fp8_e4m3fn": torch.float8_e4m3fn,
"fp8_e5m2": torch.float8_e5m2,
"fp16": torch.float16,
"bf16": torch.bfloat16,
"fp32": torch.float32
}
model_options = {}
if dtype := DTYPE_MAP.get(weight_dtype):
model_options["dtype"] = dtype
logging.info(f"Setting {model_name} weight dtype to {dtype}")
if weight_dtype == "fp8_e4m3fn_fast":
model_options["dtype"] = torch.float8_e4m3fn
model_options["fp8_optimizations"] = True
if enable_fp16_accumulation:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = True
else:
raise RuntimeError("Failed to set fp16 accumulation, this requires pytorch 2.7.1 or higher")
else:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = False
if patch_cublaslinear:
args.fast.add("cublas_ops")
else:
args.fast.discard("cublas_ops")
unet_path = folder_paths.get_full_path_or_raise("diffusion_models", model_name)
sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
if extra_state_dict is not None:
extra_sd = comfy.utils.load_torch_file(extra_state_dict)
sd.update(extra_sd)
del extra_sd
diffusion_model_prefix = comfy.sd.model_detection.unet_prefix_from_state_dict(sd)
sd = comfy.utils.state_dict_prefix_replace(sd, {diffusion_model_prefix: ""}, filter_keys=False)
model = comfy.sd.load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata)
if dtype := DTYPE_MAP.get(compute_dtype):
model.set_model_compute_dtype(dtype)
model.force_cast_weights = False
logging.info(f"Setting {model_name} compute dtype to {dtype}")
if sage_attention != "disabled":
new_attention = get_sage_func(sage_attention)
def attention_override_sage(func, *args, **kwargs):
return new_attention.__wrapped__(*args, **kwargs)
# attention override
model.model_options["transformer_options"]["optimized_attention_override"] = attention_override_sage
return (model,)
class ModelPatchTorchSettings:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"enable_fp16_accumulation": ("BOOLEAN", {"default": False, "tooltip": "Enable torch.backends.cuda.matmul.allow_fp16_accumulation, requires pytorch 2.7.0 nightly."}),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
DESCRIPTION = "Adds callbacks to model to set torch settings before and after running the model."
EXPERIMENTAL = True
CATEGORY = "KJNodes/experimental"
def patch(self, model, enable_fp16_accumulation):
model_clone = model.clone()
def patch_enable_fp16_accum(model):
logging.info("Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True")
torch.backends.cuda.matmul.allow_fp16_accumulation = True
def patch_disable_fp16_accum(model):
logging.info("Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False")
torch.backends.cuda.matmul.allow_fp16_accumulation = False
if enable_fp16_accumulation:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
model_clone.add_callback(CallbacksMP.ON_PRE_RUN, patch_enable_fp16_accum)
model_clone.add_callback(CallbacksMP.ON_CLEANUP, patch_disable_fp16_accum)
else:
raise RuntimeError("Failed to set fp16 accumulation, this requires pytorch 2.7.1 or higher")
else:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
model_clone.add_callback(CallbacksMP.ON_PRE_RUN, patch_disable_fp16_accum)
else:
raise RuntimeError("Failed to set fp16 accumulation, this requires pytorch 2.7.1 or higher")
return (model_clone,)
class PatchModelPatcherOrder:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"patch_order": (["object_patch_first", "weight_patch_first"], {"default": "weight_patch_first", "tooltip": "Patch the comfy patch_model function to load weight patches (LoRAs) before compiling the model"}),
"full_load": (["enabled", "disabled", "auto"], {"default": "auto", "tooltip": "Disabling may help with memory issues when loading large models, when changing this you should probably force model reload to avoid issues!"}),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
CATEGORY = "KJNodes/deprecated"
DESCRIPTION = "NO LONGER NECESSARY OR FUNCTIONAL, keeping node for backwards compatibility. Use the TorchCompileModelAdvanced to use LoRA with torch.compile."
DEPRECATED = True
def patch(self, model, patch_order, full_load):
return model,
class TorchCompileModelFluxAdvancedV2:
def __init__(self):
self._compiled = False
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"backend": (["inductor", "cudagraphs"],),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"double_blocks": ("BOOLEAN", {"default": True, "tooltip": "Compile double blocks"}),
"single_blocks": ("BOOLEAN", {"default": True, "tooltip": "Compile single blocks"}),
"dynamic": ("BOOLEAN", {"default": False, "tooltip": "Enable dynamic mode"}),
},
"optional": {
"dynamo_cache_size_limit": ("INT", {"default": 64, "min": 0, "max": 1024, "step": 1, "tooltip": "torch._dynamo.config.cache_size_limit"}),
"force_parameter_static_shapes": ("BOOLEAN", {"default": True, "tooltip": "torch._dynamo.config.force_parameter_static_shapes"}),
}
}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
CATEGORY = "KJNodes/torchcompile"
EXPERIMENTAL = True
DEPRECATED = True
DESCRIPTION = "Deprecated, use TorchCompileModelAdvanced instead."
def patch(self, model, backend, mode, fullgraph, single_blocks, double_blocks, dynamic, dynamo_cache_size_limit=64, force_parameter_static_shapes=True):
from comfy_api.torch_helpers import set_torch_compile_wrapper
m = model.clone()
diffusion_model = m.get_model_object("diffusion_model")
torch._dynamo.config.cache_size_limit = dynamo_cache_size_limit
torch._dynamo.config.force_parameter_static_shapes = force_parameter_static_shapes
compile_key_list = []
try:
if double_blocks:
for i, block in enumerate(diffusion_model.double_blocks):
print("Adding double block to compile list", i)
compile_key_list.append(f"diffusion_model.double_blocks.{i}")
if single_blocks:
for i, block in enumerate(diffusion_model.single_blocks):
compile_key_list.append(f"diffusion_model.single_blocks.{i}")
set_torch_compile_wrapper(model=m, keys=compile_key_list, backend=backend, mode=mode, dynamic=dynamic, fullgraph=fullgraph)
except:
raise RuntimeError("Failed to compile model")
return (m, )
class TorchCompileModelWanVideoV2:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"backend": (["inductor","cudagraphs"], {"default": "inductor"}),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"dynamic": ("BOOLEAN", {"default": False, "tooltip": "Enable dynamic mode"}),
"compile_transformer_blocks_only": ("BOOLEAN", {"default": True, "tooltip": "Compile only transformer blocks, faster compile and less error prone"}),
"dynamo_cache_size_limit": ("INT", {"default": 64, "min": 0, "max": 1024, "step": 1, "tooltip": "torch._dynamo.config.cache_size_limit"}),
},
"optional": {
"force_parameter_static_shapes": ("BOOLEAN", {"default": True, "tooltip": "torch._dynamo.config.force_parameter_static_shapes"}),
},
}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
CATEGORY = "KJNodes/torchcompile"
EXPERIMENTAL = True
DEPRECATED = True
DESCRIPTION = "Deprecated, use TorchCompileModelAdvanced instead."
def patch(self, model, backend, fullgraph, mode, dynamic, dynamo_cache_size_limit, compile_transformer_blocks_only, force_parameter_static_shapes=True):
from comfy_api.torch_helpers import set_torch_compile_wrapper
m = model.clone()
diffusion_model = m.get_model_object("diffusion_model")
torch._dynamo.config.cache_size_limit = dynamo_cache_size_limit
torch._dynamo.config.force_parameter_static_shapes = force_parameter_static_shapes
try:
if compile_transformer_blocks_only:
compile_key_list = []
for i, block in enumerate(diffusion_model.blocks):
compile_key_list.append(f"diffusion_model.blocks.{i}")
else:
compile_key_list =["diffusion_model"]
set_torch_compile_wrapper(model=m, keys=compile_key_list, backend=backend, mode=mode, dynamic=dynamic, fullgraph=fullgraph)
except:
raise RuntimeError("Failed to compile model")
return (m, )
class TorchCompileModelAdvanced:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"backend": (["inductor","cudagraphs"], {"default": "inductor"}),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"dynamic": (
["auto", "true", "false"],
{"default": "auto", "tooltip": "Use dynamic shape tracing."},
),
"compile_transformer_blocks_only": ("BOOLEAN", {"default": True, "tooltip": "Compile only transformer blocks, faster compile and less error prone"}),
"dynamo_cache_size_limit": ("INT", {"default": 64, "min": 0, "max": 1024, "step": 1, "tooltip": "torch._dynamo.config.cache_size_limit"}),
"debug_compile_keys": ("BOOLEAN", {"default": False, "tooltip": "Print the compile keys used for torch.compile"}),
},
"optional": {
"disable_dynamic_vram": ("BOOLEAN", {"default": False, "tooltip": "Disable dynamic VRAM feature as it can cause issues with compile"}),
}
}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
CATEGORY = "KJNodes/torchcompile"
DESCRIPTION = "Advanced torch.compile patching for diffusion models."
EXPERIMENTAL = True
def patch(self, model, backend, fullgraph, mode, dynamic, dynamo_cache_size_limit, compile_transformer_blocks_only, debug_compile_keys, disable_dynamic_vram=False):
from comfy_api.torch_helpers import set_torch_compile_wrapper
if disable_dynamic_vram:
try:
m = model.clone(disable_dynamic=True)
except TypeError:
logging.warning("This ComfyUI version do not support disabling dynamic VRAM through a node. This may cause issues with torch.compile.")
m = model.clone()
else:
m = model.clone()
diffusion_model = m.get_model_object("diffusion_model")
torch._dynamo.config.cache_size_limit = dynamo_cache_size_limit
try:
if compile_transformer_blocks_only:
layer_types = ["double_blocks", "single_blocks", "layers", "transformer_blocks", "blocks", "visual_transformer_blocks", "text_transformer_blocks"]
compile_key_list = []
for layer_name in layer_types:
if hasattr(diffusion_model, layer_name):
blocks = getattr(diffusion_model, layer_name)
for i in range(len(blocks)):
compile_key_list.append(f"diffusion_model.{layer_name}.{i}")
if not compile_key_list:
logging.warning("No known transformer blocks found to compile, compiling entire diffusion model instead")
elif debug_compile_keys:
logging.info("TorchCompileModelAdvanced: Compile key list:")
for key in compile_key_list:
logging.info(f" - {key}")
if not compile_key_list:
compile_key_list =["diffusion_model"]
dynamic_kv = {"true": True, "false": False, "auto": None}
try:
dynamic = dynamic_kv[dynamic]
except KeyError:
raise ValueError(f"Invalid dynamic arg {dynamic}")
set_torch_compile_wrapper(model=m, keys=compile_key_list, backend=backend, mode=mode, dynamic=dynamic, fullgraph=fullgraph)
except:
raise RuntimeError("Failed to compile model")
return (m, )
class TorchCompileModelQwenImage:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"backend": (["inductor","cudagraphs"], {"default": "inductor"}),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"dynamic": ("BOOLEAN", {"default": False, "tooltip": "Enable dynamic mode"}),
"compile_transformer_blocks_only": ("BOOLEAN", {"default": True, "tooltip": "Compile only transformer blocks, faster compile and less error prone"}),
"dynamo_cache_size_limit": ("INT", {"default": 64, "min": 0, "max": 1024, "step": 1, "tooltip": "torch._dynamo.config.cache_size_limit"}),
},
}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
CATEGORY = "KJNodes/torchcompile"
EXPERIMENTAL = True
DEPRECATED = True
DESCRIPTION = "Deprecated, use TorchCompileModelAdvanced instead."
def patch(self, model, backend, fullgraph, mode, dynamic, dynamo_cache_size_limit, compile_transformer_blocks_only):
from comfy_api.torch_helpers import set_torch_compile_wrapper
m = model.clone()
diffusion_model = m.get_model_object("diffusion_model")
torch._dynamo.config.cache_size_limit = dynamo_cache_size_limit
try:
if compile_transformer_blocks_only:
compile_key_list = []
for i, block in enumerate(diffusion_model.transformer_blocks):
compile_key_list.append(f"diffusion_model.transformer_blocks.{i}")
else:
compile_key_list =["diffusion_model"]
set_torch_compile_wrapper(model=m, keys=compile_key_list, backend=backend, mode=mode, dynamic=dynamic, fullgraph=fullgraph)
except:
raise RuntimeError("Failed to compile model")
return (m, )
class TorchCompileVAE:
def __init__(self):
self._compiled_encoder = False
self._compiled_decoder = False
@classmethod
def INPUT_TYPES(s):
return {"required": {
"vae": ("VAE",),
"backend": (["inductor", "cudagraphs"],),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"compile_encoder": ("BOOLEAN", {"default": True, "tooltip": "Compile encoder"}),
"compile_decoder": ("BOOLEAN", {"default": True, "tooltip": "Compile decoder"}),
}}
RETURN_TYPES = ("VAE",)
FUNCTION = "compile"
CATEGORY = "KJNodes/torchcompile"
EXPERIMENTAL = True
def compile(self, vae, backend, mode, fullgraph, compile_encoder, compile_decoder):
if compile_encoder:
if not self._compiled_encoder:
encoder_name = "encoder"
if hasattr(vae.first_stage_model, "taesd_encoder"):
encoder_name = "taesd_encoder"
try:
setattr(
vae.first_stage_model,
encoder_name,
torch.compile(
getattr(vae.first_stage_model, encoder_name),
mode=mode,
fullgraph=fullgraph,
backend=backend,
),
)
self._compiled_encoder = True
except:
raise RuntimeError("Failed to compile model")
if compile_decoder:
if not self._compiled_decoder:
decoder_name = "decoder"
if hasattr(vae.first_stage_model, "taesd_decoder"):
decoder_name = "taesd_decoder"
try:
setattr(
vae.first_stage_model,
decoder_name,
torch.compile(
getattr(vae.first_stage_model, decoder_name),
mode=mode,
fullgraph=fullgraph,
backend=backend,
),
)
self._compiled_decoder = True
except:
raise RuntimeError("Failed to compile model")
return (vae, )
class TorchCompileControlNet:
def __init__(self):
self._compiled= False
@classmethod
def INPUT_TYPES(s):
return {"required": {
"controlnet": ("CONTROL_NET",),
"backend": (["inductor", "cudagraphs"],),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
}}
RETURN_TYPES = ("CONTROL_NET",)
FUNCTION = "compile"
CATEGORY = "KJNodes/torchcompile"
EXPERIMENTAL = True
def compile(self, controlnet, backend, mode, fullgraph):
if not self._compiled:
try:
# for i, block in enumerate(controlnet.control_model.double_blocks):
# print("Compiling controlnet double_block", i)
# controlnet.control_model.double_blocks[i] = torch.compile(block, mode=mode, fullgraph=fullgraph, backend=backend)
controlnet.control_model = torch.compile(controlnet.control_model, mode=mode, fullgraph=fullgraph, backend=backend)
self._compiled = True
except:
self._compiled = False
raise RuntimeError("Failed to compile model")
return (controlnet, )
#teacache
try:
from comfy.ldm.wan.model import sinusoidal_embedding_1d
except:
pass
from unittest.mock import patch
import numpy as np
def relative_l1_distance(last_tensor, current_tensor):
l1_distance = torch.abs(last_tensor - current_tensor).mean()
norm = torch.abs(last_tensor).mean()
relative_l1_distance = l1_distance / norm
return relative_l1_distance.to(torch.float32)
@torch.compiler.disable()
def tea_cache(self, x, e0, e, transformer_options):
#teacache for cond and uncond separately
rel_l1_thresh = transformer_options["rel_l1_thresh"]
is_cond = True if transformer_options["cond_or_uncond"] == [0] else False
should_calc = True
suffix = "cond" if is_cond else "uncond"
# Init cache dict if not exists
if not hasattr(self, 'teacache_state'):
self.teacache_state = {
'cond': {'accumulated_rel_l1_distance': 0, 'prev_input': None,
'teacache_skipped_steps': 0, 'previous_residual': None},
'uncond': {'accumulated_rel_l1_distance': 0, 'prev_input': None,
'teacache_skipped_steps': 0, 'previous_residual': None}
}
logging.info("\nTeaCache: Initialized")
cache = self.teacache_state[suffix]
if cache['prev_input'] is not None:
if transformer_options["coefficients"] == []:
temb_relative_l1 = relative_l1_distance(cache['prev_input'], e0)
curr_acc_dist = cache['accumulated_rel_l1_distance'] + temb_relative_l1
else:
rescale_func = np.poly1d(transformer_options["coefficients"])
curr_acc_dist = cache['accumulated_rel_l1_distance'] + rescale_func(((e-cache['prev_input']).abs().mean() / cache['prev_input'].abs().mean()).cpu().item())
try:
if curr_acc_dist < rel_l1_thresh:
should_calc = False
cache['accumulated_rel_l1_distance'] = curr_acc_dist
else:
should_calc = True
cache['accumulated_rel_l1_distance'] = 0
except:
should_calc = True
cache['accumulated_rel_l1_distance'] = 0
if transformer_options["coefficients"] == []:
cache['prev_input'] = e0.clone().detach()
else:
cache['prev_input'] = e.clone().detach()
if not should_calc:
x += cache['previous_residual'].to(x.device)
cache['teacache_skipped_steps'] += 1
#print(f"TeaCache: Skipping {suffix} step")
return should_calc, cache
def teacache_wanvideo_vace_forward_orig(self, x, t, context, vace_context, vace_strength, clip_fea=None, freqs=None, transformer_options={}, **kwargs):
# embeddings
x = self.patch_embedding(x.float()).to(x.dtype)
grid_sizes = x.shape[2:]
x = x.flatten(2).transpose(1, 2)
# time embeddings
e = self.time_embedding(
sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
e0 = self.time_projection(e).unflatten(1, (6, self.dim))
# context
context = self.text_embedding(context)
context_img_len = None
if clip_fea is not None:
if self.img_emb is not None:
context_clip = self.img_emb(clip_fea) # bs x 257 x dim
context = torch.concat([context_clip, context], dim=1)
context_img_len = clip_fea.shape[-2]
orig_shape = list(vace_context.shape)
vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:])
c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype)
c = c.flatten(2).transpose(1, 2)
c = list(c.split(orig_shape[0], dim=0))
if not transformer_options:
raise RuntimeError("Can't access transformer_options, this requires ComfyUI nightly version from Mar 14, 2025 or later")
teacache_enabled = transformer_options.get("teacache_enabled", False)
if not teacache_enabled:
should_calc = True
else:
should_calc, cache = tea_cache(self, x, e0, e, transformer_options)
if should_calc:
original_x = x.clone().detach()
patches_replace = transformer_options.get("patches_replace", {})
blocks_replace = patches_replace.get("dit", {})
for i, block in enumerate(self.blocks):
if ("double_block", i) in blocks_replace:
def block_wrap(args):
out = {}
out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
return out
out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap, "transformer_options": transformer_options})
x = out["img"]
else:
x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
ii = self.vace_layers_mapping.get(i, None)
if ii is not None:
for iii in range(len(c)):
c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=original_x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
x += c_skip * vace_strength[iii]
del c_skip
if teacache_enabled:
cache['previous_residual'] = (x - original_x).to(transformer_options["teacache_device"])
# head
x = self.head(x, e)
# unpatchify
x = self.unpatchify(x, grid_sizes)
return x
def teacache_wanvideo_forward_orig(self, x, t, context, clip_fea=None, freqs=None, transformer_options={}, **kwargs):
# embeddings
x = self.patch_embedding(x.float()).to(x.dtype)
grid_sizes = x.shape[2:]
x = x.flatten(2).transpose(1, 2)
# time embeddings
e = self.time_embedding(
sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
e0 = self.time_projection(e).unflatten(1, (6, self.dim))
# context
context = self.text_embedding(context)
context_img_len = None
if clip_fea is not None:
if self.img_emb is not None:
context_clip = self.img_emb(clip_fea) # bs x 257 x dim
context = torch.concat([context_clip, context], dim=1)
context_img_len = clip_fea.shape[-2]
teacache_enabled = transformer_options.get("teacache_enabled", False)
if not teacache_enabled:
should_calc = True
else:
should_calc, cache = tea_cache(self, x, e0, e, transformer_options)
if should_calc:
original_x = x.clone().detach()
patches_replace = transformer_options.get("patches_replace", {})
blocks_replace = patches_replace.get("dit", {})
for i, block in enumerate(self.blocks):
if ("double_block", i) in blocks_replace:
def block_wrap(args):
out = {}
out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
return out
out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap, "transformer_options": transformer_options})
x = out["img"]
else:
x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
if teacache_enabled:
cache['previous_residual'] = (x - original_x).to(transformer_options["teacache_device"])
# head
x = self.head(x, e)
# unpatchify
x = self.unpatchify(x, grid_sizes)
return x
class WanVideoTeaCacheKJ:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"rel_l1_thresh": ("FLOAT", {"default": 0.275, "min": 0.0, "max": 10.0, "step": 0.001, "tooltip": "Threshold for to determine when to apply the cache, compromise between speed and accuracy. When using coefficients a good value range is something between 0.2-0.4 for all but 1.3B model, which should be about 10 times smaller, same as when not using coefficients."}),
"start_percent": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1.0, "step": 0.01, "tooltip": "The start percentage of the steps to use with TeaCache."}),
"end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01, "tooltip": "The end percentage of the steps to use with TeaCache."}),
"cache_device": (["main_device", "offload_device"], {"default": "offload_device", "tooltip": "Device to cache to"}),
"coefficients": (["disabled", "1.3B", "14B", "i2v_480", "i2v_720"], {"default": "i2v_480", "tooltip": "Coefficients for rescaling the relative l1 distance, if disabled the threshold value should be about 10 times smaller than the value used with coefficients."}),
}
}
RETURN_TYPES = ("MODEL",)
RETURN_NAMES = ("model",)
FUNCTION = "patch_teacache"
CATEGORY = "KJNodes/deprecated"
DEPRECATED = True
DESCRIPTION = """
Patch WanVideo model to use TeaCache. Speeds up inference by caching the output and
applying it instead of doing the step. Best results are achieved by choosing the
appropriate coefficients for the model. Early steps should never be skipped, with too
aggressive values this can happen and the motion suffers. Starting later can help with that too.
When NOT using coefficients, the threshold value should be
about 10 times smaller than the value used with coefficients.
Official recommended values https://github.com/ali-vilab/TeaCache/tree/main/TeaCache4Wan2.1
"""
EXPERIMENTAL = True
def patch_teacache(self, model, rel_l1_thresh, start_percent, end_percent, cache_device, coefficients):
if rel_l1_thresh == 0:
return (model,)
if coefficients == "disabled" and rel_l1_thresh > 0.1:
logging.warning("Threshold value is too high for TeaCache without coefficients, consider using coefficients for better results.")
if coefficients != "disabled" and rel_l1_thresh < 0.1 and "1.3B" not in coefficients:
logging.warning("Threshold value is too low for TeaCache with coefficients, consider using higher threshold value for better results.")
# type_str = str(type(model.model.model_config).__name__)
#if model.model.diffusion_model.dim == 1536:
# model_type ="1.3B"
# else:
# if "WAN21_T2V" in type_str:
# model_type = "14B"
# elif "WAN21_I2V" in type_str:
# model_type = "i2v_480"
# else:
# model_type = "i2v_720" #how to detect this?
teacache_coefficients_map = {
"disabled": [],
"1.3B": [2.39676752e+03, -1.31110545e+03, 2.01331979e+02, -8.29855975e+00, 1.37887774e-01],
"14B": [-5784.54975374, 5449.50911966, -1811.16591783, 256.27178429, -13.02252404],
"i2v_480": [-3.02331670e+02, 2.23948934e+02, -5.25463970e+01, 5.87348440e+00, -2.01973289e-01],
"i2v_720": [-114.36346466, 65.26524496, -18.82220707, 4.91518089, -0.23412683],
}
coefficients = teacache_coefficients_map[coefficients]
teacache_device = mm.get_torch_device() if cache_device == "main_device" else mm.unet_offload_device()
model_clone = model.clone()
if 'transformer_options' not in model_clone.model_options:
model_clone.model_options['transformer_options'] = {}
model_clone.model_options["transformer_options"]["rel_l1_thresh"] = rel_l1_thresh
model_clone.model_options["transformer_options"]["teacache_device"] = teacache_device
model_clone.model_options["transformer_options"]["coefficients"] = coefficients
diffusion_model = model_clone.get_model_object("diffusion_model")
def outer_wrapper(start_percent, end_percent):
def unet_wrapper_function(model_function, kwargs):
input = kwargs["input"]
timestep = kwargs["timestep"]
c = kwargs["c"]
sigmas = c["transformer_options"]["sample_sigmas"]
cond_or_uncond = kwargs["cond_or_uncond"]
last_step = (len(sigmas) - 1)
matched_step_index = (sigmas == timestep[0] ).nonzero()
if len(matched_step_index) > 0:
current_step_index = matched_step_index.item()
else:
for i in range(len(sigmas) - 1):
# walk from beginning of steps until crossing the timestep
if (sigmas[i] - timestep[0]) * (sigmas[i + 1] - timestep[0]) <= 0:
current_step_index = i
break
else:
current_step_index = 0
if current_step_index == 0:
if (len(cond_or_uncond) == 1 and cond_or_uncond[0] == 1) or len(cond_or_uncond) == 2:
if hasattr(diffusion_model, "teacache_state"):
delattr(diffusion_model, "teacache_state")
logging.info("\nResetting TeaCache state")
current_percent = current_step_index / (len(sigmas) - 1)
c["transformer_options"]["current_percent"] = current_percent
if start_percent <= current_percent <= end_percent:
c["transformer_options"]["teacache_enabled"] = True
forward_function = teacache_wanvideo_vace_forward_orig if hasattr(diffusion_model, "vace_layers") else teacache_wanvideo_forward_orig
context = patch.multiple(
diffusion_model,
forward_orig=forward_function.__get__(diffusion_model, diffusion_model.__class__)
)
with context:
out = model_function(input, timestep, **c)
if current_step_index+1 == last_step and hasattr(diffusion_model, "teacache_state"):
if len(cond_or_uncond) == 1 and cond_or_uncond[0] == 0:
skipped_steps_cond = diffusion_model.teacache_state["cond"]["teacache_skipped_steps"]
skipped_steps_uncond = diffusion_model.teacache_state["uncond"]["teacache_skipped_steps"]
logging.info("-----------------------------------")
logging.info(f"TeaCache skipped:")
logging.info(f"{skipped_steps_cond} cond steps")
logging.info(f"{skipped_steps_uncond} uncond step")
logging.info(f"out of {last_step} steps")
logging.info("-----------------------------------")
elif len(cond_or_uncond) == 2:
skipped_steps_cond = diffusion_model.teacache_state["uncond"]["teacache_skipped_steps"]
logging.info("-----------------------------------")
logging.info(f"TeaCache skipped:")
logging.info(f"{skipped_steps_cond} cond steps")
logging.info(f"out of {last_step} steps")
logging.info("-----------------------------------")
return out
return unet_wrapper_function
model_clone.set_model_unet_function_wrapper(outer_wrapper(start_percent=start_percent, end_percent=end_percent))
return (model_clone,)
from comfy.ldm.flux.math import apply_rope
def modified_wan_self_attention_forward(self, x, freqs, transformer_options={}):
r"""
Args:
x(Tensor): Shape [B, L, num_heads, C / num_heads]
freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
"""
b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
# query, key, value function
def qkv_fn(x):
q = self.norm_q(self.q(x)).view(b, s, n, d)
k = self.norm_k(self.k(x)).view(b, s, n, d)
v = self.v(x).view(b, s, n * d)
return q, k, v
q, k, v = qkv_fn(x)
q, k = apply_rope(q, k, freqs)
feta_scores = get_feta_scores(q, k, self.num_frames, self.enhance_weight)
try:
x = comfy.ldm.modules.attention.optimized_attention(
q.view(b, s, n * d),
k.view(b, s, n * d),
v,
heads=self.num_heads,
transformer_options=transformer_options,
)
except:
# backward compatibility for now
x = comfy.ldm.modules.attention.attention(
q.view(b, s, n * d),
k.view(b, s, n * d),
v,
heads=self.num_heads,
)
x = self.o(x)
x *= feta_scores
return x
from einops import rearrange
def get_feta_scores(query, key, num_frames, enhance_weight, num_heads=12):
img_q, img_k = query, key #torch.Size([2, 9216, 12, 128])
if img_q.ndim == 4:
B, ST, num_heads, head_dim = img_q.shape
elif img_q.ndim == 3:
B, ST, hidden_dim = img_q.shape
head_dim = hidden_dim // num_heads
# Reshape from [B, ST, hidden_dim] to [B, ST, num_heads, head_dim]
img_q = img_q.view(B, ST, num_heads, head_dim)
img_k = img_k.view(B, ST, num_heads, head_dim)
spatial_dim = ST // num_frames
query_image = rearrange(
img_q, "B (T S) N C -> (B S) N T C", T=num_frames, S=spatial_dim, N=num_heads, C=head_dim
)
key_image = rearrange(
img_k, "B (T S) N C -> (B S) N T C", T=num_frames, S=spatial_dim, N=num_heads, C=head_dim
)
return feta_score(query_image, key_image, head_dim, num_frames, enhance_weight)
def feta_score(query_image, key_image, head_dim, num_frames, enhance_weight):
scale = head_dim**-0.5
query_image = query_image * scale
attn_temp = query_image @ key_image.transpose(-2, -1) # translate attn to float32
attn_temp = attn_temp.to(torch.float32)
attn_temp = attn_temp.softmax(dim=-1)
# Reshape to [batch_size * num_tokens, num_frames, num_frames]
attn_temp = attn_temp.reshape(-1, num_frames, num_frames)
# Create a mask for diagonal elements
diag_mask = torch.eye(num_frames, device=attn_temp.device).bool()
diag_mask = diag_mask.unsqueeze(0).expand(attn_temp.shape[0], -1, -1)
# Zero out diagonal elements
attn_wo_diag = attn_temp.masked_fill(diag_mask, 0)
# Calculate mean for each token's attention matrix
# Number of off-diagonal elements per matrix is n*n - n
num_off_diag = num_frames * num_frames - num_frames
mean_scores = attn_wo_diag.sum(dim=(1, 2)) / num_off_diag
enhance_scores = mean_scores.mean() * (num_frames + enhance_weight)
enhance_scores = enhance_scores.clamp(min=1)
return enhance_scores
import types
class WanAttentionPatch:
def __init__(self, num_frames, weight):
self.num_frames = num_frames
self.enhance_weight = weight
def __get__(self, obj, objtype=None):
# Create bound method with stored parameters
def wrapped_attention(self_module, *args, **kwargs):
self_module.num_frames = self.num_frames
self_module.enhance_weight = self.enhance_weight
return modified_wan_self_attention_forward(self_module, *args, **kwargs)
return types.MethodType(wrapped_attention, obj)
class WanVideoEnhanceAVideoKJ:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"latent": ("LATENT", {"tooltip": "Only used to get the latent count"}),
"weight": ("FLOAT", {"default": 2.0, "min": 0.0, "max": 10.0, "step": 0.001, "tooltip": "Strength of the enhance effect"}),
}
}
RETURN_TYPES = ("MODEL",)
RETURN_NAMES = ("model",)
FUNCTION = "enhance"
CATEGORY = "KJNodes/wan"
DESCRIPTION = "https://github.com/NUS-HPC-AI-Lab/Enhance-A-Video"
EXPERIMENTAL = True
def enhance(self, model, weight, latent):
if weight == 0:
return (model,)
num_frames = latent["samples"].shape[2]
model_clone = model.clone()
if 'transformer_options' not in model_clone.model_options:
model_clone.model_options['transformer_options'] = {}
model_clone.model_options["transformer_options"]["enhance_weight"] = weight
diffusion_model = model_clone.get_model_object("diffusion_model")
compile_settings = getattr(model.model, "compile_settings", None)
for idx, block in enumerate(diffusion_model.blocks):
patched_attn = WanAttentionPatch(num_frames, weight).__get__(block.self_attn, block.__class__)
if compile_settings is not None:
patched_attn = torch.compile(patched_attn, mode=compile_settings["mode"], dynamic=compile_settings["dynamic"], fullgraph=compile_settings["fullgraph"], backend=compile_settings["backend"])
model_clone.add_object_patch(f"diffusion_model.blocks.{idx}.self_attn.forward", patched_attn)
return (model_clone,)
try:
from comfy.ldm.lightricks.model import apply_rotary_emb
except:
apply_rotary_emb = None
def ltxv_feta_forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}):
q = self.to_q(x)
context = x if context is None else context
k = self.to_k(context)
v = self.to_v(context)
q = self.q_norm(q)
k = self.k_norm(k)
if pe is not None:
q = apply_rotary_emb(q, pe)
k = apply_rotary_emb(k, pe if k_pe is None else k_pe)
feta_scores = get_feta_scores(q, k, self.num_frames, self.enhance_weight, self.heads)
if mask is None:
out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
else:
out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options)
if self.to_gate_logits is not None:
gate_logits = self.to_gate_logits(x) # (B, T, H)
b, t, _ = out.shape
out = out.view(b, t, self.heads, self.dim_head)
gates = 2.0 * torch.sigmoid(gate_logits) # zero-init -> identity
out = out * gates.unsqueeze(-1)
out = out.view(b, t, self.heads * self.dim_head)
return self.to_out(out) * feta_scores
class LTXCrossAttentionPatch:
def __init__(self, num_frames, weight):
self.num_frames = num_frames
self.enhance_weight = weight
def __get__(self, obj, objtype=None):
# Create bound method with stored parameters
def wrapped_attention(self_module, *args, **kwargs):
self_module.num_frames = self.num_frames
self_module.enhance_weight = self.enhance_weight
return ltxv_feta_forward(self_module, *args, **kwargs)
return types.MethodType(wrapped_attention, obj)
class LTXVEnhanceAVideoKJ:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"latent": ("LATENT", {"tooltip": "Only used to get the latent count"}),
"weight": ("FLOAT", {"default": 4.0, "min": 0.0, "max": 100.0, "step": 0.001, "tooltip": "Strength of the enhance effect"}),
}
}
RETURN_TYPES = ("MODEL",)
RETURN_NAMES = ("model",)
FUNCTION = "enhance"
CATEGORY = "KJNodes/ltxv"
DESCRIPTION = "https://github.com/NUS-HPC-AI-Lab/Enhance-A-Video"
EXPERIMENTAL = True
def enhance(self, model, weight, latent):
if weight == 0:
return (model,)
num_frames = latent["samples"].shape[2]
model_clone = model.clone()
if 'transformer_options' not in model_clone.model_options:
model_clone.model_options['transformer_options'] = {}
model_clone.model_options["transformer_options"]["enhance_weight"] = weight
diffusion_model = model_clone.get_model_object("diffusion_model")
for idx, block in enumerate(diffusion_model.transformer_blocks):
patched_attn1 = LTXCrossAttentionPatch(num_frames, weight).__get__(block.attn1, block.__class__)
model_clone.add_object_patch(f"diffusion_model.transformer_blocks.{idx}.attn1.forward", patched_attn1)
return (model_clone,)
def normalized_attention_guidance(self, query, context_positive, context_negative, transformer_options={}):
k_positive = self.norm_k(self.k(context_positive))
v_positive = self.v(context_positive)
k_negative = self.norm_k(self.k(context_negative))
v_negative = self.v(context_negative)
try:
x_positive = comfy.ldm.modules.attention.optimized_attention(query, k_positive, v_positive, heads=self.num_heads, transformer_options=transformer_options).flatten(2)
x_negative = comfy.ldm.modules.attention.optimized_attention(query, k_negative, v_negative, heads=self.num_heads, transformer_options=transformer_options).flatten(2)
except: #backwards compatibility for now
x_positive = comfy.ldm.modules.attention.optimized_attention(query, k_positive, v_positive, heads=self.num_heads).flatten(2)
x_negative = comfy.ldm.modules.attention.optimized_attention(query, k_negative, v_negative, heads=self.num_heads).flatten(2)
nag_guidance = x_positive * self.nag_scale - x_negative * (self.nag_scale - 1)
norm_positive = torch.norm(x_positive, p=1, dim=-1, keepdim=True).expand_as(x_positive)
norm_guidance = torch.norm(nag_guidance, p=1, dim=-1, keepdim=True).expand_as(nag_guidance)
scale = torch.nan_to_num(norm_guidance / norm_positive, nan=10.0)
mask = scale > self.nag_tau
adjustment = (norm_positive * self.nag_tau) / (norm_guidance + 1e-7)
nag_guidance = torch.where(mask, nag_guidance * adjustment, nag_guidance)
x = nag_guidance * self.nag_alpha + x_positive * (1 - self.nag_alpha)
del nag_guidance
return x
#region NAG
def wan_crossattn_forward_nag(self, x, context, transformer_options={}, **kwargs):
r"""
Args:
x(Tensor): Shape [B, L1, C]
context(Tensor): Shape [B, L2, C]
"""
# Determine batch splitting and context handling
if self.input_type == "default":
# Single or [pos, neg] pair
if context.shape[0] == 1:
x_pos, context_pos = x, context
x_neg, context_neg = None, None
else:
x_pos, x_neg = torch.chunk(x, 2, dim=0)
context_pos, context_neg = torch.chunk(context, 2, dim=0)
elif self.input_type == "batch":
# Standard batch, no CFG
x_pos, context_pos = x, context
x_neg, context_neg = None, None
# Positive branch
q_pos = self.norm_q(self.q(x_pos))
nag_context = self.nag_context
if self.input_type == "batch":
nag_context = nag_context.repeat(x_pos.shape[0], 1, 1)
try:
x_pos_out = normalized_attention_guidance(self, q_pos, context_pos, nag_context, transformer_options=transformer_options)
except: #backwards compatibility for now
x_pos_out = normalized_attention_guidance(self, q_pos, context_pos, nag_context)
# Negative branch
if x_neg is not None and context_neg is not None:
q_neg = self.norm_q(self.q(x_neg))
k_neg = self.norm_k(self.k(context_neg))
v_neg = self.v(context_neg)
try:
x_neg_out = comfy.ldm.modules.attention.optimized_attention(q_neg, k_neg, v_neg, heads=self.num_heads, transformer_options=transformer_options)
except: #backwards compatibility for now
x_neg_out = comfy.ldm.modules.attention.optimized_attention(q_neg, k_neg, v_neg, heads=self.num_heads)
x = torch.cat([x_pos_out, x_neg_out], dim=0)
else:
x = x_pos_out
return self.o(x)
def wan_i2v_crossattn_forward_nag(self, x, context, context_img_len, transformer_options={}, **kwargs):
r"""
Args:
x(Tensor): Shape [B, L1, C]
context(Tensor): Shape [B, L2, C]
"""
context_img = context[:, :context_img_len]
context = context[:, context_img_len:]
q_img = self.norm_q(self.q(x))
k_img = self.norm_k_img(self.k_img(context_img))
v_img = self.v_img(context_img)
try:
img_x = comfy.ldm.modules.attention.optimized_attention(q_img, k_img, v_img, heads=self.num_heads, transformer_options=transformer_options)
except: #backwards compatibility for now
img_x = comfy.ldm.modules.attention.optimized_attention(q_img, k_img, v_img, heads=self.num_heads)
if context.shape[0] == 2:
x, x_real_negative = torch.chunk(x, 2, dim=0)
context_positive, context_negative = torch.chunk(context, 2, dim=0)
else:
context_positive = context
context_negative = None
q = self.norm_q(self.q(x))
x = normalized_attention_guidance(self, q, context_positive, self.nag_context, transformer_options=transformer_options)
if context_negative is not None:
q_real_negative = self.norm_q(self.q(x_real_negative))
k_real_negative = self.norm_k(self.k(context_negative))
v_real_negative = self.v(context_negative)
try:
x_real_negative = comfy.ldm.modules.attention.optimized_attention(q_real_negative, k_real_negative, v_real_negative, heads=self.num_heads, transformer_options=transformer_options)
except: #backwards compatibility for now
x_real_negative = comfy.ldm.modules.attention.optimized_attention(q_real_negative, k_real_negative, v_real_negative, heads=self.num_heads)
x = torch.cat([x, x_real_negative], dim=0)
# output
x = x + img_x
x = self.o(x)
return x
class WanCrossAttentionPatch:
def __init__(self, context, nag_scale, nag_alpha, nag_tau, i2v=False, input_type="default"):
self.nag_context = context
self.nag_scale = nag_scale
self.nag_alpha = nag_alpha
self.nag_tau = nag_tau
self.i2v = i2v
self.input_type = input_type
def __get__(self, obj, objtype=None):
# Create bound method with stored parameters
def wrapped_attention(self_module, *args, **kwargs):
self_module.nag_context = self.nag_context
self_module.nag_scale = self.nag_scale
self_module.nag_alpha = self.nag_alpha
self_module.nag_tau = self.nag_tau
self_module.input_type = self.input_type
if self.i2v:
return wan_i2v_crossattn_forward_nag(self_module, *args, **kwargs)
else:
return wan_crossattn_forward_nag(self_module, *args, **kwargs)
return types.MethodType(wrapped_attention, obj)
class WanVideoNAG:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"conditioning": ("CONDITIONING",),
"nag_scale": ("FLOAT", {"default": 11.0, "min": 0.0, "max": 100.0, "step": 0.001, "tooltip": "Strength of negative guidance effect"}),
"nag_alpha": ("FLOAT", {"default": 0.25, "min": 0.0, "max": 1.0, "step": 0.001, "tooltip": "Mixing coefficient in that controls the balance between the normalized guided representation and the original positive representation."}),
"nag_tau": ("FLOAT", {"default": 2.5, "min": 0.0, "max": 10.0, "step": 0.001, "tooltip": "Clipping threshold that controls how much the guided attention can deviate from the positive attention."}),
},
"optional": {
"input_type": (["default", "batch"], {"tooltip": "Type of the model input"}),
},
}
RETURN_TYPES = ("MODEL",)
RETURN_NAMES = ("model",)
FUNCTION = "patch"
CATEGORY = "KJNodes/wan"
DESCRIPTION = "https://github.com/ChenDarYen/Normalized-Attention-Guidance"
EXPERIMENTAL = True
def patch(self, model, conditioning, nag_scale, nag_alpha, nag_tau, input_type="default"):
if nag_scale == 0:
return (model,)
device = mm.get_torch_device()
dtype = mm.unet_dtype()
model_clone = model.clone()
diffusion_model = model_clone.get_model_object("diffusion_model")
diffusion_model.text_embedding.to(device)
context = diffusion_model.text_embedding(conditioning[0][0].to(device, dtype))
type_str = str(type(model.model.model_config).__name__)
i2v = True if "WAN21_I2V" in type_str else False
for idx, block in enumerate(diffusion_model.blocks):
patched_attn = WanCrossAttentionPatch(context, nag_scale, nag_alpha, nag_tau, i2v, input_type=input_type).__get__(block.cross_attn, block.__class__)
model_clone.add_object_patch(f"diffusion_model.blocks.{idx}.cross_attn.forward", patched_attn)
return (model_clone,)
class SkipLayerGuidanceWanVideo:
@classmethod
def INPUT_TYPES(s):
return {"required": {"model": ("MODEL", ),
"blocks": ("STRING", {"default": "10", "multiline": False}),
"start_percent": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.001}),
"end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001}),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "slg"
EXPERIMENTAL = True
DESCRIPTION = "Simplified skip layer guidance that only skips the uncond on selected blocks"
DEPRECATED = True
CATEGORY = "advanced/guidance"
def slg(self, model, start_percent, end_percent, blocks):
def skip(args, extra_args):
transformer_options = extra_args.get("transformer_options", {})
original_block = extra_args["original_block"]
if not transformer_options:
raise ValueError("transformer_options not found in extra_args, currently SkipLayerGuidanceWanVideo only works with TeaCacheKJ")
if start_percent <= transformer_options["current_percent"] <= end_percent:
if args["img"].shape[0] == 2:
prev_img_uncond = args["img"][0].unsqueeze(0)
new_args = {
"img": args["img"][1].unsqueeze(0),
"txt": args["txt"][1].unsqueeze(0),
"vec": args["vec"][1].unsqueeze(0),
"pe": args["pe"][1].unsqueeze(0)
}
block_out = original_block(new_args)
out = {
"img": torch.cat([prev_img_uncond, block_out["img"]], dim=0),
"txt": args["txt"],
"vec": args["vec"],
"pe": args["pe"]
}
else:
if transformer_options.get("cond_or_uncond") == [0]:
out = original_block(args)
else:
out = args
else:
out = original_block(args)
return out
block_list = [int(x.strip()) for x in blocks.split(",")]
blocks = [int(i) for i in block_list]
logging.info(f"Selected blocks to skip uncond on: {blocks}")
m = model.clone()
for b in blocks:
#m.set_model_patch_replace(skip, "dit", "double_block", b)
model_options = m.model_options["transformer_options"].copy()
if "patches_replace" not in model_options:
model_options["patches_replace"] = {}
else:
model_options["patches_replace"] = model_options["patches_replace"].copy()
if "dit" not in model_options["patches_replace"]:
model_options["patches_replace"]["dit"] = {}
else:
model_options["patches_replace"]["dit"] = model_options["patches_replace"]["dit"].copy()
block = ("double_block", b)
model_options["patches_replace"]["dit"][block] = skip
m.model_options["transformer_options"] = model_options
return (m, )
class CFGZeroStarAndInit:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"use_zero_init": ("BOOLEAN", {"default": True}),
"zero_init_steps": ("INT", {"default": 0, "min": 0, "tooltip": "for zero init, starts from 0 so first step is always zeroed out if use_zero_init enabled"}),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
DESCRIPTION = "https://github.com/WeichenFan/CFG-Zero-star"
CATEGORY = "KJNodes/experimental"
EXPERIMENTAL = True
def patch(self, model, use_zero_init, zero_init_steps):
def cfg_zerostar(args):
#zero init
cond = args["cond"]
timestep = args["timestep"]
sigmas = args["model_options"]["transformer_options"]["sample_sigmas"]
matched_step_index = (sigmas == timestep[0]).nonzero()
if len(matched_step_index) > 0:
current_step_index = matched_step_index.item()
else:
for i in range(len(sigmas) - 1):
if (sigmas[i] - timestep[0]) * (sigmas[i + 1] - timestep[0]) <= 0:
current_step_index = i
break
else:
current_step_index = 0
if (current_step_index <= zero_init_steps) and use_zero_init:
return cond * 0
uncond = args["uncond"]
cond_scale = args["cond_scale"]
batch_size = cond.shape[0]
positive_flat = cond.view(batch_size, -1)
negative_flat = uncond.view(batch_size, -1)
dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)
squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8
alpha = dot_product / squared_norm
alpha = alpha.view(batch_size, *([1] * (len(cond.shape) - 1)))
noise_pred = uncond * alpha + cond_scale * (cond - uncond * alpha)
return noise_pred
m = model.clone()
m.set_model_sampler_cfg_function(cfg_zerostar)
return (m, )
class GGUFLoaderKJ(io.ComfyNode):
@classmethod
def define_schema(cls):
# Get GGUF models safely, fallback to empty list if unet_gguf folder doesn't exist
try:
gguf_models = folder_paths.get_filename_list("unet_gguf")
ltx2_connector_models = folder_paths.get_filename_list("text_encoders")
ltx2_connector_models = [m for m in ltx2_connector_models if "connector" in m.lower()]
except KeyError:
gguf_models = []
ltx2_connector_models = []
return io.Schema(
node_id="GGUFLoaderKJ",
category="KJNodes/model_loaders",
description="Loads a GGUF model with advanced options, requires [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF) to be installed.",
is_experimental=True,
inputs=[
io.Combo.Input("model_name", options=gguf_models),
io.Combo.Input("extra_model_name", options=gguf_models + ltx2_connector_models + ["none"], default="none", tooltip="An extra gguf model to load and merge into the main model, for example VACE module"),
io.Combo.Input("dequant_dtype", options=["default", "target", "float32", "float16", "bfloat16"], default="default"),
io.Combo.Input("patch_dtype", options=["default", "target", "float32", "float16", "bfloat16"], default="default"),
io.Boolean.Input("patch_on_device", default=False),
io.Boolean.Input("enable_fp16_accumulation", default=False, tooltip="Enable torch.backends.cuda.matmul.allow_fp16_accumulation, required minimum pytorch version 2.7.1"),
io.Combo.Input("attention_override", options=["none", "sdpa", "sageattn", "xformers", "flashattn"], default="none", tooltip="Overrides the used attention implementation, requires the respective library to be installed"),
],
outputs=[io.Model.Output(),],
)
def attention_override_pytorch(func, *args, **kwargs):
new_attention = comfy.ldm.modules.attention.attention_pytorch
return new_attention.__wrapped__(*args, **kwargs)
def attention_override_sage(func, *args, **kwargs):
new_attention = comfy.ldm.modules.attention.attention_sage
return new_attention.__wrapped__(*args, **kwargs)
def attention_override_xformers(func, *args, **kwargs):
new_attention = comfy.ldm.modules.attention.attention_xformers
return new_attention.__wrapped__(*args, **kwargs)
def attention_override_flash(func, *args, **kwargs):
new_attention = comfy.ldm.modules.attention.attention_flash
return new_attention.__wrapped__(*args, **kwargs)
ATTENTION_OVERRIDES = {
"sdpa": attention_override_pytorch,
"sageattn": attention_override_sage,
"xformers": attention_override_xformers,
"flashattn": attention_override_flash,
}
@classmethod
def _get_gguf_module(cls):
"""Import GGUF module with version validation"""
for key, mod in sys.modules.items():
if key.endswith("ComfyUI-GGUF") or key.endswith("comfyui-gguf"):
if hasattr(mod, "ops") and hasattr(mod, "nodes"):
return mod
gguf_path = os.path.join(folder_paths.folder_names_and_paths["custom_nodes"][0][0], "ComfyUI-GGUF")
for module_name in ["ComfyUI-GGUF", "custom_nodes.ComfyUI-GGUF", "comfyui-gguf", "custom_nodes.comfyui-gguf", gguf_path, gguf_path.lower()]:
try:
module = importlib.import_module(module_name)
return module
except ImportError:
continue
raise ImportError(
"Compatible ComfyUI-GGUF not found. "
"Please install/update from: https://github.com/city96/ComfyUI-GGUF"
)
@classmethod
def execute(cls, model_name, extra_model_name, dequant_dtype, patch_dtype, patch_on_device, attention_override, enable_fp16_accumulation):
gguf_nodes = cls._get_gguf_module()
ops = gguf_nodes.ops.GGMLOps()
def set_linear_dtype(attr, value):
if value == "default":
setattr(ops.Linear, attr, None)
elif value == "target":
setattr(ops.Linear, attr, value)
else:
setattr(ops.Linear, attr, getattr(torch, value))
set_linear_dtype("dequant_dtype", dequant_dtype)
set_linear_dtype("patch_dtype", patch_dtype)
# init model
extra = {}
model_path = folder_paths.get_full_path("unet", model_name)
try:
sd, extra = gguf_nodes.loader.gguf_sd_loader(model_path)
except:
sd = gguf_nodes.loader.gguf_sd_loader(model_path)
if extra_model_name is not None and extra_model_name != "none":
if extra_model_name.endswith(".gguf"):
extra_model_full_path = folder_paths.get_full_path("unet", extra_model_name)
extra_model = gguf_nodes.loader.gguf_sd_loader(extra_model_full_path)
elif "connector" in extra_model_name.lower():
extra_model_full_path = folder_paths.get_full_path("text_encoders", extra_model_name)
extra_model = comfy.utils.load_torch_file(extra_model_full_path)
diffusion_model_prefix = comfy.model_detection.unet_prefix_from_state_dict(extra_model)
if diffusion_model_prefix == "model.diffusion_model.":
temp_sd = comfy.utils.state_dict_prefix_replace(extra_model, {diffusion_model_prefix: ""}, filter_keys=True)
if len(temp_sd) > 0:
extra_model = temp_sd
else:
raise ValueError("Extra model must also be a .gguf file")
sd.update(extra_model)
model = comfy.sd.load_diffusion_model_state_dict(
sd, model_options={"custom_operations": ops}, metadata=extra.get("metadata", {})
)
if model is None:
raise RuntimeError(f"ERROR: Could not detect model type of: {model_path}")
model = gguf_nodes.nodes.GGUFModelPatcher.clone(model)
model.patch_on_device = patch_on_device
# attention override
if attention_override in cls.ATTENTION_OVERRIDES:
model.model_options["transformer_options"]["optimized_attention_override"] = cls.ATTENTION_OVERRIDES[attention_override]
if enable_fp16_accumulation:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = True
else:
raise RuntimeError("Failed to set fp16 accumulation, requires pytorch version 2.7.1 or higher")
else:
if hasattr(torch.backends.cuda.matmul, "allow_fp16_accumulation"):
torch.backends.cuda.matmul.allow_fp16_accumulation = False
return io.NodeOutput(model,)
try:
from torch.nn.attention.flex_attention import flex_attention, BlockMask
except:
flex_attention = None
BlockMask = None
class NABLA_AttentionKJ():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"latent": ("LATENT", {"tooltip": "Only used to get the latent shape"}),
"window_time": ("INT", {"default": 11, "min": 1, "tooltip": "Temporal attention window size"}),
"window_width": ("INT", {"default": 3, "min": 1, "tooltip": "Spatial attention window size"}),
"window_height": ("INT", {"default": 3, "min": 1, "tooltip": "Spatial attention window size"}),
"sparsity": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 1.0, "step": 0.01}),
"torch_compile": ("BOOLEAN", {"default": True, "tooltip": "Most likely required for reasonable memory usage"})
},
}
RETURN_TYPES = ("MODEL", )
FUNCTION = "patch"
DESCRIPTION = "Experimental node for patching attention mode to use NABLA sparse attention for video models, currently only works with Kadinsky5"
CATEGORY = "KJNodes/experimental"
def patch(self, model, latent, window_time, window_width, window_height, sparsity, torch_compile):
if flex_attention is None or BlockMask is None:
raise RuntimeError("can't import flex_attention from torch.nn.attention, requires newer pytorch version")
model_clone = model.clone()
samples = latent["samples"]
sparse_params = get_sparse_params(samples, window_time, window_height, window_width, sparsity)
nabla_attention = NABLA_Attention(sparse_params)
def attention_override_nabla(func, *args, **kwargs):
return nabla_attention(*args, **kwargs)
if torch_compile:
attention_override_nabla = torch.compile(attention_override_nabla, mode="max-autotune-no-cudagraphs", dynamic=True)
# attention override
model_clone.model_options["transformer_options"]["optimized_attention_override"] = attention_override_nabla
return model_clone,
class NABLA_Attention():
def __init__(self, sparse_params):
self.sparse_params = sparse_params
def __call__(self, q, k, v, heads, **kwargs):
if q.shape[-2] < 3000 or k.shape[-2] < 3000:
return optimized_attention(q, k, v, heads, **kwargs)
block_mask = self.nablaT_v2(q, k, self.sparse_params["sta_mask"], thr=self.sparse_params["P"])
out = flex_attention(q, k, v, block_mask=block_mask).transpose(1, 2).contiguous().flatten(-2, -1)
return out
def nablaT_v2(self, q, k, sta, thr=0.9):
# Map estimation
BLOCK_SIZE = 64
B, h, S, D = q.shape
s1 = S // BLOCK_SIZE
qa = q.reshape(B, h, s1, BLOCK_SIZE, D).mean(-2)
ka = k.reshape(B, h, s1, BLOCK_SIZE, D).mean(-2).transpose(-2, -1)
map = qa @ ka
map = torch.softmax(map / math.sqrt(D), dim=-1)
# Map binarization
vals, inds = map.sort(-1)
cvals = vals.cumsum_(-1)
mask = (cvals >= 1 - thr).int()
mask = mask.gather(-1, inds.argsort(-1))
mask = torch.logical_or(mask, sta)
# BlockMask creation
kv_nb = mask.sum(-1).to(torch.int32)
kv_inds = mask.argsort(dim=-1, descending=True).to(torch.int32)
return BlockMask.from_kv_blocks(torch.zeros_like(kv_nb), kv_inds, kv_nb, kv_inds, BLOCK_SIZE=BLOCK_SIZE, mask_mod=None)
def fast_sta_nabla(T, H, W, wT=3, wH=3, wW=3):
l = torch.Tensor([T, H, W]).amax()
r = torch.arange(0, l, 1, dtype=torch.int16, device=mm.get_torch_device())
mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs()
sta_t, sta_h, sta_w = (
mat[:T, :T].flatten(),
mat[:H, :H].flatten(),
mat[:W, :W].flatten(),
)
sta_t = sta_t <= wT // 2
sta_h = sta_h <= wH // 2
sta_w = sta_w <= wW // 2
sta_hw = (sta_h.unsqueeze(1) * sta_w.unsqueeze(0)).reshape(H, H, W, W).transpose(1, 2).flatten()
sta = (sta_t.unsqueeze(1) * sta_hw.unsqueeze(0)).reshape(T, T, H * W, H * W).transpose(1, 2)
return sta.reshape(T * H * W, T * H * W)
def get_sparse_params(x, wT, wH, wW, sparsity=0.9):
B, C, T, H, W = x.shape
#print("x shape:", x.shape)
patch_size = (1, 2, 2)
T, H, W = (
T // patch_size[0],
H // patch_size[1],
W // patch_size[2],
)
sta_mask = fast_sta_nabla(T, H // 8, W // 8, wT, wH, wW)
sparse_params = {
"sta_mask": sta_mask.unsqueeze_(0).unsqueeze_(0),
"to_fractal": True,
"P": sparsity,
"wT": wT,
"wH": wH,
"wW": wW,
"add_sta": True,
"visual_shape": (T, H, W),
"method": "topcdf",
}
return sparse_params
from comfy.comfy_types.node_typing import IO
class StartRecordCUDAMemoryHistory():
# @classmethod
# def IS_CHANGED(s):
# return True
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"input": (IO.ANY,),
"enabled": (["all", "state", "None"], {"default": "all", "tooltip": "None: disable, 'state': keep info for allocated memory, 'all': keep history of all alloc/free calls"}),
"context": (["all", "state", "alloc", "None"], {"default": "all", "tooltip": "None: no tracebacks, 'state': tracebacks for allocated memory, 'alloc': for alloc calls, 'all': for free calls"}),
"stacks": (["python", "all"], {"default": "all", "tooltip": "'python': Python/TorchScript/inductor frames, 'all': also C++ frames"}),
"max_entries": ("INT", {"default": 100000, "min": 1000, "max": 10000000, "tooltip": "Maximum number of entries to record"}),
},
}
RETURN_TYPES = (IO.ANY, )
RETURN_NAMES = ("input", "output_path",)
FUNCTION = "start"
CATEGORY = "KJNodes/memory"
DESCRIPTION = "THIS NODE ALWAYS RUNS. Starts recording CUDA memory allocation history, can be ended and saved with EndRecordCUDAMemoryHistory. "
def start(self, input, enabled, context, stacks, max_entries):
mm.soft_empty_cache()
torch.cuda.reset_peak_memory_stats(mm.get_torch_device())
torch.cuda.memory._record_memory_history(
max_entries=max_entries,
enabled=enabled if enabled != "None" else None,
context=context if context != "None" else None,
stacks=stacks
)
return input,
class EndRecordCUDAMemoryHistory():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"input": (IO.ANY,),
"output_path": ("STRING", {"default": "comfy_cuda_memory_history"}, "Base path for saving the CUDA memory history file, timestamp and .pt extension will be added"),
},
}
RETURN_TYPES = (IO.ANY, "STRING",)
RETURN_NAMES = ("input", "output_path",)
FUNCTION = "end"
CATEGORY = "KJNodes/memory"
DESCRIPTION = "Records CUDA memory allocation history between start and end, saves to a file that can be analyzed here: https://docs.pytorch.org/memory_viz or with VisualizeCUDAMemoryHistory node"
def end(self, input, output_path):
mm.soft_empty_cache()
time = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
output_path = f"{output_path}{time}.pt"
torch.cuda.memory._dump_snapshot(output_path)
torch.cuda.memory._record_memory_history(enabled=None)
return input, output_path
try:
from server import PromptServer
except:
PromptServer = None
class VisualizeCUDAMemoryHistory():
@classmethod
def INPUT_TYPES(s):
return {"required": {
"snapshot_path": ("STRING", ),
},
"hidden": {
"unique_id": "UNIQUE_ID",
},
}
RETURN_TYPES = ("STRING",)
RETURN_NAMES = ("output_path",)
FUNCTION = "visualize"
CATEGORY = "KJNodes/memory"
DESCRIPTION = "Visualizes a CUDA memory allocation history file, opens in browser"
OUTPUT_NODE = True
def visualize(self, snapshot_path, unique_id):
import pickle
from torch.cuda import _memory_viz
import uuid
from folder_paths import get_output_directory
output_dir = get_output_directory()
with open(snapshot_path, "rb") as f:
snapshot = pickle.load(f)
html = _memory_viz.trace_plot(snapshot)
html_filename = f"cuda_memory_history_{uuid.uuid4().hex}.html"
output_path = os.path.join(output_dir, "memory_history", html_filename)
os.makedirs(os.path.dirname(output_path), exist_ok=True)
with open(output_path, "w", encoding="utf-8") as f:
f.write(html)
api_url = f"http://localhost:8188/api/view?type=output&filename={html_filename}&subfolder=memory_history"
# Progress UI
if unique_id and PromptServer is not None:
try:
PromptServer.instance.send_progress_text(
api_url,
unique_id
)
except:
pass
return api_url,
class ModelMemoryUseReportPatch:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
DESCRIPTION = "Adds callbacks to model to report memory usage during after sampling"
EXPERIMENTAL = True
CATEGORY = "KJNodes/memory"
def patch(self, model):
model_clone = model.clone()
device = mm.get_torch_device()
def reset_mem_usage(model):
torch.cuda.reset_peak_memory_stats(device)
def report_mem_usage(model):
max_memory = torch.cuda.max_memory_allocated(device) / 1024**3
max_reserved = torch.cuda.max_memory_reserved(device) / 1024**3
logging.info(f"Sampling max allocated memory: {max_memory=:.3f} GB")
logging.info(f"Sampling max reserved memory: {max_reserved=:.3f} GB")
model_clone.add_callback(CallbacksMP.ON_PRE_RUN, reset_mem_usage)
model_clone.add_callback(CallbacksMP.ON_CLEANUP, report_mem_usage)
return (model_clone,)
class MemoryUsageFactorAdjustWrapper:
def __init__(self, memory_usage_factor, original_factor):
self.memory_usage_factor = memory_usage_factor
self.original_factor = original_factor
def __call__(self, executor, model, noise_shape: torch.Tensor, *args, **kwargs):
m = model.clone()
m.model.memory_usage_factor = self.memory_usage_factor
logging.info(f"Temporarily set memory usage factor to {self.memory_usage_factor}")
try:
result = executor(m, noise_shape, *args, **kwargs)
finally:
logging.info(f"Model memory usage calculated, restoring original memory usage factor: {self.original_factor}")
m.model.memory_usage_factor = self.original_factor
return result
class ModelMemoryUsageFactorOverride:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"model": ("MODEL",),
"memory_usage_factor": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.001}),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
DESCRIPTION = "Overrides the memory usage factor of the model during sampling."
EXPERIMENTAL = True
CATEGORY = "KJNodes/memory"
def patch(self, model, memory_usage_factor):
model_clone = model.clone()
original_memory_usage_factor = model_clone.model.memory_usage_factor
logging.info(f"Original memory usage factor: {original_memory_usage_factor}")
wrapper = MemoryUsageFactorAdjustWrapper(memory_usage_factor, original_memory_usage_factor)
model_clone.add_wrapper_with_key(
comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING,
"memory_usage_factor_adjust_prepare_sampling",
wrapper
)
return (model_clone,)
def wan_ffn_chunked_forward(self, x):
if x.shape[1] > self.dim_threshold:
chunks = torch.chunk(x, self.num_chunks, dim=1)
output_chunks = []
for chunk in chunks:
output_chunks.append(torch.nn.Sequential.forward(self, chunk))
chunked = torch.cat(output_chunks, dim=1)
return chunked
else:
return torch.nn.Sequential.forward(self, x)
class WanffnChunkPatch:
def __init__(self, num_chunks, dim_threshold=4096):
self.num_chunks = num_chunks
self.dim_threshold = dim_threshold
def __get__(self, obj, objtype=None):
def wrapped_forward(self_module, *args, **kwargs):
self_module.num_chunks = self.num_chunks
self_module.dim_threshold = self.dim_threshold
return wan_ffn_chunked_forward(self_module, *args, **kwargs)
return types.MethodType(wrapped_forward, obj)
class WanChunkFeedForward(io.ComfyNode):
@classmethod
def define_schema(cls):
return io.Schema(
node_id="WanChunkFeedForward",
display_name="Wan Chunk FeedForward",
category="KJNodes/wan",
description="EXPERIMENTAL AND MAY CHANGE THE MODEL OUTPUT!! Chunks feedforward activations to reduce peak VRAM usage.",
is_experimental=True,
inputs=[
io.Model.Input("model"),
io.Int.Input("chunks", default=2, min=1, max=100, step=1, tooltip="Number of chunks to split the feedforward activations into to reduce peak VRAM usage."),
io.Int.Input("dim_threshold", default=4096, min=1024, max=16384, step=256, tooltip="Dimension threshold above which to apply chunking."),
],
outputs=[
io.Model.Output(display_name="model"),
],
)
@classmethod
def execute(cls, model, chunks, dim_threshold) -> io.NodeOutput:
if chunks == 1:
return io.NodeOutput(model)
model_clone = model.clone()
diffusion_model = model_clone.get_model_object("diffusion_model")
for idx, block in enumerate(diffusion_model.blocks):
patched_ffn = WanffnChunkPatch(chunks, dim_threshold).__get__(block.ffn, block.__class__)
model_clone.add_object_patch(f"diffusion_model.blocks.{idx}.ffn.forward", patched_ffn)
return io.NodeOutput(model_clone)
from comfy.samplers import KSAMPLER
from comfy.k_diffusion.sampling import to_d
def sample_selfrefinevideo(model, x, sigmas, stochastic_step_map, certain_percentage=0.999, uncertainty_threshold=0.25, extra_args=None, callback=None, disable=None, verbose=False, video_shape=None, seed=None):
extra_args = {} if extra_args is None else extra_args
sigma_in = x.new_ones([x.shape[0]])
if seed is not None:
generator = torch.Generator(torch.device("cpu")).manual_seed(seed)
pbar = tqdm(total=len(sigmas) - 1, disable=disable, desc="Sampling")
for i in range(len(sigmas) - 1):
# Get stochastic steps for this noise level
current_num_anneal_steps = stochastic_step_map.get(i, 0)
use_stochastic = current_num_anneal_steps > 0
m = current_num_anneal_steps + 1 if use_stochastic else 1
sigma, sigma_next = sigmas[i], sigmas[i + 1]
prev_certain_mask = None
prev_denoised = None
prev_denoised_full = None
prev_x_next = None
prev_x_next_video = None
is_certain = False
for ii in range(m):
if m > 1:
pbar.set_description(f"Step {i}/{len(sigmas)-1} (substep {ii+1}/{m})")
# Early exit if certain threshold reached
if is_certain:
x = prev_x_next
break
# Determine input
noise = torch.randn(x.shape, device=torch.device("cpu"), generator=generator).to(x)
x_in = x if ii == 0 else (1.0 - sigma) * prev_denoised_full + sigma * noise
if ii > 0:
x = x_in
denoised = model(x_in, sigmas[i] * sigma_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
# Compute next latents
d = to_d(x, sigma, denoised)
x_next = x + (sigma_next - sigma) * d
# Separate video and audio if joint model
if d.ndim == 3 and video_shape is not None:
cut = math.prod(video_shape[1:])
denoised_video = denoised[:, :, :cut].reshape([denoised.shape[0]] + list(video_shape)[1:])
x_next_video = x_next[:, :, :cut].reshape([denoised.shape[0]] + list(video_shape)[1:])
denoised_audio = denoised[:, :, cut:]
x_next_audio = x_next[:, :, cut:]
if verbose:
tqdm.write(f"Video shape: {denoised_video.shape}, Audio shape: {denoised_audio.shape}")
else:
denoised_video = denoised
x_next_video = x_next
denoised_audio = None
x_next_audio = None
# Stochastic sampling with uncertainty masking
if use_stochastic and prev_denoised is not None:
# Compute uncertainty and masking on video part
diff = denoised_video - prev_denoised
uncertainty = torch.sqrt(torch.sum(diff ** 2, dim=1)) / denoised_video.shape[1]
certain_mask = uncertainty < uncertainty_threshold
if verbose:
tqdm.write(f"Step {i}/{len(sigmas)-1} substep {ii+1}/{m}:")
tqdm.write(f"Uncertainty: min {uncertainty.min():.4f}, max {uncertainty.max():.4f}, threshold {uncertainty_threshold}")
tqdm.write(f"Certain pixels: {certain_mask.sum()}/{certain_mask.numel()} = {certain_mask.sum()/certain_mask.numel():.4f}")
# Update certain mask (union with previous)
if prev_certain_mask is not None:
certain_mask = certain_mask | prev_certain_mask
# Check certainty threshold
if certain_mask.sum() / certain_mask.numel() > certain_percentage:
is_certain = True
if verbose:
tqdm.write(f"{ii}/{current_num_anneal_steps}: Certain region is more than {certain_percentage}, we are certain")
# Apply masking to video
certain_mask_float = certain_mask.float().unsqueeze(1)
x_next_video = certain_mask_float * prev_x_next_video + (1.0 - certain_mask_float) * x_next_video
denoised_video = certain_mask_float * prev_denoised + (1.0 - certain_mask_float) * denoised_video
# Reconstruct full latents by replacing the video portion
if x_next_audio is not None:
# Flatten masked video back to match original format and replace video portion
x_next = x_next.clone()
x_next[:, :, :cut] = x_next_video.reshape([x_next_video.shape[0], x_next.shape[1], -1])
# Also reconstruct full denoised for next iteration input
denoised_full = denoised.clone()
denoised_full[:, :, :cut] = denoised_video.reshape([denoised_video.shape[0], denoised.shape[1], -1])
else:
# No audio separation
x_next = x_next_video
denoised_full = denoised_video
prev_certain_mask = certain_mask
prev_denoised = denoised_video
prev_denoised_full = denoised_full
prev_x_next_video = x_next_video
prev_x_next = x_next
elif use_stochastic:
# For first stochastic step, create denoised_full if we have audio
if x_next_audio is not None:
denoised_full = denoised.clone()
denoised_full[:, :, :cut] = denoised_video.reshape([denoised_video.shape[0], denoised.shape[1], -1])
else:
denoised_full = denoised_video
prev_certain_mask = None
prev_denoised = denoised_video
prev_denoised_full = denoised_full
prev_x_next_video = x_next_video
prev_x_next = x_next
# Update x for final step
if use_stochastic and ii == m - 1:
x = prev_x_next
elif not use_stochastic:
x = x_next
pbar.update(1)
if m == 1:
pbar.set_description("Sampling")
pbar.close()
return x
class SamplerSelfRefineVideo(io.ComfyNode):
@classmethod
def define_schema(cls):
default_ranges = [
(2, 5, 3), # Range 1
(6, 14, 1), # Range 2
]
options = []
# Option 1: 2 ranges
range_inputs_2 = []
for i in range(1, 3):
start_default, end_default, steps_default = default_ranges[i - 1]
range_inputs_2.extend([
io.Int.Input(f"start_step{i}", default=start_default, min=0, max=999, step=1, tooltip=f"Start step for range {i}"),
io.Int.Input(f"end_step{i}", default=end_default, min=0, max=999, step=1, tooltip=f"End step for range {i}"),
io.Int.Input(f"steps_{i}", default=steps_default, min=1, max=100, step=1, tooltip=f"Number of P&P steps for range {i}"),
])
options.append(io.DynamicCombo.Option(key="2 ranges", inputs=range_inputs_2))
# Option 2: 1 range
range_inputs_1 = []
for i in range(1, 2):
start_default, end_default, steps_default = default_ranges[i - 1]
range_inputs_1.extend([
io.Int.Input(f"start_step{i}", default=start_default, min=0, max=999, step=1, tooltip=f"Start step for range {i}"),
io.Int.Input(f"end_step{i}", default=end_default, min=0, max=999, step=1, tooltip=f"End step for range {i}"),
io.Int.Input(f"steps_{i}", default=steps_default, min=1, max=100, step=1, tooltip=f"Number of P&P steps for range {i}"),
])
options.append(io.DynamicCombo.Option(key="1 range", inputs=range_inputs_1))
# Option 3: Manual string input
options.append(io.DynamicCombo.Option(
key="from_string",
inputs=[
io.String.Input(
"stochastic_plan",
default="2-5:3,6-14:1",
multiline=True,
tooltip="Format: 'start-end:steps,start-end:steps' e.g. '2-5:3,6-14:1'"
)
]
))
return io.Schema(
node_id="SamplerSelfRefineVideo",
category="KJNodes/samplers",
description="Attempt to implement https://github.com/agwmon/self-refine-video, for testing only, MAY NOT WORK AS INTENDED.",
is_experimental=True,
inputs=[
io.DynamicCombo.Input("input_mode", options=options, tooltip="How to configure the step plan"),
io.Float.Input("certain_percentage", default=0.999, min=0.0, max=1.0, step=0.001, round=False, tooltip="Percentage of certain pixels to consider the frame as certain and skip further refinement"),
io.Float.Input("uncertainty_threshold", default=0.2, min=0.0, max=1.0, step=0.01, round=False, tooltip="Threshold of uncertainty to consider a pixel uncertain"),
io.Boolean.Input("verbose", default=False, tooltip="Enable verbose logging during sampling"),
io.Latent.Input("latent", optional=True, tooltip="Optional latent input to get input shape for LTX2 audio/video separation"),
io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, step=1, tooltip="Seed for stochastic sampling"),
],
outputs=[io.Sampler.Output()]
)
@classmethod
def execute(cls, input_mode, certain_percentage, uncertainty_threshold, seed, verbose, latent=None) -> io.NodeOutput:
video_shape = None
if latent is not None:
video_shape = latent["samples"].shape
range_keys = sorted([k for k in input_mode.keys() if k.startswith('start_step')])
stochastic_step_map = {}
if "stochastic_plan" in input_mode:
# Parse manual string format: "2-5:3,6-14:1"
plan_str = input_mode["stochastic_plan"]
ranges = plan_str.split(",")
for range_spec in ranges:
range_spec = range_spec.strip()
if not range_spec:
continue
try:
range_part, steps_part = range_spec.split(":")
start, end = range_part.split("-")
start, end, steps = int(start), int(end), int(steps_part)
for idx in range(start, end + 1):
stochastic_step_map[idx] = steps
except ValueError:
raise ValueError(f"Invalid format in stochastic_plan: '{range_spec}'. Expected format: 'start-end:steps'")
else:
range_keys = [k for k in input_mode.keys() if k.startswith('start_step')]
for start_key in range_keys:
i = start_key.replace('start_step', '')
start = input_mode.get(f"start_step{i}")
end = input_mode.get(f"end_step{i}")
steps = input_mode.get(f"steps_{i}")
if start is not None and end is not None and steps is not None:
for idx in range(start, end + 1):
stochastic_step_map[idx] = steps
sampler = KSAMPLER(sample_selfrefinevideo, {
"stochastic_step_map": stochastic_step_map,
"certain_percentage": certain_percentage,
"uncertainty_threshold": uncertainty_threshold,
"verbose": verbose,
"video_shape": video_shape,
"seed": seed,
})
return io.NodeOutput(sampler)
|